diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml index 5af497a3c..bc295d52d 100644 --- a/.github/workflows/build.yml +++ b/.github/workflows/build.yml @@ -288,6 +288,7 @@ jobs: OPENBLAS_VERSION: 0.3.23 OPENCL_VERSION: 2023.04.17 CLBLAST_VERSION: 1.6.0 + SDE_VERSION: 9.21.1-2023-04-24 strategy: matrix: @@ -383,11 +384,23 @@ jobs: - name: Test id: cmake_test - if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible + if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512 run: | cd build ctest -C Release --verbose --timeout 900 + - name: Test (Intel SDE) + id: cmake_test_sde + if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation + run: | + curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/777395/sde-external-${env:SDE_VERSION}-win.tar.xz" + # for some weird reason windows tar doesn't like sde tar.xz + 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz + 7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar + $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe) + cd build + & $sde -future -- ctest -C Release --verbose --timeout 900 + - name: Determine tag name id: tag shell: bash diff --git a/.github/workflows/python-lint.yml b/.github/workflows/python-lint.yml new file mode 100644 index 000000000..56d17b66c --- /dev/null +++ b/.github/workflows/python-lint.yml @@ -0,0 +1,20 @@ +name: flake8 Lint + +on: [push, pull_request] + +jobs: + flake8-lint: + runs-on: ubuntu-latest + name: Lint + steps: + - name: Check out source repository + uses: actions/checkout@v3 + - name: Set up Python environment + uses: actions/setup-python@v4 + with: + python-version: "3.11" + - name: flake8 Lint + uses: py-actions/flake8@v2 + with: + ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704" + exclude: "examples/*,examples/*/**,*/**/__init__.py" diff --git a/.gitignore b/.gitignore index 545c28726..41259a12f 100644 --- a/.gitignore +++ b/.gitignore @@ -15,6 +15,7 @@ .DS_Store .build/ .cache/ +.ccls-cache/ .direnv/ .envrc .swiftpm @@ -45,7 +46,7 @@ models-mnt /infill /libllama.so /llama-bench -/llava +/llava-cli /main /metal /perplexity @@ -63,8 +64,9 @@ models-mnt /speculative /parallel /train-text-from-scratch +/tokenize /vdot -build-info.h +/common/build-info.cpp arm_neon.h compile_commands.json CMakeSettings.json diff --git a/CMakeLists.txt b/CMakeLists.txt index 3659279e2..f32df5fe5 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -10,7 +10,7 @@ endif() set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin) -if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR) +if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR) set(LLAMA_STANDALONE ON) # configure project version @@ -100,39 +100,6 @@ option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALO option(LLAMA_BUILD_EXAMPLES "llama: build examples" ${LLAMA_STANDALONE}) option(LLAMA_BUILD_SERVER "llama: build server example" ON) -# -# Build info header -# - -# Generate initial build-info.h -include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake) - -if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git") - set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git") - - # Is git submodule - if(NOT IS_DIRECTORY "${GIT_DIR}") - file(READ ${GIT_DIR} REAL_GIT_DIR_LINK) - string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK}) - set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}") - endif() - - # Add a custom target for build-info.h - add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h") - - # Add a custom command to rebuild build-info.h when .git/index changes - add_custom_command( - OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h" - COMMENT "Generating build details from Git" - COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake" - WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} - DEPENDS "${GIT_DIR}/index" - VERBATIM - ) -else() - message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.") -endif() - # # Compile flags # @@ -491,6 +458,15 @@ if (LLAMA_LTO) endif() endif() +# this version of Apple ld64 is buggy +execute_process( + COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v + ERROR_VARIABLE output +) +if (output MATCHES "dyld-1015\.7") + add_compile_definitions(HAVE_BUGGY_APPLE_LINKER) +endif() + # Architecture specific # TODO: probably these flags need to be tweaked on some architectures # feel free to update the Makefile for your architecture and send a pull request or issue @@ -543,6 +519,10 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" ) message(STATUS "x86 detected") if (MSVC) + # instruction set detection for MSVC only + if (LLAMA_NATIVE) + include(cmake/FindSIMD.cmake) + endif () if (LLAMA_AVX512) add_compile_options($<$:/arch:AVX512>) add_compile_options($<$:/arch:AVX512>) @@ -594,8 +574,12 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE endif() elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64") message(STATUS "PowerPC detected") - add_compile_options(-mcpu=native -mtune=native) - #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) + if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le") + add_compile_options(-mcpu=powerpc64le) + else() + add_compile_options(-mcpu=native -mtune=native) + #TODO: Add targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be) + endif() else() message(STATUS "Unknown architecture") endif() diff --git a/Makefile b/Makefile index c53c1e726..a6d2c2ec0 100644 --- a/Makefile +++ b/Makefile @@ -1,8 +1,8 @@ # Define the default target now so that it is always the first target BUILD_TARGETS = \ main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ - simple batched batched-bench save-load-state server gguf llama-bench llava baby-llama beam-search \ - speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o + simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search \ + speculative infill tokenize benchmark-matmult parallel finetune export-lora tests/test-c.o # Binaries only useful for tests TEST_TARGETS = \ @@ -239,6 +239,11 @@ else endif endif +# this version of Apple ld64 is buggy +ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))' + MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER +endif + # OS specific # TODO: support Windows ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)' @@ -337,6 +342,12 @@ ifneq ($(filter ppc64%,$(UNAME_M)),) endif endif +ifneq ($(filter ppc64le%,$(UNAME_M)),) + MK_CFLAGS += -mcpu=powerpc64le + MK_CXXFLAGS += -mcpu=powerpc64le + CUDA_POWER_ARCH = 1 +endif + else MK_CFLAGS += -march=rv64gcv -mabi=lp64d MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d @@ -387,6 +398,8 @@ else endif #LLAMA_CUDA_NVCC ifdef CUDA_DOCKER_ARCH NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH) +else ifdef CUDA_POWER_ARCH + NVCCFLAGS += else NVCCFLAGS += -arch=native endif # CUDA_DOCKER_ARCH @@ -542,9 +555,9 @@ llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h l $(CXX) $(CXXFLAGS) -c $< -o $@ COMMON_H_DEPS = common/common.h common/sampling.h common/log.h -COMMON_DEPS = common.o sampling.o grammar-parser.o +COMMON_DEPS = common.o sampling.o grammar-parser.o build-info.o -common.o: common/common.cpp build-info.h $(COMMON_H_DEPS) +common.o: common/common.cpp $(COMMON_H_DEPS) $(CXX) $(CXXFLAGS) -c $< -o $@ sampling.o: common/sampling.cpp $(COMMON_H_DEPS) @@ -563,46 +576,49 @@ libllama.so: llama.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) clean: - rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS) + rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS) # # Examples # -main: examples/main/main.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) +main: examples/main/main.cpp ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) @echo @echo '==== Run ./main -h for help. ====' @echo -infill: examples/infill/infill.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) +infill: examples/infill/infill.cpp ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -simple: examples/simple/simple.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +simple: examples/simple/simple.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -batched: examples/batched/batched.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tokenize: examples/tokenize/tokenize.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -batched-bench: examples/batched-bench/batched-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS) +batched: examples/batched/batched.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS) +batched-bench: examples/batched-bench/batched-bench.cpp build-info.o ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.h ggml.o llama.o $(OBJS) +quantize: examples/quantize/quantize.cpp build-info.o ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -perplexity: examples/perplexity/perplexity.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +quantize-stats: examples/quantize-stats/quantize-stats.cpp build-info.o ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -embedding: examples/embedding/embedding.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +embedding: examples/embedding/embedding.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + +server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS) @@ -614,28 +630,31 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual + +llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS) +finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) ifdef LLAMA_METAL @@ -648,7 +667,7 @@ swift: examples/batched.swift (cd examples/batched.swift; make build) endif -build-info.h: $(wildcard .git/index) scripts/build-info.sh +common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh @sh scripts/build-info.sh $(CC) > $@.tmp @if ! cmp -s $@.tmp $@; then \ mv $@.tmp $@; \ @@ -656,13 +675,16 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh rm $@.tmp; \ fi +build-info.o: common/build-info.cpp + $(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@ + # # Tests # tests: $(TEST_TARGETS) -benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS) +benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) run-benchmark-matmult: benchmark-matmult @@ -676,40 +698,40 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS) q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) -tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) +tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS) +tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) tests/test-c.o: tests/test-c.c llama.h diff --git a/README.md b/README.md index b56ecaec7..5189e1255 100644 --- a/README.md +++ b/README.md @@ -2,7 +2,6 @@ ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png) -[![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions) [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml) @@ -11,8 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++ ### Hot topics -- LLaVA support: https://github.com/ggerganov/llama.cpp/pull/3436 -- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252) +- Collecting Apple Silicon performance stats: https://github.com/ggerganov/llama.cpp/discussions/4167 ---- @@ -95,6 +93,7 @@ as the main playground for developing new features for the [ggml](https://github - [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410) - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417) - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553) +- [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586) **Bindings:** @@ -411,22 +410,31 @@ Building the program with BLAS support may lead to some performance improvements This provides BLAS acceleration on HIP-supported AMD GPUs. Make sure to have ROCm installed. You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html). - Windows support is coming soon... - Using `make`: ```bash make LLAMA_HIPBLAS=1 ``` - - Using `CMake`: + - Using `CMake` for Linux: ```bash mkdir build cd build CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON cmake --build . ``` + - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS): + ```bash + set PATH=%HIP_PATH%\bin;%PATH% + mkdir build + cd build + cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ .. + cmake --build . + ``` + Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors) + The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used. - If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3. + If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3. The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above): | Option | Legal values | Default | Description | diff --git a/build.zig b/build.zig index 9b58b74ca..699738f3d 100644 --- a/build.zig +++ b/build.zig @@ -10,7 +10,6 @@ const Maker = struct { builder: *std.build.Builder, target: CrossTarget, optimize: Mode, - config_header: *ConfigHeader, enable_lto: bool, include_dirs: ArrayList([]const u8), @@ -41,26 +40,24 @@ const Maker = struct { const commit_hash = try std.ChildProcess.exec( .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } }, ); - const config_header = builder.addConfigHeader( - .{ .style = .blank, .include_path = "build-info.h" }, - .{ - .BUILD_NUMBER = 0, - .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline - .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}), - .BUILD_TARGET = try target.allocDescription(builder.allocator), - }, - ); + try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt( + \\int LLAMA_BUILD_NUMBER = {}; + \\char const *LLAMA_COMMIT = "{s}"; + \\char const *LLAMA_COMPILER = "Zig {s}"; + \\char const *LLAMA_BUILD_TARGET = "{s}"; + \\ + , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) })); var m = Maker{ .builder = builder, .target = target, .optimize = builder.standardOptimizeOption(.{}), - .config_header = config_header, .enable_lto = false, .include_dirs = ArrayList([]const u8).init(builder.allocator), .cflags = ArrayList([]const u8).init(builder.allocator), .cxxflags = ArrayList([]const u8).init(builder.allocator), .objs = ArrayList(*Compile).init(builder.allocator), }; + try m.addCFlag("-std=c11"); try m.addCxxFlag("-std=c++11"); try m.addProjectInclude(&.{}); @@ -72,7 +69,7 @@ const Maker = struct { const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize }); if (o.target.getAbi() != .msvc) o.defineCMacro("_GNU_SOURCE", null); - o.addConfigHeader(m.config_header); + if (std.mem.endsWith(u8, src, ".c")) { o.addCSourceFiles(&.{src}, m.cflags.items); o.linkLibC(); @@ -85,7 +82,6 @@ const Maker = struct { o.linkLibCpp(); } } - o.addConfigHeader(m.config_header); for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i }); o.want_lto = m.enable_lto; return o; @@ -105,7 +101,6 @@ const Maker = struct { // linkLibCpp already add (libc++ + libunwind + libc) e.linkLibCpp(); } - e.addConfigHeader(m.config_header); m.builder.installArtifact(e); e.want_lto = m.enable_lto; return e; @@ -121,6 +116,7 @@ pub fn build(b: *std.build.Builder) !void { const ggml_backend = make.obj("ggml-backend", "ggml-backend.c"); const ggml_quants = make.obj("ggml-quants", "ggml-quants.c"); const llama = make.obj("llama", "llama.cpp"); + const buildinfo = make.obj("common", "common/build-info.cpp"); const common = make.obj("common", "common/common.cpp"); const console = make.obj("console", "common/console.cpp"); const sampling = make.obj("sampling", "common/sampling.cpp"); @@ -128,14 +124,14 @@ pub fn build(b: *std.build.Builder) !void { const train = make.obj("train", "common/train.cpp"); const clip = make.obj("clip", "examples/llava/clip.cpp"); - _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, console, grammar_parser }); - _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); - _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); - _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common }); - _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train }); - _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train }); + _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser }); + _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo }); + _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo }); + _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo }); + _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train }); + _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train }); - const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip }); + const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip }); if (server.target.isWindows()) { server.linkSystemLibrary("ws2_32"); } diff --git a/cmake/FindSIMD.cmake b/cmake/FindSIMD.cmake new file mode 100644 index 000000000..33377ec44 --- /dev/null +++ b/cmake/FindSIMD.cmake @@ -0,0 +1,100 @@ +include(CheckCSourceRuns) + +set(AVX_CODE " + #include + int main() + { + __m256 a; + a = _mm256_set1_ps(0); + return 0; + } +") + +set(AVX512_CODE " + #include + int main() + { + __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0, + 0, 0, 0, 0, 0, 0, 0, 0); + __m512i b = a; + __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ); + return 0; + } +") + +set(AVX2_CODE " + #include + int main() + { + __m256i a = {0}; + a = _mm256_abs_epi16(a); + __m256i x; + _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code + return 0; + } +") + +set(FMA_CODE " + #include + int main() + { + __m256 acc = _mm256_setzero_ps(); + const __m256 d = _mm256_setzero_ps(); + const __m256 p = _mm256_setzero_ps(); + acc = _mm256_fmadd_ps( d, p, acc ); + return 0; + } +") + +macro(check_sse type flags) + set(__FLAG_I 1) + set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS}) + foreach (__FLAG ${flags}) + if (NOT ${type}_FOUND) + set(CMAKE_REQUIRED_FLAGS ${__FLAG}) + check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I}) + if (HAS_${type}_${__FLAG_I}) + set(${type}_FOUND TRUE CACHE BOOL "${type} support") + set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags") + endif() + math(EXPR __FLAG_I "${__FLAG_I}+1") + endif() + endforeach() + set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE}) + + if (NOT ${type}_FOUND) + set(${type}_FOUND FALSE CACHE BOOL "${type} support") + set(${type}_FLAGS "" CACHE STRING "${type} flags") + endif() + + mark_as_advanced(${type}_FOUND ${type}_FLAGS) +endmacro() + +# flags are for MSVC only! +check_sse("AVX" " ;/arch:AVX") +if (NOT ${AVX_FOUND}) + set(LLAMA_AVX OFF) +else() + set(LLAMA_AVX ON) +endif() + +check_sse("AVX2" " ;/arch:AVX2") +check_sse("FMA" " ;/arch:AVX2") +if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND})) + set(LLAMA_AVX2 OFF) +else() + set(LLAMA_AVX2 ON) +endif() + +check_sse("AVX512" " ;/arch:AVX512") +if (NOT ${AVX512_FOUND}) + set(LLAMA_AVX512 OFF) +else() + set(LLAMA_AVX512 ON) +endif() diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt index fbb0ff095..4f930bdc5 100644 --- a/common/CMakeLists.txt +++ b/common/CMakeLists.txt @@ -1,8 +1,47 @@ # common + +# Build info header +# + +if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git") + set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git") + + # Is git submodule + if(NOT IS_DIRECTORY "${GIT_DIR}") + file(READ ${GIT_DIR} REAL_GIT_DIR_LINK) + string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK}) + set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}") + endif() + + set(GIT_INDEX "${GIT_DIR}/index") +else() + message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.") + set(GIT_INDEX "") +endif() + +# Add a custom command to rebuild build-info.cpp when .git/index changes +add_custom_command( + OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp" + COMMENT "Generating build details from Git" + COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} + -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} + -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake" + WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.." + DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX} + VERBATIM +) +set(TARGET build_info) +add_library(${TARGET} OBJECT build-info.cpp) +if (BUILD_SHARED_LIBS) + set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON) +endif() + + set(TARGET common) -add_library(${TARGET} OBJECT +add_library(${TARGET} STATIC + base64.hpp common.h common.cpp sampling.h @@ -21,4 +60,4 @@ endif() target_include_directories(${TARGET} PUBLIC .) target_compile_features(${TARGET} PUBLIC cxx_std_11) -target_link_libraries(${TARGET} PRIVATE llama) +target_link_libraries(${TARGET} PRIVATE llama build_info) diff --git a/common/base64.hpp b/common/base64.hpp new file mode 100644 index 000000000..563247a6e --- /dev/null +++ b/common/base64.hpp @@ -0,0 +1,392 @@ +/* +This is free and unencumbered software released into the public domain. + +Anyone is free to copy, modify, publish, use, compile, sell, or +distribute this software, either in source code form or as a compiled +binary, for any purpose, commercial or non-commercial, and by any +means. + +In jurisdictions that recognize copyright laws, the author or authors +of this software dedicate any and all copyright interest in the +software to the public domain. We make this dedication for the benefit +of the public at large and to the detriment of our heirs and +successors. We intend this dedication to be an overt act of +relinquishment in perpetuity of all present and future rights to this +software under copyright law. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, +EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF +MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. +IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR +OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, +ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR +OTHER DEALINGS IN THE SOFTWARE. + +For more information, please refer to +*/ + +#ifndef PUBLIC_DOMAIN_BASE64_HPP_ +#define PUBLIC_DOMAIN_BASE64_HPP_ + +#include +#include +#include +#include + +class base64_error : public std::runtime_error +{ +public: + using std::runtime_error::runtime_error; +}; + +class base64 +{ +public: + enum class alphabet + { + /** the alphabet is detected automatically */ + auto_, + /** the standard base64 alphabet is used */ + standard, + /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/ + url_filename_safe + }; + + enum class decoding_behavior + { + /** if the input is not padded, the remaining bits are ignored */ + moderate, + /** if a padding character is encounter decoding is finished */ + loose + }; + + /** + Encodes all the elements from `in_begin` to `in_end` to `out`. + + @warning The source and destination cannot overlap. The destination must be able to hold at least + `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator. + + @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than + 8 bits + @tparam Output_iterator the destination; the elements written to it are from the type `char` + @param in_begin the beginning of the source + @param in_end the ending of the source + @param out the destination iterator + @param alphabet which alphabet should be used + @returns the iterator to the next element past the last element copied + @throws see `Input_iterator` and `Output_iterator` + */ + template + static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out, + alphabet alphabet = alphabet::standard) + { + constexpr auto pad = '='; + const char* alpha = alphabet == alphabet::url_filename_safe + ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_" + : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/"; + + while (in_begin != in_end) { + std::uint8_t i0 = 0, i1 = 0, i2 = 0; + + // first character + i0 = static_cast(*in_begin); + ++in_begin; + + *out = alpha[i0 >> 2 & 0x3f]; + ++out; + + // part of first character and second + if (in_begin != in_end) { + i1 = static_cast(*in_begin); + ++in_begin; + + *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)]; + ++out; + } else { + *out = alpha[(i0 & 0x3) << 4]; + ++out; + + // last padding + *out = pad; + ++out; + + // last padding + *out = pad; + ++out; + + break; + } + + // part of second character and third + if (in_begin != in_end) { + i2 = static_cast(*in_begin); + ++in_begin; + + *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)]; + ++out; + } else { + *out = alpha[(i1 & 0xf) << 2]; + ++out; + + // last padding + *out = pad; + ++out; + + break; + } + + // rest of third + *out = alpha[i2 & 0x3f]; + ++out; + } + + return out; + } + /** + Encodes a string. + + @param str the string that should be encoded + @param alphabet which alphabet should be used + @returns the encoded base64 string + @throws see base64::encode() + */ + static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard) + { + std::string result; + + result.reserve(required_encode_size(str.length()) + 1); + + encode(str.begin(), str.end(), std::back_inserter(result), alphabet); + + return result; + } + /** + Encodes a char array. + + @param buffer the char array + @param size the size of the array + @param alphabet which alphabet should be used + @returns the encoded string + */ + static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard) + { + std::string result; + + result.reserve(required_encode_size(size) + 1); + + encode(buffer, buffer + size, std::back_inserter(result), alphabet); + + return result; + } + /** + Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`, + in other words: inplace decoding is possible. + + @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`, + otherwise the behavior depends on the output iterator. + + @tparam Input_iterator the source; the returned elements are cast to `char` + @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t` + @param in_begin the beginning of the source + @param in_end the ending of the source + @param out the destination iterator + @param alphabet which alphabet should be used + @param behavior the behavior when an error was detected + @returns the iterator to the next element past the last element copied + @throws base64_error depending on the set behavior + @throws see `Input_iterator` and `Output_iterator` + */ + template + static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out, + alphabet alphabet = alphabet::auto_, + decoding_behavior behavior = decoding_behavior::moderate) + { + //constexpr auto pad = '='; + std::uint8_t last = 0; + auto bits = 0; + + while (in_begin != in_end) { + auto c = *in_begin; + ++in_begin; + + if (c == '=') { + break; + } + + auto part = _base64_value(alphabet, c); + + // enough bits for one byte + if (bits + 6 >= 8) { + *out = (last << (8 - bits)) | (part >> (bits - 2)); + ++out; + + bits -= 2; + } else { + bits += 6; + } + + last = part; + } + + // check padding + if (behavior != decoding_behavior::loose) { + while (in_begin != in_end) { + auto c = *in_begin; + ++in_begin; + + if (c != '=') { + throw base64_error("invalid base64 character."); + } + } + } + + return out; + } + /** + Decodes a string. + + @param str the base64 encoded string + @param alphabet which alphabet should be used + @param behavior the behavior when an error was detected + @returns the decoded string + @throws see base64::decode() + */ + static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_, + decoding_behavior behavior = decoding_behavior::moderate) + { + std::string result; + + result.reserve(max_decode_size(str.length())); + + decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior); + + return result; + } + /** + Decodes a string. + + @param buffer the base64 encoded buffer + @param size the size of the buffer + @param alphabet which alphabet should be used + @param behavior the behavior when an error was detected + @returns the decoded string + @throws see base64::decode() + */ + static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_, + decoding_behavior behavior = decoding_behavior::moderate) + { + std::string result; + + result.reserve(max_decode_size(size)); + + decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior); + + return result; + } + /** + Decodes a string inplace. + + @param[in,out] str the base64 encoded string + @param alphabet which alphabet should be used + @param behavior the behavior when an error was detected + @throws base64::decode_inplace() + */ + static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_, + decoding_behavior behavior = decoding_behavior::moderate) + { + str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin()); + } + /** + Decodes a char array inplace. + + @param[in,out] str the string array + @param size the length of the array + @param alphabet which alphabet should be used + @param behavior the behavior when an error was detected + @returns the pointer to the next element past the last element decoded + @throws base64::decode_inplace() + */ + static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_, + decoding_behavior behavior = decoding_behavior::moderate) + { + return decode(str, str + size, str, alphabet, behavior); + } + /** + Returns the required decoding size for a given size. The value is calculated with the following formula: + + $$ + \lceil \frac{size}{4} \rceil \cdot 3 + $$ + + @param size the size of the encoded input + @returns the size of the resulting decoded buffer; this the absolute maximum + */ + static std::size_t max_decode_size(std::size_t size) noexcept + { + return (size / 4 + (size % 4 ? 1 : 0)) * 3; + } + /** + Returns the required encoding size for a given size. The value is calculated with the following formula: + + $$ + \lceil \frac{size}{3} \rceil \cdot 4 + $$ + + @param size the size of the decoded input + @returns the size of the resulting encoded buffer + */ + static std::size_t required_encode_size(std::size_t size) noexcept + { + return (size / 3 + (size % 3 ? 1 : 0)) * 4; + } + +private: + static std::uint8_t _base64_value(alphabet& alphabet, char c) + { + if (c >= 'A' && c <= 'Z') { + return c - 'A'; + } else if (c >= 'a' && c <= 'z') { + return c - 'a' + 26; + } else if (c >= '0' && c <= '9') { + return c - '0' + 52; + } + + // comes down to alphabet + if (alphabet == alphabet::standard) { + if (c == '+') { + return 62; + } else if (c == '/') { + return 63; + } + } else if (alphabet == alphabet::url_filename_safe) { + if (c == '-') { + return 62; + } else if (c == '_') { + return 63; + } + } // auto detect + else { + if (c == '+') { + alphabet = alphabet::standard; + + return 62; + } else if (c == '/') { + alphabet = alphabet::standard; + + return 63; + } else if (c == '-') { + alphabet = alphabet::url_filename_safe; + + return 62; + } else if (c == '_') { + alphabet = alphabet::url_filename_safe; + + return 63; + } + } + + throw base64_error("invalid base64 character."); + } +}; + +#endif // !PUBLIC_DOMAIN_BASE64_HPP_ diff --git a/common/build-info.cpp.in b/common/build-info.cpp.in new file mode 100644 index 000000000..0b945aa68 --- /dev/null +++ b/common/build-info.cpp.in @@ -0,0 +1,4 @@ +int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@; +char const *LLAMA_COMMIT = "@BUILD_COMMIT@"; +char const *LLAMA_COMPILER = "@BUILD_COMPILER@"; +char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@"; diff --git a/common/common.cpp b/common/common.cpp index c187128d6..1dcc235ea 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1,5 +1,4 @@ #include "common.h" -#include "build-info.h" #include "llama.h" #include @@ -13,6 +12,7 @@ #include #include #include +#include #include #include #include @@ -91,6 +91,19 @@ void process_escapes(std::string& input) { case '\'': input[output_idx++] = '\''; break; case '\"': input[output_idx++] = '\"'; break; case '\\': input[output_idx++] = '\\'; break; + case 'x': + // Handle \x12, etc + if (input_idx + 2 < input_len) { + const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 }; + char *err_p = nullptr; + const long val = std::strtol(x, &err_p, 16); + if (err_p == x + 2) { + input_idx += 2; + input[output_idx++] = char(val); + break; + } + } + // fall through default: input[output_idx++] = '\\'; input[output_idx++] = input[input_idx]; break; } @@ -103,9 +116,24 @@ void process_escapes(std::string& input) { } bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { + bool result = true; + try { + if (!gpt_params_parse_ex(argc, argv, params)) { + gpt_print_usage(argc, argv, gpt_params()); + exit(0); + } + } + catch (const std::invalid_argument & ex) { + fprintf(stderr, "%s\n", ex.what()); + gpt_print_usage(argc, argv, gpt_params()); + exit(1); + } + return result; +} + +bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) { bool invalid_param = false; std::string arg; - gpt_params default_params; const std::string arg_prefix = "--"; llama_sampling_params & sparams = params.sparams; @@ -204,12 +232,52 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.rope_freq_scale = std::stof(argv[i]); + } else if (arg == "--rope-scaling") { + if (++i >= argc) { + invalid_param = true; + break; + } + std::string value(argv[i]); + /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; } + else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; } + else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; } + else { invalid_param = true; break; } } else if (arg == "--rope-scale") { if (++i >= argc) { invalid_param = true; break; } params.rope_freq_scale = 1.0f/std::stof(argv[i]); + } else if (arg == "--yarn-orig-ctx") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_orig_ctx = std::stoi(argv[i]); + } else if (arg == "--yarn-ext-factor") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_ext_factor = std::stof(argv[i]); + } else if (arg == "--yarn-attn-factor") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_attn_factor = std::stof(argv[i]); + } else if (arg == "--yarn-beta-fast") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_beta_fast = std::stof(argv[i]); + } else if (arg == "--yarn-beta-slow") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_beta_slow = std::stof(argv[i]); } else if (arg == "--memory-f32") { params.memory_f16 = false; } else if (arg == "--top-p") { @@ -218,6 +286,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } sparams.top_p = std::stof(argv[i]); + } else if (arg == "--min-p") { + if (++i >= argc) { + invalid_param = true; + break; + } + sparams.min_p = std::stof(argv[i]); } else if (arg == "--temp") { if (++i >= argc) { invalid_param = true; @@ -343,6 +417,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } params.n_sequences = std::stoi(argv[i]); + } else if (arg == "--p-accept" || arg == "-pa") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.p_accept = std::stof(argv[i]); + } else if (arg == "--p-split" || arg == "-ps") { + if (++i >= argc) { + invalid_param = true; + break; + } + params.p_split = std::stof(argv[i]); } else if (arg == "-m" || arg == "--model") { if (++i >= argc) { invalid_param = true; @@ -406,8 +492,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { params.interactive_first = true; } else if (arg == "-ins" || arg == "--instruct") { params.instruct = true; + } else if (arg == "-cml" || arg == "--chatml") { + params.chatml = true; } else if (arg == "--infill") { params.infill = true; + } else if (arg == "-dkvc" || arg == "--dump-kv-cache") { + params.dump_kv_cache = true; } else if (arg == "--multiline-input") { params.multiline_input = true; } else if (arg == "--simple-io") { @@ -548,11 +638,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { break; } } else if (arg == "-h" || arg == "--help") { - gpt_print_usage(argc, argv, default_params); -#ifndef LOG_DISABLE_LOGS - log_print_usage(); -#endif // LOG_DISABLE_LOGS - exit(0); + return false; + } else if (arg == "--random-prompt") { params.random_prompt = true; } else if (arg == "--in-prefix-bos") { @@ -611,22 +698,17 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { // End of Parse args for logging parameters #endif // LOG_DISABLE_LOGS } else { - fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); - gpt_print_usage(argc, argv, default_params); - exit(1); + throw std::invalid_argument("error: unknown argument: " + arg); } } if (invalid_param) { - fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str()); - gpt_print_usage(argc, argv, default_params); - exit(1); + throw std::invalid_argument("error: invalid parameter for argument: " + arg); } if (params.prompt_cache_all && (params.interactive || params.interactive_first || params.instruct)) { - fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n"); - gpt_print_usage(argc, argv, default_params); - exit(1); + + throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n"); } if (params.escape) { @@ -645,6 +727,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) { void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { const llama_sampling_params & sparams = params.sparams; + printf("\n"); printf("usage: %s [options]\n", argv[0]); printf("\n"); printf("options:\n"); @@ -652,6 +735,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -i, --interactive run in interactive mode\n"); printf(" --interactive-first run in interactive mode and wait for input right away\n"); printf(" -ins, --instruct run in instruction mode (use with Alpaca models)\n"); + printf(" -cml, --chatml run in chatml mode (use with ChatML-compatible models)\n"); printf(" --multiline-input allows you to write or paste multiple lines without ending each in '\\'\n"); printf(" -r PROMPT, --reverse-prompt PROMPT\n"); printf(" halt generation at PROMPT, return control in interactive mode\n"); @@ -679,6 +763,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); printf(" --top-k N top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k); printf(" --top-p N top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p); + printf(" --min-p N min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p); printf(" --tfs N tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z); printf(" --typical N locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p); printf(" --repeat-last-n N last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n); @@ -701,9 +786,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --cfg-negative-prompt-file FNAME\n"); printf(" negative prompt file to use for guidance. (default: empty)\n"); printf(" --cfg-scale N strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale); - printf(" --rope-scale N RoPE context linear scaling factor, inverse of --rope-freq-scale\n"); + printf(" --rope-scaling {none,linear,yarn}\n"); + printf(" RoPE frequency scaling method, defaults to linear unless specified by the model\n"); + printf(" --rope-scale N RoPE context scaling factor, expands context by a factor of N\n"); printf(" --rope-freq-base N RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n"); - printf(" --rope-freq-scale N RoPE frequency linear scaling factor (default: loaded from model)\n"); + printf(" --rope-freq-scale N RoPE frequency scaling factor, expands context by a factor of 1/N\n"); + printf(" --yarn-orig-ctx N YaRN: original context size of model (default: 0 = model training context size)\n"); + printf(" --yarn-ext-factor N YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n"); + printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n"); + printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow); + printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast); printf(" --ignore-eos ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n"); printf(" --no-penalize-nl do not penalize newline token\n"); printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n"); @@ -717,6 +809,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" --chunks N max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks); printf(" -np N, --parallel N number of parallel sequences to decode (default: %d)\n", params.n_parallel); printf(" -ns N, --sequences N number of sequences to decode (default: %d)\n", params.n_sequences); + printf(" -pa N, --p-accept N speculative decoding accept probability (default: %.1f)\n", (double)params.p_accept); + printf(" -ps N, --p-split N speculative decoding split probability (default: %.1f)\n", (double)params.p_split); printf(" -cb, --cont-batching enable continuous batching (a.k.a dynamic batching) (default: disabled)\n"); printf(" --mmproj MMPROJ_FILE path to a multimodal projector file for LLaVA. see examples/llava/README.md\n"); printf(" --image IMAGE_FILE path to an image file. use with multimodal models\n"); @@ -744,6 +838,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { #endif // GGML_USE_CUBLAS #endif printf(" --verbose-prompt print prompt before generation\n"); + printf(" -dkvc, --dump-kv-cache\n"); + printf(" verbose print of the KV cache\n"); printf(" --simple-io use basic IO for better compatibility in subprocesses and limited consoles\n"); printf(" --lora FNAME apply LoRA adapter (implies --no-mmap)\n"); printf(" --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n"); @@ -755,6 +851,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) { printf(" -ld LOGDIR, --logdir LOGDIR\n"); printf(" path under which to save YAML logs (no logging if unset)\n"); printf("\n"); +#ifndef LOG_DISABLE_LOGS + log_print_usage(); +#endif // LOG_DISABLE_LOGS } std::string get_system_info(const gpt_params & params) { @@ -808,17 +907,23 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params & struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) { auto cparams = llama_context_default_params(); - cparams.n_ctx = params.n_ctx; - cparams.n_batch = params.n_batch; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; - cparams.mul_mat_q = params.mul_mat_q; - cparams.seed = params.seed; - cparams.f16_kv = params.memory_f16; - cparams.logits_all = params.logits_all; - cparams.embedding = params.embedding; - cparams.rope_freq_base = params.rope_freq_base; - cparams.rope_freq_scale = params.rope_freq_scale; + cparams.n_ctx = params.n_ctx; + cparams.n_batch = params.n_batch; + cparams.n_threads = params.n_threads; + cparams.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; + cparams.mul_mat_q = params.mul_mat_q; + cparams.seed = params.seed; + cparams.f16_kv = params.memory_f16; + cparams.logits_all = params.logits_all; + cparams.embedding = params.embedding; + cparams.rope_scaling_type = params.rope_scaling_type; + cparams.rope_freq_base = params.rope_freq_base; + cparams.rope_freq_scale = params.rope_freq_scale; + cparams.yarn_ext_factor = params.yarn_ext_factor; + cparams.yarn_attn_factor = params.yarn_attn_factor; + cparams.yarn_beta_fast = params.yarn_beta_fast; + cparams.yarn_beta_slow = params.yarn_beta_slow; + cparams.yarn_orig_ctx = params.yarn_orig_ctx; return cparams; } @@ -834,7 +939,7 @@ void llama_batch_add( const std::vector & seq_ids, bool logits) { batch.token [batch.n_tokens] = id; - batch.pos [batch.n_tokens] = pos, + batch.pos [batch.n_tokens] = pos; batch.n_seq_id[batch.n_tokens] = seq_ids.size(); for (size_t i = 0; i < seq_ids.size(); ++i) { batch.seq_id[batch.n_tokens][i] = seq_ids[i]; @@ -975,6 +1080,12 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector & prompt_tokens, const char * model_desc) { const llama_sampling_params & sparams = params.sparams; - fprintf(stream, "build_commit: %s\n", BUILD_COMMIT); - fprintf(stream, "build_number: %d\n", BUILD_NUMBER); + fprintf(stream, "build_commit: %s\n", LLAMA_COMMIT); + fprintf(stream, "build_number: %d\n", LLAMA_BUILD_NUMBER); fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false"); fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false"); fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false"); @@ -1275,6 +1387,81 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency()); fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k); fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p); + fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p); fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p); fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false"); } + +// +// KV cache utils +// + +void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) { + static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+"; + + printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d", + view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx); + + llama_kv_cache_view_cell * c_curr = view.cells; + llama_seq_id * cs_curr = view.cells_sequences; + + for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) { + if (i % row_size == 0) { + printf("\n%5d: ", i); + } + int seq_count = 0; + for (int j = 0; j < view.n_max_seq; j++) { + if (cs_curr[j] >= 0) { seq_count++; } + } + putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]); + } + + printf("\n=== Done dumping\n"); +} + +void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) { + static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"; + + printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n", + view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx); + + std::unordered_map seqs; + llama_kv_cache_view_cell * c_curr = view.cells; + llama_seq_id * cs_curr = view.cells_sequences; + + for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) { + for (int j = 0; j < view.n_max_seq; j++) { + if (cs_curr[j] < 0) { continue; } + if (seqs.find(cs_curr[j]) == seqs.end()) { + if (seqs.size() + 1 >= sizeof(slot_chars)) { break; } + seqs[cs_curr[j]] = seqs.size(); + } + } + if (seqs.size() + 1 >= sizeof(slot_chars)) { break; } + } + + printf("=== Sequence legend: "); + for (const auto & it : seqs) { + printf("%zu=%d, ", it.second, it.first); + } + printf("'+'=other sequence ids"); + + c_curr = view.cells; + cs_curr = view.cells_sequences; + for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) { + if (i % row_size == 0) { + printf("\n%5d: ", i); + } + for (int j = 0; j < view.n_max_seq; j++) { + if (cs_curr[j] >= 0) { + const auto & it = seqs.find(cs_curr[j]); + putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+'); + } else { + putchar('.'); + } + } + putchar(' '); + } + + printf("\n=== Done dumping\n"); +} diff --git a/common/common.h b/common/common.h index 84523a4fb..2f6fe48ab 100644 --- a/common/common.h +++ b/common/common.h @@ -9,6 +9,7 @@ #define LOG_NO_FILE_LINE_FUNCTION #include "log.h" +#include #include #include #include @@ -25,35 +26,51 @@ #define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0) #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0) -#define print_build_info() do { \ - fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); \ - fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET); \ +#define print_build_info() do { \ + fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT); \ + fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET); \ } while(0) +// build info +extern int LLAMA_BUILD_NUMBER; +extern char const *LLAMA_COMMIT; +extern char const *LLAMA_COMPILER; +extern char const *LLAMA_BUILD_TARGET; + // // CLI argument parsing // int32_t get_num_physical_cores(); struct gpt_params { - uint32_t seed = -1; // RNG seed + uint32_t seed = -1; // RNG seed + int32_t n_threads = get_num_physical_cores(); - int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads) - int32_t n_predict = -1; // new tokens to predict - int32_t n_ctx = 512; // context size - int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) - int32_t n_keep = 0; // number of tokens to keep from initial prompt - int32_t n_draft = 16; // number of tokens to draft during speculative decoding - int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited) - int32_t n_parallel = 1; // number of parallel sequences to decode - int32_t n_sequences = 1; // number of sequences to decode - int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default) - int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default) - int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors - float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs - int32_t n_beams = 0; // if non-zero then use beam search of given width. - float rope_freq_base = 0.0f; // RoPE base frequency - float rope_freq_scale = 0.0f; // RoPE frequency scaling factor + int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads) + int32_t n_predict = -1; // new tokens to predict + int32_t n_ctx = 512; // context size + int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS) + int32_t n_keep = 0; // number of tokens to keep from initial prompt + int32_t n_draft = 16; // number of tokens to draft during speculative decoding + int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited) + int32_t n_parallel = 1; // number of parallel sequences to decode + int32_t n_sequences = 1; // number of sequences to decode + float p_accept = 0.5f; // speculative decoding accept probability + float p_split = 0.1f; // speculative decoding split probability + int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default) + int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default) + int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors + float tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs + int32_t n_beams = 0; // if non-zero then use beam search of given width. + float rope_freq_base = 0.0f; // RoPE base frequency + float rope_freq_scale = 0.0f; // RoPE frequency scaling factor + float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor + float yarn_attn_factor = 1.0f; // YaRN magnitude scaling factor + float yarn_beta_fast = 32.0f; // YaRN low correction dim + float yarn_beta_slow = 1.0f; // YaRN high correction dim + int32_t yarn_orig_ctx = 0; // YaRN original context length + int8_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment + // pinging @cebtenzzre // // sampling parameters struct llama_sampling_params sparams; @@ -77,7 +94,7 @@ struct gpt_params { int ppl_output_type = 0; // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line // (which is more convenient to use for plotting) // - bool hellaswag = false; // compute HellaSwag score over random tasks from datafile supplied in prompt + bool hellaswag = false; // compute HellaSwag score over random tasks from datafile supplied in prompt size_t hellaswag_tasks = 400; // number of tasks to use when computing the HellaSwag score bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS @@ -85,6 +102,7 @@ struct gpt_params { bool random_prompt = false; // do not randomize prompt if none provided bool use_color = false; // use color to distinguish generations and inputs bool interactive = false; // interactive mode + bool chatml = false; // chatml mode (used for models trained on chatml syntax) bool prompt_cache_all = false; // save user input and generations to prompt cache bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it @@ -104,12 +122,15 @@ struct gpt_params { bool numa = false; // attempt optimizations that help on some NUMA systems bool verbose_prompt = false; // print prompt tokens before generation bool infill = false; // use infill mode + bool dump_kv_cache = false; // dump the KV cache contents for debugging purposes // multimodal models (see examples/llava) std::string mmproj = ""; // path to multimodal projector std::string image = ""; // path to an image file }; +bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params); + bool gpt_params_parse(int argc, char ** argv, gpt_params & params); void gpt_print_usage(int argc, char ** argv, const gpt_params & params); @@ -181,6 +202,10 @@ std::string llama_detokenize_bpe( llama_context * ctx, const std::vector & tokens); +// Uses the value from the model metadata if possible, otherwise +// defaults to true when model type is SPM, otherwise false. +bool llama_should_add_bos_token(const llama_model * model); + // // YAML utils // @@ -194,3 +219,13 @@ std::string get_sortable_timestamp(); void dump_non_result_info_yaml( FILE * stream, const gpt_params & params, const llama_context * lctx, const std::string & timestamp, const std::vector & prompt_tokens, const char * model_desc); + +// +// KV cache utils +// + +// Dump the KV cache view with the number of sequences per cell. +void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80); + +// Dump the KV cache view showing individual sequences in each cell (long output). +void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40); diff --git a/common/log.h b/common/log.h index d2c864cea..c0e814861 100644 --- a/common/log.h +++ b/common/log.h @@ -97,38 +97,56 @@ #define LOG_TEE_TARGET stderr #endif -// NOTE: currently disabled as it produces too many log files +// Utility for synchronizing log configuration state +// since std::optional was introduced only in c++17 +enum LogTriState +{ + LogTriStateSame, + LogTriStateFalse, + LogTriStateTrue +}; + // Utility to obtain "pid" like unique process id and use it when creating log files. -//inline std::string log_get_pid() -//{ -// static std::string pid; -// if (pid.empty()) -// { -// // std::this_thread::get_id() is the most portable way of obtaining a "process id" -// // it's not the same as "pid" but is unique enough to solve multiple instances -// // trying to write to the same log. -// std::stringstream ss; -// ss << std::this_thread::get_id(); -// pid = ss.str(); -// } -// -// return pid; -//} +inline std::string log_get_pid() +{ + static std::string pid; + if (pid.empty()) + { + // std::this_thread::get_id() is the most portable way of obtaining a "process id" + // it's not the same as "pid" but is unique enough to solve multiple instances + // trying to write to the same log. + std::stringstream ss; + ss << std::this_thread::get_id(); + pid = ss.str(); + } + + return pid; +} // Utility function for generating log file names with unique id based on thread id. // invocation with log_filename_generator( "llama", "log" ) creates a string "llama..log" // where the number is a runtime id of the current thread. -#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension) +#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension) // INTERNAL, DO NOT USE -inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension) +inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension) { + static bool _multilog = false; + + if (multilog != LogTriStateSame) + { + _multilog = multilog == LogTriStateTrue; + } + std::stringstream buf; buf << log_file_basename; - //buf << "."; - //buf << log_get_pid(); + if (_multilog) + { + buf << "."; + buf << log_get_pid(); + } buf << "."; buf << log_file_extension; @@ -213,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base #define LOG_TEE_FLF_VAL ,"" #endif -// Utility for synchronizing log configuration state -// since std::optional was introduced only in c++17 -enum LogTriState -{ - LogTriStateSame, - LogTriStateFalse, - LogTriStateTrue -}; - // INTERNAL, DO NOT USE // USE LOG() INSTEAD // @@ -315,16 +324,23 @@ enum LogTriState #endif // INTERNAL, DO NOT USE -inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr) +inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr) { - static bool _initialized{false}; - static bool _disabled{(filename.empty() && target == nullptr)}; + static bool _initialized = false; + static bool _append = false; + static bool _disabled = filename.empty() && target == nullptr; static std::string log_current_filename{filename}; static FILE *log_current_target{target}; static FILE *logfile = nullptr; if (change) { + if (append != LogTriStateSame) + { + _append = append == LogTriStateTrue; + return logfile; + } + if (disable == LogTriStateTrue) { // Disable primary target @@ -377,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri } } - logfile = fopen(filename.c_str(), "w"); + logfile = fopen(filename.c_str(), _append ? "a" : "w"); } if (!logfile) @@ -398,9 +414,9 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri } // INTERNAL, DO NOT USE -inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME) +inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME) { - return log_handler1_impl(change, disable, filename, target); + return log_handler1_impl(change, append, disable, filename, target); } // Disables logs entirely at runtime. @@ -411,7 +427,7 @@ inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTri // INTERNAL, DO NOT USE inline FILE *log_disable_impl() { - return log_handler1_impl(true, LogTriStateTrue); + return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue); } // Enables logs at runtime. @@ -420,19 +436,31 @@ inline FILE *log_disable_impl() // INTERNAL, DO NOT USE inline FILE *log_enable_impl() { - return log_handler1_impl(true, LogTriStateFalse); + return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse); } // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*) #define log_set_target(target) log_set_target_impl(target) // INTERNAL, DO NOT USE -inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); } -inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); } +inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); } +inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); } // INTERNAL, DO NOT USE inline FILE *log_handler() { return log_handler1_impl(); } +// Enable or disable creating separate log files for each run. +// can ONLY be invoked BEFORE first log use. +#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "") +// Enable or disable append mode for log file. +// can ONLY be invoked BEFORE first log use. +#define log_append(enable) log_append_impl(enable) +// INTERNAL, DO NOT USE +inline FILE *log_append_impl(bool enable) +{ + return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame); +} + inline void log_test() { log_disable(); @@ -494,6 +522,18 @@ inline bool log_param_single_parse(const std::string & param) return true; } + if (param == "--log-new") + { + log_multilog(true); + return true; + } + + if (param == "--log-append") + { + log_append(true); + return true; + } + return false; } @@ -523,7 +563,9 @@ inline void log_print_usage() printf(" --log-disable Disable trace logs\n"); printf(" --log-enable Enable trace logs\n"); printf(" --log-file Specify a log filename (without extension)\n"); - printf(" Log file will be tagged with unique ID and written as \"..log\"\n"); /* */ + printf(" --log-new Create a separate new log file on start. " + "Each log file will have unique name: \"..log\"\n"); + printf(" --log-append Don't truncate the old log file.\n"); } #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv) diff --git a/common/sampling.cpp b/common/sampling.cpp index c4996c985..1317024c2 100644 --- a/common/sampling.cpp +++ b/common/sampling.cpp @@ -39,6 +39,7 @@ void llama_sampling_free(struct llama_sampling_context * ctx) { void llama_sampling_reset(llama_sampling_context * ctx) { if (ctx->grammar != NULL) { llama_grammar_free(ctx->grammar); + ctx->grammar = NULL; } if (!ctx->parsed_grammar.rules.empty()) { @@ -89,10 +90,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) { snprintf(result, sizeof(result), "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n" - "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, typical_p = %.3f, temp = %.3f\n" + "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n" "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f", params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present, - params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp, + params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp, params.mirostat, params.mirostat_eta, params.mirostat_tau); return std::string(result); @@ -110,6 +111,7 @@ llama_token llama_sampling_sample( const float temp = params.temp; const int32_t top_k = params.top_k <= 0 ? n_vocab : params.top_k; const float top_p = params.top_p; + const float min_p = params.min_p; const float tfs_z = params.tfs_z; const float typical_p = params.typical_p; const int32_t penalty_last_n = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n; @@ -190,6 +192,7 @@ llama_token llama_sampling_sample( llama_sample_tail_free(ctx_main, &cur_p, tfs_z, min_keep); llama_sample_typical (ctx_main, &cur_p, typical_p, min_keep); llama_sample_top_p (ctx_main, &cur_p, top_p, min_keep); + llama_sample_min_p (ctx_main, &cur_p, min_p, min_keep); llama_sample_temp (ctx_main, &cur_p, temp); id = llama_sample_token(ctx_main, &cur_p); diff --git a/common/sampling.h b/common/sampling.h index 62ea6d4cf..7c9b8dcf2 100644 --- a/common/sampling.h +++ b/common/sampling.h @@ -14,6 +14,7 @@ typedef struct llama_sampling_params { int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens. int32_t top_k = 40; // <= 0 to use vocab size float top_p = 0.95f; // 1.0 = disabled + float min_p = 0.05f; // 0.0 = disabled float tfs_z = 1.00f; // 1.0 = disabled float typical_p = 1.00f; // 1.0 = disabled float temp = 0.80f; // 1.0 = disabled diff --git a/common/train.cpp b/common/train.cpp index 3cce5da26..773e2c59c 100644 --- a/common/train.cpp +++ b/common/train.cpp @@ -32,6 +32,7 @@ struct train_state * init_train_state() { state->opt = new struct ggml_opt_context; state->opt->ctx = NULL; state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM); + state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; state->opt->loss_after = 0.0f; return state; @@ -1045,6 +1046,7 @@ struct train_params_common get_default_train_params_common() { params.n_batch = 8; params.n_gradient_accumulation = 1; params.n_epochs = -1; + params.n_gpu_layers = 0; params.custom_n_ctx = false; @@ -1080,6 +1082,7 @@ struct train_params_common get_default_train_params_common() { params.adam_beta2 = 0.999f; params.adam_gclip = 1.0f; params.adam_eps_f = 0.0f; + return params; } @@ -1133,6 +1136,7 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train fprintf(stderr, " --adam-beta2 N AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2); fprintf(stderr, " --adam-gclip N AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip); fprintf(stderr, " --adam-epsf N AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f); + fprintf(stderr, " -ngl N, --n-gpu-layers N Number of model layers to offload to GPU (default %d)", params->n_gpu_layers); fprintf(stderr, "\n"); } @@ -1352,6 +1356,17 @@ bool consume_common_train_arg( return true; } params->adam_gclip = std::stof(argv[i]); + } else if (arg == "-ngl" || arg == "--n-gpu-layers") { + if (++i >= argc) { + *invalid_param = true; + return true; + } +#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD + params->n_gpu_layers = std::stoi(argv[i]); +#else + fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n"); + fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n"); +#endif } else if (arg == "-h" || arg == "--help") { params->print_usage = true; return true; diff --git a/common/train.h b/common/train.h index 42fa704b8..263d940c0 100644 --- a/common/train.h +++ b/common/train.h @@ -9,6 +9,8 @@ #include "ggml.h" #include "llama.h" +#define LLAMA_TRAIN_MAX_NODES 16384 + typedef std::string mt19937_state; struct train_state { @@ -44,6 +46,7 @@ struct train_params_common { int n_batch; int n_gradient_accumulation; int n_epochs; + int n_gpu_layers; bool custom_n_ctx; diff --git a/convert-baichuan-hf-to-gguf.py b/convert-baichuan-hf-to-gguf.py deleted file mode 100755 index 5ee99be73..000000000 --- a/convert-baichuan-hf-to-gguf.py +++ /dev/null @@ -1,316 +0,0 @@ -#!/usr/bin/env python3 -# HF baichuan --> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import TYPE_CHECKING, Any -import itertools -import numpy as np -import torch -from sentencepiece import SentencePieceProcessor # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -if TYPE_CHECKING: - from typing import TypeAlias - -NDArray: TypeAlias = 'np.ndarray[Any, Any]' - -# reverse HF permute back to original pth layout - - -def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray: - if n_kv_head is not None and n_head != n_kv_head: - n_head //= n_kv_head - - return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape)) - -def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray: - r = weights.shape[0] // 3 - return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv)) - -def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray: - r = weights.shape[0] // 3 - return weights[r * n_part : r * n_part + r, ...] - -def count_model_parts(dir_model: str) -> int: - num_parts = 0 - - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - - return num_parts - - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file") - parser.add_argument( - "--vocab-only", action="store_true", - help="extract only the vocab", - ) - parser.add_argument( - "--outfile", type=Path, - help="path to write to; default: based on input", - ) - parser.add_argument( - "model", type=Path, - help="directory containing model file, or model file itself (*.bin)", - ) - parser.add_argument( - "ftype", type=int, choices=[0, 1], default=1, nargs='?', - help="output format - use 0 for float32, 1 for float16", - ) - parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine") - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -endianess = gguf.GGUFEndian.LITTLE -if args.bigendian: - endianess = gguf.GGUFEndian.BIG -endianess_str = "Big Endian" if args.bigendian else "Little Endian" -print(f"gguf: Conversion Endianess {endianess}") -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) -print("hello print: ",hparams["architectures"][0]) -if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectures"][0] != "BaiChuanForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit() - -# get number of model parts -num_parts = count_model_parts(dir_model) -print(f"num_parts:{num_parts}\n") -ARCH=gguf.MODEL_ARCH.BAICHUAN -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) - -print("gguf: get model metadata") - -block_count = hparams["num_hidden_layers"] -head_count = hparams["num_attention_heads"] - -if "num_key_value_heads" in hparams: - head_count_kv = hparams["num_key_value_heads"] -else: - head_count_kv = head_count - -if "_name_or_path" in hparams: - hf_repo = hparams["_name_or_path"] -else: - hf_repo = "" - -if "max_sequence_length" in hparams: - ctx_length = hparams["max_sequence_length"] -elif "max_position_embeddings" in hparams: - ctx_length = hparams["max_position_embeddings"] -elif "model_max_length" in hparams: - ctx_length = hparams["model_max_length"] -else: - print("gguf: can not find ctx length parameter.") - - sys.exit() - - -gguf_writer.add_name(dir_model.name) -gguf_writer.add_source_hf_repo(hf_repo) -gguf_writer.add_tensor_data_layout("Meta AI original pth") -gguf_writer.add_context_length(ctx_length) -gguf_writer.add_embedding_length(hparams["hidden_size"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) -gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"]) -gguf_writer.add_head_count(head_count) -gguf_writer.add_head_count_kv(head_count_kv) -gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"]) - -if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]: - if "type" in hparams["rope_scaling"]: - if hparams["rope_scaling"]["type"] == "linear": - gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"]) - - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytes] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -tokenizer_model_file = dir_model / 'tokenizer.model' -if not tokenizer_model_file.is_file(): - print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr) - sys.exit(1) - -# vocab type sentencepiece -print("gguf: get sentencepiece tokenizer vocab, scores and token types") - -tokenizer = SentencePieceProcessor(str(tokenizer_model_file)) -vocab_size = hparams.get('vocab_size') -if vocab_size is None: - vocab_size = tokenizer.vocab_size() - -for i in range(vocab_size): - text: bytes - score: float - - piece = tokenizer.id_to_piece(i) - text = piece.encode("utf-8") - score = tokenizer.get_score(i) - - toktype = 1 # defualt to normal token type - if tokenizer.is_unknown(i): - toktype = 2 - if tokenizer.is_control(i): - toktype = 3 - - # toktype = 4 is user-defined = tokens from added_tokens.json - - if tokenizer.is_unused(i): - toktype = 5 - if tokenizer.is_byte(i): - toktype = 6 - - tokens.append(text) - scores.append(score) - toktypes.append(toktype) - -added_tokens_file = dir_model / 'added_tokens.json' -if added_tokens_file.is_file(): - with open(added_tokens_file, "r", encoding="utf-8") as f: - addtokens_json = json.load(f) - - print("gguf: get added tokens") - - for key in addtokens_json: - tokens.append( key.encode("utf-8") ) - scores.append(-1000.0) - toktypes.append(4) # user-defined token type - - -gguf_writer.add_tokenizer_model("llama") -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_scores(scores) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") - - tmp=model_part - for i in range(block_count): - if f"model.layers.{i}.self_attn.W_pack.weight" in model_part: - print(f"Unpacking and permuting layer {i}") - tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count) - tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv) - tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2) - del tmp[f"model.layers.{i}.self_attn.W_pack.weight"] - - for name in model_part.keys(): - data = model_part[name] - # we don't need these - if name.endswith(".rotary_emb.inv_freq"): - continue - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(name + " -> " + new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-bloom-hf-to-gguf.py b/convert-bloom-hf-to-gguf.py deleted file mode 100755 index 6e866d943..000000000 --- a/convert-bloom-hf-to-gguf.py +++ /dev/null @@ -1,247 +0,0 @@ -#!/usr/bin/env python3 -# HF bloom --> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import re -import struct -import sys -from pathlib import Path -from typing import Any - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -# Supported Models: -# https://huggingface.co/bigscience/bloom-1b7 -# https://huggingface.co/bigscience/bloom-3b -# https://huggingface.co/bigscience/bloom-7b1 -# https://huggingface.co/Langboat/bloom-1b4-zh -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a Bloom model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "BloomForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - sys.exit(1) - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH=gguf.MODEL_ARCH.BLOOM -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams["n_layer"] - -gguf_writer.add_name("Bloom") -n_embed = hparams.get("hidden_size", hparams.get("n_embed")) -n_head = hparams.get("n_head", hparams.get("num_attention_heads")) -gguf_writer.add_context_length(hparams.get("seq_length", n_embed)) -gguf_writer.add_embedding_length(n_embed) -gguf_writer.add_feed_forward_length(4 * n_embed) -gguf_writer.add_block_count(block_count) -gguf_writer.add_head_count(n_head) -gguf_writer.add_head_count_kv(n_head) -gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) -gguf_writer.add_file_type(ftype) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -# The number of tokens in tokenizer.json can differ from the expected vocab size. -# This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) -assert max(tokenizer.vocab.values()) < vocab_size - -added_vocab = tokenizer.get_added_vocab() -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - if i not in reverse_vocab: - tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH, block_count) - -# params for qkv transform -n_head_kv = hparams.get("n_head_kv", n_head) -head_dim = n_embed // n_head - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(dir_model / part_name, map_location="cpu") - - has_lm_head = True - if "lm_head.weight" not in model_part.keys() and "output.weight" not in model_part.keys(): - has_lm_head = False - - for original_name in model_part.keys(): - data = model_part[original_name] - name = re.sub(r'transformer\.', '', original_name) - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name): - # Map bloom-style qkv_linear to gpt-style qkv_linear - # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252 # noqa - # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312 # noqa - qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed)) - data = np.concatenate( - (qkv_weights[:, 0, :, :].reshape((-1, n_embed)), - qkv_weights[:, 1, :, :].reshape((-1, n_embed)), - qkv_weights[:, 2, :, :].reshape((-1, n_embed))), - axis=0 - ) - print("re-format attention.linear_qkv.weight") - elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name): - qkv_bias = data.reshape((n_head, 3, n_embed // n_head)) - data = np.concatenate( - (qkv_bias[:, 0, :].reshape((n_embed,)), - qkv_bias[:, 1, :].reshape((n_embed,)), - qkv_bias[:, 2, :].reshape((n_embed,))), - axis=0 - ) - print("re-format attention.linear_qkv.bias") - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - if not has_lm_head and name == "word_embeddings.weight": - gguf_writer.add_tensor("output.weight", data) - print(name, "=>", "output.weight" + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype)) # noqa - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-falcon-hf-to-gguf.py b/convert-falcon-hf-to-gguf.py deleted file mode 100755 index 8e8f3c3f8..000000000 --- a/convert-falcon-hf-to-gguf.py +++ /dev/null @@ -1,253 +0,0 @@ -#!/usr/bin/env python3 -# HF falcon--> gguf conversion - -from __future__ import annotations - -import argparse -import contextlib -import json -import os -import struct -import sys -from pathlib import Path -from typing import Any - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -def count_model_parts(dir_model: Path, prefix: str) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith(prefix): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file") - parser.add_argument( - "--vocab-only", action="store_true", - help="extract only the vocab", - ) - parser.add_argument( - "--outfile", type=Path, - help="path to write to; default: based on input", - ) - parser.add_argument( - "model", type=Path, - help="directory containing model file, or model file itself (*.bin)", - ) - parser.add_argument( - "ftype", type=int, choices=[0, 1], default=1, nargs='?', - help="output format - use 0 for float32, 1 for float16", - ) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] not in ("RWForCausalLM", "FalconForCausalLM"): - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit(1) - -# get number of model parts -num_parts = count_model_parts(dir_model, "model-00") -if num_parts: - is_safetensors = True - from safetensors import safe_open -else: - is_safetensors = False - num_parts = count_model_parts(dir_model, "pytorch_model-") - -ARCH=gguf.MODEL_ARCH.FALCON -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams.get("num_hidden_layers") -if block_count is None: - block_count = hparams["n_layer"] # old name - -n_head = hparams.get("num_attention_heads") -if n_head is None: - n_head = hparams["n_head"] # old name - -n_head_kv = hparams.get("num_kv_heads") -if n_head_kv is None: - n_head_kv = hparams.get("n_head_kv", 1) # old name - -gguf_writer.add_name("Falcon") -gguf_writer.add_context_length(2048) # not in config.json -gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform -gguf_writer.add_embedding_length(hparams["hidden_size"]) -gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_head_count(n_head) -gguf_writer.add_head_count_kv(n_head_kv) -gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) -gguf_writer.add_file_type(ftype) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -# The number of tokens in tokenizer.json can differ from the expected vocab size. -# This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) -assert max(tokenizer.vocab.values()) < vocab_size - -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - tokens.append(reverse_vocab[i]) - scores.append(0.0) # dummy - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_scores(scores) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -head_dim = hparams["hidden_size"] // n_head - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -elif is_safetensors: - part_names = ( - f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1) - ) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - if is_safetensors: - ctx = safe_open(dir_model / part_name, framework="pt", device="cpu") - else: - ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu")) - - with ctx as model_part: - for name in model_part.keys(): - data = model_part.get_tensor(name) if is_safetensors else model_part[name] - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - # QKV tensor transform - # The original query_key_value tensor contains n_head_kv "kv groups", - # each consisting of n_head/n_head_kv query weights followed by one key - # and one value weight (shared by all query heads in the kv group). - # This layout makes it a big pain to work with in GGML. - # So we rearrange them here,, so that we have n_head query weights - # followed by n_head_kv key weights followed by n_head_kv value weights, - # in contiguous fashion. - # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py - - if "query_key_value" in name: - qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head) - q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head) - k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head) - v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head) - data = torch.cat((q,k,v)).reshape_as(data) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-gptneox-hf-to-gguf.py b/convert-gptneox-hf-to-gguf.py deleted file mode 100755 index 02d1fdf16..000000000 --- a/convert-gptneox-hf-to-gguf.py +++ /dev/null @@ -1,221 +0,0 @@ -#!/usr/bin/env python3 -# HF gptneox--> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import Any - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file") - parser.add_argument( - "--vocab-only", action="store_true", - help="extract only the vocab", - ) - parser.add_argument( - "--outfile", type=Path, - help="path to write to; default: based on input", - ) - parser.add_argument( - "model", type=Path, - help="directory containing model file, or model file itself (*.bin)", - ) - parser.add_argument( - "ftype", type=int, choices=[0, 1], default=1, nargs='?', - help="output format - use 0 for float32, 1 for float16", - ) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "GPTNeoXForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit() - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH=gguf.MODEL_ARCH.GPTNEOX -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams["num_hidden_layers"] - -gguf_writer.add_name(dir_model.name) -gguf_writer.add_context_length(hparams["max_position_embeddings"]) -gguf_writer.add_embedding_length(hparams["hidden_size"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) -gguf_writer.add_rope_dimension_count(int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"]))) -gguf_writer.add_head_count(hparams["num_attention_heads"]) -gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True) -gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"]) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -# The number of tokens in tokenizer.json can differ from the expected vocab size. -# This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) -assert max(tokenizer.vocab.values()) < vocab_size - -added_vocab = tokenizer.get_added_vocab() -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - if i not in reverse_vocab: - tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") - - for name in model_part.keys(): - data = model_part[name] - - # we don't need these - if name.endswith(".attention.masked_bias") or name.endswith(".attention.bias") or name.endswith(".attention.rotary_emb.inv_freq"): - continue - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py new file mode 100755 index 000000000..147d5717e --- /dev/null +++ b/convert-hf-to-gguf.py @@ -0,0 +1,900 @@ +#!/usr/bin/env python3 + +from __future__ import annotations + +import argparse +import contextlib +import json +import os +import re +import sys +from enum import IntEnum +from pathlib import Path +from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast + +import numpy as np +import torch + +if TYPE_CHECKING: + from torch import Tensor + +if 'NO_LOCAL_GGUF' not in os.environ: + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) +import gguf + + +###### MODEL DEFINITIONS ###### + +class SentencePieceTokenTypes(IntEnum): + NORMAL = 1 + UNKNOWN = 2 + CONTROL = 3 + USER_DEFINED = 4 + UNUSED = 5 + BYTE = 6 + + +class Model: + def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool): + self.dir_model = dir_model + self.ftype = ftype + self.fname_out = fname_out + self.is_big_endian = is_big_endian + self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE + self.is_safetensors = self._is_model_safetensors() + self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin") + self.part_names = self._get_part_names() + self.hparams = Model.load_hparams(self.dir_model) + self.model_arch = self._get_model_architecture() + self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess) + + def set_vocab(self): + self._set_vocab_gpt2() + + def get_tensors(self) -> Iterator[tuple[str, Tensor]]: + for part_name in self.part_names: + print(f"gguf: loading model part '{part_name}'") + ctx: ContextManager[Any] + if self.is_safetensors: + from safetensors import safe_open + ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu")) + else: + ctx = contextlib.nullcontext(torch.load(self.dir_model / part_name, map_location="cpu")) + + with ctx as model_part: + for name in model_part.keys(): + data = model_part.get_tensor(name) if self.is_safetensors else model_part[name] + yield name, data + + def set_gguf_parameters(self): + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_block_count(self.hparams.get( + "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")), + )) + if (n_ctx := self.hparams.get("max_position_embeddings")) is not None: + self.gguf_writer.add_context_length(n_ctx) + if (n_embd := self.hparams.get("hidden_size")) is not None: + self.gguf_writer.add_embedding_length(n_embd) + if (n_ff := self.hparams.get("intermediate_size")) is not None: + self.gguf_writer.add_feed_forward_length(n_ff) + if (n_head := self.hparams.get("num_attention_head")) is not None: + self.gguf_writer.add_head_count(n_head) + self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True)) + + def write_tensors(self): + block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer"))) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + for name, data_torch in self.get_tensors(): + # we don't need these + if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + def write(self): + self.write_tensors() + self.gguf_writer.write_header_to_file() + self.gguf_writer.write_kv_data_to_file() + self.gguf_writer.write_tensors_to_file() + self.gguf_writer.close() + + def write_vocab(self): + self.gguf_writer.write_header_to_file() + self.gguf_writer.write_kv_data_to_file() + self.gguf_writer.close() + + @staticmethod + def count_model_parts(dir_model: Path, prefix: str) -> int: + num_parts = 0 + for filename in os.listdir(dir_model): + if filename.endswith(prefix): + num_parts += 1 + + return num_parts + + @staticmethod + def load_hparams(dir_model): + with open(dir_model / "config.json", "r", encoding="utf-8") as f: + return json.load(f) + + @staticmethod + def from_model_architecture(model_architecture): + if model_architecture == "GPTNeoXForCausalLM": + return GPTNeoXModel + if model_architecture == "BloomForCausalLM": + return BloomModel + if model_architecture == "MPTForCausalLM": + return MPTModel + if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"): + return BaichuanModel + if model_architecture in ("FalconForCausalLM", "RWForCausalLM"): + return FalconModel + if model_architecture == "GPTBigCodeForCausalLM": + return StarCoderModel + if model_architecture == "GPTRefactForCausalLM": + return RefactModel + if model_architecture == "PersimmonForCausalLM": + return PersimmonModel + if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"): + return StableLMModel + return Model + + def _is_model_safetensors(self) -> bool: + return Model.count_model_parts(self.dir_model, ".safetensors") > 0 + + def _get_part_names(self): + if self.is_safetensors: + if self.num_parts == 1: # there's only one .safetensors file + return ("model.safetensors",) + return (f"model-{n:05}-of-{self.num_parts:05}.safetensors" for n in range(1, self.num_parts + 1)) + + if self.num_parts == 1: # there's only one .bin file + return ("pytorch_model.bin",) + return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1)) + + def _get_model_architecture(self) -> gguf.MODEL_ARCH: + arch = self.hparams["architectures"][0] + if arch == "GPTNeoXForCausalLM": + return gguf.MODEL_ARCH.GPTNEOX + if arch == "BloomForCausalLM": + return gguf.MODEL_ARCH.BLOOM + if arch == "MPTForCausalLM": + return gguf.MODEL_ARCH.MPT + if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"): + return gguf.MODEL_ARCH.BAICHUAN + if arch in ("FalconForCausalLM", "RWForCausalLM"): + return gguf.MODEL_ARCH.FALCON + if arch == "GPTBigCodeForCausalLM": + return gguf.MODEL_ARCH.STARCODER + if arch == "GPTRefactForCausalLM": + return gguf.MODEL_ARCH.REFACT + if arch == "PersimmonForCausalLM": + return gguf.MODEL_ARCH.PERSIMMON + if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"): + return gguf.MODEL_ARCH.STABLELM + + raise NotImplementedError(f'Architecture "{arch}" not supported!') + + def _set_vocab_gpt2(self): + dir_model = self.dir_model + hparams = self.hparams + tokens: list[bytearray] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer # type: ignore[attr-defined] + tokenizer = AutoTokenizer.from_pretrained(dir_model) + vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) + assert max(tokenizer.vocab.values()) < vocab_size + + reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()} + added_vocab = tokenizer.get_added_vocab() + + for i in range(vocab_size): + if i not in reverse_vocab: + pad_token = f"[PAD{i}]".encode('utf-8') + tokens.append(bytearray(pad_token)) + toktypes.append(gguf.TokenType.USER_DEFINED) + elif reverse_vocab[i] in added_vocab: + tokens.append(reverse_vocab[i]) + if tokenizer.added_tokens_decoder[i].special: + toktypes.append(gguf.TokenType.CONTROL) + else: + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.NORMAL) + + self.gguf_writer.add_tokenizer_model("gpt2") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) + special_vocab.add_to_gguf(self.gguf_writer) + + def _set_vocab_sentencepiece(self): + from sentencepiece import SentencePieceProcessor + + tokenizer_path = self.dir_model / 'tokenizer.model' + + tokens: list[bytes] = [] + scores: list[float] = [] + toktypes: list[int] = [] + + if not tokenizer_path.is_file(): + print(f'Error: Missing {tokenizer_path}', file=sys.stderr) + sys.exit(1) + + tokenizer = SentencePieceProcessor(str(tokenizer_path)) + vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) + + for token_id in range(vocab_size): + piece = tokenizer.id_to_piece(token_id) + text = piece.encode("utf-8") + score = tokenizer.get_score(token_id) + + toktype = SentencePieceTokenTypes.NORMAL + if tokenizer.is_unknown(token_id): + toktype = SentencePieceTokenTypes.UNKNOWN + elif tokenizer.is_control(token_id): + toktype = SentencePieceTokenTypes.CONTROL + elif tokenizer.is_unused(token_id): + toktype = SentencePieceTokenTypes.UNUSED + elif tokenizer.is_byte(token_id): + toktype = SentencePieceTokenTypes.BYTE + + tokens.append(text) + scores.append(score) + toktypes.append(toktype) + + added_tokens_file = self.dir_model / 'added_tokens.json' + if added_tokens_file.is_file(): + with open(added_tokens_file, "r", encoding="utf-8") as f: + added_tokens_json = json.load(f) + + for key in added_tokens_json: + tokens.append(key.encode("utf-8")) + scores.append(-1000.0) + toktypes.append(SentencePieceTokenTypes.USER_DEFINED) + + self.gguf_writer.add_tokenizer_model("llama") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_scores(scores) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens)) + special_vocab.add_to_gguf(self.gguf_writer) + + +class GPTNeoXModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"]) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_rope_dimension_count( + int(self.hparams["rotary_pct"] * (self.hparams["hidden_size"] // self.hparams["num_attention_heads"])), + ) + self.gguf_writer.add_head_count(self.hparams["num_attention_heads"]) + self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True)) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"]) + + +class BloomModel(Model): + def set_gguf_parameters(self): + self.gguf_writer.add_name("Bloom") + n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) + n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) + self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed)) + self.gguf_writer.add_embedding_length(n_embed) + self.gguf_writer.add_feed_forward_length(4 * n_embed) + self.gguf_writer.add_block_count(self.hparams["n_layer"]) + self.gguf_writer.add_head_count(n_head) + self.gguf_writer.add_head_count_kv(n_head) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def write_tensors(self): + block_count = self.hparams["n_layer"] + tensors = dict(self.get_tensors()) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + has_lm_head = True + n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads")) + n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed")) + + for name, data_torch in tensors.items(): + if "lm_head.weight" not in tensors.keys() and "output.weight" not in tensors.keys(): + has_lm_head = False + + name = re.sub(r'transformer\.', '', name) + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name): + # Map bloom-style qkv_linear to gpt-style qkv_linear + # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252 # noqa + # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312 # noqa + qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed)) + data = np.concatenate( + ( + qkv_weights[:, 0, :, :].reshape((-1, n_embed)), + qkv_weights[:, 1, :, :].reshape((-1, n_embed)), + qkv_weights[:, 2, :, :].reshape((-1, n_embed)), + ), + axis=0, + ) + print("re-format attention.linear_qkv.weight") + elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name): + qkv_bias = data.reshape((n_head, 3, n_embed // n_head)) + data = np.concatenate( + ( + qkv_bias[:, 0, :].reshape((n_embed,)), + qkv_bias[:, 1, :].reshape((n_embed,)), + qkv_bias[:, 2, :].reshape((n_embed,)), + ), + axis=0, + ) + print("re-format attention.linear_qkv.bias") + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + if not has_lm_head and name == "word_embeddings.weight": + self.gguf_writer.add_tensor("output.weight", data) + print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}") + + +class MPTModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams["n_layers"] + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_context_length(self.hparams["max_seq_len"]) + self.gguf_writer.add_embedding_length(self.hparams["d_model"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(4 * self.hparams["d_model"]) + self.gguf_writer.add_head_count(self.hparams["n_heads"]) + if kv_n_heads := self.hparams["attn_config"].get("kv_n_heads"): + self.gguf_writer.add_head_count_kv(kv_n_heads) + self.gguf_writer.add_layer_norm_eps(1e-5) + if self.hparams["attn_config"]["clip_qkv"] is not None: + self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"]) + self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"]) + + def write_tensors(self): + block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers")) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + for name, data_torch in self.get_tensors(): + # we don't need these + if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + # note: MPT output is tied to (same as) wte in original model; + # for easier implementation in llama.cpp it's duplicated in GGUF, though :/ + if new_name == "token_embd.weight": + self.gguf_writer.add_tensor("output.weight", data) + + +class BaichuanModel(Model): + def set_vocab(self): + self._set_vocab_sentencepiece() + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + hf_repo = self.hparams.get("_name_or_path", "") + + ctx_length = 0 + if "max_sequence_length" in self.hparams: + ctx_length = self.hparams["max_sequence_length"] + elif "max_position_embeddings" in self.hparams: + ctx_length = self.hparams["max_position_embeddings"] + elif "model_max_length" in self.hparams: + ctx_length = self.hparams["model_max_length"] + else: + print("gguf: can not find ctx length parameter.") + sys.exit() + + self.gguf_writer.add_name(self.dir_model.name) + self.gguf_writer.add_source_hf_repo(hf_repo) + self.gguf_writer.add_tensor_data_layout("Meta AI original pth") + self.gguf_writer.add_context_length(ctx_length) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"]) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) + + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + + def write_tensors(self): + # Collect tensors from generator object + model_kv = dict(self.get_tensors()) + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + + for i in range(block_count): + if (w := model_kv.get(f"model.layers.{i}.self_attn.W_pack.weight")) is not None: + print(f"Unpacking and permuting layer {i}") + model_kv[f"model.layers.{i}.self_attn.q_proj.weight"] = \ + self._reverse_hf_permute_part(w, 0, head_count, head_count) + model_kv[f"model.layers.{i}.self_attn.k_proj.weight"] = \ + self._reverse_hf_permute_part(w, 1, head_count, head_count_kv) + model_kv[f"model.layers.{i}.self_attn.v_proj.weight"] = \ + self._reverse_hf_part(w, 2) + del model_kv[f"model.layers.{i}.self_attn.W_pack.weight"] + + for name, data_torch in model_kv.items(): + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: + if n_kv_head is not None and n_head != n_kv_head: + n_head //= n_kv_head + + return ( + weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape) + ) + + def _reverse_hf_permute_part( + self, weights: Tensor, n_part: int, n_head: int, n_head_kv: int | None = None, + ) -> Tensor: + r = weights.shape[0] // 3 + return self._reverse_hf_permute(weights[r * n_part:r * n_part + r, ...], n_head, n_head_kv) + + def _reverse_hf_part(self, weights: Tensor, n_part: int) -> Tensor: + r = weights.shape[0] // 3 + return weights[r * n_part:r * n_part + r, ...] + + +class FalconModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams.get("num_hidden_layers") + if block_count is None: + block_count = self.hparams["n_layer"] # old name + + n_head = self.hparams.get("num_attention_heads") + if n_head is None: + n_head = self.hparams["n_head"] # old name + + n_head_kv = self.hparams.get("num_kv_heads") + if n_head_kv is None: + n_head_kv = self.hparams.get("n_head_kv", 1) # old name + + self.gguf_writer.add_name("Falcon") + self.gguf_writer.add_context_length(2048) # not in config.json + self.gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_feed_forward_length(4 * self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_head_count(n_head) + self.gguf_writer.add_head_count_kv(n_head_kv) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def write_tensors(self): + block_count = self.hparams.get("num_hidden_layers") + if block_count is None: + block_count = self.hparams["n_layer"] # old name + + n_head = self.hparams.get("num_attention_heads") + if n_head is None: + n_head = self.hparams["n_head"] # old name + + n_head_kv = self.hparams.get("num_kv_heads") + if n_head_kv is None: + n_head_kv = self.hparams.get("n_head_kv", 1) # old name + + head_dim = self.hparams["hidden_size"] // n_head + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in self.get_tensors(): + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + # QKV tensor transform + # The original query_key_value tensor contains n_head_kv "kv groups", + # each consisting of n_head/n_head_kv query weights followed by one key + # and one value weight (shared by all query heads in the kv group). + # This layout makes it a big pain to work with in GGML. + # So we rearrange them here,, so that we have n_head query weights + # followed by n_head_kv key weights followed by n_head_kv value weights, + # in contiguous fashion. + # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py + + if "query_key_value" in name: + qkv = data_torch.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head) + q = qkv[:, :-2].reshape(n_head * head_dim, head_dim * n_head) + k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head) + v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head) + data_torch = torch.cat((q, k, v)).reshape_as(data_torch) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + +class StarCoderModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams["n_layer"] + + self.gguf_writer.add_name("StarCoder") + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_head_count_kv(1) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + +class RefactModel(Model): + def set_gguf_parameters(self): + hidden_dim = self.hparams["n_embd"] + inner_dim = 4 * hidden_dim + hidden_dim = int(2 * inner_dim / 3) + multiple_of = 256 + ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) + + block_count = self.hparams["n_layer"] + + self.gguf_writer.add_name("Refact") + # refact uses Alibi. So this is from config.json which might be used by training. + self.gguf_writer.add_context_length(self.hparams["n_positions"]) + self.gguf_writer.add_embedding_length(self.hparams["n_embd"]) + + self.gguf_writer.add_feed_forward_length(ff_dim) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_head_count(self.hparams["n_head"]) + self.gguf_writer.add_head_count_kv(1) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"]) + self.gguf_writer.add_file_type(self.ftype) + + def write_tensors(self): + hidden_dim = self.hparams["n_embd"] + inner_dim = 4 * hidden_dim + hidden_dim = int(2 * inner_dim / 3) + multiple_of = 256 + ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) + n_head = self.hparams["n_head"] + n_head_kv = 1 + head_dim = self.hparams["n_embd"] // n_head + block_count = self.hparams["n_layer"] + + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + tensors = dict(self.get_tensors()) + for i in range(block_count): + if (w := tensors.get(f"transformer.h.{i}.attn.kv.weight")) is not None: + tensors[f"model.layers.{i}.self_attn.k_proj.weight"] = w[:n_head_kv * head_dim] + tensors[f"model.layers.{i}.self_attn.v_proj.weight"] = w[n_head_kv * head_dim:] + del tensors[f"transformer.h.{i}.attn.kv.weight"] + if (w := tensors.get(f"transformer.h.{i}.attn.q.weight")) is not None: + tensors[f"model.layers.{i}.self_attn.q_proj.weight"] = w + del tensors[f"transformer.h.{i}.attn.q.weight"] + if (w := tensors.get(f"transformer.h.{i}.mlp.gate_up_proj.weight")) is not None: + tensors[f"model.layers.{i}.mlp.gate_proj.weight"] = w[:ff_dim] + tensors[f"model.layers.{i}.mlp.up_proj.weight"] = w[ff_dim:] + del tensors[f"transformer.h.{i}.mlp.gate_up_proj.weight"] + + for name, data_torch in tensors.items(): + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight",)) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + + self.gguf_writer.add_tensor(new_name, data) + + +class PersimmonModel(Model): + def set_gguf_parameters(self): + block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers")) + head_count = self.hparams["num_attention_heads"] + head_count_kv = head_count + hidden_size = self.hparams["hidden_size"] + + self.gguf_writer.add_name('persimmon-8b-chat') + self.gguf_writer.add_embedding_length(hidden_size) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_rope_dimension_count(hidden_size // head_count) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"]) + self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"]) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) + + def set_vocab(self): + self._set_vocab_sentencepiece() + # self.gguf_writer.add_bos_token_id(71013) + # self.gguf_writer.add_eos_token_id(71013) + + def write_tensors(self): + block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers")) + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + + for name, data_torch in self.get_tensors(): + if name.endswith(".self_attention.rotary_emb.inv_freq"): + continue + old_dtype = data_torch.dtype + # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?) + data = data_torch.to(torch.float32).squeeze().numpy() + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + n_dims = len(data.shape) + print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + +class StableLMModel(Model): + def set_gguf_parameters(self): + hparams = self.hparams + block_count = hparams["num_hidden_layers"] + + self.gguf_writer.add_name(dir_model.name) + self.gguf_writer.add_context_length(hparams["max_position_embeddings"]) + self.gguf_writer.add_embedding_length(hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"]) + self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"]))) + self.gguf_writer.add_head_count(hparams["num_attention_heads"]) + self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True) + self.gguf_writer.add_layer_norm_eps(1e-5) + +###### CONVERSION LOGIC ###### + + +def parse_args() -> argparse.Namespace: + parser = argparse.ArgumentParser(description="Convert a huggingface model to a GGML compatible file") + parser.add_argument( + "--vocab-only", action="store_true", + help="extract only the vocab", + ) + parser.add_argument( + "--outfile", type=Path, + help="path to write to; default: based on input", + ) + parser.add_argument( + "--outtype", type=str, choices=["f32", "f16"], default="f16", + help="output format - use f32 for float32, f16 for float16", + ) + parser.add_argument("--bigendian", action="store_true", help="model is executed on big endian machine") + parser.add_argument( + "model", type=Path, + help="directory containing model file", + ) + + return parser.parse_args() + + +args = parse_args() + +dir_model = args.model +if not dir_model.is_dir(): + print(f'Error: {args.model} is not a directory', file=sys.stderr) + sys.exit(1) + +ftype_map = { + "f32": gguf.GGMLQuantizationType.F32, + "f16": gguf.GGMLQuantizationType.F16, +} + +if args.outfile is not None: + fname_out = args.outfile +else: + # output in the same directory as the model by default + fname_out = dir_model / f'ggml-model-{args.outtype}.gguf' + +print(f"Loading model: {dir_model.name}") + +hparams = Model.load_hparams(dir_model) + +with torch.inference_mode(): + model_class = Model.from_model_architecture(hparams["architectures"][0]) + model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian) + + print("Set model parameters") + model_instance.set_gguf_parameters() + + print("Set model tokenizer") + model_instance.set_vocab() + + if args.vocab_only: + print(f"Exporting model vocab to '{fname_out}'") + model_instance.write_vocab() + else: + print(f"Exporting model to '{fname_out}'") + model_instance.write() + + print(f"Model successfully exported to '{fname_out}'") diff --git a/convert-llama-ggml-to-gguf.py b/convert-llama-ggml-to-gguf.py index 871add64d..e359330af 100755 --- a/convert-llama-ggml-to-gguf.py +++ b/convert-llama-ggml-to-gguf.py @@ -2,7 +2,6 @@ from __future__ import annotations import argparse -import math import struct import sys from enum import IntEnum @@ -12,34 +11,16 @@ import numpy as np import os if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf -# Note: Does not support GGML_QKK_64 -QK_K = 256 -# Items here are (block size, type size) -GGML_QUANT_SIZES = { - gguf.GGMLQuantizationType.F32 : (1, 4), - gguf.GGMLQuantizationType.F16 : (1, 2), - gguf.GGMLQuantizationType.Q4_0 : (32, 2 + 16), - gguf.GGMLQuantizationType.Q4_1 : (32, 2 + 2 + 16), - gguf.GGMLQuantizationType.Q5_0 : (32, 2 + 4 + 16), - gguf.GGMLQuantizationType.Q5_1 : (32, 2 + 2 + 4 + 16), - gguf.GGMLQuantizationType.Q8_0 : (32, 2 + 32), - gguf.GGMLQuantizationType.Q8_1 : (32, 4 + 4 + 32), - gguf.GGMLQuantizationType.Q2_K : (256, 2 + 2 + QK_K // 16 + QK_K // 4), - gguf.GGMLQuantizationType.Q3_K : (256, 2 + QK_K // 4 + QK_K // 8 + 12), - gguf.GGMLQuantizationType.Q4_K : (256, 2 + 2 + QK_K // 2 + 12), - gguf.GGMLQuantizationType.Q5_K : (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12), - gguf.GGMLQuantizationType.Q6_K : (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16), - gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8), -} class GGMLFormat(IntEnum): GGML = 0 GGMF = 1 GGJT = 2 + class GGMLFType(IntEnum): ALL_F32 = 0 MOSTLY_F16 = 1 @@ -59,6 +40,7 @@ class GGMLFType(IntEnum): MOSTLY_Q5_K_M = 17 MOSTLY_Q6_K = 18 + class Hyperparameters: def __init__(self): self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0 @@ -90,6 +72,7 @@ class Hyperparameters: def __str__(self): return f'' + class Vocab: def __init__(self, load_scores = True): self.items = [] @@ -111,6 +94,7 @@ class Vocab: self.items.append((item_text, item_score)) return offset - orig_offset + class Tensor: def __init__(self, use_padding = True): self.name = None @@ -125,7 +109,7 @@ class Tensor: (n_dims, name_len, dtype) = struct.unpack('<3I', data[offset:offset + 12]) assert n_dims >= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}' assert name_len < 4096, 'Absurd tensor name length' - quant = GGML_QUANT_SIZES.get(dtype) + quant = gguf.GGML_QUANT_SIZES.get(dtype) assert quant is not None, 'Unknown tensor type' (blksize, tysize) = quant offset += 12 @@ -144,6 +128,7 @@ class Tensor: # print(n_dims, name_len, dtype, self.dims, self.name, pad) return offset - orig_offset + class GGMLModel: def __init__(self): self.hyperparameters = None @@ -180,8 +165,8 @@ class GGMLModel: if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16): err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.' elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2): - if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1, - GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0): + if ftype in (GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1, + GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0): err = 'Q4 and Q8 quantizations changed in GGJTv3.' if len(err) > 0: raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.') @@ -208,6 +193,7 @@ class GGMLModel: hp.set_n_ff(self) return offset + class GGMLToGGUF: def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None): hp = ggml_model.hyperparameters @@ -238,7 +224,7 @@ class GGMLToGGUF: gguf_writer = gguf.GGUFWriter( self.cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], - use_temp_file = False ) + use_temp_file = False) self.add_params(gguf_writer) self.add_vocab(gguf_writer) if self.special_vocab is not None: @@ -362,7 +348,8 @@ class GGMLToGGUF: mapped_name, data[tensor.start_offset:tensor.start_offset + tensor.len_bytes], raw_shape = tempdims, - raw_dtype = tensor.dtype ) + raw_dtype = tensor.dtype) + def handle_metadata(cfg, hp): import convert @@ -386,38 +373,40 @@ def handle_metadata(cfg, hp): raise ValueError('Unable to load metadata') vocab = convert.load_vocab( cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir, - cfg.vocabtype ) + cfg.vocabtype) # FIXME: Respect cfg.vocab_dir? svocab = gguf.SpecialVocab(cfg.model_metadata_dir, - load_merges = cfg.vocabtype == 'bpe', - n_vocab = vocab.vocab_size) + load_merges = cfg.vocabtype == 'bpe', + n_vocab = vocab.vocab_size) convert.check_vocab_size(params, vocab) return (params, vocab, svocab) + def handle_args(): parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF') parser.add_argument('--input', '-i', type = Path, required = True, - help = 'Input GGMLv3 filename') + help = 'Input GGMLv3 filename') parser.add_argument('--output', '-o', type = Path, required = True, - help ='Output GGUF filename') + help ='Output GGUF filename') parser.add_argument('--name', - help = 'Set model name') + help = 'Set model name') parser.add_argument('--desc', - help = 'Set model description') + help = 'Set model description') parser.add_argument('--gqa', type = int, default = 1, - help = 'grouped-query attention factor (use 8 for LLaMA2 70B)') + help = 'grouped-query attention factor (use 8 for LLaMA2 70B)') parser.add_argument('--eps', default = '5.0e-06', - help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2') + help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2') parser.add_argument('--context-length', '-c', type=int, default = 2048, - help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096') + help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096') parser.add_argument('--model-metadata-dir', '-m', type = Path, - help ='Load HuggingFace/.pth vocab and metadata from the specified directory') + help ='Load HuggingFace/.pth vocab and metadata from the specified directory') parser.add_argument("--vocab-dir", type=Path, - help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir") + help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir") parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm", - help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)") + help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)") return parser.parse_args() + def main(): cfg = handle_args() print(f'* Using config: {cfg}') @@ -427,7 +416,7 @@ def main(): data = np.memmap(cfg.input, mode = 'r') model = GGMLModel() print('* Scanning GGML input file') - offset = model.load(data, 0) + offset = model.load(data, 0) # noqa print(f'* GGML model hyperparameters: {model.hyperparameters}') vocab_override = None params_override = None @@ -442,12 +431,15 @@ def main(): print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n') if model.file_format == GGMLFormat.GGML: print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!') - converter = GGMLToGGUF(model, data, cfg, + converter = GGMLToGGUF( + model, data, cfg, params_override = params_override, vocab_override = vocab_override, - special_vocab = special_vocab ) + special_vocab = special_vocab + ) converter.save() print(f'* Successful completion. Output saved to: {cfg.output}') + if __name__ == '__main__': main() diff --git a/convert-mpt-hf-to-gguf.py b/convert-mpt-hf-to-gguf.py deleted file mode 100755 index 70d154b3f..000000000 --- a/convert-mpt-hf-to-gguf.py +++ /dev/null @@ -1,227 +0,0 @@ -#!/usr/bin/env python3 -# HF mpt--> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import Any - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file") - parser.add_argument( - "--vocab-only", action="store_true", - help="extract only the vocab", - ) - parser.add_argument( - "--outfile", type=Path, - help="path to write to; default: based on input", - ) - parser.add_argument( - "model", type=Path, - help="directory containing model file, or model file itself (*.bin)", - ) - parser.add_argument( - "ftype", type=int, choices=[0, 1], default=1, nargs='?', - help="output format - use 0 for float32, 1 for float16", - ) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "MPTForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit() - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH=gguf.MODEL_ARCH.MPT -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams["n_layers"] - -gguf_writer.add_name(dir_model.name) -gguf_writer.add_context_length(hparams["max_seq_len"]) -gguf_writer.add_embedding_length(hparams["d_model"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_feed_forward_length(4 * hparams["d_model"]) -gguf_writer.add_head_count(hparams["n_heads"]) -if kv_n_heads := hparams["attn_config"].get("kv_n_heads"): - gguf_writer.add_head_count_kv(kv_n_heads) -gguf_writer.add_layer_norm_eps(1e-05) -if hparams["attn_config"]["clip_qkv"] is not None: - gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"]) -gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"]) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but -# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to -# accomodate some "reserved" tokens; this is causing problems down the line in -# llama.cpp, so we pad the vocab with dummy tokens: - -vocab_size = hparams["vocab_size"] - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -added_vocab = tokenizer.get_added_vocab() -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - if i not in reverse_vocab: - tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu") - - for name in model_part.keys(): - data = model_part[name] - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Cannot map tensor '" + name + "'") - continue # for the sake of compatibility with some old published models, don't quit - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - # note: MPT output is tied to (same as) wte in original model; - # for easier implementation in llama.cpp it's duplicated in GGUF, though :/ - if new_name == "token_embd.weight": - gguf_writer.add_tensor("output.weight", data) - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-persimmon-to-gguf.py b/convert-persimmon-to-gguf.py index e022ffe46..206b7d5ff 100644 --- a/convert-persimmon-to-gguf.py +++ b/convert-persimmon-to-gguf.py @@ -6,9 +6,10 @@ import argparse from pathlib import Path from sentencepiece import SentencePieceProcessor if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf + def _flatten_dict(dct, tensors, prefix=None): assert isinstance(dct, dict) for key in dct.keys(): @@ -21,6 +22,7 @@ def _flatten_dict(dct, tensors, prefix=None): raise ValueError(type(dct[key])) return None + def _get_sentencepiece_tokenizer_info(dir_model: Path): tokenizer_path = dir_model / 'adept_vocab.model' print('gguf: getting sentencepiece tokenizer from', tokenizer_path) @@ -54,6 +56,7 @@ def _get_sentencepiece_tokenizer_info(dir_model: Path): pass return tokens, scores, toktypes + def main(): parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file") parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") @@ -125,6 +128,5 @@ def main(): print("") - if __name__ == '__main__': main() diff --git a/convert-refact-hf-to-gguf.py b/convert-refact-hf-to-gguf.py deleted file mode 100755 index f0cfe84d8..000000000 --- a/convert-refact-hf-to-gguf.py +++ /dev/null @@ -1,272 +0,0 @@ -#!/usr/bin/env python3 -# HF refact--> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import sys -from pathlib import Path - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if "NO_LOCAL_GGUF" not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf")) -import gguf - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser( - description="Convert a Refact model to a GGML compatible file" - ) - parser.add_argument( - "--vocab-only", - action="store_true", - help="extract only the vocab", - ) - parser.add_argument( - "--outfile", - type=Path, - help="path to write to; default: based on input", - ) - parser.add_argument( - "model", - type=Path, - help="directory containing model file, or model file itself (*.bin)", - ) - parser.add_argument( - "ftype", - type=int, - choices=[0, 1], - default=1, - nargs="?", - help="output format - use 0 for float32, 1 for float16", - ) - return parser.parse_args() - - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f"Error: {args.model} is not a directory", file=sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf" - -print("gguf: loading model " + dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "GPTRefactForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit(1) - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH = gguf.MODEL_ARCH.REFACT -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -# Get refact feed forward dimension -hidden_dim = hparams["n_embd"] -inner_dim = 4 * hidden_dim -hidden_dim = int(2 * inner_dim / 3) -multiple_of = 256 -ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of) - -block_count = hparams["n_layer"] - -gguf_writer.add_name("Refact") -# refact uses Alibi. So this is from config.json which might be used by training. -gguf_writer.add_context_length(hparams["n_positions"]) -gguf_writer.add_embedding_length(hparams["n_embd"]) - -gguf_writer.add_feed_forward_length(ff_dim) -gguf_writer.add_block_count(block_count) -gguf_writer.add_head_count(hparams["n_head"]) -gguf_writer.add_head_count_kv(1) -gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"]) -gguf_writer.add_file_type(ftype) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -# The number of tokens in tokenizer.json can differ from the expected vocab size. -# This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) -assert max(tokenizer.vocab.values()) < vocab_size - -added_vocab = tokenizer.get_added_vocab() -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - if i not in reverse_vocab: - tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_types(toktypes) - -special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH, block_count) - -# params for qkv transform -n_head = hparams["n_head"] -n_head_kv = 1 - -head_dim = hparams["n_embd"] // n_head - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(dir_model / part_name, map_location="cpu") - - for i in range(block_count): - if f"transformer.h.{i}.attn.kv.weight" in model_part: - data = model_part[f"transformer.h.{i}.attn.kv.weight"] - model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[ - : n_head_kv * head_dim - ] - model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[ - n_head_kv * head_dim : - ] - del model_part[f"transformer.h.{i}.attn.kv.weight"] - if f"transformer.h.{i}.attn.q.weight" in model_part: - model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[ - f"transformer.h.{i}.attn.q.weight" - ] - del model_part[f"transformer.h.{i}.attn.q.weight"] - if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part: - data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"] - model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim] - model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:] - del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"] - - for name in model_part.keys(): - data = model_part[name] - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes=(".weight",)) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ( - ftype == 1 - and data_dtype == np.float32 - and name.endswith(".weight") - and n_dims == 2 - ): - data = data.astype(np.float16) - - print( - new_name - + ", n_dims = " - + str(n_dims) - + ", " - + str(old_dtype) - + " --> " - + str(data.dtype) - ) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert-starcoder-hf-to-gguf.py b/convert-starcoder-hf-to-gguf.py deleted file mode 100755 index a9bfed85e..000000000 --- a/convert-starcoder-hf-to-gguf.py +++ /dev/null @@ -1,210 +0,0 @@ -#!/usr/bin/env python3 -# HF starcoder --> gguf conversion - -from __future__ import annotations - -import argparse -import json -import os -import struct -import sys -from pathlib import Path -from typing import Any - -import numpy as np -import torch -from transformers import AutoTokenizer # type: ignore[import] - -if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) -import gguf - - -def count_model_parts(dir_model: Path) -> int: - num_parts = 0 - for filename in os.listdir(dir_model): - if filename.startswith("pytorch_model-"): - num_parts += 1 - - if num_parts > 0: - print("gguf: found " + str(num_parts) + " model parts") - return num_parts - - -def parse_args() -> argparse.Namespace: - parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file") - parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.bin)") - parser.add_argument("ftype", type=int, help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1) - return parser.parse_args() - -args = parse_args() - -dir_model = args.model -ftype = args.ftype -if not dir_model.is_dir(): - print(f'Error: {args.model} is not a directory', file = sys.stderr) - sys.exit(1) - -# possible tensor data types -# ftype == 0 -> float32 -# ftype == 1 -> float16 - -# map from ftype to string -ftype_str = ["f32", "f16"] - -if args.outfile is not None: - fname_out = args.outfile -else: - # output in the same directory as the model by default - fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf' - -print("gguf: loading model "+dir_model.name) - -with open(dir_model / "config.json", "r", encoding="utf-8") as f: - hparams = json.load(f) - -if hparams["architectures"][0] != "GPTBigCodeForCausalLM": - print("Model architecture not supported: " + hparams["architectures"][0]) - - sys.exit(1) - -# get number of model parts -num_parts = count_model_parts(dir_model) - -ARCH=gguf.MODEL_ARCH.STARCODER -gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH]) - -print("gguf: get model metadata") - -block_count = hparams["n_layer"] - -gguf_writer.add_name("StarCoder") -gguf_writer.add_context_length(hparams["n_positions"]) -gguf_writer.add_embedding_length(hparams["n_embd"]) -gguf_writer.add_feed_forward_length(4 * hparams["n_embd"]) -gguf_writer.add_block_count(block_count) -gguf_writer.add_head_count(hparams["n_head"]) -gguf_writer.add_head_count_kv(1) -gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"]) -gguf_writer.add_file_type(ftype) - -# TOKENIZATION - -print("gguf: get tokenizer metadata") - -tokens: list[bytearray] = [] -scores: list[float] = [] -toktypes: list[int] = [] - -# gpt2 tokenizer -gguf_writer.add_tokenizer_model("gpt2") - -print("gguf: get gpt2 tokenizer vocab") - -# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py -tokenizer = AutoTokenizer.from_pretrained(dir_model) - -# The number of tokens in tokenizer.json can differ from the expected vocab size. -# This causes downstream issues with mismatched tensor sizes when running the inference -vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) -assert max(tokenizer.vocab.values()) < vocab_size - -added_vocab = tokenizer.get_added_vocab() -reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()} - -for i in range(vocab_size): - if i not in reverse_vocab: - tokens.append(f"[PAD{i}]") - toktypes.append(gguf.TokenType.USER_DEFINED) - elif reverse_vocab[i] in added_vocab: - tokens.append(reverse_vocab[i]) - if tokenizer.added_tokens_decoder[i].special: - toktypes.append(gguf.TokenType.CONTROL) - else: - toktypes.append(gguf.TokenType.USER_DEFINED) - else: - tokens.append(reverse_vocab[i]) - toktypes.append(gguf.TokenType.NORMAL) - -gguf_writer.add_token_list(tokens) -gguf_writer.add_token_types(toktypes) -special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens)) -special_vocab.add_to_gguf(gguf_writer) - -# TENSORS - -tensor_map = gguf.get_tensor_name_map(ARCH,block_count) - -# params for qkv transform -n_head = hparams["n_head"] -n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1 - -head_dim = hparams["n_embd"] // n_head - -# tensor info -print("gguf: get tensor metadata") - -if num_parts == 0: - part_names = iter(("pytorch_model.bin",)) -else: - part_names = ( - f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1) - ) - -for part_name in part_names: - if args.vocab_only: - break - print("gguf: loading model part '" + part_name + "'") - model_part = torch.load(dir_model / part_name, map_location="cpu") - - for name in model_part.keys(): - data = model_part[name] - - old_dtype = data.dtype - - # convert any unsupported data types to float32 - if data.dtype != torch.float16 and data.dtype != torch.float32: - data = data.to(torch.float32) - - data = data.squeeze().numpy() - - # map tensor names - new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias")) - if new_name is None: - print("Can not map tensor '" + name + "'") - sys.exit() - - n_dims = len(data.shape) - data_dtype = data.dtype - - # if f32 desired, convert any float16 to float32 - if ftype == 0 and data_dtype == np.float16: - data = data.astype(np.float32) - - # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 - if ftype == 1 and data_dtype == np.float16 and n_dims == 1: - data = data.astype(np.float32) - - # if f16 desired, convert any float32 2-dim weight tensors to float16 - if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: - data = data.astype(np.float16) - - print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype)) - - gguf_writer.add_tensor(new_name, data) - - -print("gguf: write header") -gguf_writer.write_header_to_file() -print("gguf: write metadata") -gguf_writer.write_kv_data_to_file() -if not args.vocab_only: - print("gguf: write tensors") - gguf_writer.write_tensors_to_file() - -gguf_writer.close() - -print(f"gguf: model successfully exported to '{fname_out}'") -print("") diff --git a/convert.py b/convert.py old mode 100755 new mode 100644 index bfbfab283..3ad836ce0 --- a/convert.py +++ b/convert.py @@ -3,11 +3,9 @@ from __future__ import annotations import argparse import concurrent.futures -import copy import enum import faulthandler import functools -import io import itertools import json import math @@ -23,14 +21,14 @@ from abc import ABCMeta, abstractmethod from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor from dataclasses import dataclass from pathlib import Path -from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar +from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar import numpy as np -from sentencepiece import SentencePieceProcessor # type: ignore[import] +from sentencepiece import SentencePieceProcessor import os if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf')) + sys.path.insert(1, str(Path(__file__).parent / 'gguf-py')) import gguf if TYPE_CHECKING: @@ -48,6 +46,7 @@ DEFAULT_CONCURRENCY = 8 # data types # + @dataclass(frozen=True) class DataType: name: str @@ -57,15 +56,18 @@ class DataType: def elements_to_bytes(self, n_elements: int) -> int: return n_elements * self.dtype.itemsize + @dataclass(frozen=True) class UnquantizedDataType(DataType): pass + DT_F16 = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0']) DT_F32 = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0']) DT_I32 = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = []) DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0']) + @dataclass(frozen=True) class QuantizedDataType(DataType): block_size: int @@ -79,6 +81,7 @@ class QuantizedDataType(DataType): assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}' return self.quantized_dtype.itemsize * (n_elements // self.block_size) + @dataclass(frozen=True) class Q8_0QuantizedDataType(QuantizedDataType): # Mini Q8_0 quantization in Python! @@ -88,6 +91,7 @@ class Q8_0QuantizedDataType(QuantizedDataType): n_blocks = arr.size // self.block_size blocks = arr.reshape((n_blocks, self.block_size)) # Much faster implementation of block quantization contributed by @Cebtenzzre + def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]: d = abs(blocks).max(axis = 1) / np.float32(127) with np.errstate(divide = 'ignore'): @@ -96,10 +100,11 @@ class Q8_0QuantizedDataType(QuantizedDataType): yield from zip(d, qs) return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype) + DT_Q8_0 = Q8_0QuantizedDataType('Q8_0', - dtype = np.dtype(np.float32), valid_conversions = [], - ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32, - quantized_dtype = np.dtype([('d', ' 1 else DT_F32 + GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { GGMLFileType.AllF32 : DT_F32, GGMLFileType.MostlyF16 : DT_F16, @@ -140,6 +148,7 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = { # hparams loading # + @dataclass class Params: n_vocab: int @@ -151,8 +160,11 @@ class Params: n_head_kv: int f_norm_eps: float + rope_scaling_type: gguf.RopeScalingType | None = None f_rope_freq_base: float | None = None f_rope_scale: float | None = None + n_orig_ctx: int | None = None + rope_finetuned: bool | None = None ftype: GGMLFileType | None = None @@ -166,11 +178,11 @@ class Params: # try transformer naming first if "model.layers.0.self_attn.q_proj.weight" in model: - n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model) elif "model.layers.0.self_attn.W_pack.weight" in model: # next: try baichuan naming - n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) + n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model) else: - n_layer=next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) + n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model) if n_layer < 1: raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n" @@ -198,20 +210,20 @@ class Params: def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) - n_vocab = config["vocab_size"] - n_embd = config["hidden_size"] - n_layer = config["num_hidden_layers"] - n_ff = config["intermediate_size"] - n_head = config["num_attention_heads"] - n_head_kv = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head - f_norm_eps = config["rms_norm_eps"] - f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None - + rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None rope_scaling = config.get("rope_scaling") - if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear": - f_rope_scale = config["rope_scaling"].get("factor") - else: - f_rope_scale = None + + if rope_scaling is not None and (typ := rope_scaling.get("type")): + rope_factor = rope_scaling.get("factor") + f_rope_scale = rope_factor + if typ == "linear": + rope_scaling_type = gguf.RopeScalingType.LINEAR + elif typ == "yarn": + rope_scaling_type = gguf.RopeScalingType.YARN + n_orig_ctx = rope_scaling['original_max_position_embeddings'] + rope_finetuned = rope_scaling['finetuned'] + else: + raise NotImplementedError(f'Unknown rope scaling type: {typ}') if "max_sequence_length" in config: n_ctx = config["max_sequence_length"] @@ -222,16 +234,19 @@ class Params: "Suggestion: provide 'config.json' of the model in the same directory containing model files.") return Params( - n_vocab = n_vocab, - n_embd = n_embd, - n_layer = n_layer, - n_ctx = n_ctx, - n_ff = n_ff, - n_head = n_head, - n_head_kv = n_head_kv, - f_norm_eps = f_norm_eps, - f_rope_freq_base = f_rope_freq_base, - f_rope_scale = f_rope_scale, + n_vocab = config["vocab_size"], + n_embd = config["hidden_size"], + n_layer = config["num_hidden_layers"], + n_ctx = n_ctx, + n_ff = config["intermediate_size"], + n_head = (n_head := config["num_attention_heads"]), + n_head_kv = config.get("num_key_value_heads", n_head), + f_norm_eps = config["rms_norm_eps"], + f_rope_freq_base = config.get("rope_theta"), + rope_scaling_type = rope_scaling_type, + f_rope_scale = f_rope_scale, + n_orig_ctx = n_orig_ctx, + rope_finetuned = rope_finetuned, ) # LLaMA v2 70B params.json @@ -240,17 +255,8 @@ class Params: def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params: config = json.load(open(config_path)) - n_vocab = config["vocab_size"] if "vocab_size" in config else -1 - n_embd = config["dim"] - n_layer = config["n_layers"] - n_ff = -1 - n_head = config["n_heads"] - n_head_kv = config["n_kv_heads"] if "n_kv_heads" in config else n_head - f_norm_eps = config["norm_eps"] - f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None - # hack to determine LLaMA v1 vs v2 vs CodeLlama - if f_rope_freq_base == 1000000: + if config.get("rope_theta") == 1000000: # CodeLlama n_ctx = 16384 elif config["norm_eps"] == 1e-05: @@ -260,22 +266,16 @@ class Params: # LLaMA v1 n_ctx = 2048 - if n_vocab == -1: - n_vocab = model["tok_embeddings.weight"].shape[0] - - if n_ff == -1: - n_ff = model["layers.0.feed_forward.w1.weight"].shape[0] - return Params( - n_vocab = n_vocab, - n_embd = n_embd, - n_layer = n_layer, + n_vocab = config.get("vocab_size", model["tok_embeddings.weight"].shape[0]), + n_embd = config["dim"], + n_layer = config["n_layers"], n_ctx = n_ctx, - n_ff = n_ff, - n_head = n_head, - n_head_kv = n_head_kv, - f_norm_eps = f_norm_eps, - f_rope_freq_base = f_rope_freq_base, + n_ff = model["layers.0.feed_forward.w1.weight"].shape[0], + n_head = (n_head := config["n_heads"]), + n_head_kv = config.get("n_kv_heads", n_head), + f_norm_eps = config["norm_eps"], + f_rope_freq_base = config.get("rope_theta"), ) @staticmethod @@ -319,7 +319,7 @@ class BpeVocab: (item['content'], item['id']) for item in tokenizer_json.get('added_tokens', []) # Added tokens here can be duplicates of the main vocabulary. - if item['content'] not in self.bpe_tokenizer ) + if item['content'] not in self.bpe_tokenizer) vocab_size: int = len(self.bpe_tokenizer) expected_ids = list(range(vocab_size, vocab_size + len(added_tokens))) @@ -337,7 +337,6 @@ class BpeVocab: def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: tokenizer = self.bpe_tokenizer - from transformers.models.gpt2 import tokenization_gpt2 # type: ignore[import] reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()} for i, _ in enumerate(tokenizer): @@ -417,6 +416,7 @@ class SentencePieceVocab: def __repr__(self) -> str: return f"" + Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab' # @@ -424,13 +424,14 @@ Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab' # TODO: reuse (probably move to gguf.py?) # + def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray: - #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) ) + # print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) ) if n_head_kv is not None and n_head != n_head_kv: n_head = n_head_kv return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) - .swapaxes(1, 2) - .reshape(weights.shape)) + .swapaxes(1, 2) + .reshape(weights.shape)) class Tensor(metaclass=ABCMeta): @@ -511,7 +512,7 @@ class LazyTensor: ret = self._load() # Should be okay if it maps to the same numpy type? assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \ - (self.data_type, ret.data_type, self.description) + (self.data_type, ret.data_type, self.description) return ret def astype(self, data_type: DataType) -> LazyTensor: @@ -599,6 +600,7 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe return lazy_tensor.load().permute(n_head, n_head_kv) return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description) + def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor: def load() -> Tensor: return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv) @@ -606,6 +608,7 @@ def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_ s[0] = s[0] // 3 return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description) + def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor: def load() -> Tensor: return lazy_tensor.load().part(n_part) @@ -701,6 +704,7 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus: data_base_path=pickle_paths[0][:-4], zip_file=zf) model = unpickler.load() + if 'model' in model: model = model['model'] as_dict = dict(model.items()) return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None) @@ -754,6 +758,7 @@ def lazy_load_file(path: Path) -> ModelPlus: In = TypeVar('In') Out = TypeVar('Out') + def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]: '''Parallel map, but with backpressure. If the caller doesn't call `next` fast enough, this will stop calling `func` at some point rather than @@ -788,6 +793,7 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc break yield result + def check_vocab_size(params: Params, vocab: Vocab) -> None: if params.n_vocab != vocab.vocab_size: assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab) @@ -806,7 +812,7 @@ def check_vocab_size(params: Params, vocab: Vocab) -> None: class OutputFile: - def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None: + def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess) def add_meta_arch(self, params: Params) -> None: @@ -831,8 +837,16 @@ class OutputFile: if params.f_rope_freq_base is not None: self.gguf.add_rope_freq_base(params.f_rope_freq_base) - if params.f_rope_scale is not None: - self.gguf.add_rope_scale_linear(params.f_rope_scale) + if params.rope_scaling_type: + assert params.f_rope_scale is not None + self.gguf.add_rope_scaling_type(params.rope_scaling_type) + self.gguf.add_rope_scaling_factor(params.f_rope_scale) + + if params.n_orig_ctx is not None: + self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx) + + if params.rope_finetuned is not None: + self.gguf.add_rope_scaling_finetuned(params.rope_finetuned) if params.ftype is not None: self.gguf.add_file_type(params.ftype) @@ -852,7 +866,7 @@ class OutputFile: elif isinstance(vocab, BpeVocab): self.gguf.add_tokenizer_model("gpt2") else: - raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab') + raise ValueError('Unknown vocab type: Not BpeVocab or SentencePieceVocab') self.gguf.add_token_list(tokens) self.gguf.add_token_scores(scores) self.gguf.add_token_types(toktypes) @@ -878,7 +892,7 @@ class OutputFile: self.gguf.close() @staticmethod - def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None: + def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: check_vocab_size(params, vocab) of = OutputFile(fname_out, endianess=endianess) @@ -906,7 +920,7 @@ class OutputFile: return dt.quantize(arr) @staticmethod - def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess=gguf.GGUFEndian.LITTLE) -> None: + def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None: check_vocab_size(params, vocab) of = OutputFile(fname_out, endianess=endianess) @@ -940,8 +954,9 @@ class OutputFile: of.close() + def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType: - wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type + wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) +".weight"].data_type if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32): return GGMLFileType.AllF32 @@ -954,10 +969,12 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT raise Exception(f"Unexpected combination of types: {name_to_type}") + def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel: return {name: tensor.astype(output_type.type_for_tensor(name, tensor)) for (name, tensor) in model.items()} + def convert_model_names(model: LazyModel, params: Params) -> LazyModel: tmap = gguf.TensorNameMap(ARCH, params.n_layer) should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, [])) @@ -970,7 +987,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: print(f"Permuting layer {i}") tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head) tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv) - #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] + # tmp[f"model.layers.{i}.self_attn.v_proj.weight"] = model[f"model.layers.{i}.self_attn.v_proj.weight"] elif f"model.layers.{i}.self_attn.W_pack.weight" in model: print(f"Unpacking and permuting layer {i}") tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head) @@ -995,6 +1012,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel: return out + def nth_multifile_path(path: Path, n: int) -> Path | None: '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return the nth path in the model. @@ -1039,7 +1057,8 @@ def load_some_model(path: Path) -> ModelPlus: # Be extra-friendly and accept either a file or a directory: if path.is_dir(): # Check if it's a set of safetensors files first - files = list(path.glob("model-00001-of-*.safetensors")) + globs = ["model-00001-of-*.safetensors", "model.safetensors"] + files = [file for glob in globs for file in path.glob(glob)] if not files: # Try the PyTorch patterns too, with lower priority globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"] @@ -1115,14 +1134,18 @@ def do_dump_model(model_plus: ModelPlus) -> None: def main(args_in: list[str] | None = None) -> None: + output_choices = ["f32", "f16"] + if np.uint32(1) == np.uint32(1).newbyteorder("<"): + # We currently only support Q8_0 output on little endian systems. + output_choices.append("q8_0") parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file") parser.add_argument("--dump", action="store_true", help="don't convert, just show what's in the model") parser.add_argument("--dump-single", action="store_true", help="don't convert, just show what's in a single model file") parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab") - parser.add_argument("--outtype", choices=["f32", "f16", "q8_0"], help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") + parser.add_argument("--outtype", choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)") parser.add_argument("--vocab-dir", type=Path, help="directory containing tokenizer.model, if separate from model file") parser.add_argument("--outfile", type=Path, help="path to write to; default: based on input") - parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)") + parser.add_argument("model", type=Path, help="directory containing model file, or model file itself (*.pth, *.pt, *.bin, *.safetensors)") parser.add_argument("--vocabtype", choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm") parser.add_argument("--ctx", type=int, help="model training context (default: based on input)") parser.add_argument("--concurrency", type=int, help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY) @@ -1171,8 +1194,8 @@ def main(args_in: list[str] | None = None) -> None: # FIXME: Try to respect vocab_dir somehow? vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype) special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, - load_merges = args.vocabtype == 'bpe', - n_vocab = vocab.vocab_size) + load_merges = args.vocabtype == 'bpe', + n_vocab = vocab.vocab_size) outfile = args.outfile OutputFile.write_vocab_only(outfile, params, vocab, special_vocab) print(f"Wrote {outfile}") @@ -1185,8 +1208,8 @@ def main(args_in: list[str] | None = None) -> None: vocab = load_vocab(vocab_dir, args.vocabtype) # FIXME: Try to respect vocab_dir somehow? special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent, - load_merges = args.vocabtype == 'bpe', - n_vocab = vocab.vocab_size) + load_merges = args.vocabtype == 'bpe', + n_vocab = vocab.vocab_size) model = model_plus.model model = convert_model_names(model, params) diff --git a/docs/llama-star/idea-arch.key b/docs/llama-star/idea-arch.key new file mode 100755 index 000000000..3e068e707 Binary files /dev/null and b/docs/llama-star/idea-arch.key differ diff --git a/docs/llama-star/idea-arch.pdf b/docs/llama-star/idea-arch.pdf new file mode 100644 index 000000000..4fa92c71d Binary files /dev/null and b/docs/llama-star/idea-arch.pdf differ diff --git a/docs/token_generation_performance_tips.md b/docs/token_generation_performance_tips.md index c9acff7d4..d7e863dff 100644 --- a/docs/token_generation_performance_tips.md +++ b/docs/token_generation_performance_tips.md @@ -17,7 +17,7 @@ llama_model_load_internal: [cublas] total VRAM used: 17223 MB If you see these lines, then the GPU is being used. ## Verifying that the CPU is not oversaturated -llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physicial CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down. +llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physical CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down. # Example of runtime flags effect on inference speed benchmark These runs were tested on the following machine: diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index 75b8df676..71bcb6893 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -24,6 +24,7 @@ else() add_subdirectory(llama-bench) add_subdirectory(llava) add_subdirectory(main) + add_subdirectory(tokenize) add_subdirectory(parallel) add_subdirectory(perplexity) add_subdirectory(quantize) diff --git a/examples/batched.swift/Sources/main.swift b/examples/batched.swift/Sources/main.swift index 772730382..ba15197ae 100644 --- a/examples/batched.swift/Sources/main.swift +++ b/examples/batched.swift/Sources/main.swift @@ -153,7 +153,7 @@ while n_cur <= n_len { // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p); // is it an end of stream? -> mark the stream as finished - if new_token_id == llama_token_eos(context) || n_cur == n_len { + if new_token_id == llama_token_eos(model) || n_cur == n_len { i_batch[i] = -1 // print("") if n_parallel > 1 { diff --git a/examples/benchmark/CMakeLists.txt b/examples/benchmark/CMakeLists.txt index 14916d831..2bb47bab5 100644 --- a/examples/benchmark/CMakeLists.txt +++ b/examples/benchmark/CMakeLists.txt @@ -1,9 +1,6 @@ set(TARGET benchmark) add_executable(${TARGET} benchmark-matmult.cpp) install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT}) target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/benchmark/benchmark-matmult.cpp b/examples/benchmark/benchmark-matmult.cpp index f1c382aa9..284733b10 100644 --- a/examples/benchmark/benchmark-matmult.cpp +++ b/examples/benchmark/benchmark-matmult.cpp @@ -1,4 +1,3 @@ -#include "build-info.h" #include "common.h" #include "ggml.h" @@ -172,7 +171,8 @@ int main(int argc, char ** argv) { struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2); // printf("Creating compute graph\n"); - struct ggml_cgraph gf = ggml_build_forward(m11xm2); + struct ggml_cgraph * gf = ggml_new_graph(ctx); + ggml_build_forward_expand(gf, m11xm2); printf("n_threads=%i\n", benchmark_params.n_threads); @@ -181,9 +181,9 @@ int main(int argc, char ** argv) { std::vector work_buffer; - ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads); + ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads); - TENSOR_DUMP(gf.nodes[0]); + TENSOR_DUMP(gf->nodes[0]); printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype)); @@ -201,7 +201,8 @@ int main(int argc, char ** argv) { struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2); // printf("Creating compute graph\n"); - struct ggml_cgraph gf31 = ggml_build_forward(q31); + struct ggml_cgraph * gf31 = ggml_new_graph(ctx); + ggml_build_forward_expand(gf31, q31); // Set up a second graph computation to make sure we override the CPU cache lines // printf("Creating new tensor q12 & Running quantize\n"); @@ -212,7 +213,8 @@ int main(int argc, char ** argv) { struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2); //printf("Creating compute graph\n"); - struct ggml_cgraph gf32 = ggml_build_forward(q32); + struct ggml_cgraph * gf32 = ggml_new_graph(ctx); + ggml_build_forward_expand(gf32, q32); printf("n_threads=%i\n", benchmark_params.n_threads); const int dimx = sizex; @@ -224,7 +226,7 @@ int main(int argc, char ** argv) { // Let's use the F32 result from above as a reference for the quantized multiplication - float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]); + float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]); printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n"); printf("=====================================================================================\n"); @@ -234,7 +236,7 @@ int main(int argc, char ** argv) { long long int start = ggml_time_us(); //printf("Running ggml_graph_compute\n"); - ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads); + ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads); long long int stop = ggml_time_us(); long long int usec = stop-start; @@ -252,7 +254,7 @@ int main(int argc, char ** argv) { // Check that the matrix multiplication result is in the right ballpark // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different - float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]); + float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]); float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference); float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; // Let's accept an epsilon of 10^-6 @@ -267,7 +269,7 @@ int main(int argc, char ** argv) { } // Running a different graph computation to make sure we override the CPU cache lines - ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads); + ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads); } printf("\n"); printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations)); diff --git a/examples/embedding/CMakeLists.txt b/examples/embedding/CMakeLists.txt index 0c752c7bb..8ffc33868 100644 --- a/examples/embedding/CMakeLists.txt +++ b/examples/embedding/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} embedding.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 14075609e..3295cd240 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -1,4 +1,3 @@ -#include "build-info.h" #include "common.h" #include "llama.h" diff --git a/examples/export-lora/export-lora.cpp b/examples/export-lora/export-lora.cpp index d803cfd5c..c8754ce70 100644 --- a/examples/export-lora/export-lora.cpp +++ b/examples/export-lora/export-lora.cpp @@ -240,7 +240,7 @@ static struct lora_data * load_lora(struct lora_info * info) { } struct ggml_init_params params_ggml; - params_ggml.mem_size = ggml_tensor_overhead() * GGML_MAX_NODES; + params_ggml.mem_size = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE; params_ggml.mem_buffer = NULL; params_ggml.no_alloc = true; result->ctx = ggml_init(params_ggml); @@ -334,7 +334,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r; struct ggml_init_params params; - params.mem_size = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5; + params.mem_size = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5; params.mem_buffer = NULL; params.no_alloc = true; struct ggml_context * ctx = NULL; diff --git a/examples/finetune/README.md b/examples/finetune/README.md index 36e62578c..a2a2c1281 100644 --- a/examples/finetune/README.md +++ b/examples/finetune/README.md @@ -21,7 +21,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s ./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin ``` -Finetune output files will be saved every N iterations (config with `--save-every N`). +**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`). The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output. So in above example after 10 iterations these files will be written: - chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf diff --git a/examples/finetune/convert-finetune-checkpoint-to-gguf.py b/examples/finetune/convert-finetune-checkpoint-to-gguf.py index c8e14da87..c89090918 100644 --- a/examples/finetune/convert-finetune-checkpoint-to-gguf.py +++ b/examples/finetune/convert-finetune-checkpoint-to-gguf.py @@ -3,9 +3,7 @@ import argparse import gguf -import os import struct -import sys import numpy as np from pathlib import Path diff --git a/examples/finetune/finetune.cpp b/examples/finetune/finetune.cpp index 35824cd2d..af46e44a6 100644 --- a/examples/finetune/finetune.cpp +++ b/examples/finetune/finetune.cpp @@ -548,35 +548,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max); randomize_tensor_normal(lora->tok_embeddings_a, rnd); - randomize_tensor_normal(lora->tok_embeddings_b, rnd); + ggml_set_zero(lora->tok_embeddings_b); randomize_tensor_normal(lora->norm_a, rnd); - randomize_tensor_normal(lora->norm_b, rnd); + ggml_set_zero(lora->norm_b); randomize_tensor_normal(lora->output_a, rnd); - randomize_tensor_normal(lora->output_b, rnd); + ggml_set_zero(lora->output_b); for (uint32_t i = 0; i < n_layer; ++i) { auto & layer = lora->layers[i]; randomize_tensor_normal(layer.attention_norm_a, rnd); - randomize_tensor_normal(layer.attention_norm_b, rnd); + ggml_set_zero(layer.attention_norm_b); randomize_tensor_normal(layer.wq_a, rnd); - randomize_tensor_normal(layer.wq_b, rnd); + ggml_set_zero(layer.wq_b); randomize_tensor_normal(layer.wk_a, rnd); - randomize_tensor_normal(layer.wk_b, rnd); + ggml_set_zero(layer.wk_b); randomize_tensor_normal(layer.wv_a, rnd); - randomize_tensor_normal(layer.wv_b, rnd); + ggml_set_zero(layer.wv_b); randomize_tensor_normal(layer.wo_a, rnd); - randomize_tensor_normal(layer.wo_b, rnd); + ggml_set_zero(layer.wo_b); randomize_tensor_normal(layer.ffn_norm_a, rnd); - randomize_tensor_normal(layer.ffn_norm_b, rnd); + ggml_set_zero(layer.ffn_norm_b); randomize_tensor_normal(layer.w1_a, rnd); - randomize_tensor_normal(layer.w1_b, rnd); + ggml_set_zero(layer.w1_b); randomize_tensor_normal(layer.w2_a, rnd); - randomize_tensor_normal(layer.w2_b, rnd); + ggml_set_zero(layer.w2_b); randomize_tensor_normal(layer.w3_a, rnd); - randomize_tensor_normal(layer.w3_b, rnd); + ggml_set_zero(layer.w3_b); } free_random_normal_distribution(rnd); @@ -642,8 +642,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( const int rope_mode = 0; return ggml_rope_custom(ctx, - t, KQ_pos, n_rot, rope_mode, n_ctx, - rope_freq_base, rope_freq_scale); + t, KQ_pos, n_rot, rope_mode, n_ctx, 0, + rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f + ); }; set_name(tokens_input, "tokens_input"); @@ -652,7 +653,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( GGML_ASSERT(tokens_input->type == GGML_TYPE_I32); auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) { - if (ggml_is_quantized(a->type)) { + if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) { return ggml_add_cast(ctx, a, b, GGML_TYPE_F32); } else if (a->type == GGML_TYPE_F32) { return ggml_add(ctx, a, b); @@ -771,7 +772,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs( if (enable_checkpointing) { ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); } else { - *gb = *gf; + ggml_graph_cpy(gf, gb); ggml_build_backward_expand(ctx, gf, gb, true); } @@ -1545,6 +1546,7 @@ int main(int argc, char ** argv) { srand(params.common.seed); struct llama_model_params llama_mparams = llama_model_default_params(); + llama_mparams.n_gpu_layers = params.common.n_gpu_layers; llama_mparams.vocab_only = false; printf("%s: model base = '%s'\n", __func__, params.fn_model_base); @@ -1602,6 +1604,7 @@ int main(int argc, char ** argv) { opt->params = ggml_opt_default_params(GGML_OPT_ADAM); opt->params.print_forward_graph = false; opt->params.print_backward_graph = false; + opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; opt->params.n_threads = params.common.n_threads; opt->params.past = params.common.opt_past; opt->params.delta = params.common.opt_delta; @@ -1728,11 +1731,9 @@ int main(int argc, char ** argv) { ggml_allocr_free(alloc); // context for compute tensors without their data - size_t estimated_compute_size_wo_data = ( - ggml_tensor_overhead()*GGML_MAX_NODES*2 - + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*( - params.common.use_checkpointing ? 3 : 2 - ) + const size_t estimated_compute_size_wo_data = ( + 2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() + + (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true)) ); struct ggml_init_params ctx_compute_params = { estimated_compute_size_wo_data, // mem_size @@ -1755,11 +1756,11 @@ int main(int argc, char ** argv) { for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new_measure(tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = (enum ggml_cgraph_eval_order) order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_lora_finetune_graphs( &model, &lora, alloc, ctx_compute, @@ -1788,11 +1789,11 @@ int main(int argc, char ** argv) { mem_compute_data.resize(max_compute_size); ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = best_order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_lora_finetune_graphs( &model, &lora, alloc, ctx_compute, diff --git a/examples/finetune/finetune.sh b/examples/finetune/finetune.sh new file mode 100644 index 000000000..079bfa113 --- /dev/null +++ b/examples/finetune/finetune.sh @@ -0,0 +1,34 @@ +#!/bin/bash +cd `dirname $0` +cd ../.. + +EXE="./finetune" + +if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi +if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi + +# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses. +MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing. + +while getopts "dg" opt; do + case $opt in + d) + DEBUGGER="gdb --args" + ;; + g) + EXE="./build/bin/Release/finetune" + GPUARG="--gpu-layers 25" + ;; + esac +done + +$DEBUGGER $EXE \ + --model-base $MODEL \ + $GPUARG \ + --checkpoint-in chk-ol3b-shakespeare-LATEST.gguf \ + --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \ + --lora-out lora-ol3b-shakespeare-ITERATION.bin \ + --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \ + --save-every 10 \ + --threads 10 --adam-iter 30 --batch 4 --ctx 64 \ + --use-checkpointing diff --git a/examples/infill/CMakeLists.txt b/examples/infill/CMakeLists.txt index 57d01cb0b..e4e8028da 100644 --- a/examples/infill/CMakeLists.txt +++ b/examples/infill/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} infill.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/infill/infill.cpp b/examples/infill/infill.cpp index 9c52b7bba..4a7827876 100644 --- a/examples/infill/infill.cpp +++ b/examples/infill/infill.cpp @@ -2,7 +2,6 @@ #include "console.h" #include "llama.h" -#include "build-info.h" #include "grammar-parser.h" #include @@ -147,6 +146,13 @@ int main(int argc, char ** argv) { return 0; } + if (params.chatml) { + printf("\n************\n"); + printf("%s: please use the 'main' tool for chatml mode\n", __func__); + printf("************\n\n"); + + return 0; + } if (!params.antiprompt.empty()) { printf("\n************\n"); printf("%s: please use the 'main' tool for antiprompt mode\n", __func__); @@ -184,8 +190,8 @@ int main(int argc, char ** argv) { LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale); } - LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); - LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET); + LOG_TEE("%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT); + LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -231,7 +237,7 @@ int main(int argc, char ** argv) { LOG_TEE("\n"); LOG_TEE("%s\n", get_system_info(params).c_str()); } - const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM; + const bool add_bos = llama_should_add_bos_token(model); LOG("add_bos: %d\n", add_bos); bool suff_rm_leading_spc = params.escape; diff --git a/examples/llama-bench/CMakeLists.txt b/examples/llama-bench/CMakeLists.txt index 7e395afd0..5bdbea4e2 100644 --- a/examples/llama-bench/CMakeLists.txt +++ b/examples/llama-bench/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} llama-bench.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 780398184..9bd82d565 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -19,7 +19,6 @@ #include "ggml.h" #include "llama.h" #include "common.h" -#include "build-info.h" #include "ggml-cuda.h" // utils @@ -641,8 +640,8 @@ struct test { } }; -const std::string test::build_commit = BUILD_COMMIT; -const int test::build_number = BUILD_NUMBER; +const std::string test::build_commit = LLAMA_COMMIT; +const int test::build_number = LLAMA_BUILD_NUMBER; const bool test::cuda = !!ggml_cpu_has_cublas(); const bool test::opencl = !!ggml_cpu_has_clblast(); const bool test::metal = !!ggml_cpu_has_metal(); diff --git a/examples/llava/CMakeLists.txt b/examples/llava/CMakeLists.txt index 2d7979ecd..8ea3e5c83 100644 --- a/examples/llava/CMakeLists.txt +++ b/examples/llava/CMakeLists.txt @@ -1,20 +1,36 @@ -set(TARGET clip) -add_library(${TARGET} clip.cpp clip.h) -install(TARGETS ${TARGET} LIBRARY) -target_link_libraries(${TARGET} PRIVATE common ggml ${CMAKE_THREAD_LIBS_INIT}) -target_compile_features(${TARGET} PRIVATE cxx_std_11) -if (NOT MSVC) - target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h - endif() -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) +add_library(llava OBJECT + llava.cpp + llava.h + clip.cpp + clip.h + ) + +target_link_libraries(llava PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT}) + +target_include_directories(llava PUBLIC .) +target_include_directories(llava PUBLIC ../..) +target_include_directories(llava PUBLIC ../../common) + +target_compile_features(llava PRIVATE cxx_std_11) + +add_library(llava_static STATIC $) +if (BUILD_SHARED_LIBS) + set_target_properties(llava PROPERTIES POSITION_INDEPENDENT_CODE ON) + target_compile_definitions(llava PRIVATE LLAMA_SHARED LLAMA_BUILD) + add_library(llava_shared SHARED $) + target_link_libraries(llava_shared PRIVATE ggml llama ${CMAKE_THREAD_LIBS_INIT}) + install(TARGETS llava_shared LIBRARY) endif() -set(TARGET llava) -add_executable(${TARGET} llava.cpp) -install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT}) -target_compile_features(${TARGET} PRIVATE cxx_std_11) +if (NOT MSVC) + target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h + endif() if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) + add_dependencies(llava BUILD_INFO) endif() + +set(TARGET llava-cli) +add_executable(llava-cli llava-cli.cpp) +install(TARGETS llava-cli RUNTIME) +target_link_libraries(llava-cli PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(llava PRIVATE cxx_std_11) diff --git a/examples/llava/README.md b/examples/llava/README.md index fc3446b60..323c5fdd0 100644 --- a/examples/llava/README.md +++ b/examples/llava/README.md @@ -9,12 +9,12 @@ models are available. After API is confirmed, more models will be supported / uploaded. ## Usage -Build with cmake or run `make llava` to build it. +Build with cmake or run `make llava-cli` to build it. -After building, run: `./llava` to see the usage. For example: +After building, run: `./llava-cli` to see the usage. For example: ```sh -./llava -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg +./llava-cli -m llava-v1.5-7b/ggml-model-q5_k.gguf --mmproj llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg ``` **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so. @@ -51,7 +51,6 @@ Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` director ## TODO -- [ ] Support server mode. - [ ] Support non-CPU backend for the image encoding part. - [ ] Support different sampling methods. - [ ] Support more model variants. diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp index 61932e659..fc0656c23 100644 --- a/examples/llava/clip.cpp +++ b/examples/llava/clip.cpp @@ -664,7 +664,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { // measure mem requirement and allocate { static const size_t tensor_alignment = 32; - new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead()); + new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead()); new_clip->alloc = ggml_allocr_new_measure(tensor_alignment); clip_image_f32_batch batch; batch.size = 1; @@ -680,26 +680,44 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) { return new_clip; } -clip_image_u8 * make_clip_image_u8() { return new clip_image_u8(); } - +clip_image_u8 * make_clip_image_u8() { + auto img = new clip_image_u8(); + return img; +} clip_image_f32 * make_clip_image_f32() { return new clip_image_f32(); } -bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) { - int nx, ny, nc; - auto data = stbi_load(fname, &nx, &ny, &nc, 3); - if (!data) { - fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname); - return false; - } +void clip_image_u8_free(clip_image_u8 * img) { if (img->data) { delete[] img->data; } delete img; } +void clip_image_f32_free(clip_image_f32 * img) { if (img->data) { delete[] img->data; } delete img; } +static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) { img->nx = nx; img->ny = ny; img->size = nx * ny * 3; img->data = new uint8_t[img->size](); memcpy(img->data, data, img->size); +} +bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) { + int nx, ny, nc; + auto data = stbi_load(fname, &nx, &ny, &nc, 3); + if (!data) { + fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname); + return false; + } + build_clip_img_from_data(data, nx, ny, img); stbi_image_free(data); + return true; +} +bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) { + int nx, ny, nc; + auto data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3); + if (!data) { + fprintf(stderr, "%s: failed to decode image bytes\n", __func__); + return false; + } + build_clip_img_from_data(data, nx, ny, img); + stbi_image_free(data); return true; } @@ -714,39 +732,40 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104) // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156 - clip_image_u8 temp; // we will keep the input image data here temporarily + clip_image_u8 * temp = make_clip_image_u8(); // we will keep the input image data here temporarily if (pad2square && img->nx != img->ny) { int longer_side = std::max(img->nx, img->ny); - temp.nx = longer_side; - temp.ny = longer_side; - temp.size = 3 * longer_side * longer_side; - temp.data = new uint8_t[temp.size](); + temp->nx = longer_side; + temp->ny = longer_side; + temp->size = 3 * longer_side * longer_side; + temp->data = new uint8_t[temp->size](); uint8_t bc[3] = {122, 116, 104}; // bakground color in RGB from LLaVA // fill with background color - for (size_t i = 0; i < temp.size; i++) { - temp.data[i] = bc[i % 3]; + for (size_t i = 0; i < temp->size; i++) { + temp->data[i] = bc[i % 3]; } // copy from the input image for (int y = 0; y < img->ny; y++) { for (int x = 0; x < img->nx; x++) { const int i = 3 * (y * img->nx + x); - const int j = 3 * (y * temp.nx + x); - temp.data[j] = img->data[i]; - temp.data[j+1] = img->data[i+1]; - temp.data[j+2] = img->data[i+2]; + const int j = 3 * (y * temp->nx + x); + temp->data[j] = img->data[i]; + temp->data[j+1] = img->data[i+1]; + temp->data[j+2] = img->data[i+2]; } } } else { - temp.nx = img->nx; - temp.ny = img->ny; - temp.size = img->size; - temp.data = img->data; + temp->nx = img->nx; + temp->ny = img->ny; + temp->size = img->size; + temp->data = new uint8_t[temp->size](); + memcpy(&temp->data[0], &img->data[0], temp->size); // copy } - const int nx = temp.nx; - const int ny = temp.ny; + const int nx = temp->nx; + const int ny = temp->ny; const int nx2 = ctx->vision_model.hparams.image_size; const int ny2 = ctx->vision_model.hparams.image_size; @@ -785,10 +804,10 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip const int j10 = 3 * (y1 * nx + x0) + c; const int j11 = 3 * (y1 * nx + x1) + c; - const float v00 = temp.data[j00]; - const float v01 = temp.data[j01]; - const float v10 = temp.data[j10]; - const float v11 = temp.data[j11]; + const float v00 = temp->data[j00]; + const float v01 = temp->data[j01]; + const float v10 = temp->data[j10]; + const float v11 = temp->data[j11]; const float v0 = v00 * (1.0f - dx) + v01 * dx; const float v1 = v10 * (1.0f - dx) + v11 * dx; @@ -803,6 +822,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip } } } + clip_image_u8_free(temp); return true; } @@ -1049,16 +1069,16 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i return true; } -int clip_n_mmproj_embd(struct clip_ctx * ctx) { +int clip_n_mmproj_embd(const struct clip_ctx * ctx) { return ctx->vision_model.mm_2_b->ne[0]; } -int clip_n_patches(struct clip_ctx * ctx) { +int clip_n_patches(const struct clip_ctx * ctx) { auto & params = ctx->vision_model.hparams; return (params.image_size / params.patch_size) * (params.image_size / params.patch_size); } -size_t clip_embd_nbytes(struct clip_ctx * ctx) { +size_t clip_embd_nbytes(const struct clip_ctx * ctx) { return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float); } diff --git a/examples/llava/clip.h b/examples/llava/clip.h index 3d7261e29..f11df85de 100644 --- a/examples/llava/clip.h +++ b/examples/llava/clip.h @@ -1,7 +1,22 @@ #ifndef CLIP_H #define CLIP_H -#include "ggml.h" +#include +#include + +#ifdef LLAMA_SHARED +# if defined(_WIN32) && !defined(__MINGW32__) +# ifdef LLAMA_BUILD +# define CLIP_API __declspec(dllexport) +# else +# define CLIP_API __declspec(dllimport) +# endif +# else +# define CLIP_API __attribute__ ((visibility ("default"))) +# endif +#else +# define CLIP_API +#endif struct clip_ctx; @@ -20,19 +35,20 @@ struct clip_vision_hparams { float eps; }; -struct clip_ctx * clip_model_load(const char * fname, const int verbosity); +/** load mmproj model */ +CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity); +/** free mmproj model */ +CLIP_API void clip_free(struct clip_ctx * ctx); -void clip_free(struct clip_ctx * ctx); - -size_t clip_embd_nbytes(struct clip_ctx * ctx); -int clip_n_patches(struct clip_ctx * ctx); -int clip_n_mmproj_embd(struct clip_ctx * ctx); +size_t clip_embd_nbytes(const struct clip_ctx * ctx); +int clip_n_patches(const struct clip_ctx * ctx); +int clip_n_mmproj_embd(const struct clip_ctx * ctx); // RGB uint8 image struct clip_image_u8 { int nx; int ny; - uint8_t * data; + uint8_t * data = NULL; size_t size; }; @@ -41,7 +57,7 @@ struct clip_image_u8 { struct clip_image_f32 { int nx; int ny; - float * data; + float * data = NULL; size_t size; }; @@ -57,7 +73,12 @@ struct clip_image_f32_batch { struct clip_image_u8 * make_clip_image_u8(); struct clip_image_f32 * make_clip_image_f32(); -bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); +CLIP_API void clip_image_u8_free(clip_image_u8 * img); +CLIP_API void clip_image_f32_free(clip_image_f32 * img); +CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img); +/** interpret bytes as an image file with length bytes_length, and use the result to populate img */ +CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img); + bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square); bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec); diff --git a/examples/llava/llava-cli.cpp b/examples/llava/llava-cli.cpp new file mode 100644 index 000000000..31f8cd8e0 --- /dev/null +++ b/examples/llava/llava-cli.cpp @@ -0,0 +1,314 @@ +#include "ggml.h" +#include "common.h" +#include "clip.h" +#include "llava.h" +#include "llama.h" + +#include "base64.hpp" + +#include +#include +#include + +static bool eval_tokens(struct llama_context * ctx_llama, std::vector tokens, int n_batch, int * n_past) { + int N = (int) tokens.size(); + for (int i = 0; i < N; i += n_batch) { + int n_eval = (int) tokens.size() - i; + if (n_eval > n_batch) { + n_eval = n_batch; + } + if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) { + fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past); + return false; + } + *n_past += n_eval; + } + return true; +} + +static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) { + std::vector tokens; + tokens.push_back(id); + return eval_tokens(ctx_llama, tokens, 1, n_past); +} + +static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){ + std::string str2 = str; + std::vector embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos); + eval_tokens(ctx_llama, embd_inp, n_batch, n_past); + return true; +} + +// TODO: use common/sampling.h +static llama_token sample_id(llama_context * ctx_llama, gpt_params & params) { + auto & sparams = params.sparams; + + // out of user input, sample next token + const float temp = sparams.temp; + const int32_t top_k = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k; + const float top_p = sparams.top_p; + const float tfs_z = sparams.tfs_z; + const float typical_p = sparams.typical_p; + // const int32_t repeat_last_n = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n; + // const float repeat_penalty = sparams.repeat_penalty; + // const float alpha_presence = sparams.presence_penalty; + // const float alpha_frequency = sparams.frequency_penalty; + const int mirostat = sparams.mirostat; + const float mirostat_tau = sparams.mirostat_tau; + const float mirostat_eta = sparams.mirostat_eta; + // const bool penalize_nl = sparams.penalize_nl; + + llama_token id = 0; + { + auto logits = llama_get_logits(ctx_llama); + auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama)); + + // Apply params.logit_bias map + for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) { + logits[it->first] += it->second; + } + + std::vector candidates; + candidates.reserve(n_vocab); + for (llama_token token_id = 0; token_id < n_vocab; token_id++) { + candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); + } + + llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; + + if (temp <= 0) { + // Greedy sampling + id = llama_sample_token_greedy(ctx_llama, &candidates_p); + } else { + if (mirostat == 1) { + static float mirostat_mu = 2.0f * mirostat_tau; + const int mirostat_m = 100; + llama_sample_temp(ctx_llama, &candidates_p, temp); + id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); + } else if (mirostat == 2) { + static float mirostat_mu = 2.0f * mirostat_tau; + llama_sample_temp(ctx_llama, &candidates_p, temp); + id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu); + } else { + // Temperature sampling + llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1); + llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1); + llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1); + llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1); + llama_sample_temp(ctx_llama, &candidates_p, temp); + id = llama_sample_token(ctx_llama, &candidates_p); + } + } + } + + return id; +} + +static const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) { + int id = sample_id(ctx_llama, params); + static std::string ret; + if (id == llama_token_eos(llama_get_model(ctx_llama))) { + ret = ""; + } else { + ret = llama_token_to_piece(ctx_llama, id); + } + eval_id(ctx_llama, id, n_past); + return ret.c_str(); +} + +static const char* IMG_BASE64_TAG_BEGIN = ""; + +static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) { + begin_out = prompt.find(IMG_BASE64_TAG_BEGIN); + end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out); +} + +static bool prompt_contains_image(const std::string& prompt) { + size_t begin, end; + find_image_tag_in_prompt(prompt, begin, end); + return (begin != std::string::npos); +} + +// replaces the base64 image tag in the prompt with `replacement` +static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) { + size_t img_base64_str_start, img_base64_str_end; + find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end); + if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) { + fprintf(stderr, "%s: invalid base64 image tag. must be %s%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END); + return NULL; + } + + auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN); + auto base64_bytes_count = img_base64_str_end - base64_bytes_start; + auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count ); + + auto required_bytes = base64::required_encode_size(base64_str.size()); + auto img_bytes = std::vector(required_bytes); + base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin()); + + auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size()); + if (!embed) { + fprintf(stderr, "%s: could not load image from base64 string.\n", __func__); + return NULL; + } + + return embed; +} + +static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") { + size_t begin, end; + find_image_tag_in_prompt(prompt, begin, end); + if (begin == std::string::npos || end == std::string::npos) { + return prompt; + } + auto pre = prompt.substr(0, begin); + auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END)); + return pre + replacement + post; +} + +struct llava_context { + struct clip_ctx * ctx_clip = NULL; + struct llama_context * ctx_llama = NULL; + struct llama_model * model = NULL; +}; + +static void show_additional_info(int /*argc*/, char ** argv) { + printf("\n example usage: %s -m --mmproj --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); + printf(" note: a lower temperature value like 0.1 is recommended for better quality.\n"); +} + +static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) { + + // load and preprocess the image + llava_image_embed * embed = NULL; + auto prompt = params->prompt; + if (prompt_contains_image(prompt)) { + if (!params->image.empty()) { + printf("using base64 encoded image instead of command line image path\n"); + } + embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt); + if (!embed) { + fprintf(stderr, "%s: can't load image from prompt\n", __func__); + return NULL; + } + params->prompt = remove_image_from_prompt(prompt); + } else { + embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str()); + if (!embed) { + fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str()); + return NULL; + } + } + + return embed; +} + +static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) { + int n_past = 0; + + const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama)); + + // llava chat format is "\nUSER:\n\nASSISTANT:" + eval_string(ctx_llava->ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params->n_batch, &n_past, add_bos); + llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past); + eval_string(ctx_llava->ctx_llama, (prompt + "\nASSISTANT:").c_str(), params->n_batch, &n_past, false); + + // generate the response + + printf("\n"); + + for (int i = 0; i < max_tgt_len; i++) { + const char * tmp = sample(ctx_llava->ctx_llama, *params, &n_past); + if (strcmp(tmp, "") == 0) break; + + printf("%s", tmp); + fflush(stdout); + } + + printf("\n"); +} + + +static struct llava_context * llava_init(gpt_params * params) { + const char * clip_path = params->mmproj.c_str(); + + auto prompt = params->prompt; + if (prompt.empty()) { + prompt = "describe the image in detail."; + } + + auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); + + llama_backend_init(params->numa); + + llama_model_params model_params = llama_model_params_from_gpt_params(*params); + + llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params); + if (model == NULL) { + fprintf(stderr , "%s: error: unable to load model\n" , __func__); + return NULL; + } + + llama_context_params ctx_params = llama_context_params_from_gpt_params(*params); + ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings + + llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params); + + if (ctx_llama == NULL) { + fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__); + return NULL; + } + + auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context)); + + ctx_llava->ctx_llama = ctx_llama; + ctx_llava->ctx_clip = ctx_clip; + ctx_llava->model = model; + return ctx_llava; +} + +static void llava_free(struct llava_context * ctx_llava) { + if (ctx_llava->ctx_clip) { + clip_free(ctx_llava->ctx_clip); + ctx_llava->ctx_clip = NULL; + } + + llama_free(ctx_llava->ctx_llama); + llama_free_model(ctx_llava->model); + llama_backend_free(); +} + +int main(int argc, char ** argv) { + ggml_time_init(); + + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + show_additional_info(argc, argv); + return 1; + } + if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) { + gpt_print_usage(argc, argv, params); + show_additional_info(argc, argv); + return 1; + } + + auto ctx_llava = llava_init(¶ms); + if (ctx_llava == NULL) { + fprintf(stderr, "%s: error: failed to init llava\n", __func__); + return 1; + } + + auto image_embed = load_image(ctx_llava, ¶ms); + + // process the prompt + process_prompt(ctx_llava, image_embed, ¶ms, params.prompt); + + llama_print_timings(ctx_llava->ctx_llama); + + llava_image_embed_free(image_embed); + llava_free(ctx_llava); + return 0; +} diff --git a/examples/llava/llava-utils.h b/examples/llava/llava-utils.h deleted file mode 100644 index 320c71967..000000000 --- a/examples/llava/llava-utils.h +++ /dev/null @@ -1,147 +0,0 @@ -#pragma once - -// this one and clip lib will be eventually merged to a single lib, let's keep it this way for now - -#include "common.h" -#include "llama.h" - -#include -#include -#include - -inline bool eval_image_embd(llama_context * ctx_llama, float * embd, int N, int n_batch, int * n_past) { - int n_embd = llama_n_embd(llama_get_model(ctx_llama)); - - for (int i = 0; i < N; i += n_batch) { - int n_eval = N - i; - if (n_eval > n_batch) { - n_eval = n_batch; - } - llama_batch batch = {int32_t(n_eval), nullptr, (embd+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, }; - if (llama_decode(ctx_llama, batch)) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return false; - } - *n_past += n_eval; - } - return true; -} - -inline bool eval_tokens(struct llama_context * ctx_llama, std::vector tokens, int n_batch, int * n_past) { - int N = (int) tokens.size(); - for (int i = 0; i < N; i += n_batch) { - int n_eval = (int) tokens.size() - i; - if (n_eval > n_batch) { - n_eval = n_batch; - } - if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) { - fprintf(stderr, "%s : failed to eval\n", __func__); - return false; - } - *n_past += n_eval; - } - return true; -} - -inline bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) { - std::vector tokens; - tokens.push_back(id); - return eval_tokens(ctx_llama, tokens, 1, n_past); -} - -inline bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){ - std::string str2 = str; - std::vector embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos); - eval_tokens(ctx_llama, embd_inp, n_batch, n_past); - return true; -} - -// TODO: use common/sampling.h -inline llama_token sample_id(llama_context * ctx_llama, gpt_params & params) { - auto & sparams = params.sparams; - - // out of user input, sample next token - const float temp = sparams.temp; - const int32_t top_k = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k; - const float top_p = sparams.top_p; - const float tfs_z = sparams.tfs_z; - const float typical_p = sparams.typical_p; - // const int32_t repeat_last_n = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n; - // const float repeat_penalty = sparams.repeat_penalty; - // const float alpha_presence = sparams.presence_penalty; - // const float alpha_frequency = sparams.frequency_penalty; - const int mirostat = sparams.mirostat; - const float mirostat_tau = sparams.mirostat_tau; - const float mirostat_eta = sparams.mirostat_eta; - // const bool penalize_nl = sparams.penalize_nl; - - llama_token id = 0; - { - auto logits = llama_get_logits(ctx_llama); - auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama)); - - // Apply params.logit_bias map - for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) { - logits[it->first] += it->second; - } - - std::vector candidates; - candidates.reserve(n_vocab); - for (llama_token token_id = 0; token_id < n_vocab; token_id++) { - candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f}); - } - - llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false }; - - // TODO: Apply penalties - // float nl_logit = logits[llama_token_nl(ctx)]; - // auto last_n_repeat = std::min(std::min((int)last_n_tokens.size(), repeat_last_n), n_ctx); - // llama_sample_repetition_penalty(ctx, &candidates_p, - // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, - // last_n_repeat, repeat_penalty); - // llama_sample_frequency_and_presence_penalties(ctx, &candidates_p, - // last_n_tokens.data() + last_n_tokens.size() - last_n_repeat, - // last_n_repeat, alpha_frequency, alpha_presence); - // if (!penalize_nl) { - // logits[llama_token_nl(ctx)] = nl_logit; - // } - - if (temp <= 0) { - // Greedy sampling - id = llama_sample_token_greedy(ctx_llama, &candidates_p); - } else { - if (mirostat == 1) { - static float mirostat_mu = 2.0f * mirostat_tau; - const int mirostat_m = 100; - llama_sample_temp(ctx_llama, &candidates_p, temp); - id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu); - } else if (mirostat == 2) { - static float mirostat_mu = 2.0f * mirostat_tau; - llama_sample_temp(ctx_llama, &candidates_p, temp); - id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu); - } else { - // Temperature sampling - llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1); - llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1); - llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1); - llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1); - llama_sample_temp(ctx_llama, &candidates_p, temp); - id = llama_sample_token(ctx_llama, &candidates_p); - } - } - } - - return id; -} - -inline const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) { - int id = sample_id(ctx_llama, params); - static std::string ret; - if (id == llama_token_eos(llama_get_model(ctx_llama))) { - ret = ""; - } else { - ret = llama_token_to_piece(ctx_llama, id); - } - eval_id(ctx_llama, id, n_past); - return ret.c_str(); -} diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp index f0974d5bc..0cae8c4b1 100644 --- a/examples/llava/llava.cpp +++ b/examples/llava/llava.cpp @@ -1,164 +1,163 @@ #include "clip.h" -#include "llava-utils.h" #include "common.h" #include "llama.h" +#include "llava.h" #include #include #include -static void show_additional_info(int /*argc*/, char ** argv) { - printf("\n example usage: %s -m --mmproj --image [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]); - printf(" note: a lower temperature value like 0.1 is recommended for better quality.\n"); -} +#include "base64.hpp" -int main(int argc, char ** argv) { - ggml_time_init(); - - gpt_params params; - - if (!gpt_params_parse(argc, argv, params)) { - show_additional_info(argc, argv); - return 1; +static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) { + clip_image_f32 * img_res = make_clip_image_f32(); + if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) { + fprintf(stderr, "%s: unable to preprocess image\n", __func__); + clip_image_f32_free(img_res); + return false; } - if (params.mmproj.empty() || params.image.empty()) { - gpt_print_usage(argc, argv, params); - show_additional_info(argc, argv); - return 1; - } - - const char * clip_path = params.mmproj.c_str(); - const char * img_path = params.image.c_str(); - - if (params.prompt.empty()) { - params.prompt = "describe the image in detail."; - } - - auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1); - - // load and preprocess the image - clip_image_u8 img; - clip_image_f32 img_res; - - if (!clip_image_load_from_file(img_path, &img)) { - fprintf(stderr, "%s: is %s really an image file?\n", __func__, img_path); - - clip_free(ctx_clip); - return 1; - } - - if (!clip_image_preprocess(ctx_clip, &img, &img_res, /*pad2square =*/ true)) { - fprintf(stderr, "%s: unable to preprocess %s\n", __func__, img_path); - - clip_free(ctx_clip); - return 1; - } - - int n_img_pos = clip_n_patches(ctx_clip); - int n_img_embd = clip_n_mmproj_embd(ctx_clip); - - float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)); - - if (!image_embd) { - fprintf(stderr, "Unable to allocate memory for image embeddings\n"); - - return 1; - } + *n_img_pos = clip_n_patches(ctx_clip); const int64_t t_img_enc_start_us = ggml_time_us(); - if (!clip_image_encode(ctx_clip, params.n_threads, &img_res, image_embd)) { + bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); + clip_image_f32_free(img_res); + if (!encoded) { fprintf(stderr, "Unable to encode image\n"); - return 1; + return false; } + const int64_t t_img_enc_end_us = ggml_time_us(); + float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0; - // we get the embeddings, free up the memory required for CLIP - clip_free(ctx_clip); + printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos); - llama_backend_init(params.numa); - - llama_model_params model_params = llama_model_default_params(); - model_params.n_gpu_layers = params.n_gpu_layers; - model_params.main_gpu = params.main_gpu; - model_params.tensor_split = params.tensor_split; - model_params.use_mmap = params.use_mmap; - model_params.use_mlock = params.use_mlock; - - llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params); - if (model == NULL) { - fprintf(stderr , "%s: error: unable to load model\n" , __func__); - return 1; - } - - llama_context_params ctx_params = llama_context_default_params(); - - ctx_params.n_ctx = params.n_ctx < 2048 ? 2048 : params.n_ctx; // we need a longer context size to process image embeddings - ctx_params.n_threads = params.n_threads; - ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; - ctx_params.seed = params.seed; - - llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params); - - if (ctx_llama == NULL) { - fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__); - return 1; - } - - // make sure that the correct mmproj was used, i.e., compare apples to apples - const int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama)); - - if (n_img_embd != n_llama_embd) { - printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_img_embd, n_llama_embd); - - llama_free(ctx_llama); - llama_free_model(model); - llama_backend_free(); - free(image_embd); - - return 1; - } - - // process the prompt - // llava chat format is "USER: \n\nASSISTANT:" - - int n_past = 0; - - const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict; - - eval_string(ctx_llama, "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:", params.n_batch, &n_past, true); - eval_image_embd(ctx_llama, image_embd, n_img_pos, params.n_batch, &n_past); - eval_string(ctx_llama, (params.prompt + "\nASSISTANT:").c_str(), params.n_batch, &n_past, false); - - // generate the response - - printf("\n"); - printf("prompt: '%s'\n", params.prompt.c_str()); - printf("\n"); - - for (int i = 0; i < max_tgt_len; i++) { - const char * tmp = sample(ctx_llama, params, &n_past); - if (strcmp(tmp, "") == 0) break; - - printf("%s", tmp); - fflush(stdout); - } - - printf("\n"); - - { - const float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0; - - printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / n_img_pos); - } - - llama_print_timings(ctx_llama); - - llama_free(ctx_llama); - llama_free_model(model); - llama_backend_free(); - free(image_embd); - - return 0; + return true; +} + +bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) { + // make sure that the correct mmproj was used, i.e., compare apples to apples + int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama)); + auto n_image_embd = clip_n_mmproj_embd(ctx_clip); + if (n_image_embd != n_llama_embd) { + printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd); + return false; + } + return true; +} + +static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) { + float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)); + if (!image_embd) { + fprintf(stderr, "Unable to allocate memory for image embeddings\n"); + free(image_embd); + return false; + } + + int n_img_pos; + if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) { + fprintf(stderr, "%s: cannot encode image, aborting\n", __func__); + free(image_embd); + return false; + } + *image_embd_out = image_embd; + *n_img_pos_out = n_img_pos; + + return true; +} + +bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) { + int n_embd = llama_n_embd(llama_get_model(ctx_llama)); + + for (int i = 0; i < image_embed->n_image_pos; i += n_batch) { + int n_eval = image_embed->n_image_pos - i; + if (n_eval > n_batch) { + n_eval = n_batch; + } + llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, }; + if (llama_decode(ctx_llama, batch)) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return false; + } + *n_past += n_eval; + } + return true; +} + +LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) { + clip_image_u8 * img = make_clip_image_u8(); + if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) { + clip_image_u8_free(img); + fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__); + return NULL; + } + + float* image_embed = NULL; + int n_image_pos = 0; + bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos); + if (!image_embed_result) { + clip_image_u8_free(img); + fprintf(stderr, "%s: coulnd't embed the image\n", __func__); + return NULL; + } + + clip_image_u8_free(img); + auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed)); + result->embed = image_embed; + result->n_image_pos = n_image_pos; + return result; +} + +static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) { + auto file = fopen(path, "rb"); + if (file == NULL) { + fprintf(stderr, "%s: can't read file %s\n", __func__, path); + return false; + } + + fseek(file, 0, SEEK_END); + auto fileSize = ftell(file); + fseek(file, 0, SEEK_SET); + + auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data + if (buffer == NULL) { + fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path); + perror("Memory allocation error"); + fclose(file); + return false; + } + errno = 0; + size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer + if (ferror(file)) { + die_fmt("read error: %s", strerror(errno)); + } + if (ret != (size_t) fileSize) { + die("unexpectedly reached end of file"); + } + fclose(file); // Close the file + + *bytesOut = buffer; + *sizeOut = fileSize; + return true; +} + +LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) { + unsigned char* image_bytes; + long image_bytes_length; + auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length); + if (!loaded) { + fprintf(stderr, "%s: failed to load %s\n", __func__, image_path); + return NULL; + } + + auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length); + free(image_bytes); + + return embed; +} + +LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) { + free(embed->embed); + free(embed); } diff --git a/examples/llava/llava.h b/examples/llava/llava.h new file mode 100644 index 000000000..e08ce7883 --- /dev/null +++ b/examples/llava/llava.h @@ -0,0 +1,50 @@ +#ifndef LLAVA_H +#define LLAVA_H + +#include "ggml.h" + + +#ifdef LLAMA_SHARED +# if defined(_WIN32) && !defined(__MINGW32__) +# ifdef LLAMA_BUILD +# define LLAVA_API __declspec(dllexport) +# else +# define LLAVA_API __declspec(dllimport) +# endif +# else +# define LLAVA_API __attribute__ ((visibility ("default"))) +# endif +#else +# define LLAVA_API +#endif + +struct clip_ctx; + +#ifdef __cplusplus +extern "C" { +#endif + +struct llava_image_embed { + float * embed; + int n_image_pos; +}; + +/** sanity check for clip <-> llava embed size match */ +LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip); + +/** build an image embed from image file bytes */ +LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length); +/** build an image embed from a path to an image filename */ +LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path); +LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed); +/** free an embedding made with llava_image_embed_make_* */ + +/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */ +LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past); + + +#ifdef __cplusplus +} +#endif + +#endif diff --git a/examples/main/CMakeLists.txt b/examples/main/CMakeLists.txt index cc1888948..d532980b7 100644 --- a/examples/main/CMakeLists.txt +++ b/examples/main/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} main.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/main/README.md b/examples/main/README.md index a9561c383..c7997f665 100644 --- a/examples/main/README.md +++ b/examples/main/README.md @@ -142,7 +142,7 @@ The `--ctx-size` option allows you to set the size of the prompt context used by ### Extended Context Size -Some fine-tuned models have extened the context length by scaling RoPE. For example, if the original pretrained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8. +Some fine-tuned models have extended the context length by scaling RoPE. For example, if the original pre-trained model have a context length (max sequence length) of 4096 (4k) and the fine-tuned model have 32k. That is a scaling factor of 8, and should work by setting the above `--ctx-size` to 32768 (32k) and `--rope-scale` to 8. - `--rope-scale N`: Where N is the linear scaling factor used by the fine-tuned model. @@ -208,6 +208,14 @@ Top-p sampling, also known as nucleus sampling, is another text generation metho Example usage: `--top-p 0.95` +### Min P Sampling + +- `--min-p N`: Sets a minimum base probability threshold for token selection (default: 0.05). + +The Min-P sampling method was designed as an alternative to Top-P, and aims to ensure a balance of quality and variety. The parameter *p* represents the minimum probability for a token to be considered, relative to the probability of the most likely token. For example, with *p*=0.05 and the most likely token having a probability of 0.9, logits with a value less than 0.045 are filtered out. + +Example usage: `--min-p 0.05` + ### Tail Free Sampling (TFS) - `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled). diff --git a/examples/main/main.cpp b/examples/main/main.cpp index 8a43b6ab8..31ec8cade 100644 --- a/examples/main/main.cpp +++ b/examples/main/main.cpp @@ -2,7 +2,6 @@ #include "console.h" #include "llama.h" -#include "build-info.h" #include #include @@ -153,8 +152,8 @@ int main(int argc, char ** argv) { LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale); } - LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT); - LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET); + LOG_TEE("%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT); + LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET); if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -230,13 +229,16 @@ int main(int argc, char ** argv) { } } - const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM; + const bool add_bos = llama_should_add_bos_token(model); LOG("add_bos: %d\n", add_bos); std::vector embd_inp; - if (params.interactive_first || params.instruct || !params.prompt.empty() || session_tokens.empty()) { + if (params.interactive_first || params.instruct || params.chatml || !params.prompt.empty() || session_tokens.empty()) { LOG("tokenize the prompt\n"); + if (params.chatml) { + params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>"; + } embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true); } else { LOG("use session tokens\n"); @@ -314,7 +316,7 @@ int main(int argc, char ** argv) { } // number of tokens to keep when resetting context - if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) { + if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) { params.n_keep = (int)embd_inp.size(); } @@ -325,11 +327,23 @@ int main(int argc, char ** argv) { LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str()); LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str()); + // chatml prefix & suffix + const auto cml_pfx = ::llama_tokenize(ctx, "\n<|im_start|>user\n", add_bos, true); + const auto cml_sfx = ::llama_tokenize(ctx, "<|im_end|>\n<|im_start|>assistant\n", false, true); + + LOG("cml_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_pfx).c_str()); + LOG("cml_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_sfx).c_str()); + // in instruct mode, we inject a prefix and a suffix to each input by the user if (params.instruct) { params.interactive_first = true; params.antiprompt.push_back("### Instruction:\n\n"); } + // similar for chatml mode + else if (params.chatml) { + params.interactive_first = true; + params.antiprompt.push_back("<|im_start|>user\n"); + } // enable interactive mode if interactive start is specified if (params.interactive_first) { @@ -706,7 +720,7 @@ int main(int argc, char ** argv) { is_interacting = true; printf("\n"); - } else if (params.instruct) { + } else if (params.instruct || params.chatml) { is_interacting = true; } } @@ -714,7 +728,7 @@ int main(int argc, char ** argv) { if (n_past > 0 && is_interacting) { LOG("waiting for user input\n"); - if (params.instruct) { + if (params.instruct || params.chatml) { printf("\n> "); } @@ -761,6 +775,12 @@ int main(int argc, char ** argv) { n_consumed = embd_inp.size(); embd_inp.insert(embd_inp.end(), inp_pfx.begin(), inp_pfx.end()); } + // chatml mode: insert user chat prefix + if (params.chatml && !is_antiprompt) { + LOG("inserting chatml prefix\n"); + n_consumed = embd_inp.size(); + embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end()); + } if (params.escape) { process_escapes(buffer); } @@ -779,6 +799,11 @@ int main(int argc, char ** argv) { LOG("inserting instruction suffix\n"); embd_inp.insert(embd_inp.end(), inp_sfx.begin(), inp_sfx.end()); } + // chatml mode: insert assistant chat suffix + if (params.chatml) { + LOG("inserting chatml suffix\n"); + embd_inp.insert(embd_inp.end(), cml_sfx.begin(), cml_sfx.end()); + } for (size_t i = original_size; i < embd_inp.size(); ++i) { const llama_token token = embd_inp[i]; @@ -804,7 +829,7 @@ int main(int argc, char ** argv) { } // end of text token - if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive)) { + if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive || params.chatml)) { LOG_TEE(" [end of text]\n"); break; } diff --git a/examples/metal/metal.cpp b/examples/metal/metal.cpp index c05a4fa93..16c1146f9 100644 --- a/examples/metal/metal.cpp +++ b/examples/metal/metal.cpp @@ -34,7 +34,7 @@ int main(int argc, char ** argv) { struct ggml_context * ctx_data = NULL; struct ggml_context * ctx_eval = NULL; - struct ggml_cgraph gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval); + struct ggml_cgraph * gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval); // this allocates all Metal resources and memory buffers auto * ctx_metal = ggml_metal_init(1); @@ -46,13 +46,13 @@ int main(int argc, char ** argv) { // main { - struct ggml_tensor * input = ggml_graph_get_tensor(&gf, "embd"); + struct ggml_tensor * input = ggml_graph_get_tensor(gf, "embd"); *(int32_t *) input->data = 1; // BOS ggml_metal_set_tensor(ctx_metal, input); // warmup - ggml_metal_graph_compute(ctx_metal, &gf); + ggml_metal_graph_compute(ctx_metal, gf); const int n_iter = 16; @@ -60,7 +60,7 @@ int main(int argc, char ** argv) { // the actual inference happens here for (int i = 0; i < n_iter; ++i) { - ggml_metal_graph_compute(ctx_metal, &gf); + ggml_metal_graph_compute(ctx_metal, gf); } const int64_t t1 = ggml_time_us(); @@ -70,7 +70,7 @@ int main(int argc, char ** argv) { // debug output { - struct ggml_tensor * logits = gf.nodes[gf.n_nodes - 1]; + struct ggml_tensor * logits = gf->nodes[gf->n_nodes - 1]; ggml_metal_get_tensor(ctx_metal, logits); float * ptr = (float *) ggml_get_data(logits); diff --git a/examples/parallel/CMakeLists.txt b/examples/parallel/CMakeLists.txt index 0bbf89eae..319535a6e 100644 --- a/examples/parallel/CMakeLists.txt +++ b/examples/parallel/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} parallel.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/parallel/README.md b/examples/parallel/README.md index 4d0fe5cef..df0456733 100644 --- a/examples/parallel/README.md +++ b/examples/parallel/README.md @@ -1,3 +1,3 @@ # llama.cpp/example/parallel -Simplified simluation for serving incoming requests in parallel +Simplified simulation of serving incoming requests in parallel diff --git a/examples/parallel/parallel.cpp b/examples/parallel/parallel.cpp index 9a0b9c183..d2e074d9e 100644 --- a/examples/parallel/parallel.cpp +++ b/examples/parallel/parallel.cpp @@ -1,7 +1,5 @@ // A basic application simulating a server with multiple clients. -// The clients submite requests to the server and they are processed in parallel. - -#include "build-info.h" +// The clients submit requests to the server and they are processed in parallel. #include "common.h" #include "llama.h" @@ -115,6 +113,8 @@ int main(int argc, char ** argv) { // insert new requests as soon as the previous one is done const bool cont_batching = params.cont_batching; + const bool dump_kv_cache = params.dump_kv_cache; + #ifndef LOG_DISABLE_LOGS log_set_target(log_filename_generator("parallel", "log")); LOG_TEE("Log start\n"); @@ -174,6 +174,8 @@ int main(int argc, char ** argv) { int32_t n_total_gen = 0; int32_t n_cache_miss = 0; + struct llama_kv_cache_view kvc_view = llama_kv_cache_view_init(ctx, n_clients); + const auto t_main_start = ggml_time_us(); LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__); @@ -203,6 +205,11 @@ int main(int argc, char ** argv) { LOG_TEE("Processing requests ...\n\n"); while (true) { + if (dump_kv_cache) { + llama_kv_cache_view_update(ctx, &kvc_view); + dump_kv_cache_view_seqs(kvc_view, 40); + } + llama_batch_clear(batch); // decode any currently ongoing sequences diff --git a/examples/perplexity/CMakeLists.txt b/examples/perplexity/CMakeLists.txt index af00b4e16..3c76d3221 100644 --- a/examples/perplexity/CMakeLists.txt +++ b/examples/perplexity/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} perplexity.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/perplexity/perplexity.cpp b/examples/perplexity/perplexity.cpp index bd2c73d87..9a77beca6 100644 --- a/examples/perplexity/perplexity.cpp +++ b/examples/perplexity/perplexity.cpp @@ -1,4 +1,3 @@ -#include "build-info.h" #include "common.h" #include "llama.h" @@ -150,8 +149,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM; - const bool add_bos = is_spm; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); fprintf(stderr, "%s: tokenizing the input ..\n", __func__); @@ -289,8 +287,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par // Output: `perplexity: 13.5106 [114/114]` // BOS tokens will be added for each chunk before eval - const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM; - const bool add_bos = is_spm; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); const int n_ctx = llama_n_ctx(ctx); auto tim1 = std::chrono::high_resolution_clock::now(); @@ -482,7 +479,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) { fprintf(stderr, "================================= is_spm = %d\n", is_spm); // This is needed as usual for LLaMA models - const bool add_bos = is_spm; + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); // Number of tasks to use when computing the score if ( params.hellaswag_tasks < hs_task_count ) { diff --git a/examples/quantize-stats/CMakeLists.txt b/examples/quantize-stats/CMakeLists.txt index db182e263..e31cf5e38 100644 --- a/examples/quantize-stats/CMakeLists.txt +++ b/examples/quantize-stats/CMakeLists.txt @@ -1,6 +1,6 @@ set(TARGET quantize-stats) add_executable(${TARGET} quantize-stats.cpp) install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT}) target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/quantize-stats/quantize-stats.cpp b/examples/quantize-stats/quantize-stats.cpp index dd76b1cee..271282477 100644 --- a/examples/quantize-stats/quantize-stats.cpp +++ b/examples/quantize-stats/quantize-stats.cpp @@ -1,5 +1,4 @@ #define LLAMA_API_INTERNAL -#include "build-info.h" #include "common.h" #include "ggml.h" #include "llama.h" diff --git a/examples/quantize/CMakeLists.txt b/examples/quantize/CMakeLists.txt index 4a8eed544..6f374a2bd 100644 --- a/examples/quantize/CMakeLists.txt +++ b/examples/quantize/CMakeLists.txt @@ -1,9 +1,6 @@ set(TARGET quantize) add_executable(${TARGET} quantize.cpp) install(TARGETS ${TARGET} RUNTIME) -target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT}) target_include_directories(${TARGET} PRIVATE ../../common) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/quantize/quantize.cpp b/examples/quantize/quantize.cpp index be0b2fe1e..d27ea5e91 100644 --- a/examples/quantize/quantize.cpp +++ b/examples/quantize/quantize.cpp @@ -1,4 +1,3 @@ -#include "build-info.h" #include "common.h" #include "llama.h" diff --git a/examples/save-load-state/CMakeLists.txt b/examples/save-load-state/CMakeLists.txt index eadd13cdf..cc6ed8554 100644 --- a/examples/save-load-state/CMakeLists.txt +++ b/examples/save-load-state/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} save-load-state.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/save-load-state/save-load-state.cpp b/examples/save-load-state/save-load-state.cpp index 38d05f4d3..48d801110 100644 --- a/examples/save-load-state/save-load-state.cpp +++ b/examples/save-load-state/save-load-state.cpp @@ -1,4 +1,3 @@ -#include "build-info.h" #include "common.h" #include "llama.h" diff --git a/examples/server/CMakeLists.txt b/examples/server/CMakeLists.txt index a23ddcc55..859cd12c6 100644 --- a/examples/server/CMakeLists.txt +++ b/examples/server/CMakeLists.txt @@ -6,11 +6,8 @@ install(TARGETS ${TARGET} RUNTIME) target_compile_definitions(${TARGET} PRIVATE SERVER_VERBOSE=$ ) -target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT}) +target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT}) if (WIN32) TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32) endif() target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/server/README.md b/examples/server/README.md index 715007735..a6eda3b32 100644 --- a/examples/server/README.md +++ b/examples/server/README.md @@ -7,7 +7,7 @@ Command line options: - `--threads N`, `-t N`: Set the number of threads to use during generation. - `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`). -- `-m ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses. +- `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses. - `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096. - `-ngl N`, `--n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance. - `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS. @@ -122,6 +122,8 @@ node index.js `top_p`: Limit the next token selection to a subset of tokens with a cumulative probability above a threshold P (default: 0.95). + `min_p`: The minimum probability for a token to be considered, relative to the probability of the most likely token (default: 0.05). + `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. (default: -1, -1 = infinity). `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. diff --git a/examples/server/index.html.hpp b/examples/server/index.html.hpp index 5d3bdfbdd..f22b77e7f 100644 --- a/examples/server/index.html.hpp +++ b/examples/server/index.html.hpp @@ -229,247 +229,241 @@ unsigned char index_html[] = { 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x65, 0x69, 0x67, 0x68, 0x74, 0x3a, 0x20, 0x31, 0x30, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, 0x6b, 0x65, 0x79, 0x66, 0x72, - 0x61, 0x6d, 0x65, 0x73, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, - 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, 0x70, 0x65, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, - 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, - 0x6e, 0x3a, 0x20, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x31, 0x30, - 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, - 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x31, 0x30, - 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, - 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, - 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, - 0x6e, 0x64, 0x2d, 0x73, 0x69, 0x7a, 0x65, 0x3a, 0x20, 0x35, 0x30, 0x25, - 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, - 0x69, 0x6d, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, - 0x72, 0x2d, 0x67, 0x72, 0x61, 0x64, 0x69, 0x65, 0x6e, 0x74, 0x28, 0x39, - 0x30, 0x64, 0x65, 0x67, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x5b, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x65, 0x64, 0x69, 0x74, 0x61, 0x62, 0x6c, 0x65, 0x5d, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x69, 0x73, 0x70, 0x6c, + 0x61, 0x79, 0x3a, 0x20, 0x69, 0x6e, 0x6c, 0x69, 0x6e, 0x65, 0x2d, 0x62, + 0x6c, 0x6f, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x77, 0x68, 0x69, 0x74, 0x65, 0x2d, 0x73, 0x70, 0x61, 0x63, 0x65, 0x3a, + 0x20, 0x70, 0x72, 0x65, 0x2d, 0x77, 0x72, 0x61, 0x70, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6f, 0x75, 0x74, 0x6c, 0x69, 0x6e, 0x65, + 0x3a, 0x20, 0x30, 0x70, 0x78, 0x20, 0x73, 0x6f, 0x6c, 0x69, 0x64, 0x20, + 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x40, 0x6b, 0x65, 0x79, 0x66, 0x72, 0x61, 0x6d, 0x65, 0x73, 0x20, 0x6c, + 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, + 0x70, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x30, + 0x25, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, + 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x30, 0x25, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x31, 0x30, 0x30, 0x25, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, + 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, + 0x6f, 0x6e, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x31, 0x29, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, + 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, + 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x32, 0x29, 0x2c, 0x20, 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, + 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, 0x65, 0x65, 0x65, 0x65, 0x65, 0x65, + 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, + 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x73, 0x69, 0x7a, + 0x65, 0x3a, 0x20, 0x35, 0x30, 0x25, 0x20, 0x31, 0x30, 0x30, 0x25, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, + 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x3a, + 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, 0x72, 0x2d, 0x67, 0x72, 0x61, 0x64, + 0x69, 0x65, 0x6e, 0x74, 0x28, 0x39, 0x30, 0x64, 0x65, 0x67, 0x2c, 0x20, + 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x2c, 0x20, + 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x29, 0x2c, 0x20, + 0x76, 0x61, 0x72, 0x28, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x29, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6e, 0x69, 0x6d, 0x61, + 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, + 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, 0x69, 0x70, 0x65, 0x20, 0x32, 0x73, + 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, 0x72, 0x20, 0x69, 0x6e, 0x66, 0x69, + 0x6e, 0x69, 0x74, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, 0x6d, 0x65, 0x64, 0x69, 0x61, 0x20, + 0x28, 0x70, 0x72, 0x65, 0x66, 0x65, 0x72, 0x73, 0x2d, 0x63, 0x6f, 0x6c, + 0x6f, 0x72, 0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x65, 0x3a, 0x20, 0x64, + 0x61, 0x72, 0x6b, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, + 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, + 0x31, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x30, 0x30, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, - 0x72, 0x2d, 0x31, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x61, 0x6e, 0x69, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x3a, 0x20, - 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, 0x62, 0x67, 0x2d, 0x77, - 0x69, 0x70, 0x65, 0x20, 0x32, 0x73, 0x20, 0x6c, 0x69, 0x6e, 0x65, 0x61, - 0x72, 0x20, 0x69, 0x6e, 0x66, 0x69, 0x6e, 0x69, 0x74, 0x65, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x40, - 0x6d, 0x65, 0x64, 0x69, 0x61, 0x20, 0x28, 0x70, 0x72, 0x65, 0x66, 0x65, - 0x72, 0x73, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x73, 0x63, 0x68, - 0x65, 0x6d, 0x65, 0x3a, 0x20, 0x64, 0x61, 0x72, 0x6b, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x6c, 0x6f, 0x61, 0x64, - 0x69, 0x6e, 0x67, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, 0x67, 0x2d, - 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x31, 0x3a, 0x20, 0x23, 0x32, 0x32, - 0x32, 0x32, 0x32, 0x32, 0x30, 0x30, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2d, 0x2d, 0x6c, 0x6f, 0x61, 0x64, 0x69, 0x6e, - 0x67, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, - 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, - 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, - 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x62, 0x6c, - 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, - 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x3c, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x6d, - 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x2c, 0x20, 0x68, 0x2c, 0x20, - 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x65, 0x66, 0x66, 0x65, - 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, - 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x2c, 0x20, 0x75, 0x73, - 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x2c, 0x20, 0x75, 0x73, 0x65, - 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x52, - 0x65, 0x66, 0x2c, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, - 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, - 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x2e, 0x6a, 0x73, 0x27, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, - 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x7d, 0x20, - 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, - 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, - 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, - 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, - 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x2e, - 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, - 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, - 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x76, 0x61, 0x72, 0x20, 0x73, 0x6c, - 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, 0x3d, 0x20, 0x2d, 0x31, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, - 0x73, 0x20, 0x69, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, - 0x72, 0x73, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, - 0x65, 0x65, 0x6e, 0x20, 0x55, 0x73, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, - 0x69, 0x65, 0x6e, 0x64, 0x6c, 0x79, 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, - 0x6f, 0x74, 0x2e, 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, - 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, - 0x6e, 0x64, 0x2c, 0x20, 0x68, 0x6f, 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, - 0x67, 0x6f, 0x6f, 0x64, 0x20, 0x61, 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, - 0x69, 0x6e, 0x67, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, - 0x65, 0x72, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, - 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72, 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, - 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, - 0x64, 0x69, 0x61, 0x74, 0x65, 0x6c, 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, - 0x77, 0x69, 0x74, 0x68, 0x20, 0x70, 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, - 0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, - 0x6e, 0x7b, 0x7b, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, - 0x5c, 0x6e, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, - 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, - 0x20, 0x22, 0x7b, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, - 0x7b, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, - 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, - 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x2c, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x7c, 0x20, 0x22, 0x63, 0x6f, - 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, - 0x6c, 0x61, 0x6d, 0x61, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x75, 0x73, 0x65, 0x72, 0x3a, 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, - 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, - 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, - 0x20, 0x27, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, - 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, - 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, - 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, - 0x35, 0x36, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, - 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, - 0x74, 0x79, 0x2c, 0x20, 0x2d, 0x31, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x78, 0x74, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, - 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, + 0x72, 0x2d, 0x32, 0x3a, 0x20, 0x23, 0x32, 0x32, 0x32, 0x32, 0x32, 0x32, + 0x66, 0x66, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x70, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, + 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, 0x2d, 0x63, 0x6f, 0x6c, + 0x6f, 0x72, 0x3a, 0x20, 0x62, 0x6c, 0x61, 0x63, 0x6b, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3e, 0x0a, + 0x0a, 0x20, 0x20, 0x3c, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x74, + 0x79, 0x70, 0x65, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x75, 0x6c, 0x65, 0x22, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x2c, 0x20, 0x68, 0x2c, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, + 0x2c, 0x20, 0x65, 0x66, 0x66, 0x65, 0x63, 0x74, 0x2c, 0x20, 0x63, 0x6f, + 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, + 0x6c, 0x2c, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, + 0x2c, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x2c, 0x20, 0x43, 0x6f, + 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x69, 0x6e, 0x64, + 0x65, 0x78, 0x2e, 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x6c, 0x6c, + 0x61, 0x6d, 0x61, 0x20, 0x7d, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, + 0x2f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x2e, + 0x6a, 0x73, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x70, + 0x6f, 0x72, 0x74, 0x20, 0x7b, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, + 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x7d, 0x20, + 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x27, 0x2f, 0x6a, 0x73, 0x6f, 0x6e, 0x2d, + 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2d, 0x74, 0x6f, 0x2d, 0x67, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x2e, 0x6d, 0x6a, 0x73, 0x27, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, + 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x3d, + 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x76, 0x61, 0x72, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, + 0x3d, 0x20, 0x2d, 0x31, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, + 0x3a, 0x20, 0x22, 0x54, 0x68, 0x69, 0x73, 0x20, 0x69, 0x73, 0x20, 0x61, + 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20, 0x55, 0x73, + 0x65, 0x72, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x4c, 0x6c, 0x61, 0x6d, 0x61, + 0x2c, 0x20, 0x61, 0x20, 0x66, 0x72, 0x69, 0x65, 0x6e, 0x64, 0x6c, 0x79, + 0x20, 0x63, 0x68, 0x61, 0x74, 0x62, 0x6f, 0x74, 0x2e, 0x20, 0x4c, 0x6c, + 0x61, 0x6d, 0x61, 0x20, 0x69, 0x73, 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, + 0x75, 0x6c, 0x2c, 0x20, 0x6b, 0x69, 0x6e, 0x64, 0x2c, 0x20, 0x68, 0x6f, + 0x6e, 0x65, 0x73, 0x74, 0x2c, 0x20, 0x67, 0x6f, 0x6f, 0x64, 0x20, 0x61, + 0x74, 0x20, 0x77, 0x72, 0x69, 0x74, 0x69, 0x6e, 0x67, 0x2c, 0x20, 0x61, + 0x6e, 0x64, 0x20, 0x6e, 0x65, 0x76, 0x65, 0x72, 0x20, 0x66, 0x61, 0x69, + 0x6c, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72, + 0x20, 0x61, 0x6e, 0x79, 0x20, 0x72, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, + 0x73, 0x20, 0x69, 0x6d, 0x6d, 0x65, 0x64, 0x69, 0x61, 0x74, 0x65, 0x6c, + 0x79, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x77, 0x69, 0x74, 0x68, 0x20, 0x70, + 0x72, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x22, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, 0x70, 0x72, 0x6f, 0x6d, 0x70, + 0x74, 0x7d, 0x7d, 0x5c, 0x6e, 0x5c, 0x6e, 0x7b, 0x7b, 0x68, 0x69, 0x73, + 0x74, 0x6f, 0x72, 0x79, 0x7d, 0x7d, 0x5c, 0x6e, 0x7b, 0x7b, 0x63, 0x68, + 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x22, 0x7b, 0x7b, 0x6e, 0x61, + 0x6d, 0x65, 0x7d, 0x7d, 0x3a, 0x20, 0x7b, 0x7b, 0x6d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x7d, 0x7d, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, + 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x79, 0x70, 0x65, 0x3a, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, + 0x2c, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, + 0x20, 0x7c, 0x20, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x68, + 0x61, 0x72, 0x3a, 0x20, 0x22, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x22, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x3a, + 0x20, 0x22, 0x55, 0x73, 0x65, 0x72, 0x22, 0x2c, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, + 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, + 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, + 0x74, 0x3a, 0x20, 0x34, 0x30, 0x30, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, + 0x65, 0x3a, 0x20, 0x30, 0x2e, 0x37, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, + 0x74, 0x5f, 0x6e, 0x3a, 0x20, 0x32, 0x35, 0x36, 0x2c, 0x20, 0x2f, 0x2f, + 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x2c, 0x20, 0x2d, 0x31, + 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x20, 0x73, + 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, + 0x3a, 0x20, 0x31, 0x2e, 0x31, 0x38, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, + 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, + 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, + 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x3c, 0x3d, + 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, 0x75, 0x73, 0x65, 0x20, 0x76, 0x6f, + 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, + 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6d, 0x69, 0x6e, 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x20, 0x3d, 0x20, 0x64, + 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x3a, 0x20, 0x34, 0x30, 0x2c, - 0x20, 0x2f, 0x2f, 0x20, 0x3c, 0x3d, 0x20, 0x30, 0x20, 0x74, 0x6f, 0x20, - 0x75, 0x73, 0x65, 0x20, 0x76, 0x6f, 0x63, 0x61, 0x62, 0x20, 0x73, 0x69, - 0x7a, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, - 0x5f, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, - 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, - 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x66, 0x73, - 0x5f, 0x7a, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, - 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, - 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x79, 0x70, - 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, - 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, - 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, - 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, - 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, - 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, - 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3a, 0x20, 0x30, 0x2c, - 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2f, 0x31, 0x2f, 0x32, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, - 0x5f, 0x74, 0x61, 0x75, 0x3a, 0x20, 0x35, 0x2c, 0x20, 0x2f, 0x2f, 0x20, - 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, 0x65, 0x6e, 0x74, 0x72, 0x6f, - 0x70, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, - 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x3a, 0x20, 0x30, - 0x2e, 0x31, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x61, 0x72, 0x6e, - 0x69, 0x6e, 0x67, 0x20, 0x72, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, - 0x27, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x5f, - 0x70, 0x72, 0x6f, 0x62, 0x73, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, - 0x20, 0x6e, 0x6f, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, - 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x61, - 0x63, 0x68, 0x65, 0x5f, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3a, 0x20, - 0x74, 0x72, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x53, 0x54, 0x41, 0x52, - 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, - 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x6f, - 0x72, 0x77, 0x73, 0x65, 0x72, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x3d, 0x20, - 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x63, 0x70, 0x70, 0x5f, 0x73, 0x65, - 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, - 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x22, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, - 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, - 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, - 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, - 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x20, 0x20, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x3a, + 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x31, 0x2e, 0x30, + 0x20, 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, + 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, + 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, + 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2e, 0x30, 0x20, + 0x3d, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, + 0x74, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x30, 0x2f, 0x31, + 0x2f, 0x32, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x69, 0x72, + 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x3a, 0x20, 0x35, + 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x20, + 0x65, 0x6e, 0x74, 0x72, 0x6f, 0x70, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, + 0x74, 0x61, 0x3a, 0x20, 0x30, 0x2e, 0x31, 0x2c, 0x20, 0x2f, 0x2f, 0x20, + 0x6c, 0x65, 0x61, 0x72, 0x6e, 0x69, 0x6e, 0x67, 0x20, 0x72, 0x61, 0x74, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, + 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x3a, 0x20, + 0x30, 0x2c, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x63, 0x6f, 0x6d, + 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, + 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, + 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x61, 0x63, 0x68, 0x65, 0x5f, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x3a, 0x20, 0x74, 0x72, 0x75, 0x65, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, + 0x20, 0x53, 0x54, 0x41, 0x52, 0x54, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, + 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, + 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, + 0x69, 0x6e, 0x20, 0x62, 0x6f, 0x72, 0x77, 0x73, 0x65, 0x72, 0x20, 0x4c, + 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, + 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, - 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, - 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x73, 0x74, 0x72, - 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, - 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, - 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, - 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, - 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, - 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, - 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, - 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, - 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, + 0x65, 0x79, 0x20, 0x3d, 0x20, 0x22, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x63, + 0x70, 0x70, 0x5f, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x5f, 0x6c, 0x6f, + 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x22, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, + 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, + 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, + 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, + 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, - 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, - 0x6d, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, - 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x4a, 0x53, 0x4f, 0x4e, - 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x69, 0x74, 0x65, 0x6d, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, - 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, - 0x61, 0x74, 0x61, 0x41, 0x73, 0x52, 0x61, 0x77, 0x54, 0x65, 0x78, 0x74, + 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x4a, 0x53, 0x4f, + 0x4e, 0x2e, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x69, 0x66, 0x79, 0x28, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, + 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x52, 0x61, 0x77, 0x54, + 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x2c, 0x20, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x2e, 0x73, 0x65, 0x74, 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, + 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, 0x2b, 0x20, 0x74, 0x61, 0x67, 0x2c, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, + 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, + 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, + 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, @@ -483,90 +477,135 @@ unsigned char index_html[] = { 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, + 0x69, 0x74, 0x65, 0x6d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, - 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x20, - 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, - 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, - 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, - 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6c, 0x65, 0x74, 0x27, - 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, - 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x69, 0x66, - 0x20, 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, 0x61, 0x72, 0x65, 0x20, 0x61, - 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x75, 0x73, - 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, - 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, - 0x73, 0x20, 0x61, 0x72, 0x65, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, - 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e, 0x65, 0x20, 0x6f, 0x62, 0x6a, 0x65, - 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x69, 0x6e, - 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, 0x66, 0x20, 0x7b, 0x20, 0x22, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, - 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x20, 0x61, 0x6e, 0x64, 0x20, - 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, - 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x64, 0x61, - 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, - 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x52, 0x61, + 0x77, 0x54, 0x65, 0x78, 0x74, 0x28, 0x74, 0x61, 0x67, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x69, 0x74, 0x65, 0x6d, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, + 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x2e, 0x67, 0x65, 0x74, + 0x49, 0x74, 0x65, 0x6d, 0x28, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x4b, 0x65, 0x79, 0x20, 0x2b, 0x20, 0x27, 0x2f, 0x27, 0x20, + 0x2b, 0x20, 0x74, 0x61, 0x67, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x69, 0x74, 0x65, 0x6d, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x74, 0x65, 0x6d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x72, + 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x61, + 0x69, 0x6e, 0x65, 0x72, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, + 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, + 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, + 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, + 0x7b, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x27, 0x27, 0x2c, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3a, 0x20, 0x7b, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x7b, 0x7d, 0x2c, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x7b, 0x7d, 0x20, + 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, + 0x74, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x6c, 0x79, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, - 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, - 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, - 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, - 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, + 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, + 0x67, 0x73, 0x20, 0x69, 0x66, 0x20, 0x74, 0x68, 0x65, 0x72, 0x65, 0x20, + 0x61, 0x72, 0x65, 0x20, 0x61, 0x6e, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x65, + 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x61, 0x72, 0x65, 0x20, 0x73, + 0x74, 0x6f, 0x72, 0x65, 0x64, 0x20, 0x69, 0x6e, 0x20, 0x6f, 0x6e, 0x65, + 0x20, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x69, 0x6e, 0x20, 0x66, 0x6f, 0x72, 0x6d, 0x20, 0x6f, + 0x66, 0x20, 0x7b, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, + 0x20, 0x61, 0x6e, 0x64, 0x20, 0x7b, 0x20, 0x22, 0x73, 0x65, 0x74, 0x74, + 0x69, 0x6e, 0x67, 0x73, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x6e, 0x61, 0x6d, 0x65, 0x22, 0x3a, 0x22, 0x73, 0x65, 0x74, 0x74, 0x69, + 0x6e, 0x67, 0x73, 0x64, 0x61, 0x74, 0x61, 0x22, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, + 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x69, + 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, - 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x20, 0x77, 0x65, 0x72, 0x65, 0x20, 0x73, 0x75, 0x63, - 0x63, 0x65, 0x73, 0x73, 0x66, 0x75, 0x6c, 0x79, 0x20, 0x69, 0x6d, 0x70, - 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, - 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f, 0x63, 0x65, 0x73, 0x73, 0x69, 0x6e, - 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, - 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, - 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, - 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, - 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, - 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x69, 0x6d, 0x70, - 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, - 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x6f, 0x76, 0x65, 0x72, 0x72, 0x69, - 0x64, 0x65, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x2e, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x3d, - 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, + 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, + 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, + 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, + 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, + 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, + 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x69, + 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x77, 0x65, 0x72, + 0x65, 0x20, 0x73, 0x75, 0x63, 0x63, 0x65, 0x73, 0x73, 0x66, 0x75, 0x6c, + 0x79, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, + 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x50, 0x72, 0x6f, 0x63, + 0x65, 0x73, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, + 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, + 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, + 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, + 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, + 0x67, 0x28, 0x69, 0x6d, 0x70, 0x6f, 0x72, 0x74, 0x65, 0x64, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x69, 0x6d, 0x70, 0x6f, 0x72, + 0x74, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x6f, + 0x76, 0x65, 0x72, 0x72, 0x69, 0x64, 0x65, 0x20, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, + 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, + 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, + 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, + 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, + 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x64, 0x65, + 0x74, 0x65, 0x63, 0x74, 0x65, 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, + 0x6f, 0x67, 0x28, 0x27, 0x49, 0x6e, 0x69, 0x74, 0x69, 0x61, 0x6c, 0x69, + 0x7a, 0x69, 0x6e, 0x67, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, + 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x61, + 0x76, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, + 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, + 0x22, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x22, 0x3a, 0x20, 0x7b, + 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, @@ -574,502 +613,384 @@ unsigned char index_html[] = { 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, - 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x20, 0x64, 0x65, 0x74, 0x65, 0x63, 0x74, 0x65, - 0x64, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x49, - 0x6e, 0x69, 0x74, 0x69, 0x61, 0x6c, 0x69, 0x7a, 0x69, 0x6e, 0x67, 0x20, - 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, - 0x20, 0x61, 0x6e, 0x64, 0x20, 0x73, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, - 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x22, 0x64, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x22, 0x3a, 0x20, 0x7b, 0x20, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, - 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, - 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, - 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x27, 0x2c, 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, - 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x52, 0x65, 0x73, 0x65, 0x74, 0x69, 0x6e, - 0x67, 0x20, 0x74, 0x68, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, - 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, - 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, - 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x73, 0x61, 0x76, 0x65, - 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5b, 0x27, 0x64, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x74, - 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, - 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, - 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, + 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, + 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x68, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x64, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, + 0x20, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x5b, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x5d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, + 0x70, 0x6c, 0x79, 0x28, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x74, 0x2e, 0x64, 0x61, 0x74, - 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x64, 0x61, 0x74, 0x61, - 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, - 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, - 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, - 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, - 0x76, 0x65, 0x64, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67, 0x65, 0x74, 0x20, 0x61, 0x75, 0x74, - 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, - 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, - 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, - 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x67, 0x65, 0x74, - 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, - 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x6c, - 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, - 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, - 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x72, 0x65, 0x73, 0x74, 0x6f, - 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, - 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x61, 0x73, - 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x4e, 0x6f, 0x20, - 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, - 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, - 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, - 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x77, 0x61, - 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x73, 0x6f, 0x20, - 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x64, 0x65, - 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, - 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x69, 0x6e, 0x67, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x6e, 0x61, - 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x66, 0x72, 0x6f, 0x6d, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x0a, 0x0a, 0x20, + 0x61, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, + 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x27, 0x20, + 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x74, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x69, 0x6d, 0x61, 0x67, 0x65, + 0x5f, 0x64, 0x61, 0x74, 0x61, 0x3a, 0x20, 0x5b, 0x5d, 0x20, 0x7d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, + 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, + 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, + 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61, 0x76, 0x65, 0x64, 0x55, 0x73, 0x65, - 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, - 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, - 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, - 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x41, - 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x2e, 0x2e, 0x2e, 0x27, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, - 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, - 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x77, 0x65, - 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, 0x20, 0x77, 0x61, 0x6e, 0x74, 0x20, - 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x6f, 0x76, 0x65, 0x72, - 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x65, - 0x74, 0x27, 0x73, 0x20, 0x63, 0x72, 0x65, 0x61, 0x74, 0x65, 0x20, 0x61, - 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f, 0x6e, 0x65, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, - 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27, 0x20, 0x2b, 0x20, 0x44, 0x61, 0x74, - 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, 0x2e, 0x74, 0x6f, 0x53, 0x74, - 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x27, - 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x2c, 0x20, - 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, - 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, - 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x73, 0x20, 0x27, - 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x61, 0x76, 0x65, - 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x65, 0x20, 0x61, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, - 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, - 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, - 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, - 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x69, - 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, - 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, - 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, - 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, + 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, + 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x67, 0x65, + 0x74, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, + 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, + 0x64, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, - 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, - 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, - 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, - 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, - 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, - 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, - 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, 0x20, 0x7b, 0x20, - 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, 0x3a, 0x20, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x27, 0x3a, 0x20, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x65, 0x5f, 0x67, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x41, 0x73, 0x4f, + 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, + 0x73, 0x74, 0x27, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x20, 0x7b, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x75, + 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, + 0x6c, 0x61, 0x74, 0x65, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, + 0x72, 0x65, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x29, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x6c, + 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x55, 0x73, 0x65, 0x64, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, - 0x28, 0x27, 0x43, 0x68, 0x65, 0x63, 0x6b, 0x69, 0x6e, 0x67, 0x20, 0x66, - 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, - 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, - 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x28, - 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x20, 0x45, 0x4e, - 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, 0x6f, 0x72, 0x74, 0x20, 0x66, - 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, 0x69, 0x6e, 0x20, 0x62, 0x72, - 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, 0x4c, 0x6f, 0x63, 0x61, 0x6c, - 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x20, 0x2a, 0x2f, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x6c, 0x6c, - 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x20, 0x3d, 0x20, 0x73, - 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, - 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, - 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x75, 0x72, 0x72, - 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, - 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, - 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, - 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x63, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x68, 0x61, 0x73, 0x20, 0x74, - 0x68, 0x65, 0x20, 0x75, 0x73, 0x65, 0x72, 0x20, 0x73, 0x74, 0x61, 0x72, - 0x74, 0x65, 0x64, 0x20, 0x61, 0x20, 0x63, 0x68, 0x61, 0x74, 0x3f, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, - 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, - 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, - 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, - 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, - 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x74, - 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, - 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x73, 0x74, 0x72, - 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, - 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x73, 0x65, 0x74, 0x74, - 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x65, 0x78, 0x74, 0x72, - 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x74, - 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, - 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x2c, 0x20, 0x2e, - 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, - 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x73, - 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x41, - 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, 0x7b, 0x28, 0x2e, 0x2a, 0x3f, - 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, 0x2c, 0x20, 0x28, 0x5f, 0x2c, - 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, - 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, 0x29, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x61, 0x73, - 0x79, 0x6e, 0x63, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, - 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x63, 0x68, 0x61, 0x72, 0x29, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x3d, 0x20, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, - 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, 0x77, 0x20, 0x6e, 0x65, 0x77, - 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, 0x22, 0x61, 0x6c, 0x72, 0x65, - 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x22, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, 0x72, 0x74, 0x43, 0x6f, 0x6e, - 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x77, 0x61, - 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, - 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, 0x7b, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x3a, 0x20, - 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, 0x20, 0x63, 0x68, 0x75, 0x6e, - 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, - 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x77, 0x68, 0x69, 0x6c, - 0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, - 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, - 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, - 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, - 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x24, - 0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x6f, 0x70, 0x28, 0x29, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, - 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x65, 0x64, - 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6d, 0x61, - 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, - 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, - 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2c, 0x20, 0x22, 0x27, 0x2c, 0x20, - 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3a, 0x20, 0x22, 0x2c, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x20, 0x3d, 0x20, 0x64, 0x61, - 0x74, 0x61, 0x2e, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, - 0x6d, 0x61, 0x67, 0x65, 0x20, 0x26, 0x26, 0x20, 0x21, 0x64, 0x61, 0x74, - 0x61, 0x2e, 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, + 0x28, 0x27, 0x4e, 0x6f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, + 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, + 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2c, 0x20, 0x75, 0x73, 0x69, 0x6e, 0x67, + 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x6e, 0x6f, 0x20, 0x61, 0x75, + 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, + 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x20, 0x77, 0x61, 0x73, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, + 0x2c, 0x20, 0x73, 0x6f, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x20, 0x66, 0x72, + 0x6f, 0x6d, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x2e, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, + 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, + 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x69, + 0x6e, 0x67, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, + 0x6e, 0x64, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x20, 0x69, 0x6e, + 0x74, 0x65, 0x72, 0x6e, 0x61, 0x6c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, + 0x66, 0x72, 0x6f, 0x6d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x73, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x70, + 0x70, 0x6c, 0x79, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, + 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x61, 0x76, + 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, + 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x75, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, + 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x20, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, + 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, + 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, + 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x22, 0x54, - 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x20, 0x77, 0x61, - 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x69, 0x6c, - 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x6d, 0x75, 0x6c, 0x74, 0x69, - 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x20, 0x6f, 0x72, 0x20, 0x74, 0x68, 0x65, - 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, 0x70, 0x72, 0x6f, 0x6a, 0x65, - 0x63, 0x74, 0x6f, 0x72, 0x20, 0x63, 0x61, 0x6e, 0x27, 0x74, 0x20, 0x62, - 0x65, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x65, 0x64, 0x2e, 0x22, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x77, 0x65, 0x20, 0x64, 0x6f, 0x6e, 0x27, 0x74, 0x20, + 0x77, 0x61, 0x6e, 0x74, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x61, 0x76, 0x65, + 0x20, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, + 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, + 0x73, 0x6f, 0x20, 0x6c, 0x65, 0x74, 0x27, 0x73, 0x20, 0x63, 0x72, 0x65, + 0x61, 0x74, 0x65, 0x20, 0x61, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x6f, 0x6e, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x20, 0x27, 0x55, 0x73, 0x65, + 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2d, 0x27, 0x20, + 0x2b, 0x20, 0x44, 0x61, 0x74, 0x65, 0x2e, 0x6e, 0x6f, 0x77, 0x28, 0x29, + 0x2e, 0x74, 0x6f, 0x53, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x28, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, + 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, + 0x3d, 0x20, 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, + 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, + 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, 0x27, 0x3a, + 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x27, + 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, + 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x53, 0x61, 0x76, 0x69, 0x6e, 0x67, + 0x20, 0x61, 0x73, 0x20, 0x27, 0x20, 0x2b, 0x20, 0x6e, 0x65, 0x77, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4e, 0x61, 0x6d, 0x65, 0x29, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x73, 0x61, 0x76, 0x65, 0x20, 0x69, 0x6e, 0x20, 0x74, 0x68, 0x65, + 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x73, 0x6c, + 0x6f, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, + 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, + 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, + 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, + 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, 0x6e, 0x65, 0x77, 0x54, 0x65, 0x6d, + 0x70, 0x6c, 0x61, 0x74, 0x65, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x6c, + 0x6f, 0x61, 0x64, 0x20, 0x69, 0x74, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x20, + 0x61, 0x6e, 0x64, 0x20, 0x61, 0x70, 0x70, 0x6c, 0x79, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, 0x6e, + 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, + 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6f, 0x63, 0x61, 0x6c, 0x5f, 0x73, + 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x74, 0x44, 0x61, + 0x74, 0x61, 0x46, 0x72, 0x6f, 0x6d, 0x4f, 0x62, 0x6a, 0x65, 0x63, 0x74, + 0x28, 0x27, 0x75, 0x73, 0x65, 0x72, 0x5f, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x73, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x27, 0x2c, 0x20, + 0x7b, 0x20, 0x27, 0x6e, 0x61, 0x6d, 0x65, 0x27, 0x3a, 0x20, 0x73, 0x65, + 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x27, 0x64, 0x61, 0x74, 0x61, + 0x27, 0x3a, 0x20, 0x7b, 0x20, 0x27, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x27, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x27, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x27, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x20, 0x7d, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, + 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x43, 0x68, 0x65, 0x63, 0x6b, + 0x69, 0x6e, 0x67, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x61, 0x75, 0x74, 0x6f, + 0x73, 0x61, 0x76, 0x65, 0x64, 0x20, 0x6c, 0x61, 0x73, 0x74, 0x20, 0x75, + 0x73, 0x65, 0x64, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x4c, 0x6f, 0x61, 0x64, 0x41, + 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x41, 0x75, 0x74, 0x6f, 0x73, + 0x61, 0x76, 0x65, 0x64, 0x28, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2a, 0x20, 0x45, 0x4e, 0x44, 0x3a, 0x20, 0x53, 0x75, 0x70, 0x70, + 0x6f, 0x72, 0x74, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x73, 0x74, 0x6f, 0x72, + 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x73, 0x20, 0x61, 0x6e, 0x64, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x65, 0x74, 0x65, 0x72, 0x73, 0x20, + 0x69, 0x6e, 0x20, 0x62, 0x72, 0x6f, 0x77, 0x73, 0x65, 0x72, 0x73, 0x20, + 0x4c, 0x6f, 0x63, 0x61, 0x6c, 0x53, 0x74, 0x6f, 0x72, 0x61, 0x67, 0x65, + 0x20, 0x2a, 0x2f, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, + 0x73, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, + 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, + 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x6e, + 0x75, 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x6c, 0x79, 0x20, 0x67, + 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x61, 0x20, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3f, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x65, + 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6e, 0x67, 0x20, 0x3d, 0x20, 0x63, + 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x21, 0x3d, 0x20, 0x6e, 0x75, + 0x6c, 0x6c, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, + 0x68, 0x61, 0x73, 0x20, 0x74, 0x68, 0x65, 0x20, 0x75, 0x73, 0x65, 0x72, + 0x20, 0x73, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x20, 0x61, 0x20, 0x63, + 0x68, 0x61, 0x74, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, 0x61, 0x72, 0x74, + 0x65, 0x64, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x75, 0x74, 0x65, + 0x64, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x29, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, + 0x20, 0x3d, 0x20, 0x28, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, + 0x70, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, - 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, - 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, - 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, - 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, - 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, + 0x63, 0x72, 0x69, 0x70, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x74, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x72, 0x65, 0x70, 0x6c, + 0x61, 0x63, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x3d, + 0x20, 0x28, 0x73, 0x74, 0x72, 0x2c, 0x20, 0x65, 0x78, 0x74, 0x72, 0x61, + 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, + 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x65, 0x78, 0x74, 0x72, 0x61, 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, - 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x64, - 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x75, 0x6c, - 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x65, 0x6e, 0x64, 0x20, 0x6d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x74, 0x6f, 0x20, 0x73, 0x65, 0x72, - 0x76, 0x65, 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, - 0x6e, 0x63, 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, - 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x6f, 0x6c, 0x65, 0x2e, - 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, - 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, - 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, - 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, - 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x7b, - 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x22, 0x2c, 0x20, 0x6d, 0x73, - 0x67, 0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x6c, 0x65, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, - 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, - 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, - 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x28, - 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, - 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, - 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, - 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, - 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, - 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, - 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, - 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, - 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2f, 0x2c, 0x20, - 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, - 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, 0x29, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, - 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x72, 0x6f, 0x6d, - 0x70, 0x74, 0x20, 0x3d, 0x20, 0x60, 0x41, 0x20, 0x63, 0x68, 0x61, 0x74, - 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, 0x6e, 0x20, 0x61, 0x20, 0x63, - 0x75, 0x72, 0x69, 0x6f, 0x75, 0x73, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, - 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x6e, 0x20, 0x61, 0x72, 0x74, 0x69, - 0x66, 0x69, 0x63, 0x69, 0x61, 0x6c, 0x20, 0x69, 0x6e, 0x74, 0x65, 0x6c, - 0x6c, 0x69, 0x67, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, - 0x73, 0x74, 0x61, 0x6e, 0x74, 0x2e, 0x20, 0x54, 0x68, 0x65, 0x20, 0x61, - 0x73, 0x73, 0x69, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x20, 0x67, 0x69, 0x76, - 0x65, 0x73, 0x20, 0x68, 0x65, 0x6c, 0x70, 0x66, 0x75, 0x6c, 0x2c, 0x20, - 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x2c, 0x20, 0x61, 0x6e, - 0x64, 0x20, 0x70, 0x6f, 0x6c, 0x69, 0x74, 0x65, 0x20, 0x61, 0x6e, 0x73, - 0x77, 0x65, 0x72, 0x73, 0x20, 0x74, 0x6f, 0x20, 0x74, 0x68, 0x65, 0x20, - 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x27, 0x73, 0x20, 0x71, 0x75, 0x65, 0x73, - 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x2e, 0x5c, 0x6e, 0x55, 0x53, 0x45, 0x52, - 0x3a, 0x5b, 0x69, 0x6d, 0x67, 0x2d, 0x31, 0x30, 0x5d, 0x24, 0x7b, 0x6d, - 0x73, 0x67, 0x7d, 0x5c, 0x6e, 0x41, 0x53, 0x53, 0x49, 0x53, 0x54, 0x41, - 0x4e, 0x54, 0x3a, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, - 0x74, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, - 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, - 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x2c, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x63, - 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x2c, 0x20, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x22, 0x7b, 0x7b, 0x75, 0x73, - 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, 0x5d, 0x2c, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x7b, 0x7b, 0x63, 0x68, - 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x72, 0x75, 0x6e, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, + 0x20, 0x20, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x3d, + 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x74, 0x74, 0x69, 0x6e, + 0x67, 0x73, 0x2c, 0x20, 0x2e, 0x2e, 0x2e, 0x65, 0x78, 0x74, 0x72, 0x61, + 0x53, 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x20, 0x7d, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x53, 0x74, 0x72, + 0x69, 0x6e, 0x67, 0x28, 0x73, 0x74, 0x72, 0x29, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x41, 0x6c, 0x6c, 0x28, 0x2f, 0x5c, 0x7b, 0x5c, + 0x7b, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, 0x7d, 0x5c, 0x7d, 0x2f, 0x67, + 0x2c, 0x20, 0x28, 0x5f, 0x2c, 0x20, 0x6b, 0x65, 0x79, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x73, + 0x65, 0x74, 0x74, 0x69, 0x6e, 0x67, 0x73, 0x5b, 0x6b, 0x65, 0x79, 0x5d, + 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x66, 0x75, 0x6e, + 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, + 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x6c, + 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2c, 0x20, + 0x63, 0x68, 0x61, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x20, + 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, + 0x79, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, + 0x72, 0x69, 0x70, 0x74, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, + 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x72, 0x6f, + 0x77, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x45, 0x72, 0x72, 0x6f, 0x72, 0x28, + 0x22, 0x61, 0x6c, 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, + 0x6e, 0x69, 0x6e, 0x67, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x41, 0x62, 0x6f, + 0x72, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, + 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, + 0x72, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x28, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x20, 0x6f, 0x66, 0x20, + 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, + 0x2c, 0x20, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x50, 0x61, 0x72, 0x61, 0x6d, + 0x73, 0x2c, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x64, 0x61, 0x74, 0x61, 0x20, 0x3d, + 0x20, 0x63, 0x68, 0x75, 0x6e, 0x6b, 0x2e, 0x64, 0x61, 0x74, 0x61, 0x3b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, + 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x74, 0x6f, 0x70, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x77, 0x68, 0x69, 0x6c, 0x65, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, 0x30, 0x20, + 0x26, 0x26, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5b, 0x63, 0x75, 0x72, 0x72, 0x65, + 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x6c, + 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x2d, 0x20, 0x31, 0x5d, 0x2e, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, + 0x28, 0x2f, 0x5c, 0x6e, 0x24, 0x2f, 0x29, 0x20, 0x21, 0x3d, 0x20, 0x6e, + 0x75, 0x6c, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, + 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x6f, + 0x70, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, + 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, + 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, + 0x61, 0x72, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x22, 0x43, 0x6f, + 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x66, 0x69, 0x6e, + 0x69, 0x73, 0x68, 0x65, 0x64, 0x3a, 0x20, 0x27, 0x22, 0x2c, 0x20, 0x63, + 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, + 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2c, + 0x20, 0x22, 0x27, 0x2c, 0x20, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, + 0x3a, 0x20, 0x22, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x70, 0x75, 0x73, 0x68, 0x28, + 0x64, 0x61, 0x74, 0x61, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, + 0x20, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x73, 0x6c, 0x6f, 0x74, + 0x5f, 0x69, 0x64, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, + 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x20, 0x26, 0x26, + 0x20, 0x21, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x75, 0x6c, 0x74, 0x69, + 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, + 0x72, 0x74, 0x28, 0x22, 0x54, 0x68, 0x65, 0x20, 0x73, 0x65, 0x72, 0x76, + 0x65, 0x72, 0x20, 0x77, 0x61, 0x73, 0x20, 0x6e, 0x6f, 0x74, 0x20, 0x63, + 0x6f, 0x6d, 0x70, 0x69, 0x6c, 0x65, 0x64, 0x20, 0x66, 0x6f, 0x72, 0x20, + 0x6d, 0x75, 0x6c, 0x74, 0x69, 0x6d, 0x6f, 0x64, 0x61, 0x6c, 0x20, 0x6f, + 0x72, 0x20, 0x74, 0x68, 0x65, 0x20, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x20, + 0x70, 0x72, 0x6f, 0x6a, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x63, 0x61, + 0x6e, 0x27, 0x74, 0x20, 0x62, 0x65, 0x20, 0x6c, 0x6f, 0x61, 0x64, 0x65, + 0x64, 0x2e, 0x22, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, 0x68, 0x69, 0x73, + 0x74, 0x6f, 0x72, 0x79, 0x2c, 0x20, 0x5b, 0x63, 0x68, 0x61, 0x72, 0x2c, + 0x20, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x4d, 0x65, 0x73, 0x73, + 0x61, 0x67, 0x65, 0x73, 0x5d, 0x5d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x64, 0x61, 0x74, 0x61, 0x2e, + 0x74, 0x69, 0x6d, 0x69, 0x6e, 0x67, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x74, 0x69, 0x6d, + 0x69, 0x6e, 0x67, 0x73, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x72, + 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x3d, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, 0x65, + 0x6e, 0x64, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x74, + 0x6f, 0x20, 0x73, 0x65, 0x72, 0x76, 0x65, 0x72, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, + 0x3d, 0x20, 0x61, 0x73, 0x79, 0x6e, 0x63, 0x20, 0x28, 0x6d, 0x73, 0x67, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, @@ -1078,26 +999,137 @@ unsigned char index_html[] = { 0x72, 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, - 0x72, 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, - 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, - 0x69, 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x22, 0x2c, 0x20, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x5d, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, - 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, - 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x5d, 0x2c, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x22, 0x22, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, + 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, 0x2e, 0x2e, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, + 0x2c, 0x20, 0x5b, 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, + 0x22, 0x2c, 0x20, 0x6d, 0x73, 0x67, 0x5d, 0x5d, 0x29, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x28, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x6d, 0x73, + 0x67, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, + 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x3a, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, + 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x66, 0x6c, 0x61, + 0x74, 0x4d, 0x61, 0x70, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x28, 0x5b, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, 0x3d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, + 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, + 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x3a, 0x20, 0x41, 0x72, 0x72, 0x61, + 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, + 0x74, 0x61, 0x29, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, + 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, + 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, + 0x5c, 0x73, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, 0x2c, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x22, 0x5c, 0x6e, 0x22, + 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x69, 0x6d, 0x61, 0x67, + 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x60, 0x41, + 0x20, 0x63, 0x68, 0x61, 0x74, 0x20, 0x62, 0x65, 0x74, 0x77, 0x65, 0x65, + 0x6e, 0x20, 0x61, 0x20, 0x63, 0x75, 0x72, 0x69, 0x6f, 0x75, 0x73, 0x20, + 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x61, 0x6e, + 0x20, 0x61, 0x72, 0x74, 0x69, 0x66, 0x69, 0x63, 0x69, 0x61, 0x6c, 0x20, + 0x69, 0x6e, 0x74, 0x65, 0x6c, 0x6c, 0x69, 0x67, 0x65, 0x6e, 0x63, 0x65, + 0x20, 0x61, 0x73, 0x73, 0x69, 0x73, 0x74, 0x61, 0x6e, 0x74, 0x2e, 0x20, + 0x54, 0x68, 0x65, 0x20, 0x61, 0x73, 0x73, 0x69, 0x73, 0x74, 0x61, 0x6e, + 0x74, 0x20, 0x67, 0x69, 0x76, 0x65, 0x73, 0x20, 0x68, 0x65, 0x6c, 0x70, + 0x66, 0x75, 0x6c, 0x2c, 0x20, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x65, + 0x64, 0x2c, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x70, 0x6f, 0x6c, 0x69, 0x74, + 0x65, 0x20, 0x61, 0x6e, 0x73, 0x77, 0x65, 0x72, 0x73, 0x20, 0x74, 0x6f, + 0x20, 0x74, 0x68, 0x65, 0x20, 0x68, 0x75, 0x6d, 0x61, 0x6e, 0x27, 0x73, + 0x20, 0x71, 0x75, 0x65, 0x73, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x2e, 0x5c, + 0x6e, 0x55, 0x53, 0x45, 0x52, 0x3a, 0x5b, 0x69, 0x6d, 0x67, 0x2d, 0x31, + 0x30, 0x5d, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x7d, 0x5c, 0x6e, 0x41, 0x53, + 0x53, 0x49, 0x53, 0x54, 0x41, 0x4e, 0x54, 0x3a, 0x60, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x61, 0x77, 0x61, 0x69, 0x74, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, + 0x61, 0x6d, 0x61, 0x28, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, + 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, + 0x5f, 0x69, 0x64, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x73, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x5b, 0x22, 0x3c, 0x2f, 0x73, + 0x3e, 0x22, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x28, 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x3a, 0x22, + 0x29, 0x2c, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, + 0x22, 0x7b, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x7d, 0x3a, 0x22, 0x29, + 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, + 0x22, 0x7b, 0x7b, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x7d, 0x22, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x75, 0x6e, 0x43, 0x6f, 0x6d, + 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x6c, + 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x6f, 0x6c, 0x65, 0x2e, 0x6c, 0x6f, 0x67, 0x28, 0x27, 0x61, 0x6c, 0x72, + 0x65, 0x61, 0x64, 0x79, 0x20, 0x72, 0x75, 0x6e, 0x6e, 0x69, 0x6e, 0x67, + 0x2e, 0x2e, 0x2e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, + 0x69, 0x70, 0x74, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x28, 0x5b, 0x2e, + 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, + 0x70, 0x74, 0x2c, 0x20, 0x5b, 0x22, 0x22, 0x2c, 0x20, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x5d, 0x5d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x75, 0x6e, 0x4c, 0x6c, 0x61, 0x6d, 0x61, 0x28, 0x70, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x2c, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, + 0x64, 0x3a, 0x20, 0x73, 0x6c, 0x6f, 0x74, 0x5f, 0x69, 0x64, 0x2c, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x6f, 0x70, + 0x3a, 0x20, 0x5b, 0x5d, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x2c, 0x20, 0x22, 0x22, 0x29, 0x2e, 0x66, 0x69, 0x6e, 0x61, 0x6c, + 0x6c, 0x79, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, + 0x6d, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, + 0x5b, 0x5f, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, 0x29, 0x20, 0x3d, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, + 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x20, 0x3f, 0x20, 0x64, 0x61, + 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, + 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x20, + 0x3a, 0x20, 0x64, 0x61, 0x74, 0x61, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, + 0x3d, 0x20, 0x5b, 0x5d, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x74, 0x6f, 0x70, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, @@ -1342,1338 +1374,1389 @@ unsigned char index_html[] = { 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, - 0x65, 0x20, 0x3d, 0x20, 0x28, 0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, - 0x64, 0x61, 0x74, 0x61, 0x5d, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x6c, 0x65, 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, - 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, - 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x41, 0x72, - 0x72, 0x61, 0x79, 0x2e, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, - 0x64, 0x61, 0x74, 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, - 0x62, 0x73, 0x20, 0x3e, 0x20, 0x30, 0x20, 0x26, 0x26, 0x20, 0x69, 0x73, - 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, - 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, - 0x61, 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x64, 0x61, 0x74, 0x61, 0x7d, 0x20, - 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3d, 0x20, 0x69, 0x73, 0x41, 0x72, - 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3f, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x64, 0x61, 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, - 0x67, 0x20, 0x3d, 0x3e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x6e, 0x74, 0x29, 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, - 0x27, 0x29, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x5e, 0x5c, 0x73, 0x2b, 0x2f, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x64, 0x61, 0x74, 0x61, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, - 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x4d, - 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x7d, 0x20, - 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, 0x20, 0x2f, - 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x20, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, + 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x27, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x20, 0x3d, 0x20, 0x28, + 0x5b, 0x75, 0x73, 0x65, 0x72, 0x2c, 0x20, 0x64, 0x61, 0x74, 0x61, 0x5d, + 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, + 0x74, 0x20, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x20, 0x3d, 0x20, 0x41, 0x72, 0x72, 0x61, 0x79, 0x2e, 0x69, + 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x28, 0x64, 0x61, 0x74, 0x61, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, - 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, - 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c, - 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d, - 0x3a, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24, + 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x20, 0x3e, 0x20, + 0x30, 0x20, 0x26, 0x26, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, + 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6d, 0x65, 0x73, + 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, + 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, 0x3d, 0x24, + 0x7b, 0x64, 0x61, 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, 0x73, + 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x65, 0x78, 0x74, + 0x20, 0x3d, 0x20, 0x69, 0x73, 0x41, 0x72, 0x72, 0x61, 0x79, 0x4d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, + 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, + 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, + 0x2e, 0x6a, 0x6f, 0x69, 0x6e, 0x28, 0x27, 0x27, 0x29, 0x2e, 0x72, 0x65, + 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x5c, 0x73, 0x2b, 0x2f, + 0x2c, 0x20, 0x27, 0x27, 0x29, 0x20, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x61, 0x74, 0x61, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x20, 0x3d, 0x20, 0x69, 0x73, + 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, + 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x65, 0x78, 0x74, 0x20, 0x3a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x4d, 0x61, 0x72, 0x6b, + 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x7d, 0x20, 0x74, 0x65, 0x78, + 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x28, 0x74, 0x65, 0x78, 0x74, 0x29, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x75, 0x73, + 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, + 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x3c, 0x73, 0x74, 0x72, + 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x7b, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x28, 0x75, 0x73, 0x65, 0x72, 0x29, 0x7d, 0x3a, 0x3c, 0x2f, + 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x20, 0x24, 0x7b, 0x6d, 0x65, + 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x65, 0x6c, + 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x69, 0x73, + 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, + 0x64, 0x65, 0x20, 0x3f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, + 0x70, 0x61, 0x6e, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, + 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, + 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x60, 0x20, + 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, + 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x20, 0x65, 0x6c, 0x73, 0x65, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x70, 0x20, 0x6b, 0x65, 0x79, - 0x3d, 0x24, 0x7b, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x3e, 0x24, 0x7b, - 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x3c, 0x2f, 0x70, 0x3e, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, + 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, + 0x69, 0x6f, 0x6e, 0x45, 0x64, 0x69, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, + 0x3d, 0x20, 0x65, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x69, + 0x6e, 0x6e, 0x65, 0x72, 0x54, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, + 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x20, 0x3d, 0x20, 0x5b, 0x5d, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, - 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, - 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x68, + 0x61, 0x74, 0x22, 0x20, 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, 0x6f, + 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x7d, 0x20, 0x6b, 0x65, 0x79, + 0x3d, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, + 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6d, 0x67, 0x20, + 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x77, 0x69, 0x64, 0x74, 0x68, + 0x3a, 0x20, 0x36, 0x30, 0x25, 0x3b, 0x24, 0x7b, 0x21, 0x73, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x69, + 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, + 0x64, 0x20, 0x3f, 0x20, 0x60, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, + 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x60, 0x20, 0x3a, 0x20, 0x60, + 0x60, 0x7d, 0x22, 0x20, 0x73, 0x72, 0x63, 0x3d, 0x22, 0x24, 0x7b, 0x73, + 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, 0x6c, 0x65, 0x63, + 0x74, 0x65, 0x64, 0x7d, 0x22, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, + 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x65, 0x64, 0x69, 0x74, 0x61, + 0x62, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x69, 0x73, 0x43, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x4d, 0x6f, 0x64, 0x65, 0x7d, 0x20, + 0x72, 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, + 0x6e, 0x65, 0x72, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6f, 0x6d, + 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x45, 0x64, 0x69, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x69, 0x6d, 0x67, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, - 0x77, 0x69, 0x64, 0x74, 0x68, 0x3a, 0x20, 0x36, 0x30, 0x25, 0x3b, 0x24, - 0x7b, 0x21, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, 0x73, 0x65, - 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x20, 0x3f, 0x20, 0x60, 0x64, 0x69, - 0x73, 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, - 0x60, 0x20, 0x3a, 0x20, 0x60, 0x60, 0x7d, 0x22, 0x20, 0x73, 0x72, 0x63, - 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x5f, - 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x7d, 0x22, 0x2f, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, - 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, 0x2e, 0x66, 0x6c, - 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, 0x74, 0x4c, 0x69, - 0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x3e, 0x60, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6e, 0x66, - 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, - 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x20, - 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x73, 0x65, - 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x73, 0x65, 0x73, 0x73, 0x69, - 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, - 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, - 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, - 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, - 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, - 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, - 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, - 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, - 0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, - 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, - 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, - 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, - 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, - 0x5d, 0x3a, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, - 0x72, 0x28, 0x70, 0x61, 0x72, 0x73, 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, - 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x29, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, - 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x27, 0x27, - 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, - 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x67, 0x72, 0x61, - 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, - 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x65, 0x6c, 0x2e, 0x74, - 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, + 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x73, + 0x2e, 0x66, 0x6c, 0x61, 0x74, 0x4d, 0x61, 0x70, 0x28, 0x63, 0x68, 0x61, + 0x74, 0x4c, 0x69, 0x6e, 0x65, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, + 0x64, 0x69, 0x76, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, + 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, + 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, + 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, + 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x6e, + 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x72, 0x79, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x20, - 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, 0x70, 0x61, 0x72, 0x73, 0x65, - 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, - 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x28, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, - 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, - 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, + 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x73, + 0x65, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, + 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2c, 0x20, 0x5b, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, + 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x5d, 0x3a, 0x20, 0x4d, 0x61, 0x74, 0x68, + 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x61, 0x72, 0x73, 0x65, + 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x28, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, + 0x67, 0x65, 0x74, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x29, 0x20, + 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, + 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, + 0x4f, 0x72, 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x73, 0x69, 0x67, 0x6e, + 0x61, 0x6c, 0x28, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, + 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, + 0x64, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x6c, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, + 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, + 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, + 0x20, 0x65, 0x6c, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, + 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x72, 0x79, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x73, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x20, 0x3d, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x2e, + 0x70, 0x61, 0x72, 0x73, 0x65, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, + 0x61, 0x72, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x63, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x6e, 0x65, 0x77, 0x20, + 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, + 0x74, 0x65, 0x72, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, + 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, + 0x2c, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, + 0x65, 0x28, 0x28, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, + 0x20, 0x69, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x20, 0x2e, 0x2e, + 0x2e, 0x61, 0x63, 0x63, 0x2c, 0x20, 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, + 0x72, 0x69, 0x6d, 0x28, 0x29, 0x5d, 0x3a, 0x20, 0x69, 0x20, 0x7d, 0x29, + 0x2c, 0x20, 0x7b, 0x7d, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, + 0x72, 0x2e, 0x76, 0x69, 0x73, 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, + 0x6d, 0x61, 0x2c, 0x20, 0x27, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, - 0x73, 0x70, 0x6c, 0x69, 0x74, 0x28, 0x27, 0x2c, 0x27, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x2e, 0x72, 0x65, 0x64, 0x75, 0x63, 0x65, 0x28, 0x28, 0x61, 0x63, - 0x63, 0x2c, 0x20, 0x63, 0x75, 0x72, 0x2c, 0x20, 0x69, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x28, 0x7b, 0x20, 0x2e, 0x2e, 0x2e, 0x61, 0x63, 0x63, 0x2c, - 0x20, 0x5b, 0x63, 0x75, 0x72, 0x2e, 0x74, 0x72, 0x69, 0x6d, 0x28, 0x29, - 0x5d, 0x3a, 0x20, 0x69, 0x20, 0x7d, 0x29, 0x2c, 0x20, 0x7b, 0x7d, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x76, 0x69, 0x73, - 0x69, 0x74, 0x28, 0x73, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2c, 0x20, 0x27, - 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x2e, 0x2e, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, - 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, 0x20, 0x63, 0x6f, 0x6e, 0x76, - 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x28, 0x29, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x20, 0x63, 0x61, 0x74, - 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x61, 0x6c, 0x65, 0x72, 0x74, - 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x66, 0x61, - 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, 0x24, 0x7b, 0x65, 0x2e, 0x6d, 0x65, - 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, - 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x2c, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, - 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, - 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, - 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, - 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, - 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, - 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, - 0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3d, 0x22, 0x24, - 0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, - 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x49, - 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, + 0x2e, 0x2e, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3a, + 0x20, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x65, 0x72, 0x2e, 0x66, + 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x28, 0x29, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x20, 0x63, 0x61, 0x74, 0x63, 0x68, 0x20, 0x28, 0x65, 0x29, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x61, 0x6c, 0x65, 0x72, 0x74, 0x28, 0x60, 0x43, 0x6f, 0x6e, 0x76, 0x65, + 0x72, 0x74, 0x20, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x3a, 0x20, 0x24, + 0x7b, 0x65, 0x2e, 0x6d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x7d, 0x60, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x46, 0x6c, 0x6f, + 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x2c, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, - 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, - 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, - 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, - 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, - 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, - 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, - 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, - 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, - 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, - 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, - 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, - 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, - 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x3d, - 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x76, - 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x28, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, - 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, - 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, 0x79, 0x28, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x55, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, - 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x28, 0x29, - 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, - 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, - 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x61, 0x6d, - 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, - 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x73, 0x74, 0x65, 0x70, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, + 0x7d, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, 0x65, 0x64, 0x3e, 0x55, 0x73, - 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x20, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x62, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x6f, 0x6e, 0x63, 0x6c, - 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x7d, - 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, 0x61, 0x6c, 0x6c, 0x20, 0x74, - 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x3c, 0x2f, 0x62, - 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, - 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, - 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x20, 0x6f, 0x6e, 0x20, 0x65, - 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, 0x68, 0x61, 0x6e, 0x67, 0x65, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, - 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x41, 0x75, 0x74, 0x6f, - 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x5d, 0x29, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, - 0x6f, 0x6c, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, - 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, - 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x67, - 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, 0x20, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, - 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, 0x66, 0x20, 0x6f, 0x72, 0x20, - 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x2b, - 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x67, 0x72, 0x61, 0x6d, 0x6d, - 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, - 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, - 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x7d, 0x2f, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, 0x2d, 0x6f, 0x72, 0x64, 0x65, - 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, 0x64, - 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x3a, 0x20, 0x70, - 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x32, 0x2c, 0x70, - 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x47, 0x72, - 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, 0x6f, 0x6e, 0x53, 0x63, 0x68, - 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, 0x4f, 0x72, 0x64, 0x65, 0x72, + 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, + 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x72, 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, + 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, + 0x22, 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, + 0x3d, 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x73, 0x74, + 0x65, 0x70, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x74, 0x65, 0x70, 0x7d, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, + 0x65, 0x7d, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, + 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, + 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, - 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x62, 0x75, 0x74, 0x74, 0x6f, - 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, - 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x4a, 0x53, 0x4f, 0x4e, - 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, - 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, 0x20, 0x4a, - 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x3c, 0x2f, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, - 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, - 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, + 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, + 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, + 0x20, 0x3d, 0x20, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x2c, + 0x20, 0x6d, 0x61, 0x78, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x2c, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x7d, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, + 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x46, 0x6f, 0x72, 0x3d, 0x22, - 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, - 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, + 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, + 0x6d, 0x65, 0x7d, 0x22, 0x3e, 0x24, 0x7b, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x7d, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, + 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, + 0x61, 0x6e, 0x67, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x24, 0x7b, + 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, 0x6d, 0x69, 0x6e, 0x3d, 0x22, + 0x24, 0x7b, 0x6d, 0x69, 0x6e, 0x7d, 0x22, 0x20, 0x6d, 0x61, 0x78, 0x3d, + 0x22, 0x24, 0x7b, 0x6d, 0x61, 0x78, 0x7d, 0x22, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x6e, 0x61, 0x6d, 0x65, 0x7d, 0x22, 0x20, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, - 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, - 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, - 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, - 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, - 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, - 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, - 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x3e, 0x55, - 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, - 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, - 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x75, 0x73, 0x65, 0x72, - 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, - 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, - 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, 0x22, 0x3e, 0x42, 0x6f, 0x74, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, - 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x72, 0x22, 0x20, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x63, 0x68, - 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, - 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, - 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, - 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, - 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, - 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, - 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x43, 0x68, - 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x20, 0x74, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, - 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, - 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, - 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, - 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x54, - 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x7d, 0x22, 0x20, 0x72, 0x6f, - 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, - 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, - 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, - 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, - 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, - 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, - 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x20, - 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, 0x28, 0x29, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x24, 0x7b, 0x47, - 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, - 0x6c, 0x28, 0x29, 0x7d, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, + 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x7d, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x75, 0x73, + 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, + 0x73, 0x65, 0x74, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x65, + 0x2e, 0x70, 0x72, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x44, 0x65, 0x66, 0x61, + 0x75, 0x6c, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x54, 0x6f, 0x44, 0x65, + 0x66, 0x61, 0x75, 0x6c, 0x74, 0x41, 0x6e, 0x64, 0x41, 0x70, 0x70, 0x6c, + 0x79, 0x28, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, + 0x20, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x73, + 0x65, 0x6c, 0x65, 0x63, 0x74, 0x65, 0x64, 0x55, 0x73, 0x65, 0x72, 0x54, + 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x6e, 0x61, 0x6d, 0x65, 0x20, 0x3d, 0x3d, 0x20, 0x27, 0x64, + 0x65, 0x66, 0x61, 0x75, 0x6c, 0x74, 0x27, 0x29, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x64, 0x69, 0x73, 0x61, 0x62, 0x6c, + 0x65, 0x64, 0x3e, 0x55, 0x73, 0x69, 0x6e, 0x67, 0x20, 0x64, 0x65, 0x66, + 0x61, 0x75, 0x6c, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, - 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x55, 0x73, - 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, 0x65, - 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x20, 0x6f, 0x6e, 0x63, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x75, + 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x52, + 0x65, 0x73, 0x65, 0x74, 0x7d, 0x3e, 0x52, 0x65, 0x73, 0x65, 0x74, 0x20, + 0x61, 0x6c, 0x6c, 0x20, 0x74, 0x6f, 0x20, 0x64, 0x65, 0x66, 0x61, 0x75, + 0x6c, 0x74, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, + 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x61, 0x75, 0x74, 0x6f, 0x73, + 0x61, 0x76, 0x65, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x20, 0x6f, 0x6e, 0x20, 0x65, 0x76, 0x65, 0x72, 0x79, 0x20, 0x63, 0x68, + 0x61, 0x6e, 0x67, 0x65, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x41, 0x75, 0x74, 0x6f, 0x73, 0x61, 0x76, 0x65, 0x28, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2c, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x5d, 0x29, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, + 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x20, 0x3d, 0x20, 0x28, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, + 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x47, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, + 0x69, 0x64, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x67, 0x72, 0x61, 0x6d, 0x6d, + 0x61, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x68, 0x6f, 0x6c, + 0x64, 0x65, 0x72, 0x3d, 0x22, 0x55, 0x73, 0x65, 0x20, 0x67, 0x62, 0x6e, + 0x66, 0x20, 0x6f, 0x72, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, + 0x68, 0x65, 0x6d, 0x61, 0x2b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x74, + 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x67, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x72, 0x6f, + 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x70, + 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x22, 0x20, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x68, 0x6f, 0x6c, 0x64, 0x65, 0x72, 0x3d, 0x22, 0x6f, 0x72, 0x64, + 0x65, 0x72, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x70, 0x31, 0x2c, 0x70, 0x72, + 0x6f, 0x70, 0x32, 0x2c, 0x70, 0x72, 0x6f, 0x70, 0x33, 0x22, 0x20, 0x6f, + 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x4a, 0x73, + 0x6f, 0x6e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x50, 0x72, 0x6f, 0x70, + 0x4f, 0x72, 0x64, 0x65, 0x72, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x62, + 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x22, 0x20, 0x6f, 0x6e, 0x63, 0x6c, + 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, + 0x74, 0x4a, 0x53, 0x4f, 0x4e, 0x53, 0x63, 0x68, 0x65, 0x6d, 0x61, 0x47, + 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x7d, 0x3e, 0x43, 0x6f, 0x6e, 0x76, + 0x65, 0x72, 0x74, 0x20, 0x4a, 0x53, 0x4f, 0x4e, 0x20, 0x53, 0x63, 0x68, + 0x65, 0x6d, 0x61, 0x3c, 0x2f, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, + 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, + 0x74, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x46, 0x6f, 0x72, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x22, + 0x3e, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, + 0x61, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x78, 0x74, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x70, 0x72, 0x6f, 0x6d, + 0x70, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, + 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x7d, 0x22, 0x20, + 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, + 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, + 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, + 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x28, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, 0x6f, 0x6e, + 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, 0x65, 0x74, + 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x75, 0x73, + 0x65, 0x72, 0x22, 0x3e, 0x55, 0x73, 0x65, 0x72, 0x20, 0x6e, 0x61, 0x6d, + 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, + 0x22, 0x74, 0x65, 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, + 0x22, 0x75, 0x73, 0x65, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x75, 0x73, 0x65, 0x72, 0x7d, 0x22, + 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, + 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, - 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, - 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x68, 0x61, 0x74, - 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, - 0x63, 0x68, 0x61, 0x74, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, - 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, - 0x68, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, + 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x62, 0x6f, 0x74, + 0x22, 0x3e, 0x42, 0x6f, 0x74, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3c, 0x2f, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, + 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x74, 0x65, + 0x78, 0x74, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x63, 0x68, + 0x61, 0x72, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, + 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x63, 0x68, 0x61, 0x72, 0x7d, 0x22, 0x20, 0x6f, 0x6e, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, + 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, + 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x66, 0x6f, 0x72, 0x3d, + 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x3e, 0x50, + 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, + 0x74, 0x65, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, + 0x64, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, + 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, + 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x34, 0x20, 0x6f, 0x6e, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, + 0x66, 0x6f, 0x72, 0x3d, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, + 0x65, 0x22, 0x3e, 0x43, 0x68, 0x61, 0x74, 0x20, 0x68, 0x69, 0x73, 0x74, + 0x6f, 0x72, 0x79, 0x20, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x74, 0x65, 0x78, 0x74, 0x61, 0x72, 0x65, 0x61, 0x20, 0x69, 0x64, 0x3d, + 0x22, 0x74, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, + 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, 0x22, 0x20, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, + 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x68, 0x69, 0x73, + 0x74, 0x6f, 0x72, 0x79, 0x54, 0x65, 0x6d, 0x70, 0x6c, 0x61, 0x74, 0x65, + 0x7d, 0x22, 0x20, 0x72, 0x6f, 0x77, 0x73, 0x3d, 0x31, 0x20, 0x6f, 0x6e, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, + 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x47, 0x72, + 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, + 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x43, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, + 0x46, 0x6f, 0x72, 0x6d, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x6d, 0x70, 0x74, 0x43, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x53, + 0x65, 0x74, 0x28, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, + 0x74, 0x3e, 0x24, 0x7b, 0x47, 0x72, 0x61, 0x6d, 0x6d, 0x61, 0x72, 0x43, + 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, 0x28, 0x29, 0x7d, 0x3c, 0x2f, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x72, + 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, + 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x24, 0x7b, 0x55, 0x73, 0x65, 0x72, 0x54, 0x65, 0x6d, 0x70, 0x6c, + 0x61, 0x74, 0x65, 0x52, 0x65, 0x73, 0x65, 0x74, 0x42, 0x75, 0x74, 0x74, + 0x6f, 0x6e, 0x7d, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, - 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, - 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x73, - 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x22, 0x63, - 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x7d, 0x20, + 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, + 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, + 0x3d, 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x68, 0x61, 0x74, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x7d, - 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, - 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, 0x27, 0x20, 0x3f, - 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, - 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, - 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, - 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, - 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, - 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, - 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, - 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, - 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, 0x70, 0x72, - 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, - 0x74, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, - 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x65, - 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, - 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x65, 0x6d, 0x70, - 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x20, 0x7d, 0x29, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, - 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x20, 0x72, 0x65, 0x70, 0x65, - 0x61, 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x65, 0x22, - 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x2e, 0x30, 0x2c, 0x20, - 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, - 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, - 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, - 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, - 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, - 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x43, 0x6f, - 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, 0x4e, 0x20, 0x74, 0x6f, 0x6b, - 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, 0x70, - 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x22, 0x2c, + 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x68, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, + 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x73, 0x6c, 0x69, + 0x6d, 0x22, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, + 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x74, 0x79, 0x70, 0x65, 0x22, 0x20, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, + 0x64, 0x3d, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, + 0x3d, 0x3d, 0x20, 0x22, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x22, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x53, 0x65, 0x73, + 0x73, 0x69, 0x6f, 0x6e, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x3c, 0x2f, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, + 0x61, 0x74, 0x27, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x43, 0x6f, + 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x20, 0x3a, + 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, + 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x28, 0x29, 0x7d, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, + 0x61, 0x73, 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, + 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x72, 0x65, 0x64, + 0x69, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, 0x69, 0x6e, + 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, + 0x22, 0x6e, 0x5f, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, 0x65, 0x70, - 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x6e, 0x20, 0x7d, - 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, 0x6c, - 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x54, 0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, - 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, - 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x2d, 0x31, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x6b, - 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x6f, 0x70, 0x5f, 0x6b, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, - 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, - 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, - 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, - 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, - 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x22, 0x2c, 0x20, - 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, - 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, 0x6f, 0x72, - 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, 0x2f, 0x73, - 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, - 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, - 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, - 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, - 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x46, 0x53, 0x2d, - 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, - 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, 0x73, 0x5f, 0x7a, - 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, - 0x66, 0x73, 0x5f, 0x7a, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, + 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, - 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x79, - 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, 0x6d, 0x61, - 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, - 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, - 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x22, 0x2c, - 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, - 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, - 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, + 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x65, + 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x35, 0x2c, 0x20, 0x6d, 0x69, + 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3a, 0x20, 0x22, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, + 0x72, 0x65, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x65, 0x6d, 0x70, 0x65, 0x72, 0x61, 0x74, 0x75, 0x72, 0x65, + 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, + 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, + 0x6c, 0x3a, 0x20, 0x22, 0x50, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, + 0x20, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x20, 0x73, 0x65, 0x71, 0x75, + 0x65, 0x6e, 0x63, 0x65, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, + 0x32, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x72, 0x65, + 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x72, + 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, + 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x43, 0x6f, 0x6e, 0x73, 0x69, 0x64, 0x65, 0x72, 0x20, + 0x4e, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x66, 0x6f, 0x72, + 0x20, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x69, 0x7a, 0x65, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x32, 0x30, 0x34, 0x38, 0x2c, 0x20, 0x6d, + 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, + 0x20, 0x22, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, + 0x74, 0x5f, 0x6e, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, + 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x72, 0x65, 0x70, 0x65, 0x61, 0x74, 0x5f, 0x6c, 0x61, 0x73, + 0x74, 0x5f, 0x6e, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, + 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, + 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x4b, 0x20, 0x73, + 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x31, 0x30, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x2d, 0x31, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x6b, 0x20, 0x7d, 0x29, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, + 0x22, 0x54, 0x6f, 0x70, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, + 0x69, 0x6e, 0x67, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, + 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x6f, 0x70, + 0x5f, 0x70, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x6f, 0x70, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, + 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, + 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, + 0x6e, 0x2d, 0x50, 0x20, 0x73, 0x61, 0x6d, 0x70, 0x6c, 0x69, 0x6e, 0x67, + 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, + 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, + 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, 0x69, 0x6e, 0x5f, 0x70, 0x22, + 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, + 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, + 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, + 0x6e, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x4d, + 0x6f, 0x72, 0x65, 0x20, 0x6f, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x73, 0x3c, + 0x2f, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, + 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, + 0x73, 0x3d, 0x22, 0x74, 0x77, 0x6f, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, + 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, + 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x54, 0x46, + 0x53, 0x2d, 0x5a, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, + 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x74, 0x66, 0x73, + 0x5f, 0x7a, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x74, 0x66, 0x73, 0x5f, 0x7a, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, - 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x70, 0x72, - 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, - 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, - 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, - 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x46, - 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, 0x65, 0x6e, - 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, 0x20, - 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, 0x2e, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x66, 0x72, - 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, - 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, + 0x54, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x20, 0x50, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, + 0x6e, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, + 0x3a, 0x20, 0x22, 0x74, 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, + 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, + 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, + 0x79, 0x70, 0x69, 0x63, 0x61, 0x6c, 0x5f, 0x70, 0x20, 0x7d, 0x29, 0x7d, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, + 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, + 0x20, 0x22, 0x50, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x20, 0x70, + 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, + 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, + 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x73, 0x65, 0x6e, 0x63, 0x65, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, + 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, + 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, + 0x22, 0x46, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x20, 0x70, + 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, + 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, + 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, 0x79, 0x5f, 0x70, 0x65, + 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, + 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x66, 0x72, 0x65, 0x71, 0x75, 0x65, 0x6e, 0x63, + 0x79, 0x5f, 0x70, 0x65, 0x6e, 0x61, 0x6c, 0x74, 0x79, 0x20, 0x7d, 0x29, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, + 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, + 0x3d, 0x22, 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x68, 0x72, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, - 0x64, 0x73, 0x65, 0x74, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, - 0x74, 0x68, 0x72, 0x65, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, - 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, - 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, - 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x30, 0x22, - 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, 0x20, - 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, - 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, 0x6d, - 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, 0x6f, 0x20, - 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, - 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, - 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, - 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, - 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, 0x31, - 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, - 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, 0x3d, - 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, - 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, 0x61, - 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, 0x2f, 0x6c, - 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, 0x22, - 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, + 0x30, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x3d, - 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x3d, 0x20, 0x30, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x4d, - 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, 0x3c, 0x2f, + 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, 0x6e, + 0x6f, 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, - 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, - 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, - 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, - 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, - 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, - 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, - 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, - 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, 0x75, 0x20, - 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, 0x6f, 0x61, - 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, - 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, - 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, 0x78, 0x3a, - 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, 0x30, - 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6d, - 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x22, - 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x31, - 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, - 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, - 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x20, 0x7d, - 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, - 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, 0x65, 0x74, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, 0x69, 0x65, - 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, - 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, - 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, 0x20, 0x6d, 0x61, - 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, 0x20, - 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, 0x6e, 0x5f, - 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, - 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x20, - 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, - 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, - 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x43, 0x6f, - 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, 0x3d, 0x3e, - 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, - 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, - 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x67, 0x20, - 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, - 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, - 0x60, 0x72, 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, 0x2c, 0x24, - 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, 0x60, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, - 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, 0x28, 0x70, - 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, 0x74, 0x61, - 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x6d, - 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, - 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, - 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, - 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, - 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, 0x3e, 0x20, - 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, 0x66, 0x6f, - 0x72, 0x20, 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, 0x72, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, - 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, - 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, 0x74, 0x68, - 0x28, 0x27, 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, 0x27, 0x29, - 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, - 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, - 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, 0x20, - 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, - 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, - 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x62, 0x20, - 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, - 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, 0x6e, 0x74, - 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, - 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, - 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, 0x6f, 0x62, - 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, - 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, 0x74, 0x61, - 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, 0x61, - 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, 0x72, 0x6f, - 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, - 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, - 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, - 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x66, 0x6f, - 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, - 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, 0x70, 0x2e, - 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, - 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x29, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, - 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, - 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, 0x62, - 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x2e, - 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, - 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, - 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, - 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x28, 0x70, - 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, 0x3e, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, 0x74, + 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, 0x6f, + 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, 0x6f, + 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3d, + 0x22, 0x31, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, 0x3d, + 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, + 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, + 0x3d, 0x3d, 0x20, 0x31, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, 0x61, + 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, 0x20, + 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x31, 0x3c, + 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, 0x69, 0x6e, - 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, - 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, 0x6f, 0x62, - 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x7d, 0x60, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, - 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, 0x64, 0x64, - 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, 0x6d, 0x27, - 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, - 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x2e, 0x74, - 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x63, - 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, 0x72, 0x6f, - 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, - 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, 0x73, 0x70, - 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, - 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, - 0x73, 0x74, 0x72, 0x7d, 0x3a, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, - 0x6e, 0x3e, 0x24, 0x7b, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, - 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x20, 0x2a, 0x20, - 0x31, 0x30, 0x30, 0x29, 0x7d, 0x25, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, - 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x3c, 0x69, 0x6e, 0x70, 0x75, + 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x72, 0x61, 0x64, 0x69, + 0x6f, 0x22, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3d, 0x22, 0x6d, 0x69, 0x72, + 0x6f, 0x73, 0x74, 0x61, 0x74, 0x22, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3d, 0x22, 0x32, 0x22, 0x20, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x65, 0x64, + 0x3d, 0x24, 0x7b, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, + 0x20, 0x3d, 0x3d, 0x20, 0x32, 0x7d, 0x20, 0x6f, 0x6e, 0x69, 0x6e, 0x70, + 0x75, 0x74, 0x3d, 0x24, 0x7b, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x50, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x49, 0x6e, 0x74, 0x7d, 0x20, 0x2f, 0x3e, + 0x20, 0x4d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x20, 0x76, 0x32, + 0x3c, 0x2f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, + 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, + 0x74, 0x61, 0x74, 0x20, 0x74, 0x61, 0x75, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, + 0x3a, 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, + 0x20, 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, + 0x61, 0x75, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, + 0x2e, 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, + 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x2e, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x74, 0x61, + 0x75, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x46, 0x6c, + 0x6f, 0x61, 0x74, 0x46, 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, + 0x61, 0x62, 0x65, 0x6c, 0x3a, 0x20, 0x22, 0x4d, 0x69, 0x72, 0x6f, 0x73, + 0x74, 0x61, 0x74, 0x20, 0x65, 0x74, 0x61, 0x22, 0x2c, 0x20, 0x6d, 0x61, + 0x78, 0x3a, 0x20, 0x31, 0x2e, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, 0x3a, + 0x20, 0x30, 0x2e, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, + 0x22, 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, + 0x61, 0x22, 0x2c, 0x20, 0x73, 0x74, 0x65, 0x70, 0x3a, 0x20, 0x30, 0x2e, + 0x30, 0x31, 0x2c, 0x20, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, + 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x6d, 0x69, 0x72, 0x6f, 0x73, 0x74, 0x61, 0x74, 0x5f, 0x65, 0x74, 0x61, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, 0x64, + 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x69, 0x65, 0x6c, 0x64, 0x73, + 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x49, 0x6e, 0x74, 0x46, + 0x69, 0x65, 0x6c, 0x64, 0x28, 0x7b, 0x20, 0x6c, 0x61, 0x62, 0x65, 0x6c, + 0x3a, 0x20, 0x22, 0x53, 0x68, 0x6f, 0x77, 0x20, 0x50, 0x72, 0x6f, 0x62, + 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x22, 0x2c, 0x20, + 0x6d, 0x61, 0x78, 0x3a, 0x20, 0x31, 0x30, 0x2c, 0x20, 0x6d, 0x69, 0x6e, + 0x3a, 0x20, 0x30, 0x2c, 0x20, 0x6e, 0x61, 0x6d, 0x65, 0x3a, 0x20, 0x22, + 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x22, 0x2c, 0x20, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x3a, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, + 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, + 0x73, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x66, 0x69, 0x65, 0x6c, + 0x64, 0x73, 0x65, 0x74, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, + 0x73, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x2f, 0x66, 0x6f, 0x72, 0x6d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x72, 0x6f, 0x62, + 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x29, 0x20, + 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, + 0x68, 0x2e, 0x66, 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, + 0x2a, 0x20, 0x28, 0x31, 0x20, 0x2d, 0x20, 0x70, 0x29, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x67, 0x20, 0x3d, 0x20, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, 0x6c, 0x6f, + 0x6f, 0x72, 0x28, 0x31, 0x39, 0x32, 0x20, 0x2a, 0x20, 0x70, 0x29, 0x3b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x60, 0x72, 0x67, 0x62, 0x61, 0x28, 0x24, 0x7b, 0x72, 0x7d, + 0x2c, 0x24, 0x7b, 0x67, 0x7d, 0x2c, 0x30, 0x2c, 0x30, 0x2e, 0x33, 0x29, + 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x72, 0x6f, 0x62, + 0x61, 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x3d, 0x20, + 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x2e, 0x64, 0x61, + 0x74, 0x61, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x6d, 0x73, 0x67, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7d, 0x20, 0x3d, + 0x20, 0x6d, 0x73, 0x67, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x21, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x20, 0x7c, 0x7c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, + 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, + 0x67, 0x74, 0x68, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x30, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, + 0x72, 0x6e, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6c, 0x65, 0x6e, 0x67, 0x74, 0x68, 0x20, + 0x3e, 0x20, 0x31, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x4e, 0x6f, 0x74, 0x20, + 0x66, 0x6f, 0x72, 0x20, 0x62, 0x79, 0x74, 0x65, 0x20, 0x70, 0x61, 0x69, + 0x72, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x69, 0x66, 0x20, 0x28, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x2e, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x2e, 0x73, 0x74, 0x61, 0x72, 0x74, 0x73, 0x57, 0x69, + 0x74, 0x68, 0x28, 0x27, 0x62, 0x79, 0x74, 0x65, 0x3a, 0x20, 0x5c, 0x5c, + 0x27, 0x29, 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6d, + 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, + 0x6e, 0x73, 0x74, 0x20, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, 0x74, + 0x61, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x69, + 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, 0x6c, 0x69, + 0x74, 0x69, 0x65, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, 0x70, 0x72, 0x6f, + 0x62, 0x20, 0x3d, 0x3e, 0x20, 0x28, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x74, + 0x65, 0x6e, 0x74, 0x3a, 0x20, 0x70, 0x72, 0x6f, 0x62, 0x2e, 0x63, 0x6f, + 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, + 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, + 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x3a, 0x20, 0x5b, 0x70, 0x72, + 0x6f, 0x62, 0x5d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x7d, 0x29, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x72, 0x6f, 0x62, 0x61, + 0x62, 0x69, 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x7d, 0x20, 0x64, 0x61, + 0x74, 0x61, 0x3d, 0x24, 0x7b, 0x73, 0x70, 0x6c, 0x69, 0x74, 0x44, 0x61, + 0x74, 0x61, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x7b, 0x20, 0x70, + 0x72, 0x6f, 0x62, 0x73, 0x2c, 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x20, 0x7d, 0x20, 0x3d, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, + 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x70, 0x72, 0x6f, 0x62, 0x61, 0x62, 0x69, + 0x6c, 0x69, 0x74, 0x69, 0x65, 0x73, 0x5b, 0x30, 0x5d, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x72, 0x6f, 0x62, + 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x28, 0x70, 0x20, 0x3d, 0x3e, 0x20, + 0x70, 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, + 0x3d, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, + 0x74, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, + 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x20, + 0x3d, 0x20, 0x66, 0x6f, 0x75, 0x6e, 0x64, 0x20, 0x3f, 0x20, 0x70, 0x72, + 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x66, 0x6f, 0x75, 0x6e, + 0x64, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, + 0x72, 0x61, 0x6e, 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, + 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x3d, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x70, 0x72, 0x6f, 0x62, 0x2d, 0x73, 0x65, 0x74, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x62, 0x73, 0x2e, 0x6d, 0x61, 0x70, 0x28, + 0x28, 0x70, 0x2c, 0x20, 0x69, 0x6e, 0x64, 0x65, 0x78, 0x29, 0x20, 0x3d, + 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, + 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6b, 0x65, 0x79, 0x3d, 0x24, 0x7b, + 0x69, 0x6e, 0x64, 0x65, 0x78, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x60, 0x70, 0x72, + 0x6f, 0x62, 0x3a, 0x20, 0x24, 0x7b, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, + 0x7d, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, + 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x61, + 0x64, 0x64, 0x69, 0x6e, 0x67, 0x3a, 0x20, 0x27, 0x30, 0x2e, 0x33, 0x65, + 0x6d, 0x27, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, + 0x6f, 0x75, 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, + 0x2e, 0x74, 0x6f, 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x20, 0x3d, 0x3d, 0x3d, + 0x20, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x20, 0x3f, 0x20, 0x70, + 0x72, 0x6f, 0x62, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, + 0x72, 0x6f, 0x62, 0x29, 0x20, 0x3a, 0x20, 0x27, 0x74, 0x72, 0x61, 0x6e, + 0x73, 0x70, 0x61, 0x72, 0x65, 0x6e, 0x74, 0x27, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x70, 0x2e, 0x74, 0x6f, + 0x6b, 0x5f, 0x73, 0x74, 0x72, 0x7d, 0x3a, 0x20, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, + 0x70, 0x61, 0x6e, 0x3e, 0x24, 0x7b, 0x4d, 0x61, 0x74, 0x68, 0x2e, 0x66, + 0x6c, 0x6f, 0x6f, 0x72, 0x28, 0x70, 0x2e, 0x70, 0x72, 0x6f, 0x62, 0x20, + 0x2a, 0x20, 0x31, 0x30, 0x30, 0x29, 0x7d, 0x25, 0x3c, 0x2f, 0x73, 0x70, + 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, + 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x29, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, + 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, + 0x24, 0x7b, 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, + 0x6e, 0x64, 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, + 0x6c, 0x6f, 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, + 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, + 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, + 0x63, 0x68, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, + 0x3f, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, + 0x3e, 0x60, 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, + 0x61, 0x6e, 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, + 0x20, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, + 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, + 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, + 0x6d, 0x73, 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, + 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, + 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, + 0x67, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, + 0x5e, 0x23, 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, + 0x24, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, + 0x24, 0x31, 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, + 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, + 0x29, 0x5c, 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, + 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, + 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, + 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, + 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, + 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, + 0x2a, 0x3f, 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, + 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, + 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, + 0x5f, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, + 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, + 0x28, 0x2f, 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, + 0x5c, 0x73, 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, + 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, + 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, + 0x3c, 0x2f, 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, + 0x2c, 0x20, 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, + 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, + 0x65, 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, + 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, + 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, + 0x6e, 0x67, 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, + 0x49, 0x6e, 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, + 0x7b, 0x20, 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, + 0x20, 0x7d, 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, + 0x73, 0x74, 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, + 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, + 0x20, 0x28, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x21, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, + 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, + 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, + 0x3e, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x70, 0x6f, 0x76, - 0x65, 0x72, 0x7d, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, - 0x7b, 0x20, 0x62, 0x61, 0x63, 0x6b, 0x67, 0x72, 0x6f, 0x75, 0x6e, 0x64, - 0x43, 0x6f, 0x6c, 0x6f, 0x72, 0x3a, 0x20, 0x70, 0x43, 0x6f, 0x6c, 0x6f, - 0x72, 0x20, 0x7d, 0x7d, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, - 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x3d, 0x24, 0x7b, 0x70, - 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, - 0x65, 0x6e, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6d, 0x73, 0x67, 0x2e, 0x63, - 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x2e, 0x6d, 0x61, 0x74, 0x63, 0x68, - 0x28, 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x29, 0x20, 0x3f, 0x20, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x62, 0x72, 0x20, 0x2f, 0x3e, 0x60, - 0x20, 0x3a, 0x20, 0x6d, 0x73, 0x67, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x6e, 0x74, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x2f, 0x2f, 0x20, 0x70, 0x6f, 0x6f, 0x72, 0x20, 0x6d, 0x61, 0x6e, - 0x73, 0x20, 0x6d, 0x61, 0x72, 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x20, 0x72, - 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x6d, 0x65, 0x6e, 0x74, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x4d, 0x61, 0x72, - 0x6b, 0x64, 0x6f, 0x77, 0x6e, 0x69, 0x73, 0x68, 0x20, 0x3d, 0x20, 0x28, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x6d, 0x64, 0x20, 0x3d, 0x20, 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, - 0x2e, 0x74, 0x65, 0x78, 0x74, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x26, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x61, 0x6d, 0x70, 0x3b, 0x27, - 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, - 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x3c, 0x2f, 0x67, 0x2c, - 0x20, 0x27, 0x26, 0x6c, 0x74, 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, - 0x65, 0x28, 0x2f, 0x3e, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x26, 0x67, 0x74, - 0x3b, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5e, 0x23, - 0x7b, 0x31, 0x2c, 0x36, 0x7d, 0x20, 0x28, 0x2e, 0x2a, 0x29, 0x24, 0x2f, - 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x68, 0x33, 0x3e, 0x24, 0x31, - 0x3c, 0x2f, 0x68, 0x33, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, - 0x28, 0x2f, 0x5c, 0x2a, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5c, - 0x2a, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, - 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, - 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x5f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x5f, 0x2f, 0x67, 0x2c, - 0x20, 0x27, 0x3c, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x24, 0x31, - 0x3c, 0x2f, 0x73, 0x74, 0x72, 0x6f, 0x6e, 0x67, 0x3e, 0x27, 0x29, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, - 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, 0x5c, 0x2a, 0x28, 0x2e, 0x2a, 0x3f, - 0x29, 0x5c, 0x2a, 0x2f, 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, - 0x24, 0x31, 0x3c, 0x2f, 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, - 0x63, 0x65, 0x28, 0x2f, 0x5f, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x5f, 0x2f, - 0x67, 0x2c, 0x20, 0x27, 0x3c, 0x65, 0x6d, 0x3e, 0x24, 0x31, 0x3c, 0x2f, - 0x65, 0x6d, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, 0x2f, - 0x60, 0x60, 0x60, 0x2e, 0x2a, 0x3f, 0x5c, 0x6e, 0x28, 0x5b, 0x5c, 0x73, - 0x5c, 0x53, 0x5d, 0x2a, 0x3f, 0x29, 0x60, 0x60, 0x60, 0x2f, 0x67, 0x2c, - 0x20, 0x27, 0x3c, 0x70, 0x72, 0x65, 0x3e, 0x3c, 0x63, 0x6f, 0x64, 0x65, - 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x3c, 0x2f, - 0x70, 0x72, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x60, 0x28, 0x2e, 0x2a, 0x3f, 0x29, 0x60, 0x2f, 0x67, 0x2c, 0x20, - 0x27, 0x3c, 0x63, 0x6f, 0x64, 0x65, 0x3e, 0x24, 0x31, 0x3c, 0x2f, 0x63, - 0x6f, 0x64, 0x65, 0x3e, 0x27, 0x29, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x2e, 0x72, 0x65, 0x70, 0x6c, 0x61, 0x63, 0x65, 0x28, - 0x2f, 0x5c, 0x6e, 0x2f, 0x67, 0x69, 0x6d, 0x2c, 0x20, 0x27, 0x3c, 0x62, - 0x72, 0x20, 0x2f, 0x3e, 0x27, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, - 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, 0x64, 0x61, 0x6e, 0x67, - 0x65, 0x72, 0x6f, 0x75, 0x73, 0x6c, 0x79, 0x53, 0x65, 0x74, 0x49, 0x6e, - 0x6e, 0x65, 0x72, 0x48, 0x54, 0x4d, 0x4c, 0x3d, 0x24, 0x7b, 0x7b, 0x20, - 0x5f, 0x5f, 0x68, 0x74, 0x6d, 0x6c, 0x3a, 0x20, 0x6d, 0x64, 0x20, 0x7d, - 0x7d, 0x20, 0x2f, 0x3e, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, - 0x20, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, - 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, 0x20, 0x3d, 0x20, 0x28, - 0x70, 0x61, 0x72, 0x61, 0x6d, 0x73, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, - 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, - 0x61, 0x6c, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, - 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x2f, 0x3e, 0x60, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, - 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, - 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, - 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, 0x2e, 0x74, - 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, 0x73, 0x20, - 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, 0x20, 0x24, - 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, - 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, - 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, 0x63, 0x6f, - 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x32, - 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, 0x70, 0x65, - 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, 0x6e, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x73, - 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, - 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, - 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x20, 0x3d, - 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x66, - 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, - 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, - 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x27, - 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, - 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x62, 0x75, - 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, 0x75, 0x73, - 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, + 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, + 0x70, 0x65, 0x72, 0x5f, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x5f, 0x6d, 0x73, + 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, 0x28, 0x29, 0x7d, 0x6d, + 0x73, 0x20, 0x70, 0x65, 0x72, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x2c, + 0x20, 0x24, 0x7b, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x53, 0x74, 0x61, 0x74, + 0x73, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x70, 0x72, 0x65, 0x64, + 0x69, 0x63, 0x74, 0x65, 0x64, 0x5f, 0x70, 0x65, 0x72, 0x5f, 0x73, 0x65, + 0x63, 0x6f, 0x6e, 0x64, 0x2e, 0x74, 0x6f, 0x46, 0x69, 0x78, 0x65, 0x64, + 0x28, 0x32, 0x29, 0x7d, 0x20, 0x74, 0x6f, 0x6b, 0x65, 0x6e, 0x73, 0x20, + 0x70, 0x65, 0x72, 0x20, 0x73, 0x65, 0x63, 0x6f, 0x6e, 0x64, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x70, 0x61, + 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, + 0x20, 0x73, 0x69, 0x6d, 0x70, 0x6c, 0x65, 0x20, 0x70, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x20, 0x69, 0x6d, 0x70, 0x6c, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x50, 0x6f, 0x70, 0x6f, 0x76, + 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, + 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, 0x69, 0x67, 0x6e, 0x61, 0x6c, + 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x70, 0x6f, 0x73, + 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x53, + 0x69, 0x67, 0x6e, 0x61, 0x6c, 0x28, 0x7b, 0x20, 0x74, 0x6f, 0x70, 0x3a, + 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, 0x2c, 0x20, 0x6c, 0x65, 0x66, 0x74, + 0x3a, 0x20, 0x27, 0x30, 0x70, 0x78, 0x27, 0x20, 0x7d, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, - 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x20, 0x3d, - 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, - 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, - 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, - 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, 0x20, 0x3d, - 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, - 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, - 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, 0x74, 0x20, 0x3d, 0x20, - 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, - 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, 0x6f, 0x75, - 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, 0x74, 0x52, - 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, - 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, - 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, 0x77, 0x69, - 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x59, - 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, - 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, 0x66, 0x74, - 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, - 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, - 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, 0x69, 0x73, - 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x68, 0x61, - 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, - 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, 0x65, 0x6e, - 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, 0x70, 0x6f, - 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, - 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, 0x6f, 0x76, - 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, - 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, - 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, - 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, - 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x63, - 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, 0x65, 0x6e, - 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, - 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x75, - 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, 0x29, 0x20, - 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x61, 0x64, - 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, - 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, - 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, - 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x64, 0x6f, - 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x76, - 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, 0x65, 0x6e, - 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, 0x6f, 0x77, - 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, - 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x29, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, 0x5d, 0x29, - 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, - 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, 0x6e, 0x20, - 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, 0x65, 0x66, - 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, - 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, 0x24, 0x7b, - 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, - 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, - 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, 0x73, 0x70, - 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, - 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, - 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, 0x6e, 0x74, - 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x66, 0x3d, - 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x70, - 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, 0x74, 0x65, - 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, - 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, 0x6f, 0x73, - 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, - 0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, 0x20, 0x70, - 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, 0x6f, 0x70, - 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, - 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, - 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, 0x72, 0x63, - 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, - 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, - 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, - 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, 0x2f, 0x70, - 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, - 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, 0x65, 0x72, - 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, - 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, 0x69, 0x72, - 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, - 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, 0x63, 0x65, 0x6e, 0x64, - 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x74, 0x68, - 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, 0x53, 0x20, - 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, 0x2f, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, 0x50, 0x6f, - 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, - 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, - 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, 0x70, 0x64, 0x61, 0x74, - 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, 0x20, 0x28, - 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, 0x65, 0x74, - 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, - 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, - 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, - 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, - 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, - 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, 0x3d, 0x20, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, - 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, 0x74, 0x68, - 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, - 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, 0x69, 0x6c, - 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, 0x20, 0x7b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, - 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, - 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, 0x20, 0x66, - 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, - 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, - 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, 0x61, 0x72, - 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, 0x6d, 0x6f, - 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x6e, 0x6f, - 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, 0x79, 0x70, - 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, 0x3d, 0x3d, - 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, 0x3f, 0x20, - 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, - 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x6e, - 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x3b, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, - 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, 0x20, 0x74, - 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x29, 0x20, - 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, 0x65, 0x61, - 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, 0x6f, 0x64, - 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, 0x67, 0x20, - 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x21, - 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, - 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, - 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, - 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, - 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, - 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, 0x26, 0x20, - 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, + 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x20, 0x3d, 0x20, + 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, 0x6c, 0x6c, 0x29, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, + 0x74, 0x20, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, + 0x20, 0x3d, 0x20, 0x75, 0x73, 0x65, 0x52, 0x65, 0x66, 0x28, 0x6e, 0x75, + 0x6c, 0x6c, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, + 0x50, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x28, 0x29, + 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, + 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, - 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, - 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, 0x3e, 0x60, - 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x2c, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, - 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x3d, - 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, - 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, - 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, - 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, - 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, 0x6f, 0x6e, - 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, 0x73, 0x2e, - 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x24, 0x7b, - 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, - 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, - 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x7d, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, - 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, - 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, - 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x6e, 0x75, - 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, - 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, 0x72, 0x20, - 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x74, 0x68, - 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, 0x20, 0x69, - 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x63, 0x68, 0x69, - 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, 0x20, 0x65, 0x78, 0x69, - 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, - 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, 0x63, 0x6f, - 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, 0x72, 0x65, - 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, 0x6f, 0x78, - 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, - 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, - 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, - 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x43, - 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, - 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, 0x68, 0x69, - 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, - 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, - 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, - 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, - 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, - 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x63, 0x6c, - 0x61, 0x73, 0x73, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x65, 0x2d, 0x24, 0x7b, - 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, - 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x22, 0x3e, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x65, 0x61, - 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, 0x6c, 0x61, - 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, - 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, 0x69, 0x6e, - 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, + 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, 0x72, 0x65, 0x63, 0x74, 0x20, + 0x3d, 0x20, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, 0x65, 0x66, 0x2e, + 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, 0x2e, 0x67, 0x65, 0x74, 0x42, + 0x6f, 0x75, 0x6e, 0x64, 0x69, 0x6e, 0x67, 0x43, 0x6c, 0x69, 0x65, 0x6e, + 0x74, 0x52, 0x65, 0x63, 0x74, 0x28, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, + 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, + 0x63, 0x74, 0x2e, 0x62, 0x6f, 0x74, 0x74, 0x6f, 0x6d, 0x20, 0x2b, 0x20, + 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, 0x73, 0x63, 0x72, 0x6f, 0x6c, + 0x6c, 0x59, 0x7d, 0x70, 0x78, 0x60, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, + 0x3a, 0x20, 0x60, 0x24, 0x7b, 0x72, 0x65, 0x63, 0x74, 0x2e, 0x6c, 0x65, + 0x66, 0x74, 0x20, 0x2b, 0x20, 0x77, 0x69, 0x6e, 0x64, 0x6f, 0x77, 0x2e, + 0x73, 0x63, 0x72, 0x6f, 0x6c, 0x6c, 0x58, 0x7d, 0x70, 0x78, 0x60, 0x2c, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x73, 0x4f, 0x70, + 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, 0x3d, 0x20, 0x21, + 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6e, 0x73, 0x74, 0x20, + 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, + 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, 0x20, 0x3d, 0x20, 0x28, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x6f, + 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, + 0x72, 0x65, 0x6e, 0x74, 0x20, 0x26, 0x26, 0x20, 0x21, 0x70, 0x6f, 0x70, + 0x6f, 0x76, 0x65, 0x72, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, + 0x65, 0x6e, 0x74, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, + 0x28, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, + 0x74, 0x29, 0x20, 0x26, 0x26, 0x20, 0x21, 0x62, 0x75, 0x74, 0x74, 0x6f, + 0x6e, 0x52, 0x65, 0x66, 0x2e, 0x63, 0x75, 0x72, 0x72, 0x65, 0x6e, 0x74, + 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x73, 0x28, 0x65, 0x76, + 0x65, 0x6e, 0x74, 0x2e, 0x74, 0x61, 0x72, 0x67, 0x65, 0x74, 0x29, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3d, 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x75, 0x73, 0x65, 0x45, 0x66, 0x66, 0x65, 0x63, 0x74, 0x28, 0x28, + 0x29, 0x20, 0x3d, 0x3e, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, + 0x61, 0x64, 0x64, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, + 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, + 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, + 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x28, 0x29, 0x20, 0x3d, 0x3e, 0x20, + 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x72, 0x65, 0x6d, + 0x6f, 0x76, 0x65, 0x45, 0x76, 0x65, 0x6e, 0x74, 0x4c, 0x69, 0x73, 0x74, + 0x65, 0x6e, 0x65, 0x72, 0x28, 0x27, 0x6d, 0x6f, 0x75, 0x73, 0x65, 0x64, + 0x6f, 0x77, 0x6e, 0x27, 0x2c, 0x20, 0x68, 0x61, 0x6e, 0x64, 0x6c, 0x65, + 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x4f, 0x75, 0x74, 0x73, 0x69, 0x64, 0x65, + 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, + 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x2c, 0x20, 0x5b, + 0x5d, 0x29, 0x3b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, + 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x70, 0x61, + 0x6e, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x24, 0x7b, 0x70, 0x72, + 0x6f, 0x70, 0x73, 0x2e, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x7d, 0x20, 0x72, + 0x65, 0x66, 0x3d, 0x24, 0x7b, 0x62, 0x75, 0x74, 0x74, 0x6f, 0x6e, 0x52, + 0x65, 0x66, 0x7d, 0x20, 0x6f, 0x6e, 0x43, 0x6c, 0x69, 0x63, 0x6b, 0x3d, + 0x24, 0x7b, 0x74, 0x6f, 0x67, 0x67, 0x6c, 0x65, 0x50, 0x6f, 0x70, 0x6f, + 0x76, 0x65, 0x72, 0x7d, 0x3e, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, + 0x2e, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x7d, 0x3c, 0x2f, + 0x73, 0x70, 0x61, 0x6e, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x24, 0x7b, 0x69, 0x73, 0x4f, 0x70, 0x65, 0x6e, 0x2e, 0x76, + 0x61, 0x6c, 0x75, 0x65, 0x20, 0x26, 0x26, 0x20, 0x68, 0x74, 0x6d, 0x6c, + 0x60, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x20, 0x69, + 0x6e, 0x74, 0x6f, 0x3d, 0x22, 0x23, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, 0x53, 0x74, - 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x20, - 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, 0x3a, 0x20, - 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, 0x7d, 0x20, - 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, 0x65, 0x63, - 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, 0x72, 0x69, - 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, - 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, - 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, 0x61, 0x74, - 0x27, 0x20, 0x3f, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x49, - 0x6e, 0x70, 0x75, 0x74, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6c, - 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x74, 0x72, 0x6f, 0x6c, - 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, 0x69, 0x6f, - 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x47, 0x65, - 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, 0x66, 0x6f, - 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x70, 0x3e, - 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, 0x20, 0x3c, - 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, + 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x66, 0x3d, 0x24, 0x7b, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x52, + 0x65, 0x66, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, + 0x22, 0x70, 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x2d, 0x63, 0x6f, 0x6e, + 0x74, 0x65, 0x6e, 0x74, 0x22, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x73, 0x74, 0x79, 0x6c, + 0x65, 0x3d, 0x24, 0x7b, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x6f, 0x70, 0x3a, 0x20, 0x70, + 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x2e, 0x74, 0x6f, 0x70, 0x2c, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x6c, 0x65, 0x66, 0x74, 0x3a, + 0x20, 0x70, 0x6f, 0x73, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x6c, 0x65, 0x66, 0x74, 0x2c, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x7d, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x24, 0x7b, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x70, + 0x6f, 0x70, 0x6f, 0x76, 0x65, 0x72, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x72, + 0x65, 0x6e, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x24, + 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x7d, 0x3e, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x7d, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x3b, + 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x53, 0x6f, 0x75, + 0x72, 0x63, 0x65, 0x3a, 0x20, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, + 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, 0x28, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, - 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, 0x6f, 0x76, - 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x22, 0x3e, - 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x61, - 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, - 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, - 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, 0x6d, 0x6c, - 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e, 0x3c, 0x2f, 0x70, 0x3e, - 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, - 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, - 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, 0x20, 0x20, - 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, - 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70, 0x29, 0x2c, 0x20, 0x64, - 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, 0x65, 0x72, - 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, 0x27, 0x23, - 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x27, 0x29, 0x29, - 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, - 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x3e, 0x0a, 0x0a, 0x3c, - 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, - 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, - 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x69, 0x6e, - 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, 0x66, 0x69, - 0x6c, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x66, 0x69, 0x6c, 0x65, - 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x20, 0x61, 0x63, 0x63, 0x65, 0x70, - 0x74, 0x3d, 0x22, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x2f, 0x2a, 0x22, 0x20, - 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x64, 0x69, 0x73, 0x70, 0x6c, - 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x22, 0x3e, 0x0a, - 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, 0x20, 0x3c, - 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x70, 0x6f, 0x72, 0x74, - 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x3c, - 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, 0x68, 0x74, - 0x6d, 0x6c, 0x3e, 0x0a, 0x0a + 0x6f, 0x6d, 0x2f, 0x64, 0x65, 0x76, 0x65, 0x6c, 0x6f, 0x70, 0x69, 0x74, + 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, + 0x61, 0x6c, 0x2f, 0x62, 0x6c, 0x6f, 0x62, 0x2f, 0x6d, 0x61, 0x73, 0x74, + 0x65, 0x72, 0x2f, 0x73, 0x72, 0x63, 0x2f, 0x70, 0x72, 0x65, 0x61, 0x63, + 0x74, 0x2d, 0x70, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x2e, 0x6a, 0x73, 0x29, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2a, 0x2a, 0x20, 0x52, 0x65, 0x64, + 0x69, 0x72, 0x65, 0x63, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x69, 0x6e, 0x67, 0x20, 0x6f, 0x66, 0x20, 0x64, 0x65, 0x73, 0x63, 0x65, + 0x6e, 0x64, 0x61, 0x6e, 0x74, 0x73, 0x20, 0x69, 0x6e, 0x74, 0x6f, 0x20, + 0x74, 0x68, 0x65, 0x20, 0x67, 0x69, 0x76, 0x65, 0x6e, 0x20, 0x43, 0x53, + 0x53, 0x20, 0x73, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x20, 0x2a, + 0x2f, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, 0x73, 0x73, 0x20, + 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, + 0x64, 0x73, 0x20, 0x43, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, + 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x55, 0x70, 0x64, + 0x61, 0x74, 0x65, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x66, 0x6f, 0x72, + 0x20, 0x28, 0x6c, 0x65, 0x74, 0x20, 0x69, 0x20, 0x69, 0x6e, 0x20, 0x70, + 0x72, 0x6f, 0x70, 0x73, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x70, 0x72, + 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x5b, 0x69, 0x5d, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x73, + 0x65, 0x74, 0x54, 0x69, 0x6d, 0x65, 0x6f, 0x75, 0x74, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, + 0x65, 0x72, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, + 0x65, 0x6e, 0x74, 0x44, 0x69, 0x64, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x28, + 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, + 0x65, 0x64, 0x20, 0x3d, 0x20, 0x74, 0x72, 0x75, 0x65, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, + 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x20, + 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, + 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x2e, 0x62, 0x69, 0x6e, 0x64, 0x28, + 0x74, 0x68, 0x69, 0x73, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x57, + 0x69, 0x6c, 0x6c, 0x55, 0x6e, 0x6d, 0x6f, 0x75, 0x6e, 0x74, 0x28, 0x29, + 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x4c, 0x61, + 0x79, 0x65, 0x72, 0x28, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, 0x20, 0x3d, + 0x20, 0x66, 0x61, 0x6c, 0x73, 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, 0x73, + 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x26, 0x26, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, + 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x2e, 0x70, + 0x61, 0x72, 0x65, 0x6e, 0x74, 0x4e, 0x6f, 0x64, 0x65, 0x2e, 0x72, 0x65, + 0x6d, 0x6f, 0x76, 0x65, 0x43, 0x68, 0x69, 0x6c, 0x64, 0x28, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x66, 0x69, 0x6e, 0x64, 0x4e, 0x6f, 0x64, 0x65, 0x28, + 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, 0x74, + 0x79, 0x70, 0x65, 0x6f, 0x66, 0x20, 0x6e, 0x6f, 0x64, 0x65, 0x20, 0x3d, + 0x3d, 0x3d, 0x20, 0x27, 0x73, 0x74, 0x72, 0x69, 0x6e, 0x67, 0x27, 0x20, + 0x3f, 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, + 0x75, 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, + 0x28, 0x6e, 0x6f, 0x64, 0x65, 0x29, 0x20, 0x3a, 0x20, 0x6e, 0x6f, 0x64, + 0x65, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, + 0x4c, 0x61, 0x79, 0x65, 0x72, 0x28, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x3d, + 0x20, 0x74, 0x72, 0x75, 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x21, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x69, 0x73, 0x4d, 0x6f, 0x75, 0x6e, 0x74, 0x65, 0x64, + 0x29, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x3b, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x2f, 0x2f, 0x20, 0x63, 0x6c, + 0x65, 0x61, 0x6e, 0x20, 0x75, 0x70, 0x20, 0x6f, 0x6c, 0x64, 0x20, 0x6e, + 0x6f, 0x64, 0x65, 0x20, 0x69, 0x66, 0x20, 0x6d, 0x6f, 0x76, 0x69, 0x6e, + 0x67, 0x20, 0x62, 0x61, 0x73, 0x65, 0x73, 0x3a, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, 0x28, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, + 0x20, 0x21, 0x3d, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, + 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, 0x74, 0x65, 0x72, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, + 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x50, 0x6f, 0x69, 0x6e, + 0x74, 0x65, 0x72, 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, + 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x69, 0x66, 0x20, + 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x20, 0x26, + 0x26, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, 0x74, + 0x65, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, + 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, + 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x3c, 0x24, 0x7b, 0x50, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x2f, + 0x3e, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, + 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, 0x6d, 0x6f, + 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, + 0x20, 0x3d, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x66, 0x69, 0x6e, 0x64, + 0x4e, 0x6f, 0x64, 0x65, 0x28, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, + 0x6f, 0x70, 0x73, 0x2e, 0x69, 0x6e, 0x74, 0x6f, 0x29, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, + 0x65, 0x6d, 0x6f, 0x74, 0x65, 0x20, 0x3d, 0x20, 0x72, 0x65, 0x6e, 0x64, + 0x65, 0x72, 0x28, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x50, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x20, 0x63, + 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3d, 0x24, 0x7b, 0x74, 0x68, 0x69, + 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x7d, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x24, 0x7b, 0x73, 0x68, 0x6f, 0x77, 0x20, 0x26, 0x26, 0x20, 0x74, 0x68, + 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x2e, 0x63, 0x68, 0x69, + 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, + 0x6c, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x24, 0x7b, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, + 0x72, 0x6f, 0x78, 0x79, 0x7d, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x60, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x69, + 0x6e, 0x74, 0x6f, 0x2c, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x72, 0x65, + 0x6d, 0x6f, 0x74, 0x65, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x28, 0x29, 0x20, 0x7b, 0x0a, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, 0x75, 0x72, 0x6e, 0x20, + 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x2f, 0x2f, 0x20, 0x68, 0x69, 0x67, 0x68, 0x2d, 0x6f, 0x72, 0x64, 0x65, + 0x72, 0x20, 0x63, 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, + 0x74, 0x68, 0x61, 0x74, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x73, + 0x20, 0x69, 0x74, 0x73, 0x20, 0x66, 0x69, 0x72, 0x73, 0x74, 0x20, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x20, 0x69, 0x66, 0x20, 0x69, 0x74, 0x20, 0x65, + 0x78, 0x69, 0x73, 0x74, 0x73, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x2f, + 0x2f, 0x20, 0x75, 0x73, 0x65, 0x64, 0x20, 0x61, 0x73, 0x20, 0x61, 0x20, + 0x63, 0x6f, 0x6e, 0x64, 0x69, 0x74, 0x69, 0x6f, 0x6e, 0x61, 0x6c, 0x20, + 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x69, 0x6e, 0x67, 0x20, 0x70, 0x72, + 0x6f, 0x78, 0x79, 0x2e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x63, 0x6c, 0x61, + 0x73, 0x73, 0x20, 0x50, 0x6f, 0x72, 0x74, 0x61, 0x6c, 0x50, 0x72, 0x6f, + 0x78, 0x79, 0x20, 0x65, 0x78, 0x74, 0x65, 0x6e, 0x64, 0x73, 0x20, 0x43, + 0x6f, 0x6d, 0x70, 0x6f, 0x6e, 0x65, 0x6e, 0x74, 0x20, 0x7b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x67, 0x65, 0x74, 0x43, 0x68, 0x69, 0x6c, + 0x64, 0x43, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x28, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x74, 0x68, 0x69, 0x73, 0x2e, 0x70, 0x72, 0x6f, + 0x70, 0x73, 0x2e, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x78, 0x74, 0x3b, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x72, 0x65, 0x6e, 0x64, 0x65, 0x72, 0x28, 0x7b, 0x20, 0x63, + 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x20, 0x7d, 0x29, 0x20, 0x7b, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, 0x74, + 0x75, 0x72, 0x6e, 0x20, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, + 0x20, 0x7c, 0x7c, 0x20, 0x6e, 0x75, 0x6c, 0x6c, 0x3b, 0x0a, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x7d, 0x0a, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x66, 0x75, 0x6e, 0x63, 0x74, 0x69, 0x6f, + 0x6e, 0x20, 0x41, 0x70, 0x70, 0x28, 0x70, 0x72, 0x6f, 0x70, 0x73, 0x29, + 0x20, 0x7b, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x74, 0x75, 0x72, 0x6e, 0x20, 0x68, 0x74, 0x6d, 0x6c, 0x60, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, + 0x63, 0x6c, 0x61, 0x73, 0x73, 0x3d, 0x22, 0x6d, 0x6f, 0x64, 0x65, 0x2d, + 0x24, 0x7b, 0x73, 0x65, 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, + 0x6c, 0x75, 0x65, 0x2e, 0x74, 0x79, 0x70, 0x65, 0x7d, 0x22, 0x3e, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, + 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x68, 0x31, 0x3e, 0x6c, + 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, 0x2f, 0x68, 0x31, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x65, 0x72, 0x3e, 0x0a, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x6d, 0x61, + 0x69, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x65, + 0x6e, 0x74, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x63, 0x68, 0x61, 0x74, + 0x53, 0x74, 0x61, 0x72, 0x74, 0x65, 0x64, 0x2e, 0x76, 0x61, 0x6c, 0x75, + 0x65, 0x20, 0x3f, 0x20, 0x43, 0x68, 0x61, 0x74, 0x4c, 0x6f, 0x67, 0x20, + 0x3a, 0x20, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x46, 0x6f, 0x72, 0x6d, + 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x6d, 0x61, 0x69, 0x6e, 0x3e, 0x0a, 0x0a, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x73, + 0x65, 0x63, 0x74, 0x69, 0x6f, 0x6e, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x77, + 0x72, 0x69, 0x74, 0x65, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x24, 0x7b, 0x73, 0x65, + 0x73, 0x73, 0x69, 0x6f, 0x6e, 0x2e, 0x76, 0x61, 0x6c, 0x75, 0x65, 0x2e, + 0x74, 0x79, 0x70, 0x65, 0x20, 0x3d, 0x3d, 0x3d, 0x20, 0x27, 0x63, 0x68, + 0x61, 0x74, 0x27, 0x20, 0x3f, 0x20, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, + 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x3a, 0x20, 0x43, 0x6f, 0x6d, + 0x70, 0x6c, 0x65, 0x74, 0x69, 0x6f, 0x6e, 0x43, 0x6f, 0x6e, 0x74, 0x72, + 0x6f, 0x6c, 0x73, 0x7d, 0x20, 0x2f, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x65, 0x63, 0x74, + 0x69, 0x6f, 0x6e, 0x3e, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x3c, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, + 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x70, 0x3e, 0x3c, 0x24, 0x7b, 0x4d, 0x6f, 0x64, 0x65, 0x6c, + 0x47, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x49, 0x6e, + 0x66, 0x6f, 0x7d, 0x20, 0x2f, 0x3e, 0x3c, 0x2f, 0x70, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x70, 0x3e, 0x50, 0x6f, 0x77, 0x65, 0x72, 0x65, 0x64, 0x20, 0x62, 0x79, + 0x20, 0x3c, 0x61, 0x20, 0x68, 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, + 0x74, 0x70, 0x73, 0x3a, 0x2f, 0x2f, 0x67, 0x69, 0x74, 0x68, 0x75, 0x62, + 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x67, 0x67, 0x65, 0x72, 0x67, 0x61, 0x6e, + 0x6f, 0x76, 0x2f, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, + 0x22, 0x3e, 0x6c, 0x6c, 0x61, 0x6d, 0x61, 0x2e, 0x63, 0x70, 0x70, 0x3c, + 0x2f, 0x61, 0x3e, 0x20, 0x61, 0x6e, 0x64, 0x20, 0x3c, 0x61, 0x20, 0x68, + 0x72, 0x65, 0x66, 0x3d, 0x22, 0x68, 0x74, 0x74, 0x70, 0x73, 0x3a, 0x2f, + 0x2f, 0x67, 0x67, 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x22, 0x3e, 0x67, 0x67, + 0x6d, 0x6c, 0x2e, 0x61, 0x69, 0x3c, 0x2f, 0x61, 0x3e, 0x2e, 0x3c, 0x2f, + 0x70, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, + 0x20, 0x3c, 0x2f, 0x66, 0x6f, 0x6f, 0x74, 0x65, 0x72, 0x3e, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, + 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x20, 0x20, 0x60, 0x3b, 0x0a, 0x20, + 0x20, 0x20, 0x20, 0x7d, 0x0a, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x72, 0x65, + 0x6e, 0x64, 0x65, 0x72, 0x28, 0x68, 0x28, 0x41, 0x70, 0x70, 0x29, 0x2c, + 0x20, 0x64, 0x6f, 0x63, 0x75, 0x6d, 0x65, 0x6e, 0x74, 0x2e, 0x71, 0x75, + 0x65, 0x72, 0x79, 0x53, 0x65, 0x6c, 0x65, 0x63, 0x74, 0x6f, 0x72, 0x28, + 0x27, 0x23, 0x63, 0x6f, 0x6e, 0x74, 0x61, 0x69, 0x6e, 0x65, 0x72, 0x27, + 0x29, 0x29, 0x3b, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x73, 0x63, 0x72, 0x69, + 0x70, 0x74, 0x3e, 0x0a, 0x3c, 0x2f, 0x68, 0x65, 0x61, 0x64, 0x3e, 0x0a, + 0x0a, 0x3c, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x64, + 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x63, 0x6f, 0x6e, 0x74, 0x61, + 0x69, 0x6e, 0x65, 0x72, 0x22, 0x3e, 0x0a, 0x20, 0x20, 0x20, 0x20, 0x3c, + 0x69, 0x6e, 0x70, 0x75, 0x74, 0x20, 0x74, 0x79, 0x70, 0x65, 0x3d, 0x22, + 0x66, 0x69, 0x6c, 0x65, 0x22, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x66, 0x69, + 0x6c, 0x65, 0x49, 0x6e, 0x70, 0x75, 0x74, 0x22, 0x20, 0x61, 0x63, 0x63, + 0x65, 0x70, 0x74, 0x3d, 0x22, 0x69, 0x6d, 0x61, 0x67, 0x65, 0x2f, 0x2a, + 0x22, 0x20, 0x73, 0x74, 0x79, 0x6c, 0x65, 0x3d, 0x22, 0x64, 0x69, 0x73, + 0x70, 0x6c, 0x61, 0x79, 0x3a, 0x20, 0x6e, 0x6f, 0x6e, 0x65, 0x3b, 0x22, + 0x3e, 0x0a, 0x20, 0x20, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, 0x0a, 0x20, + 0x20, 0x3c, 0x64, 0x69, 0x76, 0x20, 0x69, 0x64, 0x3d, 0x22, 0x70, 0x6f, + 0x72, 0x74, 0x61, 0x6c, 0x22, 0x3e, 0x3c, 0x2f, 0x64, 0x69, 0x76, 0x3e, + 0x0a, 0x3c, 0x2f, 0x62, 0x6f, 0x64, 0x79, 0x3e, 0x0a, 0x0a, 0x3c, 0x2f, + 0x68, 0x74, 0x6d, 0x6c, 0x3e, 0x0a, 0x0a }; -unsigned int index_html_len = 32105; +unsigned int index_html_len = 33103; diff --git a/examples/server/public/completion.js b/examples/server/public/completion.js index 0c9bd5f10..b9c442509 100644 --- a/examples/server/public/completion.js +++ b/examples/server/public/completion.js @@ -94,6 +94,10 @@ export async function* llama(prompt, params = {}, config = {}) { break; } } + if (result.error) { + result.error = JSON.parse(result.error); + console.error(`llama.cpp error: ${result.error.content}`); + } } } } diff --git a/examples/server/public/index.html b/examples/server/public/index.html index 39d7bb93d..175c52478 100644 --- a/examples/server/public/index.html +++ b/examples/server/public/index.html @@ -160,6 +160,11 @@ height: 10em; } + [contenteditable] { + display: inline-block; + white-space: pre-wrap; + outline: 0px solid transparent; + } @keyframes loading-bg-wipe { 0% { @@ -219,6 +224,7 @@ repeat_penalty: 1.18, // 1.0 = disabled top_k: 40, // <= 0 to use vocab size top_p: 0.5, // 1.0 = disabled + min_p: 0.05, // 0 = disabled tfs_z: 1.0, // 1.0 = disabled typical_p: 1.0, // 1.0 = disabled presence_penalty: 0.0, // 0.0 = disabled @@ -461,18 +467,23 @@ }, "{{char}}"); } - const runCompletion = async () => { + const runCompletion = () => { if (controller.value) { console.log('already running...'); return; } const { prompt } = session.value; transcriptUpdate([...session.value.transcript, ["", prompt]]); - await runLlama(prompt, { + runLlama(prompt, { ...params.value, slot_id: slot_id, stop: [], - }, ""); + }, "").finally(() => { + session.value.prompt = session.value.transcript.map(([_, data]) => + Array.isArray(data) ? data.map(msg => msg.content).join('') : data + ).join(''); + session.value.transcript = []; + }) } const stop = (e) => { @@ -572,6 +583,7 @@ } }, [messages]) + const isCompletionMode = session.value.type === 'completion' const chatLine = ([user, data], index) => { let message const isArrayMessage = Array.isArray(data) @@ -581,20 +593,31 @@ const text = isArrayMessage ? data.map(msg => msg.content).join('').replace(/^\s+/, '') : data; - message = html`<${Markdownish} text=${template(text)} />` + message = isCompletionMode ? + text : + html`<${Markdownish} text=${template(text)} />` } if (user) { return html`

${template(user)}: ${message}

` } else { - return html`

${message}

` + return isCompletionMode ? + html`${message}` : + html`

${message}

` } }; + const handleCompletionEdit = (e) => { + session.value.prompt = e.target.innerText; + session.value.transcript = []; + } + return html` -
+
- ${messages.flatMap(chatLine)} -
`; + + ${messages.flatMap(chatLine)} + + `; }; const ConfigForm = (props) => { @@ -744,6 +767,7 @@ ${IntField({ label: "Consider N tokens for penalize", max: 2048, min: 0, name: "repeat_last_n", value: params.value.repeat_last_n })} ${IntField({ label: "Top-K sampling", max: 100, min: -1, name: "top_k", value: params.value.top_k })} ${FloatField({ label: "Top-P sampling", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })} + ${FloatField({ label: "Min-P sampling", max: 1.0, min: 0.0, name: "min_p", step: 0.01, value: params.value.min_p })}
More options diff --git a/examples/server/server.cpp b/examples/server/server.cpp index c163c7f8e..be23ad169 100644 --- a/examples/server/server.cpp +++ b/examples/server/server.cpp @@ -1,6 +1,5 @@ #include "common.h" #include "llama.h" -#include "build-info.h" #include "grammar-parser.h" #include "../llava/clip.h" @@ -149,6 +148,7 @@ struct task_server { task_type type; json data; bool infill_mode = false; + bool embedding_mode = false; }; struct task_result { @@ -371,6 +371,7 @@ struct llama_client_slot std::vector generated_token_probs; bool infill = false; + bool embedding = false; bool has_next_token = true; bool truncated = false; bool stopped_eos = false; @@ -500,6 +501,7 @@ struct llama_server_context bool multimodal = false; bool clean_kv_cache = true; bool all_slots_are_idle = false; + bool add_bos_token = true; int32_t id_gen; int32_t n_ctx; // total context for all clients / slots @@ -572,6 +574,8 @@ struct llama_server_context n_ctx = llama_n_ctx(ctx); + add_bos_token = llama_should_add_bos_token(model); + return true; } @@ -678,6 +682,7 @@ struct llama_server_context slot->params.n_predict = json_value(data, "n_predict", default_params.n_predict); slot->sparams.top_k = json_value(data, "top_k", default_sparams.top_k); slot->sparams.top_p = json_value(data, "top_p", default_sparams.top_p); + slot->sparams.min_p = json_value(data, "min_p", default_sparams.min_p); slot->sparams.tfs_z = json_value(data, "tfs_z", default_sparams.tfs_z); slot->sparams.typical_p = json_value(data, "typical_p", default_sparams.typical_p); slot->sparams.temp = json_value(data, "temperature", default_sparams.temp); @@ -862,7 +867,7 @@ struct llama_server_context } void update_system_prompt() { - system_tokens = ::llama_tokenize(ctx, system_prompt, true); + system_tokens = ::llama_tokenize(ctx, system_prompt, add_bos_token); llama_batch_clear(batch); @@ -1090,6 +1095,7 @@ struct llama_server_context std::lock_guard lock(mutex_results); task_result res; res.id = id; + res.stop = false; res.error = true; res.result_json = { { "content", error } }; queue_results.push_back(res); @@ -1112,6 +1118,7 @@ struct llama_server_context {"temp", slot.sparams.temp}, {"top_k", slot.sparams.top_k}, {"top_p", slot.sparams.top_p}, + {"min_p", slot.sparams.min_p}, {"tfs_z", slot.sparams.tfs_z}, {"typical_p", slot.sparams.typical_p}, {"repeat_last_n", slot.sparams.penalty_last_n}, @@ -1244,13 +1251,15 @@ struct llama_server_context queue_results.push_back(res); } - int request_completion(json data, bool infill) + int request_completion(json data, bool infill, bool embedding) { std::lock_guard lock(mutex_tasks); task_server task; task.id = id_gen++; + task.target_id = 0; task.data = data; task.infill_mode = infill; + task.embedding_mode = embedding; task.type = COMPLETION_TASK; queue_tasks.push_back(task); return task.id; @@ -1376,7 +1385,7 @@ struct llama_server_context { LOG_TEE("slot unavailable\n"); // send error result - send_error(task.id, "slot unavaliable"); + send_error(task.id, "slot unavailable"); return; } @@ -1388,6 +1397,7 @@ struct llama_server_context slot->reset(); slot->infill = task.infill_mode; + slot->embedding = task.embedding_mode; slot->task_id = task.id; if (!launch_slot_with_data(slot, task.data)) @@ -1547,11 +1557,40 @@ struct llama_server_context } else { - prompt_tokens = tokenize(slot.prompt, system_prompt.empty()); // add BOS if there isn't system prompt + prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token); // add BOS if there isn't system prompt } slot.num_prompt_tokens = prompt_tokens.size(); + if (slot.params.n_keep < 0) + { + slot.params.n_keep = slot.num_prompt_tokens; + } + slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); + + // if input prompt is too big, truncate it + if (slot.num_prompt_tokens >= slot.n_ctx) + { + const int n_left = slot.n_ctx - slot.params.n_keep; + const int n_block_size = n_left / 2; + const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + + std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); + new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); + + LOG_VERBOSE("input truncated", { + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, + }); + slot.truncated = true; + prompt_tokens = new_tokens; + + slot.num_prompt_tokens = prompt_tokens.size(); + GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx); + } + if (!slot.params.cache_prompt) { llama_sampling_reset(slot.ctx_sampling); @@ -1561,35 +1600,6 @@ struct llama_server_context } else { - if (slot.params.n_keep < 0) - { - slot.params.n_keep = slot.num_prompt_tokens; - } - slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); - - // if input prompt is too big, truncate it - if (slot.num_prompt_tokens >= slot.n_ctx) - { - const int n_left = slot.n_ctx - slot.params.n_keep; - const int n_block_size = n_left / 2; - const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; - - std::vector new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); - new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end()); - - LOG_VERBOSE("input truncated", { - {"n_ctx", slot.n_ctx}, - {"n_keep", slot.params.n_keep}, - {"n_left", n_left}, - {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, - }); - slot.truncated = true; - prompt_tokens = new_tokens; - - slot.num_prompt_tokens = prompt_tokens.size(); - GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx); - } - // push the prompt into the sampling context (do not apply grammar) for (auto &token : prompt_tokens) { @@ -1624,7 +1634,7 @@ struct llama_server_context const bool has_images = process_images(slot); // process the prefix of first image - std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, true) : prompt_tokens; + std::vector prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens; for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past) { llama_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot.n_past, { slot.id }, false); @@ -1695,7 +1705,7 @@ struct llama_server_context } // prompt evaluated for embedding - if (params.embedding) + if (slot.embedding) { send_embedding(slot); slot.release(); @@ -1751,12 +1761,18 @@ static void server_print_usage(const char *argv0, const gpt_params ¶ms, printf("options:\n"); printf(" -h, --help show this help message and exit\n"); printf(" -v, --verbose verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled"); - printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); + printf(" -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads); printf(" -tb N, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads)\n"); - printf(" -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx); + printf(" -c N, --ctx-size N size of the prompt context (default: %d)\n", params.n_ctx); + printf(" --rope-scaling {none,linear,yarn}\n"); + printf(" RoPE frequency scaling method, defaults to linear unless specified by the model\n"); printf(" --rope-freq-base N RoPE base frequency (default: loaded from model)\n"); - printf(" --rope-freq-scale N RoPE frequency scaling factor (default: loaded from model)\n"); - printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); + printf(" --rope-freq-scale N RoPE frequency scaling factor, expands context by a factor of 1/N\n"); + printf(" --yarn-ext-factor N YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n"); + printf(" --yarn-attn-factor N YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n"); + printf(" --yarn-beta-slow N YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow); + printf(" --yarn-beta-fast N YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast); + printf(" -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch); printf(" --memory-f32 use f32 instead of f16 for memory key+value (default: disabled)\n"); printf(" not recommended: doubles context memory required and no measurable increase in quality\n"); if (llama_mlock_supported()) @@ -1877,6 +1893,19 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } params.n_ctx = std::stoi(argv[i]); } + else if (arg == "--rope-scaling") + { + if (++i >= argc) + { + invalid_param = true; + break; + } + std::string value(argv[i]); + /**/ if (value == "none") { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; } + else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; } + else if (value == "yarn") { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; } + else { invalid_param = true; break; } + } else if (arg == "--rope-freq-base") { if (++i >= argc) @@ -1895,6 +1924,38 @@ static void server_params_parse(int argc, char **argv, server_params &sparams, } params.rope_freq_scale = std::stof(argv[i]); } + else if (arg == "--yarn-ext-factor") + { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_ext_factor = std::stof(argv[i]); + } + else if (arg == "--yarn-attn-factor") + { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_attn_factor = std::stof(argv[i]); + } + else if (arg == "--yarn-beta-fast") + { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_beta_fast = std::stof(argv[i]); + } + else if (arg == "--yarn-beta-slow") + { + if (++i >= argc) { + invalid_param = true; + break; + } + params.yarn_beta_slow = std::stof(argv[i]); + } else if (arg == "--memory-f32" || arg == "--memory_f32") { params.memory_f16 = false; @@ -2209,8 +2270,8 @@ int main(int argc, char **argv) llama_backend_init(params.numa); - LOG_INFO("build info", {{"build", BUILD_NUMBER}, - {"commit", BUILD_COMMIT}}); + LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER}, + {"commit", LLAMA_COMMIT}}); LOG_INFO("system info", { {"n_threads", params.n_threads}, @@ -2274,7 +2335,7 @@ int main(int argc, char **argv) svr.Post("/completion", [&llama](const httplib::Request &req, httplib::Response &res) { json data = json::parse(req.body); - const int task_id = llama.request_completion(data, false); + const int task_id = llama.request_completion(data, false, false); if (!json_value(data, "stream", false)) { std::string completion_text; task_result result = llama.next_result(task_id); @@ -2309,6 +2370,17 @@ int main(int argc, char **argv) break; } } else { + const std::string str = + "error: " + + result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) + + "\n\n"; + LOG_VERBOSE("data stream", { + { "to_send", str } + }); + if (!sink.write(str.c_str(), str.size())) + { + return false; + } break; } } @@ -2329,7 +2401,7 @@ int main(int argc, char **argv) svr.Post("/infill", [&llama](const httplib::Request &req, httplib::Response &res) { json data = json::parse(req.body); - const int task_id = llama.request_completion(data, true); + const int task_id = llama.request_completion(data, true, false); if (!json_value(data, "stream", false)) { std::string completion_text; task_result result = llama.next_result(task_id); @@ -2433,7 +2505,7 @@ int main(int argc, char **argv) { prompt = ""; } - const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false); + const int task_id = llama.request_completion({ {"prompt", prompt}, { "n_predict", 0} }, false, true); task_result result = llama.next_result(task_id); return res.set_content(result.result_json.dump(), "application/json"); }); diff --git a/examples/speculative/CMakeLists.txt b/examples/speculative/CMakeLists.txt index 6c5c9456e..810f3c46a 100644 --- a/examples/speculative/CMakeLists.txt +++ b/examples/speculative/CMakeLists.txt @@ -3,6 +3,3 @@ add_executable(${TARGET} speculative.cpp) install(TARGETS ${TARGET} RUNTIME) target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) target_compile_features(${TARGET} PRIVATE cxx_std_11) -if(TARGET BUILD_INFO) - add_dependencies(${TARGET} BUILD_INFO) -endif() diff --git a/examples/speculative/speculative.cpp b/examples/speculative/speculative.cpp index 323c74652..ace755c51 100644 --- a/examples/speculative/speculative.cpp +++ b/examples/speculative/speculative.cpp @@ -1,5 +1,3 @@ -#include "build-info.h" - #include "common.h" #include "llama.h" @@ -39,9 +37,11 @@ int main(int argc, char ** argv) { // max number of parallel drafting sequences (i.e. tree branches) const int n_seq_dft = params.n_parallel; - // TODO: make this configurable - const float p_accept = 0.80f; - const float p_split = 0.10f; + // probability threshold for accepting a token from the draft model + const float p_accept = params.p_accept; + + // probability threshold for splitting a draft branch (only for n_seq_dft > 1) + const float p_split = params.p_split; #ifndef LOG_DISABLE_LOGS log_set_target(log_filename_generator("speculative", "log")); @@ -94,9 +94,22 @@ int main(int argc, char ** argv) { } } - // tokenize the prompt + + // Tokenize the prompt + const bool add_bos_tgt = llama_should_add_bos_token(model_tgt); + LOG("add_bos tgt: %d\n", add_bos_tgt); + + const bool add_bos_dft = llama_should_add_bos_token(model_dft); + LOG("add_bos dft: %d\n", add_bos_dft); + + if (add_bos_tgt != add_bos_dft) { + fprintf(stderr, "%s: error: draft model add_bos must match target model to use speculation but ", __func__); + fprintf(stderr, "add_bos_dft = %d while add_bos_tgt = %d\n", add_bos_dft, add_bos_tgt); + return 1; + } + std::vector inp; - inp = ::llama_tokenize(ctx_tgt, params.prompt, true); + inp = ::llama_tokenize(ctx_tgt, params.prompt, add_bos_tgt, true); const int max_context_size = llama_n_ctx(ctx_tgt); const int max_tokens_list_size = max_context_size - 4; diff --git a/examples/tokenize/CMakeLists.txt b/examples/tokenize/CMakeLists.txt new file mode 100644 index 000000000..5e6654d7e --- /dev/null +++ b/examples/tokenize/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET tokenize) +add_executable(${TARGET} tokenize.cpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/tokenize/tokenize.cpp b/examples/tokenize/tokenize.cpp new file mode 100644 index 000000000..4ff8e3fa7 --- /dev/null +++ b/examples/tokenize/tokenize.cpp @@ -0,0 +1,44 @@ +#include "common.h" +#include "llama.h" + +#include +#include +#include +#include + +int main(int argc, char ** argv) { + if (argc < 3 || argv[1][0] == '-') { + printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]); + return 1; + } + + const char * model_path = argv[1]; + const char * prompt = argv[2]; + + const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids"; + + llama_backend_init(false); + + llama_model_params model_params = llama_model_default_params(); + model_params.vocab_only = true; + llama_model * model = llama_load_model_from_file(model_path, model_params); + + llama_context_params ctx_params = llama_context_default_params(); + llama_context * ctx = llama_new_context_with_model(model, ctx_params); + + const bool add_bos = llama_should_add_bos_token(model); + + std::vector tokens; + + tokens = ::llama_tokenize(model, prompt, add_bos, true); + + for (int i = 0; i < (int) tokens.size(); i++) { + if (printing_ids) { + printf("%d\n", tokens[i]); + } else { + printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str()); + } + } + + return 0; +} diff --git a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py index 887ed2e21..ed93673bc 100644 --- a/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py +++ b/examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py @@ -9,7 +9,7 @@ import numpy as np from pathlib import Path if 'NO_LOCAL_GGUF' not in os.environ: - sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py' / 'gguf')) + sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py')) import gguf # gguf constants diff --git a/examples/train-text-from-scratch/train-text-from-scratch.cpp b/examples/train-text-from-scratch/train-text-from-scratch.cpp index 1ce6cef29..f049a3923 100644 --- a/examples/train-text-from-scratch/train-text-from-scratch.cpp +++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp @@ -349,9 +349,9 @@ static struct ggml_tensor * llama_build_train_graphs( // not capturing these, to silcence warnings const int rope_mode = 0; - return ggml_rope_custom(ctx, - t, KQ_pos, n_rot, rope_mode, n_ctx, - rope_freq_base, rope_freq_scale); + return ggml_rope_custom( + ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f + ); }; set_name(tokens_input, "tokens_input"); @@ -436,7 +436,7 @@ static struct ggml_tensor * llama_build_train_graphs( if (enable_checkpointing) { ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size()); } else { - *gb = *gf; + ggml_graph_cpy(gf, gb); ggml_build_backward_expand(ctx, gf, gb, true); } @@ -1006,6 +1006,7 @@ int main(int argc, char ** argv) { opt->params = ggml_opt_default_params(GGML_OPT_ADAM); opt->params.print_forward_graph = false; opt->params.print_backward_graph = false; + opt->params.graph_size = LLAMA_TRAIN_MAX_NODES; opt->params.n_threads = params.common.n_threads; opt->params.past = params.common.opt_past; opt->params.delta = params.common.opt_delta; @@ -1108,11 +1109,9 @@ int main(int argc, char ** argv) { ggml_allocr_free(alloc); // context for compute tensors without their data - size_t estimated_compute_size_wo_data = ( - ggml_tensor_overhead()*GGML_MAX_NODES*2 - + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*( - params.common.use_checkpointing ? 3 : 2 - ) + const size_t estimated_compute_size_wo_data = ( + 2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() + + (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true)) ); struct ggml_init_params ctx_compute_params = { estimated_compute_size_wo_data, // mem_size @@ -1135,11 +1134,11 @@ int main(int argc, char ** argv) { for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) { ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new_measure(tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = (enum ggml_cgraph_eval_order) order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_train_graphs( &model, alloc, ctx_compute, @@ -1168,11 +1167,11 @@ int main(int argc, char ** argv) { mem_compute_data.resize(max_compute_size); ctx_compute = ggml_init(ctx_compute_params); alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment); - gf = ggml_new_graph(ctx_compute); + gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gf->order = best_order; - gb = ggml_new_graph(ctx_compute); + gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true); gb_tmp = params.common.use_checkpointing - ? ggml_new_graph(ctx_compute) + ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true) : NULL; loss = llama_build_train_graphs( &model, alloc, ctx_compute, diff --git a/flake.lock b/flake.lock index 070f0e161..0455f6561 100644 --- a/flake.lock +++ b/flake.lock @@ -5,11 +5,11 @@ "systems": "systems" }, "locked": { - "lastModified": 1692799911, - "narHash": "sha256-3eihraek4qL744EvQXsK1Ha6C3CR7nnT8X2qWap4RNk=", + "lastModified": 1694529238, + "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=", "owner": "numtide", "repo": "flake-utils", - "rev": "f9e7cf818399d17d347f847525c5a5a8032e4e44", + "rev": "ff7b65b44d01cf9ba6a71320833626af21126384", "type": "github" }, "original": { @@ -20,11 +20,11 @@ }, "nixpkgs": { "locked": { - "lastModified": 1698134075, - "narHash": "sha256-foCD+nuKzfh49bIoiCBur4+Fx1nozo+4C/6k8BYk4sg=", + "lastModified": 1698318101, + "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=", "owner": "NixOS", "repo": "nixpkgs", - "rev": "8efd5d1e283604f75a808a20e6cde0ef313d07d4", + "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c", "type": "github" }, "original": { diff --git a/flake.nix b/flake.nix index fa34394b2..4cf28d5c1 100644 --- a/flake.nix +++ b/flake.nix @@ -11,8 +11,7 @@ meta.mainProgram = "llama"; inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin; buildInputs = with pkgs; [ openmpi ]; - osSpecific = with pkgs; buildInputs ++ - ( + osSpecific = with pkgs; buildInputs ++ ( if isAarch64 && isDarwin then with pkgs.darwin.apple_sdk_11_0.frameworks; [ Accelerate @@ -96,12 +95,15 @@ }; packages.rocm = pkgs.stdenv.mkDerivation { inherit name src meta postPatch nativeBuildInputs postInstall; - buildInputs = with pkgs; buildInputs ++ [ hip hipblas rocblas ]; + buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ]; cmakeFlags = cmakeFlags ++ [ "-DLLAMA_HIPBLAS=1" "-DCMAKE_C_COMPILER=hipcc" "-DCMAKE_CXX_COMPILER=hipcc" - "-DCMAKE_POSITION_INDEPENDENT_CODE=ON" + # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM + # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt + # and select the line that matches the current nixpkgs version of rocBLAS. + "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102" ]; }; apps.llama-server = { diff --git a/ggml-alloc.c b/ggml-alloc.c index 34eba3f83..cdfe4caf6 100644 --- a/ggml-alloc.c +++ b/ggml-alloc.c @@ -1,51 +1,21 @@ #include "ggml-alloc.h" -#include "ggml-backend.h" +#include "ggml-backend-impl.h" #include "ggml.h" +#include "ggml-impl.h" #include +#include #include #include #include #include - -#define UNUSED(x) (void)(x) #define MAX(a, b) ((a) > (b) ? (a) : (b)) -#define GGML_MAX_CONCUR (2*GGML_MAX_NODES) +#define MAX_FREE_BLOCKS 256 //#define GGML_ALLOCATOR_DEBUG -//#define AT_PRINTF printf -#define AT_PRINTF(...) ((void)0) - -struct hash_node { - struct ggml_tensor * t; - int n_children; - int n_views; -}; - -static size_t hash(void * p) { - return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; -} - -static struct hash_node * hash_get(struct hash_node hash_table[], struct ggml_tensor * t) { - size_t h = hash(t); - - // linear probing - size_t i = h; - while (hash_table[i].t != NULL) { - if (hash_table[i].t == t) { - return &hash_table[i]; - } - i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; - if (i == h) { - // hash table is full - GGML_ASSERT(false); - } - } - - hash_table[i].t = t; - return &hash_table[i]; -} +//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__) +#define AT_PRINTF(...) // TODO: GGML_PAD ? static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) { @@ -59,20 +29,18 @@ struct free_block { size_t size; }; -#define MAX_FREE_BLOCKS 256 - -struct ggml_allocr { +struct ggml_tallocr { struct ggml_backend_buffer * buffer; bool buffer_owned; - void * data; + void * base; size_t alignment; + int n_free_blocks; struct free_block free_blocks[MAX_FREE_BLOCKS]; - struct hash_node hash_table[GGML_GRAPH_HASHTABLE_SIZE]; + size_t max_size; + bool measure; - int parse_seq[GGML_MAX_CONCUR]; - int parse_seq_len; #ifdef GGML_ALLOCATOR_DEBUG struct ggml_tensor * allocated_tensors[1024]; @@ -80,7 +48,7 @@ struct ggml_allocr { }; #ifdef GGML_ALLOCATOR_DEBUG -static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static void add_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { for (int i = 0; i < 1024; i++) { if (alloc->allocated_tensors[i] == NULL) { alloc->allocated_tensors[i] = tensor; @@ -89,7 +57,7 @@ static void add_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor } GGML_ASSERT(!"out of allocated_tensors"); } -static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { for (int i = 0; i < 1024; i++) { if (alloc->allocated_tensors[i] == tensor || (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) { @@ -103,7 +71,7 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens #endif // check if a tensor is allocated by this buffer -static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) { +static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) { return tensor->buffer == alloc->buffer; } @@ -111,7 +79,7 @@ static bool ggml_is_view(struct ggml_tensor * t) { return t->view_src != NULL; } -void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { +void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated @@ -162,9 +130,10 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) } tensor->data = addr; - AT_PRINTF("%s: allocated data at %p\n", __func__, tensor->data); tensor->buffer = alloc->buffer; - ggml_backend_buffer_init_tensor(alloc->buffer, tensor); + if (!alloc->measure) { + ggml_backend_buffer_init_tensor(alloc->buffer, tensor); + } #ifdef GGML_ALLOCATOR_DEBUG add_allocated_tensor(alloc, tensor); @@ -180,16 +149,16 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) } #endif - alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size); + alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->base + size); } // this is a very naive implementation, but for our case the number of free blocks should be very small -static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) { - if (ggml_allocr_is_own(alloc, tensor) == false) { +static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) { + if (ggml_tallocr_is_own(alloc, tensor) == false) { // the tensor was not allocated in this buffer // this can happen because the graph allocator will try to free weights and other tensors from different buffers // the easiest way to deal with this is just to ignore it - AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer); + // AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer); return; } @@ -199,7 +168,9 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens size = aligned_offset(NULL, size, alloc->alignment); AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks); - ggml_backend_buffer_free_tensor(alloc->buffer, tensor); + if (!alloc->measure) { + ggml_backend_buffer_free_tensor(alloc->buffer, tensor); + } #ifdef GGML_ALLOCATOR_DEBUG remove_allocated_tensor(alloc, tensor); @@ -253,91 +224,180 @@ static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tens alloc->n_free_blocks++; } -void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n) { - for (int i = 0; i < n; i++) { - alloc->parse_seq[i] = list[i]; - } - alloc->parse_seq_len = n; -} - -void ggml_allocr_reset(struct ggml_allocr * alloc) { +void ggml_tallocr_reset(ggml_tallocr_t alloc) { alloc->n_free_blocks = 1; - size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment); - alloc->free_blocks[0].addr = (char *)alloc->data + align_offset; - alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset; + size_t align_offset = aligned_offset(alloc->base, 0, alloc->alignment); + alloc->free_blocks[0].addr = (char *)alloc->base + align_offset; + + if (alloc->measure) { + alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows + } else { + alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset; + } } -struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) { +ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) { struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(NULL, data, size); - struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr)); + ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr)); - *alloc = (struct ggml_allocr){ + *alloc = (struct ggml_tallocr) { /*.buffer = */ buffer, /*.buffer_owned = */ true, /*.base = */ ggml_backend_buffer_get_base(buffer), /*.alignment = */ alignment, /*.n_free_blocks = */ 0, /*.free_blocks = */ {{0}}, - /*.hash_table = */ {{0}}, /*.max_size = */ 0, /*.measure = */ false, - /*.parse_seq = */ {0}, - /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG /*.allocated_tensors = */ {0}, #endif }; - ggml_allocr_reset(alloc); + ggml_tallocr_reset(alloc); return alloc; } -struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) { - struct ggml_allocr * alloc = ggml_allocr_new((void *)0x1000, (size_t)-0x1001, alignment); +ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) { + ggml_tallocr_t alloc = ggml_tallocr_new((void *)0x1000, SIZE_MAX/2, alignment); alloc->measure = true; return alloc; } -struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) { - struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr)); +ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) { + // create a backend buffer to get the correct tensor allocation sizes + ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1); - *alloc = (struct ggml_allocr){ + // TODO: move alloc initialization to a common ggml_tallocr_new_impl function + ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer); + alloc->buffer_owned = true; + alloc->measure = true; + ggml_tallocr_reset(alloc); + return alloc; +} + +ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) { + ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size); + ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer); + alloc->buffer_owned = true; + return alloc; +} + +ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) { + ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr)); + + *alloc = (struct ggml_tallocr) { /*.buffer = */ buffer, /*.buffer_owned = */ false, /*.base = */ ggml_backend_buffer_get_base(buffer), /*.alignment = */ ggml_backend_buffer_get_alignment(buffer), /*.n_free_blocks = */ 0, /*.free_blocks = */ {{0}}, - /*.hash_table = */ {{0}}, /*.max_size = */ 0, /*.measure = */ false, - /*.parse_seq = */ {0}, - /*.parse_seq_len = */ 0, #ifdef GGML_ALLOCATOR_DEBUG /*.allocated_tensors = */ {0}, #endif }; - ggml_allocr_reset(alloc); + ggml_tallocr_reset(alloc); return alloc; } -void ggml_allocr_free(struct ggml_allocr * alloc) { +struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t alloc) { + return alloc->buffer; +} + +void ggml_tallocr_free(ggml_tallocr_t alloc) { + if (alloc == NULL) { + return; + } + if (alloc->buffer_owned) { ggml_backend_buffer_free(alloc->buffer); } free(alloc); } -bool ggml_allocr_is_measure(struct ggml_allocr * alloc) { +bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) { return alloc->measure; } -//////////// compute graph allocator +size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) { + return alloc->max_size; +} + +// graph allocator + +struct hash_node { + int n_children; + int n_views; +}; + +struct ggml_gallocr { + ggml_tallocr_t talloc; + struct ggml_hash_set hash_set; + struct hash_node * hash_values; + size_t hash_values_size; + ggml_tallocr_t * hash_allocs; + int * parse_seq; + int parse_seq_len; +}; + +ggml_gallocr_t ggml_gallocr_new(void) { + ggml_gallocr_t galloc = (ggml_gallocr_t)malloc(sizeof(struct ggml_gallocr)); + + *galloc = (struct ggml_gallocr) { + /*.talloc = */ NULL, + /*.hash_set = */ {0}, + /*.hash_values = */ NULL, + /*.hash_values_size = */ 0, + /*.hash_allocs = */ NULL, + /*.parse_seq = */ NULL, + /*.parse_seq_len = */ 0, + }; + + return galloc; +} + +void ggml_gallocr_free(ggml_gallocr_t galloc) { + if (galloc == NULL) { + return; + } + + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + if (galloc->hash_values != NULL) { + free(galloc->hash_values); + } + if (galloc->hash_allocs != NULL) { + free(galloc->hash_allocs); + } + if (galloc->parse_seq != NULL) { + free(galloc->parse_seq); + } + free(galloc); +} + +void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n) { + free(galloc->parse_seq); + galloc->parse_seq = malloc(sizeof(int) * n); + + for (int i = 0; i < n; i++) { + galloc->parse_seq[i] = list[i]; + } + galloc->parse_seq_len = n; +} + +static struct hash_node * hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) { + size_t i = ggml_hash_find_or_insert(galloc->hash_set, t); + return &galloc->hash_values[i]; +} static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) { if (a->type != b->type) { @@ -378,23 +438,40 @@ static bool ggml_op_can_inplace(enum ggml_op op) { } } -static void init_view(struct ggml_allocr * alloc, struct ggml_tensor * view) { - assert(view->view_src != NULL && view->view_src->data != NULL); - view->backend = view->view_src->backend; +static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * node) { + if (galloc->talloc != NULL) { + return galloc->talloc; + } + + return galloc->hash_allocs[ggml_hash_find_or_insert(galloc->hash_set, node)]; +} + +static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) { + ggml_tallocr_t alloc = node_tallocr(galloc, view); + + //printf("init_view: %s from src %s\n", view->name, view->view_src->name); + GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL); + if (update_backend) { + view->backend = view->view_src->backend; + } view->buffer = view->view_src->buffer; view->data = (char *)view->view_src->data + view->view_offs; // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras - assert(ggml_allocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend); - ggml_backend_buffer_init_tensor(alloc->buffer, view); + assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->backend == alloc->buffer->backend); + + if (!alloc->measure) { + ggml_backend_buffer_init_tensor(alloc->buffer, view); + } } -static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) { - struct hash_node * ht = alloc->hash_table; +static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) { + ggml_tallocr_t alloc = node_tallocr(galloc, node); + if (node->data == NULL) { if (ggml_is_view(node)) { - init_view(alloc, node); + init_view(galloc, node, true); } else { // see if we can reuse a parent's buffer (inplace) if (ggml_op_can_inplace(node->op)) { @@ -405,16 +482,16 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) } // if the node's data is external, then we cannot re-use it - if (ggml_allocr_is_own(alloc, parent) == false) { + if (ggml_tallocr_is_own(alloc, parent) == false) { AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data); continue; } - struct hash_node * p_hn = hash_get(ht, parent); + struct hash_node * p_hn = hash_get(galloc, parent); if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) { if (ggml_is_view(parent)) { struct ggml_tensor * view_src = parent->view_src; - struct hash_node * view_src_hn = hash_get(ht, view_src); + struct hash_node * view_src_hn = hash_get(galloc, view_src); if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) { // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite // the parent's data that it will need later (same layout requirement). the problem is that then @@ -424,171 +501,267 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node) AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name); node->view_src = view_src; view_src_hn->n_views += 1; - init_view(alloc, node); + init_view(galloc, node, false); return; } - } - else { + } else { AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name); node->view_src = parent; p_hn->n_views += 1; - init_view(alloc, node); + init_view(galloc, node, false); return; } } } } - ggml_allocr_alloc(alloc, node); + ggml_tallocr_alloc(alloc, node); } } } -size_t ggml_allocr_alloc_graph_n( - struct ggml_allocr * alloc, - struct ggml_cgraph ** graphs, int n_graphs, - struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) { +static void free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) { + ggml_tallocr_t alloc = node_tallocr(galloc, node); - // reset hash table - struct hash_node * ht = alloc->hash_table; - memset(ht, 0, sizeof(struct hash_node) * GGML_GRAPH_HASHTABLE_SIZE); + ggml_tallocr_free_tensor(alloc, node); +} + +static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * gf) { + const int * parse_seq = galloc->parse_seq; + int parse_seq_len = galloc->parse_seq_len; // count number of children and views - for (int g = 0; g < n_graphs; g++) { - struct ggml_cgraph * gf = graphs[g]; - for (int i = 0; i < gf->n_nodes; i++) { + for (int i = 0; i < gf->n_nodes; i++) { + struct ggml_tensor * node = gf->nodes[i]; + + if (ggml_is_view(node)) { + struct ggml_tensor * view_src = node->view_src; + hash_get(galloc, view_src)->n_views += 1; + if (node->buffer == NULL && node->data != NULL) { + // view of a pre-allocated tensor, didn't call init_view() yet + init_view(galloc, node, true); + } + } + + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + hash_get(galloc, parent)->n_children += 1; + if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) { + init_view(galloc, parent, true); + } + } + } + + // allocate tensors + // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers + int last_barrier_pos = 0; + int n_nodes = parse_seq_len ? parse_seq_len : gf->n_nodes; + + for (int ind = 0; ind < n_nodes; ind++) { + // allocate a node if there is no parse_seq or this is not a barrier + if (parse_seq_len == 0 || parse_seq[ind] != -1) { + int i = parse_seq_len ? parse_seq[ind] : ind; struct ggml_tensor * node = gf->nodes[i]; - if (ggml_is_view(node)) { - struct ggml_tensor * view_src = node->view_src; - hash_get(ht, view_src)->n_views += 1; - if (node->buffer == NULL && node->data != NULL) { - // view of a pre-allocated tensor, didn't call init_view() yet - init_view(alloc, node); - } - } - + // allocate parents (leafs) for (int j = 0; j < GGML_MAX_SRC; j++) { struct ggml_tensor * parent = node->src[j]; if (parent == NULL) { break; } - hash_get(ht, parent)->n_children += 1; - if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) { - init_view(alloc, parent); + allocate_node(galloc, parent); + } + + // allocate node + allocate_node(galloc, node); + + AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * parent = node->src[j]; + if (parent == NULL) { + break; + } + AT_PRINTF("%s", parent->name); + if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { + AT_PRINTF(", "); } } + AT_PRINTF("\n"); } - } - // allocate tensors - for (int g = 0; g < n_graphs; g++) { - struct ggml_cgraph * gf = graphs[g]; - AT_PRINTF("####### graph %d/%d\n", g, n_graphs); - // graph inputs are allocated first to ensure that they are not overwritten by each other - if (inputs != NULL && inputs[g] != NULL) { - for (int i = 0; inputs[g][i] != NULL; i++) { - struct ggml_tensor * input = inputs[g][i]; - AT_PRINTF("input: %s\n", input->name); - allocate_node(alloc, input); - } - } - // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers - int last_barrier_pos = 0; - int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes; + // update parents + // update immediately if there is no parse_seq + // update only at barriers if there is parse_seq + if ((parse_seq_len == 0) || parse_seq[ind] == -1) { + int update_start = parse_seq_len ? last_barrier_pos : ind; + int update_end = parse_seq_len ? ind : ind + 1; + for (int i = update_start; i < update_end; i++) { + int node_i = parse_seq_len ? parse_seq[i] : i; + struct ggml_tensor * node = gf->nodes[node_i]; - for (int ind = 0; ind < n_nodes; ind++) { - // allocate a node if there is no parse_seq or this is not a barrier - if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) { - int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind; - struct ggml_tensor * node = gf->nodes[i]; - - // allocate parents (leafs) for (int j = 0; j < GGML_MAX_SRC; j++) { struct ggml_tensor * parent = node->src[j]; if (parent == NULL) { break; } - allocate_node(alloc, parent); - } + struct hash_node * p_hn = hash_get(galloc, parent); + p_hn->n_children -= 1; - // allocate node - allocate_node(alloc, node); + //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); - AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name); - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; - } - AT_PRINTF("%s", parent->name); - if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) { - AT_PRINTF(", "); - } - } - AT_PRINTF("\n"); - } - - // update parents - // update immediately if there is no parse_seq - // update only at barriers if there is parse_seq - if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) { - int update_start = alloc->parse_seq_len ? last_barrier_pos : ind; - int update_end = alloc->parse_seq_len ? ind : ind + 1; - for (int i = update_start; i < update_end; i++) { - int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i; - struct ggml_tensor * node = gf->nodes[node_i]; - - for (int j = 0; j < GGML_MAX_SRC; j++) { - struct ggml_tensor * parent = node->src[j]; - if (parent == NULL) { - break; + if (p_hn->n_children == 0 && p_hn->n_views == 0) { + if (ggml_is_view(parent)) { + struct ggml_tensor * view_src = parent->view_src; + struct hash_node * view_src_hn = hash_get(galloc, view_src); + view_src_hn->n_views -= 1; + AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); + if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0) { + free_node(galloc, view_src); + } } - struct hash_node * p_hn = hash_get(ht, parent); - p_hn->n_children -= 1; - - //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views); - - if (p_hn->n_children == 0 && p_hn->n_views == 0) { - if (ggml_is_view(parent)) { - struct ggml_tensor * view_src = parent->view_src; - struct hash_node * view_src_hn = hash_get(ht, view_src); - view_src_hn->n_views -= 1; - AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views); - if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) { - ggml_allocr_free_tensor(alloc, view_src); - } - } - else { - if (parent->data != node->data) { - ggml_allocr_free_tensor(alloc, parent); - } - } + else { + free_node(galloc, parent); } } } - AT_PRINTF("\n"); - if (alloc->parse_seq_len) { - last_barrier_pos = ind + 1; - } } - } - // free graph outputs here that wouldn't be freed otherwise because they have no children - if (outputs != NULL && outputs[g] != NULL) { - for (int i = 0; outputs[g][i] != NULL; i++) { - struct ggml_tensor * output = outputs[g][i]; - AT_PRINTF("output: %s\n", output->name); - ggml_allocr_free_tensor(alloc, output); + AT_PRINTF("\n"); + if (parse_seq_len) { + last_barrier_pos = ind + 1; } } } - - return alloc->max_size; } -size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) { - return ggml_allocr_alloc_graph_n(alloc, &graph, 1, NULL, NULL); +size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph) { + size_t hash_size = graph->visited_hash_table.size; + + // check if the hash table is initialized and large enough + if (galloc->hash_set.size < hash_size) { + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + if (galloc->hash_values != NULL) { + free(galloc->hash_values); + } + galloc->hash_set.keys = malloc(sizeof(struct ggml_tensor *) * hash_size); + galloc->hash_set.size = hash_size; + galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size); + } + + // reset hash table + memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * hash_size); + memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size); + + galloc->talloc = talloc; + ggml_tallocr_alloc_graph_impl(galloc, graph); + galloc->talloc = NULL; + + size_t max_size = ggml_tallocr_max_size(talloc); + + return max_size; } -size_t ggml_allocr_max_size(struct ggml_allocr * alloc) { - return alloc->max_size; +void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_talloc) { + const size_t hash_size = hash_set.size; + + GGML_ASSERT(hash_size >= (size_t)(graph->n_nodes + graph->n_leafs)); + + galloc->talloc = NULL; + + // alloc hash_values if needed + if (galloc->hash_values == NULL || galloc->hash_values_size < hash_size) { + free(galloc->hash_values); + galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size); + galloc->hash_values_size = hash_size; + } + + // free hash_set.keys if needed + if (galloc->hash_set.keys != NULL) { + free(galloc->hash_set.keys); + } + galloc->hash_set = hash_set; + + // reset hash values + memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size); + + galloc->hash_allocs = hash_node_talloc; + + ggml_tallocr_alloc_graph_impl(galloc, graph); + + // remove unowned resources + galloc->hash_set.keys = NULL; + galloc->hash_allocs = NULL; +} + +// legacy API wrapper + +struct ggml_allocr { + ggml_tallocr_t talloc; + ggml_gallocr_t galloc; +}; + +static ggml_allocr_t ggml_allocr_new_impl(ggml_tallocr_t talloc) { + ggml_allocr_t alloc = (ggml_allocr_t)malloc(sizeof(struct ggml_allocr)); + *alloc = (struct ggml_allocr) { + /*.talloc = */ talloc, + /*.galloc = */ ggml_gallocr_new(), + }; + return alloc; +} + +ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment) { + return ggml_allocr_new_impl(ggml_tallocr_new(data, size, alignment)); +} + +ggml_allocr_t ggml_allocr_new_measure(size_t alignment) { + return ggml_allocr_new_impl(ggml_tallocr_new_measure(alignment)); +} + +ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) { + return ggml_allocr_new_impl(ggml_tallocr_new_from_buffer(buffer)); +} + +ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size) { + return ggml_allocr_new_impl(ggml_tallocr_new_from_backend(backend, size)); +} + +ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend) { + return ggml_allocr_new_impl(ggml_tallocr_new_measure_from_backend(backend)); +} + +struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc) { + return ggml_tallocr_get_buffer(alloc->talloc); +} + +void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) { + ggml_gallocr_set_parse_seq(alloc->galloc, list, n); +} + +void ggml_allocr_free(ggml_allocr_t alloc) { + ggml_gallocr_free(alloc->galloc); + ggml_tallocr_free(alloc->talloc); + free(alloc); +} + +bool ggml_allocr_is_measure(ggml_allocr_t alloc) { + return ggml_tallocr_is_measure(alloc->talloc); +} + +void ggml_allocr_reset(ggml_allocr_t alloc) { + ggml_tallocr_reset(alloc->talloc); +} + +void ggml_allocr_alloc(ggml_allocr_t alloc, struct ggml_tensor * tensor) { + ggml_tallocr_alloc(alloc->talloc, tensor); +} + +size_t ggml_allocr_max_size(ggml_allocr_t alloc) { + return ggml_tallocr_max_size(alloc->talloc); +} + +size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) { + return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph); } diff --git a/ggml-alloc.h b/ggml-alloc.h index e38758878..dde2a06bf 100644 --- a/ggml-alloc.h +++ b/ggml-alloc.h @@ -6,27 +6,79 @@ extern "C" { #endif +struct ggml_backend; struct ggml_backend_buffer; -GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment); -GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment); -GGML_API struct ggml_allocr * ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer); +// +// Legacy API +// + +typedef struct ggml_allocr * ggml_allocr_t; + +// initialize allocator for use with CPU backend only +GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment); +GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment); + +// initialize allocator for use with ggml-backend +GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer); +GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer +GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend); + +GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc); // tell the allocator to parse nodes following the order described in the list // you should call this if your graph are optimized to execute out-of-order -GGML_API void ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, const int * list, int n); +GGML_API void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n); -GGML_API void ggml_allocr_free (struct ggml_allocr * alloc); -GGML_API bool ggml_allocr_is_measure (struct ggml_allocr * alloc); -GGML_API void ggml_allocr_reset (struct ggml_allocr * alloc); -GGML_API void ggml_allocr_alloc (struct ggml_allocr * alloc, struct ggml_tensor * tensor); -GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph); -GGML_API size_t ggml_allocr_max_size (struct ggml_allocr * alloc); +GGML_API void ggml_allocr_free (ggml_allocr_t alloc); +GGML_API bool ggml_allocr_is_measure (ggml_allocr_t alloc); +GGML_API void ggml_allocr_reset (ggml_allocr_t alloc); +GGML_API void ggml_allocr_alloc (ggml_allocr_t alloc, struct ggml_tensor * tensor); +GGML_API size_t ggml_allocr_max_size (ggml_allocr_t alloc); -GGML_API size_t ggml_allocr_alloc_graph_n( - struct ggml_allocr * alloc, - struct ggml_cgraph ** graphs, int n_graphs, - struct ggml_tensor *** inputs, struct ggml_tensor *** outputs); +GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph); + +// +// ggml-backend v2 API +// + +// Seperate tensor and graph allocator objects +// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators +// The original API is kept as a wrapper around the new API + +// Tensor allocator +typedef struct ggml_tallocr * ggml_tallocr_t; + +GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment); +GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment); +GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer); +GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer +GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend); + +GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc); + +GGML_API void ggml_tallocr_free (ggml_tallocr_t talloc); +GGML_API bool ggml_tallocr_is_measure (ggml_tallocr_t talloc); +GGML_API void ggml_tallocr_reset (ggml_tallocr_t talloc); +GGML_API void ggml_tallocr_alloc (ggml_tallocr_t talloc, struct ggml_tensor * tensor); +GGML_API size_t ggml_tallocr_max_size (ggml_tallocr_t talloc); + + +// Graph allocator +typedef struct ggml_gallocr * ggml_gallocr_t; + +GGML_API ggml_gallocr_t ggml_gallocr_new(void); +GGML_API void ggml_gallocr_free(ggml_gallocr_t galloc); + +GGML_API void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n); +GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph); + +// Allocate tensors from the allocators given by the hash table +GGML_API void ggml_gallocr_alloc_graph_n( + ggml_gallocr_t galloc, + struct ggml_cgraph * graph, + struct ggml_hash_set hash_set, + ggml_tallocr_t * hash_node_talloc); #ifdef __cplusplus } diff --git a/ggml-backend-impl.h b/ggml-backend-impl.h new file mode 100644 index 000000000..211e3d424 --- /dev/null +++ b/ggml-backend-impl.h @@ -0,0 +1,87 @@ +#pragma once + +// ggml-backend internal header + +#include "ggml-backend.h" + +#ifdef __cplusplus +extern "C" { +#endif + + // + // Backend buffer + // + + typedef void * ggml_backend_buffer_context_t; + + struct ggml_backend_buffer_i { + void (*free_buffer) (ggml_backend_buffer_t buffer); + void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer + size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback + void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback + void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback + }; + + struct ggml_backend_buffer { + struct ggml_backend_buffer_i iface; + + ggml_backend_t backend; + ggml_backend_buffer_context_t context; + + size_t size; + }; + + GGML_API ggml_backend_buffer_t ggml_backend_buffer_init( + struct ggml_backend * backend, + struct ggml_backend_buffer_i iface, + ggml_backend_buffer_context_t context, + size_t size); + + // + // Backend + // + + typedef void * ggml_backend_context_t; + + struct ggml_backend_i { + const char * (*get_name)(ggml_backend_t backend); + + void (*free)(ggml_backend_t backend); + + // buffer allocation + ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size); + + // get buffer alignment + size_t (*get_alignment)(ggml_backend_t backend); + + // tensor data access + // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize + void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); + void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); + void (*synchronize) (ggml_backend_t backend); + + // (optional) copy tensor between different backends, allow for single-copy tranfers + void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); + void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); + + // compute graph with a plan + ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph); + void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); + void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); + + // compute graph without a plan + void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); + + // check if the backend supports an operation + bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); + }; + + struct ggml_backend { + struct ggml_backend_i iface; + + ggml_backend_context_t context; + }; + +#ifdef __cplusplus +} +#endif diff --git a/ggml-backend.c b/ggml-backend.c index ca8d83daf..f6e5fceed 100644 --- a/ggml-backend.c +++ b/ggml-backend.c @@ -1,7 +1,9 @@ -#include "ggml-backend.h" +#include "ggml-backend-impl.h" #include "ggml-alloc.h" +#include "ggml-impl.h" #include +#include #include #include #include @@ -33,6 +35,10 @@ ggml_backend_buffer_t ggml_backend_buffer_init( } void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) { + if (buffer == NULL) { + return; + } + if (buffer->iface.free_buffer != NULL) { buffer->iface.free_buffer(buffer); } @@ -43,15 +49,20 @@ size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) { return ggml_backend_get_alignment(buffer->backend); } -void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { - return buffer->iface.get_base(buffer); -} - size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) { return buffer->size; } +void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) { + void * base = buffer->iface.get_base(buffer); + + GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL"); + + return base; +} + size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // get_alloc_size is optional, defaults to ggml_nbytes if (buffer->iface.get_alloc_size) { return buffer->iface.get_alloc_size(buffer, tensor); } @@ -59,12 +70,14 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g } void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // init_tensor is optional if (buffer->iface.init_tensor) { buffer->iface.init_tensor(buffer, tensor); } } void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { + // free_tensor is optional if (buffer->iface.free_tensor) { buffer->iface.free_tensor(buffer, tensor); } @@ -73,14 +86,21 @@ void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_t // backend ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) { - return tensor->buffer->backend; + return tensor->buffer ? tensor->buffer->backend : NULL; } const char * ggml_backend_name(ggml_backend_t backend) { + if (backend == NULL) { + return "NULL"; + } return backend->iface.get_name(backend); } void ggml_backend_free(ggml_backend_t backend) { + if (backend == NULL) { + return; + } + backend->iface.free(backend); } @@ -101,13 +121,23 @@ void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * dat } void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); - ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor)); + ggml_backend_t backend = ggml_get_backend(tensor); + + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(backend != NULL && "tensor backend not set"); + + backend->iface.set_tensor_async(backend, tensor, data, offset, size); + backend->iface.synchronize(backend); } void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) { - ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size); - ggml_get_backend(tensor)->iface.synchronize(ggml_get_backend(tensor)); + ggml_backend_t backend = ggml_get_backend(tensor); + + GGML_ASSERT(tensor->data != NULL && "tensor not allocated"); + GGML_ASSERT(backend != NULL && "tensor backend not set"); + + backend->iface.get_tensor_async(backend, tensor, data, offset, size); + backend->iface.synchronize(backend); } void ggml_backend_synchronize(ggml_backend_t backend) { @@ -156,7 +186,7 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]); GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts"); - // printf("cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src)); + // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src)); if (src == dst) { return; @@ -234,6 +264,8 @@ static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backen size += TENSOR_ALIGNMENT; // malloc may return an address that is not aligned void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC? + GGML_ASSERT(data != NULL && "failed to allocate buffer"); + return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size); } @@ -271,8 +303,7 @@ static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml } static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) { - // for a backend such as CUDA that can queue async calls, it is ok to do this asynchronously, but it may not be the case for other backends - ggml_backend_tensor_set_async(dst, src->data, 0, ggml_nbytes(src)); + ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src)); UNUSED(backend); } @@ -383,3 +414,537 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) { ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) { return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size); } + +// scheduler + +#define GGML_MAX_BACKENDS 4 +#define GGML_MAX_SPLITS 256 +#define GGML_MAX_SPLIT_INPUTS 16 + +struct ggml_backend_sched_split { + ggml_tallocr_t tallocr; + int i_start; + int i_end; + struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS]; + int n_inputs; + struct ggml_cgraph * graph; +}; + +struct ggml_backend_sched { + int n_backends; + ggml_backend_t backends[GGML_MAX_BACKENDS]; + ggml_tallocr_t tallocs[GGML_MAX_BACKENDS]; + + ggml_gallocr_t galloc; + + struct ggml_hash_set hash_set; + ggml_tallocr_t * node_talloc; // [hash_set.size] + struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS] + + struct ggml_cgraph * graph; + struct ggml_backend_sched_split splits[GGML_MAX_SPLITS]; + int n_splits; + + struct ggml_context * ctx; + + // align context_buffer to GGML_MEM_ALIGN + #ifdef _MSC_VER + __declspec(align(GGML_MEM_ALIGN)) + #else + __attribute__((aligned(GGML_MEM_ALIGN))) + #endif + char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)]; +}; + +#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node) +#define node_allocr(node) sched->node_talloc[hash_id(node)] + +static bool ggml_is_view_op(enum ggml_op op) { + return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE; +} + +// returns the priority of the backend, lower is better +static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) { + for (int i = 0; i < sched->n_backends; i++) { + if (sched->backends[i] == backend) { + return i; + } + } + return INT_MAX; +} + +static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) { + for (int i = 0; i < sched->n_backends; i++) { + if (sched->tallocs[i] == allocr) { + return i; + } + } + return INT_MAX; +} + +// returns the backend that should be used for the node based on the current locations +char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove +static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) { + // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there + // ie. kv cache updates + // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend. + // dst + ggml_backend_t cur_backend = ggml_get_backend(node); + if (cur_backend != NULL) { + sprintf(causes[hash_id(node)], "1.dst"); + return cur_backend; + } + + // view_src + if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) { + sprintf(causes[hash_id(node)], "1.vsrc"); + return ggml_get_backend(node->view_src); + } + + // src + int cur_prio = INT_MAX; + size_t cur_size = 0; + + for (int i = 0; i < GGML_MAX_SRC; i++) { + const struct ggml_tensor * src = node->src[i]; + if (src == NULL) { + break; + } + ggml_backend_t src_backend = ggml_get_backend(src); + if (src_backend != NULL) { + int src_prio = sched_backend_prio(sched, src_backend); + size_t src_size = ggml_nbytes(src); + if (src_prio < cur_prio && src_size >= cur_size) { + cur_prio = src_prio; + cur_size = src_size; + cur_backend = src_backend; + sprintf(causes[hash_id(node)], "1.src%d", i); + } + } + } + return cur_backend; +} + +static char * fmt_size(size_t size) { + static char buffer[128]; + if (size >= 1024*1024) { + sprintf(buffer, "%zuM", size/1024/1024); + } else { + sprintf(buffer, "%zuK", size/1024); + } + return buffer; +} + +static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + int cur_split = 0; + for (int i = 0; i < graph->n_nodes; i++) { + if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) { + ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend; + fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs); + for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) { + fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j]))); + } + fprintf(stderr, "\n"); + cur_split++; + } + struct ggml_tensor * node = graph->nodes[i]; + if (ggml_is_view_op(node->op)) { + continue; + } + ggml_tallocr_t node_allocr = node_allocr(node); + ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL; + fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL; + fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]); + } + fprintf(stderr, "\n"); + } +} + +// creates a copy of the tensor with the same memory layout +static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) { + struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor); + for (int i = 0; i < GGML_MAX_DIMS; i++) { + dup->nb[i] = tensor->nb[i]; + } + return dup; +} + +// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend +// TODO: merge passes +static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + // reset state + size_t hash_size = sched->hash_set.size; + memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size); + memset(sched->node_talloc, 0, sizeof(sched->node_talloc[0]) * hash_size); + memset(sched->node_copies, 0, sizeof(sched->node_copies[0]) * hash_size); + sched->n_splits = 0; + + struct ggml_init_params params = { + /*.mem_size = */ sizeof(sched->context_buffer), + /*.mem_buffer = */ sched->context_buffer, + /*.no_alloc = */ true + }; + + if (sched->ctx != NULL) { + ggml_free(sched->ctx); + } + + sched->ctx = ggml_init(params); + + // pass 1: assign backends to ops with allocated inputs + for (int i = 0; i < graph->n_leafs; i++) { + struct ggml_tensor * leaf = graph->leafs[i]; + if (node_allocr(leaf) != NULL) { + // do not overwrite user assignments + continue; + } + ggml_backend_t leaf_backend = ggml_get_backend(leaf); + if (leaf_backend == NULL && leaf->view_src != NULL) { + leaf_backend = ggml_get_backend(leaf->view_src); + } + if (leaf_backend != NULL) { + node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend); + } + } + + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (node_allocr(node) != NULL) { + // do not overwrite user assignments + continue; + } + ggml_backend_t node_backend = sched_backend_from_cur(sched, node); + if (node_backend != NULL) { + node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend); + } + } + //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 2: assign backends to ops from current assignments + // TODO: + // - reuse sched_backend_from_cur + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + if (node_allocr == NULL) { + int cur_prio = INT_MAX; + size_t cur_size = 0; + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != NULL) { + int src_prio = sched_allocr_prio(sched, src_allocr); + size_t src_size = ggml_nbytes(src); + if (src_prio < cur_prio && src_size >= cur_size) { + cur_prio = src_prio; + cur_size = src_size; + node_allocr = src_allocr; + sprintf(causes[hash_id(node)], "2.src%d", j); + } + } + } + if (node_allocr != NULL) { + node_allocr(node) = node_allocr; + } + } + } + //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 3: assign backends to remaining src from dst (should only be leafs) + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr == NULL) { + node_allocr(src) = node_allocr; + } + } + } + //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); + + // pass 4: split graph, find tensors that need to be copied + // TODO: + // - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost + // find first backend + int cur_split = 0; + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + if (node->view_src == NULL) { + sched->splits[0].tallocr = node_allocr(node); + break; + } + } + sched->splits[0].i_start = 0; + sched->splits[0].n_inputs = 0; + memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK + ggml_tallocr_t cur_allocr = sched->splits[0].tallocr; + size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr); + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + + if (ggml_is_view_op(node->op)) { + continue; + } + + ggml_tallocr_t node_allocr = node_allocr(node); + + if (node_allocr != cur_allocr) { + sched->splits[cur_split].i_end = i; + cur_split++; + GGML_ASSERT(cur_split < GGML_MAX_SPLITS); + sched->splits[cur_split].tallocr = node_allocr; + sched->splits[cur_split].i_start = i; + sched->splits[cur_split].n_inputs = 0; + memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK + cur_allocr = node_allocr; + cur_backend_id = sched_allocr_prio(sched, cur_allocr); + } + + // find inputs that are not on the same backend + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != node_allocr) { + int n_inputs = sched->splits[cur_split].n_inputs++; + GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS); + sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src; + + // create copies + size_t id = hash_id(src); + if (sched->node_copies[id][cur_backend_id] == NULL) { + struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src); + sched->node_copies[id][cur_backend_id] = tensor_copy; + node_allocr(tensor_copy) = cur_allocr; + ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend; + ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name); + } + node->src[j] = sched->node_copies[id][cur_backend_id]; + } + } + } + sched->splits[cur_split].i_end = graph->n_nodes; + sched->n_splits = cur_split + 1; + + //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout); + +#if 1 + // sanity check: all sources should have the same backend as the node + for (int i = 0; i < graph->n_nodes; i++) { + struct ggml_tensor * node = graph->nodes[i]; + ggml_tallocr_t node_allocr = node_allocr(node); + if (node_allocr == NULL) { + fprintf(stderr, "!!!!!!! %s has no backend\n", node->name); + } + for (int j = 0; j < GGML_MAX_SRC; j++) { + struct ggml_tensor * src = node->src[j]; + if (src == NULL) { + break; + } + ggml_tallocr_t src_allocr = node_allocr(src); + if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now + fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n", + node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL", + j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL"); + } + } + } +#endif + + // create copies of the graph for each split + // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way + struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false); + for (int i = 0; i < sched->n_splits; i++) { + struct ggml_backend_sched_split * split = &sched->splits[i]; + split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end); + + // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split + for (int j = 0; j < split->n_inputs; j++) { + struct ggml_tensor * input = split->inputs[j]; + struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)]; + input_cpy->src[0] = input; + graph_copy->nodes[graph_copy->n_nodes++] = input_cpy; + } + + for (int j = split->i_start; j < split->i_end; j++) { + graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j]; + } + } + sched->graph = graph_copy; +} + +static void sched_alloc_splits(ggml_backend_sched_t sched) { + ggml_gallocr_alloc_graph_n( + sched->galloc, + sched->graph, + sched->hash_set, + sched->node_talloc); +} + +static void sched_compute_splits(ggml_backend_sched_t sched) { + uint64_t copy_us[GGML_MAX_BACKENDS] = {0}; + uint64_t compute_us[GGML_MAX_BACKENDS] = {0}; + + struct ggml_backend_sched_split * splits = sched->splits; + + for (int i = 0; i < sched->n_splits; i++) { + struct ggml_backend_sched_split * split = &splits[i]; + ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend; + int split_backend_id = sched_backend_prio(sched, split_backend); + + // copy the input tensors to the split backend + uint64_t copy_start_us = ggml_time_us(); + for (int j = 0; j < split->n_inputs; j++) { + struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)]; + if (split->inputs[j]->buffer == NULL) { + if (split->inputs[j]->view_src == NULL) { + fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name); + exit(1); + } + struct ggml_tensor * view = split->inputs[j]; + view->backend = view->view_src->backend; + view->buffer = view->view_src->buffer; + view->data = (char *)view->view_src->data + view->view_offs; + ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view); + } + if (input_cpy->buffer == NULL) { + fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name); + exit(1); + } + GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend); + GGML_ASSERT(input_cpy->buffer->backend == split_backend); + ggml_backend_tensor_copy(split->inputs[j], input_cpy); + } + // ggml_backend_synchronize(split_backend); + int64_t copy_end_us = ggml_time_us(); + copy_us[split_backend_id] += copy_end_us - copy_start_us; + +#if 0 + char split_filename[GGML_MAX_NAME]; + snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend)); + ggml_graph_dump_dot(split->graph, NULL, split_filename); +#endif + + uint64_t compute_start_us = ggml_time_us(); + ggml_backend_graph_compute(split_backend, split->graph); + // ggml_backend_synchronize(split_backend); + uint64_t compute_end_us = ggml_time_us(); + compute_us[split_backend_id] += compute_end_us - compute_start_us; + } + +#if 0 + // per-backend timings + fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits); + for (int i = 0; i < sched->n_backends; i++) { + if (copy_us[i] > 0 || compute_us[i] > 0) { + fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]); + } + } +#endif +} + +static void sched_reset(ggml_backend_sched_t sched) { + for (int i = 0; i < sched->n_backends; i++) { + ggml_tallocr_reset(sched->tallocs[i]); + } +} + +ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) { + GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS); + + struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched)); + memset(sched, 0, sizeof(struct ggml_backend_sched)); + + fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024); + + sched->n_backends = n_backends; + for (int i = 0; i < n_backends; i++) { + sched->backends[i] = backends[i]; + } + + sched->galloc = ggml_gallocr_new(); + + // init measure allocs for each backend + for (int i = 0; i < n_backends; i++) { + sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]); + } + + return sched; +} + +void ggml_backend_sched_free(ggml_backend_sched_t sched) { + if (sched == NULL) { + return; + } + for (int i = 0; i < sched->n_backends; i++) { + ggml_tallocr_free(sched->tallocs[i]); + } + ggml_gallocr_free(sched->galloc); + free(sched->hash_set.keys); + free(sched->node_talloc); + free(sched->node_copies); + free(sched); +} + +void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) { + // initialize hash tables + size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS; + sched->hash_set.size = hash_size; + sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size); + sched->node_talloc = malloc(sizeof(sched->node_talloc[0]) * hash_size); + sched->node_copies = malloc(sizeof(sched->node_copies[0]) * hash_size); + + sched_split_graph(sched, measure_graph); + sched_alloc_splits(sched); + + // allocate buffers and reset allocators + for (int i = 0; i < sched->n_backends; i++) { + size_t size = ggml_tallocr_max_size(sched->tallocs[i]); + ggml_tallocr_free(sched->tallocs[i]); + sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size); + } + + sched_reset(sched); +} + +void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) { + GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS); + + sched_split_graph(sched, graph); + sched_alloc_splits(sched); + sched_compute_splits(sched); + sched_reset(sched); +} + +ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + return sched->tallocs[backend_index]; +} + +ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + return ggml_tallocr_get_buffer(sched->tallocs[backend_index]); +} + +void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) { + int backend_index = sched_backend_prio(sched, backend); + GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends); + node_allocr(node) = sched->tallocs[backend_index]; +} diff --git a/ggml-backend.h b/ggml-backend.h index da134b0db..966687320 100644 --- a/ggml-backend.h +++ b/ggml-backend.h @@ -1,51 +1,20 @@ #pragma once #include "ggml.h" +#include "ggml-alloc.h" #ifdef __cplusplus extern "C" { #endif - struct ggml_backend; + + // + // Backend buffer + // + struct ggml_backend_buffer; - - // type-erased backend-specific types / wrappers - typedef void * ggml_backend_context_t; - typedef void * ggml_backend_graph_plan_t; - typedef void * ggml_backend_buffer_context_t; - - // avoid accessing internals of these types - typedef struct ggml_backend * ggml_backend_t; typedef struct ggml_backend_buffer * ggml_backend_buffer_t; - // - // backend buffer - // - - struct ggml_backend_buffer_i { - void (*free_buffer) (ggml_backend_buffer_t buffer); - void * (*get_base) (ggml_backend_buffer_t buffer); // get base pointer - size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback - void (*init_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback - void (*free_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback - }; - - // TODO: hide behind API - struct ggml_backend_buffer { - struct ggml_backend_buffer_i iface; - - ggml_backend_t backend; - ggml_backend_buffer_context_t context; - - size_t size; - }; - // backend buffer functions - GGML_API ggml_backend_buffer_t ggml_backend_buffer_init( - struct ggml_backend * backend, - struct ggml_backend_buffer_i iface, - ggml_backend_buffer_context_t context, - size_t size); - GGML_API void ggml_backend_buffer_free (ggml_backend_buffer_t buffer); GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer); GGML_API void * ggml_backend_buffer_get_base (ggml_backend_buffer_t buffer); @@ -55,50 +24,13 @@ extern "C" { GGML_API void ggml_backend_buffer_free_tensor (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // - // backend + // Backend // - struct ggml_backend_i { - const char * (*get_name)(ggml_backend_t backend); + struct ggml_backend; + typedef struct ggml_backend * ggml_backend_t; + typedef void * ggml_backend_graph_plan_t; - void (*free)(ggml_backend_t backend); - - // buffer allocation - ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size); - - // get buffer alignment - size_t (*get_alignment)(ggml_backend_t backend); - - // tensor data access - // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize - void (*set_tensor_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size); - void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size); - void (*synchronize) (ggml_backend_t backend); - - // (optional) copy tensor between different backends, allow for single-copy tranfers - void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); - void (*cpy_tensor_to) (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); - - // compute graph with a plan - ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph); - void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan); - void (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan); - - // compute graph without a plan - void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph); - - // check if the backend supports an operation - bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op); - }; - - // TODO: hide behind API - struct ggml_backend { - struct ggml_backend_i iface; - - ggml_backend_context_t context; - }; - - // backend helper functions GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor); GGML_API const char * ggml_backend_name(ggml_backend_t backend); @@ -133,11 +65,72 @@ extern "C" { GGML_API ggml_backend_t ggml_backend_cpu_init(void); GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend); - GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads); + // Create a backend buffer from an existing pointer GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size); + + // + // Backend scheduler + // + + // The backend scheduler allows for multiple backends to be used together + // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends + // The backends are selected based on: + // - the backend that supports the operation + // - the location of the pre-allocated tensors (e.g. the weights) + /* + Example usage: + + sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends); + // sched is initialized with measure allocators and cannot be used until allocated with a measure graph + + // initialize buffers from a measure graph + measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed + + // in build_graph: + build_graph(...) { + // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer) + alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu); + ggml_allocr_alloc(alloc_cpu, tensor); + + // manually assigning nodes to a backend (optional, shouldn't be needed in most cases) + struct ggml_tensor * node = ggml_mul_mat(ctx, ...); + ggml_backend_sched_set_node_backend(sched, node, backend_gpu); + } + + // allocate backend buffers from measure graph + ggml_backend_sched_init_measure(sched, measure_graph); + + // the scheduler is now ready to compute graphs + + // compute + graph = build_graph(sched); + ggml_backend_sched_graph_compute(sched, graph); + */ + + struct ggml_backend_sched; + typedef struct ggml_backend_sched * ggml_backend_sched_t; + + // Initialize a backend scheduler + GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends); + + GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched); + + // Initialize backend buffers from a measure graph + GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); + + GGML_API ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend); + GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend); + + GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend); + + // Allocate a graph on the backend scheduler + GGML_API void ggml_backend_sched_graph_compute( + ggml_backend_sched_t sched, + struct ggml_cgraph * graph); + #ifdef __cplusplus } #endif diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 1ba951f68..5b80e4ae3 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -1,4 +1,5 @@ #include +#include #include #include #include @@ -81,12 +82,15 @@ #include "ggml-cuda.h" #include "ggml.h" +#include "ggml-backend-impl.h" #define MIN_CC_DP4A 610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products #define CC_VOLTA 700 #define CC_OFFSET_AMD 1000000 #define CC_RDNA2 (CC_OFFSET_AMD + 1030) +#define GGML_CUDA_MAX_NODES 8192 + // define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication // on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant // for large computational tasks. the drawback is that this requires some extra amount of VRAM: @@ -232,7 +236,7 @@ typedef float2 dfloat2; #endif //GGML_CUDA_F16 static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) { - const uint16_t * x16 = (uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment int x32 = 0; x32 |= x16[0] << 0; @@ -242,7 +246,7 @@ static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const } static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) { - const uint16_t * x16 = (uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment + const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment int x32 = 0; x32 |= x16[0] << 0; @@ -252,11 +256,11 @@ static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, con } static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) { - return *((int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment } static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) { - return *((int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment + return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment } template @@ -433,6 +437,8 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_ #define CUDA_MUL_BLOCK_SIZE 256 #define CUDA_GELU_BLOCK_SIZE 256 #define CUDA_SILU_BLOCK_SIZE 256 +#define CUDA_RELU_BLOCK_SIZE 256 +#define CUDA_SQR_BLOCK_SIZE 256 #define CUDA_CPY_BLOCK_SIZE 32 #define CUDA_SCALE_BLOCK_SIZE 256 #define CUDA_CLAMP_BLOCK_SIZE 256 @@ -464,7 +470,7 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA #define MUL_MAT_SRC1_COL_STRIDE 128 #define MAX_STREAMS 8 -static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { nullptr }; +static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } }; struct ggml_tensor_extra_gpu { void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors @@ -513,6 +519,15 @@ static __global__ void add_f16_f32_f16(const half * x, const float * y, half * d dst[i] = __hadd(x[i], __float2half(y[i])); } +static __global__ void add_f16_f32_f32(const half * x, const float * y, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = __half2float(x[i]) + y[i]; +} + static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) { const int i = blockDim.x*blockIdx.x + threadIdx.x; @@ -544,6 +559,24 @@ static __global__ void silu_f32(const float * x, float * dst, const int k) { dst[i] = x[i] / (1.0f + expf(-x[i])); } +static __global__ void relu_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = fmaxf(x[i], 0); +} + +static __global__ void sqr_f32(const float * x, float * dst, const int k) { + const int i = blockDim.x*blockIdx.x + threadIdx.x; + + if (i >= k) { + return; + } + dst[i] = x[i] * x[i]; +} + static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) { #pragma unroll for (int mask = 16; mask > 0; mask >>= 1) { @@ -973,7 +1006,7 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row > nrows) return; const int num_blocks_per_row = ncols / QK_K; @@ -1077,7 +1110,7 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row > nrows) return; const int num_blocks_per_row = ncols / QK_K; @@ -1181,7 +1214,7 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) { - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row > nrows) return; const int num_blocks_per_row = ncols / QK_K; const int ib0 = row*num_blocks_per_row; @@ -1435,7 +1468,7 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION"); - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row > nrows) return; const int num_blocks_per_row = ncols / QK_K; @@ -2216,6 +2249,7 @@ static __device__ __forceinline__ float vec_dot_q4_0_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; (void)x_sc; __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0]; @@ -2227,7 +2261,7 @@ template static __device__ __forceinline__ void allocate_tiles_q4_0( template static __device__ __forceinline__ void load_tiles_q4_0( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { - + (void)x_qh; (void)x_sc; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); GGML_CUDA_ASSUME(k >= 0); @@ -2236,7 +2270,7 @@ template static __device__ __forceinlin const int kbx = k / QI4_0; const int kqsx = k % QI4_0; - const block_q4_0 * bx0 = (block_q4_0 *) vx; + const block_q4_0 * bx0 = (const block_q4_0 *) vx; float * x_dmf = (float *) x_dm; @@ -2274,9 +2308,10 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; (void)x_sc; const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); - const float * x_dmf = (float *) x_dm; + const float * x_dmf = (const float *) x_dm; int u[2*VDR_Q4_0_Q8_1_MMQ]; @@ -2310,6 +2345,7 @@ static __device__ __forceinline__ float vec_dot_q4_1_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; (void)x_sc; __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1]; @@ -2321,6 +2357,7 @@ template static __device__ __forceinline__ void allocate_tiles_q4_1( template static __device__ __forceinline__ void load_tiles_q4_1( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; (void)x_sc; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -2330,7 +2367,7 @@ template static __device__ __forceinlin const int kbx = k / QI4_1; const int kqsx = k % QI4_1; - const block_q4_1 * bx0 = (block_q4_1 *) vx; + const block_q4_1 * bx0 = (const block_q4_1 *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2365,6 +2402,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; (void)x_sc; const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); @@ -2402,6 +2440,7 @@ static __device__ __forceinline__ float vec_dot_q5_0_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; (void)x_sc; __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0]; @@ -2413,6 +2452,7 @@ template static __device__ __forceinline__ void allocate_tiles_q5_0( template static __device__ __forceinline__ void load_tiles_q5_0( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; (void)x_sc; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -2422,7 +2462,7 @@ template static __device__ __forceinlin const int kbx = k / QI5_0; const int kqsx = k % QI5_0; - const block_q5_0 * bx0 = (block_q5_0 *) vx; + const block_q5_0 * bx0 = (const block_q5_0 *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2477,6 +2517,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; (void)x_sc; const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0; @@ -2516,6 +2557,7 @@ static __device__ __forceinline__ float vec_dot_q5_1_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; (void)x_sc; __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1]; @@ -2527,6 +2569,7 @@ template static __device__ __forceinline__ void allocate_tiles_q5_1( template static __device__ __forceinline__ void load_tiles_q5_1( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; (void)x_sc; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -2536,7 +2579,7 @@ template static __device__ __forceinlin const int kbx = k / QI5_1; const int kqsx = k % QI5_1; - const block_q5_1 * bx0 = (block_q5_1 *) vx; + const block_q5_1 * bx0 = (const block_q5_1 *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2588,6 +2631,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; (void)x_sc; const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2)); const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1; @@ -2622,6 +2666,7 @@ static __device__ __forceinline__ float vec_dot_q8_0_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; (void)x_sc; __shared__ int tile_x_qs[mmq_y * (WARP_SIZE) + mmq_y]; __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0]; @@ -2633,6 +2678,7 @@ template static __device__ __forceinline__ void allocate_tiles_q8_0( template static __device__ __forceinline__ void load_tiles_q8_0( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; (void)x_sc; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -2643,7 +2689,7 @@ template static __device__ __forceinlin const int kqsx = k % QI8_0; float * x_dmf = (float *) x_dm; - const block_q8_0 * bx0 = (block_q8_0 *) vx; + const block_q8_0 * bx0 = (const block_q8_0 *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2678,6 +2724,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; (void)x_sc; const float * x_dmf = (const float *) x_dm; const float * y_df = (const float *) y_ds; @@ -2711,6 +2758,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K]; @@ -2724,6 +2772,7 @@ template static __device__ __forceinline__ void allocate_tiles_q2_K( template static __device__ __forceinline__ void load_tiles_q2_K( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -2733,7 +2782,7 @@ template static __device__ __forceinlin const int kbx = k / QI2_K; const int kqsx = k % QI2_K; - const block_q2_K * bx0 = (block_q2_K *) vx; + const block_q2_K * bx0 = (const block_q2_K *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2781,6 +2830,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; const int kbx = k / QI2_K; const int ky = (k % QI2_K) * QR2_K; @@ -2854,7 +2904,7 @@ template static __device__ __forceinlin const int kbx = k / QI3_K; const int kqsx = k % QI3_K; - const block_q3_K * bx0 = (block_q3_K *) vx; + const block_q3_K * bx0 = (const block_q3_K *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -2935,7 +2985,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat( const float * x_dmf = (const float *) x_dm; const float * y_df = (const float *) y_ds; - const int8_t * scales = ((int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; + const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4; int v[QR3_K*VDR_Q3_K_Q8_1_MMQ]; @@ -3050,6 +3100,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; __shared__ int tile_x_ql[mmq_y * (WARP_SIZE) + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K]; @@ -3063,6 +3114,7 @@ template static __device__ __forceinline__ void allocate_tiles_q4_K( template static __device__ __forceinline__ void load_tiles_q4_K( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -3072,7 +3124,7 @@ template static __device__ __forceinlin const int kbx = k / QI4_K; // == 0 if QK_K == 256 const int kqsx = k % QI4_K; // == k if QK_K == 256 - const block_q4_K * bx0 = (block_q4_K *) vx; + const block_q4_K * bx0 = (const block_q4_K *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -3117,7 +3169,7 @@ template static __device__ __forceinlin const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8); - const int * scales = (int *) bxi->scales; + const int * scales = (const int *) bxi->scales; const int ksc = k % (WARP_SIZE/8); @@ -3132,6 +3184,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8); @@ -3231,6 +3284,7 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K]; @@ -3244,6 +3298,7 @@ template static __device__ __forceinline__ void allocate_tiles_q5_K( template static __device__ __forceinline__ void load_tiles_q5_K( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -3253,7 +3308,7 @@ template static __device__ __forceinlin const int kbx = k / QI5_K; // == 0 if QK_K == 256 const int kqsx = k % QI5_K; // == k if QK_K == 256 - const block_q5_K * bx0 = (block_q5_K *) vx; + const block_q5_K * bx0 = (const block_q5_K *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -3309,7 +3364,7 @@ template static __device__ __forceinlin const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8); - const int * scales = (int *) bxi->scales; + const int * scales = (const int *) bxi->scales; const int ksc = k % (WARP_SIZE/8); @@ -3324,6 +3379,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8); @@ -3360,6 +3416,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1( } template static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) { + (void)x_qh; __shared__ int tile_x_ql[mmq_y * (2*WARP_SIZE) + mmq_y]; __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K]; @@ -3373,6 +3430,7 @@ template static __device__ __forceinline__ void allocate_tiles_q6_K( template static __device__ __forceinline__ void load_tiles_q6_K( const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh, int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) { + (void)x_qh; GGML_CUDA_ASSUME(i_offset >= 0); GGML_CUDA_ASSUME(i_offset < nwarps); @@ -3382,7 +3440,7 @@ template static __device__ __forceinlin const int kbx = k / QI6_K; // == 0 if QK_K == 256 const int kqsx = k % QI6_K; // == k if QK_K == 256 - const block_q6_K * bx0 = (block_q6_K *) vx; + const block_q6_K * bx0 = (const block_q6_K *) vx; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps) { @@ -3444,6 +3502,7 @@ template static __device__ __forceinlin static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat( const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc, const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) { + (void)x_qh; const float * x_dmf = (const float *) x_dm; const float * y_df = (const float *) y_ds; @@ -3486,7 +3545,7 @@ static __device__ __forceinline__ void mul_mat_q( __shared__ int tile_y_qs[mmq_x * WARP_SIZE]; __shared__ half2 tile_y_ds[mmq_x * WARP_SIZE/QI8_1]; - float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {0.0f}; + float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}}; for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) { @@ -4245,7 +4304,7 @@ template static __global__ void template static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows) { - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row >= nrows) { return; @@ -4285,7 +4344,7 @@ template static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) { // qk = quantized weights per x block // qr = number of quantized weights per data value in x block - const int row = blockIdx.y*blockDim.y + threadIdx.y; + const int row = blockIdx.x*blockDim.y + threadIdx.y; if (row >= nrows) { return; @@ -4459,6 +4518,13 @@ static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) { *dsti = __float2half(*xi); } +static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) { + const half * xi = (const half *) cxi; + half * dsti = (half *) cdsti; + + *dsti = *xi; +} + template static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, @@ -4484,11 +4550,41 @@ static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne, cpy_1(cx + x_offset, cdst + dst_offset); } -// rope == RoPE == rotary positional embedding +static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / max(0.001f, high - low); + return 1.0f - min(1.0f, max(0.0f, y)); +} +struct rope_corr_dims { + float v[4]; +}; + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static __device__ void rope_yarn( + float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +// rope == RoPE == rotary positional embedding template -static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale) { +static __global__ void rope( + const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + float ext_factor, float attn_factor, rope_corr_dims corr_dims +) { const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); if (col >= ncols) { @@ -4500,10 +4596,10 @@ static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t const int i2 = row/p_delta_rows; const int p = has_pos ? pos[i2] : 0; - const float p0 = p*freq_scale; - const float theta = p0*powf(theta_scale, col/2); - const float sin_theta = sinf(theta); - const float cos_theta = cosf(theta); + const float theta_base = p*powf(freq_base, -float(col)/ncols); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta); const float x0 = x[i + 0]; const float x1 = x[i + 1]; @@ -4513,8 +4609,10 @@ static __global__ void rope(const T * x, T * dst, const int ncols, const int32_t } template -static __global__ void rope_neox(const T * x, T * dst, const int ncols, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale) { +static __global__ void rope_neox( + const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows, + float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims +) { const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y); if (col >= ncols) { @@ -4522,24 +4620,31 @@ static __global__ void rope_neox(const T * x, T * dst, const int ncols, const in } const int row = blockDim.x*blockIdx.x + threadIdx.x; - const int i = row*ncols + col/2; + const int ib = col / n_dims; + const int ic = col % n_dims; + + const int i = row*ncols + ib*n_dims + ic/2; const int i2 = row/p_delta_rows; + float cur_rot = inv_ndims * ic - ib; + const int p = has_pos ? pos[i2] : 0; - const float p0 = p*freq_scale; - const float theta = p0*powf(theta_scale, col/2); - const float sin_theta = sinf(theta); - const float cos_theta = cosf(theta); + const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f); + + float cos_theta, sin_theta; + rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); const float x0 = x[i + 0]; - const float x1 = x[i + ncols/2]; + const float x1 = x[i + n_dims/2]; - dst[i + 0] = x0*cos_theta - x1*sin_theta; - dst[i + ncols/2] = x0*sin_theta + x1*cos_theta; + dst[i + 0] = x0*cos_theta - x1*sin_theta; + dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; } -static __global__ void rope_glm_f32(const float * x, float * dst, const int ncols, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale, const int n_ctx) { +static __global__ void rope_glm_f32( + const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base, + int n_ctx +) { const int col = blockDim.x*blockIdx.x + threadIdx.x; const int half_n_dims = ncols/4; @@ -4551,7 +4656,7 @@ static __global__ void rope_glm_f32(const float * x, float * dst, const int ncol const int i = row*ncols + col; const int i2 = row/p_delta_rows; - const float col_theta_scale = powf(theta_scale, col); + const float col_theta_scale = powf(freq_base, -2.0f*col/ncols); // FIXME: this is likely wrong const int p = pos != nullptr ? pos[i2] : 0; @@ -4675,6 +4780,25 @@ static __global__ void clamp_f32(const float * x, float * dst, const float min, dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]); } +static __global__ void im2col_f32_f16( + const float * x, half * dst, + int ofs0, int ofs1, int IW, int IH, int CHW, + int s0, int s1, int p0, int p1, int d0, int d1) { + const int iiw = blockIdx.z * s0 + threadIdx.z * d0 - p0; + const int iih = blockIdx.y * s1 + threadIdx.y * d1 - p1; + + const int offset_dst = + (threadIdx.x * gridDim.y * gridDim.z + blockIdx.y * gridDim.z + blockIdx.z) * CHW + + (blockIdx.x * (blockDim.y * blockDim.z) + threadIdx.y * blockDim.z + threadIdx.z); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = __float2half(0.0f); + } else { + const int offset_src = threadIdx.x * ofs0 + blockIdx.x * ofs1; + dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]); + } +} + template static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) { const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1); @@ -4693,6 +4817,11 @@ static void add_f16_f32_f16_cuda(const half * x, const float * y, half * dst, co add_f16_f32_f16<<>>(x, y, dst, k); } +static void add_f16_f32_f32_cuda(const half * x, const float * y, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_ADD_BLOCK_SIZE - 1) / CUDA_ADD_BLOCK_SIZE; + add_f16_f32_f32<<>>(x, y, dst, k); +} + static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) { const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE; mul_f32<<>>(x, y, dst, kx, ky); @@ -4708,6 +4837,16 @@ static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_ silu_f32<<>>(x, dst, k); } +static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE; + relu_f32<<>>(x, dst, k); +} + +static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) { + const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE; + sqr_f32<<>>(x, dst, k); +} + static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % WARP_SIZE == 0); if (ncols < 1024) { @@ -4816,7 +4955,8 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec <<>>(vx, y, dst, ncols, nrows); @@ -4825,7 +4965,7 @@ static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec <<>>(vx, y, dst, ncols, nrows); @@ -4834,7 +4974,7 @@ static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec <<>>(vx, y, dst, ncols, nrows); @@ -4843,7 +4983,7 @@ static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec <<>>(vx, y, dst, ncols, nrows); @@ -4852,7 +4992,7 @@ static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec <<>>(vx, y, dst, ncols, nrows); @@ -4862,7 +5002,7 @@ static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, f GGML_ASSERT(ncols % QK_K == 0); const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2 const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(32, ny, 1); dequantize_mul_mat_vec_q2_k<<>>(vx, y, dst, ncols, nrows); } @@ -4871,7 +5011,7 @@ static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, f GGML_ASSERT(ncols % QK_K == 0); const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(32, ny, 1); dequantize_mul_mat_vec_q3_k<<>>(vx, y, dst, ncols, nrows); } @@ -4880,7 +5020,7 @@ static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, f GGML_ASSERT(ncols % QK_K == 0); const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(32, ny, 1); dequantize_mul_mat_vec_q4_k<<>>(vx, y, dst, ncols, nrows); } @@ -4895,7 +5035,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f GGML_ASSERT(ncols % QK_K == 0); const int ny = 2 / K_QUANTS_PER_ITERATION; const int block_num_y = (nrows + ny - 1) / ny; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(32, ny, 1); dequantize_mul_mat_vec_q6_k<<>>(vx, y, dst, ncols, nrows); } @@ -4903,7 +5043,7 @@ static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, f static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK4_0 == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4912,7 +5052,7 @@ static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK4_1 == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4921,7 +5061,7 @@ static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK5_0 == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4930,7 +5070,7 @@ static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK5_1 == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4939,7 +5079,7 @@ static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK8_0 == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4948,7 +5088,7 @@ static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4957,7 +5097,7 @@ static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4966,7 +5106,7 @@ static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4975,7 +5115,7 @@ static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -4984,7 +5124,7 @@ static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float * static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % QK_K == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); mul_mat_vec_q <<>>(vx, vy, dst, ncols, nrows); @@ -5003,7 +5143,7 @@ static void convert_fp32_to_fp16_cuda(const void * vx, half * y, const int k, cu static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) { GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0); const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y; - const dim3 block_nums(1, block_num_y, 1); + const dim3 block_nums(block_num_y, 1, 1); const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1); dequantize_mul_mat_vec<1, 1, convert_f16> <<>>(vx, y, dst, ncols, nrows); @@ -5559,6 +5699,16 @@ static void ggml_cpy_f32_f16_cuda( (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12); } +static void ggml_cpy_f16_f16_cuda( + const char * cx, char * cdst, const int ne, + const int ne00, const int ne01, const int nb00, const int nb01, const int nb02, + const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) { + + const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE; + cpy_f32_f16<<>> + (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12); +} + static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) { const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE; scale_f32<<>>(x, dst, scale, k); @@ -5570,40 +5720,60 @@ static void clamp_f32_cuda(const float * x, float * dst, const float min, const } template -static void rope_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale, cudaStream_t stream) { +static void rope_cuda( + const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream +) { GGML_ASSERT(ncols % 2 == 0); const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); const dim3 block_nums(nrows, num_blocks_x, 1); if (pos == nullptr) { - rope<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale); + rope<<>>( + x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims + ); } else { - rope<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale); + rope<<>>( + x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims + ); } } template -static void rope_neox_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale, cudaStream_t stream) { +static void rope_neox_cuda( + const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream +) { GGML_ASSERT(ncols % 2 == 0); const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1); const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE); const dim3 block_nums(nrows, num_blocks_x, 1); + + const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.0f / n_dims; + if (pos == nullptr) { - rope_neox<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale); + rope_neox<<>>( + x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, + theta_scale, inv_ndims + ); } else { - rope_neox<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale); + rope_neox<<>>( + x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims, + theta_scale, inv_ndims + ); } } -static void rope_glm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale, - const int p_delta_rows, const float theta_scale, const int n_ctx, cudaStream_t stream) { +static void rope_glm_f32_cuda( + const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows, + float freq_base, int n_ctx, cudaStream_t stream +) { GGML_ASSERT(ncols % 4 == 0); const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1); const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE; const dim3 block_nums(num_blocks_x, nrows, 1); - rope_glm_f32<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, theta_scale, n_ctx); + rope_glm_f32<<>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx); } static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, @@ -5628,6 +5798,15 @@ static void soft_max_f32_cuda(const float * x, float * dst, const int ncols_x, c soft_max_f32<<>>(x, dst, ncols_x); } +static void im2col_f32_f16_cuda(const float * x, half * dst, + int OH, int IW, int IH, int OW, int IC, + int KH, int KW, int N, int ofs0, int ofs1, + int s0, int s1, int p0, int p1, int d0, int d1, cudaStream_t stream) { + dim3 block_nums(IC, OH, OW); + dim3 block_dims(N, KH, KW); + im2col_f32_f16<<>>(x, dst, ofs0, ofs1, IW, IH, (IC * KH * KW), s0, s1, p0, p1, d0, d1); +} + // buffer pool for cuda #define MAX_CUDA_BUFFERS 256 @@ -5696,7 +5875,7 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) { return ptr; } #ifdef DEBUG_CUDA_MALLOC - fprintf(stderr, "%s: %d buffers, max_size = %u MB, tot_size = %u MB, requested %u MB\n", __func__, nnz, + fprintf(stderr, "%s: %d buffers, max_size = %u MiB, tot_size = %u MiB, requested %u MiB\n", __func__, nnz, (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024)); #endif void * ptr; @@ -5724,6 +5903,11 @@ static void ggml_cuda_pool_free(void * ptr, size_t size) { CUDA_CHECK(cudaFree(ptr)); } +static bool g_cublas_loaded = false; + +bool ggml_cublas_loaded(void) { + return g_cublas_loaded; +} void ggml_init_cublas() { static bool initialized = false; @@ -5737,7 +5921,12 @@ void ggml_init_cublas() { CUDA_CHECK(cudaDeviceSynchronize()); #endif - CUDA_CHECK(cudaGetDeviceCount(&g_device_count)); + if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) { + initialized = true; + g_cublas_loaded = false; + return; + } + GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES); int64_t total_vram = 0; #if defined(GGML_CUDA_FORCE_MMQ) @@ -5785,6 +5974,7 @@ void ggml_init_cublas() { // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr)); initialized = true; + g_cublas_loaded = true; } } @@ -5823,7 +6013,7 @@ void * ggml_cuda_host_malloc(size_t size) { // The allocation error can be bypassed. A null ptr will assigned out of this function. // This can fixed the OOM error in WSL. cudaGetLastError(); - fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n", + fprintf(stderr, "WARNING: failed to allocate %.2f MiB of pinned memory: %s\n", size/1024.0/1024.0, cudaGetErrorString(err)); return nullptr; } @@ -5868,18 +6058,18 @@ static cudaError_t ggml_cuda_cpy_tensor_2d( const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3; if (nb0 == ts && nb1 == ts*ne0/bs) { return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream); - } else if (nb0 == ts) { - return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream); - } else { - for (int64_t i1 = 0; i1 < i1_diff; i1++) { - const void * rx = (const void *) ((const char *) x + i1*nb1); - void * rd = (void *) (dst_ptr + i1*ts*ne0/bs); - // pretend the row is a matrix with cols=1 - cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream); - if (r != cudaSuccess) return r; - } - return cudaSuccess; } + if (nb0 == ts) { + return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream); + } + for (int64_t i1 = 0; i1 < i1_diff; i1++) { + const void * rx = (const void *) ((const char *) x + i1*nb1); + void * rd = (void *) (dst_ptr + i1*ts*ne0/bs); + // pretend the row is a matrix with cols=1 + cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream); + if (r != cudaSuccess) { return r; } + } + return cudaSuccess; } static void ggml_cuda_op_repeat( @@ -5996,7 +6186,10 @@ inline void ggml_cuda_op_add( add_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(src0), ne10*ne11, main_stream); } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) { add_f16_f32_f16_cuda((const half *) src0_dd, src1_dd, (half *) dst_dd, ggml_nelements(src0), main_stream); + } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) { + add_f16_f32_f32_cuda((const half *) src0_dd, src1_dd, dst_dd, ggml_nelements(src0), main_stream); } else { + fprintf(stderr, "src0->type: %d dst->type: %d\n", src0->type, dst->type); GGML_ASSERT(false); } @@ -6048,6 +6241,34 @@ inline void ggml_cuda_op_silu( (void) src1_dd; } +inline void ggml_cuda_op_relu( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + +inline void ggml_cuda_op_sqr( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F32); + + sqr_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream); + + (void) src1; + (void) dst; + (void) src1_dd; +} + inline void ggml_cuda_op_norm( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { @@ -6170,6 +6391,7 @@ static int64_t get_row_rounding(ggml_type type) { case GGML_TYPE_Q8_0: return max_compute_capability >= CC_RDNA2 ? 128 : 64; case GGML_TYPE_F16: + case GGML_TYPE_F32: return 1; case GGML_TYPE_Q2_K: return max_compute_capability >= CC_RDNA2 ? 128 : 32; @@ -6192,6 +6414,7 @@ static int64_t get_row_rounding(ggml_type type) { case GGML_TYPE_Q8_0: return 64; case GGML_TYPE_F16: + case GGML_TYPE_F32: return 1; case GGML_TYPE_Q2_K: case GGML_TYPE_Q3_K: @@ -6383,8 +6606,7 @@ inline void ggml_cuda_op_mul_mat_cublas( src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &src1_as); to_fp16_cuda(src1_ddf_i, src1_as_f16, ne, stream); } - const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16; - + const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16; size_t dst_as = 0; half * dst_f16 = (half *) ggml_cuda_pool_malloc(row_diff*src1_ncols * sizeof(half), &dst_as); @@ -6460,17 +6682,20 @@ inline void ggml_cuda_op_rope( const int64_t ne2 = dst->ne[2]; const int64_t nrows = ggml_nrows(src0); - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + // RoPE alteration for extended context - - float freq_base, freq_scale; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); - - const float theta_scale = powf(freq_base, -2.0f/n_dims); + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); const int32_t * pos = nullptr; if ((mode & 1) == 0) { @@ -6482,24 +6707,38 @@ inline void ggml_cuda_op_rope( const bool is_neox = mode & 2; const bool is_glm = mode & 4; + rope_corr_dims corr_dims; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v); + // compute if (is_glm) { GGML_ASSERT(false); - rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, n_ctx, main_stream); + rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream); } else if (is_neox) { - GGML_ASSERT(ne00 == n_dims && "ne00 != n_dims is not implemented for CUDA yet"); if (src0->type == GGML_TYPE_F32) { - rope_neox_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream); + rope_neox_cuda( + (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); } else if (src0->type == GGML_TYPE_F16) { - rope_neox_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream); + rope_neox_cuda( + (const half *)src0_dd, (half *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); } else { GGML_ASSERT(false); } } else { if (src0->type == GGML_TYPE_F32) { - rope_cuda((const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream); + rope_cuda( + (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); } else if (src0->type == GGML_TYPE_F16) { - rope_cuda((const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, theta_scale, main_stream); + rope_cuda( + (const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor, + attn_factor, corr_dims, main_stream + ); } else { GGML_ASSERT(false); } @@ -6541,6 +6780,45 @@ inline void ggml_cuda_op_alibi( (void) src1_dd; } +inline void ggml_cuda_op_im2col( + const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, + const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { + + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + + const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1; + + const int64_t N = src1->ne[is_2D ? 3 : 2]; + const int64_t IC = src1->ne[is_2D ? 2 : 1]; + const int64_t IH = is_2D ? src1->ne[1] : 1; + const int64_t IW = src1->ne[0]; + + const int64_t KH = is_2D ? src0->ne[1] : 1; + const int64_t KW = src0->ne[0]; + + const int64_t OH = is_2D ? dst->ne[2] : 1; + const int64_t OW = dst->ne[1]; + + const size_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; // nb is byte offset, src is type float32 + const size_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32 + + im2col_f32_f16_cuda(src1_dd, (half*) dst_dd, + OH, IW, IH, OW, IC, KH, KW, N, + ofs0, ofs1, s0, s1, p0, p1, d0, d1, main_stream); + + (void) src0; + (void) src0_dd; +} + inline void ggml_cuda_op_diag_mask_inf( const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) { @@ -6610,8 +6888,10 @@ inline void ggml_cuda_op_clamp( GGML_ASSERT(src0->type == GGML_TYPE_F32); GGML_ASSERT( dst->type == GGML_TYPE_F32); - const float min = ((float *) dst->op_params)[0]; - const float max = ((float *) dst->op_params)[1]; + float min; + float max; + memcpy(&min, dst->op_params, sizeof(float)); + memcpy(&max, (float *) dst->op_params + 1, sizeof(float)); clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream); CUDA_CHECK(cudaGetLastError()); @@ -6743,7 +7023,7 @@ static void ggml_cuda_op_mul_mat( const int64_t ne01 = src0->ne[1]; const int64_t ne02 = src0->ne[2]; const int64_t ne03 = src0->ne[3]; - const int64_t nrows0 = ggml_nrows(src0); + // const int64_t nrows0 = ggml_nrows(src0); const int64_t ne10 = src1->ne[0]; const int64_t ne11 = src1->ne[1]; @@ -6804,6 +7084,8 @@ static void ggml_cuda_op_mul_mat( int64_t row_low[GGML_CUDA_MAX_DEVICES]; int64_t row_high[GGML_CUDA_MAX_DEVICES]; + int used_devices = 0; + for (int64_t id = 0; id < g_device_count; ++id) { // by default, use all rows row_low[id] = 0; @@ -6831,6 +7113,8 @@ static void ggml_cuda_op_mul_mat( continue; } + used_devices++; + const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device; const bool dst_on_device = dst->backend == GGML_BACKEND_GPU && id == g_main_device; @@ -6840,7 +7124,7 @@ static void ggml_cuda_op_mul_mat( if (src0_on_device && src0_is_contiguous) { src0_dd[id] = (char *) src0_extra->data_device[id]; } else { - const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); + // const size_t size_src0_ddq = split ? (row_high[id]-row_low[id])*ne00 * src0_ts/src0_bs : ggml_nbytes(src0); src0_dd[id] = (char *) ggml_cuda_pool_malloc(ggml_nbytes(src0), &src0_as[id]); } @@ -6869,12 +7153,12 @@ static void ggml_cuda_op_mul_mat( // if multiple devices are used they need to wait for the main device // here an event is recorded that signals that the main device has finished calculating the input data - if (split && g_device_count > 1) { + if (split && used_devices > 1) { CUDA_CHECK(ggml_cuda_set_device(g_main_device)); CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0])); } - const int64_t src1_col_stride = split && g_device_count > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; + const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11; for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) { const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0; const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride; @@ -6990,6 +7274,9 @@ static void ggml_cuda_op_mul_mat( } for (int64_t id = 0; id < g_device_count; ++id) { + if ((!split && id != g_main_device) || row_low[id] == row_high[id]) { + continue; + } CUDA_CHECK(ggml_cuda_set_device(id)); // free buffers again when done @@ -7014,6 +7301,9 @@ static void ggml_cuda_op_mul_mat( CUDA_CHECK(ggml_cuda_set_device(g_main_device)); for (int64_t id = 0; id < g_device_count; ++id) { + if (row_low[id] == row_high[id]) { + continue; + } for (int64_t is = 0; is < is_max; ++is) { CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0)); } @@ -7050,6 +7340,14 @@ static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, g ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu); } +static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu); +} + +static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr); +} + static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm); } @@ -7059,6 +7357,8 @@ static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src } bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) { + if (!g_cublas_loaded) { return false; } + const int64_t ne10 = src1->ne[0]; const int64_t ne0 = dst->ne[0]; @@ -7135,6 +7435,30 @@ static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream); } +__global__ static void k_compute_batched_ptrs( + const half * src0_as_f16, const half * src1_as_f16, half * dst_f16, + const void ** ptrs_src, void ** ptrs_dst, + int ne12, int ne13, + int ne23, + int nb02, int nb03, + int nb12, int nb13, + int nb2, int nb3, + int r2, int r3) { + int i13 = blockIdx.x * blockDim.x + threadIdx.x; + int i12 = blockIdx.y * blockDim.y + threadIdx.y; + + if (i13 >= ne13 || i12 >= ne12) { + return; + } + + int i03 = i13 / r3; + int i02 = i12 / r2; + + ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03; + ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12/2 + i13*nb13/2; + ptrs_dst[0*ne23 + i12 + i13*ne12] = ( char *) dst_f16 + i12* nb2/2 + i13* nb3/2; +} + static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { GGML_ASSERT(!ggml_is_transposed(src0)); GGML_ASSERT(!ggml_is_transposed(src1)); @@ -7236,49 +7560,45 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const CUBLAS_GEMM_DEFAULT_TENSOR_OP)); } else { // use cublasGemmBatchedEx - // TODO: https://github.com/ggerganov/llama.cpp/pull/3749#discussion_r1369997000 const int ne23 = ne12*ne13; - // TODO: avoid this alloc - void ** ptrs = (void **) malloc(3*ne23*sizeof(void *)); + const void ** ptrs_src = nullptr; + void ** ptrs_dst = nullptr; - for (int i13 = 0; i13 < ne13; ++i13) { - for (int i12 = 0; i12 < ne12; ++i12) { - int i03 = i13 / r3; - int i02 = i12 / r2; + size_t ptrs_src_s = 0; + size_t ptrs_dst_s = 0; - ptrs[0*ne23 + i12 + i13*ne12] = (char *) src0_as_f16 + i02*src0->nb[2] + i03*src0->nb[3]; - ptrs[1*ne23 + i12 + i13*ne12] = (char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2; - ptrs[2*ne23 + i12 + i13*ne12] = (char *) dst_f16 + i12* dst->nb[2]/2 + i13* dst->nb[3]/2; - } - } + ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s); + ptrs_dst = ( void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s); - // allocate device memory for pointers - void ** ptrs_as = nullptr; - CUDA_CHECK(cudaMalloc(&ptrs_as, 3*ne23*sizeof(void *))); - - // TODO: this does not work for some reason -- not sure why? - //size_t ptrs_s = 0; - //ptrs_as = (void **) ggml_cuda_pool_malloc(3*ne23*sizeof(void *), &ptrs_s); - - // copy pointers to device - CUDA_CHECK(cudaMemcpy(ptrs_as, ptrs, 3*ne23*sizeof(void *), cudaMemcpyHostToDevice)); - - free(ptrs); + dim3 block_dims(ne13, ne12); + k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>( + src0_as_f16, src1_as_f16, dst_f16, + ptrs_src, ptrs_dst, + ne12, ne13, + ne23, + nb02, nb03, + nb12, nb13, + dst->nb[2], dst->nb[3], + r2, r3); + CUDA_CHECK(cudaGetLastError()); CUBLAS_CHECK( cublasGemmBatchedEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N, ne01, ne11, ne10, - &alpha_f16, (const void **) (ptrs_as + 0*ne23), CUDA_R_16F, nb01/sizeof(half), - (const void **) (ptrs_as + 1*ne23), CUDA_R_16F, nb11/sizeof(float), - &beta_f16, ( void **) (ptrs_as + 2*ne23), CUDA_R_16F, ne01, + &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, nb01/sizeof(half), + (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, nb11/sizeof(float), + &beta_f16, ( void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01, ne23, CUBLAS_COMPUTE_16F, CUBLAS_GEMM_DEFAULT_TENSOR_OP)); - // free device memory for pointers - CUDA_CHECK(cudaFree(ptrs_as)); - //ggml_cuda_pool_free(ptrs_as, ptrs_s); + if (ptrs_src_s != 0) { + ggml_cuda_pool_free(ptrs_src, ptrs_src_s); + } + if (ptrs_dst_s != 0) { + ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s); + } } #endif @@ -7291,10 +7611,12 @@ static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { const bool all_on_device = - (src0->backend == GGML_BACKEND_GPU) && + (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) && (src1->backend == GGML_BACKEND_GPU) && ( dst->backend == GGML_BACKEND_GPU); + const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT; + int64_t min_compute_capability = INT_MAX; for (int64_t id = 0; id < g_device_count; ++id) { if (min_compute_capability > g_compute_capabilities[id] && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) { @@ -7316,13 +7638,13 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1 //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name); //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name); - if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { + if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) { // KQ single-batch ggml_cuda_mul_mat_vec_p021(src0, src1, dst); - } else if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { + } else if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) { // KQV single-batch ggml_cuda_mul_mat_vec_nc(src0, src1, dst); - } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { + } else if (!split && all_on_device && use_tensor_cores && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) { // KQ + KQV multi-batch ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst); } else if (src0->type == GGML_TYPE_F32) { @@ -7409,6 +7731,9 @@ static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, gg } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) { ggml_cpy_f32_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream); + } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) { + ggml_cpy_f16_f16_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, + ne10, ne11, nb10, nb11, nb12, main_stream); } else { fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type), ggml_type_name(src1->type)); @@ -7440,6 +7765,10 @@ static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi); } +static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { + ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col); +} + static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) { (void) src0; (void) src1; @@ -7551,11 +7880,11 @@ static size_t g_temp_tensor_extra_index = 0; static ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() { if (g_temp_tensor_extras == nullptr) { - g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_MAX_NODES]; + g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES]; } size_t alloc_index = g_temp_tensor_extra_index; - g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_MAX_NODES; + g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES; ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index]; memset(extra, 0, sizeof(*extra)); @@ -7722,6 +8051,8 @@ void ggml_cuda_free_scratch() { } bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) { + if (!g_cublas_loaded) { return false; } + ggml_cuda_func_t func; const bool any_on_device = tensor->backend == GGML_BACKEND_GPU || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) @@ -7731,6 +8062,15 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_ return false; } + if (tensor->op == GGML_OP_MUL_MAT) { + if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) { +#ifndef NDEBUG + fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = " PRId64 ", src1->ne[3] = " PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]); +#endif + return false; + } + } + switch (tensor->op) { case GGML_OP_REPEAT: func = ggml_cuda_repeat; @@ -7755,6 +8095,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_UNARY_OP_SILU: func = ggml_cuda_silu; break; + case GGML_UNARY_OP_RELU: + func = ggml_cuda_relu; + break; default: return false; } break; @@ -7773,6 +8116,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_OP_SCALE: func = ggml_cuda_scale; break; + case GGML_OP_SQR: + func = ggml_cuda_sqr; + break; case GGML_OP_CLAMP: if (!any_on_device) { return false; @@ -7803,6 +8149,9 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_ case GGML_OP_ALIBI: func = ggml_cuda_alibi; break; + case GGML_OP_IM2COL: + func = ggml_cuda_im2col; + break; default: return false; } @@ -7862,11 +8211,11 @@ struct ggml_backend_buffer_context_cuda { ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() { if (temp_tensor_extras == nullptr) { - temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_MAX_NODES]; + temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES]; } size_t alloc_index = temp_tensor_extra_index; - temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_MAX_NODES; + temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES; ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index]; memset(extra, 0, sizeof(*extra)); @@ -7952,7 +8301,12 @@ static ggml_backend_buffer_t ggml_backend_cuda_alloc_buffer(ggml_backend_t backe ggml_cuda_set_device(g_main_device); ggml_backend_buffer_context_cuda * ctx = new ggml_backend_buffer_context_cuda; + + size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0 + + ggml_cuda_set_device(g_main_device); CUDA_CHECK(cudaMalloc(&ctx->device, size)); + return ggml_backend_buffer_init(backend, cuda_backend_buffer_interface, ctx, size); } @@ -7996,14 +8350,14 @@ static ggml_backend_graph_plan_t ggml_backend_cuda_graph_plan_create(ggml_backen UNUSED(cgraph); } -static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +[[noreturn]] static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { GGML_ASSERT(!"not implemented"); UNUSED(backend); UNUSED(plan); } -static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { +[[noreturn]] static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) { GGML_ASSERT(!"not implemented"); UNUSED(backend); @@ -8019,6 +8373,9 @@ static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph for (int i = 0; i < cgraph->n_nodes; i++) { ggml_tensor * node = cgraph->nodes[i]; + if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE) { + continue; + } assert(node->backend == GGML_BACKEND_GPU); for (int j = 0; j < GGML_MAX_SRC; j++) { if (node->src[j] != nullptr) { diff --git a/ggml-cuda.h b/ggml-cuda.h index 57adc9cf3..528e66c33 100644 --- a/ggml-cuda.h +++ b/ggml-cuda.h @@ -17,7 +17,12 @@ extern "C" { #define GGML_CUDA_MAX_DEVICES 16 +// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`. GGML_API void ggml_init_cublas(void); + +// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`. +GGML_API bool ggml_cublas_loaded(void); + GGML_API void * ggml_cuda_host_malloc(size_t size); GGML_API void ggml_cuda_host_free(void * ptr); diff --git a/ggml-impl.h b/ggml-impl.h index 5ec18a50c..06c07339e 100644 --- a/ggml-impl.h +++ b/ggml-impl.h @@ -39,12 +39,6 @@ extern "C" { #endif #endif -#undef MIN -#undef MAX - -#define MIN(a, b) ((a) < (b) ? (a) : (b)) -#define MAX(a, b) ((a) > (b) ? (a) : (b)) - // 16-bit float // on Arm, we use __fp16 // on x86, we use uint16_t @@ -230,7 +224,19 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { #endif - // TODO: backend v2 PR +#define GGML_HASHTABLE_FULL ((size_t)-1) +#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2) + +bool ggml_hash_contains (const struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted +size_t ggml_hash_find (const struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full +size_t ggml_hash_insert ( struct ggml_hash_set hash_set, struct ggml_tensor * key); + +// return index, asserts if table is full +size_t ggml_hash_find_or_insert( struct ggml_hash_set hash_set, struct ggml_tensor * key); #ifdef __cplusplus } diff --git a/ggml-metal.h b/ggml-metal.h index 096b844e3..be2731f8b 100644 --- a/ggml-metal.h +++ b/ggml-metal.h @@ -26,7 +26,7 @@ #include // max memory buffers that can be mapped to the device -#define GGML_METAL_MAX_BUFFERS 16 +#define GGML_METAL_MAX_BUFFERS 64 #define GGML_METAL_MAX_COMMAND_BUFFERS 32 struct ggml_tensor; diff --git a/ggml-metal.m b/ggml-metal.m index 2380c4310..a9fdd3903 100644 --- a/ggml-metal.m +++ b/ggml-metal.m @@ -1,5 +1,6 @@ #import "ggml-metal.h" +#import "ggml-backend-impl.h" #import "ggml.h" #import @@ -23,7 +24,7 @@ #define UNUSED(x) (void)(x) -#define GGML_MAX_CONCUR (2*GGML_MAX_NODES) +#define GGML_MAX_CONCUR (2*GGML_DEFAULT_GRAPH_SIZE) struct ggml_metal_buffer { const char * name; @@ -85,6 +86,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(rms_norm); GGML_METAL_DECL_KERNEL(norm); GGML_METAL_DECL_KERNEL(mul_mv_f32_f32); + GGML_METAL_DECL_KERNEL(mul_mv_f16_f16); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4); @@ -113,6 +115,7 @@ struct ggml_metal_context { GGML_METAL_DECL_KERNEL(rope_f32); GGML_METAL_DECL_KERNEL(rope_f16); GGML_METAL_DECL_KERNEL(alibi_f32); + GGML_METAL_DECL_KERNEL(im2col_f16); GGML_METAL_DECL_KERNEL(cpy_f32_f16); GGML_METAL_DECL_KERNEL(cpy_f32_f32); GGML_METAL_DECL_KERNEL(cpy_f16_f16); @@ -125,7 +128,7 @@ struct ggml_metal_context { // MSL code // TODO: move the contents here when ready // for now it is easier to work in a separate file -static NSString * const msl_library_source = @"see metal.metal"; +//static NSString * const msl_library_source = @"see metal.metal"; // Here to assist with NSBundle Path Hack @interface GGMLMetalClass : NSObject @@ -141,7 +144,8 @@ void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_dat ggml_metal_log_user_data = user_data; } -static void ggml_metal_log(enum ggml_log_level level, const char* format, ...){ +GGML_ATTRIBUTE_FORMAT(2, 3) +static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){ if (ggml_metal_log_callback != NULL) { va_list args; va_start(args, format); @@ -209,7 +213,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { } else { GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__); - NSString * sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + NSString * sourcePath; + NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"]; + if (ggmlMetalPathResources) { + sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"]; + } else { + sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"]; + } if (sourcePath == nil) { GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__); sourcePath = @"ggml-metal.metal"; @@ -238,14 +248,17 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { // load kernels { NSError * error = nil; -#define GGML_METAL_ADD_KERNEL(name) \ - ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ - ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ + + /* GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \ (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \ (int) ctx->pipeline_##name.threadExecutionWidth); \ + */ +#define GGML_METAL_ADD_KERNEL(name) \ + ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \ + ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \ if (error) { \ - GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ + GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \ return NULL; \ } @@ -277,6 +290,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(rms_norm); GGML_METAL_ADD_KERNEL(norm); GGML_METAL_ADD_KERNEL(mul_mv_f32_f32); + GGML_METAL_ADD_KERNEL(mul_mv_f16_f16); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4); @@ -307,6 +321,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { GGML_METAL_ADD_KERNEL(rope_f32); GGML_METAL_ADD_KERNEL(rope_f16); GGML_METAL_ADD_KERNEL(alibi_f32); + GGML_METAL_ADD_KERNEL(im2col_f16); GGML_METAL_ADD_KERNEL(cpy_f32_f16); GGML_METAL_ADD_KERNEL(cpy_f32_f32); GGML_METAL_ADD_KERNEL(cpy_f16_f16); @@ -325,15 +340,15 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) { // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) { if ([ctx->device supportsFamily:i]) { - GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - MTLGPUFamilyApple1 + 1, i); + GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i); break; } } - GGML_METAL_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); - GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); + GGML_METAL_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false"); + GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MiB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); if (ctx->device.maxTransferRate != 0) { - GGML_METAL_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); + GGML_METAL_LOG_INFO("%s: maxTransferRate = %8.2f MiB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0); } else { GGML_METAL_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__); } @@ -376,6 +391,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(rms_norm); GGML_METAL_DEL_KERNEL(norm); GGML_METAL_DEL_KERNEL(mul_mv_f32_f32); + GGML_METAL_DEL_KERNEL(mul_mv_f16_f16); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row); GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4); @@ -406,6 +422,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) { GGML_METAL_DEL_KERNEL(rope_f32); GGML_METAL_DEL_KERNEL(rope_f16); GGML_METAL_DEL_KERNEL(alibi_f32); + GGML_METAL_DEL_KERNEL(im2col_f16); GGML_METAL_DEL_KERNEL(cpy_f32_f16); GGML_METAL_DEL_KERNEL(cpy_f32_f32); GGML_METAL_DEL_KERNEL(cpy_f16_f16); @@ -463,6 +480,10 @@ static id ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru const int64_t tsize = ggml_nbytes(t); + if (t->buffer && t->buffer->backend && t->buffer->backend->context) { + ctx = t->buffer->backend->context; + } + // find the view that contains the tensor fully for (int i = 0; i < ctx->n_buffers; ++i) { const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data; @@ -520,11 +541,11 @@ bool ggml_metal_add_buffer( ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0); + GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_aligned / 1024.0 / 1024.0); return false; } - GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0); + GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB", __func__, name, size_aligned / 1024.0 / 1024.0); ++ctx->n_buffers; } else { @@ -544,11 +565,11 @@ bool ggml_metal_add_buffer( ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil]; if (ctx->buffers[ctx->n_buffers].metal == nil) { - GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0); + GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0); return false; } - GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i); + GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i); if (i + size_step < size) { GGML_METAL_LOG_INFO("\n"); } @@ -563,7 +584,7 @@ bool ggml_metal_add_buffer( ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0); if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) { - GGML_METAL_LOG_WARN(", warning: current allocated size is greater than the recommended max working set size\n", __func__); + GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__); } else { GGML_METAL_LOG_INFO("\n"); } @@ -741,6 +762,20 @@ void ggml_metal_graph_compute( struct ggml_tensor * src1 = gf->nodes[i]->src[1]; struct ggml_tensor * dst = gf->nodes[i]; + switch (dst->op) { + case GGML_OP_NONE: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_TRANSPOSE: + case GGML_OP_PERMUTE: + { + // noop -> next node + } continue; + default: + { + } break; + } + const int64_t ne00 = src0 ? src0->ne[0] : 0; const int64_t ne01 = src0 ? src0->ne[1] : 0; const int64_t ne02 = src0 ? src0->ne[2] : 0; @@ -794,14 +829,6 @@ void ggml_metal_graph_compute( //} switch (dst->op) { - case GGML_OP_NONE: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_TRANSPOSE: - case GGML_OP_PERMUTE: - { - // noop - } break; case GGML_OP_CONCAT: { const int64_t nb = ne00; @@ -998,11 +1025,15 @@ void ggml_metal_graph_compute( } break; case GGML_OP_SOFT_MAX: { - const int nth = MIN(32, ne00); + int nth = 32; // SIMD width if (ne00%4 == 0) { [encoder setComputePipelineState:ctx->pipeline_soft_max_4]; } else { + do { + nth *= 2; + } while (nth <= ne00 && nth <= 1024); + nth /= 2; [encoder setComputePipelineState:ctx->pipeline_soft_max]; } [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; @@ -1010,8 +1041,9 @@ void ggml_metal_graph_compute( [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2]; [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3]; [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4]; + [encoder setThreadgroupMemoryLength:GGML_PAD(nth/32*sizeof(float), 16) atIndex:0]; - [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; + [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; case GGML_OP_DIAG_MASK_INF: { @@ -1118,6 +1150,7 @@ void ggml_metal_graph_compute( switch (src0t) { case GGML_TYPE_F32: { + GGML_ASSERT(src1t == GGML_TYPE_F32); [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32]; nrows = 4; } break; @@ -1125,13 +1158,18 @@ void ggml_metal_graph_compute( { nth0 = 32; nth1 = 1; - if (ne11 * ne12 < 4) { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row]; - } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4]; - nrows = ne11; + if (src1t == GGML_TYPE_F32) { + if (ne11 * ne12 < 4) { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row]; + } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4]; + nrows = ne11; + } else { + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32]; + nrows = 4; + } } else { - [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32]; + [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f16]; nrows = 4; } } break; @@ -1321,7 +1359,7 @@ void ggml_metal_graph_compute( [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; [encoder setBytes:&eps length:sizeof( float) atIndex:4]; - [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0]; + [encoder setThreadgroupMemoryLength:GGML_PAD(nth/32*sizeof(float), 16) atIndex:0]; const int64_t nrows = ggml_nrows(src0); @@ -1340,7 +1378,7 @@ void ggml_metal_graph_compute( [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2]; [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3]; [encoder setBytes:&eps length:sizeof( float) atIndex:4]; - [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0]; + [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0]; const int64_t nrows = ggml_nrows(src0); @@ -1392,14 +1430,18 @@ void ggml_metal_graph_compute( const int nth = MIN(1024, ne00); - const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; + const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[3]; - float freq_base; - float freq_scale; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); switch (src0->type) { case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break; @@ -1407,33 +1449,90 @@ void ggml_metal_graph_compute( default: GGML_ASSERT(false); }; - [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; - [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; - [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; - [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; - [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:4]; - [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:5]; - [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:6]; - [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:7]; - [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; - [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; - [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; - [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:11]; - [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:12]; - [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:13]; - [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:14]; - [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:15]; - [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:16]; - [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:17]; - [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:18]; - [encoder setBytes:&n_past length:sizeof( int) atIndex:19]; - [encoder setBytes:&n_dims length:sizeof( int) atIndex:20]; - [encoder setBytes:&mode length:sizeof( int) atIndex:21]; - [encoder setBytes:&freq_base length:sizeof(float) atIndex:22]; - [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23]; + [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0]; + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:2]; + [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3]; + [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:4]; + [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:5]; + [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:6]; + [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:7]; + [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:8]; + [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:9]; + [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:10]; + [encoder setBytes:&ne0 length:sizeof( int64_t) atIndex:11]; + [encoder setBytes:&ne1 length:sizeof( int64_t) atIndex:12]; + [encoder setBytes:&ne2 length:sizeof( int64_t) atIndex:13]; + [encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:14]; + [encoder setBytes:&nb0 length:sizeof(uint64_t) atIndex:15]; + [encoder setBytes:&nb1 length:sizeof(uint64_t) atIndex:16]; + [encoder setBytes:&nb2 length:sizeof(uint64_t) atIndex:17]; + [encoder setBytes:&nb3 length:sizeof(uint64_t) atIndex:18]; + [encoder setBytes:&n_past length:sizeof( int) atIndex:19]; + [encoder setBytes:&n_dims length:sizeof( int) atIndex:20]; + [encoder setBytes:&mode length:sizeof( int) atIndex:21]; + [encoder setBytes:&n_orig_ctx length:sizeof( int) atIndex:22]; + [encoder setBytes:&freq_base length:sizeof( float) atIndex:23]; + [encoder setBytes:&freq_scale length:sizeof( float) atIndex:24]; + [encoder setBytes:&ext_factor length:sizeof( float) atIndex:25]; + [encoder setBytes:&attn_factor length:sizeof( float) atIndex:26]; + [encoder setBytes:&beta_fast length:sizeof( float) atIndex:27]; + [encoder setBytes:&beta_slow length:sizeof( float) atIndex:28]; [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)]; } break; + case GGML_OP_IM2COL: + { + GGML_ASSERT(src0->type == GGML_TYPE_F16); + GGML_ASSERT(src1->type == GGML_TYPE_F32); + GGML_ASSERT( dst->type == GGML_TYPE_F16); + + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; + + const int32_t N = src1->ne[is_2D ? 3 : 2]; + const int32_t IC = src1->ne[is_2D ? 2 : 1]; + const int32_t IH = is_2D ? src1->ne[1] : 1; + const int32_t IW = src1->ne[0]; + + const int32_t KH = is_2D ? src0->ne[1] : 1; + const int32_t KW = src0->ne[0]; + + const int32_t OH = is_2D ? dst->ne[2] : 1; + const int32_t OW = dst->ne[1]; + + const int32_t CHW = IC * KH * KW; + + const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4; + const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4; + + switch (src0->type) { + case GGML_TYPE_F32: GGML_ASSERT(false && "not implemented"); break; + case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_im2col_f16]; break; + default: GGML_ASSERT(false); + }; + + [encoder setBuffer:id_src1 offset:offs_src1 atIndex:0]; + [encoder setBuffer:id_dst offset:offs_dst atIndex:1]; + [encoder setBytes:&ofs0 length:sizeof( int32_t) atIndex:2]; + [encoder setBytes:&ofs1 length:sizeof( int32_t) atIndex:3]; + [encoder setBytes:&IW length:sizeof( int32_t) atIndex:4]; + [encoder setBytes:&IH length:sizeof( int32_t) atIndex:5]; + [encoder setBytes:&CHW length:sizeof( int32_t) atIndex:6]; + [encoder setBytes:&s0 length:sizeof( int32_t) atIndex:7]; + [encoder setBytes:&s1 length:sizeof( int32_t) atIndex:8]; + [encoder setBytes:&p0 length:sizeof( int32_t) atIndex:9]; + [encoder setBytes:&p1 length:sizeof( int32_t) atIndex:10]; + [encoder setBytes:&d0 length:sizeof( int32_t) atIndex:11]; + [encoder setBytes:&d1 length:sizeof( int32_t) atIndex:12]; + + [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)]; + } break; case GGML_OP_DUP: case GGML_OP_CPY: case GGML_OP_CONT: diff --git a/ggml-metal.metal b/ggml-metal.metal index f4b460564..5d1357cd7 100644 --- a/ggml-metal.metal +++ b/ggml-metal.metal @@ -184,36 +184,73 @@ kernel void kernel_soft_max( constant int64_t & ne00, constant int64_t & ne01, constant int64_t & ne02, - uint3 tgpig[[threadgroup_position_in_grid]], - uint3 tpitg[[thread_position_in_threadgroup]], - uint3 ntg[[threads_per_threadgroup]]) { - const int64_t i03 = tgpig[2]; - const int64_t i02 = tgpig[1]; - const int64_t i01 = tgpig[0]; + threadgroup float * buf [[threadgroup(0)]], + uint tgpig[[threadgroup_position_in_grid]], + uint tpitg[[thread_position_in_threadgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]], + uint tiisg[[thread_index_in_simdgroup]], + uint ntg[[threads_per_threadgroup]]) { + const int64_t i03 = (tgpig) / (ne02*ne01); + const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; + const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); device const float * psrc0 = src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; device float * pdst = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00; // parallel max - float lmax = tpitg[0] < ne00 ? psrc0[tpitg[0]] : -INFINITY; - for (int i00 = tpitg[0] + ntg[0]; i00 < ne00; i00 += ntg[0]) { + float lmax = tpitg < ne00 ? psrc0[tpitg] : -INFINITY; + + for (int i00 = tpitg + ntg; i00 < ne00; i00 += ntg) { lmax = MAX(lmax, psrc0[i00]); } - const float max = simd_max(lmax); + + float max = simd_max(lmax); + if (tiisg == 0) { + buf[sgitg] = max; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + // broadcast, simd group number is ntg / 32 + for (uint i = ntg / 32 / 2; i > 0; i /= 2) { + if (tpitg < i) { + buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]); + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + max = buf[0]; // parallel sum float lsum = 0.0f; - for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { + for (int i00 = tpitg; i00 < ne00; i00 += ntg) { const float exp_psrc0 = exp(psrc0[i00] - max); lsum += exp_psrc0; // Remember the result of exp here. exp is expensive, so we really do not - // whish to compute it twice. + // wish to compute it twice. pdst[i00] = exp_psrc0; } - const float sum = simd_sum(lsum); + float sum = simd_sum(lsum); + if (tiisg == 0) { + buf[sgitg] = sum; + } - for (int i00 = tpitg[0]; i00 < ne00; i00 += ntg[0]) { + threadgroup_barrier(mem_flags::mem_threadgroup); + + // broadcast, simd group number is ntg / 32 + for (uint i = ntg / 32 / 2; i > 0; i /= 2) { + if (tpitg < i) { + buf[tpitg] += buf[tpitg + i]; + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + sum = buf[0]; + + for (int i00 = tpitg; i00 < ne00; i00 += ntg) { pdst[i00] /= sum; } } @@ -224,37 +261,73 @@ kernel void kernel_soft_max_4( constant int64_t & ne00, constant int64_t & ne01, constant int64_t & ne02, - uint3 tgpig[[threadgroup_position_in_grid]], - uint3 tpitg[[thread_position_in_threadgroup]], - uint3 ntg[[threads_per_threadgroup]]) { - const int64_t i03 = tgpig[2]; - const int64_t i02 = tgpig[1]; - const int64_t i01 = tgpig[0]; + threadgroup float * buf [[threadgroup(0)]], + uint tgpig[[threadgroup_position_in_grid]], + uint tpitg[[thread_position_in_threadgroup]], + uint sgitg[[simdgroup_index_in_threadgroup]], + uint tiisg[[thread_index_in_simdgroup]], + uint ntg[[threads_per_threadgroup]]) { + const int64_t i03 = (tgpig) / (ne02*ne01); + const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01; + const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01); device const float4 * psrc4 = (device const float4 *)(src0 + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); device float4 * pdst4 = (device float4 *)(dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00); // parallel max - float4 lmax4 = tpitg[0] < ne00/4 ? psrc4[tpitg[0]] : -INFINITY; - for (int i00 = tpitg[0] + ntg[0]; i00 < ne00/4; i00 += ntg[0]) { + float4 lmax4 = tpitg < ne00/4 ? psrc4[tpitg] : -INFINITY; + + for (int i00 = tpitg + ntg; i00 < ne00/4; i00 += ntg) { lmax4 = fmax(lmax4, psrc4[i00]); } - float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); - const float max = simd_max(lmax); + const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3])); + float max = simd_max(lmax); + if (tiisg == 0) { + buf[sgitg] = max; + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + // broadcast, simd group number is ntg / 32 + for (uint i = ntg / 32 / 2; i > 0; i /= 2) { + if (tpitg < i) { + buf[tpitg] = MAX(buf[tpitg], buf[tpitg + i]); + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + max = buf[0]; // parallel sum float4 lsum4 = 0.0f; - for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) { + for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { const float4 exp_psrc4 = exp(psrc4[i00] - max); lsum4 += exp_psrc4; pdst4[i00] = exp_psrc4; } - float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3]; - const float sum = simd_sum(lsum); + const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3]; + float sum = simd_sum(lsum); + if (tiisg == 0) { + buf[sgitg] = sum; + } - for (int i00 = tpitg[0]; i00 < ne00/4; i00 += ntg[0]) { + threadgroup_barrier(mem_flags::mem_threadgroup); + + // broadcast, simd group number is ntg / 32 + for (uint i = ntg / 32 / 2; i > 0; i /= 2) { + if (tpitg < i) { + buf[tpitg] += buf[tpitg + i]; + } + } + + threadgroup_barrier(mem_flags::mem_threadgroup); + + sum = buf[0]; + + for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) { pdst4[i00] /= sum; } } @@ -274,7 +347,7 @@ kernel void kernel_diag_mask_inf( dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY; } else { dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00]; - } + } } kernel void kernel_diag_mask_inf_8( @@ -719,7 +792,7 @@ kernel void kernel_mul_mv_f32_f32( constant int64_t & ne0, constant int64_t & ne1, uint3 tgpig[[threadgroup_position_in_grid]], - uint tiisg[[thread_index_in_simdgroup]]) { + uint tiisg[[thread_index_in_simdgroup]]) { const int64_t r0 = tgpig.x; const int64_t rb = tgpig.y*N_F32_F32; @@ -771,6 +844,79 @@ kernel void kernel_mul_mv_f32_f32( } } +#define N_F16_F16 4 + +kernel void kernel_mul_mv_f16_f16( + device const char * src0, + device const char * src1, + device float * dst, + constant int64_t & ne00, + constant int64_t & ne01, + constant int64_t & ne02, + constant uint64_t & nb00, + constant uint64_t & nb01, + constant uint64_t & nb02, + constant int64_t & ne10, + constant int64_t & ne11, + constant int64_t & ne12, + constant uint64_t & nb10, + constant uint64_t & nb11, + constant uint64_t & nb12, + constant int64_t & ne0, + constant int64_t & ne1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint tiisg[[thread_index_in_simdgroup]]) { + + const int64_t r0 = tgpig.x; + const int64_t rb = tgpig.y*N_F16_F16; + const int64_t im = tgpig.z; + + device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02); + + if (ne00 < 128) { + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12); + + float sumf = 0; + for (int i = tiisg; i < ne00; i += 32) { + sumf += (half) x[i] * (half) y[i]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } else { + device const half4 * x4 = (device const half4 *)x; + for (int row = 0; row < N_F16_F16; ++row) { + int r1 = rb + row; + if (r1 >= ne11) { + break; + } + + device const half * y = (device const half *) (src1 + r1*nb11 + im*nb12); + device const half4 * y4 = (device const half4 *) y; + + float sumf = 0; + for (int i = tiisg; i < ne00/4; i += 32) { + for (int k = 0; k < 4; ++k) sumf += (half) x4[i][k] * y4[i][k]; + } + + float all_sum = simd_sum(sumf); + if (tiisg == 0) { + for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (half) x[i] * y[i]; + dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum; + } + } + } +} + kernel void kernel_mul_mv_f16_f32_1row( device const char * src0, device const char * src1, @@ -988,6 +1134,45 @@ kernel void kernel_alibi_f32( } } +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / max(0.001f, high - low); + return 1.0f - min(1.0f, max(0.0f, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, + thread float * cos_theta, thread float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * log(1.0f / freq_scale); + } + *cos_theta = cos(theta) * mscale; + *sin_theta = sin(theta) * mscale; +} + +// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get +// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))` +static float rope_yarn_corr_factor(int n_dims, int n_orig_ctx, float n_rot, float base) { + return n_dims * log(n_orig_ctx / (n_rot * 2 * M_PI_F)) / (2 * log(base)); +} + +static void rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2] +) { + // start and end correction dims + dims[0] = max(0.0f, floor(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_fast, freq_base))); + dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_orig_ctx, beta_slow, freq_base))); +} + typedef void (rope_t)( device const void * src0, device const int32_t * src1, @@ -1011,8 +1196,13 @@ typedef void (rope_t)( constant int & n_past, constant int & n_dims, constant int & mode, + constant int & n_orig_ctx, constant float & freq_base, constant float & freq_scale, + constant float & ext_factor, + constant float & attn_factor, + constant float & beta_fast, + constant float & beta_slow, uint tiitg[[thread_index_in_threadgroup]], uint3 tptg[[threads_per_threadgroup]], uint3 tgpig[[threadgroup_position_in_grid]]); @@ -1041,8 +1231,13 @@ kernel void kernel_rope( constant int & n_past, constant int & n_dims, constant int & mode, + constant int & n_orig_ctx, constant float & freq_base, constant float & freq_scale, + constant float & ext_factor, + constant float & attn_factor, + constant float & beta_fast, + constant float & beta_slow, uint tiitg[[thread_index_in_threadgroup]], uint3 tptg[[threads_per_threadgroup]], uint3 tgpig[[threadgroup_position_in_grid]]) { @@ -1052,19 +1247,22 @@ kernel void kernel_rope( const bool is_neox = mode & 2; + float corr_dims[2]; + rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); + device const int32_t * pos = src1; const int64_t p = pos[i2]; - const float theta_0 = freq_scale * (float)p; + const float theta_0 = (float)p; const float inv_ndims = -1.f/n_dims; if (!is_neox) { for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) { const float theta = theta_0 * pow(freq_base, inv_ndims*i0); - const float cos_theta = cos(theta); - const float sin_theta = sin(theta); + float cos_theta, sin_theta; + rope_yarn(theta, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); device T * dst_data = (device T *)((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -1079,9 +1277,12 @@ kernel void kernel_rope( for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 2*tiitg; ic < n_dims; ic += 2*tptg.x) { - const float theta = theta_0 * pow(freq_base, inv_ndims*ic - ib); - const float cos_theta = cos(theta); - const float sin_theta = sin(theta); + // simplified from `(ib * n_dims + ic) * inv_ndims` + const float cur_rot = inv_ndims*ic - ib; + + const float theta = theta_0 * pow(freq_base, cur_rot); + float cos_theta, sin_theta; + rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta); const int64_t i0 = ib*n_dims + ic/2; @@ -1101,6 +1302,39 @@ kernel void kernel_rope( template [[host_name("kernel_rope_f32")]] kernel rope_t kernel_rope; template [[host_name("kernel_rope_f16")]] kernel rope_t kernel_rope; +kernel void kernel_im2col_f16( + device const float * x, + device half * dst, + constant int32_t & ofs0, + constant int32_t & ofs1, + constant int32_t & IW, + constant int32_t & IH, + constant int32_t & CHW, + constant int32_t & s0, + constant int32_t & s1, + constant int32_t & p0, + constant int32_t & p1, + constant int32_t & d0, + constant int32_t & d1, + uint3 tgpig[[threadgroup_position_in_grid]], + uint3 tgpg[[threadgroups_per_grid]], + uint3 tpitg[[thread_position_in_threadgroup]], + uint3 ntg[[threads_per_threadgroup]]) { + const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0; + const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1; + + const int32_t offset_dst = + (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW + + (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]); + + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst[offset_dst] = 0.0f; + } else { + const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1; + dst[offset_dst] = x[offset_src + iih * IW + iiw]; + } +} + kernel void kernel_cpy_f16_f16( device const half * src0, device half * dst, diff --git a/ggml-quants.c b/ggml-quants.c index 721594467..7285d5f7f 100644 --- a/ggml-quants.c +++ b/ggml-quants.c @@ -14,32 +14,12 @@ // #include -#if !defined(__aarch64__) -inline static int32_t vaddvq_s16(int16x8_t v) { - return - (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + - (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + - (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + - (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); -} - -inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { - int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); - int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); - return vcombine_s16(a0, b0); -} - -inline static int32_t vaddvq_s32(int32x4_t v) { - return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); -} -#endif - #else #ifdef __wasm_simd128__ #include #else -#ifdef __POWER9_VECTOR__ +#if defined(__POWER9_VECTOR__) || defined(__powerpc64__) #include #undef bool #define bool _Bool @@ -47,13 +27,15 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #if defined(_MSC_VER) || defined(__MINGW32__) #include #else -#if !defined(__riscv) && !defined(__s390__) +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) +#if !defined(__riscv) #include #endif #endif #endif #endif #endif +#endif #ifdef __riscv_v_intrinsic #include @@ -61,6 +43,7 @@ inline static int32_t vaddvq_s32(int32x4_t v) { #undef MIN #undef MAX + #define MIN(a, b) ((a) < (b) ? (a) : (b)) #define MAX(a, b) ((a) > (b) ? (a) : (b)) @@ -283,9 +266,31 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) #if defined(__ARM_NEON) - #if !defined(__aarch64__) +// 64-bit compatibility + +// vaddvq_s16 +// vpaddq_s16 +// vaddvq_s32 +// vaddvq_f32 +// vmaxvq_f32 +// vcvtnq_s32_f32 + +inline static int32_t vaddvq_s16(int16x8_t v) { + return + (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) + + (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) + + (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) + + (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7); +} + +inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) { + int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a)); + int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b)); + return vcombine_s16(a0, b0); +} + inline static int32_t vaddvq_s32(int32x4_t v) { return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3); } @@ -311,6 +316,96 @@ inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) { return res; } +// vld1q_s16_x2 +// vld1q_u8_x2 +// vld1q_u8_x4 +// vld1q_s8_x2 +// vld1q_s8_x4 +// TODO: double-check these work correctly + +typedef struct ggml_int16x8x2_t { + int16x8_t val[2]; +} ggml_int16x8x2_t; + +inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) { + ggml_int16x8x2_t res; + + res.val[0] = vld1q_s16(ptr + 0); + res.val[1] = vld1q_s16(ptr + 8); + + return res; +} + +typedef struct ggml_uint8x16x2_t { + uint8x16_t val[2]; +} ggml_uint8x16x2_t; + +inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) { + ggml_uint8x16x2_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + + return res; +} + +typedef struct ggml_uint8x16x4_t { + uint8x16_t val[4]; +} ggml_uint8x16x4_t; + +inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) { + ggml_uint8x16x4_t res; + + res.val[0] = vld1q_u8(ptr + 0); + res.val[1] = vld1q_u8(ptr + 16); + res.val[2] = vld1q_u8(ptr + 32); + res.val[3] = vld1q_u8(ptr + 48); + + return res; +} + +typedef struct ggml_int8x16x2_t { + int8x16_t val[2]; +} ggml_int8x16x2_t; + +inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) { + ggml_int8x16x2_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + + return res; +} + +typedef struct ggml_int8x16x4_t { + int8x16_t val[4]; +} ggml_int8x16x4_t; + +inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) { + ggml_int8x16x4_t res; + + res.val[0] = vld1q_s8(ptr + 0); + res.val[1] = vld1q_s8(ptr + 16); + res.val[2] = vld1q_s8(ptr + 32); + res.val[3] = vld1q_s8(ptr + 48); + + return res; +} + +#else + +#define ggml_int16x8x2_t int16x8x2_t +#define ggml_uint8x16x2_t uint8x16x2_t +#define ggml_uint8x16x4_t uint8x16x4_t +#define ggml_int8x16x2_t int8x16x2_t +#define ggml_int8x16x4_t int8x16x4_t + +#define ggml_vld1q_s16_x2 vld1q_s16_x2 +#define ggml_vld1q_u8_x2 vld1q_u8_x2 +#define ggml_vld1q_u8_x4 vld1q_u8_x4 +#define ggml_vld1q_s8_x2 vld1q_s8_x2 +#define ggml_vld1q_s8_x4 vld1q_s8_x4 + #endif #endif @@ -716,6 +811,7 @@ void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) { __riscv_vse8_v_i8m1(y[i].qs , vs, vl); } #else + GGML_UNUSED(nb); // scalar quantize_row_q8_0_reference(x, y, k); #endif @@ -969,6 +1065,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) { y[i].s = sum*d; } #else + GGML_UNUSED(nb); // scalar quantize_row_q8_1_reference(x, y, k); #endif @@ -1271,7 +1368,12 @@ static float make_qkx2_quants(int n, int nmax, const float * restrict x, const f float max = x[0]; float sum_w = weights[0]; float sum_x = sum_w * x[0]; +#ifdef HAVE_BUGGY_APPLE_LINKER + // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7 + for (volatile int i = 1; i < n; ++i) { +#else for (int i = 1; i < n; ++i) { +#endif if (x[i] < min) min = x[i]; if (x[i] > max) max = x[i]; float w = weights[i]; @@ -3555,7 +3657,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t vzero = vdupq_n_s32(0); #endif - int8x16x2_t q2bytes; + ggml_int8x16x2_t q2bytes; uint8_t aux[16]; float sum = 0; @@ -3574,8 +3676,8 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri vst1q_u8(aux, scales); const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4); - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); - const int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t mins16 = {vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}; const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])), vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0]))); const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])), @@ -3603,7 +3705,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri #endif #define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\ - q8bytes = vld1q_s8_x2(q8); q8 += 32;\ + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\ q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\ q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\ MULTIPLY_ACCUM_WITH_SCALE((index)); @@ -3611,9 +3713,9 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - const uint8x16x2_t q2bits = vld1q_u8_x2(q2); q2 += 32; + const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32; - int8x16x2_t q8bytes = vld1q_s8_x2(q8); q8 += 32; + ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3)); MULTIPLY_ACCUM_WITH_SCALE(0); @@ -3947,7 +4049,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t vzero = vdupq_n_s32(0); #endif - int8x16x4_t q2bytes; + ggml_int8x16x4_t q2bytes; uint32_t aux32[2]; const uint8_t * scales = (const uint8_t *)aux32; @@ -3972,7 +4074,7 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t q2bits = vld1q_u8(q2); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3)); q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3)); @@ -4236,7 +4338,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3 = vshlq_n_u8(m0, 3); const int8_t m32 = 32; - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; float sum = 0; @@ -4248,9 +4350,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].hmask; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; int32_t isum = 0; @@ -4266,9 +4368,9 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - const uint8x16x2_t q3bits = vld1q_u8_x2(q3); q3 += 32; - const int8x16x4_t q8bytes_1 = vld1q_s8_x4(q8); q8 += 64; - const int8x16x4_t q8bytes_2 = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32; + const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64; + const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64; q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2); q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2); @@ -4770,7 +4872,7 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t m3b = vdupq_n_u8(0x3); const uint8x16_t mh = vdupq_n_u8(4); - int8x16x4_t q3bytes; + ggml_int8x16x4_t q3bytes; uint16_t aux16[2]; int8_t * scales = (int8_t *)aux16; @@ -4779,11 +4881,11 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri for (int i = 0; i < nb; ++i) { - uint8x16x4_t q3h; + ggml_uint8x16x4_t q3h; const uint8x8_t hbits = vld1_u8(x[i].hmask); const uint8x16_t q3bits = vld1q_u8(x[i].qs); - const int8x16x4_t q8bytes = vld1q_s8_x4(y[i].qs); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(y[i].qs); const uint16_t a = *(const uint16_t *)x[i].scales; aux16[0] = a & 0x0f0f; @@ -5132,8 +5234,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x2_t q4bytes; - int8x16x2_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x2_t q8bytes; float sumf = 0; @@ -5168,17 +5270,17 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/64; ++j) { - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); q4 += 32; + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32; #ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int32x4_t p1 = vdotq_s32(vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]); sumi1 += vaddvq_s32(p1) * scales[2*j+0]; - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); @@ -5186,7 +5288,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri sumi2 += vaddvq_s32(p2) * scales[2*j+1]; #else - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5195,7 +5297,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri vmull_s8(vget_high_s8(q4bytes.val[1]), vget_high_s8(q8bytes.val[1]))); sumi1 += vaddvq_s16(vaddq_s16(p0, p1)) * scales[2*j+0]; - q8bytes = vld1q_s8_x2(q8); q8 += 32; + q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32; q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4)); q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4)); const int16x8_t p2 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5510,8 +5612,8 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri float sumf = 0; - int8x16x2_t q4bytes; - int8x16x4_t q8bytes; + ggml_int8x16x2_t q4bytes; + ggml_int8x16x4_t q8bytes; float sum_mins = 0.f; @@ -5532,10 +5634,10 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const float d = y[i].d * (float)x[i].d[0]; - const uint8x16x2_t q4bits = vld1q_u8_x2(q4); + const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); #ifdef __ARM_FEATURE_DOTPROD - q8bytes = vld1q_s8_x4(q8); + q8bytes = ggml_vld1q_s8_x4(q8); q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); @@ -5549,7 +5651,7 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri const int32_t sumi2 = vaddvq_s32(p2) * scales[1]; #else - q8bytes = vld1q_s8_x4(q8); + q8bytes = ggml_vld1q_s8_x4(q8); q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[0], m4b)); q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8 (q4bits.val[1], m4b)); const int16x8_t p0 = vaddq_s16(vmull_s8(vget_low_s8 (q4bytes.val[0]), vget_low_s8 (q8bytes.val[0])), @@ -5783,7 +5885,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x4_t q5bytes; + ggml_int8x16x4_t q5bytes; float sumf = 0; @@ -5813,16 +5915,16 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8_t * restrict qh = x[i].qh; const int8_t * restrict q8 = y[i].qs; - uint8x16x2_t qhbits = vld1q_u8_x2(qh); + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); - uint8x16x4_t q5h; + ggml_uint8x16x4_t q5h; int32_t sumi = 0; for (int j = 0; j < QK_K/64; ++j) { - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); q5 += 32; - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32; + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -6216,8 +6318,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const int32x4_t mzero = vdupq_n_s32(0); #endif - int8x16x4_t q5bytes; - uint8x16x4_t q5h; + ggml_int8x16x4_t q5bytes; + ggml_uint8x16x4_t q5h; float sumf = 0; @@ -6232,8 +6334,8 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri const uint8x8_t qhbits = vld1_u8(qh); - const uint8x16x2_t q5bits = vld1q_u8_x2(q5); - const int8x16x4_t q8bytes = vld1q_s8_x4(q8); + const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); + const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1)); q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4)); @@ -6509,8 +6611,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { @@ -6522,9 +6624,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const int8_t * restrict scale = x[i].scales; - const int16x8x2_t q8sums = vld1q_s16_x2(y[i].bsums); + const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums); const int8x16_t scales = vld1q_s8(scale); - const int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; + const ggml_int16x8x2_t q6scales = {vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}; const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])), vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))), @@ -6536,9 +6638,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri for (int j = 0; j < QK_K/128; ++j) { - uint8x16x2_t qhbits = vld1q_u8_x2(qh); qh += 32; - uint8x16x4_t q6bits = vld1q_u8_x4(q6); q6 += 64; - int8x16x4_t q8bytes = vld1q_s8_x4(q8); q8 += 64; + ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32; + ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64; + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4); q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4); @@ -6581,7 +6683,7 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri scale += 2; #endif - q8bytes = vld1q_s8_x4(q8); q8 += 64; + q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64; shifted = vshrq_n_u8(qhbits.val[0], 4); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4); @@ -6985,8 +7087,8 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri const uint8x16_t mone = vdupq_n_u8(3); - int8x16x4_t q6bytes; - uint8x16x4_t q6h; + ggml_int8x16x4_t q6bytes; + ggml_uint8x16x4_t q6h; for (int i = 0; i < nb; ++i) { @@ -7000,9 +7102,9 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri int32_t isum = 0; - uint8x16_t qhbits = vld1q_u8(qh); - uint8x16x2_t q6bits = vld1q_u8_x2(q6); - int8x16x4_t q8bytes = vld1q_s8_x4(q8); + uint8x16_t qhbits = vld1q_u8(qh); + ggml_uint8x16x2_t q6bits = ggml_vld1q_u8_x2(q6); + ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4); uint8x16_t shifted = vshrq_n_u8(qhbits, 2); diff --git a/ggml.c b/ggml.c index 84407b122..f92292b39 100644 --- a/ggml.c +++ b/ggml.c @@ -1,4 +1,5 @@ #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows +#define _USE_MATH_DEFINES // For M_PI on MSVC #include "ggml-impl.h" #include "ggml-quants.h" @@ -99,6 +100,49 @@ typedef void * thread_ret_t; #include #endif +#if defined(__APPLE__) +#include +#endif + +#if (defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)) && \ + (!defined(TARGET_OS_TV) && !defined(TARGET_OS_WATCH)) + +#include + +void ggml_print_backtrace(void) { + /* + #include + #include + + void * trace[100]; + + int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0])); + + backtrace_symbols_fd(trace, nptrs, STDERR_FILENO); + */ + + // backtrack_symbols does not show line numbers, use gdb instead + char attach[32]; + snprintf(attach, sizeof(attach), "attach %d", getpid()); + int pid = fork(); + if (pid == 0) { + execlp("gdb", "gdb", "--batch", + "-ex", "set style enabled on", + "-ex", attach, + "-ex", "bt -frame-info source-and-location", + "-ex", "detach", + "-ex", "quit", + NULL); + } else { + waitpid(pid, NULL, 0); + } +} +#else +void ggml_print_backtrace(void) { + // platform not supported +} +#endif + /*#define GGML_PERF*/ #define GGML_DEBUG 0 #define GGML_GELU_FP16 @@ -227,6 +271,12 @@ inline static void * ggml_aligned_malloc(size_t size) { // floating point type used to accumulate sums typedef double ggml_float; +#undef MIN +#undef MAX + +#define MIN(a, b) ((a) < (b) ? (a) : (b)) +#define MAX(a, b) ((a) > (b) ? (a) : (b)) + // // global data // @@ -560,6 +610,18 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) { // simd mappings // +#if defined(__ARM_NEON) +#if !defined(__aarch64__) + +// 64-bit compatibility + +inline static float vaddvq_f32(float32x4_t v) { + return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3); +} + +#endif +#endif + // we define a common set of C macros which map to specific intrinsics based on the current architecture // we then implement the fundamental computation operations below using only these macros // adding support for new architectures requires to define the corresponding SIMD macros @@ -1351,6 +1413,7 @@ inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); } inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } +inline static void ggml_vec_leaky_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.1f*x[i]; } static const float GELU_COEF_A = 0.044715f; static const float GELU_QUICK_COEF = -1.702f; @@ -1571,13 +1634,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "ROPE_BACK", "ALIBI", "CLAMP", - "CONV_1D", - "CONV_1D_STAGE_0", - "CONV_1D_STAGE_1", "CONV_TRANSPOSE_1D", - "CONV_2D", - "CONV_2D_STAGE_0", - "CONV_2D_STAGE_1", + "IM2COL", "CONV_TRANSPOSE_2D", "POOL_1D", "POOL_2D", @@ -1608,7 +1666,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "CROSS_ENTROPY_LOSS_BACK", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "none", @@ -1658,13 +1716,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "rope_back(x)", "alibi(x)", "clamp(x)", - "conv_1d(x)", - "conv_1d_stage_0(x)", - "conv_1d_stage_1(x)", "conv_transpose_1d(x)", - "conv_2d(x)", - "conv_2d_stage_0(x)", - "conv_2d_stage_1(x)", + "im2col(x)", "conv_transpose_2d(x)", "pool_1d(x)", "pool_2d(x)", @@ -1695,7 +1748,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = { "cross_entropy_loss_back(x,y)", }; -static_assert(GGML_OP_COUNT == 73, "GGML_OP_COUNT != 73"); +static_assert(GGML_OP_COUNT == 68, "GGML_OP_COUNT != 68"); static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2"); @@ -1723,13 +1776,7 @@ static void ggml_setup_op_has_task_pass(void) { p[GGML_OP_GET_ROWS_BACK ] = true; p[GGML_OP_DIAG_MASK_INF ] = true; p[GGML_OP_DIAG_MASK_ZERO ] = true; - p[GGML_OP_CONV_1D ] = true; - p[GGML_OP_CONV_1D_STAGE_0 ] = true; - p[GGML_OP_CONV_1D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_1D ] = true; - p[GGML_OP_CONV_2D ] = true; - p[GGML_OP_CONV_2D_STAGE_0 ] = true; - p[GGML_OP_CONV_2D_STAGE_1 ] = true; p[GGML_OP_CONV_TRANSPOSE_2D ] = true; p[GGML_OP_FLASH_ATTN_BACK ] = true; p[GGML_OP_CROSS_ENTROPY_LOSS ] = true; @@ -3153,7 +3200,7 @@ static struct ggml_tensor * ggml_add_cast_impl( // TODO: support less-strict constraint // GGML_ASSERT(ggml_can_repeat(b, a)); GGML_ASSERT(ggml_can_repeat_rows(b, a)); - GGML_ASSERT(ggml_is_quantized(a->type)); // currently only supported for quantized input + GGML_ASSERT(ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16); // currently only supported for quantized input and f16 bool is_node = false; @@ -3768,6 +3815,14 @@ struct ggml_tensor * ggml_relu_inplace( return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_RELU); } +// ggml_leaky + +struct ggml_tensor * ggml_leaky( + struct ggml_context * ctx, + struct ggml_tensor * a) { + return ggml_unary(ctx, a, GGML_UNARY_OP_LEAKY); +} + // ggml_gelu struct ggml_tensor * ggml_gelu( @@ -4845,8 +4900,13 @@ static struct ggml_tensor * ggml_rope_impl( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, float xpos_base, bool xpos_down, bool inplace) { @@ -4862,11 +4922,15 @@ static struct ggml_tensor * ggml_rope_impl( struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a); - int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx }; - memcpy(params + 4, &freq_base, sizeof(float)); - memcpy(params + 5, &freq_scale, sizeof(float)); - memcpy(params + 6, &xpos_base, sizeof(float)); - memcpy(params + 7, &xpos_down, sizeof(bool)); + int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx }; + memcpy(params + 5, &freq_base, sizeof(float)); + memcpy(params + 6, &freq_scale, sizeof(float)); + memcpy(params + 7, &ext_factor, sizeof(float)); + memcpy(params + 8, &attn_factor, sizeof(float)); + memcpy(params + 9, &beta_fast, sizeof(float)); + memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, &xpos_base, sizeof(float)); + memcpy(params + 12, &xpos_down, sizeof(bool)); ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE; @@ -4884,7 +4948,9 @@ struct ggml_tensor * ggml_rope( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, false); + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false + ); } struct ggml_tensor * ggml_rope_inplace( @@ -4894,7 +4960,9 @@ struct ggml_tensor * ggml_rope_inplace( int n_dims, int mode, int n_ctx) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, 10000.0f, 1.0f, 0.0f, false, true); + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true + ); } struct ggml_tensor * ggml_rope_custom( @@ -4904,9 +4972,17 @@ struct ggml_tensor * ggml_rope_custom( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, false); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false + ); } struct ggml_tensor * ggml_rope_custom_inplace( @@ -4916,9 +4992,17 @@ struct ggml_tensor * ggml_rope_custom_inplace( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale) { - return ggml_rope_impl(ctx, a, b, n_dims, mode, n_ctx, freq_base, freq_scale, 0.0f, false, true); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow) { + return ggml_rope_impl( + ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true + ); } struct ggml_tensor * ggml_rope_xpos_inplace( @@ -4928,7 +5012,7 @@ struct ggml_tensor * ggml_rope_xpos_inplace( int n_dims, float base, bool down) { - return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 10000.0f, 1.0f, base, down, true); + return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true); } // ggml_rope_back @@ -4940,8 +5024,13 @@ struct ggml_tensor * ggml_rope_back( int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, float xpos_base, bool xpos_down) { GGML_ASSERT(ggml_is_vector(b)); @@ -4958,11 +5047,15 @@ struct ggml_tensor * ggml_rope_back( struct ggml_tensor * result = ggml_dup_tensor(ctx, a); - int32_t params[8] = { /*n_past*/ 0, n_dims, mode, n_ctx }; - memcpy(params + 4, &freq_base, sizeof(float)); - memcpy(params + 5, &freq_scale, sizeof(float)); - memcpy(params + 6, &xpos_base, sizeof(float)); - memcpy(params + 7, &xpos_down, sizeof(bool)); + int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx }; + memcpy(params + 5, &freq_base, sizeof(float)); + memcpy(params + 6, &freq_scale, sizeof(float)); + memcpy(params + 7, &ext_factor, sizeof(float)); + memcpy(params + 8, &attn_factor, sizeof(float)); + memcpy(params + 9, &beta_fast, sizeof(float)); + memcpy(params + 10, &beta_slow, sizeof(float)); + memcpy(params + 11, &xpos_base, sizeof(float)); + memcpy(params + 12, &xpos_down, sizeof(bool)); ggml_set_op_params(result, params, sizeof(params)); result->op = GGML_OP_ROPE_BACK; @@ -5037,82 +5130,6 @@ static int64_t ggml_calc_conv_output_size(int64_t ins, int64_t ks, int s, int p, return (ins + 2 * p - d * (ks - 1) - 1) / s + 1; } -// im2col: [N, IC, IL] => [N, OL, IC*K] -// a: [OC，IC, K] -// b: [N, IC, IL] -// result: [N, OL, IC*K] -static struct ggml_tensor * ggml_conv_1d_stage_0( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int p0, - int d0) { - GGML_ASSERT(a->ne[1] == b->ne[1]); - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t OL = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); - - const int64_t ne[4] = { - a->ne[1] * a->ne[0], - OL, - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - - int32_t params[] = { s0, p0, d0 }; - ggml_set_op_params(result, params, sizeof(params)); - - result->op = GGML_OP_CONV_1D_STAGE_0; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -// ggml_conv_1d_stage_1 - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// a: [OC, IC, K] -// b: [N, OL, IC * K] -// result: [N, OC, OL] -static struct ggml_tensor * ggml_conv_1d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - a->ne[2], - b->ne[2], - 1, - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_1D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; -} - -// ggml_conv_1d - GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -5120,44 +5137,18 @@ GGML_API struct ggml_tensor * ggml_conv_1d( int s0, int p0, int d0) { - struct ggml_tensor * result = ggml_conv_1d_stage_0(ctx, a, b, s0, p0, d0); - result = ggml_conv_1d_stage_1(ctx, a, result); + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false); // [N, OL, IC * K] + + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, im2col, im2col->ne[0], (im2col->ne[2] * im2col->ne[1])), // [N, OL, IC * K] => [N*OL, IC * K] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1]), a->ne[2])); // [OC，IC, K] => [OC, IC * K] + + result = ggml_reshape_3d(ctx, result, im2col->ne[1], a->ne[2], im2col->ne[2]); // [N, OC, OL] + return result; } -// GGML_API struct ggml_tensor * ggml_conv_1d( -// struct ggml_context * ctx, -// struct ggml_tensor * a, -// struct ggml_tensor * b, -// int s0, -// int p0, -// int d0) { -// GGML_ASSERT(ggml_is_matrix(b)); -// GGML_ASSERT(a->ne[1] == b->ne[1]); -// bool is_node = false; - -// if (a->grad || b->grad) { -// GGML_ASSERT(false); // TODO: implement backward -// is_node = true; -// } - -// const int64_t ne[4] = { -// ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0), -// a->ne[2], 1, 1, -// }; -// struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne); - -// int32_t params[] = { s0, p0, d0 }; -// ggml_set_op_params(result, params, sizeof(params)); - -// result->op = GGML_OP_CONV_1D; -// result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; -// result->src[0] = a; -// result->src[1] = b; - -// return result; -// } - // ggml_conv_1d_ph struct ggml_tensor* ggml_conv_1d_ph( @@ -5219,7 +5210,7 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d( // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OH, OW, IC*KH*KW] -static struct ggml_tensor * ggml_conv_2d_stage_0( +struct ggml_tensor * ggml_im2col( struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, @@ -5228,9 +5219,14 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( int p0, int p1, int d0, - int d1) { + int d1, + bool is_2D) { - GGML_ASSERT(a->ne[2] == b->ne[2]); + if(is_2D) { + GGML_ASSERT(a->ne[2] == b->ne[2]); + } else { + GGML_ASSERT(a->ne[1] == b->ne[1]); + } bool is_node = false; if (a->grad || b->grad) { @@ -5238,81 +5234,51 @@ static struct ggml_tensor * ggml_conv_2d_stage_0( is_node = true; } - const int64_t OH = ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1); - const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); + const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0; + const int64_t OW = ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0); const int64_t ne[4] = { - a->ne[2] * a->ne[1] * a->ne[0], + is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0], OW, - OH, - b->ne[3], + is_2D ? OH : b->ne[2], + is_2D ? b->ne[3] : 1, }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); - int32_t params[] = { s0, s1, p0, p1, d0, d1 }; + struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne); + int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) }; ggml_set_op_params(result, params, sizeof(params)); - result->op = GGML_OP_CONV_2D_STAGE_0; + result->op = GGML_OP_IM2COL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->src[0] = a; result->src[1] = b; return result; - -} - -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// a: [OC, IC, KH, KW] -// b: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static struct ggml_tensor * ggml_conv_2d_stage_1( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b) { - - bool is_node = false; - - if (a->grad || b->grad) { - GGML_ASSERT(false); // TODO: implement backward - is_node = true; - } - - const int64_t ne[4] = { - b->ne[1], - b->ne[2], - a->ne[3], - b->ne[3], - }; - struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne); - - result->op = GGML_OP_CONV_2D_STAGE_1; - result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; - result->src[0] = a; - result->src[1] = b; - - return result; - } // a: [OC，IC, KH, KW] // b: [N, IC, IH, IW] // result: [N, OC, OH, OW] struct ggml_tensor * ggml_conv_2d( - struct ggml_context * ctx, - struct ggml_tensor * a, - struct ggml_tensor * b, - int s0, - int s1, - int p0, - int p1, - int d0, - int d1) { + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1) { + struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true); // [N, OH, OW, IC * KH * KW] - struct ggml_tensor * result = ggml_conv_2d_stage_0(ctx, a, b, s0, s1, p0, p1, d0, d1); // [N, OH, OW, IC * KH * KW] - result = ggml_conv_2d_stage_1(ctx, a, result); + struct ggml_tensor * result = + ggml_mul_mat(ctx, + ggml_reshape_2d(ctx, im2col, im2col->ne[0], im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW] + ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]), a->ne[3])); // [OC，IC, KH, KW] => [OC, IC * KH * KW] + + result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], a->ne[3], im2col->ne[3]); // [N, OC, OH, OW] return result; - } // ggml_conv_2d_sk_p0 @@ -5372,7 +5338,7 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0( // ggml_pool_* -static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, int p) { +static int64_t ggml_calc_pool_output_size(int64_t ins, int ks, int s, float p) { return (ins + 2 * p - ks) / s + 1; } @@ -5419,8 +5385,8 @@ struct ggml_tensor * ggml_pool_2d( int k1, int s0, int s1, - int p0, - int p1) { + float p0, + float p1) { bool is_node = false; @@ -6927,9 +6893,15 @@ static void ggml_compute_forward_add_f16_f32( GGML_ASSERT(src0->type == GGML_TYPE_F16); GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT(dst->type == GGML_TYPE_F16); - GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + if (dst->type == GGML_TYPE_F32) { + GGML_ASSERT( nb0 == sizeof(float)); + } + else { + GGML_ASSERT(dst->type == GGML_TYPE_F16); + GGML_ASSERT( nb0 == sizeof(ggml_fp16_t)); + } + GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); // rows per thread @@ -6940,18 +6912,35 @@ static void ggml_compute_forward_add_f16_f32( const int ir1 = MIN(ir0 + dr, nr); if (nb10 == sizeof(float)) { - for (int ir = ir0; ir < ir1; ++ir) { - // src0, src1 and dst are same shape => same indices - const int i3 = ir/(ne2*ne1); - const int i2 = (ir - i3*ne2*ne1)/ne1; - const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + if (dst->type == GGML_TYPE_F16) { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); - ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); - ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); - float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + ggml_fp16_t * dst_ptr = (ggml_fp16_t *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); - for (int i = 0; i < ne0; i++) { - dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]); + } + } + } else { + for (int ir = ir0; ir < ir1; ++ir) { + // src0, src1 and dst are same shape => same indices + const int i3 = ir/(ne2*ne1); + const int i2 = (ir - i3*ne2*ne1)/ne1; + const int i1 = (ir - i3*ne2*ne1 - i2*ne1); + + float * dst_ptr = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1); + ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01); + float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11); + + for (int i = 0; i < ne0; i++) { + dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]; + } } } } @@ -8859,6 +8848,48 @@ static void ggml_compute_forward_silu( } } +// ggml_compute_forward_leaky + +static void ggml_compute_forward_leaky_f32( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + assert(params->ith == 0); + assert(ggml_are_same_shape(src0, dst)); + + if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { + return; + } + + const int n = ggml_nrows(src0); + const int nc = src0->ne[0]; + + assert(dst->nb[0] == sizeof(float)); + assert(src0->nb[0] == sizeof(float)); + + for (int i = 0; i < n; i++) { + ggml_vec_leaky_f32(nc, + (float *) ((char *) dst->data + i*( dst->nb[1])), + (float *) ((char *) src0->data + i*(src0->nb[1]))); + } +} + +static void ggml_compute_forward_leaky( + const struct ggml_compute_params * params, + const struct ggml_tensor * src0, + struct ggml_tensor * dst) { + switch (src0->type) { + case GGML_TYPE_F32: + { + ggml_compute_forward_leaky_f32(params, src0, dst); + } break; + default: + { + GGML_ASSERT(false); + } break; + } +} + // ggml_compute_forward_silu_back static void ggml_compute_forward_silu_back_f32( @@ -9342,6 +9373,8 @@ static bool ggml_compute_forward_mul_mat_use_blas( // TODO: find the optimal values for these if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && + src0->type == GGML_TYPE_F32 && + src1->type == GGML_TYPE_F32 && (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) { /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/ @@ -9380,7 +9413,7 @@ static void ggml_compute_forward_mul_mat( // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == ggml_type_size(type)); - GGML_ASSERT(nb10 == sizeof(float)); + GGML_ASSERT(nb10 == ggml_type_size(src1->type)); // dst cannot be transposed or permuted GGML_ASSERT(nb0 == sizeof(float)); @@ -9578,10 +9611,12 @@ static void ggml_compute_forward_out_prod_f32( const int ith = params->ith; const int nth = params->nth; + GGML_ASSERT(ne0 == ne00); + GGML_ASSERT(ne1 == ne10); + GGML_ASSERT(ne2 == ne02); GGML_ASSERT(ne02 == ne12); - GGML_ASSERT(ne03 == ne13); - GGML_ASSERT(ne2 == ne12); GGML_ASSERT(ne3 == ne13); + GGML_ASSERT(ne03 == ne13); // we don't support permuted src0 or src1 GGML_ASSERT(nb00 == sizeof(float)); @@ -9592,18 +9627,25 @@ static void ggml_compute_forward_out_prod_f32( // GGML_ASSERT(nb1 <= nb2); // GGML_ASSERT(nb2 <= nb3); - GGML_ASSERT(ne0 == ne00); - GGML_ASSERT(ne1 == ne10); - GGML_ASSERT(ne2 == ne02); - GGML_ASSERT(ne3 == ne03); - // nb01 >= nb00 - src0 is not transposed // compute by src0 rows // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod - // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST) + // TODO: #if defined(GGML_USE_CLBLAST) + +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + bool use_blas = ggml_is_matrix(src0) && + ggml_is_matrix(src1) && + ggml_is_contiguous(src0) && + (ggml_is_contiguous(src1) || ggml_is_transposed(src1)); +#endif if (params->type == GGML_TASK_INIT) { +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) // gemm beta will zero dst + if (use_blas) { + return; + } +#endif ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0); return; } @@ -9612,6 +9654,50 @@ static void ggml_compute_forward_out_prod_f32( return; } +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + if (use_blas) { + if (params->ith != 0) { // All threads other than the first do no work. + return; + } + // Arguments to ggml_compute_forward_out_prod (expressed as major,minor) + // src0: (k,n) + // src1: (k,m) + // dst: (m,n) + // + // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f) + // Also expressed as (major,minor) + // a: (m,k): so src1 transposed + // b: (k,n): so src0 + // c: (m,n) + // + // However, if ggml_is_transposed(src1) is true, then + // src1->data already contains a transposed version, so sgemm mustn't + // transpose it further. + + int n = src0->ne[0]; + int k = src0->ne[1]; + int m = src1->ne[0]; + + int transposeA, lda; + + if (!ggml_is_transposed(src1)) { + transposeA = CblasTrans; + lda = m; + } else { + transposeA = CblasNoTrans; + lda = k; + } + + float * a = (float *) ((char *) src1->data); + float * b = (float *) ((char *) src0->data); + float * c = (float *) ((char *) dst->data); + + cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n); + + return; + } +#endif + // dst[:,:,:,:] = 0 // for i2,i3: // for i1: @@ -10878,30 +10964,75 @@ static void ggml_compute_forward_clamp( // ggml_compute_forward_rope +static float rope_yarn_ramp(const float low, const float high, const int i0) { + const float y = (i0 / 2 - low) / MAX(0.001f, high - low); + return 1 - MIN(1, MAX(0, y)); +} + +// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn +// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng. +static void rope_yarn( + float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale, + float * cos_theta, float * sin_theta +) { + // Get n-d rotational scaling corrected for extrapolation + float theta_interp = freq_scale * theta_extrap; + float theta = theta_interp; + if (ext_factor != 0.0f) { + float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor; + theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix; + + // Get n-d magnitude scaling corrected for interpolation + mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale); + } + *cos_theta = cosf(theta) * mscale; + *sin_theta = sinf(theta) * mscale; +} + +// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get +// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))` +static float ggml_rope_yarn_corr_dim(int n_dims, int n_orig_ctx, float n_rot, float base) { + return n_dims * logf(n_orig_ctx / (n_rot * 2 * (float)M_PI)) / (2 * logf(base)); +} + +void ggml_rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2] +) { + // start and end correction dims + dims[0] = MAX(0, floorf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_fast, freq_base))); + dims[1] = MIN(n_dims - 1, ceilf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_slow, freq_base))); +} + static void ggml_compute_forward_rope_f32( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst, + const bool forward) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - float freq_base; - float freq_scale; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; // these two only relevant for xPos RoPE: float xpos_base; bool xpos_down; - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) dst->op_params + 7, sizeof(bool)); + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) dst->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) dst->op_params + 12, sizeof(bool)); GGML_TENSOR_UNARY_OP_LOCALS @@ -10929,10 +11060,18 @@ static void ggml_compute_forward_rope_f32( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + const int32_t * pos = (const int32_t *) src1->data; for (int64_t i3 = 0; i3 < ne3; i3++) { @@ -10942,18 +11081,18 @@ static void ggml_compute_forward_rope_f32( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = freq_scale * (float)p; + float theta_base = (float)p; if (is_glm) { - theta = MIN(p, n_ctx - 2); + theta_base = MIN(p, n_ctx - 2); float block_theta = MAX(p - (n_ctx - 2), 0); for (int64_t i0 = 0; i0 < ne0 / 4; i0++) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta_base); + const float sin_theta = sinf(theta_base) * sin_sign; const float cos_block_theta = cosf(block_theta); - const float sin_block_theta = sinf(block_theta); + const float sin_block_theta = sinf(block_theta) * sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -10971,13 +11110,17 @@ static void ggml_compute_forward_rope_f32( } } else if (!is_neox) { for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + // zeta scaling for xPos only: float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f; if (xpos_down) zeta = 1.0f / zeta; - theta *= theta_scale; + theta_base *= theta_scale; const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); float * dst_data = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -10991,12 +11134,20 @@ static void ggml_compute_forward_rope_f32( } else { // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + theta_base *= freq_scale; for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + // simplified from `(ib * n_dims + ic) * inv_ndims` + float cur_rot = inv_ndims * ic - ib; - theta *= theta_scale; + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + + theta_base *= theta_scale; const int64_t i0 = ib*n_dims + ic/2; @@ -11020,20 +11171,25 @@ static void ggml_compute_forward_rope_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, - struct ggml_tensor * dst) { + struct ggml_tensor * dst, + const bool forward) { if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { return; } - float freq_base; - float freq_scale; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow; - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); + //const int n_past = ((int32_t *) dst->op_params)[0]; + const int n_dims = ((int32_t *) dst->op_params)[1]; + const int mode = ((int32_t *) dst->op_params)[2]; + const int n_ctx = ((int32_t *) dst->op_params)[3]; + const int n_orig_ctx = ((int32_t *) dst->op_params)[4]; + memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); GGML_TENSOR_UNARY_OP_LOCALS @@ -11061,10 +11217,18 @@ static void ggml_compute_forward_rope_f16( int ir = 0; const float theta_scale = powf(freq_base, -2.0f/n_dims); + const float inv_ndims = -1.f/n_dims; + float corr_dims[2]; + ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims); const bool is_neox = mode & 2; const bool is_glm = mode & 4; + // backward process uses inverse rotation by cos and sin. + // cos and sin build a rotation matrix, where the inverse is the transpose. + // this essentially just switches the sign of sin. + const float sin_sign = forward ? 1.0f : -1.0f; + const int32_t * pos = (const int32_t *) src1->data; for (int64_t i3 = 0; i3 < ne3; i3++) { @@ -11074,18 +11238,18 @@ static void ggml_compute_forward_rope_f16( if (ir++ < ir0) continue; if (ir > ir1) break; - float theta = freq_scale * (float)p; + float theta_base = (float)p; if (is_glm) { - theta = MIN(p, n_ctx - 2); + theta_base = MIN(p, n_ctx - 2); float block_theta = MAX(p - (n_ctx - 2), 0); for (int64_t i0 = 0; i0 < ne0 / 4; i0++) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + const float cos_theta = cosf(theta_base); + const float sin_theta = sinf(theta_base) * sin_sign; const float cos_block_theta = cosf(block_theta); - const float sin_block_theta = sinf(block_theta); + const float sin_block_theta = sinf(block_theta) * sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; block_theta *= theta_scale; const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); @@ -11103,10 +11267,13 @@ static void ggml_compute_forward_rope_f16( } } else if (!is_neox) { for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; - theta *= theta_scale; + theta_base *= theta_scale; const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); ggml_fp16_t * dst_data = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); @@ -11120,12 +11287,20 @@ static void ggml_compute_forward_rope_f16( } else { // TODO: this might be wrong for ne0 != n_dims - need double check // ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py#LL251C1-L294C28 + theta_base *= freq_scale; for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); + // simplified from `(ib * n_dims + ic) * inv_ndims` + float cur_rot = inv_ndims * ic - ib; - theta *= theta_scale; + float cos_theta, sin_theta; + rope_yarn( + theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, + &cos_theta, &sin_theta + ); + sin_theta *= sin_sign; + + theta_base *= theta_scale; const int64_t i0 = ib*n_dims + ic/2; @@ -11153,11 +11328,11 @@ static void ggml_compute_forward_rope( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f16(params, src0, src1, dst, true); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_f32(params, src0, src1, dst); + ggml_compute_forward_rope_f32(params, src0, src1, dst, true); } break; default: { @@ -11168,215 +11343,6 @@ static void ggml_compute_forward_rope( // ggml_compute_forward_rope_back -static void ggml_compute_forward_rope_back_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { - return; - } - - // y = rope(x, src1) - // dx = rope_back(dy, src1) - // src0 is dy, src1 contains options - - float freq_base; - float freq_scale; - - // these two only relevant for xPos RoPE: - float xpos_base; - bool xpos_down; - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - const int n_ctx = ((int32_t *) dst->op_params)[3]; UNUSED(n_ctx); - memcpy(&freq_base, (int32_t *) dst->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) dst->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) dst->op_params + 7, sizeof(bool)); - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - assert(nb0 == sizeof(float)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(freq_base, -2.0f/n_dims); - - const bool is_neox = mode & 2; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - const int64_t p = pos[i2]; - for (int64_t i1 = 0; i1 < ne1; i1++) { - if (ir++ < ir0) continue; - if (ir > ir1) break; - - float theta = freq_scale * (float)p; - - if (!is_neox) { - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - // zeta scaling for xPos only: - float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f; - if (xpos_down) zeta = 1.0f / zeta; - - theta *= theta_scale; - - const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = dy[0]; - const float dy1 = dy[1]; - - dx[0] = dy0*cos_theta*zeta + dy1*sin_theta*zeta; - dx[1] = - dy0*sin_theta*zeta + dy1*cos_theta*zeta; - } - } else { - for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { - for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const int64_t i0 = ib*n_dims + ic/2; - - const float * const dy = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - float * dx = (float *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = dy[0]; - const float dy1 = dy[n_dims/2]; - - dx[0] = dy0*cos_theta + dy1*sin_theta; - dx[n_dims/2] = - dy0*sin_theta + dy1*cos_theta; - } - } - } - } - } - } -} - -static void ggml_compute_forward_rope_back_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - - if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) { - return; - } - - // y = rope(x, src1) - // dx = rope_back(dy, src1) - // src0 is dy, src1 contains options - - //const int n_past = ((int32_t *) dst->op_params)[0]; - const int n_dims = ((int32_t *) dst->op_params)[1]; - const int mode = ((int32_t *) dst->op_params)[2]; - - GGML_TENSOR_UNARY_OP_LOCALS - - //printf("ne0: %d, ne1: %d, ne2: %d, ne3: %d\n", ne0, ne1, ne2, ne3); - //printf("n_past = %d, ne2 = %d\n", n_past, ne2); - - assert(nb0 == sizeof(ggml_fp16_t)); - - const int ith = params->ith; - const int nth = params->nth; - - const int nr = ggml_nrows(dst); - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - // row index used to determine which thread to use - int ir = 0; - - const float theta_scale = powf(10000.0, -2.0f/n_dims); - - const bool is_neox = mode & 2; - - const int32_t * pos = (const int32_t *) src1->data; - - for (int64_t i3 = 0; i3 < ne3; i3++) { - for (int64_t i2 = 0; i2 < ne2; i2++) { - const int64_t p = pos[i2]; - for (int64_t i1 = 0; i1 < ne1; i1++) { - if (ir++ < ir0) continue; - if (ir > ir1) break; - - float theta = (float)p; - - if (!is_neox) { - for (int64_t i0 = 0; i0 < ne0; i0 += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = GGML_FP16_TO_FP32(dy[0]); - const float dy1 = GGML_FP16_TO_FP32(dy[1]); - - dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); - dx[1] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); - } - } else { - for (int64_t ib = 0; ib < ne0/n_dims; ++ib) { - for (int64_t ic = 0; ic < n_dims; ic += 2) { - const float cos_theta = cosf(theta); - const float sin_theta = sinf(theta); - - theta *= theta_scale; - - const int64_t i0 = ib*n_dims + ic/2; - - const ggml_fp16_t * const dy = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00); - ggml_fp16_t * dx = (ggml_fp16_t *)((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0); - - const float dy0 = GGML_FP16_TO_FP32(dy[0]); - const float dy1 = GGML_FP16_TO_FP32(dy[n_dims/2]); - - dx[0] = GGML_FP32_TO_FP16( dy0*cos_theta + dy1*sin_theta); - dx[n_dims/2] = GGML_FP32_TO_FP16(-dy0*sin_theta + dy1*cos_theta); - } - } - } - } - } - } -} - static void ggml_compute_forward_rope_back( const struct ggml_compute_params * params, const struct ggml_tensor * src0, @@ -11385,421 +11351,11 @@ static void ggml_compute_forward_rope_back( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_rope_back_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f16(params, src0, src1, dst, false); } break; case GGML_TYPE_F32: { - ggml_compute_forward_rope_back_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -// ggml_compute_forward_conv_1d - -static void ggml_compute_forward_conv_1d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - // size of the convolution row - the kernel size unrolled across all input channels - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - ggml_fp16_t * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = GGML_FP32_TO_FP16(src[idx0]); - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne2; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f16(ew0, dst_data + i0, - (ggml_fp16_t *) ((char *) src0->data + i1*nb02), - (ggml_fp16_t *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -static void ggml_compute_forward_conv_1d_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F32); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - const int ith = params->ith; - const int nth = params->nth; - - const int nk = ne00; - - const int ew0 = nk*ne01; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(float)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - float * const wdata = (float *) params->wdata + 0; - - for (int64_t i11 = 0; i11 < ne11; i11++) { - const float * const src = (float *)((char *) src1->data + i11*nb11); - float * dst_data = wdata; - - for (int64_t i0 = 0; i0 < ne0; i0++) { - for (int64_t ik = 0; ik < nk; ik++) { - const int idx0 = i0*s0 + ik*d0 - p0; - - if(!(idx0 < 0 || idx0 >= ne10)) { - dst_data[i0*ew0 + i11*nk + ik] = src[idx0]; - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // total rows in dst - const int nr = ne02; - - // rows per thread - const int dr = (nr + nth - 1)/nth; - - // row range for this thread - const int ir0 = dr*ith; - const int ir1 = MIN(ir0 + dr, nr); - - float * const wdata = (float *) params->wdata + 0; - - for (int i2 = 0; i2 < ne2; i2++) { - for (int i1 = ir0; i1 < ir1; i1++) { - float * dst_data = (float *)((char *) dst->data + i2*nb2 + i1*nb1); - - for (int i0 = 0; i0 < ne0; i0++) { - ggml_vec_dot_f32(ew0, dst_data + i0, - (float *) ((char *) src0->data + i1*nb02), - (float *) wdata + i2*nb2 + i0*ew0); - } - } - } -} - -// TODO: reuse ggml_mul_mat or implement ggml_im2col and remove stage_0 and stage_1 -static void gemm_f16_out_f32(int64_t m, int64_t n, int64_t k, - ggml_fp16_t * A, - ggml_fp16_t * B, - float * C, - const int ith, const int nth) { - // does not seem to make a difference - int64_t m0, m1, n0, n1; - // patches per thread - if (m > n) { - n0 = 0; - n1 = n; - - // total patches in dst - const int np = m; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - m0 = dp*ith; - m1 = MIN(m0 + dp, np); - } else { - m0 = 0; - m1 = m; - - // total patches in dst - const int np = n; - - // patches per thread - const int dp = (np + nth - 1)/nth; - - // patch range for this thread - n0 = dp*ith; - n1 = MIN(n0 + dp, np); - } - - // block-tiling attempt - int64_t blck_n = 16; - int64_t blck_m = 16; - - // int64_t CACHE_SIZE = 2 * 1024 * 1024; // 2MB - // int64_t blck_size = CACHE_SIZE / (sizeof(float) + 2 * sizeof(ggml_fp16_t) * K); - // if (blck_size > 0) { - // blck_0 = 4; - // blck_1 = blck_size / blck_0; - // if (blck_1 < 0) { - // blck_1 = 1; - // } - // // blck_0 = (int64_t)sqrt(blck_size); - // // blck_1 = blck_0; - // } - // // printf("%zd %zd %zd %zd\n", blck_size, K, blck_0, blck_1); - - for (int j = n0; j < n1; j+=blck_n) { - for (int i = m0; i < m1; i+=blck_m) { - // printf("i j k => %d %d %d\n", i, j, K); - for (int ii = i; ii < i + blck_m && ii < m1; ii++) { - for (int jj = j; jj < j + blck_n && jj < n1; jj++) { - ggml_vec_dot_f16(k, - C + ii*n + jj, - A + ii * k, - B + jj * k); - } - } - } - } -} - -// src0: kernel [OC, IC, K] -// src1: signal [N, IC, IL] -// dst: result [N, OL, IC*K] -static void ggml_compute_forward_conv_1d_stage_0_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F16); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS; - - const int64_t N = ne12; - const int64_t IC = ne11; - const int64_t IL = ne10; - - const int64_t K = ne00; - - const int64_t OL = ne1; - - const int ith = params->ith; - const int nth = params->nth; - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[1]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[2]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(dst->data, 0, ggml_nbytes(dst)); - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - // im2col: [N, IC, IL] => [N, OL, IC*K] - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; - - for (int64_t in = 0; in < N; in++) { - for (int64_t iol = 0; iol < OL; iol++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OL + iol)*(IC*K); // [IC, K] - const float * const src_data = (float *)((char *) src1->data + in*nb12 + iic*nb11); // [IL] - - for (int64_t ik = 0; ik < K; ik++) { - const int64_t iil = iol*s0 + ik*d0 - p0; - - if (!(iil < 0 || iil >= IL)) { - dst_data[iic*K + ik] = GGML_FP32_TO_FP16(src_data[iil]); - } - } - } - } - } - } -} - -// gemm: [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] -// src0: [OC, IC, K] -// src1: [N, OL, IC * K] -// result: [N, OC, OL] -static void ggml_compute_forward_conv_1d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne12; - const int OL = ne11; - - const int OC = ne02; - const int IC = ne01; - const int K = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OL; - int64_t k = IC * K; - - // [N, OC, OL] = [OC, IC * K] x [N*OL, IC * K] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_1d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - ggml_compute_forward_conv_1d_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_0_f32(params, src0, src1, dst); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_1d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch(src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_1d_stage_1_f16(params, src0, src1, dst); + ggml_compute_forward_rope_f32(params, src0, src1, dst, false); } break; default: { @@ -12019,12 +11575,10 @@ static void ggml_compute_forward_conv_transpose_1d( } } -// ggml_compute_forward_conv_2d - // src0: kernel [OC, IC, KH, KW] // src1: image [N, IC, IH, IW] // dst: result [N, OH, OW, IC*KH*KW] -static void ggml_compute_forward_conv_2d_stage_0_f32( +static void ggml_compute_forward_im2col_f16( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12038,34 +11592,35 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( GGML_TENSOR_BINARY_OP_LOCALS; - const int64_t N = ne13; - const int64_t IC = ne12; - const int64_t IH = ne11; - const int64_t IW = ne10; - - // const int64_t OC = ne03; - // const int64_t IC = ne02; - const int64_t KH = ne01; - const int64_t KW = ne00; - - const int64_t OH = ne2; - const int64_t OW = ne1; + const int32_t s0 = ((const int32_t *)(dst->op_params))[0]; + const int32_t s1 = ((const int32_t *)(dst->op_params))[1]; + const int32_t p0 = ((const int32_t *)(dst->op_params))[2]; + const int32_t p1 = ((const int32_t *)(dst->op_params))[3]; + const int32_t d0 = ((const int32_t *)(dst->op_params))[4]; + const int32_t d1 = ((const int32_t *)(dst->op_params))[5]; + const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1; const int ith = params->ith; const int nth = params->nth; - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; + const int64_t N = is_2D ? ne13 : ne12; + const int64_t IC = is_2D ? ne12 : ne11; + const int64_t IH = is_2D ? ne11 : 1; + const int64_t IW = ne10; + + const int64_t KH = is_2D ? ne01 : 1; + const int64_t KW = ne00; + + const int64_t OH = is_2D ? ne2 : 1; + const int64_t OW = ne1; + + int ofs0 = is_2D ? nb13 : nb12; + int ofs1 = is_2D ? nb12 : nb11; GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); GGML_ASSERT(nb10 == sizeof(float)); if (params->type == GGML_TASK_INIT) { - memset(dst->data, 0, ggml_nbytes(dst)); return; } @@ -12078,20 +11633,22 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data; for (int64_t in = 0; in < N; in++) { - for (int64_t ioh = 0; ioh < OH; ioh++) { + for (int64_t ioh = 0; ioh < OH; ioh++) { // 1 for (int64_t iow = 0; iow < OW; iow++) { - for (int64_t iic = ith; iic < IC; iic+=nth) { + for (int64_t iic = ith; iic < IC; iic += nth) { // micro kernel ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] + const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW] - for (int64_t ikh = 0; ikh < KH; ikh++) { + for (int64_t ikh = 0; ikh < KH; ikh++) { // 1 for (int64_t ikw = 0; ikw < KW; ikw++) { const int64_t iiw = iow*s0 + ikw*d0 - p0; const int64_t iih = ioh*s1 + ikh*d1 - p1; - if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { + if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) { + dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0; + } else { dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); } } @@ -12103,180 +11660,7 @@ static void ggml_compute_forward_conv_2d_stage_0_f32( } } -// gemm: [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] -// src0: [OC, IC, KH, KW] -// src1: [N, OH, OW, IC * KH * KW] -// result: [N, OC, OH, OW] -static void ggml_compute_forward_conv_2d_stage_1_f16( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F16); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - if (params->type == GGML_TASK_INIT) { - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - GGML_TENSOR_BINARY_OP_LOCALS; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb0 == sizeof(float)); - - const int N = ne13; - const int OH = ne12; - const int OW = ne11; - - const int OC = ne03; - const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int ith = params->ith; - const int nth = params->nth; - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)src1->data + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m, n] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d_f16_f32( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - GGML_ASSERT(src0->type == GGML_TYPE_F16); - GGML_ASSERT(src1->type == GGML_TYPE_F32); - GGML_ASSERT( dst->type == GGML_TYPE_F32); - - int64_t t0 = ggml_perf_time_us(); - UNUSED(t0); - - GGML_TENSOR_BINARY_OP_LOCALS - - // src1: image [N, IC, IH, IW] - // src0: kernel [OC, IC, KH, KW] - // dst: result [N, OC, OH, OW] - // ne12: IC - // ne0: OW - // ne1: OH - // nk0: KW - // nk1: KH - // ne13: N - - const int N = ne13; - const int IC = ne12; - const int IH = ne11; - const int IW = ne10; - - const int OC = ne03; - // const int IC = ne02; - const int KH = ne01; - const int KW = ne00; - - const int OH = ne1; - const int OW = ne0; - - const int ith = params->ith; - const int nth = params->nth; - - // const int nk0 = ne00; - // const int nk1 = ne01; - - // size of the convolution row - the kernel size unrolled across all channels - // const int ew0 = nk0*nk1*ne02; - // ew0: IC*KH*KW - - const int32_t s0 = ((const int32_t*)(dst->op_params))[0]; - const int32_t s1 = ((const int32_t*)(dst->op_params))[1]; - const int32_t p0 = ((const int32_t*)(dst->op_params))[2]; - const int32_t p1 = ((const int32_t*)(dst->op_params))[3]; - const int32_t d0 = ((const int32_t*)(dst->op_params))[4]; - const int32_t d1 = ((const int32_t*)(dst->op_params))[5]; - - GGML_ASSERT(nb00 == sizeof(ggml_fp16_t)); - GGML_ASSERT(nb10 == sizeof(float)); - - if (params->type == GGML_TASK_INIT) { - memset(params->wdata, 0, params->wsize); - - // prepare source data (src1) - // im2col: [N, IC, IH, IW] => [N*OH*OW, IC*KH*KW] - - { - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - - for (int in = 0; in < N; in++) { - for (int iic = 0; iic < IC; iic++) { - for (int ioh = 0; ioh < OH; ioh++) { - for (int iow = 0; iow < OW; iow++) { - - // micro kernel - ggml_fp16_t * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW] - const float * const src_data = (float *)((char *) src1->data + in*nb13 + iic*nb12); // [IH, IW] - - for (int ikh = 0; ikh < KH; ikh++) { - for (int ikw = 0; ikw < KW; ikw++) { - const int iiw = iow*s0 + ikw*d0 - p0; - const int iih = ioh*s1 + ikh*d1 - p1; - - if (!(iih < 0 || iih >= IH || iiw < 0 || iiw >= IW)) { - dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]); - } - } - } - } - } - } - } - } - - return; - } - - if (params->type == GGML_TASK_FINALIZE) { - return; - } - - ggml_fp16_t * const wdata = (ggml_fp16_t *) params->wdata + 0; - // wdata: [N*OH*OW, IC*KH*KW] - // dst: result [N, OC, OH, OW] - // src0: kernel [OC, IC, KH, KW] - - int64_t m = OC; - int64_t n = OH * OW; - int64_t k = IC * KH * KW; - - // [N, OC, OH, OW] = [OC, IC * KH * KW] x [N*OH*OW, IC * KH * KW] - for (int i = 0; i < N; i++) { - ggml_fp16_t * A = (ggml_fp16_t *)src0->data; // [m, k] - ggml_fp16_t * B = (ggml_fp16_t *)wdata + i * m * k; // [n, k] - float * C = (float *)dst->data + i * m * n; // [m * k] - - gemm_f16_out_f32(m, n, k, A, B, C, ith, nth); - } -} - -static void ggml_compute_forward_conv_2d( +static void ggml_compute_forward_im2col( const struct ggml_compute_params * params, const struct ggml_tensor * src0, const struct ggml_tensor * src1, @@ -12284,50 +11668,7 @@ static void ggml_compute_forward_conv_2d( switch (src0->type) { case GGML_TYPE_F16: { - ggml_compute_forward_conv_2d_f16_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - //ggml_compute_forward_conv_2d_f32(params, src0, src1, dst); - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_0( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_0_f32(params, src0, src1, dst); - } break; - case GGML_TYPE_F32: - { - GGML_ASSERT(false); - } break; - default: - { - GGML_ASSERT(false); - } break; - } -} - -static void ggml_compute_forward_conv_2d_stage_1( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - const struct ggml_tensor * src1, - struct ggml_tensor * dst) { - switch (src0->type) { - case GGML_TYPE_F16: - { - ggml_compute_forward_conv_2d_stage_1_f16(params, src0, src1, dst); + ggml_compute_forward_im2col_f16(params, src0, src1, dst); } break; case GGML_TYPE_F32: { @@ -12512,14 +11853,11 @@ static void ggml_compute_forward_pool_1d( ggml_compute_forward_pool_1d_sk_p0(params, op, src0, k0, dst); } -// ggml_compute_forward_pool_2d_sk_p0 +// ggml_compute_forward_pool_2d -static void ggml_compute_forward_pool_2d_sk_p0( +static void ggml_compute_forward_pool_2d( const struct ggml_compute_params * params, - const enum ggml_op_pool op, const struct ggml_tensor * src, - const int k0, - const int k1, struct ggml_tensor * dst) { assert(src->type == GGML_TYPE_F32); assert(params->ith == 0); @@ -12528,6 +11866,14 @@ static void ggml_compute_forward_pool_2d_sk_p0( return; } + const int32_t * opts = (const int32_t *)dst->op_params; + enum ggml_op_pool op = opts[0]; + const int k0 = opts[1]; + const int k1 = opts[2]; + const int s0 = opts[3]; + const int s1 = opts[4]; + const int p0 = opts[5]; + const int p1 = opts[6]; const char * cdata = (const char*)src->data; const char * const data_end = cdata + ggml_nbytes(src); @@ -12538,6 +11884,8 @@ static void ggml_compute_forward_pool_2d_sk_p0( float * dplane = (float *)dst->data; const int ka = k0 * k1; + const int offset0 = -p0; + const int offset1 = -p1; while (cdata < data_end) { for (int oy = 0; oy < py; ++oy) { @@ -12550,13 +11898,15 @@ static void ggml_compute_forward_pool_2d_sk_p0( case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break; } - const int ix = ox * k0; - const int iy = oy * k1; + const int ix = offset0 + ox * s0; + const int iy = offset1 + oy * s1; for (int ky = 0; ky < k1; ++ky) { + if (iy + ky < 0 || iy + ky >= src->ne[1]) continue; const float * const srow = (const float *)(cdata + src->nb[1] * (iy + ky)); for (int kx = 0; kx < k0; ++kx) { int j = ix + kx; + if (j < 0 || j >= src->ne[0]) continue; switch (op) { case GGML_OP_POOL_AVG: *out += srow[j]; break; case GGML_OP_POOL_MAX: if (srow[j] > *out) *out = srow[j]; break; @@ -12577,29 +11927,6 @@ static void ggml_compute_forward_pool_2d_sk_p0( } } -// ggml_compute_forward_pool_2d - -static void ggml_compute_forward_pool_2d( - const struct ggml_compute_params * params, - const struct ggml_tensor * src0, - struct ggml_tensor * dst) { - - const int32_t * opts = (const int32_t *)dst->op_params; - enum ggml_op_pool op = opts[0]; - const int k0 = opts[1]; - const int k1 = opts[2]; - const int s0 = opts[3]; - const int s1 = opts[4]; - const int p0 = opts[5]; - const int p1 = opts[6]; - GGML_ASSERT(p0 == 0); - GGML_ASSERT(p1 == 0); // padding not supported - GGML_ASSERT(k0 == s0); - GGML_ASSERT(k1 == s1); // only s = k supported - - ggml_compute_forward_pool_2d_sk_p0(params, op, src0, k0, k1, dst); -} - // ggml_compute_forward_upscale static void ggml_compute_forward_upscale_f32( @@ -13801,6 +13128,10 @@ static void ggml_compute_forward_unary( { ggml_compute_forward_silu(params, src0, dst); } break; + case GGML_UNARY_OP_LEAKY: + { + ggml_compute_forward_leaky(params, src0, dst); + } break; default: { GGML_ASSERT(false); @@ -14554,33 +13885,13 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm { ggml_compute_forward_clamp(params, tensor->src[0], tensor); } break; - case GGML_OP_CONV_1D: - { - ggml_compute_forward_conv_1d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - ggml_compute_forward_conv_1d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - ggml_compute_forward_conv_1d_stage_1(params, tensor->src[0], tensor->src[1], tensor); - } break; case GGML_OP_CONV_TRANSPOSE_1D: { ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor); } break; - case GGML_OP_CONV_2D: + case GGML_OP_IM2COL: { - ggml_compute_forward_conv_2d(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - ggml_compute_forward_conv_2d_stage_0(params, tensor->src[0], tensor->src[1], tensor); - } break; - case GGML_OP_CONV_2D_STAGE_1: - { - ggml_compute_forward_conv_2d_stage_1(params, tensor->src[0], tensor->src[1], tensor); + ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor); } break; case GGML_OP_CONV_TRANSPOSE_2D: { @@ -14709,62 +14020,109 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm //////////////////////////////////////////////////////////////////////////////// -static_assert(GGML_GRAPH_HASHTABLE_SIZE > GGML_MAX_NODES * 2, "GGML_GRAPH_HT_SIZE is too small"); +static size_t ggml_hash_size(size_t min_sz) { + // next primes after powers of two + static const size_t primes[] = { + 2, 3, 5, 11, 17, 37, 67, 131, 257, 521, 1031, + 2053, 4099, 8209, 16411, 32771, 65537, 131101, + 262147, 524309, 1048583, 2097169, 4194319, 8388617, + 16777259, 33554467, 67108879, 134217757, 268435459, + 536870923, 1073741827, 2147483659 + }; + static const size_t n_primes = sizeof(primes)/sizeof(primes[0]); -static size_t hash(void * p) { - return (size_t)p % GGML_GRAPH_HASHTABLE_SIZE; + // find the smallest prime that is larger or equal to min_sz + size_t l = 0; + size_t r = n_primes; + while (l < r) { + size_t m = (l + r)/2; + if (primes[m] < min_sz) { + l = m + 1; + } else { + r = m; + } + } + size_t sz = l < n_primes ? primes[l] : min_sz | 1; + return sz; } -static size_t hash_find(void * hash_table[], void * p) { - size_t h = hash(p); +static size_t ggml_hash(const void * p) { + return (size_t)p; +} + +size_t ggml_hash_find(const struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t h = ggml_hash(key) % hash_set.size; // linear probing size_t i = h; - while (hash_table[i] != NULL && hash_table[i] != p) { - i = (i + 1) % GGML_GRAPH_HASHTABLE_SIZE; + while (hash_set.keys[i] != NULL && hash_set.keys[i] != key) { + i = (i + 1) % hash_set.size; if (i == h) { // visited all hash table entries -> not found - return GGML_GRAPH_HASHTABLE_SIZE; + return GGML_HASHTABLE_FULL; } } return i; } -static bool hash_insert(void * hash_table[], void * p) { - size_t i = hash_find(hash_table, p); +bool ggml_hash_contains(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); + return i != GGML_HASHTABLE_FULL && hash_set.keys[i] == key; +} - GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full +size_t ggml_hash_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); - if (hash_table[i] == p) { - return true; + GGML_ASSERT(i != GGML_HASHTABLE_FULL); + + if (hash_set.keys[i] == key) { + return GGML_HASHTABLE_ALREADY_EXISTS; } // insert - GGML_ASSERT(hash_table[i] == NULL); - hash_table[i] = p; - return false; + GGML_ASSERT(hash_set.keys[i] == NULL); + hash_set.keys[i] = key; + return i; } -static bool hash_contains(void * hash_table[], void * p) { - size_t i = hash_find(hash_table, p); - return (i < GGML_GRAPH_HASHTABLE_SIZE) && (hash_table[i] == p); +size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) { + size_t i = ggml_hash_find(hash_set, key); + + GGML_ASSERT(i != GGML_HASHTABLE_FULL); + + hash_set.keys[i] = key; + return i; } -struct hash_map { - void * keys[GGML_GRAPH_HASHTABLE_SIZE]; - void * vals[GGML_GRAPH_HASHTABLE_SIZE]; -}; - -static struct hash_map * new_hash_map(void) { - struct hash_map * result = malloc(sizeof(struct hash_map)); - for (int i=0; ikeys[i] = NULL; - result->vals[i] = NULL; - } +static struct ggml_hash_set ggml_hash_set_new(size_t size) { + size = ggml_hash_size(size); + struct ggml_hash_set result; + result.size = size; + result.keys = malloc(sizeof(struct ggml_tensor *) * size); + memset(result.keys, 0, sizeof(struct ggml_tensor *) * size); return result; } -static void free_hash_map(struct hash_map * map) { +static void ggml_hash_set_free(struct ggml_hash_set hash_set) { + free(hash_set.keys); +} + +struct hash_map { + struct ggml_hash_set set; + struct ggml_tensor ** vals; +}; + +static struct hash_map * ggml_new_hash_map(size_t size) { + struct hash_map * result = malloc(sizeof(struct hash_map)); + result->set = ggml_hash_set_new(size); + result->vals = malloc(sizeof(struct ggml_tensor *) * result->set.size); + memset(result->vals, 0, sizeof(struct ggml_tensor *) * result->set.size); + return result; +} + +static void ggml_hash_map_free(struct hash_map * map) { + ggml_hash_set_free(map->set); + free(map->vals); free(map); } @@ -14784,7 +14142,7 @@ static struct ggml_tensor * ggml_recompute_graph_node( return node; } - if (!hash_contains(graph->visited_hash_table, node)) { + if (!ggml_hash_contains(graph->visited_hash_table, node)) { return node; } @@ -14799,17 +14157,17 @@ static struct ggml_tensor * ggml_recompute_graph_node( return node; } - size_t i = hash_find(replacements->keys, node); - GGML_ASSERT(i < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full - if (replacements->keys[i] == node) { - return (struct ggml_tensor *) replacements->vals[i]; + size_t i = ggml_hash_find(replacements->set, node); + GGML_ASSERT(i != GGML_HASHTABLE_FULL); // assert that not full + if (replacements->set.keys[i] == node) { + return replacements->vals[i]; } struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne); // insert clone into replacements - GGML_ASSERT(replacements->keys[i] == NULL); // assert that we don't overwrite - replacements->keys[i] = node; + GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite + replacements->set.keys[i] = node; replacements->vals[i] = clone; clone->op = node->op; @@ -14846,26 +14204,26 @@ void ggml_build_backward_gradient_checkpointing( struct ggml_cgraph * gb_tmp, struct ggml_tensor * * checkpoints, int n_checkpoints) { - *gb_tmp = *gf; + ggml_graph_cpy(gf, gb_tmp); ggml_build_backward_expand(ctx, gf, gb_tmp, true); if (n_checkpoints <= 0) { - *gb = *gb_tmp; + ggml_graph_cpy(gb_tmp, gb); return; } - struct hash_map * replacements = new_hash_map(); + struct hash_map * replacements = ggml_new_hash_map(gf->n_nodes + gf->n_leafs + n_checkpoints); // insert checkpoints in replacements for (int i = 0; i < n_checkpoints; ++i) { - size_t k = hash_find(replacements->keys, checkpoints[i]); - GGML_ASSERT(k < GGML_GRAPH_HASHTABLE_SIZE); // assert that not full - GGML_ASSERT(replacements->keys[k] == NULL); // assert that we don't overwrite - replacements->keys[k] = checkpoints[i]; - replacements->vals[k] = checkpoints[i]; + size_t k = ggml_hash_find(replacements->set, checkpoints[i]); + GGML_ASSERT(k != GGML_HASHTABLE_FULL); // assert that not full + GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite + replacements->set.keys[k] = checkpoints[i]; + replacements->vals[k] = checkpoints[i]; } - *gb = *gf; + ggml_graph_cpy(gf, gb); // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes], // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]), // by recomputing them from checkpoints @@ -14882,21 +14240,21 @@ void ggml_build_backward_gradient_checkpointing( ggml_build_forward_expand(gb, node); } - free_hash_map(replacements); + ggml_hash_map_free(replacements); } // functions to change gradients considering the case that input a might be initial gradient with zero value -static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return b; } else { return ggml_add_impl(ctx, a, b, false); } } -static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { struct ggml_tensor * a_zero = ggml_scale(ctx, a, ggml_new_f32(ctx, 0)); return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false); } else { @@ -14904,23 +14262,23 @@ static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct gg } } -static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return ggml_repeat(ctx, b, a); } else { return ggml_add1_impl(ctx, a, b, false); } } -static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, void * zero_table[]) { - if (hash_contains(zero_table, a)) { +static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) { + if (ggml_hash_contains(zero_table, a)) { return ggml_neg(ctx, b); } else { return ggml_sub_impl(ctx, a, b, false); } } -static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, void * zero_table[]) { +static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) { struct ggml_tensor * src0 = tensor->src[0]; struct ggml_tensor * src1 = tensor->src[1]; @@ -15432,17 +14790,20 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor // necessary for llama if (src0->grad) { //const int n_past = ((int32_t *) tensor->op_params)[0]; - const int n_dims = ((int32_t *) tensor->op_params)[1]; - const int mode = ((int32_t *) tensor->op_params)[2]; - const int n_ctx = ((int32_t *) tensor->op_params)[3]; - float freq_base; - float freq_scale; - float xpos_base; - bool xpos_down; - memcpy(&freq_base, (int32_t *) tensor->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) tensor->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) tensor->op_params + 7, sizeof(bool)); + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; + const int n_orig_ctx = ((int32_t *) tensor->op_params)[4]; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down; + + memcpy(&freq_base, (int32_t *) tensor->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) tensor->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) tensor->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) tensor->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) tensor->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) tensor->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) tensor->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) tensor->op_params + 12, sizeof(bool)); src0->grad = ggml_add_or_set(ctx, src0->grad, @@ -15452,8 +14813,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor n_dims, mode, n_ctx, + n_orig_ctx, freq_base, freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, xpos_base, xpos_down), zero_table); @@ -15463,17 +14829,20 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { if (src0->grad) { //const int n_past = ((int32_t *) tensor->op_params)[0]; - const int n_dims = ((int32_t *) tensor->op_params)[1]; - const int mode = ((int32_t *) tensor->op_params)[2]; - const int n_ctx = ((int32_t *) tensor->op_params)[3]; - float freq_base; - float freq_scale; - float xpos_base; - bool xpos_down; - memcpy(&freq_base, (int32_t *) tensor->op_params + 4, sizeof(float)); - memcpy(&freq_scale, (int32_t *) tensor->op_params + 5, sizeof(float)); - memcpy(&xpos_base, (int32_t *) tensor->op_params + 6, sizeof(float)); - memcpy(&xpos_down, (int32_t *) tensor->op_params + 7, sizeof(bool)); + const int n_dims = ((int32_t *) tensor->op_params)[1]; + const int mode = ((int32_t *) tensor->op_params)[2]; + const int n_ctx = ((int32_t *) tensor->op_params)[3]; + const int n_orig_ctx = ((int32_t *) tensor->op_params)[4]; + float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down; + + memcpy(&freq_base, (int32_t *) tensor->op_params + 5, sizeof(float)); + memcpy(&freq_scale, (int32_t *) tensor->op_params + 6, sizeof(float)); + memcpy(&ext_factor, (int32_t *) tensor->op_params + 7, sizeof(float)); + memcpy(&attn_factor, (int32_t *) tensor->op_params + 8, sizeof(float)); + memcpy(&beta_fast, (int32_t *) tensor->op_params + 9, sizeof(float)); + memcpy(&beta_slow, (int32_t *) tensor->op_params + 10, sizeof(float)); + memcpy(&xpos_base, (int32_t *) tensor->op_params + 11, sizeof(float)); + memcpy(&xpos_down, (int32_t *) tensor->op_params + 12, sizeof(bool)); src0->grad = ggml_add_or_set(ctx, src0->grad, @@ -15483,8 +14852,13 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor n_dims, mode, n_ctx, + n_orig_ctx, freq_base, freq_scale, + ext_factor, + attn_factor, + beta_fast, + beta_slow, xpos_base, xpos_down, false), @@ -15499,31 +14873,11 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_1D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - GGML_ASSERT(false); // TODO: not implemented - } break; case GGML_OP_CONV_TRANSPOSE_1D: { GGML_ASSERT(false); // TODO: not implemented } break; - case GGML_OP_CONV_2D: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - GGML_ASSERT(false); // TODO: not implemented - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { GGML_ASSERT(false); // TODO: not implemented } break; @@ -15737,7 +15091,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * } // check if already visited - if (hash_insert(cgraph->visited_hash_table, node)) { + if (ggml_hash_insert(cgraph->visited_hash_table, node) == GGML_HASHTABLE_ALREADY_EXISTS) { return; } @@ -15753,7 +15107,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * if (node->op == GGML_OP_NONE && node->grad == NULL) { // reached a leaf node, not part of the gradient graph (e.g. a constant) - GGML_ASSERT(cgraph->n_leafs < GGML_MAX_NODES); + GGML_ASSERT(cgraph->n_leafs < cgraph->size); if (strlen(node->name) == 0) { ggml_format_name(node, "leaf_%d", cgraph->n_leafs); @@ -15762,22 +15116,24 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * cgraph->leafs[cgraph->n_leafs] = node; cgraph->n_leafs++; } else { - GGML_ASSERT(cgraph->n_nodes < GGML_MAX_NODES); + GGML_ASSERT(cgraph->n_nodes < cgraph->size); if (strlen(node->name) == 0) { ggml_format_name(node, "node_%d", cgraph->n_nodes); } cgraph->nodes[cgraph->n_nodes] = node; - cgraph->grads[cgraph->n_nodes] = node->grad; + if (cgraph->grads) { + cgraph->grads[cgraph->n_nodes] = node->grad; + } cgraph->n_nodes++; } } static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) { if (!expand) { - cgraph->n_nodes = 0; - cgraph->n_leafs = 0; + // TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand + ggml_graph_clear(cgraph); } const int n0 = cgraph->n_nodes; @@ -15798,25 +15154,6 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * ggml_build_forward_impl(cgraph, tensor, true); } -struct ggml_cgraph ggml_build_forward(struct ggml_tensor * tensor) { - struct ggml_cgraph result = { - /*.n_nodes =*/ 0, - /*.n_leafs =*/ 0, - /*.nodes =*/ { NULL }, - /*.grads =*/ { NULL }, - /*.leafs =*/ { NULL }, - /*.hash_table =*/ { NULL }, - /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, - /*.perf_runs =*/ 0, - /*.perf_cycles =*/ 0, - /*.perf_time_us =*/ 0, - }; - - ggml_build_forward_impl(&result, tensor, false); - - return result; -} - void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) { GGML_ASSERT(gf->n_nodes > 0); @@ -15833,11 +15170,10 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * } // remember original gradients which start with zero values - void ** zero_table = malloc(sizeof(void *) * GGML_GRAPH_HASHTABLE_SIZE); - memset(zero_table, 0, sizeof(void*) * GGML_GRAPH_HASHTABLE_SIZE); + struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size); for (int i = 0; i < gf->n_nodes; i++) { if (gf->grads[i]) { - hash_insert(zero_table, gf->grads[i]); + ggml_hash_insert(zero_table, gf->grads[i]); } } @@ -15860,26 +15196,54 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * } } - free(zero_table); + ggml_hash_set_free(zero_table); } -struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep) { - struct ggml_cgraph result = *gf; - ggml_build_backward_expand(ctx, gf, &result, keep); - return result; +static size_t ggml_graph_nbytes(size_t size, bool grads) { + size_t nbytes = sizeof(struct ggml_cgraph); + nbytes += size * sizeof(struct ggml_tensor *) * 2; // leafs + nodes + if (grads) { + nbytes += size * sizeof(struct ggml_tensor *); // grads + } + nbytes += ggml_hash_size(size * 2) * sizeof(struct ggml_tensor *); // hash set + return nbytes; } -struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { - struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE); +size_t ggml_graph_overhead_custom(size_t size, bool grads) { + return GGML_OBJECT_SIZE + GGML_PAD(ggml_graph_nbytes(size, grads), GGML_MEM_ALIGN); +} + +size_t ggml_graph_overhead(void) { + return ggml_graph_overhead_custom(GGML_DEFAULT_GRAPH_SIZE, false); +} + +struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads) { + const size_t obj_size = ggml_graph_nbytes(size, grads); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size); struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); + struct ggml_tensor ** data_start = (struct ggml_tensor **) (cgraph + 1); + + size_t hash_size = ggml_hash_size(size * 2); + struct ggml_tensor ** nodes_ptr = data_start; + struct ggml_tensor ** leafs_ptr = nodes_ptr + size; + struct ggml_tensor ** hash_keys_ptr = leafs_ptr + size; + struct ggml_tensor ** grads_ptr = grads ? hash_keys_ptr + hash_size : NULL; + + // check that we allocated the correct amount of memory + assert(obj_size == (size_t) ( + (grads ? (char *)(grads_ptr + size) : (char *)(hash_keys_ptr + hash_size)) - (char *)cgraph)); + + memset(hash_keys_ptr, 0, hash_size * sizeof(struct ggml_tensor *)); + *cgraph = (struct ggml_cgraph) { + /*.size =*/ size, /*.n_nodes =*/ 0, /*.n_leafs =*/ 0, - /*.nodes =*/ { NULL }, - /*.grads =*/ { NULL }, - /*.leafs =*/ { NULL }, - /*.hash_table =*/ { NULL }, + /*.nodes =*/ nodes_ptr, + /*.grads =*/ grads_ptr, + /*.leafs =*/ leafs_ptr, + /*.hash_table =*/ { hash_size, hash_keys_ptr }, /*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, /*.perf_runs =*/ 0, /*.perf_cycles =*/ 0, @@ -15889,14 +15253,85 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { return cgraph; } -struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor) { - struct ggml_cgraph * cgraph = ggml_new_graph(ctx); - ggml_build_forward_impl(cgraph, tensor, false); +struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) { + return ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, false); +} + +struct ggml_cgraph * ggml_graph_view(struct ggml_context * ctx, struct ggml_cgraph * cgraph0, int i0, int i1) { + const size_t obj_size = sizeof(struct ggml_cgraph); + struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size); + struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); + + *cgraph = (struct ggml_cgraph) { + /*.size =*/ 0, + /*.n_nodes =*/ i1 - i0, + /*.n_leafs =*/ 0, + /*.nodes =*/ cgraph0->nodes + i0, + /*.grads =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL, + /*.leafs =*/ NULL, + /*.hash_table =*/ { 0, NULL }, + /*.order =*/ cgraph0->order, + /*.perf_runs =*/ 0, + /*.perf_cycles =*/ 0, + /*.perf_time_us =*/ 0, + }; + return cgraph; } -size_t ggml_graph_overhead(void) { - return GGML_OBJECT_SIZE + GGML_PAD(GGML_GRAPH_SIZE, GGML_MEM_ALIGN); +void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) { + GGML_ASSERT(dst->size >= src->n_leafs); + GGML_ASSERT(dst->size >= src->n_nodes); + GGML_ASSERT(dst->visited_hash_table.size >= src->visited_hash_table.size); + + dst->n_leafs = src->n_leafs; + dst->n_nodes = src->n_nodes; + dst->order = src->order; + + for (int i = 0; i < src->n_leafs; ++i) { + dst->leafs[i] = src->leafs[i]; + } + + for (int i = 0; i < src->n_nodes; ++i) { + dst->nodes[i] = src->nodes[i]; + } + + if (src->grads) { + GGML_ASSERT(dst->grads != NULL); + for (int i = 0; i < src->n_nodes; ++i) { + dst->grads[i] = src->grads[i]; + } + } + + for (size_t i = 0; i < src->visited_hash_table.size; ++i) { + if (src->visited_hash_table.keys[i]) { + ggml_hash_insert(dst->visited_hash_table, src->visited_hash_table.keys[i]); + } + } +} + +struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) { + struct ggml_cgraph * result = ggml_new_graph_custom(ctx, cgraph->size, cgraph->grads != NULL); + ggml_graph_cpy(cgraph, result); + return result; +} + +void ggml_graph_reset(struct ggml_cgraph * cgraph) { + GGML_ASSERT(cgraph->grads != NULL); + + for (int i = 0; i < cgraph->n_nodes; i++) { + struct ggml_tensor * grad = cgraph->grads[i]; + + if (grad) { + ggml_set_zero(grad); + } + } +} + +void ggml_graph_clear(struct ggml_cgraph * cgraph) { + cgraph->n_leafs = 0; + cgraph->n_nodes = 0; + memset(cgraph->visited_hash_table.keys, 0, cgraph->visited_hash_table.size * sizeof(struct ggml_tensor *)); } // @@ -16049,13 +15484,233 @@ static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const node->perf_time_us += time_us_cur; } +static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) { + int n_tasks = 0; + + switch (node->op) { + case GGML_OP_CPY: + case GGML_OP_DUP: + case GGML_OP_ADD: + case GGML_OP_ADD1: + case GGML_OP_ACC: + { + n_tasks = n_threads; + } break; + case GGML_OP_SUB: + case GGML_OP_DIV: + case GGML_OP_SQR: + case GGML_OP_SQRT: + case GGML_OP_LOG: + case GGML_OP_SUM: + case GGML_OP_SUM_ROWS: + case GGML_OP_MEAN: + case GGML_OP_ARGMAX: + case GGML_OP_REPEAT: + case GGML_OP_REPEAT_BACK: + { + n_tasks = 1; + } break; + case GGML_OP_UNARY: + switch (ggml_get_unary_op(node)) { + case GGML_UNARY_OP_ABS: + case GGML_UNARY_OP_SGN: + case GGML_UNARY_OP_NEG: + case GGML_UNARY_OP_STEP: + case GGML_UNARY_OP_TANH: + case GGML_UNARY_OP_ELU: + case GGML_UNARY_OP_RELU: + case GGML_UNARY_OP_LEAKY: + { + n_tasks = 1; + } break; + + case GGML_UNARY_OP_GELU: + case GGML_UNARY_OP_GELU_QUICK: + case GGML_UNARY_OP_SILU: + { + n_tasks = n_threads; + } break; + } + break; + case GGML_OP_SILU_BACK: + case GGML_OP_MUL: + case GGML_OP_NORM: + case GGML_OP_RMS_NORM: + case GGML_OP_RMS_NORM_BACK: + case GGML_OP_GROUP_NORM: + case GGML_OP_CONCAT: + { + n_tasks = n_threads; + } break; + case GGML_OP_MUL_MAT: + { + n_tasks = n_threads; + + // TODO: use different scheduling for different matrix sizes + //const int nr0 = ggml_nrows(node->src[0]); + //const int nr1 = ggml_nrows(node->src[1]); + + //n_tasks = MIN(n_threads, MAX(1, nr0/128)); + //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks); + +#if defined(GGML_USE_CUBLAS) + if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#elif defined(GGML_USE_CLBLAST) + if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#endif +#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) + if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) { + n_tasks = 1; // TODO: this actually is doing nothing + // the threads are still spinning + } +#endif + } break; + case GGML_OP_OUT_PROD: + { + n_tasks = n_threads; + } break; + case GGML_OP_SCALE: + case GGML_OP_SET: + case GGML_OP_CONT: + case GGML_OP_RESHAPE: + case GGML_OP_VIEW: + case GGML_OP_PERMUTE: + case GGML_OP_TRANSPOSE: + case GGML_OP_GET_ROWS: + case GGML_OP_GET_ROWS_BACK: + case GGML_OP_DIAG: + { + n_tasks = 1; + } break; + case GGML_OP_DIAG_MASK_ZERO: + case GGML_OP_DIAG_MASK_INF: + case GGML_OP_SOFT_MAX: + case GGML_OP_SOFT_MAX_BACK: + case GGML_OP_ROPE: + case GGML_OP_ROPE_BACK: + case GGML_OP_ADD_REL_POS: + { + n_tasks = n_threads; + } break; + case GGML_OP_ALIBI: + { + n_tasks = 1; //TODO + } break; + case GGML_OP_CLAMP: + { + n_tasks = 1; //TODO + } break; + case GGML_OP_CONV_TRANSPOSE_1D: + { + n_tasks = n_threads; + } break; + case GGML_OP_IM2COL: + { + n_tasks = n_threads; + } break; + case GGML_OP_CONV_TRANSPOSE_2D: + { + n_tasks = n_threads; + } break; + case GGML_OP_POOL_1D: + case GGML_OP_POOL_2D: + { + n_tasks = 1; + } break; + case GGML_OP_UPSCALE: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_ATTN: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_FF: + { + n_tasks = n_threads; + } break; + case GGML_OP_FLASH_ATTN_BACK: + { + n_tasks = n_threads; + } break; + case GGML_OP_WIN_PART: + case GGML_OP_WIN_UNPART: + case GGML_OP_GET_REL_POS: + case GGML_OP_MAP_UNARY: + case GGML_OP_MAP_BINARY: + case GGML_OP_MAP_CUSTOM1_F32: + case GGML_OP_MAP_CUSTOM2_F32: + case GGML_OP_MAP_CUSTOM3_F32: + { + n_tasks = 1; + } break; + case GGML_OP_MAP_CUSTOM1: + { + struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_MAP_CUSTOM2: + { + struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_MAP_CUSTOM3: + { + struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params; + if (p->n_tasks == GGML_N_TASKS_MAX) { + n_tasks = n_threads; + } else { + n_tasks = MIN(p->n_tasks, n_threads); + } + } break; + case GGML_OP_CROSS_ENTROPY_LOSS: + { + n_tasks = n_threads; + } break; + case GGML_OP_CROSS_ENTROPY_LOSS_BACK: + { + n_tasks = n_threads; + } break; + case GGML_OP_NONE: + { + n_tasks = 1; + } break; + case GGML_OP_COUNT: + { + GGML_ASSERT(false); + } break; + default: + { + printf("%s: op %s not implemented\n", __func__, ggml_op_name(node->op)); + GGML_ASSERT(false); + } break; + } + + assert(n_tasks > 0); + + return n_tasks; +} + static thread_ret_t ggml_graph_compute_thread(void * data) { struct ggml_compute_state * state = (struct ggml_compute_state *) data; const struct ggml_cgraph * cgraph = state->shared->cgraph; const struct ggml_cplan * cplan = state->shared->cplan; - const int * n_tasks_arr = cplan->n_tasks; const int n_threads = state->shared->n_threads; set_numa_thread_affinity(state->ith, n_threads); @@ -16080,9 +15735,9 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { if (node_n != -1) { /* FINALIZE */ - struct ggml_tensor * node = state->shared->cgraph->nodes[node_n]; + struct ggml_tensor * node = cgraph->nodes[node_n]; if (GGML_OP_HAS_FINALIZE[node->op]) { - params.nth = n_tasks_arr[node_n]; + params.nth = ggml_get_n_tasks(node, n_threads); ggml_compute_forward(¶ms, node); } ggml_graph_compute_perf_stats_node(node, state->shared); @@ -16093,7 +15748,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes); struct ggml_tensor * node = cgraph->nodes[node_n]; - const int n_tasks = n_tasks_arr[node_n]; + const int n_tasks = ggml_get_n_tasks(node, n_threads); state->shared->perf_node_start_cycles = ggml_perf_cycles(); state->shared->perf_node_start_time_us = ggml_perf_time_us(); @@ -16151,7 +15806,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) { /* COMPUTE */ struct ggml_tensor * node = cgraph->nodes[node_n]; - const int n_tasks = n_tasks_arr[node_n]; + const int n_tasks = ggml_get_n_tasks(node, n_threads); struct ggml_compute_params params = { /*.type =*/ GGML_TASK_COMPUTE, @@ -16185,121 +15840,46 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { struct ggml_tensor * node = cgraph->nodes[i]; + size_t cur = 0; + switch (node->op) { case GGML_OP_CPY: case GGML_OP_DUP: { n_tasks = n_threads; - size_t cur = 0; if (ggml_is_quantized(node->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_ADD: case GGML_OP_ADD1: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_ACC: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); } break; - case GGML_OP_SUB: - case GGML_OP_DIV: - case GGML_OP_SQR: - case GGML_OP_SQRT: - case GGML_OP_LOG: - case GGML_OP_SUM: - case GGML_OP_SUM_ROWS: - case GGML_OP_MEAN: - case GGML_OP_ARGMAX: - case GGML_OP_REPEAT: - case GGML_OP_REPEAT_BACK: - { - n_tasks = 1; - } break; - - case GGML_OP_UNARY: - { - switch (ggml_get_unary_op(node)) { - case GGML_UNARY_OP_ABS: - case GGML_UNARY_OP_SGN: - case GGML_UNARY_OP_NEG: - case GGML_UNARY_OP_STEP: - case GGML_UNARY_OP_TANH: - case GGML_UNARY_OP_ELU: - case GGML_UNARY_OP_RELU: - { - n_tasks = 1; - } break; - - case GGML_UNARY_OP_GELU: - case GGML_UNARY_OP_GELU_QUICK: - case GGML_UNARY_OP_SILU: - { - n_tasks = n_threads; - } break; - } - } break; - case GGML_OP_SILU_BACK: - case GGML_OP_MUL: - case GGML_OP_NORM: - case GGML_OP_RMS_NORM: - case GGML_OP_RMS_NORM_BACK: - case GGML_OP_GROUP_NORM: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONCAT: case GGML_OP_MUL_MAT: { - n_tasks = n_threads; - - // TODO: use different scheduling for different matrix sizes - //const int nr0 = ggml_nrows(node->src[0]); - //const int nr1 = ggml_nrows(node->src[1]); - - //n_tasks = MIN(n_threads, MAX(1, nr0/128)); - //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks); - - size_t cur = 0; const enum ggml_type vec_dot_type = type_traits[node->src[0]->type].vec_dot_type; -#if defined(GGML_USE_CUBLAS) - if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning - } else -#elif defined(GGML_USE_CLBLAST) +#if defined(GGML_USE_CLBLAST) if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning cur = ggml_cl_mul_mat_get_wsize(node->src[0], node->src[1], node); } else #endif #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) { - n_tasks = 1; // TODO: this actually is doing nothing - // the threads are still spinning if (node->src[0]->type != GGML_TYPE_F32) { // here we need memory just for single 2D matrix from src0 cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]); @@ -16308,108 +15888,18 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { #endif if (node->src[1]->type != vec_dot_type) { cur = ggml_type_size(vec_dot_type)*ggml_nelements(node->src[1])/ggml_blck_size(vec_dot_type); - } else { - cur = 0; } - - work_size = MAX(work_size, cur); } break; case GGML_OP_OUT_PROD: { n_tasks = n_threads; - size_t cur = 0; - if (ggml_is_quantized(node->src[0]->type)) { cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks; } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_SCALE: - { - n_tasks = 1; - } break; - case GGML_OP_SET: - case GGML_OP_CONT: - case GGML_OP_RESHAPE: - case GGML_OP_VIEW: - case GGML_OP_PERMUTE: - case GGML_OP_TRANSPOSE: - case GGML_OP_GET_ROWS: - case GGML_OP_GET_ROWS_BACK: - case GGML_OP_DIAG: - { - n_tasks = 1; - } break; - case GGML_OP_DIAG_MASK_ZERO: - case GGML_OP_DIAG_MASK_INF: - case GGML_OP_SOFT_MAX: - case GGML_OP_SOFT_MAX_BACK: - case GGML_OP_ROPE: - case GGML_OP_ROPE_BACK: - case GGML_OP_ADD_REL_POS: - { - n_tasks = n_threads; - } break; - case GGML_OP_ALIBI: - { - n_tasks = 1; //TODO - } break; - case GGML_OP_CLAMP: - { - n_tasks = 1; //TODO - } break; - case GGML_OP_CONV_1D: - { - n_tasks = n_threads; - - GGML_ASSERT(node->src[0]->ne[3] == 1); - GGML_ASSERT(node->src[1]->ne[2] == 1); - GGML_ASSERT(node->src[1]->ne[3] == 1); - - const int64_t ne00 = node->src[0]->ne[0]; - const int64_t ne01 = node->src[0]->ne[1]; - const int64_t ne02 = node->src[0]->ne[2]; - - const int64_t ne10 = node->src[1]->ne[0]; - const int64_t ne11 = node->src[1]->ne[1]; - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t nk = ne00; - const int64_t ew0 = nk * ne01; - - UNUSED(ne02); - UNUSED(ne10); - UNUSED(ne11); - - size_t cur = 0; - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(ggml_fp16_t)*(ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)*(ne0*ne1*ew0); - } else { - GGML_ASSERT(false); - } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CONV_1D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_1D_STAGE_1: - { - n_tasks = n_threads; } break; case GGML_OP_CONV_TRANSPOSE_1D: { - n_tasks = n_threads; - GGML_ASSERT(node->src[0]->ne[3] == 1); GGML_ASSERT(node->src[1]->ne[2] == 1); GGML_ASSERT(node->src[1]->ne[3] == 1); @@ -16421,7 +15911,6 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne10 = node->src[1]->ne[0]; // L const int64_t ne11 = node->src[1]->ne[1]; // Cin - size_t cur = 0; if (node->src[0]->type == GGML_TYPE_F16 && node->src[1]->type == GGML_TYPE_F32) { cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02; @@ -16433,59 +15922,13 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { } else { GGML_ASSERT(false); } - - work_size = MAX(work_size, cur); } break; - case GGML_OP_CONV_2D: - { - n_tasks = n_threads; - - const int64_t ne00 = node->src[0]->ne[0]; // W - const int64_t ne01 = node->src[0]->ne[1]; // H - const int64_t ne02 = node->src[0]->ne[2]; // C - const int64_t ne03 = node->src[0]->ne[3]; // N - - const int64_t ne10 = node->src[1]->ne[0]; // W - const int64_t ne11 = node->src[1]->ne[1]; // H - const int64_t ne12 = node->src[1]->ne[2]; // C - - const int64_t ne0 = node->ne[0]; - const int64_t ne1 = node->ne[1]; - const int64_t ne2 = node->ne[2]; - const int64_t ne3 = node->ne[3]; - const int64_t nk = ne00*ne01; - const int64_t ew0 = nk * ne02; - - UNUSED(ne03); - UNUSED(ne2); - - size_t cur = 0; - - if (node->src[0]->type == GGML_TYPE_F16 && - node->src[1]->type == GGML_TYPE_F32) { - // im2col: [N*OH*OW, IC*KH*KW] - cur = sizeof(ggml_fp16_t)*(ne3*ne0*ne1*ew0); - } else if (node->src[0]->type == GGML_TYPE_F32 && - node->src[1]->type == GGML_TYPE_F32) { - cur = sizeof(float)* (ne10*ne11*ne12); - } else { - GGML_ASSERT(false); - } - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CONV_2D_STAGE_0: - { - n_tasks = n_threads; - } break; - case GGML_OP_CONV_2D_STAGE_1: + case GGML_OP_IM2COL: { n_tasks = n_threads; } break; case GGML_OP_CONV_TRANSPOSE_2D: { - n_tasks = n_threads; - const int64_t ne00 = node->src[0]->ne[0]; // W const int64_t ne01 = node->src[0]->ne[1]; // H const int64_t ne02 = node->src[0]->ne[2]; // Channels Out @@ -16495,141 +15938,66 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) { const int64_t ne11 = node->src[1]->ne[1]; // H const int64_t ne12 = node->src[1]->ne[2]; // Channels In - size_t cur = 0; cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03; cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12; - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_POOL_1D: - case GGML_OP_POOL_2D: - { - n_tasks = 1; - } break; - case GGML_OP_UPSCALE: - { - n_tasks = n_threads; } break; case GGML_OP_FLASH_ATTN: { n_tasks = n_threads; - size_t cur = 0; - const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2 } - - work_size = MAX(work_size, cur); } break; case GGML_OP_FLASH_FF: { n_tasks = n_threads; - size_t cur = 0; - if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2 } - - work_size = MAX(work_size, cur); } break; case GGML_OP_FLASH_ATTN_BACK: { n_tasks = n_threads; - size_t cur = 0; - const int64_t D = node->src[0]->ne[0]; const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL); const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back if (node->src[1]->type == GGML_TYPE_F32) { cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 - } - - if (node->src[1]->type == GGML_TYPE_F16) { + } else if (node->src[1]->type == GGML_TYPE_F16) { cur = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1) cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2 } + } break; - work_size = MAX(work_size, cur); - } break; - case GGML_OP_WIN_PART: - case GGML_OP_WIN_UNPART: - case GGML_OP_GET_REL_POS: - case GGML_OP_MAP_UNARY: - case GGML_OP_MAP_BINARY: - case GGML_OP_MAP_CUSTOM1_F32: - case GGML_OP_MAP_CUSTOM2_F32: - case GGML_OP_MAP_CUSTOM3_F32: - { - n_tasks = 1; - } break; - case GGML_OP_MAP_CUSTOM1: - { - struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; - case GGML_OP_MAP_CUSTOM2: - { - struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; - case GGML_OP_MAP_CUSTOM3: - { - struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params; - if (p->n_tasks == GGML_N_TASKS_MAX) { - n_tasks = n_threads; - } else { - n_tasks = MIN(p->n_tasks, n_threads); - } - } break; case GGML_OP_CROSS_ENTROPY_LOSS: { n_tasks = n_threads; - size_t cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); - - work_size = MAX(work_size, cur); - } break; - case GGML_OP_CROSS_ENTROPY_LOSS_BACK: - { - n_tasks = n_threads; - } break; - case GGML_OP_NONE: - { - n_tasks = 1; + cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks); } break; case GGML_OP_COUNT: { GGML_ASSERT(false); } break; + default: + break; } - cplan.n_tasks[i] = n_tasks; + work_size = MAX(work_size, cur); } if (work_size > 0) { @@ -16651,12 +16019,6 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { if (cplan->work_size > 0) { GGML_ASSERT(cplan->work_data); } - - for (int i = 0; i < cgraph->n_nodes; ++i) { - if (cgraph->nodes[i]->op != GGML_OP_NONE) { - GGML_ASSERT(cplan->n_tasks[i] > 0); - } - } } const int n_threads = cplan->n_threads; @@ -16729,16 +16091,6 @@ int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) { return compute_status; } -void ggml_graph_reset(struct ggml_cgraph * cgraph) { - for (int i = 0; i < cgraph->n_nodes; i++) { - struct ggml_tensor * grad = cgraph->grads[i]; - - if (grad) { - ggml_set_zero(grad); - } - } -} - void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) { struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads); @@ -16865,12 +16217,12 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { const uint32_t magic = GGML_FILE_MAGIC; const uint32_t version = GGML_FILE_VERSION; const uint32_t n_leafs = cgraph->n_leafs; - const uint32_t nodes = cgraph->n_nodes; + const uint32_t n_nodes = cgraph->n_nodes; fwrite(&magic, sizeof(uint32_t), 1, fout); fwrite(&version, sizeof(uint32_t), 1, fout); fwrite(&n_leafs, sizeof(uint32_t), 1, fout); - fwrite(&nodes, sizeof(uint32_t), 1, fout); + fwrite(&n_nodes, sizeof(uint32_t), 1, fout); fwrite(&size_eval, sizeof(uint64_t), 1, fout); } @@ -16958,7 +16310,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { if (idx == -1) { for (int k = 0; k < cgraph->n_nodes; ++k) { if (args[j] == cgraph->nodes[k]) { - idx = GGML_MAX_NODES + k; + idx = cgraph->n_leafs + k; break; } } @@ -16985,11 +16337,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) { } } -struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) { +struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) { assert(*ctx_data == NULL); assert(*ctx_eval == NULL); - struct ggml_cgraph result = { 0 }; + struct ggml_cgraph * result = NULL; struct ggml_tensor * data = NULL; @@ -17061,13 +16413,11 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** const uint32_t n_leafs = *(const uint32_t *) ptr; ptr += sizeof(n_leafs); const uint32_t n_nodes = *(const uint32_t *) ptr; ptr += sizeof(n_nodes); const uint64_t size_eval = *(const uint64_t *) ptr; ptr += sizeof(size_eval); - - result.n_leafs = n_leafs; - result.n_nodes = n_nodes; + const int graph_size = MAX(n_leafs, n_nodes); // create the data context { - const size_t overhead = (n_leafs + n_nodes)*ggml_tensor_overhead(); + const size_t overhead = (n_leafs + n_nodes)*ggml_tensor_overhead() + ggml_graph_overhead_custom(graph_size, false); struct ggml_init_params params = { .mem_size = size_eval + overhead, @@ -17083,6 +16433,12 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** } } + result = ggml_new_graph_custom(*ctx_eval, graph_size, false); + + result->n_leafs = n_leafs; + result->n_nodes = n_nodes; + + // leafs { uint32_t type; @@ -17121,7 +16477,7 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** tensor->nb[j] = nb[j]; } - result.leafs[i] = tensor; + result->leafs[i] = tensor; ptr += ggml_nbytes(tensor); @@ -17173,10 +16529,10 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** continue; } - if (arg_idx < GGML_MAX_NODES) { - args[j] = result.leafs[arg_idx]; + if (arg_idx < result->n_leafs) { + args[j] = result->leafs[arg_idx]; } else { - args[j] = result.nodes[arg_idx - GGML_MAX_NODES]; + args[j] = result->nodes[arg_idx - result->n_leafs]; } } @@ -17228,7 +16584,7 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** tensor->src[j] = args[j]; } - result.nodes[i] = tensor; + result->nodes[i] = tensor; fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor)); } @@ -18133,10 +17489,11 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) { case GGML_OPT_ADAM: { result = (struct ggml_opt_params) { - .type = GGML_OPT_ADAM, - .n_threads = 1, - .past = 0, - .delta = 1e-5f, + .type = GGML_OPT_ADAM, + .graph_size = GGML_DEFAULT_GRAPH_SIZE, + .n_threads = 1, // FIXME: GGML_DEFAULT_N_THREADS ? + .past = 0, + .delta = 1e-5f, .max_no_improvement = 100, @@ -18163,10 +17520,11 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) { case GGML_OPT_LBFGS: { result = (struct ggml_opt_params) { - .type = GGML_OPT_LBFGS, - .n_threads = 1, - .past = 0, - .delta = 1e-5f, + .type = GGML_OPT_LBFGS, + .graph_size = GGML_DEFAULT_GRAPH_SIZE, + .n_threads = 1, + .past = 0, + .delta = 1e-5f, .max_no_improvement = 0, @@ -18308,14 +17666,11 @@ enum ggml_opt_result ggml_opt_resume( struct ggml_tensor * f) { // build forward + backward compute graphs - struct ggml_tensor * gfbuf = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32)+ (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); - struct ggml_tensor * gbbuf = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(struct ggml_cgraph) / ggml_type_size(GGML_TYPE_I32)+ (sizeof(struct ggml_cgraph) % ggml_type_size(GGML_TYPE_I32) ? 1 : 0)); + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx, opt->params.graph_size, true); + ggml_build_forward_expand(gf, f); - struct ggml_cgraph * gf = (struct ggml_cgraph *) gfbuf->data; - struct ggml_cgraph * gb = (struct ggml_cgraph *) gbbuf->data; - - *gf = ggml_build_forward (f); - *gb = ggml_build_backward(ctx, gf, true); + struct ggml_cgraph * gb = ggml_graph_dup(ctx, gf); + ggml_build_backward_expand(ctx, gf, gb, true); return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL); } @@ -18679,8 +18034,7 @@ static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) return n == size; } -// NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 -static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset) { +static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) { p->n = 0; p->data = NULL; @@ -18692,19 +18046,6 @@ static bool gguf_fread_str_cur(FILE * file, struct gguf_str * p, size_t * offset return ok; } -static bool gguf_fread_str_v1(FILE * file, struct gguf_str * p, size_t * offset) { - p->n = 0; - p->data = NULL; - - bool ok = true; - - uint32_t n = 0; - ok = ok && gguf_fread_el(file, &n, sizeof(n), offset); p->data = calloc(n + 1, 1); p->n = n; - ok = ok && gguf_fread_el(file, p->data, p->n, offset); - - return ok; -} - struct gguf_context * gguf_init_empty(void) { struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context)); @@ -18763,20 +18104,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ctx->data = NULL; ok = ok && gguf_fread_el(file, &ctx->header.version, sizeof(ctx->header.version), &offset); + ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); + ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t n_tensors = 0; - uint32_t n_kv = 0; - - ok = ok && gguf_fread_el(file, &n_tensors, sizeof(n_tensors), &offset); - ok = ok && gguf_fread_el(file, &n_kv, sizeof(n_kv), &offset); - - ctx->header.n_tensors = n_tensors; - ctx->header.n_kv = n_kv; - } else { - ok = ok && gguf_fread_el(file, &ctx->header.n_tensors, sizeof(ctx->header.n_tensors), &offset); - ok = ok && gguf_fread_el(file, &ctx->header.n_kv, sizeof(ctx->header.n_kv), &offset); + fprintf(stderr, "%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__); + fclose(file); + gguf_free(ctx); + return NULL; } if (!ok) { @@ -18787,17 +18122,11 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p } } - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - bool (* gguf_fread_str)(FILE *, struct gguf_str *, size_t *) = gguf_fread_str_cur; - if (ctx->header.version == 1) { - gguf_fread_str = gguf_fread_str_v1; - } - // read the kv pairs { ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv)); - for (uint32_t i = 0; i < ctx->header.n_kv; ++i) { + for (uint64_t i = 0; i < ctx->header.n_kv; ++i) { struct gguf_kv * kv = &ctx->kv[i]; //fprintf(stderr, "%s: reading kv %d\n", __func__, i); @@ -18823,15 +18152,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_ARRAY: { ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset); - - if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t n = 0; - ok = ok && gguf_fread_el(file, &n, sizeof(n), &offset); - kv->value.arr.n = n; - } else { - ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); - } + ok = ok && gguf_fread_el(file, &kv->value.arr.n, sizeof(kv->value.arr.n), &offset); switch (kv->value.arr.type) { case GGUF_TYPE_UINT8: @@ -18852,7 +18173,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p case GGUF_TYPE_STRING: { kv->value.arr.data = malloc(kv->value.arr.n * sizeof(struct gguf_str)); - for (uint32_t j = 0; j < kv->value.arr.n; ++j) { + for (uint64_t j = 0; j < kv->value.arr.n; ++j) { ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset); } } break; @@ -18880,7 +18201,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p { ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info)); - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { struct gguf_tensor_info * info = &ctx->infos[i]; for (int j = 0; j < GGML_MAX_DIMS; ++j) { @@ -18890,14 +18211,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ok = ok && gguf_fread_str(file, &info->name, &offset); ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims), &offset); for (uint32_t j = 0; j < info->n_dims; ++j) { - if (ctx->header.version == 1) { - // NOTE: temporary handling of GGUFv1 >> remove after Oct 2023 - uint32_t t = 0; - ok = ok && gguf_fread_el(file, &t, sizeof(t), &offset); - info->ne[j] = t; - } else { - ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); - } + ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset); } ok = ok && gguf_fread_el (file, &info->type, sizeof(info->type), &offset); ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset), &offset); @@ -18934,7 +18248,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p // compute the total size of the data section, taking into account the alignment { ctx->size = 0; - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { struct gguf_tensor_info * info = &ctx->infos[i]; const int64_t ne = @@ -19003,7 +18317,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p ggml_set_no_alloc(ctx_data, true); // create the tensors - for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) { + for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) { const int64_t ne[GGML_MAX_DIMS] = { ctx->infos[i].ne[0], ctx->infos[i].ne[1], @@ -19138,24 +18452,29 @@ int gguf_find_key(const struct gguf_context * ctx, const char * key) { } const char * gguf_get_key(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); return ctx->kv[key_id].key.data; } enum gguf_type gguf_get_kv_type(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); return ctx->kv[key_id].type; } enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.type; } const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.data; } const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); struct gguf_kv * kv = &ctx->kv[key_id]; struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[i]; @@ -19163,70 +18482,90 @@ const char * gguf_get_arr_str(const struct gguf_context * ctx, int key_id, int i } int gguf_get_arr_n(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_ARRAY); return ctx->kv[key_id].value.arr.n; } uint8_t gguf_get_val_u8(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT8); return ctx->kv[key_id].value.uint8; } int8_t gguf_get_val_i8(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT8); return ctx->kv[key_id].value.int8; } uint16_t gguf_get_val_u16(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT16); return ctx->kv[key_id].value.uint16; } int16_t gguf_get_val_i16(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT16); return ctx->kv[key_id].value.int16; } uint32_t gguf_get_val_u32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT32); return ctx->kv[key_id].value.uint32; } int32_t gguf_get_val_i32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT32); return ctx->kv[key_id].value.int32; } float gguf_get_val_f32(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT32); return ctx->kv[key_id].value.float32; } uint64_t gguf_get_val_u64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_UINT64); return ctx->kv[key_id].value.uint64; } int64_t gguf_get_val_i64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_INT64); return ctx->kv[key_id].value.int64; } double gguf_get_val_f64(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_FLOAT64); return ctx->kv[key_id].value.float64; } bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_BOOL); return ctx->kv[key_id].value.bool_; } const char * gguf_get_val_str(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); GGML_ASSERT(ctx->kv[key_id].type == GGUF_TYPE_STRING); return ctx->kv[key_id].value.str.data; } +const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id) { + GGML_ASSERT(key_id >= 0 && key_id < gguf_get_n_kv(ctx)); + GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_ARRAY); + GGML_ASSERT(ctx->kv[key_id].type != GGUF_TYPE_STRING); + return &ctx->kv[key_id].value; +} + int gguf_get_n_tensors(const struct gguf_context * ctx) { return ctx->header.n_tensors; } diff --git a/ggml.h b/ggml.h index 8c954904e..f2fce0f22 100644 --- a/ggml.h +++ b/ggml.h @@ -58,7 +58,8 @@ // { // ... // -// struct ggml_cgraph gf = ggml_build_forward(f); +// struct ggml_cgraph * gf = ggml_new_graph(ctx); +// ggml_build_forward_expand(gf, f); // // // set the input variable and parameter values // ggml_set_f32(x, 2.0f); @@ -213,15 +214,14 @@ #define GGML_QNT_VERSION 2 // bump this on quantization format changes #define GGML_QNT_VERSION_FACTOR 1000 // do not change this -#define GGML_MAX_DIMS 4 -#define GGML_MAX_NODES 16384 -#define GGML_MAX_PARAMS 1024 -#define GGML_MAX_CONTEXTS 64 -#define GGML_MAX_SRC 6 -#define GGML_MAX_NAME 64 -#define GGML_MAX_OP_PARAMS 32 -#define GGML_DEFAULT_N_THREADS 4 - +#define GGML_MAX_DIMS 4 +#define GGML_MAX_PARAMS 1024 +#define GGML_MAX_CONTEXTS 64 +#define GGML_MAX_SRC 6 +#define GGML_MAX_NAME 64 +#define GGML_MAX_OP_PARAMS 64 +#define GGML_DEFAULT_N_THREADS 4 +#define GGML_DEFAULT_GRAPH_SIZE 2048 #if UINTPTR_MAX == 0xFFFFFFFF #define GGML_MEM_ALIGN 4 #else @@ -245,7 +245,10 @@ do { \ if (!(x)) { \ fprintf(stderr, "GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \ - abort(); \ + fflush(stderr); \ + fflush(stdout); \ + ggml_print_backtrace(); \ + exit(1); \ } \ } while (0) @@ -400,13 +403,8 @@ extern "C" { GGML_OP_ROPE_BACK, GGML_OP_ALIBI, GGML_OP_CLAMP, - GGML_OP_CONV_1D, - GGML_OP_CONV_1D_STAGE_0, // internal - GGML_OP_CONV_1D_STAGE_1, // internal GGML_OP_CONV_TRANSPOSE_1D, - GGML_OP_CONV_2D, - GGML_OP_CONV_2D_STAGE_0, // internal - GGML_OP_CONV_2D_STAGE_1, // internal + GGML_OP_IM2COL, GGML_OP_CONV_TRANSPOSE_2D, GGML_OP_POOL_1D, GGML_OP_POOL_2D, @@ -451,6 +449,7 @@ extern "C" { GGML_UNARY_OP_GELU, GGML_UNARY_OP_GELU_QUICK, GGML_UNARY_OP_SILU, + GGML_UNARY_OP_LEAKY }; enum ggml_object_type { @@ -531,37 +530,33 @@ extern "C" { int n_threads; - // the `n_tasks` of nodes, 1:1 mapping to cgraph nodes - int n_tasks[GGML_MAX_NODES]; - // abort ggml_graph_compute when true bool (*abort_callback)(void * data); void * abort_callback_data; }; - // next prime after GGML_MAX_NODES - // #define GGML_GRAPH_HASHTABLE_SIZE 4099 - // next prime after GGML_MAX_NODES * 2 (nodes + leafs) - // #define GGML_GRAPH_HASHTABLE_SIZE 8273 - // #define GGML_GRAPH_HASHTABLE_SIZE 16411 - #define GGML_GRAPH_HASHTABLE_SIZE 32771 - enum ggml_cgraph_eval_order { GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0, GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT, GGML_CGRAPH_EVAL_ORDER_COUNT }; + struct ggml_hash_set { + size_t size; + struct ggml_tensor ** keys; + }; + // computation graph struct ggml_cgraph { + int size; int n_nodes; int n_leafs; - struct ggml_tensor * nodes[GGML_MAX_NODES]; - struct ggml_tensor * grads[GGML_MAX_NODES]; - struct ggml_tensor * leafs[GGML_MAX_NODES]; + struct ggml_tensor ** nodes; + struct ggml_tensor ** grads; + struct ggml_tensor ** leafs; - void * visited_hash_table[GGML_GRAPH_HASHTABLE_SIZE]; + struct ggml_hash_set visited_hash_table; enum ggml_cgraph_eval_order order; @@ -571,8 +566,6 @@ extern "C" { int64_t perf_time_us; }; - static const size_t GGML_GRAPH_SIZE = sizeof(struct ggml_cgraph); - // scratch buffer struct ggml_scratch { size_t offs; @@ -617,6 +610,8 @@ extern "C" { GGML_API int64_t ggml_cycles(void); GGML_API int64_t ggml_cycles_per_ms(void); + GGML_API void ggml_print_backtrace(void); + GGML_API void ggml_numa_init(void); // call once for better performance on NUMA systems GGML_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node @@ -943,6 +938,10 @@ extern "C" { struct ggml_context * ctx, struct ggml_tensor * a); + GGML_API struct ggml_tensor * ggml_leaky( + struct ggml_context * ctx, + struct ggml_tensor * a); + GGML_API struct ggml_tensor * ggml_relu_inplace( struct ggml_context * ctx, struct ggml_tensor * a); @@ -1326,8 +1325,13 @@ extern "C" { int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); // in-place, returns view(a) GGML_API struct ggml_tensor * ggml_rope_custom_inplace( @@ -1337,8 +1341,17 @@ extern "C" { int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, - float freq_scale); + float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow); + + // compute correction dims for YaRN RoPE scaling + void ggml_rope_yarn_corr_dims( + int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]); // xPos RoPE, in-place, returns view(a) GGML_API struct ggml_tensor * ggml_rope_xpos_inplace( @@ -1358,8 +1371,13 @@ extern "C" { int n_dims, int mode, int n_ctx, + int n_orig_ctx, float freq_base, float freq_scale, + float ext_factor, + float attn_factor, + float beta_fast, + float beta_slow, float xpos_base, bool xpos_down); @@ -1380,6 +1398,18 @@ extern "C" { float min, float max); + GGML_API struct ggml_tensor * ggml_im2col( + struct ggml_context * ctx, + struct ggml_tensor * a, + struct ggml_tensor * b, + int s0, + int s1, + int p0, + int p1, + int d0, + int d1, + bool is_2D); + GGML_API struct ggml_tensor * ggml_conv_1d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -1463,6 +1493,8 @@ extern "C" { int s0, // stride int p0); // padding + // the result will have 2*p0 padding for the first dimension + // and 2*p1 padding for the second dimension GGML_API struct ggml_tensor * ggml_pool_2d( struct ggml_context * ctx, struct ggml_tensor * a, @@ -1471,8 +1503,8 @@ extern "C" { int k1, int s0, int s1, - int p0, - int p1); + float p0, + float p1); // nearest interpolate // used in stable-diffusion @@ -1713,19 +1745,22 @@ extern "C" { GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor); GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep); - GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor); - GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep); - // graph allocation in a context - GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); - GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor); + GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false + GGML_API struct ggml_cgraph * ggml_new_graph_custom (struct ggml_context * ctx, size_t size, bool grads); + GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph); + GGML_API struct ggml_cgraph * ggml_graph_view (struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i0, int i1); + GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst); + GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // zero grads + GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph); + GGML_API size_t ggml_graph_overhead(void); + GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads); // ggml_graph_plan() has to be called before ggml_graph_compute() // when plan.work_size > 0, caller must allocate memory for plan.work_data GGML_API struct ggml_cplan ggml_graph_plan (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/); - GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); - GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); + GGML_API int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan); // same as ggml_graph_compute() but the work data is allocated as a part of the context // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data @@ -1733,8 +1768,8 @@ extern "C" { GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name); - GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname); - GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval); + GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname); + GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval); // print info and performance information for the graph GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph); @@ -1797,6 +1832,8 @@ extern "C" { struct ggml_opt_params { enum ggml_opt_type type; + size_t graph_size; + int n_threads; // delta-based convergence test @@ -2008,6 +2045,7 @@ extern "C" { GGML_API double gguf_get_val_f64 (const struct gguf_context * ctx, int key_id); GGML_API bool gguf_get_val_bool(const struct gguf_context * ctx, int key_id); GGML_API const char * gguf_get_val_str (const struct gguf_context * ctx, int key_id); + GGML_API const void * gguf_get_val_data(const struct gguf_context * ctx, int key_id); GGML_API int gguf_get_arr_n (const struct gguf_context * ctx, int key_id); GGML_API const void * gguf_get_arr_data(const struct gguf_context * ctx, int key_id); GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int key_id, int i); diff --git a/gguf-py/README.md b/gguf-py/README.md index a28d8c57a..502b6a510 100644 --- a/gguf-py/README.md +++ b/gguf-py/README.md @@ -11,6 +11,16 @@ as an example for its usage. pip install gguf ``` +## API Examples/Simple Tools + +[examples/writer.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/examples/writer.py) β€” Generates `example.gguf` in the current directory to demonstrate generating a GGUF file. Note that this file cannot be used as a model. + +[scripts/gguf-dump.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-dump.py) β€” Dumps a GGUF file's metadata to the console. + +[scripts/gguf-set-metadata.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-set-metadata.py) β€” Allows changing simple metadata values in a GGUF file by key. + +[scripts/gguf-convert-endian.py](https://github.com/ggerganov/llama.cpp/blob/master/gguf-py/scripts/gguf-convert-endian.py) β€” Allows converting the endianness of GGUF files. + ## Development Maintainers who participate in development of this package are advised to install it in editable mode: diff --git a/gguf-py/examples/writer.py b/gguf-py/examples/writer.py new file mode 100755 index 000000000..f39eed1af --- /dev/null +++ b/gguf-py/examples/writer.py @@ -0,0 +1,40 @@ +#!/usr/bin/env python3 +import sys +from pathlib import Path + +import numpy as np + +# Necessary to load the local gguf package +sys.path.insert(0, str(Path(__file__).parent.parent)) + +from gguf import GGUFWriter # noqa: E402 + + +# Example usage: +def writer_example() -> None: + # Example usage with a file + gguf_writer = GGUFWriter("example.gguf", "llama") + + gguf_writer.add_architecture() + gguf_writer.add_block_count(12) + gguf_writer.add_uint32("answer", 42) # Write a 32-bit integer + gguf_writer.add_float32("answer_in_float", 42.0) # Write a 32-bit float + gguf_writer.add_custom_alignment(64) + + tensor1 = np.ones((32,), dtype=np.float32) * 100.0 + tensor2 = np.ones((64,), dtype=np.float32) * 101.0 + tensor3 = np.ones((96,), dtype=np.float32) * 102.0 + + gguf_writer.add_tensor("tensor1", tensor1) + gguf_writer.add_tensor("tensor2", tensor2) + gguf_writer.add_tensor("tensor3", tensor3) + + gguf_writer.write_header_to_file() + gguf_writer.write_kv_data_to_file() + gguf_writer.write_tensors_to_file() + + gguf_writer.close() + + +if __name__ == '__main__': + writer_example() diff --git a/gguf-py/gguf/__init__.py b/gguf-py/gguf/__init__.py index f9b70a85b..110ab342c 100644 --- a/gguf-py/gguf/__init__.py +++ b/gguf-py/gguf/__init__.py @@ -1 +1,5 @@ -from .gguf import * +from .constants import * +from .gguf_reader import * +from .gguf_writer import * +from .tensor_mapping import * +from .vocab import * diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py new file mode 100644 index 000000000..8bd82daca --- /dev/null +++ b/gguf-py/gguf/constants.py @@ -0,0 +1,488 @@ +from __future__ import annotations + +import sys +from enum import Enum, IntEnum, auto +from typing import Any + +# +# constants +# + +GGUF_MAGIC = 0x46554747 # "GGUF" +GGUF_VERSION = 3 +GGUF_DEFAULT_ALIGNMENT = 32 + +# +# metadata keys +# + + +class Keys: + class General: + ARCHITECTURE = "general.architecture" + QUANTIZATION_VERSION = "general.quantization_version" + ALIGNMENT = "general.alignment" + NAME = "general.name" + AUTHOR = "general.author" + URL = "general.url" + DESCRIPTION = "general.description" + LICENSE = "general.license" + SOURCE_URL = "general.source.url" + SOURCE_HF_REPO = "general.source.huggingface.repository" + FILE_TYPE = "general.file_type" + + class LLM: + CONTEXT_LENGTH = "{arch}.context_length" + EMBEDDING_LENGTH = "{arch}.embedding_length" + BLOCK_COUNT = "{arch}.block_count" + FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" + USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" + TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" + + class Attention: + HEAD_COUNT = "{arch}.attention.head_count" + HEAD_COUNT_KV = "{arch}.attention.head_count_kv" + MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" + CLAMP_KQV = "{arch}.attention.clamp_kqv" + LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" + LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" + + class Rope: + DIMENSION_COUNT = "{arch}.rope.dimension_count" + FREQ_BASE = "{arch}.rope.freq_base" + SCALING_TYPE = "{arch}.rope.scaling.type" + SCALING_FACTOR = "{arch}.rope.scaling.factor" + SCALING_ORIG_CTX_LEN = "{arch}.rope.scaling.original_context_length" + SCALING_FINETUNED = "{arch}.rope.scaling.finetuned" + + class Tokenizer: + MODEL = "tokenizer.ggml.model" + LIST = "tokenizer.ggml.tokens" + TOKEN_TYPE = "tokenizer.ggml.token_type" + SCORES = "tokenizer.ggml.scores" + MERGES = "tokenizer.ggml.merges" + BOS_ID = "tokenizer.ggml.bos_token_id" + EOS_ID = "tokenizer.ggml.eos_token_id" + UNK_ID = "tokenizer.ggml.unknown_token_id" + SEP_ID = "tokenizer.ggml.seperator_token_id" + PAD_ID = "tokenizer.ggml.padding_token_id" + ADD_BOS = "tokenizer.ggml.add_bos_token" + ADD_EOS = "tokenizer.ggml.add_eos_token" + HF_JSON = "tokenizer.huggingface.json" + RWKV = "tokenizer.rwkv.world" + CHAT_TEMPLATE = "tokenizer.chat_template" + + +# +# recommended mapping of model tensor names for storage in gguf +# + + +class MODEL_ARCH(IntEnum): + LLAMA = auto() + FALCON = auto() + BAICHUAN = auto() + GPT2 = auto() + GPTJ = auto() + GPTNEOX = auto() + MPT = auto() + STARCODER = auto() + PERSIMMON = auto() + REFACT = auto() + BERT = auto() + BLOOM = auto() + STABLELM = auto() + + +class MODEL_TENSOR(IntEnum): + TOKEN_EMBD = auto() + TOKEN_EMBD_NORM = auto() + TOKEN_TYPES = auto() + POS_EMBD = auto() + OUTPUT = auto() + OUTPUT_NORM = auto() + ROPE_FREQS = auto() + ATTN_Q = auto() + ATTN_K = auto() + ATTN_V = auto() + ATTN_QKV = auto() + ATTN_OUT = auto() + ATTN_NORM = auto() + ATTN_NORM_2 = auto() + ATTN_ROT_EMBD = auto() + FFN_GATE = auto() + FFN_DOWN = auto() + FFN_UP = auto() + FFN_NORM = auto() + ATTN_Q_NORM = auto() + ATTN_K_NORM = auto() + + +MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { + MODEL_ARCH.LLAMA: "llama", + MODEL_ARCH.FALCON: "falcon", + MODEL_ARCH.BAICHUAN: "baichuan", + MODEL_ARCH.GPT2: "gpt2", + MODEL_ARCH.GPTJ: "gptj", + MODEL_ARCH.GPTNEOX: "gptneox", + MODEL_ARCH.MPT: "mpt", + MODEL_ARCH.STARCODER: "starcoder", + MODEL_ARCH.PERSIMMON: "persimmon", + MODEL_ARCH.REFACT: "refact", + MODEL_ARCH.BERT: "bert", + MODEL_ARCH.BLOOM: "bloom", + MODEL_ARCH.STABLELM: "stablelm", +} + +TENSOR_NAMES: dict[MODEL_TENSOR, str] = { + MODEL_TENSOR.TOKEN_EMBD: "token_embd", + MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm", + MODEL_TENSOR.TOKEN_TYPES: "token_types", + MODEL_TENSOR.POS_EMBD: "position_embd", + MODEL_TENSOR.OUTPUT_NORM: "output_norm", + MODEL_TENSOR.OUTPUT: "output", + MODEL_TENSOR.ROPE_FREQS: "rope_freqs", + MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", + MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2", + MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", + MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q", + MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k", + MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v", + MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", + MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd", + MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm", + MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm", + MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", + MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", + MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", + MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", +} + +MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { + MODEL_ARCH.LLAMA: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GPTNEOX: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.FALCON: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_NORM_2, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BAICHUAN: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.STARCODER: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BERT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_TYPES, + MODEL_TENSOR.POS_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.MPT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GPTJ: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.PERSIMMON: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + MODEL_TENSOR.ATTN_Q_NORM, + MODEL_TENSOR.ATTN_K_NORM, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.REFACT: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.BLOOM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.TOKEN_EMBD_NORM, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_QKV, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.STABLELM: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], + MODEL_ARCH.GPT2: [ + # TODO + ], + # TODO +} + +# tensors that will not be serialized +MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { + MODEL_ARCH.LLAMA: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.BAICHUAN: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], + MODEL_ARCH.PERSIMMON: [ + MODEL_TENSOR.ROPE_FREQS, + ], +} + +# +# types +# + + +class TokenType(IntEnum): + NORMAL = 1 + UNKNOWN = 2 + CONTROL = 3 + USER_DEFINED = 4 + UNUSED = 5 + BYTE = 6 + + +class RopeScalingType(Enum): + NONE = 'none' + LINEAR = 'linear' + YARN = 'yarn' + + +class GGMLQuantizationType(IntEnum): + F32 = 0 + F16 = 1 + Q4_0 = 2 + Q4_1 = 3 + Q5_0 = 6 + Q5_1 = 7 + Q8_0 = 8 + Q8_1 = 9 + Q2_K = 10 + Q3_K = 11 + Q4_K = 12 + Q5_K = 13 + Q6_K = 14 + Q8_K = 15 + + +class GGUFEndian(IntEnum): + LITTLE = 0 + BIG = 1 + + +class GGUFValueType(IntEnum): + UINT8 = 0 + INT8 = 1 + UINT16 = 2 + INT16 = 3 + UINT32 = 4 + INT32 = 5 + FLOAT32 = 6 + BOOL = 7 + STRING = 8 + ARRAY = 9 + UINT64 = 10 + INT64 = 11 + FLOAT64 = 12 + + @staticmethod + def get_type(val: Any) -> GGUFValueType: + if isinstance(val, (str, bytes, bytearray)): + return GGUFValueType.STRING + elif isinstance(val, list): + return GGUFValueType.ARRAY + elif isinstance(val, float): + return GGUFValueType.FLOAT32 + elif isinstance(val, bool): + return GGUFValueType.BOOL + elif isinstance(val, int): + return GGUFValueType.INT32 + # TODO: need help with 64-bit types in Python + else: + print("Unknown type:", type(val)) + sys.exit() + + +# Note: Does not support GGML_QKK_64 +QK_K = 256 +# Items here are (block size, type size) +GGML_QUANT_SIZES = { + GGMLQuantizationType.F32: (1, 4), + GGMLQuantizationType.F16: (1, 2), + GGMLQuantizationType.Q4_0: (32, 2 + 16), + GGMLQuantizationType.Q4_1: (32, 2 + 2 + 16), + GGMLQuantizationType.Q5_0: (32, 2 + 4 + 16), + GGMLQuantizationType.Q5_1: (32, 2 + 2 + 4 + 16), + GGMLQuantizationType.Q8_0: (32, 2 + 32), + GGMLQuantizationType.Q8_1: (32, 4 + 4 + 32), + GGMLQuantizationType.Q2_K: (256, 2 + 2 + QK_K // 16 + QK_K // 4), + GGMLQuantizationType.Q3_K: (256, 2 + QK_K // 4 + QK_K // 8 + 12), + GGMLQuantizationType.Q4_K: (256, 2 + 2 + QK_K // 2 + 12), + GGMLQuantizationType.Q5_K: (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12), + GGMLQuantizationType.Q6_K: (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16), + GGMLQuantizationType.Q8_K: (256, 4 + QK_K + QK_K // 8), +} + + +# Aliases for backward compatibility. + +# general +KEY_GENERAL_ARCHITECTURE = Keys.General.ARCHITECTURE +KEY_GENERAL_QUANTIZATION_VERSION = Keys.General.QUANTIZATION_VERSION +KEY_GENERAL_ALIGNMENT = Keys.General.ALIGNMENT +KEY_GENERAL_NAME = Keys.General.NAME +KEY_GENERAL_AUTHOR = Keys.General.AUTHOR +KEY_GENERAL_URL = Keys.General.URL +KEY_GENERAL_DESCRIPTION = Keys.General.DESCRIPTION +KEY_GENERAL_LICENSE = Keys.General.LICENSE +KEY_GENERAL_SOURCE_URL = Keys.General.SOURCE_URL +KEY_GENERAL_SOURCE_HF_REPO = Keys.General.SOURCE_HF_REPO +KEY_GENERAL_FILE_TYPE = Keys.General.FILE_TYPE + +# LLM +KEY_CONTEXT_LENGTH = Keys.LLM.CONTEXT_LENGTH +KEY_EMBEDDING_LENGTH = Keys.LLM.EMBEDDING_LENGTH +KEY_BLOCK_COUNT = Keys.LLM.BLOCK_COUNT +KEY_FEED_FORWARD_LENGTH = Keys.LLM.FEED_FORWARD_LENGTH +KEY_USE_PARALLEL_RESIDUAL = Keys.LLM.USE_PARALLEL_RESIDUAL +KEY_TENSOR_DATA_LAYOUT = Keys.LLM.TENSOR_DATA_LAYOUT + +# attention +KEY_ATTENTION_HEAD_COUNT = Keys.Attention.HEAD_COUNT +KEY_ATTENTION_HEAD_COUNT_KV = Keys.Attention.HEAD_COUNT_KV +KEY_ATTENTION_MAX_ALIBI_BIAS = Keys.Attention.MAX_ALIBI_BIAS +KEY_ATTENTION_CLAMP_KQV = Keys.Attention.CLAMP_KQV +KEY_ATTENTION_LAYERNORM_EPS = Keys.Attention.LAYERNORM_EPS +KEY_ATTENTION_LAYERNORM_RMS_EPS = Keys.Attention.LAYERNORM_RMS_EPS + +# RoPE +KEY_ROPE_DIMENSION_COUNT = Keys.Rope.DIMENSION_COUNT +KEY_ROPE_FREQ_BASE = Keys.Rope.FREQ_BASE +KEY_ROPE_SCALING_TYPE = Keys.Rope.SCALING_TYPE +KEY_ROPE_SCALING_FACTOR = Keys.Rope.SCALING_FACTOR +KEY_ROPE_SCALING_ORIG_CTX_LEN = Keys.Rope.SCALING_ORIG_CTX_LEN +KEY_ROPE_SCALING_FINETUNED = Keys.Rope.SCALING_FINETUNED + +# tokenization +KEY_TOKENIZER_MODEL = Keys.Tokenizer.MODEL +KEY_TOKENIZER_LIST = Keys.Tokenizer.LIST +KEY_TOKENIZER_TOKEN_TYPE = Keys.Tokenizer.TOKEN_TYPE +KEY_TOKENIZER_SCORES = Keys.Tokenizer.SCORES +KEY_TOKENIZER_MERGES = Keys.Tokenizer.MERGES +KEY_TOKENIZER_BOS_ID = Keys.Tokenizer.BOS_ID +KEY_TOKENIZER_EOS_ID = Keys.Tokenizer.EOS_ID +KEY_TOKENIZER_UNK_ID = Keys.Tokenizer.UNK_ID +KEY_TOKENIZER_SEP_ID = Keys.Tokenizer.SEP_ID +KEY_TOKENIZER_PAD_ID = Keys.Tokenizer.PAD_ID +KEY_TOKENIZER_HF_JSON = Keys.Tokenizer.HF_JSON +KEY_TOKENIZER_RWKV = Keys.Tokenizer.RWKV diff --git a/gguf-py/gguf/gguf.py b/gguf-py/gguf/gguf.py index 6b7d65429..651a81eb8 100644 --- a/gguf-py/gguf/gguf.py +++ b/gguf-py/gguf/gguf.py @@ -1,1103 +1,15 @@ -#!/usr/bin/env python3 -from __future__ import annotations +# This file left for compatibility. If you want to use the GGUF API from Python +# then don't import gguf/gguf.py directly. If you're looking for examples, see the +# examples/ directory for gguf-py -import json -import os -import shutil -import struct +import importlib import sys -import tempfile -from enum import IntEnum, auto -from io import BufferedWriter from pathlib import Path -from typing import IO, Any, BinaryIO, Callable, Sequence -import numpy as np +sys.path.insert(0, str(Path(__file__).parent.parent)) -# -# constants -# +# Compatibility for people trying to import gguf/gguf.py directly instead of as a package. +importlib.invalidate_caches() +import gguf # noqa: E402 -GGUF_MAGIC = 0x46554747 -GGUF_VERSION = 3 -GGUF_DEFAULT_ALIGNMENT = 32 - - -# general -KEY_GENERAL_ARCHITECTURE = "general.architecture" -KEY_GENERAL_QUANTIZATION_VERSION = "general.quantization_version" -KEY_GENERAL_ALIGNMENT = "general.alignment" -KEY_GENERAL_NAME = "general.name" -KEY_GENERAL_AUTHOR = "general.author" -KEY_GENERAL_URL = "general.url" -KEY_GENERAL_DESCRIPTION = "general.description" -KEY_GENERAL_LICENSE = "general.license" -KEY_GENERAL_SOURCE_URL = "general.source.url" -KEY_GENERAL_SOURCE_HF_REPO = "general.source.huggingface.repository" -KEY_GENERAL_FILE_TYPE = "general.file_type" - -# LLM -KEY_CONTEXT_LENGTH = "{arch}.context_length" -KEY_EMBEDDING_LENGTH = "{arch}.embedding_length" -KEY_BLOCK_COUNT = "{arch}.block_count" -KEY_FEED_FORWARD_LENGTH = "{arch}.feed_forward_length" -KEY_USE_PARALLEL_RESIDUAL = "{arch}.use_parallel_residual" -KEY_TENSOR_DATA_LAYOUT = "{arch}.tensor_data_layout" - -# attention -KEY_ATTENTION_HEAD_COUNT = "{arch}.attention.head_count" -KEY_ATTENTION_HEAD_COUNT_KV = "{arch}.attention.head_count_kv" -KEY_ATTENTION_MAX_ALIBI_BIAS = "{arch}.attention.max_alibi_bias" -KEY_ATTENTION_CLAMP_KQV = "{arch}.attention.clamp_kqv" -KEY_ATTENTION_LAYERNORM_EPS = "{arch}.attention.layer_norm_epsilon" -KEY_ATTENTION_LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon" - -# RoPE -KEY_ROPE_DIMENSION_COUNT = "{arch}.rope.dimension_count" -KEY_ROPE_FREQ_BASE = "{arch}.rope.freq_base" -KEY_ROPE_SCALE_LINEAR = "{arch}.rope.scale_linear" - -# tokenization -KEY_TOKENIZER_MODEL = "tokenizer.ggml.model" -KEY_TOKENIZER_LIST = "tokenizer.ggml.tokens" -KEY_TOKENIZER_TOKEN_TYPE = "tokenizer.ggml.token_type" -KEY_TOKENIZER_SCORES = "tokenizer.ggml.scores" -KEY_TOKENIZER_MERGES = "tokenizer.ggml.merges" -KEY_TOKENIZER_BOS_ID = "tokenizer.ggml.bos_token_id" -KEY_TOKENIZER_EOS_ID = "tokenizer.ggml.eos_token_id" -KEY_TOKENIZER_UNK_ID = "tokenizer.ggml.unknown_token_id" -KEY_TOKENIZER_SEP_ID = "tokenizer.ggml.seperator_token_id" -KEY_TOKENIZER_PAD_ID = "tokenizer.ggml.padding_token_id" -KEY_TOKENIZER_HF_JSON = "tokenizer.huggingface.json" -KEY_TOKENIZER_RWKV = "tokenizer.rwkv.world" - - -# -# recommended mapping of model tensor names for storage in gguf -# - - -class MODEL_ARCH(IntEnum): - LLAMA : int = auto() - FALCON : int = auto() - BAICHUAN : int = auto() - GPT2 : int = auto() - GPTJ : int = auto() - GPTNEOX : int = auto() - MPT : int = auto() - STARCODER : int = auto() - PERSIMMON : int = auto() - REFACT : int = auto() - BERT : int = auto() - BLOOM : int = auto() - - -class MODEL_TENSOR(IntEnum): - TOKEN_EMBD : int = auto() - TOKEN_EMBD_NORM : int = auto() - TOKEN_TYPES : int = auto() - POS_EMBD : int = auto() - OUTPUT : int = auto() - OUTPUT_NORM : int = auto() - ROPE_FREQS : int = auto() - ATTN_Q : int = auto() - ATTN_K : int = auto() - ATTN_V : int = auto() - ATTN_QKV : int = auto() - ATTN_OUT : int = auto() - ATTN_NORM : int = auto() - ATTN_NORM_2 : int = auto() - ATTN_ROT_EMBD : int = auto() - FFN_GATE : int = auto() - FFN_DOWN : int = auto() - FFN_UP : int = auto() - FFN_NORM : int = auto() - ATTN_Q_NORM : int = auto() - ATTN_K_NORM : int = auto() - - -MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { - MODEL_ARCH.LLAMA: "llama", - MODEL_ARCH.FALCON: "falcon", - MODEL_ARCH.BAICHUAN: "baichuan", - MODEL_ARCH.GPT2: "gpt2", - MODEL_ARCH.GPTJ: "gptj", - MODEL_ARCH.GPTNEOX: "gptneox", - MODEL_ARCH.MPT: "mpt", - MODEL_ARCH.STARCODER: "starcoder", - MODEL_ARCH.PERSIMMON: "persimmon", - MODEL_ARCH.REFACT: "refact", - MODEL_ARCH.BERT: "bert", - MODEL_ARCH.BLOOM: "bloom", -} - -TENSOR_NAMES: dict[MODEL_TENSOR, str] = { - MODEL_TENSOR.TOKEN_EMBD: "token_embd", - MODEL_TENSOR.TOKEN_EMBD_NORM: "token_embd_norm", - MODEL_TENSOR.TOKEN_TYPES: "token_types", - MODEL_TENSOR.POS_EMBD: "position_embd", - MODEL_TENSOR.OUTPUT_NORM: "output_norm", - MODEL_TENSOR.OUTPUT: "output", - MODEL_TENSOR.ROPE_FREQS: "rope_freqs", - MODEL_TENSOR.ATTN_NORM: "blk.{bid}.attn_norm", - MODEL_TENSOR.ATTN_NORM_2: "blk.{bid}.attn_norm_2", - MODEL_TENSOR.ATTN_QKV: "blk.{bid}.attn_qkv", - MODEL_TENSOR.ATTN_Q: "blk.{bid}.attn_q", - MODEL_TENSOR.ATTN_K: "blk.{bid}.attn_k", - MODEL_TENSOR.ATTN_V: "blk.{bid}.attn_v", - MODEL_TENSOR.ATTN_OUT: "blk.{bid}.attn_output", - MODEL_TENSOR.ATTN_ROT_EMBD: "blk.{bid}.attn_rot_embd", - MODEL_TENSOR.ATTN_Q_NORM: "blk.{bid}.attn_q_norm", - MODEL_TENSOR.ATTN_K_NORM: "blk.{bid}.attn_k_norm", - MODEL_TENSOR.FFN_NORM: "blk.{bid}.ffn_norm", - MODEL_TENSOR.FFN_GATE: "blk.{bid}.ffn_gate", - MODEL_TENSOR.FFN_DOWN: "blk.{bid}.ffn_down", - MODEL_TENSOR.FFN_UP: "blk.{bid}.ffn_up", -} - -MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { - MODEL_ARCH.LLAMA: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ROPE_FREQS, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_Q, - MODEL_TENSOR.ATTN_K, - MODEL_TENSOR.ATTN_V, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.ATTN_ROT_EMBD, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_GATE, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.GPTNEOX: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.FALCON: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_NORM_2, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.BAICHUAN: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ROPE_FREQS, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_Q, - MODEL_TENSOR.ATTN_K, - MODEL_TENSOR.ATTN_V, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.ATTN_ROT_EMBD, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_GATE, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.STARCODER: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.POS_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.BERT: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.TOKEN_TYPES, - MODEL_TENSOR.POS_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_Q, - MODEL_TENSOR.ATTN_K, - MODEL_TENSOR.ATTN_V, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.MPT: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.GPTJ: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_Q, - MODEL_TENSOR.ATTN_K, - MODEL_TENSOR.ATTN_V, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.PERSIMMON: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - MODEL_TENSOR.ATTN_Q_NORM, - MODEL_TENSOR.ATTN_K_NORM, - MODEL_TENSOR.ATTN_ROT_EMBD, - ], - MODEL_ARCH.REFACT: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_Q, - MODEL_TENSOR.ATTN_K, - MODEL_TENSOR.ATTN_V, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_GATE, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.BLOOM: [ - MODEL_TENSOR.TOKEN_EMBD, - MODEL_TENSOR.TOKEN_EMBD_NORM, - MODEL_TENSOR.OUTPUT_NORM, - MODEL_TENSOR.OUTPUT, - MODEL_TENSOR.ATTN_NORM, - MODEL_TENSOR.ATTN_QKV, - MODEL_TENSOR.ATTN_OUT, - MODEL_TENSOR.FFN_NORM, - MODEL_TENSOR.FFN_DOWN, - MODEL_TENSOR.FFN_UP, - ], - MODEL_ARCH.GPT2: [ - # TODO - ], - # TODO -} - -# tensors that will not be serialized -MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { - MODEL_ARCH.LLAMA: [ - MODEL_TENSOR.ROPE_FREQS, - MODEL_TENSOR.ATTN_ROT_EMBD, - ], - MODEL_ARCH.BAICHUAN: [ - MODEL_TENSOR.ROPE_FREQS, - MODEL_TENSOR.ATTN_ROT_EMBD, - ], - MODEL_ARCH.PERSIMMON: [ - MODEL_TENSOR.ROPE_FREQS, - ] -} - - -class TensorNameMap: - mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { - # Token embeddings - MODEL_TENSOR.TOKEN_EMBD: ( - "gpt_neox.embed_in", # gptneox - "transformer.wte", # gpt2 gpt-j mpt refact - "transformer.word_embeddings", # falcon - "word_embeddings", # bloom - "model.embed_tokens", # llama-hf - "tok_embeddings", # llama-pth - "embeddings.word_embeddings", # bert - "language_model.embedding.word_embeddings", # persimmon - ), - - # Token type embeddings - MODEL_TENSOR.TOKEN_TYPES: ( - "embeddings.token_type_embeddings", # bert - ), - - # Normalization of token embeddings - MODEL_TENSOR.TOKEN_EMBD_NORM: ( - "word_embeddings_layernorm", # bloom - ), - - # Position embeddings - MODEL_TENSOR.POS_EMBD: ( - "transformer.wpe", # gpt2 - "embeddings.position_embeddings", # bert - ), - - # Output - MODEL_TENSOR.OUTPUT: ( - "embed_out", # gptneox - "lm_head", # gpt2 mpt falcon llama-hf baichuan - "output", # llama-pth bloom - "word_embeddings_for_head", # persimmon - ), - - # Output norm - MODEL_TENSOR.OUTPUT_NORM: ( - "gpt_neox.final_layer_norm", # gptneox - "transformer.ln_f", # gpt2 gpt-j falcon - "model.norm", # llama-hf baichuan - "norm", # llama-pth - "embeddings.LayerNorm", # bert - "transformer.norm_f", # mpt - "ln_f", # refact bloom - "language_model.encoder.final_layernorm", # persimmon - ), - - # Rope frequencies - MODEL_TENSOR.ROPE_FREQS: ( - "rope.freqs", # llama-pth - ), - } - - block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { - # Attention norm - MODEL_TENSOR.ATTN_NORM: ( - "gpt_neox.layers.{bid}.input_layernorm", # gptneox - "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact - "transformer.blocks.{bid}.norm_1", # mpt - "transformer.h.{bid}.input_layernorm", # falcon7b - "h.{bid}.input_layernorm", # bloom - "transformer.h.{bid}.ln_mlp", # falcon40b - "model.layers.{bid}.input_layernorm", # llama-hf - "layers.{bid}.attention_norm", # llama-pth - "encoder.layer.{bid}.attention.output.LayerNorm", # bert - "language_model.encoder.layers.{bid}.input_layernorm", # persimmon - ), - - # Attention norm 2 - MODEL_TENSOR.ATTN_NORM_2: ( - "transformer.h.{bid}.ln_attn", # falcon40b - ), - - # Attention query-key-value - MODEL_TENSOR.ATTN_QKV: ( - "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox - "transformer.h.{bid}.attn.c_attn", # gpt2 - "transformer.blocks.{bid}.attn.Wqkv", # mpt - "transformer.h.{bid}.self_attention.query_key_value", # falcon - "h.{bid}.self_attention.query_key_value", # bloom - "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon - ), - - # Attention query - MODEL_TENSOR.ATTN_Q: ( - "model.layers.{bid}.self_attn.q_proj", # llama-hf - "layers.{bid}.attention.wq", # llama-pth - "encoder.layer.{bid}.attention.self.query", # bert - "transformer.h.{bid}.attn.q_proj", # gpt-j - ), - - # Attention key - MODEL_TENSOR.ATTN_K: ( - "model.layers.{bid}.self_attn.k_proj", # llama-hf - "layers.{bid}.attention.wk", # llama-pth - "encoder.layer.{bid}.attention.self.key", # bert - "transformer.h.{bid}.attn.k_proj", # gpt-j - ), - - # Attention value - MODEL_TENSOR.ATTN_V: ( - "model.layers.{bid}.self_attn.v_proj", # llama-hf - "layers.{bid}.attention.wv", # llama-pth - "encoder.layer.{bid}.attention.self.value", # bert - "transformer.h.{bid}.attn.v_proj", # gpt-j - ), - - # Attention output - MODEL_TENSOR.ATTN_OUT: ( - "gpt_neox.layers.{bid}.attention.dense", # gptneox - "transformer.h.{bid}.attn.c_proj", # gpt2 refact - "transformer.blocks.{bid}.attn.out_proj", # mpt - "transformer.h.{bid}.self_attention.dense", # falcon - "h.{bid}.self_attention.dense", # bloom - "model.layers.{bid}.self_attn.o_proj", # llama-hf - "layers.{bid}.attention.wo", # llama-pth - "encoder.layer.{bid}.attention.output.dense", # bert - "transformer.h.{bid}.attn.out_proj", # gpt-j - "language_model.encoder.layers.{bid}.self_attention.dense" # persimmon - ), - - # Rotary embeddings - MODEL_TENSOR.ATTN_ROT_EMBD: ( - "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf - "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth - ), - - # Feed-forward norm - MODEL_TENSOR.FFN_NORM: ( - "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox - "transformer.h.{bid}.ln_2", # gpt2 refact - "h.{bid}.post_attention_layernorm", # bloom - "transformer.blocks.{bid}.norm_2", # mpt - "model.layers.{bid}.post_attention_layernorm", # llama-hf - "layers.{bid}.ffn_norm", # llama-pth - "encoder.layer.{bid}.output.LayerNorm", # bert - "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon - ), - - # Feed-forward up - MODEL_TENSOR.FFN_UP: ( - "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox - "transformer.h.{bid}.mlp.c_fc", # gpt2 - "transformer.blocks.{bid}.ffn.up_proj", # mpt - "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon - "h.{bid}.mlp.dense_h_to_4h", # bloom - "model.layers.{bid}.mlp.up_proj", # llama-hf refact - "layers.{bid}.feed_forward.w3", # llama-pth - "encoder.layer.{bid}.intermediate.dense", # bert - "transformer.h.{bid}.mlp.fc_in", # gpt-j - "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon - ), - - # Feed-forward gate - MODEL_TENSOR.FFN_GATE: ( - "model.layers.{bid}.mlp.gate_proj", # llama-hf refact - "layers.{bid}.feed_forward.w1", # llama-pth - ), - - # Feed-forward down - MODEL_TENSOR.FFN_DOWN: ( - "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox - "transformer.h.{bid}.mlp.c_proj", # gpt2 refact - "transformer.blocks.{bid}.ffn.down_proj", # mpt - "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon - "h.{bid}.mlp.dense_4h_to_h", # bloom - "model.layers.{bid}.mlp.down_proj", # llama-hf - "layers.{bid}.feed_forward.w2", # llama-pth - "encoder.layer.{bid}.output.dense", # bert - "transformer.h.{bid}.mlp.fc_out", # gpt-j - "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon - ), - - MODEL_TENSOR.ATTN_Q_NORM: ( - "language_model.encoder.layers.{bid}.self_attention.q_layernorm", - ), - - MODEL_TENSOR.ATTN_K_NORM: ( - "language_model.encoder.layers.{bid}.self_attention.k_layernorm", - ), - - MODEL_TENSOR.ROPE_FREQS: ( - "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon - ) - } - - mapping: dict[str, tuple[MODEL_TENSOR, str]] - - def __init__(self, arch: MODEL_ARCH, n_blocks: int): - self.mapping = {} - for tensor, keys in self.mappings_cfg.items(): - if tensor not in MODEL_TENSORS[arch]: - continue - tensor_name = TENSOR_NAMES[tensor] - self.mapping[tensor_name] = (tensor, tensor_name) - for key in keys: - self.mapping[key] = (tensor, tensor_name) - for bid in range(n_blocks): - for tensor, keys in self.block_mappings_cfg.items(): - if tensor not in MODEL_TENSORS[arch]: - continue - tensor_name = TENSOR_NAMES[tensor].format(bid = bid) - self.mapping[tensor_name] = (tensor, tensor_name) - for key in keys: - key = key.format(bid = bid) - self.mapping[key] = (tensor, tensor_name) - - def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None: - result = self.mapping.get(key) - if result is not None: - return result - for suffix in try_suffixes: - if key.endswith(suffix): - result = self.mapping.get(key[:-len(suffix)]) - if result is not None: - return (result[0], result[1] + suffix) - return None - - def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None: - result = self.get_type_and_name(key, try_suffixes = try_suffixes) - if result is None: - return None - return result[1] - - def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None: - result = self.get_type_and_name(key, try_suffixes = try_suffixes) - if result is None: - return None - return result[0] - - def __getitem__(self, key: str) -> str: - try: - return self.mapping[key][1] - except KeyError: - raise KeyError(key) - - def __contains__(self, key: str) -> bool: - return key in self.mapping - - def __repr__(self) -> str: - return repr(self.mapping) - -def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap: - return TensorNameMap(arch, n_blocks) - -class TokenType(IntEnum): - NORMAL = 1 - UNKNOWN = 2 - CONTROL = 3 - USER_DEFINED = 4 - UNUSED = 5 - BYTE = 6 - -# -# implementation -# - - -class GGMLQuantizationType(IntEnum): - F32 = 0 - F16 = 1 - Q4_0 = 2 - Q4_1 = 3 - Q5_0 = 6 - Q5_1 = 7 - Q8_0 = 8 - Q8_1 = 9 - Q2_K = 10 - Q3_K = 11 - Q4_K = 12 - Q5_K = 13 - Q6_K = 14 - Q8_K = 15 - -class GGUFEndian(IntEnum): - LITTLE = 0 - BIG = 1 - - -class GGUFValueType(IntEnum): - UINT8 = 0 - INT8 = 1 - UINT16 = 2 - INT16 = 3 - UINT32 = 4 - INT32 = 5 - FLOAT32 = 6 - BOOL = 7 - STRING = 8 - ARRAY = 9 - UINT64 = 10 - INT64 = 11 - FLOAT64 = 12 - - @staticmethod - def get_type(val): - if isinstance(val, str) or isinstance(val, bytes) or isinstance(val, bytearray): - return GGUFValueType.STRING - elif isinstance(val, list): - return GGUFValueType.ARRAY - elif isinstance(val, float): - return GGUFValueType.FLOAT32 - elif isinstance(val, bool): - return GGUFValueType.BOOL - elif isinstance(val, int): - return GGUFValueType.INT32 - # TODO: need help with 64-bit types in Python - else: - print("Unknown type: "+str(type(val))) - sys.exit() - - -class GGUFWriter: - fout: BufferedWriter - arch: str - offset_tensor = 0 - data_alignment = GGUF_DEFAULT_ALIGNMENT - kv_data = b"" - kv_data_count = 0 - ti_data = b"" - ti_data_count = 0 - use_temp_file: bool - temp_file: tempfile.SpooledTemporaryFile[bytes] | None = None - tensors: list[tuple[np.ndarray[Any, Any], int]] - - @property - def pack_prefix(self): - if self.endianess==GGUFEndian.LITTLE: - return "<" - else: - return ">" - - def __init__(self, path: os.PathLike[str] | str, arch: str, use_temp_file = True, endianess=GGUFEndian.LITTLE): - self.fout = open(path, "wb") - self.arch = arch - self.endianess = endianess - self._simple_value_packing = { - GGUFValueType.UINT8: f"{self.pack_prefix}B", - GGUFValueType.INT8: f"{self.pack_prefix}b", - GGUFValueType.UINT16: f"{self.pack_prefix}H", - GGUFValueType.INT16: f"{self.pack_prefix}h", - GGUFValueType.UINT32: f"{self.pack_prefix}I", - GGUFValueType.INT32: f"{self.pack_prefix}i", - GGUFValueType.FLOAT32: f"{self.pack_prefix}f", - GGUFValueType.UINT64: f"{self.pack_prefix}Q", - GGUFValueType.INT64: f"{self.pack_prefix}q", - GGUFValueType.FLOAT64: f"{self.pack_prefix}d", - GGUFValueType.BOOL: "?" , - } - self.add_architecture() - self.use_temp_file = use_temp_file - self.tensors = [] - endianess_str = "Big Endian" if self.endianess == GGUFEndian.BIG else "Little Endian" - print(f"This gguf file is for {endianess_str} only") - - def write_header_to_file(self): - self.fout.write(struct.pack(" 0: - ltype = GGUFValueType.get_type(val[0]) - if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]): - raise ValueError("All items in a GGUF array should be of the same type") - self.kv_data += struct.pack(f"{self.pack_prefix}I", ltype) - self.kv_data += struct.pack(f"{self.pack_prefix}Q", len(val)) - for item in val: - self.add_val(item, add_vtype=False) - else: - raise ValueError("Invalid GGUF metadata value type or value") - - @staticmethod - def ggml_pad(x: int, n: int) -> int: - return ((x + n - 1) // n) * n - - def add_tensor_info(self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None): - assert raw_dtype is not None or tensor_dtype in (np.float32, np.float16), "Only F32 and F16 tensors are supported for now" - - encoded_name = name.encode("utf8") - self.ti_data += struct.pack(f"{self.pack_prefix}Q", len(encoded_name)) - self.ti_data += encoded_name - n_dims = len(tensor_shape) - self.ti_data += struct.pack(f"{self.pack_prefix}I", n_dims) - for i in range(n_dims): - self.ti_data += struct.pack(f"{self.pack_prefix}Q", tensor_shape[n_dims - 1 - i]) - if raw_dtype is None: - dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16 - else: - dtype = raw_dtype - self.ti_data += struct.pack(f"{self.pack_prefix}I", dtype) - self.ti_data += struct.pack(f"{self.pack_prefix}Q", self.offset_tensor) - self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment) - self.ti_data_count += 1 - - def add_tensor(self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, raw_dtype: GGMLQuantizationType | None = None): - if self.endianess == GGUFEndian.BIG: - tensor.byteswap(inplace=True) - if self.use_temp_file and self.temp_file is None: - fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256*1024*1024) - fp.seek(0) - self.temp_file = fp - - shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape - self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype) - - pad = GGUFWriter.ggml_pad(tensor.nbytes, self.data_alignment) - tensor.nbytes - - if self.temp_file is None: - self.tensors.append((tensor, pad)) - return - - tensor.tofile(self.temp_file) - - if pad != 0: - self.temp_file.write(bytes([0] * pad)) - - def write_padding(self, fp: BinaryIO, n: int, align: int | None = None): - pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n - if pad != 0: - fp.write(bytes([0] * pad)) - - def write_tensor_data(self, tensor: np.ndarray[Any, Any]): - if self.endianess==GGUFEndian.BIG: - tensor.byteswap(inplace=True) - self.write_padding(self.fout, self.fout.tell()) - tensor.tofile(self.fout) - self.write_padding(self.fout, tensor.nbytes) - - def write_tensors_to_file(self): - self.write_ti_data_to_file() - - self.write_padding(self.fout, self.fout.tell()) - - if self.temp_file is None: - for (currtensor, currpad) in self.tensors: - currtensor.tofile(self.fout) - if currpad != 0: - self.fout.write(bytes([0] * currpad)) - return - - self.temp_file.seek(0) - - shutil.copyfileobj(self.temp_file, self.fout) - self.flush() - self.temp_file.close() - - def flush(self): - self.fout.flush() - - def close(self): - self.fout.close() - - def add_architecture(self): - self.add_string(KEY_GENERAL_ARCHITECTURE, self.arch) - - def add_author(self, author: str): - self.add_string(KEY_GENERAL_AUTHOR, author) - - def add_tensor_data_layout(self, layout: str): - self.add_string(KEY_TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) - - def add_url(self, url: str): - self.add_string(KEY_GENERAL_URL, url) - - def add_description(self, description: str): - self.add_string(KEY_GENERAL_DESCRIPTION, description) - - def add_source_url(self, url: str): - self.add_string(KEY_GENERAL_SOURCE_URL, url) - - def add_source_hf_repo(self, repo: str): - self.add_string(KEY_GENERAL_SOURCE_HF_REPO, repo) - - def add_file_type(self, ftype: int): - self.add_uint32(KEY_GENERAL_FILE_TYPE, ftype) - - def add_name(self, name: str): - self.add_string(KEY_GENERAL_NAME, name) - - def add_quantization_version(self, quantization_version: GGMLQuantizationType): - self.add_uint32( - KEY_GENERAL_QUANTIZATION_VERSION, quantization_version) - - def add_custom_alignment(self, alignment: int): - self.data_alignment = alignment - self.add_uint32(KEY_GENERAL_ALIGNMENT, alignment) - - def add_context_length(self, length: int): - self.add_uint32( - KEY_CONTEXT_LENGTH.format(arch=self.arch), length) - - def add_embedding_length(self, length: int): - self.add_uint32( - KEY_EMBEDDING_LENGTH.format(arch=self.arch), length) - - def add_block_count(self, length: int): - self.add_uint32( - KEY_BLOCK_COUNT.format(arch=self.arch), length) - - def add_feed_forward_length(self, length: int): - self.add_uint32( - KEY_FEED_FORWARD_LENGTH.format(arch=self.arch), length) - - def add_parallel_residual(self, use: bool): - self.add_bool( - KEY_USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) - - def add_head_count(self, count: int): - self.add_uint32( - KEY_ATTENTION_HEAD_COUNT.format(arch=self.arch), count) - - def add_head_count_kv(self, count: int): - self.add_uint32( - KEY_ATTENTION_HEAD_COUNT_KV.format(arch=self.arch), count) - - def add_max_alibi_bias(self, bias: float): - self.add_float32( - KEY_ATTENTION_MAX_ALIBI_BIAS.format(arch=self.arch), bias) - - def add_clamp_kqv(self, value: float): - self.add_float32( - KEY_ATTENTION_CLAMP_KQV.format(arch=self.arch), value) - - def add_layer_norm_eps(self, value: float): - self.add_float32( - KEY_ATTENTION_LAYERNORM_EPS.format(arch=self.arch), value) - - def add_layer_norm_rms_eps(self, value: float): - self.add_float32( - KEY_ATTENTION_LAYERNORM_RMS_EPS.format(arch=self.arch), value) - - def add_rope_dimension_count(self, count: int): - self.add_uint32( - KEY_ROPE_DIMENSION_COUNT.format(arch=self.arch), count) - - def add_rope_freq_base(self, value: float): - self.add_float32(KEY_ROPE_FREQ_BASE.format(arch=self.arch), value) - - def add_rope_scale_linear(self, value: float): - self.add_float32(KEY_ROPE_SCALE_LINEAR.format(arch=self.arch), value) - - def add_tokenizer_model(self, model: str): - self.add_string(KEY_TOKENIZER_MODEL, model) - - def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]): - self.add_array(KEY_TOKENIZER_LIST, tokens) - - def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]): - self.add_array(KEY_TOKENIZER_MERGES, merges) - - def add_token_types(self, types: Sequence[TokenType] | Sequence[int]): - self.add_array(KEY_TOKENIZER_TOKEN_TYPE, types) - - def add_token_scores(self, scores: Sequence[float]): - self.add_array(KEY_TOKENIZER_SCORES, scores) - - def add_bos_token_id(self, id: int): - self.add_uint32(KEY_TOKENIZER_BOS_ID, id) - - def add_eos_token_id(self, id: int): - self.add_uint32(KEY_TOKENIZER_EOS_ID, id) - - def add_unk_token_id(self, id: int): - self.add_uint32(KEY_TOKENIZER_UNK_ID, id) - - def add_sep_token_id(self, id: int): - self.add_uint32(KEY_TOKENIZER_SEP_ID, id) - - def add_pad_token_id(self, id: int): - self.add_uint32(KEY_TOKENIZER_PAD_ID, id) - - -class SpecialVocab: - load_merges: bool = False - merges: list[str] = [] - special_token_types: tuple[str, ...] = ('bos', 'eos', 'unk', 'sep', 'pad') - special_token_ids: dict[str, int] = {} - n_vocab: int | None = None - - def __init__( - self, path: str | os.PathLike[str], load_merges: bool = False, - special_token_types: tuple[str, ...] | None = None, - n_vocab: int | None = None, - ): - self.special_token_ids = {} - self.n_vocab = n_vocab - self.load_merges = load_merges - if special_token_types is not None: - self.special_token_types = special_token_types - self._load(Path(path)) - - def _load(self, path: Path) -> None: - if not self._try_load_from_tokenizer_json(path): - self._try_load_from_config_json(path) - - def _set_special_token(self, typ: str, tid: Any): - if not isinstance(tid, int) or tid < 0: - return - if self.n_vocab is None or tid < self.n_vocab: - self.special_token_ids[typ] = tid - return - print(f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping', - file = sys.stderr) - - - def _try_load_from_tokenizer_json(self, path: Path) -> bool: - tokenizer_file = path / 'tokenizer.json' - if not tokenizer_file.is_file(): - return False - with open(tokenizer_file, encoding = 'utf-8') as f: - tokenizer = json.load(f) - if self.load_merges: - merges = tokenizer.get('model', {}).get('merges') - if isinstance(merges, list) and len(merges) > 0 and isinstance(merges[0], str): - self.merges = merges - tokenizer_config_file = path / 'tokenizer_config.json' - added_tokens = tokenizer.get('added_tokens') - if added_tokens is None or not tokenizer_config_file.is_file(): - return True - with open(tokenizer_config_file, encoding = 'utf-8') as f: - tokenizer_config = json.load(f) - for typ in self.special_token_types: - entry = tokenizer_config.get(f'{typ}_token') - if isinstance(entry, str): - tc_content = entry - elif isinstance(entry, dict): - entry_content = entry.get('content') - if not isinstance(entry_content, str): - continue - tc_content = entry_content - else: - continue - # We only need the first match here. - maybe_token_id = next(( - atok.get('id') for atok in added_tokens - if atok.get('content') == tc_content), None) - self._set_special_token(typ, maybe_token_id) - return True - - def _try_load_from_config_json(self, path: Path) -> bool: - config_file = path / 'config.json' - if not config_file.is_file(): - return False - with open(config_file, encoding = 'utf-8') as f: - config = json.load(f) - for typ in self.special_token_types: - self._set_special_token(typ, config.get(f'{typ}_token_id')) - return True - - def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None: - if len(self.merges) > 0: - if not quiet: - print(f'gguf: Adding {len(self.merges)} merge(s).') - gw.add_token_merges(self.merges) - for typ, tokid in self.special_token_ids.items(): - handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) - if handler is None: - print(f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', file = sys.stderr) - continue - if not quiet: - print(f'gguf: Setting special token type {typ} to {tokid}') - handler(tokid) - - def __repr__(self) -> str: - return f'' - - -# Example usage: -if __name__ == "__main__": - # Example usage with a file - gguf_writer = GGUFWriter("example.gguf", "llama") - - gguf_writer.add_architecture() - gguf_writer.add_block_count(12) - gguf_writer.add_uint32("answer", 42) # Write a 32-bit integer - gguf_writer.add_float32("answer_in_float", 42.0) # Write a 32-bit float - gguf_writer.add_custom_alignment(64) - - tensor1 = np.ones((32,), dtype=np.float32) * 100.0 - tensor2 = np.ones((64,), dtype=np.float32) * 101.0 - tensor3 = np.ones((96,), dtype=np.float32) * 102.0 - - gguf_writer.add_tensor("tensor1", tensor1) - gguf_writer.add_tensor("tensor2", tensor2) - gguf_writer.add_tensor("tensor3", tensor3) - - gguf_writer.write_header_to_file() - gguf_writer.write_kv_data_to_file() - gguf_writer.write_tensors_to_file() - - gguf_writer.close() +importlib.reload(gguf) diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py new file mode 100644 index 000000000..8682765ed --- /dev/null +++ b/gguf-py/gguf/gguf_reader.py @@ -0,0 +1,264 @@ +# +# GGUF file reading/modification support. For API usage information, +# please see the files scripts/ for some fairly simple examples. +# +from __future__ import annotations + +import os +from collections import OrderedDict +from typing import Any, Literal, NamedTuple, TypeVar, Union + +import numpy as np +import numpy.typing as npt + +if __name__ == "__main__": + import sys + from pathlib import Path + + # Allow running file in package as a script. + sys.path.insert(0, str(Path(__file__).parent.parent)) + +from gguf.constants import ( + GGML_QUANT_SIZES, + GGUF_DEFAULT_ALIGNMENT, + GGUF_MAGIC, + GGUF_VERSION, + GGMLQuantizationType, + GGUFValueType, +) + + +READER_SUPPORTED_VERSIONS = [2, GGUF_VERSION] + + +class ReaderField(NamedTuple): + # Offset to start of this field. + offset: int + + # Name of the field (not necessarily from file data). + name: str + + # Data parts. Some types have multiple components, such as strings + # that consist of a length followed by the string data. + parts: list[npt.NDArray[Any]] = [] + + # Indexes into parts that we can call the actual data. For example + # an array of strings will be populated with indexes to the actual + # string data. + data: list[int] = [-1] + + types: list[GGUFValueType] = [] + + +class ReaderTensor(NamedTuple): + name: str + tensor_type: GGMLQuantizationType + shape: npt.NDArray[np.uint32] + n_elements: int + n_bytes: int + data_offset: int + data: npt.NDArray[Any] + field: ReaderField + + +class GGUFReader: + # I - same as host, S - swapped + byte_order: Literal['I' | 'S'] = 'I' + alignment: int = GGUF_DEFAULT_ALIGNMENT + + # Note: Internal helper, API may change. + gguf_scalar_to_np: dict[GGUFValueType, type[np.generic]] = { + GGUFValueType.UINT8: np.uint8, + GGUFValueType.INT8: np.int8, + GGUFValueType.UINT16: np.uint16, + GGUFValueType.INT16: np.int16, + GGUFValueType.UINT32: np.uint32, + GGUFValueType.INT32: np.int32, + GGUFValueType.FLOAT32: np.float32, + GGUFValueType.UINT64: np.uint64, + GGUFValueType.INT64: np.int64, + GGUFValueType.FLOAT64: np.float64, + GGUFValueType.BOOL: np.bool_, + } + + def __init__(self, path: os.PathLike[str] | str, mode: Literal['r' | 'r+' | 'c'] = 'r'): + self.data = np.memmap(path, mode = mode) + offs = 0 + if self._get(offs, np.uint32, override_order = '<')[0] != GGUF_MAGIC: + raise ValueError('GGUF magic invalid') + offs += 4 + temp_version = self._get(offs, np.uint32) + if temp_version[0] & 65535 == 0: + # If we get 0 here that means it's (probably) a GGUF file created for + # the opposite byte order of the machine this script is running on. + self.byte_order = 'S' + temp_version = temp_version.newbyteorder(self.byte_order) + version = temp_version[0] + if version not in READER_SUPPORTED_VERSIONS: + raise ValueError(f'Sorry, file appears to be version {version} which we cannot handle') + self.fields: OrderedDict[str, ReaderField] = OrderedDict() + self.tensors: list[ReaderTensor] = [] + offs += self._push_field(ReaderField(offs, 'GGUF.version', [temp_version], [0], [GGUFValueType.UINT32])) + temp_counts = self._get(offs, np.uint64, 2) + offs += self._push_field(ReaderField(offs, 'GGUF.tensor_count', [temp_counts[:1]], [0], [GGUFValueType.UINT64])) + offs += self._push_field(ReaderField(offs, 'GGUF.kv_count', [temp_counts[1:]], [0], [GGUFValueType.UINT64])) + tensor_count, kv_count = temp_counts + offs = self._build_fields(offs, kv_count) + offs, tensors_fields = self._build_tensors_fields(offs, tensor_count) + new_align = self.fields.get('general.alignment') + if new_align is not None: + if new_align.types != [GGUFValueType.UINT64]: + raise ValueError('Bad type for general.alignment field') + self.alignment = new_align.parts[-1][0] + padding = offs % self.alignment + if padding != 0: + offs += self.alignment - padding + self._build_tensors(offs, tensors_fields) + + _DT = TypeVar('_DT', bound = npt.DTypeLike) + + # Fetch a key/value metadata field by key. + def get_field(self, key: str) -> Union[ReaderField, None]: + return self.fields.get(key, None) + + # Fetch a tensor from the list by index. + def get_tensor(self, idx: int) -> ReaderTensor: + return self.tensors[idx] + + def _get( + self, offset: int, dtype: npt.DTypeLike, count: int = 1, override_order: None | Literal['I' | 'S' | '<'] = None, + ) -> npt.NDArray[Any]: + count = int(count) + itemsize = int(np.empty([], dtype = dtype).itemsize) + end_offs = offset + itemsize * count + return ( + self.data[offset:end_offs] + .view(dtype = dtype)[:count] + .newbyteorder(override_order or self.byte_order) + ) + + def _push_field(self, field: ReaderField, skip_sum: bool = False) -> int: + if field.name in self.fields: + raise KeyError(f'Duplicate {field.name} already in list at offset {field.offset}') + self.fields[field.name] = field + return 0 if skip_sum else sum(int(part.nbytes) for part in field.parts) + + def _get_str(self, offset: int) -> tuple[npt.NDArray[np.uint64], npt.NDArray[np.uint8]]: + slen = self._get(offset, np.uint64) + return slen, self._get(offset + 8, np.uint8, slen[0]) + + def _get_field_parts( + self, orig_offs: int, raw_type: int, + ) -> tuple[int, list[npt.NDArray[Any]], list[int], list[GGUFValueType]]: + offs = orig_offs + types: list[GGUFValueType] = [] + gtype = GGUFValueType(raw_type) + types.append(gtype) + # Handle strings. + if gtype == GGUFValueType.STRING: + sparts: list[npt.NDArray[Any]] = list(self._get_str(offs)) + size = sum(int(part.nbytes) for part in sparts) + return size, sparts, [1], types + # Check if it's a simple scalar type. + nptype = self.gguf_scalar_to_np.get(gtype) + if nptype is not None: + val = self._get(offs, nptype) + return int(val.nbytes), [val], [0], types + # Handle arrays. + if gtype == GGUFValueType.ARRAY: + raw_itype = self._get(offs, np.uint32) + offs += int(raw_itype.nbytes) + alen = self._get(offs, np.uint64) + offs += int(alen.nbytes) + aparts: list[npt.NDArray[Any]] = [raw_itype, alen] + data_idxs: list[int] = [] + for idx in range(alen[0]): + curr_size, curr_parts, curr_idxs, curr_types = self._get_field_parts(offs, raw_itype[0]) + if idx == 0: + types += curr_types + idxs_offs = len(aparts) + aparts += curr_parts + data_idxs += (idx + idxs_offs for idx in curr_idxs) + offs += curr_size + return offs - orig_offs, aparts, data_idxs, types + # We can't deal with this one. + raise ValueError('Unknown/unhandled field type {gtype}') + + def _get_tensor(self, orig_offs: int) -> ReaderField: + offs = orig_offs + name_len, name_data = self._get_str(offs) + offs += int(name_len.nbytes + name_data.nbytes) + n_dims = self._get(offs, np.uint32) + offs += int(n_dims.nbytes) + dims = self._get(offs, np.uint64, n_dims[0]) + offs += int(dims.nbytes) + raw_dtype = self._get(offs, np.uint32) + offs += int(raw_dtype.nbytes) + offset_tensor = self._get(offs, np.uint64) + offs += int(offset_tensor.nbytes) + return ReaderField( + orig_offs, + str(bytes(name_data), encoding = 'utf-8'), + [name_len, name_data, n_dims, dims, raw_dtype, offset_tensor], + [1, 3, 4, 5], + ) + + def _build_fields(self, offs: int, count: int) -> int: + for _ in range(count): + orig_offs = offs + kv_klen, kv_kdata = self._get_str(offs) + offs += int(kv_klen.nbytes + kv_kdata.nbytes) + raw_kv_type = self._get(offs, np.uint32) + offs += int(raw_kv_type.nbytes) + parts: list[npt.NDArray[Any]] = [kv_klen, kv_kdata, raw_kv_type] + idxs_offs = len(parts) + field_size, field_parts, field_idxs, field_types = self._get_field_parts(offs, raw_kv_type[0]) + parts += field_parts + self._push_field(ReaderField( + orig_offs, + str(bytes(kv_kdata), encoding = 'utf-8'), + parts, + [idx + idxs_offs for idx in field_idxs], + field_types, + ), skip_sum = True) + offs += field_size + return offs + + def _build_tensors_fields(self, offs: int, count: int) -> tuple[int, list[ReaderField]]: + tensor_fields = [] + for _ in range(count): + field = self._get_tensor(offs) + offs += sum(int(part.nbytes) for part in field.parts) + tensor_fields.append(field) + return offs, tensor_fields + + def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None: + tensors = [] + for field in fields: + _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts + ggml_type = GGMLQuantizationType(raw_dtype[0]) + n_elems = np.prod(dims) + block_size, type_size = GGML_QUANT_SIZES[ggml_type] + n_bytes = n_elems * type_size // block_size + data_offs = int(start_offs + offset_tensor[0]) + item_type: npt.DTypeLike + if ggml_type == GGMLQuantizationType.F32: + item_count = n_elems + item_type = np.float32 + elif ggml_type == GGMLQuantizationType.F16: + item_count = n_elems + item_type = np.float16 + else: + item_count = n_bytes + item_type = np.uint8 + tensors.append(ReaderTensor( + name = str(bytes(name_data), encoding = 'utf-8'), + tensor_type = ggml_type, + shape = dims, + n_elements = n_elems, + n_bytes = n_bytes, + data_offset = data_offs, + data = self._get(data_offs, item_type, item_count), + field = field, + )) + self.tensors = tensors diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py new file mode 100644 index 000000000..b8ec977c8 --- /dev/null +++ b/gguf-py/gguf/gguf_writer.py @@ -0,0 +1,412 @@ +from __future__ import annotations + +import os +import shutil +import struct +import tempfile +from enum import Enum, auto +from io import BufferedWriter +from typing import IO, Any, Sequence + +import numpy as np + +from .constants import ( + GGUF_DEFAULT_ALIGNMENT, + GGUF_MAGIC, + GGUF_VERSION, + GGMLQuantizationType, + GGUFEndian, + GGUFValueType, + Keys, + RopeScalingType, + TokenType, +) + + +class WriterState(Enum): + EMPTY = auto() + HEADER = auto() + KV_DATA = auto() + TI_DATA = auto() + + +class GGUFWriter: + fout: BufferedWriter + temp_file: tempfile.SpooledTemporaryFile[bytes] | None + tensors: list[np.ndarray[Any, Any]] + _simple_value_packing = { + GGUFValueType.UINT8: "B", + GGUFValueType.INT8: "b", + GGUFValueType.UINT16: "H", + GGUFValueType.INT16: "h", + GGUFValueType.UINT32: "I", + GGUFValueType.INT32: "i", + GGUFValueType.FLOAT32: "f", + GGUFValueType.UINT64: "Q", + GGUFValueType.INT64: "q", + GGUFValueType.FLOAT64: "d", + GGUFValueType.BOOL: "?", + } + + def __init__( + self, path: os.PathLike[str] | str, arch: str, use_temp_file: bool = True, + endianess: GGUFEndian = GGUFEndian.LITTLE, + ): + self.fout = open(path, "wb") + self.arch = arch + self.endianess = endianess + self.offset_tensor = 0 + self.data_alignment = GGUF_DEFAULT_ALIGNMENT + self.kv_data = bytearray() + self.kv_data_count = 0 + self.ti_data = bytearray() + self.ti_data_count = 0 + self.use_temp_file = use_temp_file + self.temp_file = None + self.tensors = [] + print("gguf: This GGUF file is for {0} Endian only".format( + "Big" if self.endianess == GGUFEndian.BIG else "Little", + )) + self.state = WriterState.EMPTY + + self.add_architecture() + + def write_header_to_file(self) -> None: + if self.state is not WriterState.EMPTY: + raise ValueError(f'Expected output file to be empty, got {self.state}') + + self._write_packed(" None: + if self.state is not WriterState.HEADER: + raise ValueError(f'Expected output file to contain the header, got {self.state}') + + self.fout.write(self.kv_data) + self.flush() + self.state = WriterState.KV_DATA + + def write_ti_data_to_file(self) -> None: + if self.state is not WriterState.KV_DATA: + raise ValueError(f'Expected output file to contain KV data, got {self.state}') + + self.fout.write(self.ti_data) + self.flush() + self.state = WriterState.TI_DATA + + def add_key(self, key: str) -> None: + self.add_val(key, GGUFValueType.STRING, add_vtype=False) + + def add_uint8(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.UINT8) + + def add_int8(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.INT8) + + def add_uint16(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.UINT16) + + def add_int16(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.INT16) + + def add_uint32(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.UINT32) + + def add_int32(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.INT32) + + def add_float32(self, key: str, val: float) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.FLOAT32) + + def add_uint64(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.UINT64) + + def add_int64(self, key: str, val: int) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.INT64) + + def add_float64(self, key: str, val: float) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.FLOAT64) + + def add_bool(self, key: str, val: bool) -> None: + self.add_key(key) + self.add_val(val, GGUFValueType.BOOL) + + def add_string(self, key: str, val: str) -> None: + if not val: + return + self.add_key(key) + self.add_val(val, GGUFValueType.STRING) + + def add_array(self, key: str, val: Sequence[Any]) -> None: + if not isinstance(val, Sequence): + raise ValueError("Value must be a sequence for array type") + + self.add_key(key) + self.add_val(val, GGUFValueType.ARRAY) + + def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True) -> None: + if vtype is None: + vtype = GGUFValueType.get_type(val) + + if add_vtype: + self.kv_data += self._pack("I", vtype) + self.kv_data_count += 1 + + pack_fmt = self._simple_value_packing.get(vtype) + if pack_fmt is not None: + self.kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL) + elif vtype == GGUFValueType.STRING: + encoded_val = val.encode("utf8") if isinstance(val, str) else val + self.kv_data += self._pack("Q", len(encoded_val)) + self.kv_data += encoded_val + elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val: + ltype = GGUFValueType.get_type(val[0]) + if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]): + raise ValueError("All items in a GGUF array should be of the same type") + self.kv_data += self._pack("I", ltype) + self.kv_data += self._pack("Q", len(val)) + for item in val: + self.add_val(item, add_vtype=False) + else: + raise ValueError("Invalid GGUF metadata value type or value") + + @staticmethod + def ggml_pad(x: int, n: int) -> int: + return ((x + n - 1) // n) * n + + def add_tensor_info( + self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype[np.float16] | np.dtype[np.float32], + tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None, + ) -> None: + if self.state is not WriterState.EMPTY: + raise ValueError(f'Expected output file to be empty, got {self.state}') + + if raw_dtype is None and tensor_dtype not in (np.float32, np.float16): + raise ValueError("Only F32 and F16 tensors are supported for now") + + encoded_name = name.encode("utf8") + self.ti_data += self._pack("Q", len(encoded_name)) + self.ti_data += encoded_name + n_dims = len(tensor_shape) + self.ti_data += self._pack("I", n_dims) + for i in range(n_dims): + self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i]) + if raw_dtype is None: + dtype = GGMLQuantizationType.F32 if tensor_dtype == np.float32 else GGMLQuantizationType.F16 + else: + dtype = raw_dtype + self.ti_data += self._pack("I", dtype) + self.ti_data += self._pack("Q", self.offset_tensor) + self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment) + self.ti_data_count += 1 + + def add_tensor( + self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, + raw_dtype: GGMLQuantizationType | None = None, + ) -> None: + if self.endianess == GGUFEndian.BIG: + tensor.byteswap(inplace=True) + if self.use_temp_file and self.temp_file is None: + fp = tempfile.SpooledTemporaryFile(mode="w+b", max_size=256 * 1024 * 1024) + fp.seek(0) + self.temp_file = fp + + shape: Sequence[int] = raw_shape if raw_shape is not None else tensor.shape + self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype = raw_dtype) + + if self.temp_file is None: + self.tensors.append(tensor) + return + + tensor.tofile(self.temp_file) + self.write_padding(self.temp_file, tensor.nbytes) + + def write_padding(self, fp: IO[bytes], n: int, align: int | None = None) -> None: + pad = GGUFWriter.ggml_pad(n, align if align is not None else self.data_alignment) - n + if pad != 0: + fp.write(bytes([0] * pad)) + + def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None: + if self.state is not WriterState.TI_DATA: + raise ValueError(f'Expected output file to contain tensor info, got {self.state}') + + if self.endianess == GGUFEndian.BIG: + tensor.byteswap(inplace=True) + self.write_padding(self.fout, self.fout.tell()) + tensor.tofile(self.fout) + self.write_padding(self.fout, tensor.nbytes) + + def write_tensors_to_file(self) -> None: + self.write_ti_data_to_file() + + self.write_padding(self.fout, self.fout.tell()) + + if self.temp_file is None: + while True: + try: + tensor = self.tensors.pop(0) + except IndexError: + break + tensor.tofile(self.fout) + self.write_padding(self.fout, tensor.nbytes) + return + + self.temp_file.seek(0) + + shutil.copyfileobj(self.temp_file, self.fout) + self.flush() + self.temp_file.close() + + def flush(self) -> None: + self.fout.flush() + + def close(self) -> None: + self.fout.close() + + def add_architecture(self) -> None: + self.add_string(Keys.General.ARCHITECTURE, self.arch) + + def add_author(self, author: str) -> None: + self.add_string(Keys.General.AUTHOR, author) + + def add_tensor_data_layout(self, layout: str) -> None: + self.add_string(Keys.LLM.TENSOR_DATA_LAYOUT.format(arch=self.arch), layout) + + def add_url(self, url: str) -> None: + self.add_string(Keys.General.URL, url) + + def add_description(self, description: str) -> None: + self.add_string(Keys.General.DESCRIPTION, description) + + def add_source_url(self, url: str) -> None: + self.add_string(Keys.General.SOURCE_URL, url) + + def add_source_hf_repo(self, repo: str) -> None: + self.add_string(Keys.General.SOURCE_HF_REPO, repo) + + def add_file_type(self, ftype: int) -> None: + self.add_uint32(Keys.General.FILE_TYPE, ftype) + + def add_name(self, name: str) -> None: + self.add_string(Keys.General.NAME, name) + + def add_quantization_version(self, quantization_version: GGMLQuantizationType) -> None: + self.add_uint32( + Keys.General.QUANTIZATION_VERSION, quantization_version) + + def add_custom_alignment(self, alignment: int) -> None: + self.data_alignment = alignment + self.add_uint32(Keys.General.ALIGNMENT, alignment) + + def add_context_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.CONTEXT_LENGTH.format(arch=self.arch), length) + + def add_embedding_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.EMBEDDING_LENGTH.format(arch=self.arch), length) + + def add_block_count(self, length: int) -> None: + self.add_uint32(Keys.LLM.BLOCK_COUNT.format(arch=self.arch), length) + + def add_feed_forward_length(self, length: int) -> None: + self.add_uint32(Keys.LLM.FEED_FORWARD_LENGTH.format(arch=self.arch), length) + + def add_parallel_residual(self, use: bool) -> None: + self.add_bool(Keys.LLM.USE_PARALLEL_RESIDUAL.format(arch=self.arch), use) + + def add_head_count(self, count: int) -> None: + self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count) + + def add_head_count_kv(self, count: int) -> None: + self.add_uint32(Keys.Attention.HEAD_COUNT_KV.format(arch=self.arch), count) + + def add_max_alibi_bias(self, bias: float) -> None: + self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias) + + def add_clamp_kqv(self, value: float) -> None: + self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value) + + def add_layer_norm_eps(self, value: float) -> None: + self.add_float32(Keys.Attention.LAYERNORM_EPS.format(arch=self.arch), value) + + def add_layer_norm_rms_eps(self, value: float) -> None: + self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value) + + def add_rope_dimension_count(self, count: int) -> None: + self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count) + + def add_rope_freq_base(self, value: float) -> None: + self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value) + + def add_rope_scaling_type(self, value: RopeScalingType) -> None: + self.add_string(Keys.Rope.SCALING_TYPE.format(arch=self.arch), value.value) + + def add_rope_scaling_factor(self, value: float) -> None: + self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value) + + def add_rope_scaling_orig_ctx_len(self, value: int) -> None: + self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value) + + def add_rope_scaling_finetuned(self, value: bool) -> None: + self.add_bool(Keys.Rope.SCALING_FINETUNED.format(arch=self.arch), value) + + def add_tokenizer_model(self, model: str) -> None: + self.add_string(Keys.Tokenizer.MODEL, model) + + def add_token_list(self, tokens: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None: + self.add_array(Keys.Tokenizer.LIST, tokens) + + def add_token_merges(self, merges: Sequence[str] | Sequence[bytes] | Sequence[bytearray]) -> None: + self.add_array(Keys.Tokenizer.MERGES, merges) + + def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None: + self.add_array(Keys.Tokenizer.TOKEN_TYPE, types) + + def add_token_scores(self, scores: Sequence[float]) -> None: + self.add_array(Keys.Tokenizer.SCORES, scores) + + def add_bos_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.BOS_ID, id) + + def add_eos_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.EOS_ID, id) + + def add_unk_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.UNK_ID, id) + + def add_sep_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.SEP_ID, id) + + def add_pad_token_id(self, id: int) -> None: + self.add_uint32(Keys.Tokenizer.PAD_ID, id) + + def add_add_bos_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_BOS, value) + + def add_add_eos_token(self, value: bool) -> None: + self.add_bool(Keys.Tokenizer.ADD_EOS, value) + + def add_chat_template(self, value: str) -> None: + self.add_string(Keys.Tokenizer.CHAT_TEMPLATE, value) + + def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes: + pack_prefix = '' + if not skip_pack_prefix: + pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>' + return struct.pack(f'{pack_prefix}{fmt}', value) + + def _write_packed(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> None: + self.fout.write(self._pack(fmt, value, skip_pack_prefix)) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py new file mode 100644 index 000000000..22ad8b8fc --- /dev/null +++ b/gguf-py/gguf/tensor_mapping.py @@ -0,0 +1,257 @@ +from __future__ import annotations + +from typing import Sequence + +from .constants import MODEL_ARCH, MODEL_TENSOR, MODEL_TENSORS, TENSOR_NAMES + + +class TensorNameMap: + mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { + # Token embeddings + MODEL_TENSOR.TOKEN_EMBD: ( + "gpt_neox.embed_in", # gptneox + "transformer.wte", # gpt2 gpt-j mpt refact + "transformer.word_embeddings", # falcon + "word_embeddings", # bloom + "model.embed_tokens", # llama-hf + "tok_embeddings", # llama-pth + "embeddings.word_embeddings", # bert + "language_model.embedding.word_embeddings", # persimmon + ), + + # Token type embeddings + MODEL_TENSOR.TOKEN_TYPES: ( + "embeddings.token_type_embeddings", # bert + ), + + # Normalization of token embeddings + MODEL_TENSOR.TOKEN_EMBD_NORM: ( + "word_embeddings_layernorm", # bloom + ), + + # Position embeddings + MODEL_TENSOR.POS_EMBD: ( + "transformer.wpe", # gpt2 + "embeddings.position_embeddings", # bert + ), + + # Output + MODEL_TENSOR.OUTPUT: ( + "embed_out", # gptneox + "lm_head", # gpt2 mpt falcon llama-hf baichuan + "output", # llama-pth bloom + "word_embeddings_for_head", # persimmon + ), + + # Output norm + MODEL_TENSOR.OUTPUT_NORM: ( + "gpt_neox.final_layer_norm", # gptneox + "transformer.ln_f", # gpt2 gpt-j falcon + "model.norm", # llama-hf baichuan + "norm", # llama-pth + "embeddings.LayerNorm", # bert + "transformer.norm_f", # mpt + "ln_f", # refact bloom + "language_model.encoder.final_layernorm", # persimmon + ), + + # Rope frequencies + MODEL_TENSOR.ROPE_FREQS: ( + "rope.freqs", # llama-pth + ), + } + + block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = { + # Attention norm + MODEL_TENSOR.ATTN_NORM: ( + "gpt_neox.layers.{bid}.input_layernorm", # gptneox + "transformer.h.{bid}.ln_1", # gpt2 gpt-j refact + "transformer.blocks.{bid}.norm_1", # mpt + "transformer.h.{bid}.input_layernorm", # falcon7b + "h.{bid}.input_layernorm", # bloom + "transformer.h.{bid}.ln_mlp", # falcon40b + "model.layers.{bid}.input_layernorm", # llama-hf + "layers.{bid}.attention_norm", # llama-pth + "encoder.layer.{bid}.attention.output.LayerNorm", # bert + "language_model.encoder.layers.{bid}.input_layernorm", # persimmon + "model.layers.{bid}.ln1", # yi + ), + + # Attention norm 2 + MODEL_TENSOR.ATTN_NORM_2: ( + "transformer.h.{bid}.ln_attn", # falcon40b + ), + + # Attention query-key-value + MODEL_TENSOR.ATTN_QKV: ( + "gpt_neox.layers.{bid}.attention.query_key_value", # gptneox + "transformer.h.{bid}.attn.c_attn", # gpt2 + "transformer.blocks.{bid}.attn.Wqkv", # mpt + "transformer.h.{bid}.self_attention.query_key_value", # falcon + "h.{bid}.self_attention.query_key_value", # bloom + "language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon + ), + + # Attention query + MODEL_TENSOR.ATTN_Q: ( + "model.layers.{bid}.self_attn.q_proj", # llama-hf + "layers.{bid}.attention.wq", # llama-pth + "encoder.layer.{bid}.attention.self.query", # bert + "transformer.h.{bid}.attn.q_proj", # gpt-j + ), + + # Attention key + MODEL_TENSOR.ATTN_K: ( + "model.layers.{bid}.self_attn.k_proj", # llama-hf + "layers.{bid}.attention.wk", # llama-pth + "encoder.layer.{bid}.attention.self.key", # bert + "transformer.h.{bid}.attn.k_proj", # gpt-j + ), + + # Attention value + MODEL_TENSOR.ATTN_V: ( + "model.layers.{bid}.self_attn.v_proj", # llama-hf + "layers.{bid}.attention.wv", # llama-pth + "encoder.layer.{bid}.attention.self.value", # bert + "transformer.h.{bid}.attn.v_proj", # gpt-j + ), + + # Attention output + MODEL_TENSOR.ATTN_OUT: ( + "gpt_neox.layers.{bid}.attention.dense", # gptneox + "transformer.h.{bid}.attn.c_proj", # gpt2 refact + "transformer.blocks.{bid}.attn.out_proj", # mpt + "transformer.h.{bid}.self_attention.dense", # falcon + "h.{bid}.self_attention.dense", # bloom + "model.layers.{bid}.self_attn.o_proj", # llama-hf + "layers.{bid}.attention.wo", # llama-pth + "encoder.layer.{bid}.attention.output.dense", # bert + "transformer.h.{bid}.attn.out_proj", # gpt-j + "language_model.encoder.layers.{bid}.self_attention.dense", # persimmon + ), + + # Rotary embeddings + MODEL_TENSOR.ATTN_ROT_EMBD: ( + "model.layers.{bid}.self_attn.rotary_emb.inv_freq", # llama-hf + "layers.{bid}.attention.inner_attention.rope.freqs", # llama-pth + ), + + # Feed-forward norm + MODEL_TENSOR.FFN_NORM: ( + "gpt_neox.layers.{bid}.post_attention_layernorm", # gptneox + "transformer.h.{bid}.ln_2", # gpt2 refact + "h.{bid}.post_attention_layernorm", # bloom + "transformer.blocks.{bid}.norm_2", # mpt + "model.layers.{bid}.post_attention_layernorm", # llama-hf + "layers.{bid}.ffn_norm", # llama-pth + "encoder.layer.{bid}.output.LayerNorm", # bert + "language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon + "model.layers.{bid}.ln2", # yi + ), + + # Feed-forward up + MODEL_TENSOR.FFN_UP: ( + "gpt_neox.layers.{bid}.mlp.dense_h_to_4h", # gptneox + "transformer.h.{bid}.mlp.c_fc", # gpt2 + "transformer.blocks.{bid}.ffn.up_proj", # mpt + "transformer.h.{bid}.mlp.dense_h_to_4h", # falcon + "h.{bid}.mlp.dense_h_to_4h", # bloom + "model.layers.{bid}.mlp.up_proj", # llama-hf refact + "layers.{bid}.feed_forward.w3", # llama-pth + "encoder.layer.{bid}.intermediate.dense", # bert + "transformer.h.{bid}.mlp.fc_in", # gpt-j + "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon + ), + + # Feed-forward gate + MODEL_TENSOR.FFN_GATE: ( + "model.layers.{bid}.mlp.gate_proj", # llama-hf refact + "layers.{bid}.feed_forward.w1", # llama-pth + ), + + # Feed-forward down + MODEL_TENSOR.FFN_DOWN: ( + "gpt_neox.layers.{bid}.mlp.dense_4h_to_h", # gptneox + "transformer.h.{bid}.mlp.c_proj", # gpt2 refact + "transformer.blocks.{bid}.ffn.down_proj", # mpt + "transformer.h.{bid}.mlp.dense_4h_to_h", # falcon + "h.{bid}.mlp.dense_4h_to_h", # bloom + "model.layers.{bid}.mlp.down_proj", # llama-hf + "layers.{bid}.feed_forward.w2", # llama-pth + "encoder.layer.{bid}.output.dense", # bert + "transformer.h.{bid}.mlp.fc_out", # gpt-j + "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon + ), + + MODEL_TENSOR.ATTN_Q_NORM: ( + "language_model.encoder.layers.{bid}.self_attention.q_layernorm", + ), + + MODEL_TENSOR.ATTN_K_NORM: ( + "language_model.encoder.layers.{bid}.self_attention.k_layernorm", + ), + + MODEL_TENSOR.ROPE_FREQS: ( + "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq", # persimmon + ), + } + + mapping: dict[str, tuple[MODEL_TENSOR, str]] + + def __init__(self, arch: MODEL_ARCH, n_blocks: int): + self.mapping = {} + for tensor, keys in self.mappings_cfg.items(): + if tensor not in MODEL_TENSORS[arch]: + continue + tensor_name = TENSOR_NAMES[tensor] + self.mapping[tensor_name] = (tensor, tensor_name) + for key in keys: + self.mapping[key] = (tensor, tensor_name) + for bid in range(n_blocks): + for tensor, keys in self.block_mappings_cfg.items(): + if tensor not in MODEL_TENSORS[arch]: + continue + tensor_name = TENSOR_NAMES[tensor].format(bid = bid) + self.mapping[tensor_name] = (tensor, tensor_name) + for key in keys: + key = key.format(bid = bid) + self.mapping[key] = (tensor, tensor_name) + + def get_type_and_name(self, key: str, try_suffixes: Sequence[str] = ()) -> tuple[MODEL_TENSOR, str] | None: + result = self.mapping.get(key) + if result is not None: + return result + for suffix in try_suffixes: + if key.endswith(suffix): + result = self.mapping.get(key[:-len(suffix)]) + if result is not None: + return result[0], result[1] + suffix + return None + + def get_name(self, key: str, try_suffixes: Sequence[str] = ()) -> str | None: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[1] + + def get_type(self, key: str, try_suffixes: Sequence[str] = ()) -> MODEL_TENSOR | None: + result = self.get_type_and_name(key, try_suffixes = try_suffixes) + if result is None: + return None + return result[0] + + def __getitem__(self, key: str) -> str: + try: + return self.mapping[key][1] + except KeyError: + raise KeyError(key) + + def __contains__(self, key: str) -> bool: + return key in self.mapping + + def __repr__(self) -> str: + return repr(self.mapping) + + +def get_tensor_name_map(arch: MODEL_ARCH, n_blocks: int) -> TensorNameMap: + return TensorNameMap(arch, n_blocks) diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py new file mode 100644 index 000000000..de3e5edb5 --- /dev/null +++ b/gguf-py/gguf/vocab.py @@ -0,0 +1,183 @@ +from __future__ import annotations + +import json +import os +import sys +from pathlib import Path +from typing import Any, Callable + +from .gguf_writer import GGUFWriter + + +class SpecialVocab: + merges: list[str] + add_special_token: dict[str, bool] + special_token_ids: dict[str, int] + chat_template: str | None + + def __init__( + self, path: str | os.PathLike[str], load_merges: bool = False, + special_token_types: tuple[str, ...] | None = None, + n_vocab: int | None = None, + ): + self.special_token_ids = {} + self.add_special_token = {} + self.n_vocab = n_vocab + self.load_merges = load_merges + self.merges = [] + self.chat_template = None + if special_token_types is not None: + self.special_token_types = special_token_types + else: + self.special_token_types = ('bos', 'eos', 'unk', 'sep', 'pad') + self._load(Path(path)) + + def __repr__(self) -> str: + return ''.format( + len(self.merges), self.special_token_ids or "unset", self.add_special_token or "unset", + ) + + def add_to_gguf(self, gw: GGUFWriter, quiet: bool = False) -> None: + if self.merges: + if not quiet: + print(f'gguf: Adding {len(self.merges)} merge(s).') + gw.add_token_merges(self.merges) + elif self.load_merges: + print( + 'gguf: WARNING: Adding merges requested but no merges found, output may be non-functional.', + file = sys.stderr, + ) + for typ, tokid in self.special_token_ids.items(): + id_handler: Callable[[int], None] | None = getattr(gw, f'add_{typ}_token_id', None) + if id_handler is None: + print( + f'gguf: WARNING: No handler for special token type {typ} with id {tokid} - skipping', + file = sys.stderr, + ) + continue + if not quiet: + print(f'gguf: Setting special token type {typ} to {tokid}') + id_handler(tokid) + for typ, value in self.add_special_token.items(): + add_handler: Callable[[bool], None] | None = getattr(gw, f'add_add_{typ}_token', None) + if add_handler is None: + print( + f'gguf: WARNING: No handler for add_{typ}_token with value {value} - skipping', + file = sys.stderr, + ) + continue + if not quiet: + print(f'gguf: Setting add_{typ}_token to {value}') + add_handler(value) + if self.chat_template is not None: + if not quiet: + print(f'gguf: Setting chat_template to {self.chat_template}') + gw.add_chat_template(self.chat_template) + + def _load(self, path: Path) -> None: + self._try_load_from_tokenizer_json(path) + self._try_load_from_config_json(path) + if self.load_merges and not self.merges: + self._try_load_merges_txt(path) + + def _try_load_merges_txt(self, path: Path) -> bool: + merges_file = path / 'merges.txt' + if not merges_file.is_file(): + return False + with open(merges_file, 'r') as fp: + first_line = next(fp, '').strip() + if not first_line.startswith('#'): + fp.seek(0) + line_num = 0 + else: + line_num = 1 + merges = [] + for line in fp: + line_num += 1 + line = line.strip() + if not line: + continue + parts = line.split(None, 3) + if len(parts) != 2: + print( + f'gguf: WARNING: {merges_file.name}: Line {line_num}: Entry malformed, ignoring', + file = sys.stderr, + ) + continue + merges.append(f'{parts[0]} {parts[1]}') + self.merges = merges + return True + + def _set_special_token(self, typ: str, tid: Any) -> None: + if not isinstance(tid, int) or tid < 0: + return + if self.n_vocab is None or tid < self.n_vocab: + if typ in self.special_token_ids: + return + self.special_token_ids[typ] = tid + return + print( + f'gguf: WARNING: Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping', + file = sys.stderr, + ) + + def _try_load_from_tokenizer_json(self, path: Path) -> bool: + tokenizer_file = path / 'tokenizer.json' + if tokenizer_file.is_file(): + with open(tokenizer_file, encoding = 'utf-8') as f: + tokenizer = json.load(f) + if self.load_merges: + merges = tokenizer.get('model', {}).get('merges') + if isinstance(merges, list) and merges and isinstance(merges[0], str): + self.merges = merges + added_tokens = tokenizer.get('added_tokens', {}) + else: + added_tokens = {} + tokenizer_config_file = path / 'tokenizer_config.json' + if not tokenizer_config_file.is_file(): + return True + with open(tokenizer_config_file, encoding = 'utf-8') as f: + tokenizer_config = json.load(f) + chat_template = tokenizer_config.get('chat_template') + if chat_template is None or isinstance(chat_template, str): + self.chat_template = chat_template + else: + print( + f'gguf: WARNING: Bad type for chat_template field in {tokenizer_config_file!r} - ignoring', + file = sys.stderr + ) + for typ in self.special_token_types: + add_entry = tokenizer_config.get(f'add_{typ}_token') + if isinstance(add_entry, bool): + self.add_special_token[typ] = add_entry + if not added_tokens: + # We will need this to get the content for the token, so if it's empty + # may as well just give up. + continue + entry = tokenizer_config.get(f'{typ}_token') + if isinstance(entry, str): + tc_content = entry + elif isinstance(entry, dict): + entry_content = entry.get('content') + if not isinstance(entry_content, str): + continue + tc_content = entry_content + else: + continue + # We only need the first match here. + maybe_token_id = next( + (atok.get('id') for atok in added_tokens if atok.get('content') == tc_content), + None, + ) + self._set_special_token(typ, maybe_token_id) + return True + + def _try_load_from_config_json(self, path: Path) -> bool: + config_file = path / 'config.json' + if not config_file.is_file(): + return False + with open(config_file, encoding = 'utf-8') as f: + config = json.load(f) + for typ in self.special_token_types: + self._set_special_token(typ, config.get(f'{typ}_token_id')) + return True diff --git a/gguf-py/pyproject.toml b/gguf-py/pyproject.toml index f0741a7c2..e6374bfe8 100644 --- a/gguf-py/pyproject.toml +++ b/gguf-py/pyproject.toml @@ -1,11 +1,12 @@ [tool.poetry] name = "gguf" -version = "0.4.5" -description = "Write ML models in GGUF for GGML" +version = "0.6.0" +description = "Read and write ML models in GGUF for GGML" authors = ["GGML "] packages = [ {include = "gguf"}, {include = "gguf/py.typed"}, + {include = "scripts"}, ] readme = "README.md" homepage = "https://ggml.ai" @@ -27,3 +28,8 @@ pytest = "^5.2" [build-system] requires = ["poetry-core>=1.0.0"] build-backend = "poetry.core.masonry.api" + +[tool.poetry.scripts] +gguf-convert-endian = "scripts:gguf_convert_endian_entrypoint" +gguf-dump = "scripts:gguf_dump_entrypoint" +gguf-set-metadata = "scripts:gguf_set_metadata_entrypoint" diff --git a/gguf-py/scripts/__init__.py b/gguf-py/scripts/__init__.py new file mode 100644 index 000000000..77132db7a --- /dev/null +++ b/gguf-py/scripts/__init__.py @@ -0,0 +1,12 @@ +import os + +from importlib import import_module + + +os.environ["NO_LOCAL_GGUF"] = "TRUE" + +gguf_convert_endian_entrypoint = import_module("scripts.gguf-convert-endian").main +gguf_dump_entrypoint = import_module("scripts.gguf-dump").main +gguf_set_metadata_entrypoint = import_module("scripts.gguf-set-metadata").main + +del import_module, os diff --git a/gguf-py/scripts/gguf-convert-endian.py b/gguf-py/scripts/gguf-convert-endian.py new file mode 100755 index 000000000..10a16ad06 --- /dev/null +++ b/gguf-py/scripts/gguf-convert-endian.py @@ -0,0 +1,112 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import argparse +import os +import sys +from pathlib import Path + +import numpy as np + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent)) + +import gguf + + +def convert_byteorder(reader: gguf.GGUFReader, args: argparse.Namespace) -> None: + if np.uint32(1) == np.uint32(1).newbyteorder("<"): + # Host is little endian + host_endian = "little" + swapped_endian = "big" + else: + # Sorry PDP or other weird systems that don't use BE or LE. + host_endian = "big" + swapped_endian = "little" + if reader.byte_order == "S": + file_endian = swapped_endian + else: + file_endian = host_endian + order = host_endian if args.order == "native" else args.order + print(f"* Host is {host_endian.upper()} endian, GGUF file seems to be {file_endian.upper()} endian") + if file_endian == order: + print(f"* File is already {order.upper()} endian. Nothing to do.") + sys.exit(0) + print("* Checking tensors for conversion compatibility") + for tensor in reader.tensors: + if tensor.tensor_type not in ( + gguf.GGMLQuantizationType.F32, + gguf.GGMLQuantizationType.F16, + gguf.GGMLQuantizationType.Q8_0, + ): + raise ValueError(f"Cannot handle type {tensor.tensor_type.name} for tensor {repr(tensor.name)}") + print(f"* Preparing to convert from {file_endian.upper()} to {order.upper()}") + if args.dry_run: + return + print("\n*** Warning *** Warning *** Warning **") + print("* This conversion process may damage the file. Ensure you have a backup.") + if order != host_endian: + print("* Requested endian differs from host, you will not be able to load the model on this machine.") + print("* The file will be modified immediately, so if conversion fails or is interrupted") + print("* the file will be corrupted. Enter exactly YES if you are positive you want to proceed:") + response = input("YES, I am sure> ") + if response != "YES": + print("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + print(f"\n* Converting fields ({len(reader.fields)})") + for idx, field in enumerate(reader.fields.values()): + print(f"- {idx:4}: Converting field {repr(field.name)}, part count: {len(field.parts)}") + for part in field.parts: + part.byteswap(inplace=True) + print(f"\n* Converting tensors ({len(reader.tensors)})") + for idx, tensor in enumerate(reader.tensors): + print( + f" - {idx:4}: Converting tensor {repr(tensor.name)}, type={tensor.tensor_type.name}, " + f"elements={tensor.n_elements}... ", + end="", + ) + tensor_type = tensor.tensor_type + for part in tensor.field.parts: + part.byteswap(inplace=True) + if tensor_type != gguf.GGMLQuantizationType.Q8_0: + tensor.data.byteswap(inplace=True) + print() + continue + # A Q8_0 block consists of a f16 delta followed by 32 int8 quants, so 34 bytes + block_size = 34 + n_blocks = len(tensor.data) // block_size + for block_num in range(n_blocks): + block_offs = block_num * block_size + # I know I said f16, but it doesn't matter here - any simple 16 bit type works. + delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16) + delta.byteswap(inplace=True) + if block_num % 100000 == 0: + print(f"[{(n_blocks - block_num) // 1000}K]", end="") + sys.stdout.flush() + print() + print("* Completion") + + +def main() -> None: + parser = argparse.ArgumentParser(description="Convert GGUF file byte order") + parser.add_argument( + "model", type=str, + help="GGUF format model filename", + ) + parser.add_argument( + "order", type=str, choices=['big', 'little', 'native'], + help="Requested byte order", + ) + parser.add_argument( + "--dry-run", action="store_true", + help="Don't actually change anything", + ) + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + print(f'* Loading: {args.model}') + reader = gguf.GGUFReader(args.model, 'r' if args.dry_run else 'r+') + convert_byteorder(reader, args) + + +if __name__ == "__main__": + main() diff --git a/gguf-py/scripts/gguf-dump.py b/gguf-py/scripts/gguf-dump.py new file mode 100755 index 000000000..dbf891508 --- /dev/null +++ b/gguf-py/scripts/gguf-dump.py @@ -0,0 +1,117 @@ +#!/usr/bin/env python3 +from __future__ import annotations + +import argparse +import os +import sys +from pathlib import Path +from typing import Any + +import numpy as np + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent)) + +from gguf import GGUFReader, GGUFValueType # noqa: E402 + + +def get_file_host_endian(reader: GGUFReader) -> tuple[str, str]: + host_endian = 'LITTLE' if np.uint32(1) == np.uint32(1).newbyteorder("<") else 'BIG' + if reader.byte_order == 'S': + file_endian = 'BIG' if host_endian == 'LITTLE' else 'LITTLE' + else: + file_endian = host_endian + return (host_endian, file_endian) + + +# For more information about what field.parts and field.data represent, +# please see the comments in the modify_gguf.py example. +def dump_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + host_endian, file_endian = get_file_host_endian(reader) + print(f'* File is {file_endian} endian, script is running on a {host_endian} endian host.') + print(f'\n* Dumping {len(reader.fields)} key/value pair(s)') + for n, field in enumerate(reader.fields.values(), 1): + if not field.types: + pretty_type = 'N/A' + elif field.types[0] == GGUFValueType.ARRAY: + nest_count = len(field.types) - 1 + pretty_type = '[' * nest_count + str(field.types[-1].name) + ']' * nest_count + else: + pretty_type = str(field.types[-1].name) + print(f' {n:5}: {pretty_type:10} | {len(field.data):8} | {field.name}', end = '') + if len(field.types) == 1: + curr_type = field.types[0] + if curr_type == GGUFValueType.STRING: + print(' = {0}'.format(repr(str(bytes(field.parts[-1]), encoding='utf8')[:60])), end = '') + elif field.types[0] in reader.gguf_scalar_to_np: + print(' = {0}'.format(field.parts[-1][0]), end = '') + print() + if args.no_tensors: + return + print(f'\n* Dumping {len(reader.tensors)} tensor(s)') + for n, tensor in enumerate(reader.tensors, 1): + prettydims = ', '.join('{0:5}'.format(d) for d in list(tensor.shape) + [1] * (4 - len(tensor.shape))) + print(f' {n:5}: {tensor.n_elements:10} | {prettydims} | {tensor.tensor_type.name:7} | {tensor.name}') + + +def dump_metadata_json(reader: GGUFReader, args: argparse.Namespace) -> None: + import json + host_endian, file_endian = get_file_host_endian(reader) + metadata: dict[str, Any] = {} + tensors: dict[str, Any] = {} + result = { + "filename": args.model, + "endian": file_endian, + "metadata": metadata, + "tensors": tensors, + } + for idx, field in enumerate(reader.fields.values()): + curr: dict[str, Any] = { + "index": idx, + "type": field.types[0].name if field.types else 'UNKNOWN', + "offset": field.offset, + } + metadata[field.name] = curr + if field.types[:1] == [GGUFValueType.ARRAY]: + curr["array_types"] = [t.name for t in field.types][1:] + if not args.json_array: + continue + itype = field.types[-1] + if itype == GGUFValueType.STRING: + curr["value"] = [str(bytes(field.parts[idx]), encoding="utf-8") for idx in field.data] + else: + curr["value"] = [pv for idx in field.data for pv in field.parts[idx].tolist()] + elif field.types[0] == GGUFValueType.STRING: + curr["value"] = str(bytes(field.parts[-1]), encoding="utf-8") + else: + curr["value"] = field.parts[-1].tolist()[0] + if not args.no_tensors: + for idx, tensor in enumerate(reader.tensors): + tensors[tensor.name] = { + "index": idx, + "shape": tensor.shape.tolist(), + "type": tensor.tensor_type.name, + "offset": tensor.field.offset, + } + json.dump(result, sys.stdout) + + +def main() -> None: + parser = argparse.ArgumentParser(description="Dump GGUF file metadata") + parser.add_argument("model", type=str, help="GGUF format model filename") + parser.add_argument("--no-tensors", action="store_true", help="Don't dump tensor metadata") + parser.add_argument("--json", action="store_true", help="Produce JSON output") + parser.add_argument("--json-array", action="store_true", help="Include full array values in JSON output (long)") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + if not args.json: + print(f'* Loading: {args.model}') + reader = GGUFReader(args.model, 'r') + if args.json: + dump_metadata_json(reader, args) + else: + dump_metadata(reader, args) + + +if __name__ == '__main__': + main() diff --git a/gguf-py/scripts/gguf-set-metadata.py b/gguf-py/scripts/gguf-set-metadata.py new file mode 100755 index 000000000..3ebdfa898 --- /dev/null +++ b/gguf-py/scripts/gguf-set-metadata.py @@ -0,0 +1,90 @@ +#!/usr/bin/env python3 +import argparse +import os +import sys +from pathlib import Path + +# Necessary to load the local gguf package +if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists(): + sys.path.insert(0, str(Path(__file__).parent.parent)) + +from gguf import GGUFReader # noqa: E402 + + +def minimal_example(filename: str) -> None: + reader = GGUFReader(filename, 'r+') + field = reader.fields['tokenizer.ggml.bos_token_id'] + if field is None: + return + part_index = field.data[0] + field.parts[part_index][0] = 2 # Set tokenizer.ggml.bos_token_id to 2 + # + # So what's this field.data thing? It's helpful because field.parts contains + # _every_ part of the GGUF field. For example, tokenizer.ggml.bos_token_id consists + # of: + # + # Part index 0: Key length (27) + # Part index 1: Key data ("tokenizer.ggml.bos_token_id") + # Part index 2: Field type (4, the id for GGUFValueType.UINT32) + # Part index 3: Field value + # + # Note also that each part is an NDArray slice, so even a part that + # is only a single value like the key length will be a NDArray of + # the key length type (numpy.uint32). + # + # The .data attribute in the Field is a list of relevant part indexes + # and doesn't contain internal GGUF details like the key length part. + # In this case, .data will be [3] - just the part index of the + # field value itself. + + +def set_metadata(reader: GGUFReader, args: argparse.Namespace) -> None: + field = reader.get_field(args.key) + if field is None: + print(f'! Field {repr(args.key)} not found', file = sys.stderr) + sys.exit(1) + # Note that field.types is a list of types. This is because the GGUF + # format supports arrays. For example, an array of UINT32 would + # look like [GGUFValueType.ARRAY, GGUFValueType.UINT32] + handler = reader.gguf_scalar_to_np.get(field.types[0]) if field.types else None + if handler is None: + print( + f'! This tool only supports changing simple values, {repr(args.key)} has unsupported type {field.types}', + file = sys.stderr, + ) + sys.exit(1) + current_value = field.parts[field.data[0]][0] + new_value = handler(args.value) + print(f'* Preparing to change field {repr(args.key)} from {current_value} to {new_value}') + if current_value == new_value: + print(f'- Key {repr(args.key)} already set to requested value {current_value}') + sys.exit(0) + if args.dry_run: + sys.exit(0) + if not args.force: + print('*** Warning *** Warning *** Warning **') + print('* Changing fields in a GGUF file can make it unusable. Proceed at your own risk.') + print('* Enter exactly YES if you are positive you want to proceed:') + response = input('YES, I am sure> ') + if response != 'YES': + print("You didn't enter YES. Okay then, see ya!") + sys.exit(0) + field.parts[field.data[0]][0] = new_value + print('* Field changed. Successful completion.') + + +def main() -> None: + parser = argparse.ArgumentParser(description="Set a simple value in GGUF file metadata") + parser.add_argument("model", type=str, help="GGUF format model filename") + parser.add_argument("key", type=str, help="Metadata key to set") + parser.add_argument("value", type=str, help="Metadata value to set") + parser.add_argument("--dry-run", action="store_true", help="Don't actually change anything") + parser.add_argument("--force", action="store_true", help="Change the field without confirmation") + args = parser.parse_args(None if len(sys.argv) > 1 else ["--help"]) + print(f'* Loading: {args.model}') + reader = GGUFReader(args.model, 'r' if args.dry_run else 'r+') + set_metadata(reader, args) + + +if __name__ == '__main__': + main() diff --git a/gguf-py/tests/test_gguf.py b/gguf-py/tests/test_gguf.py index 512531dd2..0adeb7d55 100644 --- a/gguf-py/tests/test_gguf.py +++ b/gguf-py/tests/test_gguf.py @@ -1,7 +1,7 @@ -import gguf +import gguf # noqa: F401 # TODO: add tests -def test_write_gguf(): +def test_write_gguf() -> None: pass diff --git a/grammars/README.md b/grammars/README.md index 7f3b11ca5..e1383fa5c 100644 --- a/grammars/README.md +++ b/grammars/README.md @@ -55,7 +55,7 @@ The order of symbols in a sequence matter. For example, in `"1. " move " " move Alternatives, denoted by `|`, give different sequences that are acceptable. For example, in `move ::= pawn | nonpawn | castle`, `move` can be a `pawn` move, a `nonpawn` move, or a `castle`. -Parentheses `()` can be used to group sequences, which allows for embedding alternatives in a larger rule or applying repetition and optptional symbols (below) to a sequence. +Parentheses `()` can be used to group sequences, which allows for embedding alternatives in a larger rule or applying repetition and optional symbols (below) to a sequence. ## Repetition and Optional Symbols @@ -67,7 +67,7 @@ Parentheses `()` can be used to group sequences, which allows for embedding alte Comments can be specified with `#`: ``` -# defines optional whitspace +# defines optional whitespace ws ::= [ \t\n]+ ``` diff --git a/llama.cpp b/llama.cpp index 08daaf889..ec23485fd 100644 --- a/llama.cpp +++ b/llama.cpp @@ -54,13 +54,16 @@ #include #include #include +#include #include #include #include #include #include #include +#include #include +#include #include #include #include @@ -69,11 +72,10 @@ #include #include #include +#include #include #include #include -#include -#include #if defined(_MSC_VER) #pragma warning(disable: 4244 4267) // possible loss of data @@ -89,6 +91,8 @@ #define LLAMA_ATTRIBUTE_FORMAT(...) #endif +#define LLAMA_MAX_NODES 8192 + // // logging // @@ -188,6 +192,7 @@ enum llm_arch { LLM_ARCH_PERSIMMON, LLM_ARCH_REFACT, LLM_ARCH_BLOOM, + LLM_ARCH_STABLELM, LLM_ARCH_UNKNOWN, }; @@ -203,6 +208,7 @@ static std::map LLM_ARCH_NAMES = { { LLM_ARCH_PERSIMMON, "persimmon" }, { LLM_ARCH_REFACT, "refact" }, { LLM_ARCH_BLOOM, "bloom" }, + { LLM_ARCH_STABLELM, "stablelm" }, }; enum llm_kv { @@ -234,6 +240,10 @@ enum llm_kv { LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_FREQ_BASE, LLM_KV_ROPE_SCALE_LINEAR, + LLM_KV_ROPE_SCALING_TYPE, + LLM_KV_ROPE_SCALING_FACTOR, + LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, + LLM_KV_ROPE_SCALING_FINETUNED, LLM_KV_TOKENIZER_MODEL, LLM_KV_TOKENIZER_LIST, @@ -245,6 +255,8 @@ enum llm_kv { LLM_KV_TOKENIZER_UNK_ID, LLM_KV_TOKENIZER_SEP_ID, LLM_KV_TOKENIZER_PAD_ID, + LLM_KV_TOKENIZER_ADD_BOS, + LLM_KV_TOKENIZER_ADD_EOS, LLM_KV_TOKENIZER_HF_JSON, LLM_KV_TOKENIZER_RWKV, }; @@ -275,9 +287,13 @@ static std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_LAYERNORM_EPS, "%s.attention.layer_norm_epsilon" }, { LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, "%s.attention.layer_norm_rms_epsilon" }, - { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, - { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, - { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, + { LLM_KV_ROPE_FREQ_BASE, "%s.rope.freq_base" }, + { LLM_KV_ROPE_SCALE_LINEAR, "%s.rope.scale_linear" }, + { LLM_KV_ROPE_SCALING_TYPE, "%s.rope.scaling.type" }, + { LLM_KV_ROPE_SCALING_FACTOR, "%s.rope.scaling.factor" }, + { LLM_KV_ROPE_SCALING_ORIG_CTX_LEN, "%s.rope.scaling.original_context_length" }, + { LLM_KV_ROPE_SCALING_FINETUNED, "%s.rope.scaling.finetuned" }, { LLM_KV_TOKENIZER_MODEL, "tokenizer.ggml.model" }, { LLM_KV_TOKENIZER_LIST, "tokenizer.ggml.tokens" }, @@ -289,6 +305,8 @@ static std::map LLM_KV_NAMES = { { LLM_KV_TOKENIZER_UNK_ID, "tokenizer.ggml.unknown_token_id" }, { LLM_KV_TOKENIZER_SEP_ID, "tokenizer.ggml.seperator_token_id" }, { LLM_KV_TOKENIZER_PAD_ID, "tokenizer.ggml.padding_token_id" }, + { LLM_KV_TOKENIZER_ADD_BOS, "tokenizer.ggml.add_bos_token" }, + { LLM_KV_TOKENIZER_ADD_EOS, "tokenizer.ggml.add_eos_token" }, { LLM_KV_TOKENIZER_HF_JSON, "tokenizer.huggingface.json" }, { LLM_KV_TOKENIZER_RWKV, "tokenizer.rwkv.world" }, }; @@ -483,6 +501,25 @@ static std::map> LLM_TENSOR_NAMES = { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, }, }, + { + LLM_ARCH_STABLELM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ROPE_FREQS, "rope_freqs" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { LLM_ARCH_UNKNOWN, { @@ -551,6 +588,76 @@ do { \ } \ } while (0) +static std::map LLAMA_ROPE_SCALING_TYPES = { + { LLAMA_ROPE_SCALING_NONE, "none" }, + { LLAMA_ROPE_SCALING_LINEAR, "linear" }, + { LLAMA_ROPE_SCALING_YARN, "yarn" }, +}; + +static int8_t llama_rope_scaling_type_from_string(const std::string & name) { + for (const auto & kv : LLAMA_ROPE_SCALING_TYPES) { + if (kv.second == name) { + return kv.first; + } + } + + return LLAMA_ROPE_SCALING_UNSPECIFIED; +} + +static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) { + switch (type) { + case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]); + case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]); + case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]); + case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]); + case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]); + case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]); + case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]); + case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]); + case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]); + case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]); + case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false"; + default: return format("unknown type %d", type); + } +} + +static std::string gguf_kv_to_str(struct gguf_context * ctx_gguf, int i) { + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + + switch (type) { + case GGUF_TYPE_STRING: + return gguf_get_val_str(ctx_gguf, i); + case GGUF_TYPE_ARRAY: + { + const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i); + int arr_n = gguf_get_arr_n(ctx_gguf, i); + const void * data = gguf_get_arr_data(ctx_gguf, i); + std::stringstream ss; + ss << "["; + for (int j = 0; j < arr_n; j++) { + if (arr_type == GGUF_TYPE_STRING) { + std::string val = gguf_get_arr_str(ctx_gguf, i, j); + // escape quotes + replace_all(val, "\\", "\\\\"); + replace_all(val, "\"", "\\\""); + ss << '"' << val << '"'; + } else if (arr_type == GGUF_TYPE_ARRAY) { + ss << "???"; + } else { + ss << gguf_data_to_str(arr_type, data, j); + } + if (j < arr_n - 1) { + ss << ", "; + } + } + ss << "]"; + return ss.str(); + } + default: + return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0); + } +} + // // ggml helpers // @@ -570,19 +677,37 @@ static void ggml_graph_compute_helper(std::vector & buf, ggml_cgraph * // llama helpers // +inline void * llama_host_malloc(size_t n) { #ifdef GGML_USE_CUBLAS -# define llama_host_malloc(n) ggml_cuda_host_malloc(n) -# define llama_host_free(data) ggml_cuda_host_free(data) + if (ggml_cublas_loaded()) { + return ggml_cuda_host_malloc(n); + } else { + return malloc(n); + } #elif GGML_USE_METAL -# define llama_host_malloc(n) ggml_metal_host_malloc(n) -# define llama_host_free(data) ggml_metal_host_free(data) + return ggml_metal_host_malloc(n); #elif GGML_USE_CPU_HBM -# define llama_host_malloc(n) hbw_malloc(n) -# define llama_host_free(data) if (data != NULL) hbw_free(data) + return hbw_malloc(n); #else -# define llama_host_malloc(n) malloc(n) -# define llama_host_free(data) free(data) + return malloc(n); #endif +} + +inline void llama_host_free(void * ptr) { +#ifdef GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + return ggml_cuda_host_free(ptr); + } else { + return free(ptr); + } +#elif GGML_USE_METAL + return ggml_metal_host_free(ptr); +#elif GGML_USE_CPU_HBM + return hbw_free(ptr); +#else + return free(ptr); +#endif +} #if defined(_WIN32) static std::string llama_format_win_err(DWORD err) { @@ -969,7 +1094,7 @@ struct llama_mlock { typedef void (*offload_func_t)(struct ggml_tensor * tensor); -static void llama_nop(struct ggml_tensor * tensor) { // don't offload by default +static void ggml_offload_nop(struct ggml_tensor * tensor) { (void) tensor; } @@ -993,6 +1118,12 @@ static std::string llama_token_to_piece(const struct llama_context * ctx, llama_ // struct llama_state { + llama_state() { +#ifdef GGML_USE_METAL + ggml_metal_log_set_callback(log_callback, log_callback_user_data); +#endif + } + // We save the log callback globally ggml_log_callback log_callback = llama_log_callback_default; void * log_callback_user_data = nullptr; @@ -1016,9 +1147,9 @@ enum e_model { MODEL_70B, }; -static const size_t kB = 1024; -static const size_t MB = 1024*kB; -static const size_t GB = 1024*MB; +static const size_t kiB = 1024; +static const size_t MiB = 1024*kiB; +static const size_t GiB = 1024*MiB; struct llama_hparams { bool vocab_only; @@ -1034,8 +1165,11 @@ struct llama_hparams { float f_norm_eps; float f_norm_rms_eps; - float rope_freq_base_train; - float rope_freq_scale_train; + float rope_freq_base_train; + float rope_freq_scale_train; + uint32_t n_yarn_orig_ctx; + int8_t rope_scaling_type_train : 3; + bool rope_finetuned : 1; float f_clamp_kqv; float f_max_alibi_bias; @@ -1050,6 +1184,8 @@ struct llama_hparams { if (this->n_layer != other.n_layer) return true; if (this->n_rot != other.n_rot) return true; if (this->n_ff != other.n_ff) return true; + if (this->rope_finetuned != other.rope_finetuned) return true; + if (this->n_yarn_orig_ctx != other.n_yarn_orig_ctx) return true; const float EPSILON = 1e-9; @@ -1080,8 +1216,16 @@ struct llama_cparams { uint32_t n_threads; // number of threads to use for generation uint32_t n_threads_batch; // number of threads to use for batch processing - float rope_freq_base; - float rope_freq_scale; + float rope_freq_base; + float rope_freq_scale; + + uint32_t n_yarn_orig_ctx; + // These hyperparameters are not exposed in GGUF, because all + // existing YaRN models use the same values for them. + float yarn_ext_factor; + float yarn_attn_factor; + float yarn_beta_fast; + float yarn_beta_slow; bool mul_mat_q; }; @@ -1113,13 +1257,13 @@ struct llama_layer { struct ggml_tensor * ffn_norm_b; // ff - struct ggml_tensor * w1; // ffn_gate - struct ggml_tensor * w2; // ffn_down - struct ggml_tensor * w3; // ffn_up + struct ggml_tensor * ffn_gate; // w1 + struct ggml_tensor * ffn_down; // w2 + struct ggml_tensor * ffn_up; // w3 // ff bias - struct ggml_tensor * b2; // ffn_down - struct ggml_tensor * b3; // ffn_up + struct ggml_tensor * ffn_down_b; // b2 + struct ggml_tensor * ffn_up_b; // b3 }; struct llama_kv_cell { @@ -1142,6 +1286,7 @@ struct llama_kv_cache { // cannot be freely changed after a slot has been allocated. uint32_t head = 0; uint32_t size = 0; + uint32_t used = 0; // used cells (i.e. at least one seq_id) // computed before each graph build uint32_t n = 0; @@ -1161,9 +1306,11 @@ struct llama_kv_cache { } #ifdef GGML_USE_CUBLAS - ggml_cuda_free_data(k); - ggml_cuda_free_data(v); -#endif // GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + ggml_cuda_free_data(k); + ggml_cuda_free_data(v); + } +#endif } }; @@ -1194,6 +1341,9 @@ struct llama_vocab { id special_sep_id = -1; id special_pad_id = -1; + int special_add_bos = -1; // -1 unknown, 1 add, 0 don't add. + int special_add_eos = -1; // -1 unknown, 1 add, 0 don't add. + id linefeed_id = 13; id special_prefix_id = 32007; id special_middle_id = 32009; @@ -1225,8 +1375,8 @@ struct llama_model { llama_hparams hparams = {}; llama_vocab vocab; - struct ggml_tensor * tok_embeddings; - struct ggml_tensor * pos_embeddings; + struct ggml_tensor * tok_embd; + struct ggml_tensor * pos_embd; struct ggml_tensor * tok_norm; struct ggml_tensor * tok_norm_b; @@ -1238,6 +1388,9 @@ struct llama_model { int n_gpu_layers; + // gguf metadata + std::unordered_map gguf_kv; + // context struct ggml_context * ctx = NULL; @@ -1263,11 +1416,15 @@ struct llama_model { } #ifdef GGML_USE_CUBLAS - for (size_t i = 0; i < tensors_by_name.size(); ++i) { - ggml_cuda_free_data(tensors_by_name[i].second); + if (ggml_cublas_loaded()) { + for (size_t i = 0; i < tensors_by_name.size(); ++i) { + ggml_cuda_free_data(tensors_by_name[i].second); + } + ggml_cuda_free_scratch(); } - ggml_cuda_free_scratch(); -#elif defined(GGML_USE_CLBLAST) +#endif + +#if defined(GGML_USE_CLBLAST) for (size_t i = 0; i < tensors_by_name.size(); ++i) { ggml_cl_free_data(tensors_by_name[i].second); } @@ -1354,6 +1511,7 @@ static bool llama_kv_cache_init( cache.head = 0; cache.size = n_ctx; + cache.used = 0; cache.cells.clear(); cache.cells.resize(n_ctx); @@ -1379,23 +1537,26 @@ static bool llama_kv_cache_init( ggml_set_name(cache.v, "cache_v"); (void) n_gpu_layers; -#ifdef GGML_USE_CUBLAS - size_t vram_kv_cache = 0; - if (n_gpu_layers > (int)n_layer + 1) { - ggml_cuda_assign_buffers_no_scratch(cache.v); - LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.v); +#ifdef GGML_USE_CUBLAS + if (ggml_cublas_loaded()) { + size_t vram_kv_cache = 0; + + if (n_gpu_layers > (int)n_layer + 1) { + ggml_cuda_assign_buffers_no_scratch(cache.v); + LLAMA_LOG_INFO("%s: offloading v cache to GPU\n", __func__); + vram_kv_cache += ggml_nbytes(cache.v); + } + if (n_gpu_layers > (int)n_layer + 2) { + ggml_cuda_assign_buffers_no_scratch(cache.k); + LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); + vram_kv_cache += ggml_nbytes(cache.k); + } + if (vram_kv_cache > 0) { + LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MiB\n", __func__, vram_kv_cache / 1024.0 / 1024.0); + } } - if (n_gpu_layers > (int)n_layer + 2) { - ggml_cuda_assign_buffers_no_scratch(cache.k); - LLAMA_LOG_INFO("%s: offloading k cache to GPU\n", __func__); - vram_kv_cache += ggml_nbytes(cache.k); - } - if (vram_kv_cache > 0) { - LLAMA_LOG_INFO("%s: VRAM kv self = %.2f MB\n", __func__, vram_kv_cache / 1024.0 / 1024.0); - } -#endif // GGML_USE_CUBLAS +#endif return true; } @@ -1452,6 +1613,8 @@ static bool llama_kv_cache_find_slot( } } + cache.used += n_tokens; + return true; } @@ -1472,6 +1635,7 @@ static void llama_kv_cache_clear(struct llama_kv_cache & cache) { cache.cells[i].seq_id.clear(); } cache.head = 0; + cache.used = 0; } static void llama_kv_cache_seq_rm( @@ -1494,6 +1658,9 @@ static void llama_kv_cache_seq_rm( continue; } if (cache.cells[i].seq_id.empty()) { + // keep count of the number of used cells + if (cache.cells[i].pos >= 0) cache.used--; + cache.cells[i].pos = -1; if (new_head == cache.size) new_head = i; } @@ -1501,7 +1668,7 @@ static void llama_kv_cache_seq_rm( } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size) cache.head = new_head; + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; } static void llama_kv_cache_seq_cp( @@ -1527,6 +1694,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id for (uint32_t i = 0; i < cache.size; ++i) { if (!cache.cells[i].has_seq_id(seq_id)) { + if (cache.cells[i].pos >= 0) cache.used--; cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; @@ -1537,7 +1705,7 @@ static void llama_kv_cache_seq_keep(struct llama_kv_cache & cache, llama_seq_id } // If we freed up a slot, set head to it so searching can start there. - if (new_head != cache.size) cache.head = new_head; + if (new_head != cache.size && new_head < cache.head) cache.head = new_head; } static void llama_kv_cache_seq_shift( @@ -1558,6 +1726,7 @@ static void llama_kv_cache_seq_shift( cache.cells[i].delta += delta; if (cache.cells[i].pos < 0) { + if (!cache.cells[i].seq_id.empty()) cache.used--; cache.cells[i].pos = -1; cache.cells[i].seq_id.clear(); if (new_head == cache.size) new_head = i; @@ -1689,10 +1858,10 @@ struct llama_model_loader { case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break; case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K; break; default: - { - LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); - ftype = LLAMA_FTYPE_ALL_F32; - } break; + { + LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max)); + ftype = LLAMA_FTYPE_ALL_F32; + } break; } // this is a way to mark that we have "guessed" the file type @@ -1706,10 +1875,21 @@ struct llama_model_loader { } for (int i = 0; i < n_kv; i++) { - const char * name = gguf_get_key(ctx_gguf, i); - const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const char * name = gguf_get_key(ctx_gguf, i); + const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i); + const std::string type_name = + type == GGUF_TYPE_ARRAY + ? format("%s[%s,%d]", gguf_type_name(type), gguf_type_name(gguf_get_arr_type(ctx_gguf, i)), gguf_get_arr_n(ctx_gguf, i)) + : gguf_type_name(type); - LLAMA_LOG_INFO("%s: - kv %3d: %42s %-8s\n", __func__, i, name, gguf_type_name(type)); + std::string value = gguf_kv_to_str(ctx_gguf, i); + const size_t MAX_VALUE_LEN = 40; + if (value.size() > MAX_VALUE_LEN) { + value = format("%s...", value.substr(0, MAX_VALUE_LEN - 3).c_str()); + } + replace_all(value, "\n", "\\n"); + + LLAMA_LOG_INFO("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str()); } // print type counts @@ -1798,6 +1978,12 @@ struct llama_model_loader { throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str())); } + if (backend == GGML_BACKEND_GPU_SPLIT) { + if (ne.size() == 1) { + throw std::runtime_error(format("%s: 1-dimensional tensor '%s' cannot be split on the GPU", __func__, name.c_str())); + } + } + { bool is_ok = true; for (size_t i = 0; i < ne.size(); ++i) { @@ -1998,6 +2184,17 @@ static void llm_load_hparams( auto & hparams = model.hparams; + // get metadata as string + for (int i = 0; i < gguf_get_n_kv(ctx); i++) { + enum gguf_type type = gguf_get_kv_type(ctx, i); + if (type == GGUF_TYPE_ARRAY) { + continue; + } + const char * name = gguf_get_key(ctx, i); + const std::string value = gguf_kv_to_str(ctx, i); + model.gguf_kv.emplace(name, value); + } + // get general kv GGUF_GET_KEY(ctx, model.name, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_GENERAL_NAME)); @@ -2013,14 +2210,30 @@ static void llm_load_hparams( hparams.n_head_kv = hparams.n_head; GGUF_GET_KEY(ctx, hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, kv(LLM_KV_ATTENTION_HEAD_COUNT_KV)); + hparams.rope_finetuned = false; + GGUF_GET_KEY(ctx, hparams.rope_finetuned, gguf_get_val_bool, GGUF_TYPE_BOOL, false, + kv(LLM_KV_ROPE_SCALING_FINETUNED)); + + hparams.n_yarn_orig_ctx = hparams.n_ctx_train; + GGUF_GET_KEY(ctx, hparams.n_yarn_orig_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, false, + kv(LLM_KV_ROPE_SCALING_ORIG_CTX_LEN)); + // rope_freq_base (optional) hparams.rope_freq_base_train = 10000.0f; GGUF_GET_KEY(ctx, hparams.rope_freq_base_train, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE)); + std::string rope_scaling("linear"); + GGUF_GET_KEY(ctx, rope_scaling, gguf_get_val_str, GGUF_TYPE_STRING, false, kv(LLM_KV_ROPE_SCALING_TYPE)); + hparams.rope_scaling_type_train = llama_rope_scaling_type_from_string(rope_scaling); + GGML_ASSERT(hparams.rope_scaling_type_train != LLAMA_ROPE_SCALING_UNSPECIFIED); + // rope_freq_scale (inverse of the kv) is optional - float ropescale = 1.0f; - GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); - hparams.rope_freq_scale_train = 1.0f/ropescale; + float ropescale = 0.0f; + GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALING_FACTOR)); + if (ropescale == 0.0f) { // try the old key name + GGUF_GET_KEY(ctx, ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR)); + } + hparams.rope_freq_scale_train = ropescale == 0.0f ? 1.0f : 1.0f/ropescale; // sanity check for n_rot (optional) { @@ -2126,6 +2339,16 @@ static void llm_load_hparams( default: model.type = e_model::MODEL_UNKNOWN; } } break; + case LLM_ARCH_STABLELM: + { + GGUF_GET_KEY(ctx, hparams.f_norm_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, kv(LLM_KV_ATTENTION_LAYERNORM_EPS)); + + switch (hparams.n_layer) { + case 32: model.type = e_model::MODEL_3B; break; + default: model.type = e_model::MODEL_UNKNOWN; + } + } break; + default: (void)0; } @@ -2267,6 +2490,23 @@ static void llm_load_vocab( __func__, key.c_str(), id, old_id); id = old_id; } + + } + + // Handle add_bos_token and add_eos_token + std::string key = kv(LLM_KV_TOKENIZER_ADD_BOS); + int kid = gguf_find_key(ctx, key.c_str()); + enum gguf_type ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid); + vocab.special_add_bos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1; + if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) { + LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str()); + } + key = kv(LLM_KV_TOKENIZER_ADD_EOS); + kid = gguf_find_key(ctx, key.c_str()); + ktype = kid < 0 ? GGUF_TYPE_COUNT : gguf_get_kv_type(ctx, kid); + vocab.special_add_eos = ktype == GGUF_TYPE_BOOL ? gguf_get_val_bool(ctx, kid) : -1; + if (ktype != GGUF_TYPE_BOOL && ktype != GGUF_TYPE_COUNT) { + LLAMA_LOG_WARN("%s: bad field type %d for '%s' - ignoring\n", __func__, ktype, key.c_str()); } } @@ -2370,6 +2610,8 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { const auto & hparams = model.hparams; const auto & vocab = model.vocab; + const auto rope_scaling_type = LLAMA_ROPE_SCALING_TYPES.at(hparams.rope_scaling_type_train); + // hparams LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml.fver)); LLAMA_LOG_INFO("%s: arch = %s\n", __func__, LLM_ARCH_NAMES.at(model.arch).c_str()); @@ -2388,13 +2630,16 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) { LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv); LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias); LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff); + LLAMA_LOG_INFO("%s: rope scaling = %s\n", __func__, rope_scaling_type.c_str()); LLAMA_LOG_INFO("%s: freq_base_train = %.1f\n", __func__, hparams.rope_freq_base_train); LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); + LLAMA_LOG_INFO("%s: n_yarn_orig_ctx = %u\n", __func__, hparams.n_yarn_orig_ctx); + LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); LLAMA_LOG_INFO("%s: model type = %s\n", __func__, llama_model_type_name(model.type)); LLAMA_LOG_INFO("%s: model ftype = %s\n", __func__, llama_model_ftype_name(model.ftype).c_str()); LLAMA_LOG_INFO("%s: model params = %.2f B\n", __func__, ml.n_elements*1e-9); - if (ml.n_bytes < GB) { - LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); + if (ml.n_bytes < GiB) { + LLAMA_LOG_INFO("%s: model size = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } else { LLAMA_LOG_INFO("%s: model size = %.2f GiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0/1024.0, ml.n_bytes*8.0/ml.n_elements); } @@ -2432,7 +2677,7 @@ static void llm_load_tensors( ml.calc_sizes(ctx_size, mmapped_size); - LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: ggml ctx size = %7.2f MiB\n", __func__, ctx_size/1024.0/1024.0); // create the ggml context { @@ -2455,18 +2700,22 @@ static void llm_load_tensors( } (void) main_gpu; + + enum ggml_backend_type llama_backend_offload = GGML_BACKEND_CPU; + enum ggml_backend_type llama_backend_offload_split = GGML_BACKEND_CPU; + #ifdef GGML_USE_CUBLAS - LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); - ggml_cuda_set_main_device(main_gpu); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU_SPLIT + if (ggml_cublas_loaded()) { + LLAMA_LOG_INFO("%s: using " GGML_CUDA_NAME " for GPU acceleration\n", __func__); + ggml_cuda_set_main_device(main_gpu); + + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU_SPLIT; + } #elif defined(GGML_USE_CLBLAST) - LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_GPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_GPU -#else -#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CPU -#define LLAMA_BACKEND_OFFLOAD_SPLIT GGML_BACKEND_CPU + LLAMA_LOG_INFO("%s: using OpenCL for GPU acceleration\n", __func__); + llama_backend_offload = GGML_BACKEND_GPU; + llama_backend_offload_split = GGML_BACKEND_GPU; #endif // prepare memory for the weights @@ -2482,7 +2731,7 @@ static void llm_load_tensors( case LLM_ARCH_LLAMA: case LLM_ARCH_REFACT: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2493,12 +2742,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2522,8 +2771,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2536,21 +2785,21 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; case LLM_ARCH_BAICHUAN: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); { ggml_backend_type backend_norm; ggml_backend_type backend_output; @@ -2559,12 +2808,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2588,8 +2837,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2602,15 +2851,15 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w1 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += - ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + - ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w1) + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; @@ -2618,7 +2867,7 @@ static void llm_load_tensors( { // TODO: CPU-only for now - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2629,12 +2878,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2660,8 +2909,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2681,21 +2930,21 @@ static void llm_load_tensors( layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.attn_norm_b) + ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3); + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); } } } break; case LLM_ARCH_STARCODER: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.pos_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU); // output { @@ -2706,12 +2955,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2737,8 +2986,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2754,11 +3003,11 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2766,29 +3015,36 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.b2) + - ggml_nbytes(layer.w3) + ggml_nbytes(layer.b3); + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b) + + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b); } } } break; case LLM_ARCH_PERSIMMON: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); { ggml_backend_type backend_norm; ggml_backend_type backend_output; if (n_gpu_layers > int(n_layer)) { +#ifdef GGML_USE_CUBLAS + if (n_gpu_layers > int(n_layer + 1)) { + LLAMA_LOG_ERROR("%s: CUDA backend missing Persimmon CUDA ops, can offload at most %ld layers. See: https://github.com/ggerganov/llama.cpp/issues/4038\n", + __func__, n_layer + 1); + throw std::runtime_error("Persimmon CUDA offload failed"); + } +#endif // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2811,34 +3067,34 @@ static void llm_load_tensors( const int i_gpu_start = n_layer - n_gpu_layers; model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; auto & layer = model.layers[i]; - layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); - layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); - layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); - layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); - layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); layer.attn_q_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend); - layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); + layer.attn_q_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i), {64}, backend); layer.attn_k_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {64}, backend); - layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); + layer.attn_k_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), {64}, backend); } } break; case LLM_ARCH_BLOOM: { // TODO: CPU-only for now - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); - model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); - model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight"), {n_embd}, GGML_BACKEND_CPU); + model.tok_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias"), {n_embd}, GGML_BACKEND_CPU); // output { @@ -2849,12 +3105,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2880,8 +3136,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2889,19 +3145,19 @@ static void llm_load_tensors( layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split); - layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend_split); + layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend); layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); - layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend_split); + layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend); layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); - layer.b2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split); + layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); - layer.b3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2909,14 +3165,14 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.bqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.bo) + ggml_nbytes(layer.ffn_norm) + ggml_nbytes(layer.ffn_norm_b) + - ggml_nbytes(layer.w3) + ggml_nbytes(layer.b3) + - ggml_nbytes(layer.w2) + ggml_nbytes(layer.b2); + ggml_nbytes(layer.ffn_up) + ggml_nbytes(layer.ffn_up_b) + + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_down_b); } } } break; case LLM_ARCH_MPT: { - model.tok_embeddings = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); // output { @@ -2927,12 +3183,12 @@ static void llm_load_tensors( // norm is not performance relevant on its own but keeping it in VRAM reduces data copying // on Windows however this is detrimental unless everything is on the GPU #ifndef _WIN32 - backend_norm = LLAMA_BACKEND_OFFLOAD; + backend_norm = llama_backend_offload; #else - backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; #endif // _WIN32 - backend_output = LLAMA_BACKEND_OFFLOAD_SPLIT; + backend_output = llama_backend_offload_split; } else { backend_norm = GGML_BACKEND_CPU; backend_output = GGML_BACKEND_CPU; @@ -2956,8 +3212,8 @@ static void llm_load_tensors( model.layers.resize(n_layer); for (uint32_t i = 0; i < n_layer; ++i) { - const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD; // NOLINT - const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD_SPLIT; // NOLINT + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT auto & layer = model.layers[i]; @@ -2967,8 +3223,8 @@ static void llm_load_tensors( layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); - layer.w2 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); - layer.w3 = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); if (backend == GGML_BACKEND_GPU) { vram_weights += @@ -2976,11 +3232,86 @@ static void llm_load_tensors( ggml_nbytes(layer.wqkv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + - ggml_nbytes(layer.w2) + - ggml_nbytes(layer.w3); + ggml_nbytes(layer.ffn_down) + + ggml_nbytes(layer.ffn_up); } } } break; + case LLM_ARCH_STABLELM: + { + model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU); + + // output + { + ggml_backend_type backend_norm; + ggml_backend_type backend_output; + + if (n_gpu_layers > int(n_layer)) { + // norm is not performance relevant on its own but keeping it in VRAM reduces data copying + // on Windows however this is detrimental unless everything is on the GPU +#ifndef _WIN32 + backend_norm = llama_backend_offload; +#else + backend_norm = n_gpu_layers <= (int) n_layer + 2 ? GGML_BACKEND_CPU : llama_backend_offload; +#endif // _WIN32 + + backend_output = llama_backend_offload_split; + } else { + backend_norm = GGML_BACKEND_CPU; + backend_output = GGML_BACKEND_CPU; + } + + model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm); + model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm); + model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output); + + if (backend_norm == GGML_BACKEND_GPU) { + vram_weights += ggml_nbytes(model.output_norm); + } + if (backend_output == GGML_BACKEND_GPU_SPLIT) { + vram_weights += ggml_nbytes(model.output); + } + } + + const uint32_t n_ff = hparams.n_ff; + + const int i_gpu_start = n_layer - n_gpu_layers; + + model.layers.resize(n_layer); + + for (uint32_t i = 0; i < n_layer; ++i) { + /* + llama_model_loader: - tensor 4: blk.0.attn_output.weight f16 [ 2560, 2560, 1, 1 ] + */ + const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT + const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT + + auto & layer = model.layers[i]; + + layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend); + layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend); + + layer.wq = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, backend_split); + layer.wk = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, backend_split); + layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split); + + layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend); + layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend); + + layer.ffn_gate = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, backend_split); + layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, backend_split); + layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split); + + if (backend == GGML_BACKEND_GPU) { + vram_weights += + ggml_nbytes(layer.attn_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk) + + ggml_nbytes(layer.wv) + ggml_nbytes(layer.wo) + ggml_nbytes(layer.ffn_norm) + + ggml_nbytes(layer.ffn_gate) + ggml_nbytes(layer.ffn_down) + ggml_nbytes(layer.ffn_up); + } + } + } break; + default: throw std::runtime_error("unknown architecture"); } @@ -2995,7 +3326,7 @@ static void llm_load_tensors( ctx_size + mmapped_size - vram_weights; // weights in VRAM not in memory - LLAMA_LOG_INFO("%s: mem required = %7.2f MB\n", __func__, mem_required / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: mem required = %7.2f MiB\n", __func__, mem_required / 1024.0 / 1024.0); #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer)); @@ -3007,14 +3338,14 @@ static void llm_load_tensors( #ifdef GGML_USE_CUBLAS const int max_backend_supported_layers = hparams.n_layer + 3; - const int max_offloadable_layers = hparams.n_layer + 3; -#elif defined(GGML_USE_CLBLAST) + const int max_offloadable_layers = hparams.n_layer + 3; +#elif GGML_USE_CLBLAST const int max_backend_supported_layers = hparams.n_layer + 1; - const int max_offloadable_layers = hparams.n_layer + 1; + const int max_offloadable_layers = hparams.n_layer + 1; #endif // GGML_USE_CUBLAS LLAMA_LOG_INFO("%s: offloaded %d/%d layers to GPU\n", __func__, std::min(n_gpu_layers, max_offloadable_layers), max_backend_supported_layers); - LLAMA_LOG_INFO("%s: VRAM used: %.2f MB\n", __func__, vram_weights / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: VRAM used: %.2f MiB\n", __func__, vram_weights / 1024.0 / 1024.0); #else (void) n_gpu_layers; #endif // defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) @@ -3046,21 +3377,11 @@ static void llm_load_tensors( model.t_load_us = ggml_time_us() - model.t_start_us; } -static bool llama_model_load( - const std::string & fname, - llama_model & model, - int n_gpu_layers, - int main_gpu, - const float * tensor_split, - bool use_mmap, - bool use_mlock, - bool vocab_only, - llama_progress_callback progress_callback, - void *progress_callback_user_data) { +static bool llama_model_load(const std::string & fname, llama_model & model, const llama_model_params & params) { try { - llama_model_loader ml(fname, use_mmap); + llama_model_loader ml(fname, params.use_mmap); - model.hparams.vocab_only = vocab_only; + model.hparams.vocab_only = params.vocab_only; llm_load_arch (ml, model); llm_load_hparams(ml, model); @@ -3072,15 +3393,15 @@ static bool llama_model_load( throw std::runtime_error("vocab size mismatch"); } - if (vocab_only) { + if (params.vocab_only) { LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__); return true; } llm_load_tensors( - ml, model, n_gpu_layers, - main_gpu, tensor_split, - use_mlock, progress_callback, progress_callback_user_data); + ml, model, params.n_gpu_layers, params.main_gpu, params.tensor_split, params.use_mlock, + params.progress_callback, params.progress_callback_user_data + ); } catch (const std::exception & err) { LLAMA_LOG_ERROR("error loading model: %s\n", err.what()); return false; @@ -3089,2818 +3410,1999 @@ static bool llama_model_load( return true; } -static struct ggml_cgraph * llm_build_llama( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; +// +// llm_build +// - const auto & kv_self = lctx.kv_self; +using llm_build_cb = std::function; - GGML_ASSERT(!!kv_self.ctx); +enum llm_rope_type { + LLM_ROPE, + LLM_ROPE_NEOX, + LLM_ROPE_GLM, +}; - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); +enum llm_ffn_op_type { + LLM_FFN_SILU, + LLM_FFN_GELU, + LLM_FFN_RELU, + LLM_FFN_RELU_SQR, +}; - GGML_ASSERT(n_embd_head == hparams.n_rot); +enum llm_ffn_gate_type { + LLM_FFN_SEQ, + LLM_FFN_PAR, // ffn_gate is parallel to ffn_up +}; - const float freq_base = cparams.rope_freq_base; - const float freq_scale = cparams.rope_freq_scale; - const float norm_rms_eps = hparams.f_norm_rms_eps; +enum llm_norm_type { + LLM_NORM, + LLM_NORM_RMS, +}; - const int n_gpu_layers = model.n_gpu_layers; +static struct ggml_tensor * llm_build_inp_embd( + struct ggml_context * ctx, + const llama_hparams & hparams, + const llama_batch & batch, + struct ggml_tensor * tok_embd, + const llm_build_cb & cb) { + const int64_t n_embd = hparams.n_embd; - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; - - //printf("n_kv = %d\n", n_kv); - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; struct ggml_tensor * inpL; if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, batch.n_tokens); + cb(inp_tokens, "inp_tokens", -1); - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); + inpL = ggml_get_rows(ctx, tok_embd, inp_tokens); } else { #ifdef GGML_USE_MPI GGML_ASSERT(false && "not implemented"); #endif - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } + inpL = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, batch.n_tokens); } - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; + return inpL; +} - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; +// Persimmon: n_rot = n_embd_head/2 +// Other: n_rot = n_embd_head +static void llm_build_k_shift( + struct ggml_context * ctx, + const llama_hparams & hparams, + const llama_cparams & cparams, + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + llm_rope_type type, + int64_t n_ctx, + int n_rot, + float freq_base, + float freq_scale, + const llm_build_cb & cb) { + const int64_t n_layer = hparams.n_layer; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_gqa = hparams.n_embd_gqa(); + const int64_t n_embd_head = hparams.n_embd_head(); + const int32_t n_orig_ctx = cparams.n_yarn_orig_ctx; + const float ext_factor = cparams.yarn_ext_factor; + const float attn_factor = cparams.yarn_attn_factor; + const float beta_fast = cparams.yarn_beta_fast; + const float beta_slow = cparams.yarn_beta_slow; -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + GGML_ASSERT(n_embd_head % n_rot == 0); - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } + struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx); + cb(K_shift, "K_shift", -1); - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); + int rope_type = 0; - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } - - // shift the entire K-cache if needed - if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } + switch (type) { + case LLM_ROPE: rope_type = 0; break; + case LLM_ROPE_NEOX: rope_type = 2; break; + case LLM_ROPE_GLM: rope_type = 4; break; } for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); + struct ggml_tensor * tmp = + // we rotate only the first n_rot dimensions + ggml_rope_custom_inplace(ctx, + ggml_view_3d(ctx, kv.k, + n_embd_head, n_head_kv, n_ctx, + ggml_element_size(kv.k)*n_embd_head, + ggml_element_size(kv.k)*n_embd_gqa, + ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il), + K_shift, n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(tmp, "K_shifted", il); + ggml_build_forward_expand(graph, tmp); + } +} - offload_func_t offload_func = llama_nop; +static void llm_build_kv_store( + struct ggml_context * ctx, + const llama_hparams & hparams, + const llama_kv_cache & kv, + struct ggml_cgraph * graph, + struct ggml_tensor * k_cur, + struct ggml_tensor * v_cur, + int64_t n_ctx, + int32_t n_tokens, + int32_t kv_head, + const llm_build_cb & cb, + int64_t il) { + const int64_t n_embd_gqa = hparams.n_embd_gqa(); -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; + // compute the transposed [n_tokens, n_embd] V matrix + struct ggml_tensor * v_cur_t = ggml_transpose(ctx, ggml_reshape_2d(ctx, v_cur, n_embd_gqa, n_tokens)); + //struct ggml_tensor * v_cur_t = ggml_transpose(ctx, v_cur); // TODO: reshape above is likely not needed + cb(v_cur_t, "v_cur_t", il); + + struct ggml_tensor * k_cache_view = ggml_view_1d(ctx, kv.k, n_tokens*n_embd_gqa, + (ggml_element_size(kv.k)*n_embd_gqa)*(il*n_ctx + kv_head)); + cb(k_cache_view, "k_cache_view", il); + + struct ggml_tensor * v_cache_view = ggml_view_2d(ctx, kv.v, n_tokens, n_embd_gqa, + ( n_ctx)*ggml_element_size(kv.v), + (il*n_ctx)*ggml_element_size(kv.v)*n_embd_gqa + kv_head*ggml_element_size(kv.v)); + cb(v_cache_view, "v_cache_view", il); + + // important: storing RoPE-ed version of K in the KV cache! + ggml_build_forward_expand(graph, ggml_cpy(ctx, k_cur, k_cache_view)); + ggml_build_forward_expand(graph, ggml_cpy(ctx, v_cur_t, v_cache_view)); +} + +static struct ggml_tensor * llm_build_norm( + struct ggml_context * ctx, + struct ggml_tensor * cur, + const llama_hparams & hparams, + struct ggml_tensor * mw, + struct ggml_tensor * mb, + llm_norm_type type, + const llm_build_cb & cb, + int il) { + switch (type) { + case LLM_NORM: cur = ggml_norm (ctx, cur, hparams.f_norm_eps); break; + case LLM_NORM_RMS: cur = ggml_rms_norm(ctx, cur, hparams.f_norm_rms_eps); break; + } + + if (mw || mb) { + cb(cur, "norm", il); + } + + if (mw) { + cur = ggml_mul(ctx, cur, mw); + if (mb) { + cb(cur, "norm_w", il); } -#endif // GGML_USE_CUBLAS + } - struct ggml_tensor * inpSA = inpL; + if (mb) { + cur = ggml_add(ctx, cur, mb); + } - // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); + return cur; +} - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); +static struct ggml_tensor * llm_build_ffn( + struct ggml_context * ctx, + struct ggml_tensor * cur, + struct ggml_tensor * up, + struct ggml_tensor * up_b, + struct ggml_tensor * gate, + struct ggml_tensor * gate_b, + struct ggml_tensor * down, + struct ggml_tensor * down_b, + llm_ffn_op_type type_op, + llm_ffn_gate_type type_gate, + const llm_build_cb & cb, + int il) { + struct ggml_tensor * tmp = ggml_mul_mat(ctx, up, cur); + cb(tmp, "ffn_up", il); + + if (up_b) { + tmp = ggml_add(ctx, tmp, up_b); + cb(tmp, "ffn_up_b", il); + } + + if (gate) { + switch (type_gate) { + case LLM_FFN_SEQ: + { + cur = ggml_mul_mat(ctx, gate, tmp); + cb(cur, "ffn_gate", il); + } break; + case LLM_FFN_PAR: + { + cur = ggml_mul_mat(ctx, gate, cur); + cb(cur, "ffn_gate", il); + } break; } - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); + if (gate_b) { + cur = ggml_add(ctx, cur, gate_b); + cb(cur, "ffn_gate_b", il); + } + } else { + cur = tmp; + } - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); - - struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); - - struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); - - // store key and value to memory + switch (type_op) { + case LLM_FFN_SILU: { - // compute the transposed [n_tokens, n_embd] V matrix + cur = ggml_silu(ctx, cur); + cb(cur, "ffn_silu", il); + } break; + case LLM_FFN_GELU: + { + cur = ggml_gelu(ctx, cur); + cb(cur, "ffn_gelu", il); + } break; + case LLM_FFN_RELU: + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); + } break; + case LLM_FFN_RELU_SQR: + { + cur = ggml_relu(ctx, cur); + cb(cur, "ffn_relu", il); - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); + cur = ggml_sqr(ctx, cur); + cb(cur, "ffn_sqr(relu)", il); + } break; + } - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); + if (type_gate == LLM_FFN_PAR) { + cur = ggml_mul(ctx, cur, tmp); + cb(cur, "ffn_gate_par", il); + } - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); + cur = ggml_mul_mat(ctx, down, cur); + if (down_b) { + cb(cur, "ffn_down", il); + } - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); + if (down_b) { + cur = ggml_add(ctx, cur, down_b); + } - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } + return cur; +} - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); +// if max_alibi_bias > 0 then apply ALiBi +static struct ggml_tensor * llm_build_kqv( + struct ggml_context * ctx, + const llama_hparams & hparams, + const llama_kv_cache & kv, + struct ggml_tensor * wo, + struct ggml_tensor * wo_b, + struct ggml_tensor * q_cur, + struct ggml_tensor * kq_scale, + struct ggml_tensor * kq_mask, + int64_t n_ctx, + int32_t n_tokens, + int32_t n_kv, + float max_alibi_bias, + const llm_build_cb & cb, + int il) { + const int64_t n_embd = hparams.n_embd; + const int64_t n_head = hparams.n_head; + const int64_t n_head_kv = hparams.n_head_kv; + const int64_t n_embd_head = hparams.n_embd_head(); + const int64_t n_embd_gqa = hparams.n_embd_gqa(); - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); + struct ggml_tensor * q = ggml_permute(ctx, q_cur, 0, 2, 1, 3); + cb(q, "q", il); - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); + struct ggml_tensor * k = + ggml_view_3d(ctx, kv.k, + n_embd_head, n_kv, n_head_kv, + ggml_element_size(kv.k)*n_embd_gqa, + ggml_element_size(kv.k)*n_embd_head, + ggml_element_size(kv.k)*n_embd_gqa*n_ctx*il); + cb(k, "k", il); - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_kv, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); + struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q); + cb(kq, "kq", il); - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); + kq = ggml_scale(ctx, kq, kq_scale); + cb(kq, "kq_scaled", il); - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); + if (max_alibi_bias > 0.0f) { + // TODO: n_head or n_head_kv + // TODO: K-shift is likely not working + // TODO: change to ggml_add + kq = ggml_alibi(ctx, kq, /*n_past*/ 0, n_head, max_alibi_bias); + cb(kq, "kq_scaled_alibi", il); + } - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); + kq = ggml_add(ctx, kq, kq_mask); + cb(kq, "kq_masked", il); -#if 1 - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); -#else - // make V contiguous in memory to speed up the matmul, however we waste time on the copy - // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation - // is there a better way? - struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head)); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); -#endif + kq = ggml_soft_max(ctx, kq); + cb(kq, "kq_soft_max", il); - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); + // split cached v into n_head heads + struct ggml_tensor * v = + ggml_view_3d(ctx, kv.v, + n_kv, n_embd_head, n_head_kv, + ggml_element_size(kv.v)*n_ctx, + ggml_element_size(kv.v)*n_ctx*n_embd_head, + ggml_element_size(kv.v)*n_ctx*n_embd_gqa*il); + cb(v, "v", il); - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); + struct ggml_tensor * kqv = ggml_mul_mat(ctx, v, kq); + cb(kqv, "kqv", il); - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + struct ggml_tensor * kqv_merged = ggml_permute(ctx, kqv, 0, 2, 1, 3); + cb(kqv_merged, "kqv_merged", il); + + struct ggml_tensor * cur = ggml_cont_2d(ctx, kqv_merged, n_embd, n_tokens); + cb(cur, "kqv_merged_cont", il); + + cur = ggml_mul_mat(ctx, wo, cur); + if (wo_b) { + cb(cur, "kqv_wo", il); + } + + if (wo_b) { + cur = ggml_add(ctx, cur, wo_b); + } + + return cur; +} + +struct llm_build_context { + const llama_model & model; + const llama_hparams & hparams; + const llama_cparams & cparams; + const llama_batch & batch; + const llama_kv_cache & kv_self; + + const int64_t n_embd; + const int64_t n_layer; + const int64_t n_ctx; // user-specified context size (can be different from n_ctx_train) + const int64_t n_head; + const int64_t n_head_kv; + const int64_t n_embd_head; + const int64_t n_embd_gqa; + + const float freq_base; + const float freq_scale; + const float ext_factor; + const float attn_factor; + const float beta_fast; + const float beta_slow; + const float norm_eps; + const float norm_rms_eps; + + const int32_t n_tokens; + const int32_t n_kv; // size of KV cache to consider (n_kv <= n_ctx) + const int32_t kv_head; // index of where we store new KV data in the cache + const int32_t n_orig_ctx; + + const bool do_rope_shift; + + const llm_build_cb & cb; + + llama_buffer & buf_compute; + + struct ggml_context * ctx0 = nullptr; + + // TODO: consider making the entire interface noexcept + llm_build_context( + llama_context & lctx, + const llama_batch & batch, + const llm_build_cb & cb, + bool worst_case) : + model (lctx.model), + hparams (model.hparams), + cparams (lctx.cparams), + batch (batch), + kv_self (lctx.kv_self), + n_embd (hparams.n_embd), + n_layer (hparams.n_layer), + n_ctx (cparams.n_ctx), + n_head (hparams.n_head), + n_head_kv (hparams.n_head_kv), + n_embd_head (hparams.n_embd_head()), + n_embd_gqa (hparams.n_embd_gqa()), + freq_base (cparams.rope_freq_base), + freq_scale (cparams.rope_freq_scale), + ext_factor (cparams.yarn_ext_factor), + attn_factor (cparams.yarn_attn_factor), + beta_fast (cparams.yarn_beta_fast), + beta_slow (cparams.yarn_beta_slow), + norm_eps (hparams.f_norm_eps), + norm_rms_eps (hparams.f_norm_rms_eps), + n_tokens (batch.n_tokens), + n_kv (worst_case ? n_ctx : kv_self.n), + kv_head (worst_case ? n_ctx - n_tokens : kv_self.head), + n_orig_ctx (cparams.n_yarn_orig_ctx), + do_rope_shift (worst_case || kv_self.has_shift), + cb (cb), + buf_compute (lctx.buf_compute) { + GGML_ASSERT(!!kv_self.ctx); + + // all initializations should be done in init() } - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); + void init() { + struct ggml_init_params params = { + /*.mem_size =*/ buf_compute.size, + /*.mem_buffer =*/ buf_compute.data, + /*.no_alloc =*/ true, + }; + + ctx0 = ggml_init(params); + } + + void free() { + if (ctx0) { + ggml_free(ctx0); + ctx0 = nullptr; + } + } + + struct ggml_cgraph * build_llama() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + GGML_ASSERT(n_embd_head == hparams.n_rot); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; - // feed-forward network - { // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); - } - - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_baichaun( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_rot); - - const float freq_base = cparams.rope_freq_base; - const float freq_scale = cparams.rope_freq_scale; - const float norm_rms_eps = hparams.f_norm_rms_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } - - // shift the entire K-cache if needed - if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 0, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } - } - - for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - struct ggml_tensor * inpSA = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); - - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); - } - - // self-attention - { - // compute Q and K and RoPE them - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); - - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); - - struct ggml_tensor * Kcur; - struct ggml_tensor * Qcur; - switch (model.type) { - case MODEL_7B: - Kcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - Qcur = ggml_rope_custom(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens), KQ_pos, n_embd_head, 0, 0, freq_base, freq_scale); - break; - case MODEL_13B: - Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, n_tokens); - Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, n_tokens); - break; - default: - GGML_ASSERT(false); - } - - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); - - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); - - // store key and value to memory - { - // compute the transposed [n_tokens, n_embd] V matrix - - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); - - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked; - struct ggml_tensor * KQ_scaled_alibi; - - switch (model.type) { - case MODEL_7B: - KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - break; - case MODEL_13B: - // TODO: replace with ggml_add() - KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - break; - default: - GGML_ASSERT(false); - } - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); - } - - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); - - // feed-forward network - { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); - - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); - } - - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_refact( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - const float norm_rms_eps = hparams.f_norm_rms_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - // printf("n_kv = %d\n", n_kv); - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - for (int il = 0; il < n_layer; ++il) { - ggml_format_name(inpL, "layer_inp_%d", il); - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - struct ggml_tensor * inpSA = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, inpL, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_0"); - - // cur = cur*attn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - ggml_set_name(cur, "attention_norm_0"); - } - - // self-attention - { - // compute Q and K - struct ggml_tensor * tmpk = ggml_mul_mat(ctx0, model.layers[il].wk, cur); - offload_func_kq(tmpk); - ggml_set_name(tmpk, "tmpk"); - - struct ggml_tensor * tmpq = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - offload_func_kq(tmpq); - ggml_set_name(tmpq, "tmpq"); - - struct ggml_tensor * Kcur = ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, n_tokens); - offload_func_kq(Kcur); - ggml_set_name(Kcur, "Kcur"); - - struct ggml_tensor * Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens); - offload_func_kq(Qcur); - ggml_set_name(Qcur, "Qcur"); - - // store key and value to memory - { - // compute the transposed [n_tokens, n_embd] V matrix - - struct ggml_tensor * tmpv = ggml_mul_mat(ctx0, model.layers[il].wv, cur); - offload_func_v(tmpv); - ggml_set_name(tmpv, "tmpv"); - - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_kv, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ 0, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - -#if 1 - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); -#else - // make V contiguous in memory to speed up the matmul, however we waste time on the copy - // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation - // is there a better way? - struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_ctx, n_embd_head, n_head)); - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max); -#endif - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - // projection (no bias) - cur = ggml_mul_mat(ctx0, - model.layers[il].wo, - cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); - } - - struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); - - // feed-forward network - { - // norm - { - cur = ggml_rms_norm(ctx0, inpFF, norm_rms_eps); - offload_func(cur); - ggml_set_name(cur, "rms_norm_1"); - - // cur = cur*ffn_norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - ggml_set_name(cur, "ffn_norm"); - } - - struct ggml_tensor * tmp = ggml_mul_mat(ctx0, - model.layers[il].w3, - cur); - offload_func(tmp); - ggml_set_name(tmp, "result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w1, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w1"); - - // SILU activation - cur = ggml_silu(ctx0, cur); - offload_func(cur); - ggml_set_name(cur, "silu"); - - cur = ggml_mul(ctx0, cur, tmp); - offload_func(cur); - ggml_set_name(cur, "silu_x_result_w3"); - - cur = ggml_mul_mat(ctx0, - model.layers[il].w2, - cur); - offload_func(cur); - ggml_set_name(cur, "result_w2"); - } - - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_result_w2"); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - { - cur = ggml_rms_norm(ctx0, cur, norm_rms_eps); - offload_func_nr(cur); - ggml_set_name(cur, "rms_norm_2"); - - // cur = cur*norm(broadcasted) - cur = ggml_mul(ctx0, cur, model.output_norm); - // offload_func_nr(cur); // TODO CPU + GPU mirrored backend - ggml_set_name(cur, "result_norm"); - } - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_falcon( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_rot); - - const float freq_base = cparams.rope_freq_base; - const float freq_scale = cparams.rope_freq_scale; - const float norm_eps = hparams.f_norm_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; - - //printf("kv_head = %d, n_kv = %d, n_tokens = %d, n_ctx = %d, is_measure = %d, has_shift = %d\n", - // kv_head, n_kv, n_tokens, n_ctx, ggml_allocr_is_measure(lctx.alloc), kv_self.has_shift); - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // KQ_pos - contains the positions - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } - - // shift the entire K-cache if needed - if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_head_kv, n_ctx, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il), - K_shift, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * attn_norm; - - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // self-attention - // TODO: refactor into common function (shared with LLaMA) - { - attn_norm = ggml_norm(ctx0, inpL, norm_eps); - offload_func(attn_norm); - - attn_norm = ggml_add(ctx0, - ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm), - model.layers[il].attn_norm_b); - offload_func(attn_norm->src[0]); - offload_func(attn_norm); - - if (model.layers[il].attn_norm_2) { // Falcon-40B - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func(cur); - - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.layers[il].attn_norm_2), - model.layers[il].attn_norm_2_b); - offload_func(cur->src[0]); - offload_func(cur); - } else { // Falcon 7B - cur = attn_norm; - } - - // compute QKV - - cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); - - // Note that the strides for Kcur, Vcur are set up so that the - // resulting views are misaligned with the tensor's storage - // (by applying the K/V offset we shift the tensor's original - // view to stick out behind the viewed QKV tensor's allocated - // memory, so to say). This is ok because no actual accesses - // happen to that out-of-range memory, but it can require some - // trickery when trying to accurately dump these views for - // debugging. - - const size_t wsize = ggml_type_size(cur->type); - - // TODO: these 2 ggml_conts are technically not needed, but we add them until CUDA support for - // non-contiguous views is added for the rope operator - struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_3d( - ctx0, cur, n_embd_head, n_head, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - 0)); - offload_func_kq(tmpq); - - struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * n_head)); - offload_func_kq(tmpk); - - struct ggml_tensor * tmpv = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * (n_head + n_head_kv)); - offload_func_v(tmpv); - - // using mode = 2 for neox mode - struct ggml_tensor * Qcur = ggml_rope_custom(ctx0, tmpq, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(Qcur); - struct ggml_tensor * Kcur = ggml_rope_custom(ctx0, tmpk, KQ_pos, n_embd_head, 2, 0, freq_base, freq_scale); - offload_func_kq(Kcur); - - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - offload_func_v(Vcur->src[0]->src[0]); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); - } - - struct ggml_tensor * attn_out = cur; - - // feed forward - { - struct ggml_tensor * inpFF = attn_norm; - - cur = ggml_mul_mat(ctx0, model.layers[il].w3, inpFF); - offload_func(cur); - - cur = ggml_gelu(ctx0, cur); - offload_func(cur); - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); - } - - cur = ggml_add(ctx0, cur, attn_out); - offload_func(cur); - cur = ggml_add(ctx0, cur, inpL); - offload_func(cur); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.output_norm), - model.output_norm_b); - ggml_set_name(cur, "result_norm"); - } - - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_starcoder( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_rot); - - const float norm_eps = hparams.f_norm_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * token; - struct ggml_tensor * position; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, token); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - { - // Compute position embeddings. - struct ggml_tensor * inp_positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - ggml_allocr_alloc(lctx.alloc, inp_positions); - if (!ggml_allocr_is_measure(lctx.alloc)) { - for (int i = 0; i < n_tokens; ++i) { - ((int32_t *) inp_positions->data)[i] = batch.pos[i]; - } - } - ggml_set_name(inp_positions, "inp_positions"); - - position = ggml_get_rows(ctx0, model.pos_embeddings, inp_positions); - } - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - ggml_set_name(KQ_mask, "KQ_mask"); - offload_func_kq(KQ_mask); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - inpL = ggml_add(ctx0, token, position); - ggml_set_name(inpL, "inpL"); - - for (int il = 0; il < n_layer; ++il) { - offload_func_t offload_func = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - { - // Norm - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func(cur); - - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); - offload_func(cur); - } - - { - // Self Attention - cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - offload_func_kq(cur); - - struct ggml_tensor * tmpq = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); - struct ggml_tensor * tmpk = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); - struct ggml_tensor * tmpv = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - - ggml_set_name(tmpq, "tmpq"); - ggml_set_name(tmpk, "tmpk"); - ggml_set_name(tmpv, "tmpv"); - - offload_func_kq(tmpq); - offload_func_kq(tmpk); - offload_func_v (tmpv); - - struct ggml_tensor * Qcur = ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, n_tokens); - struct ggml_tensor * Kcur = tmpk; - - { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, tmpv); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); - - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); - } - - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); - - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - } - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo); - offload_func(cur); - - // Add the input - cur = ggml_add(ctx0, cur, inpL); - offload_func(cur); - - struct ggml_tensor * inpFF = cur; - - // FF - { - // Norm - { - cur = ggml_norm(ctx0, inpFF, norm_eps); - offload_func_nr(cur); - - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b); - offload_func_nr(cur); - } - - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3); - offload_func(cur); - - // GELU activation - cur = ggml_gelu(ctx0, cur); - offload_func(cur); - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2); - offload_func(cur); - } - - inpL = ggml_add(ctx0, cur, inpFF); - - } - - // Output Norm - { - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func_nr(cur); - - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b); - ggml_set_name(cur, "result_norm"); - } - - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - - ggml_build_forward_expand(gf, cur); - ggml_free(ctx0); - - return gf; -} - -static struct ggml_cgraph * llm_build_persimmon( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const auto & cparams = lctx.cparams; - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_head = hparams.n_head; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - const size_t n_rot = n_embd_head / 2; - - const float freq_base = cparams.rope_freq_base; - const float freq_scale = cparams.rope_freq_scale; - const float norm_eps = hparams.f_norm_eps; - - const int n_gpu_layers = model.n_gpu_layers; - - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - const bool do_rope_shift = ggml_allocr_is_measure(lctx.alloc) || kv_self.has_shift; - - auto & buf_compute = lctx.buf_compute; - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ true, - }; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd_head))); - } - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - offload_func_kq(KQ_pos); - ggml_set_name(KQ_pos, "KQ_pos"); - ggml_allocr_alloc(lctx.alloc, KQ_pos); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) KQ_pos->data; - for (int i = 0; i < n_tokens; ++i) { - data[i] = batch.pos[i]; - } - } - if (do_rope_shift) { - struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_ctx); - offload_func_kq(K_shift); - ggml_set_name(K_shift, "K_shift"); - ggml_allocr_alloc(lctx.alloc, K_shift); - if (!ggml_allocr_is_measure(lctx.alloc)) { - int * data = (int *) K_shift->data; - for (int i = 0; i < n_ctx; ++i) { - data[i] = kv_self.cells[i].delta; - } - } - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * tmp = - // we rotate only the first n_rot dimensions. - ggml_rope_custom_inplace(ctx0, - ggml_view_3d(ctx0, kv_self.k, - n_rot, n_head, n_ctx, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*(n_embd_head*n_ctx*il) - ), - K_shift, n_rot, 2, 0, freq_base, freq_scale); - offload_func_kq(tmp); - ggml_build_forward_expand(gf, tmp); - } - } - for (int il=0; il < n_layer; ++il) { - struct ggml_tensor * residual = inpL; - offload_func_t offload_func = llama_nop; - { - cur = ggml_norm(ctx0, inpL, norm_eps); - offload_func(cur); - cur = ggml_mul(ctx0, cur, model.layers[il].attn_norm); - offload_func(cur); - cur = ggml_add(ctx0, cur, model.layers[il].attn_norm_b); - offload_func(cur); - ggml_format_name(cur, "input_layernorm_%d", il); - } - // self attention - { - cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - offload_func_kq(cur); - - // split qkv - GGML_ASSERT(n_head_kv == n_head); - ggml_set_name(cur, format("qkv_%d", il).c_str()); - struct ggml_tensor * tmpqkv = ggml_reshape_4d(ctx0, cur, n_embd_head, 3, n_head, n_tokens); - offload_func_kq(tmpqkv); - struct ggml_tensor * tmpqkv_perm = ggml_cont(ctx0, ggml_permute(ctx0, tmpqkv, 0, 3, 1, 2)); - offload_func_kq(tmpqkv_perm); - ggml_format_name(tmpqkv_perm, "tmpqkv_perm_%d", il); - struct ggml_tensor * tmpq = ggml_view_3d( - ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, - ggml_element_size(tmpqkv_perm) * n_embd_head, - ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, - 0 + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow ); - offload_func_kq(tmpq); - struct ggml_tensor * tmpk = ggml_view_3d( - ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, - ggml_element_size(tmpqkv_perm) * n_embd_head, - ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, - ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow ); - offload_func_kq(tmpk); - // Q/K Layernorm - tmpq = ggml_norm(ctx0, tmpq, norm_eps); - offload_func_kq(tmpq); - tmpq = ggml_mul(ctx0, tmpq, model.layers[il].attn_q_norm); - offload_func_kq(tmpq); - tmpq = ggml_add(ctx0, tmpq, model.layers[il].attn_q_norm_b); - offload_func_kq(tmpq); + cb(Kcur, "Kcur", il); - tmpk = ggml_norm(ctx0, tmpk, norm_eps); - offload_func_v(tmpk); - tmpk = ggml_mul(ctx0, tmpk, model.layers[il].attn_k_norm); - offload_func_v(tmpk); - tmpk = ggml_add(ctx0, tmpk, model.layers[il].attn_k_norm_b); - offload_func_v(tmpk); + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - // RoPE the first n_rot of q/k, pass the other half, and concat. - struct ggml_tensor * qrot = ggml_view_3d( - ctx0, tmpq, n_rot, n_head, n_tokens, - ggml_element_size(tmpq) * n_embd_head, - ggml_element_size(tmpq) * n_embd_head * n_head, - 0 - ); - offload_func_kq(qrot); - ggml_format_name(qrot, "qrot_%d", il); - struct ggml_tensor * krot = ggml_view_3d( - ctx0, tmpk, n_rot, n_head, n_tokens, - ggml_element_size(tmpk) * n_embd_head, - ggml_element_size(tmpk) * n_embd_head * n_head, - 0 - ); - offload_func_kq(krot); - ggml_format_name(krot, "krot_%d", il); + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); + } - // get the second half of tmpq, e.g tmpq[n_rot:, :, :] - struct ggml_tensor * qpass = ggml_view_3d( - ctx0, tmpq, n_rot, n_head, n_tokens, - ggml_element_size(tmpq) * n_embd_head, - ggml_element_size(tmpq) * n_embd_head * n_head, - ggml_element_size(tmpq) * n_rot - ); - offload_func_kq(qpass); - ggml_format_name(qpass, "qpass_%d", il); - struct ggml_tensor * kpass = ggml_view_3d( - ctx0, tmpk, n_rot, n_head, n_tokens, - ggml_element_size(tmpk) * n_embd_head, - ggml_element_size(tmpk) * n_embd_head * n_head, - ggml_element_size(tmpk) * n_rot - ); - offload_func_kq(kpass); - ggml_format_name(kpass, "kpass_%d", il); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); - struct ggml_tensor * qrotated = ggml_rope_custom( - ctx0, qrot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale - ); - offload_func_kq(qrotated); - struct ggml_tensor * krotated = ggml_rope_custom( - ctx0, krot, KQ_pos, n_rot, 2, 0, freq_base, freq_scale - ); - offload_func_kq(krotated); - // ggml currently only supports concatenation on dim=2 - // so we need to permute qrot, qpass, concat, then permute back. - qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3)); - offload_func_kq(qrotated); - krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3)); - offload_func_kq(krotated); - - qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3)); - offload_func_kq(qpass); - kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3)); - offload_func_kq(kpass); - - struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass); - offload_func_kq(Qcur); - struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass); - offload_func_kq(Kcur); - - struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 1, 2, 0, 3)); - offload_func_kq(Q); - - Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3)); - offload_func_kq(Kcur); + // feed-forward network { - struct ggml_tensor * tmpv = ggml_view_3d( + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_baichuan() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + switch (model.type) { + case MODEL_7B: + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + break; + case MODEL_13B: + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd/n_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd/n_head, n_head, n_tokens); + break; + default: + GGML_ASSERT(false); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + // apply ALiBi for 13B model + const float max_alibi_bias = model.type == MODEL_13B ? 8.0f : -1.0f; + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, max_alibi_bias, cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_falcon() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * attn_norm; + + attn_norm = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + if (model.layers[il].attn_norm_2) { + // Falcon-40B + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm_2", il); + } else { + cur = attn_norm; + } + + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + // using mode = 2 for neox mode + Qcur = ggml_rope_custom( + ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = cur; + + // feed forward + { + cur = llm_build_ffn(ctx0, attn_norm, // !! use the attn norm, not the result + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_starcoder() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * pos; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); + } + + // add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_persimmon() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + const int64_t n_rot = n_embd_head / 2; + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "imp_embd", -1); + + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * residual = inpL; + + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + // split qkv + GGML_ASSERT(n_head_kv == n_head); + + struct ggml_tensor * tmpqkv = ggml_reshape_4d(ctx0, cur, n_embd_head, 3, n_head, n_tokens); + cb(tmpqkv, "tmpqkv", il); + + struct ggml_tensor * tmpqkv_perm = ggml_cont(ctx0, ggml_permute(ctx0, tmpqkv, 0, 3, 1, 2)); + cb(tmpqkv_perm, "tmpqkv", il); + + struct ggml_tensor * tmpq = ggml_view_3d( + ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, + ggml_element_size(tmpqkv_perm) * n_embd_head, + ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, + 0 + ); + cb(tmpq, "tmpq", il); + + struct ggml_tensor * tmpk = ggml_view_3d( + ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, + ggml_element_size(tmpqkv_perm) * n_embd_head, + ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, + ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens + ); + cb(tmpk, "tmpk", il); + + // Q/K Layernorm + tmpq = llm_build_norm(ctx0, tmpq, hparams, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, cb, il); + cb(tmpq, "tmpq", il); + + tmpk = llm_build_norm(ctx0, tmpk, hparams, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, cb, il); + cb(tmpk, "tmpk", il); + + // RoPE the first n_rot of q/k, pass the other half, and concat. + struct ggml_tensor * qrot = ggml_view_3d( + ctx0, tmpq, n_rot, n_head, n_tokens, + ggml_element_size(tmpq) * n_embd_head, + ggml_element_size(tmpq) * n_embd_head * n_head, + 0 + ); + cb(qrot, "qrot", il); + + struct ggml_tensor * krot = ggml_view_3d( + ctx0, tmpk, n_rot, n_head, n_tokens, + ggml_element_size(tmpk) * n_embd_head, + ggml_element_size(tmpk) * n_embd_head * n_head, + 0 + ); + cb(krot, "krot", il); + + // get the second half of tmpq, e.g tmpq[n_rot:, :, :] + struct ggml_tensor * qpass = ggml_view_3d( + ctx0, tmpq, n_rot, n_head, n_tokens, + ggml_element_size(tmpq) * n_embd_head, + ggml_element_size(tmpq) * n_embd_head * n_head, + ggml_element_size(tmpq) * n_rot + ); + cb(qpass, "qpass", il); + + struct ggml_tensor * kpass = ggml_view_3d( + ctx0, tmpk, n_rot, n_head, n_tokens, + ggml_element_size(tmpk) * n_embd_head, + ggml_element_size(tmpk) * n_embd_head * n_head, + ggml_element_size(tmpk) * n_rot + ); + cb(kpass, "kpass", il); + + struct ggml_tensor * qrotated = ggml_rope_custom( + ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(qrotated, "qrotated", il); + + struct ggml_tensor * krotated = ggml_rope_custom( + ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx, + freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(krotated, "krotated", il); + + // ggml currently only supports concatenation on dim=2 + // so we need to permute qrot, qpass, concat, then permute back. + qrotated = ggml_cont(ctx0, ggml_permute(ctx0, qrotated, 2, 1, 0, 3)); + cb(qrotated, "qrotated", il); + + krotated = ggml_cont(ctx0, ggml_permute(ctx0, krotated, 2, 1, 0, 3)); + cb(krotated, "krotated", il); + + qpass = ggml_cont(ctx0, ggml_permute(ctx0, qpass, 2, 1, 0, 3)); + cb(qpass, "qpass", il); + + kpass = ggml_cont(ctx0, ggml_permute(ctx0, kpass, 2, 1, 0, 3)); + cb(kpass, "kpass", il); + + struct ggml_tensor * Qcur = ggml_concat(ctx0, qrotated, qpass); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_concat(ctx0, krotated, kpass); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Q = ggml_cont(ctx0, ggml_permute(ctx0, Qcur, 2, 1, 0, 3)); + cb(Q, "Q", il); + + Kcur = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 2, 1, 0, 3)); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_view_3d( ctx0, tmpqkv_perm, n_embd_head, n_head, n_tokens, ggml_element_size(tmpqkv_perm) * n_embd_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head, ggml_element_size(tmpqkv_perm) * n_embd_head * n_head * n_tokens * 2 - ); - offload_func_v(tmpv); - // store K, V in cache - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - ggml_set_name(Vcur, "Vcur"); + ); + cb(Vcur, "Vcur", il); - struct ggml_tensor * k = ggml_view_1d( - ctx0, kv_self.k, n_tokens*n_embd_gqa, - (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head) - ); - offload_func_kq(k); - ggml_set_name(k, "k"); + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - ggml_set_name(v, "v"); - - // important: storing RoPE-ed version of K in the KV cache! - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + // TODO: not tested, could be broken + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Q, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * K = ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_format_name(K, "K_%d", il); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - offload_func_kq(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - cur = ggml_add(ctx0, cur, model.layers[il].bo); - offload_func(cur); - ggml_set_name(cur, "result_wo"); - } - - struct ggml_tensor * inpFF = ggml_add(ctx0, residual, cur); - offload_func(inpFF); - ggml_set_name(inpFF, "inpFF"); - { - // MLP + // feed-forward network { - // Norm - cur = ggml_norm(ctx0, inpFF, norm_eps); - offload_func(cur); - cur = ggml_add(ctx0, - ggml_mul(ctx0, cur, model.layers[il].ffn_norm), - model.layers[il].ffn_norm_b - ); - ggml_set_name(cur, "ffn_norm"); - offload_func(cur); + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur); - offload_func(cur); - cur = ggml_add(ctx0, cur, model.layers[il].b3); - offload_func(cur); - ggml_set_name(cur, "result_ffn_up"); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); - cur = ggml_sqr(ctx0, ggml_relu(ctx0, cur)); - ggml_set_name(cur, "result_ffn_act"); - offload_func(cur); - offload_func(cur->src[0]); - - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); - cur = ggml_add(ctx0, - cur, - model.layers[il].b2); - offload_func(cur); - ggml_set_name(cur, "outFF"); - } - cur = ggml_add(ctx0, cur, inpFF); - offload_func(cur); - ggml_set_name(cur, "inpFF_+_outFF"); - inpL = cur; - } - cur = inpL; - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - cur = ggml_mul(ctx0, cur, model.output_norm); - offload_func_nr(cur); - - cur = ggml_add(ctx0, cur, model.output_norm_b); - // offload_func_nr(cur); - - ggml_set_name(cur, "result_norm"); - } - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); - ggml_build_forward_expand(gf, cur); - ggml_free(ctx0); - return gf; -} - -static struct ggml_cgraph * llm_build_bloom( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; - - const auto & kv_self = lctx.kv_self; - - GGML_ASSERT(!!kv_self.ctx); - - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_rot); - - const float norm_eps = hparams.f_norm_eps; - - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; - - auto & buf_compute = lctx.buf_compute; - - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ false, - }; - - params.no_alloc = true; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * token; - struct ggml_tensor * inpL; - - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - } - ggml_set_name(inp_tokens, "inp_tokens"); - - token = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - token = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, token); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(token->data, batch.embd, n_tokens * n_embd * ggml_element_size(token)); - } - } - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } - } - } - } - - // norm - { - inpL = ggml_norm(ctx0, token, norm_eps); - inpL = ggml_add(ctx0, ggml_mul(ctx0, inpL, model.tok_norm), model.tok_norm_b); - } - - ggml_set_name(inpL, "inpL"); - - for (int il = 0; il < n_layer; ++il) { - { - // Norm - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].attn_norm), model.layers[il].attn_norm_b); + inpL = cur; } - { - // Self Attention - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wqkv, cur), model.layers[il].bqkv); + cur = inpL; - struct ggml_tensor * tmpq = ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*n_embd); - struct ggml_tensor * tmpk = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*n_embd); - struct ggml_tensor * tmpv = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], sizeof(float)*(n_embd + n_embd_gqa)); + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); - struct ggml_tensor * Qcur = tmpq; - struct ggml_tensor * Kcur = tmpk; + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); - // store key and value to memory + ggml_build_forward_expand(gf, cur); + + return gf; + } + + struct ggml_cgraph * build_refact() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "attn_norm", il); + + // self-attention { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - ggml_set_name(Vcur, "Vcur"); + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - ggml_set_name(k, "k"); + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + cb(Kcur, "Kcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + cb(Qcur, "Qcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + cb(cur, "kqv_out", il); } - struct ggml_tensor * Q = - ggml_permute(ctx0, - ggml_cpy(ctx0, - Qcur, - ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd_head, n_head, n_tokens)), - 0, 2, 1, 3); - ggml_set_name(Q, "Q"); + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - ggml_set_name(K, "K"); - - // K * Q - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - ggml_set_name(KQ, "KQ"); - - // KQ_scaled = KQ / sqrt(n_embd_head) - // KQ_scaled shape [n_past + n_tokens, n_tokens, n_head, 1] - struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, /*n_past*/ kv_head, n_head, 8); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - // KQ_masked = mask_past(KQ_scaled) - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - ggml_set_name(KQ_masked, "KQ_masked"); - - // KQ = soft_max(KQ_masked) - struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - // split cached V into n_head heads - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - ggml_set_name(KQV, "KQV"); - - // KQV_merged = KQV.permute(0, 2, 1, 3) - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - ggml_set_name(KQV_merged, "KQV_merged"); - - // cur = KQV_merged.contiguous().view(n_embd, n_tokens) - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - ggml_set_name(cur, "KQV_merged_contiguous"); - } - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wo, cur), model.layers[il].bo); - - // Add the input - cur = ggml_add(ctx0, cur, inpL); - - struct ggml_tensor * inpFF = cur; - - // FF - { - // Norm + // feed-forward network { - cur = ggml_norm(ctx0, inpFF, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ffn_norm), model.layers[il].ffn_norm_b); + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); } - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w3, cur), model.layers[il].b3); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); - // GELU activation - cur = ggml_gelu(ctx0, cur); - - // Projection - cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].w2, cur), model.layers[il].b2); + // input for next layer + inpL = cur; } - inpL = ggml_add(ctx0, cur, inpFF); + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, NULL, + LLM_NORM_RMS, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; } - // Output Norm - { - cur = ggml_norm(ctx0, inpL, norm_eps); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.output_norm), model.output_norm_b); - } - ggml_set_name(cur, "result_norm"); + struct ggml_cgraph * build_bloom() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + struct ggml_tensor * cur; + struct ggml_tensor * inpL; - ggml_build_forward_expand(gf, cur); + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); - ggml_free(ctx0); + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); - return gf; -} + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); -static struct ggml_cgraph * llm_build_mpt( - llama_context & lctx, - const llama_batch & batch) { - const auto & model = lctx.model; - const auto & hparams = model.hparams; - const auto & cparams = lctx.cparams; + inpL = llm_build_norm(ctx0, inpL, hparams, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, cb, -1); + cb(inpL, "inp_norm", -1); - const auto & kv_self = lctx.kv_self; + for (int il = 0; il < n_layer; ++il) { + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); - GGML_ASSERT(!!kv_self.ctx); + // self-attention + { + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); - const int64_t n_embd = hparams.n_embd; - const int64_t n_layer = hparams.n_layer; - const int64_t n_ctx = cparams.n_ctx; - const int64_t n_head = hparams.n_head; - const int64_t n_head_kv = hparams.n_head_kv; - const int64_t n_embd_head = hparams.n_embd_head(); - const int64_t n_embd_gqa = hparams.n_embd_gqa(); + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); - const float norm_eps = hparams.f_norm_eps; - const float clamp_kqv = hparams.f_clamp_kqv; - const float max_alibi_bias = hparams.f_max_alibi_bias; + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); - const int n_gpu_layers = model.n_gpu_layers; + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); - const int32_t n_tokens = batch.n_tokens; - const int32_t n_kv = ggml_allocr_is_measure(lctx.alloc) ? n_ctx : kv_self.n; - const int32_t kv_head = ggml_allocr_is_measure(lctx.alloc) ? n_ctx - n_tokens : kv_self.head; + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - auto & buf_compute = lctx.buf_compute; + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); - struct ggml_init_params params = { - /*.mem_size =*/ buf_compute.size, - /*.mem_buffer =*/ buf_compute.data, - /*.no_alloc =*/ false, - }; - - params.no_alloc = true; - - struct ggml_context * ctx0 = ggml_init(params); - - ggml_cgraph * gf = ggml_new_graph(ctx0); - - struct ggml_tensor * cur; - struct ggml_tensor * inpL; - - //int warmup = 0; - if (batch.token) { - struct ggml_tensor * inp_tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inp_tokens); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inp_tokens->data, batch.token, n_tokens*ggml_element_size(inp_tokens)); - //warmup = ((uint32_t*) inp_tokens->data)[0] == 0; - } - - ggml_set_name(inp_tokens, "inp_tokens"); - - inpL = ggml_get_rows(ctx0, model.tok_embeddings, inp_tokens); - } else { -#ifdef GGML_USE_MPI - GGML_ASSERT(false && "not implemented"); -#endif - - inpL = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens); - - ggml_allocr_alloc(lctx.alloc, inpL); - if (!ggml_allocr_is_measure(lctx.alloc)) { - memcpy(inpL->data, batch.embd, n_tokens * n_embd * ggml_element_size(inpL)); - } - } - - const int i_gpu_start = n_layer - n_gpu_layers; - (void) i_gpu_start; - - // offload functions set the tensor output backend to GPU - // tensors are GPU-accelerated if any input or the output has been offloaded - offload_func_t offload_func_nr = llama_nop; // nr = non-repeating - offload_func_t offload_func_kq = llama_nop; - offload_func_t offload_func_v = llama_nop; - -#ifdef GGML_USE_CUBLAS - if (n_gpu_layers > n_layer) { - offload_func_nr = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 1) { - offload_func_v = ggml_cuda_assign_buffers_no_alloc; - } - if (n_gpu_layers > n_layer + 2) { - offload_func_kq = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS - - // KQ_scale - struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); - ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)"); - ggml_allocr_alloc(lctx.alloc, KQ_scale); - if (!ggml_allocr_is_measure(lctx.alloc)) { - ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head)); - } - - // KQ_mask (mask for 1 head, it will be broadcasted to all heads) - struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); - offload_func_kq(KQ_mask); - ggml_set_name(KQ_mask, "KQ_mask"); - ggml_allocr_alloc(lctx.alloc, KQ_mask); - if (!ggml_allocr_is_measure(lctx.alloc)) { - float * data = (float *) KQ_mask->data; - memset(data, 0, ggml_nbytes(KQ_mask)); - - for (int h = 0; h < 1; ++h) { - for (int j = 0; j < n_tokens; ++j) { - const llama_pos pos = batch.pos[j]; - const llama_seq_id seq_id = batch.seq_id[j][0]; - - for (int i = 0; i < n_kv; ++i) { - if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) { - data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; - } - } + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, 8.0f, cb, il); + cb(cur, "kqv_out", il); } + + // Add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, + NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); } + + cur = llm_build_norm(ctx0, inpL, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; } - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * attn_norm; + struct ggml_cgraph * build_mpt() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); - offload_func_t offload_func = llama_nop; + struct ggml_tensor * cur; + struct ggml_tensor * inpL; -#ifdef GGML_USE_CUBLAS - if (il >= i_gpu_start) { - offload_func = ggml_cuda_assign_buffers_no_alloc; - } -#endif // GGML_USE_CUBLAS + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); - // self-attention - // TODO: refactor into common function (shared with LLaMA) - { - attn_norm = ggml_norm(ctx0, inpL, norm_eps); - offload_func(attn_norm); + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); - attn_norm = ggml_mul(ctx0, attn_norm, model.layers[il].attn_norm); - offload_func(attn_norm); + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); - if (1) { + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * attn_norm; + + attn_norm = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + NULL, + LLM_NORM, cb, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { cur = attn_norm; + + cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + if (hparams.f_clamp_kqv > 0.0f) { + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + } + + struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd))); + struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd))); + struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa))); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, hparams.f_max_alibi_bias, cb, il); + cb(cur, "kqv_out", il); } - // compute QKV - - cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur); - offload_func_kq(cur); - - if (clamp_kqv > 0.0f) { - cur = ggml_clamp(ctx0, cur, -clamp_kqv, clamp_kqv); - offload_func_kq(cur); - } - - const size_t wsize = ggml_type_size(cur->type); - - struct ggml_tensor * Qcur = ggml_view_3d( - ctx0, cur, n_embd_head, n_head, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - 0); - offload_func_kq(Qcur); - - struct ggml_tensor * Kcur = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * n_head); - offload_func_kq(Kcur); - - struct ggml_tensor * tmpv = ggml_view_3d( - ctx0, cur, n_embd_head, n_head_kv, n_tokens, - wsize * n_embd_head, - wsize * n_embd_head * (n_head + 2 * n_head_kv), - wsize * n_embd_head * (n_head + n_head_kv)); - offload_func_kq(Kcur); - - ggml_set_name(Qcur, "Qcur"); - ggml_set_name(Kcur, "Kcur"); + // Add the input + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + // feed forward { - struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_cont(ctx0, tmpv), n_embd_gqa, n_tokens)); - offload_func_v(Vcur); - offload_func_v(Vcur->src[0]->src[0]); - ggml_set_name(Vcur, "Vcur"); + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + NULL, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); - struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + kv_head)); - offload_func_kq(k); - ggml_set_name(k, "k"); - - struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_embd_gqa, - ( n_ctx)*ggml_element_size(kv_self.v), - (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + kv_head*ggml_element_size(kv_self.v)); - offload_func_v(v); - - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k)); - ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v)); + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + NULL, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, cb, il); + cb(cur, "ffn_out", il); } - struct ggml_tensor * Q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - offload_func_kq(Q); - ggml_set_name(Q, "Q"); + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); - struct ggml_tensor * K = - ggml_view_3d(ctx0, kv_self.k, - n_embd_head, n_kv, n_head_kv, - ggml_element_size(kv_self.k)*n_embd_gqa, - ggml_element_size(kv_self.k)*n_embd_head, - ggml_element_size(kv_self.k)*n_embd_gqa*n_ctx*il); - offload_func_kq(K); - ggml_set_name(K, "K"); - - struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); - offload_func_kq(KQ); - ggml_set_name(KQ, "KQ"); - - struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale); - offload_func_kq(KQ_scaled); - ggml_set_name(KQ_scaled, "KQ_scaled"); - - // TODO: replace with ggml_add() - struct ggml_tensor * KQ_scaled_alibi = - ggml_alibi(ctx0, KQ_scaled, 0, n_head, max_alibi_bias); - offload_func_kq(KQ_scaled_alibi); - ggml_set_name(KQ_scaled_alibi, "KQ_scaled_alibi"); - - struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ_scaled_alibi, KQ_mask); - offload_func_kq(KQ_masked); - ggml_set_name(KQ_masked, "KQ_masked"); - - struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); - offload_func_v(KQ_soft_max); - ggml_set_name(KQ_soft_max, "KQ_soft_max"); - - struct ggml_tensor * V = - ggml_view_3d(ctx0, kv_self.v, - n_kv, n_embd_head, n_head_kv, - ggml_element_size(kv_self.v)*n_ctx, - ggml_element_size(kv_self.v)*n_ctx*n_embd_head, - ggml_element_size(kv_self.v)*n_ctx*n_embd_gqa*il); - offload_func_v(V); - ggml_set_name(V, "V"); - - struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max); - offload_func_v(KQV); - ggml_set_name(KQV, "KQV"); - - struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); - offload_func_v(KQV_merged); - ggml_set_name(KQV_merged, "KQV_merged"); - - cur = ggml_cont_2d(ctx0, KQV_merged, n_embd, n_tokens); - offload_func_v(cur); - ggml_set_name(cur, "KQV_merged_contiguous"); - - cur = ggml_mul_mat(ctx0, model.layers[il].wo, cur); - offload_func(cur); - ggml_set_name(cur, "result_wo"); + // input for next layer + inpL = cur; } - // Add the input - cur = ggml_add(ctx0, cur, inpL); - offload_func(cur); + cur = inpL; - struct ggml_tensor * attn_out = cur; + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + NULL, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); - // feed forward - { - // Norm - { - cur = ggml_norm(ctx0, attn_out, norm_eps); - offload_func(cur); + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm); - offload_func(cur); - } + ggml_build_forward_expand(gf, cur); - cur = ggml_mul_mat(ctx0, model.layers[il].w3, cur); - offload_func(cur); + return gf; + } - cur = ggml_gelu(ctx0, cur); - offload_func(cur); - cur = ggml_mul_mat(ctx0, model.layers[il].w2, cur); - offload_func(cur); + struct ggml_cgraph * build_stablelm() { + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + struct ggml_tensor * cur; + struct ggml_tensor * inpL; + + inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb); + cb(inpL, "inp_embd", -1); + + // inp_pos - contains the positions + struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens); + cb(inp_pos, "inp_pos", -1); + + // KQ_scale + struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1); + cb(KQ_scale, "KQ_scale", -1); + + // KQ_mask (mask for 1 head, it will be broadcasted to all heads) + struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1); + cb(KQ_mask, "KQ_mask", -1); + + // shift the entire K-cache if needed + if (do_rope_shift) { + llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb); } - cur = ggml_add(ctx0, cur, attn_out); - offload_func(cur); - // input for next layer - inpL = cur; + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = llm_build_norm(ctx0, inpL, hparams, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, cb, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Qcur, "Qcur", il); + + Kcur = ggml_rope_custom( + ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, + hparams.n_rot, 2, 0, n_orig_ctx, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(Kcur, "Kcur", il); + + llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il); + + cur = llm_build_kqv(ctx0, hparams, kv_self, + model.layers[il].wo, NULL, + Qcur, KQ_scale, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, cb, il); + cb(cur, "kqv_out", il); + } + + struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = llm_build_norm(ctx0, ffn_inp, hparams, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, cb, il); + cb(cur, "ffn_norm", il); + + cur = llm_build_ffn(ctx0, cur, + model.layers[il].ffn_up, NULL, + model.layers[il].ffn_gate, NULL, + model.layers[il].ffn_down, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, cb, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = llm_build_norm(ctx0, cur, hparams, + model.output_norm, + model.output_norm_b, + LLM_NORM, cb, -1); + cb(cur, "result_norm", -1); + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + cb(cur, "result_output", -1); + + ggml_build_forward_expand(gf, cur); + + return gf; + } +}; + +// +// tensor offloading helpers +// +// TODO: will be removed with backend v2 + +enum llm_offload_func_e { + OFFLOAD_FUNC_NOP, + OFFLOAD_FUNC, + OFFLOAD_FUNC_KQ, + OFFLOAD_FUNC_V, + OFFLOAD_FUNC_NR, + OFFLOAD_FUNC_EMB, + OFFLOAD_FUNC_OUT, +}; + +// TODO: will be removed with backend v2 +struct llm_offload_trie { + struct node { + ~node() { + for (int i = 0; i < 256; ++i) { + if (children[i]) { + delete children[i]; + } + } + } + + node * children[256] = { nullptr }; + llm_offload_func_e func = OFFLOAD_FUNC_NOP; + }; + + llm_offload_trie() { + root = new node; } - cur = inpL; + llm_offload_trie(const std::unordered_map & map) { + root = new node; - // norm - { - cur = ggml_norm(ctx0, cur, norm_eps); - offload_func_nr(cur); - - cur = ggml_mul(ctx0, cur, model.output_norm); - ggml_set_name(cur, "result_norm"); + for (const auto & kv : map) { + add(kv.first, kv.second); + } } - cur = ggml_mul_mat(ctx0, model.output, cur); - ggml_set_name(cur, "result_output"); + ~llm_offload_trie() { + delete root; + } - ggml_build_forward_expand(gf, cur); + void add(const char * name, llm_offload_func_e func) { + node * cur = root; - ggml_free(ctx0); + for (int i = 0; ; ++i) { + const uint8_t c = name[i]; - return gf; -} + if (!c) { + break; + } + + if (!cur->children[c]) { + cur->children[c] = new node; + } + + cur = cur->children[c]; + } + + cur->func = func; + } + + llm_offload_func_e find(const char * name) const { + const node * cur = root; + + for (int i = 0; ; ++i) { + const uint8_t c = name[i]; + + if (!c) { + break; + } + + if (!cur->children[c]) { + return OFFLOAD_FUNC_NOP; + } + + cur = cur->children[c]; + } + + return cur->func; + } + + node * root = nullptr; +}; + +// TODO: will be removed with backend v2 +static const std::unordered_map k_offload_map = { + //{ "inp_tokens", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + //{ "inp_embd", OFFLOAD_FUNC_NR }, // TODO: missing K-quants get_rows kernel + { "pos_embd", OFFLOAD_FUNC_NR }, + + { "inp_pos", OFFLOAD_FUNC_KQ }, // this is often used for KQ ops (e.g. rope) + { "KQ_scale", OFFLOAD_FUNC_KQ }, + { "KQ_mask", OFFLOAD_FUNC_KQ }, + { "K_shift", OFFLOAD_FUNC_KQ }, + { "K_shifted", OFFLOAD_FUNC_KQ }, + + { "inp_norm", OFFLOAD_FUNC_NR }, + { "inp_norm_w", OFFLOAD_FUNC_NR }, + { "inp_norm_wb", OFFLOAD_FUNC_NR }, + + { "norm", OFFLOAD_FUNC }, + { "norm_w", OFFLOAD_FUNC }, + { "norm_wb", OFFLOAD_FUNC }, + + { "attn_norm", OFFLOAD_FUNC }, + { "attn_norm_2", OFFLOAD_FUNC }, + + { "wqkv", OFFLOAD_FUNC_KQ }, + { "bqkv", OFFLOAD_FUNC_KQ }, + { "wqkv_clamped", OFFLOAD_FUNC_KQ }, + + { "tmpk", OFFLOAD_FUNC_KQ }, + { "tmpq", OFFLOAD_FUNC_KQ }, + { "tmpv", OFFLOAD_FUNC_V }, + { "Kcur", OFFLOAD_FUNC_KQ }, + { "Qcur", OFFLOAD_FUNC_KQ }, + { "Vcur", OFFLOAD_FUNC_V }, + + { "krot", OFFLOAD_FUNC_KQ }, + { "qrot", OFFLOAD_FUNC_KQ }, + { "kpass", OFFLOAD_FUNC_KQ }, + { "qpass", OFFLOAD_FUNC_KQ }, + { "krotated", OFFLOAD_FUNC_KQ }, + { "qrotated", OFFLOAD_FUNC_KQ }, + + { "q", OFFLOAD_FUNC_KQ }, + { "k", OFFLOAD_FUNC_KQ }, + { "kq", OFFLOAD_FUNC_KQ }, + { "kq_scaled", OFFLOAD_FUNC_KQ }, + { "kq_scaled_alibi", OFFLOAD_FUNC_KQ }, + { "kq_masked", OFFLOAD_FUNC_KQ }, + { "kq_soft_max", OFFLOAD_FUNC_V }, + { "v", OFFLOAD_FUNC_V }, + { "kqv", OFFLOAD_FUNC_V }, + { "kqv_merged", OFFLOAD_FUNC_V }, + { "kqv_merged_cont", OFFLOAD_FUNC_V }, + { "kqv_wo", OFFLOAD_FUNC_V }, + { "kqv_out", OFFLOAD_FUNC_V }, + + { "ffn_inp", OFFLOAD_FUNC }, + { "ffn_norm", OFFLOAD_FUNC }, + + { "ffn_up", OFFLOAD_FUNC }, + { "ffn_up_b", OFFLOAD_FUNC }, + { "ffn_gate", OFFLOAD_FUNC }, + { "ffn_gate_b", OFFLOAD_FUNC }, + { "ffn_gate_par", OFFLOAD_FUNC }, + { "ffn_down", OFFLOAD_FUNC }, + { "ffn_down_b", OFFLOAD_FUNC }, + { "ffn_out", OFFLOAD_FUNC }, + + { "ffn_silu", OFFLOAD_FUNC }, + { "ffn_gelu", OFFLOAD_FUNC }, + { "ffn_relu", OFFLOAD_FUNC }, + { "ffn_sqr(relu)", OFFLOAD_FUNC }, + + { "l_out", OFFLOAD_FUNC }, + + { "result_norm", OFFLOAD_FUNC_EMB }, + { "result_output", OFFLOAD_FUNC_OUT }, +}; + +static llm_offload_trie k_offload_func_trie(k_offload_map); static struct ggml_cgraph * llama_build_graph( llama_context & lctx, const llama_batch & batch) { const auto & model = lctx.model; + // check if we should build the worst-case graph (for memory measurement) + const bool worst_case = ggml_allocr_is_measure(lctx.alloc); + + // keep track of the input that has already been allocated + bool alloc_inp_tokens = false; + bool alloc_inp_embd = false; + bool alloc_inp_pos = false; + bool alloc_inp_KQ_scale = false; + bool alloc_inp_KQ_mask = false; + bool alloc_inp_K_shift = false; + +#ifdef GGML_USE_CUBLAS + const bool do_offload = true; +#else + const bool do_offload = true; // TODO: set to false after finishing refactoring +#endif + + int n_non_view = 0; // number of non-view tensors that have been processed by the callback + + // this callback allows us to apply custom logic to each tensor (e.g. ggml-alloc, offloading, etc.) + // TODO: will be removed with backend v2 + llm_build_cb cb = [&](struct ggml_tensor * cur, const char * name, int il) { + if (il >= 0) { + ggml_format_name(cur, "%s-%d", name, il); + } else { + ggml_set_name(cur, name); + } + + // + // allocate input tensors and set input data + // + // TODO: will be removed with backend v2 + + if (!alloc_inp_tokens && strcmp(name, "inp_tokens") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.token) { + const int64_t n_tokens = cur->ne[0]; + + memcpy(cur->data, batch.token, n_tokens*ggml_element_size(cur)); + } + + alloc_inp_tokens = true; + } + + if (!alloc_inp_embd && strcmp(name, "inp_embd") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.embd) { + const int64_t n_embd = cur->ne[0]; + const int64_t n_tokens = cur->ne[1]; + + memcpy(cur->data, batch.embd, n_tokens*n_embd*ggml_element_size(cur)); + } + + alloc_inp_embd = true; + } + + if (!alloc_inp_pos && strcmp(name, "inp_pos") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc) && batch.pos) { + const int64_t n_tokens = cur->ne[0]; + + int32_t * data = (int32_t *) cur->data; + + for (int i = 0; i < n_tokens; ++i) { + data[i] = batch.pos[i]; + } + } + + alloc_inp_pos = true; + } + + if (!alloc_inp_KQ_scale && strcmp(name, "KQ_scale") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_embd_head = model.hparams.n_embd_head(); + ggml_set_f32(cur, 1.0f/sqrtf(float(n_embd_head))); + } + + alloc_inp_KQ_scale = true; + } + + if (!alloc_inp_KQ_mask && strcmp(name, "KQ_mask") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_kv = cur->ne[0]; + const int64_t n_tokens = cur->ne[1]; + + float * data = (float *) cur->data; + memset(data, 0, ggml_nbytes(cur)); + + for (int h = 0; h < 1; ++h) { + for (int j = 0; j < n_tokens; ++j) { + const llama_pos pos = batch.pos[j]; + const llama_seq_id seq_id = batch.seq_id[j][0]; + + for (int i = 0; i < n_kv; ++i) { + if (!lctx.kv_self.cells[i].has_seq_id(seq_id) || lctx.kv_self.cells[i].pos > pos) { + data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY; + } + } + } + } + } + + alloc_inp_KQ_mask = true; + } + + if (!alloc_inp_K_shift && strcmp(name, "K_shift") == 0) { + ggml_allocr_alloc(lctx.alloc, cur); + + if (!ggml_allocr_is_measure(lctx.alloc)) { + const int64_t n_ctx = cur->ne[0]; + + int32_t * data = (int32_t *) cur->data; + + for (int i = 0; i < n_ctx; ++i) { + data[i] = lctx.kv_self.cells[i].delta; + } + } + + alloc_inp_K_shift = true; + } + + // view tensors are not processed further + if (cur->view_src != nullptr) { + return; + } + + if (cur->op != GGML_OP_NONE) { + n_non_view++; + } + + // + // offload layers + // + // TODO: will be removed with backend v2 + +//#define LLAMA_OFFLOAD_DEBUG + + if (!do_offload) { + return; + } + + const int n_layer = model.hparams.n_layer; + + const int n_gpu_layers = model.n_gpu_layers; + const int i_gpu_start = n_layer - n_gpu_layers; + + // should we offload the final norm? yes if we are not computing embeddings + const bool offload_emb = lctx.embedding.empty(); + + static const std::unordered_map> k_offload_func_name = { + { OFFLOAD_FUNC_NOP, "CPU" }, + { OFFLOAD_FUNC_OUT, "CPU" }, +#ifdef GGML_USE_CUBLAS + { OFFLOAD_FUNC, "GPU (CUDA)" }, + { OFFLOAD_FUNC_KQ, "GPU (CUDA) KQ" }, + { OFFLOAD_FUNC_V, "GPU (CUDA) V" }, + { OFFLOAD_FUNC_NR, "GPU (CUDA) NR" }, + { OFFLOAD_FUNC_EMB, "GPU (CUDA) EMB" }, +#else + { OFFLOAD_FUNC, "CPU" }, + { OFFLOAD_FUNC_KQ, "CPU" }, + { OFFLOAD_FUNC_V, "CPU" }, + { OFFLOAD_FUNC_NR, "CPU" }, + { OFFLOAD_FUNC_EMB, "CPU" }, +#endif // GGML_USE_CUBLAS + }; + + // check the global map for what offload function to use for this tensor + llm_offload_func_e func_e = k_offload_func_trie.find(name); + + if (func_e == OFFLOAD_FUNC_NOP) { +#ifdef LLAMA_OFFLOAD_DEBUG + // if a tensor hasn't been offloaded, we warn the user + if (worst_case) { + LLAMA_LOG_WARN("%s: %32s: not offloaded (ref: %s)\n", __func__, + cur->name, "https://github.com/ggerganov/llama.cpp/pull/3837"); + } +#endif + + return; + } + + // count the number of layers and respect the provided n_gpu_layers + switch (func_e) { + case OFFLOAD_FUNC_NOP: + case OFFLOAD_FUNC_OUT: + break; + case OFFLOAD_FUNC: + if (n_gpu_layers < n_layer) { + if (il < i_gpu_start) { + func_e = OFFLOAD_FUNC_NOP; + } + } + break; + case OFFLOAD_FUNC_NR: + if (n_gpu_layers <= n_layer + 0) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_V: + if (n_gpu_layers <= n_layer + 1) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_KQ: + if (n_gpu_layers <= n_layer + 2) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + case OFFLOAD_FUNC_EMB: + if (!offload_emb || n_gpu_layers < n_layer) { + func_e = OFFLOAD_FUNC_NOP; + } + break; + default: GGML_ASSERT(false); + } + + offload_func_t func = ggml_offload_nop; + + // this is needed for compatibility with Metal for example +#ifdef GGML_USE_CUBLAS + static offload_func_t ggml_offload_gpu = ggml_cuda_assign_buffers_no_alloc; +#else + static offload_func_t ggml_offload_gpu = ggml_offload_nop; +#endif + + switch (func_e) { + case OFFLOAD_FUNC_NOP: + case OFFLOAD_FUNC_OUT: func = ggml_offload_nop; break; + case OFFLOAD_FUNC: + case OFFLOAD_FUNC_KQ: + case OFFLOAD_FUNC_V: + case OFFLOAD_FUNC_NR: + case OFFLOAD_FUNC_EMB: func = ggml_offload_gpu; break; + default: GGML_ASSERT(false); + } + + // apply offload function to the tensor + func(cur); + +#ifdef LLAMA_OFFLOAD_DEBUG + if (worst_case) { + LLAMA_LOG_INFO("%s: %32s: %s\n", __func__, cur->name, k_offload_func_name.at(func_e).c_str()); + } +#endif + }; + struct ggml_cgraph * result = NULL; + struct llm_build_context llm(lctx, batch, cb, worst_case); + + llm.init(); + switch (model.arch) { case LLM_ARCH_LLAMA: { - result = llm_build_llama(lctx, batch); + result = llm.build_llama(); } break; case LLM_ARCH_BAICHUAN: { - result = llm_build_baichaun(lctx, batch); + result = llm.build_baichuan(); } break; case LLM_ARCH_FALCON: { - result = llm_build_falcon(lctx, batch); + result = llm.build_falcon(); } break; case LLM_ARCH_STARCODER: { - result = llm_build_starcoder(lctx, batch); + result = llm.build_starcoder(); } break; case LLM_ARCH_PERSIMMON: { - result = llm_build_persimmon(lctx, batch); + result = llm.build_persimmon(); } break; case LLM_ARCH_REFACT: { - result = llm_build_refact(lctx, batch); + result = llm.build_refact(); } break; case LLM_ARCH_BLOOM: { - result = llm_build_bloom(lctx, batch); + result = llm.build_bloom(); } break; case LLM_ARCH_MPT: { - result = llm_build_mpt(lctx, batch); + result = llm.build_mpt(); + } break; + case LLM_ARCH_STABLELM: + { + result = llm.build_stablelm(); } break; default: GGML_ASSERT(false); } + llm.free(); + + if (worst_case) { + int n_non_view_total = 0; + + for (int i = 0; i < result->n_nodes; ++i) { + if (result->nodes[i]->view_src == nullptr) { + n_non_view_total++; + } + } + + LLAMA_LOG_INFO("%s: non-view tensors processed: %d/%d\n", __func__, n_non_view, n_non_view_total); + + if (n_non_view != n_non_view_total) { + LLAMA_LOG_WARN("%s: ****************************************************************\n", __func__); + LLAMA_LOG_WARN("%s: not all non-view tensors have been processed with a callback\n", __func__); + LLAMA_LOG_WARN("%s: this can indicate an inefficiency in the graph implementation\n", __func__); + LLAMA_LOG_WARN("%s: build with LLAMA_OFFLOAD_DEBUG for more info\n", __func__); + LLAMA_LOG_WARN("%s: ref: https://github.com/ggerganov/llama.cpp/pull/3837\n", __func__); + LLAMA_LOG_WARN("%s: ****************************************************************\n", __func__); + } + } + return result; } @@ -5983,6 +5485,12 @@ static int llama_decode_internal( batch.seq_id = seq_id_arr.data(); } + // if we have enough unused cells before the current head -> + // better to start searching from the beginning of the cache, hoping to fill it + if (kv_self.head > kv_self.used + 2*n_tokens) { + kv_self.head = 0; + } + if (!llama_kv_cache_find_slot(kv_self, batch)) { return 1; } @@ -5993,7 +5501,7 @@ static int llama_decode_internal( //kv_self.n = std::max(32, GGML_PAD(llama_kv_cache_cell_max(kv_self), 32)); // TODO: this might be better for CUDA? kv_self.n = std::min((int32_t) cparams.n_ctx, std::max(32, llama_kv_cache_cell_max(kv_self))); - //printf("kv_self.n = %d\n", kv_self.n); + //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head); ggml_allocr_reset(lctx.alloc); @@ -6043,11 +5551,15 @@ static int llama_decode_internal( } // If all tensors can be run on the GPU then using more than 1 thread is detrimental. - const bool full_offload_supported = model.arch == LLM_ARCH_LLAMA || - model.arch == LLM_ARCH_BAICHUAN || - model.arch == LLM_ARCH_FALCON || - model.arch == LLM_ARCH_REFACT || - model.arch == LLM_ARCH_MPT; + const bool full_offload_supported = + model.arch == LLM_ARCH_LLAMA || + model.arch == LLM_ARCH_BAICHUAN || + model.arch == LLM_ARCH_FALCON || + model.arch == LLM_ARCH_REFACT || + model.arch == LLM_ARCH_MPT || + model.arch == LLM_ARCH_STARCODER || + model.arch == LLM_ARCH_STABLELM; + const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3; if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) { n_threads = 1; @@ -6102,6 +5614,8 @@ static int llama_decode_internal( //} // extract logits + // TODO: do not compute and extract logits if only embeddings are needed + // need to update the graphs to skip "result_output" { auto & logits_out = lctx.logits; @@ -6836,7 +6350,10 @@ static std::vector llama_tokenize_internal(const llama_vocab & // by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer // and passing 'add space prefix' as bool argument // - auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length); + auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length); + if (&fragment == &fragment_buffer.front()) { + raw_text = " " + raw_text; // prefix with space if the first token is not special + } #ifdef PRETOKENIZERDEBUG fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str()); @@ -7378,6 +6895,32 @@ void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * can } } +void llama_sample_min_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep) { + if (p <= 0.0f || !candidates->size) { + return; + } + + llama_sample_softmax(ctx, candidates); + + const int64_t t_start_sample_us = ggml_time_us(); + + float scale = candidates->data[0].p; // scale by max prob + size_t i = 1; // first token always matches + + for (; i < candidates->size; ++i) { + if (candidates->data[i].p < p * scale && i >= min_keep) { + break; // prob too small + } + } + + // Resize the output vector to keep only the matching tokens + candidates->size = i; + + if (ctx) { + ctx->t_sample_us += ggml_time_us() - t_start_sample_us; + } +} + void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep) { if (z >= 1.0f || candidates->size <= 2) { return; @@ -8472,7 +8015,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s workers.clear(); } - LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); + LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0); int64_t tot_count = 0; for (size_t i = 0; i < hist_cur.size(); i++) { hist_all[i] += hist_cur[i]; @@ -8697,14 +8240,14 @@ static int llama_apply_lora_from_file_internal( ggml_tensor * dest_t = model_tensors[base_name]; - offload_func_t offload_func = llama_nop; - offload_func_t offload_func_force_inplace = llama_nop; + offload_func_t offload_func = ggml_offload_nop; + offload_func_t offload_func_force_inplace = ggml_offload_nop; #ifdef GGML_USE_CUBLAS if (dest_t->backend == GGML_BACKEND_GPU || dest_t->backend == GGML_BACKEND_GPU_SPLIT) { if (dest_t->type != GGML_TYPE_F16) { throw std::runtime_error(format( - "%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models", __func__)); + "%s: error: the simultaneous use of LoRAs and GPU acceleration is only supported for f16 models. dest_t->type: %d", __func__, dest_t->type)); } offload_func = ggml_cuda_assign_buffers; offload_func_force_inplace = ggml_cuda_assign_buffers_force_inplace; @@ -8839,8 +8382,14 @@ struct llama_context_params llama_context_default_params() { /*.n_batch =*/ 512, /*.n_threads =*/ GGML_DEFAULT_N_THREADS, // TODO: better default /*.n_threads_batch =*/ GGML_DEFAULT_N_THREADS, + /*.rope_scaling_type =*/ LLAMA_ROPE_SCALING_UNSPECIFIED, /*.rope_freq_base =*/ 0.0f, /*.rope_freq_scale =*/ 0.0f, + /*.yarn_ext_factor =*/ -1.0f, + /*.yarn_attn_factor =*/ 1.0f, + /*.yarn_beta_fast =*/ 32.0f, + /*.yarn_beta_slow =*/ 1.0f, + /*.yarn_orig_ctx =*/ 0, /*.mul_mat_q =*/ true, /*.f16_kv =*/ true, /*.logits_all =*/ false, @@ -8927,10 +8476,7 @@ struct llama_model * llama_load_model_from_file( }; } - if (!llama_model_load(path_model, *model, params.n_gpu_layers, - params.main_gpu, params.tensor_split, - params.use_mmap, params.use_mlock, params.vocab_only, - params.progress_callback, params.progress_callback_user_data)) { + if (!llama_model_load(path_model, *model, params)) { LLAMA_LOG_ERROR("%s: failed to load model\n", __func__); delete model; return nullptr; @@ -8956,13 +8502,35 @@ struct llama_context * llama_new_context_with_model( const auto & hparams = model->hparams; auto & cparams = ctx->cparams; - cparams.n_batch = params.n_batch; - cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; - cparams.rope_freq_base = params.rope_freq_base == 0 ? hparams.rope_freq_base_train : params.rope_freq_base; - cparams.rope_freq_scale = params.rope_freq_scale == 0 ? hparams.rope_freq_scale_train : params.rope_freq_scale; - cparams.n_threads = params.n_threads; - cparams.n_threads_batch = params.n_threads_batch; - cparams.mul_mat_q = params.mul_mat_q; + cparams.n_batch = params.n_batch; + cparams.n_threads = params.n_threads; + cparams.n_threads_batch = params.n_threads_batch; + cparams.yarn_ext_factor = params.yarn_ext_factor; + cparams.yarn_attn_factor = params.yarn_attn_factor; + cparams.yarn_beta_fast = params.yarn_beta_fast; + cparams.yarn_beta_slow = params.yarn_beta_slow; + cparams.mul_mat_q = params.mul_mat_q; + + cparams.n_ctx = params.n_ctx == 0 ? hparams.n_ctx_train : params.n_ctx; + cparams.rope_freq_base = params.rope_freq_base == 0.0f ? hparams.rope_freq_base_train : params.rope_freq_base; + cparams.rope_freq_scale = params.rope_freq_scale == 0.0f ? hparams.rope_freq_scale_train : params.rope_freq_scale; + + cparams.n_yarn_orig_ctx = params.yarn_orig_ctx != 0 ? params.yarn_orig_ctx : + hparams.n_yarn_orig_ctx != 0 ? hparams.n_yarn_orig_ctx : + hparams.n_ctx_train; + + auto rope_scaling_type = params.rope_scaling_type; + if (rope_scaling_type == LLAMA_ROPE_SCALING_UNSPECIFIED) { + rope_scaling_type = hparams.rope_scaling_type_train; + } + + if (rope_scaling_type == LLAMA_ROPE_SCALING_NONE) { + cparams.rope_freq_scale = 1.0f; // never scale if scaling type is none + } + + if (cparams.yarn_ext_factor < 0.0f) { // negative indicates 'not set' + cparams.yarn_ext_factor = rope_scaling_type == LLAMA_ROPE_SCALING_YARN ? 1.0f : 0.0f; + } if (params.seed == LLAMA_DEFAULT_SEED) { params.seed = time(NULL); @@ -8987,7 +8555,7 @@ struct llama_context * llama_new_context_with_model( { const size_t memory_size = ggml_nbytes(ctx->kv_self.k) + ggml_nbytes(ctx->kv_self.v); - LLAMA_LOG_INFO("%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: kv self size = %7.2f MiB\n", __func__, memory_size / 1024.0 / 1024.0); } // resized during inference @@ -9004,7 +8572,7 @@ struct llama_context * llama_new_context_with_model( { static const size_t tensor_alignment = 32; // the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data - ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead()); + ctx->buf_compute.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead()); // create measure allocator ctx->alloc = ggml_allocr_new_measure(tensor_alignment); @@ -9017,8 +8585,6 @@ struct llama_context * llama_new_context_with_model( #ifdef GGML_USE_METAL if (model->n_gpu_layers > 0) { - ggml_metal_log_set_callback(llama_log_callback_default, NULL); - ctx->ctx_metal = ggml_metal_init(1); if (!ctx->ctx_metal) { LLAMA_LOG_ERROR("%s: ggml_metal_init() failed\n", __func__); @@ -9032,7 +8598,7 @@ struct llama_context * llama_new_context_with_model( // measure memory requirements for the graph size_t alloc_size = ggml_allocr_alloc_graph(ctx->alloc, gf) + tensor_alignment; - LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: compute buffer total size = %.2f MiB\n", __func__, (ctx->buf_compute.size + alloc_size) / 1024.0 / 1024.0); // recreate allocator with exact memory requirements ggml_allocr_free(ctx->alloc); @@ -9046,7 +8612,7 @@ struct llama_context * llama_new_context_with_model( #endif #ifdef GGML_USE_CUBLAS ggml_cuda_set_scratch_size(alloc_size); - LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MB\n", __func__, alloc_size / 1024.0 / 1024.0); + LLAMA_LOG_INFO("%s: VRAM scratch buffer: %.2f MiB\n", __func__, alloc_size / 1024.0 / 1024.0); // calculate total VRAM usage auto add_tensor = [](const ggml_tensor * t, size_t & size) { @@ -9066,10 +8632,10 @@ struct llama_context * llama_new_context_with_model( size_t ctx_vram_size = alloc_size + kv_vram_size; size_t total_vram_size = model_vram_size + ctx_vram_size; - LLAMA_LOG_INFO("%s: total VRAM used: %.2f MB (model: %.2f MB, context: %.2f MB)\n", __func__, + LLAMA_LOG_INFO("%s: total VRAM used: %.2f MiB (model: %.2f MiB, context: %.2f MiB)\n", __func__, total_vram_size / 1024.0 / 1024.0, model_vram_size / 1024.0 / 1024.0, - ctx_vram_size / 1024.0 / 1024.0); + ctx_vram_size / 1024.0 / 1024.0); #endif } @@ -9090,7 +8656,7 @@ struct llama_context * llama_new_context_with_model( const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx); - LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0); + LLAMA_LOG_INFO("%s: max tensor size = %8.2f MiB\n", __func__, max_size/1024.0/1024.0); #define LLAMA_METAL_CHECK_BUF(result) \ if (!(result)) { \ @@ -9156,6 +8722,45 @@ float llama_rope_freq_scale_train(const struct llama_model * model) { return model->hparams.rope_freq_scale_train; } +int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size) { + const auto & it = model->gguf_kv.find(key); + if (it == model->gguf_kv.end()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + +int llama_model_meta_count(const struct llama_model * model) { + return (int)model->gguf_kv.size(); +} + +int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->first.c_str()); +} + +int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size) { + if (i < 0 || i >= (int)model->gguf_kv.size()) { + if (buf_size > 0) { + buf[0] = '\0'; + } + return -1; + } + auto it = model->gguf_kv.begin(); + std::advance(it, i); + return snprintf(buf, buf_size, "%s", it->second.c_str()); +} + int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size) { return snprintf(buf, buf_size, "%s %s %s", llama_model_arch_name(model->arch).c_str(), @@ -9214,8 +8819,107 @@ int llama_model_apply_lora_from_file(const struct llama_model * model, const cha } } +struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq) { + struct llama_kv_cache_view result = { + /*.n_cells = */ 0, + /*.n_max_seq = */ n_max_seq, + /*.token_count = */ 0, + /*.used_cells = */ llama_get_kv_cache_used_cells(ctx), + /*.max_contiguous = */ 0, + /*.max_contiguous_idx = */ -1, + /*.cells = */ nullptr, + /*.cells_sequences = */ nullptr, + }; + return result; +} + +void llama_kv_cache_view_free(struct llama_kv_cache_view * view) { + if (view->cells != nullptr) { + free(view->cells); + view->cells = nullptr; + } + if (view->cells_sequences != nullptr) { + free(view->cells_sequences); + view->cells_sequences = nullptr; + } +} + +void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view) { + if (uint32_t(view->n_cells) < ctx->kv_self.size || view->cells == nullptr) { + view->n_cells = int32_t(ctx->kv_self.size); + void * p = realloc(view->cells, sizeof(struct llama_kv_cache_view_cell) * view->n_cells); + GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells"); + view->cells = (struct llama_kv_cache_view_cell *)p; + p = realloc(view->cells_sequences, sizeof(llama_seq_id) * view->n_max_seq * view->n_cells); + GGML_ASSERT(p != nullptr && "Failed to alloc kv_cache_view cells sequences"); + view->cells_sequences = (llama_seq_id *)p; + } + + const std::vector & kv_cells = ctx->kv_self.cells; + llama_kv_cache_view_cell * c_curr = view->cells; + llama_seq_id * cs_curr = view->cells_sequences; + int32_t used_cells = 0; + int32_t token_count = 0; + int32_t curr_contig_idx = -1; + uint32_t max_contig = 0; + int32_t max_contig_idx = -1; + + for (int32_t i = 0; i < int32_t(ctx->kv_self.size); i++, c_curr++, cs_curr += view->n_max_seq) { + const size_t curr_size = kv_cells[i].seq_id.size(); + token_count += curr_size; + c_curr->pos = kv_cells[i].pos + kv_cells[i].delta; + + if (curr_size > 0) { + if (curr_contig_idx >= 0 && uint32_t(i - curr_contig_idx) > max_contig) { + max_contig = i - curr_contig_idx; + max_contig_idx = curr_contig_idx; + } + curr_contig_idx = -1; + } else if (curr_contig_idx < 0) { + curr_contig_idx = i; + } + + int seq_idx = 0; + for (const llama_seq_id it : kv_cells[i].seq_id) { + if (seq_idx >= view->n_max_seq) { + break; + } + cs_curr[seq_idx] = it; + seq_idx++; + } + if (seq_idx != 0) { + used_cells++; + } + for (; seq_idx < view->n_max_seq; seq_idx++) { + cs_curr[seq_idx] = -1; + } + } + if (curr_contig_idx >= 0 && kv_cells.size() - curr_contig_idx > max_contig) { + max_contig_idx = curr_contig_idx; + max_contig = kv_cells.size() - curr_contig_idx; + } + view->max_contiguous = max_contig; + view->max_contiguous_idx = max_contig_idx; + view->token_count = token_count; + view->used_cells = used_cells; + if (uint32_t(used_cells) != ctx->kv_self.used) { + LLAMA_LOG_ERROR("%s: used cells mismatch. kv_cache says %d but we calculated %d\n", + __func__, ctx->kv_self.used, used_cells); + } +} + int llama_get_kv_cache_token_count(const struct llama_context * ctx) { - return ctx->kv_self.head; + int result = 0; + + for (uint32_t i = 0; i < ctx->kv_self.size; i++) { + result += ctx->kv_self.cells[i].seq_id.size(); + } + + return result; +} + +int llama_get_kv_cache_used_cells(const struct llama_context * ctx) { + return ctx->kv_self.used; } void llama_kv_cache_clear(struct llama_context * ctx) { @@ -9385,16 +9089,18 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat const size_t kv_buf_size = kv_self.buf.size; const uint32_t kv_head = kv_self.head; const uint32_t kv_size = kv_self.size; + const uint32_t kv_used = kv_self.used; data_ctx->write(&kv_buf_size, sizeof(kv_buf_size)); data_ctx->write(&kv_head, sizeof(kv_head)); data_ctx->write(&kv_size, sizeof(kv_size)); + data_ctx->write(&kv_used, sizeof(kv_used)); if (kv_buf_size) { const size_t elt_size = ggml_element_size(kv_self.k); - ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); - ggml_cgraph gf{}; + ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_cgraph * gf = ggml_new_graph(cpy_ctx); ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); std::vector kout3d_data(ggml_nbytes(kout3d), 0); @@ -9412,9 +9118,9 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat kv_head, n_embd, n_layer, elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d)); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d)); - ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d)); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d)); + ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); ggml_free(cpy_ctx); @@ -9511,18 +9217,20 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { size_t kv_buf_size; uint32_t kv_head; uint32_t kv_size; + uint32_t kv_used; memcpy(&kv_buf_size, inp, sizeof(kv_buf_size)); inp += sizeof(kv_buf_size); memcpy(&kv_head, inp, sizeof(kv_head)); inp += sizeof(kv_head); memcpy(&kv_size, inp, sizeof(kv_size)); inp += sizeof(kv_size); + memcpy(&kv_used, inp, sizeof(kv_used)); inp += sizeof(kv_used); if (kv_buf_size) { GGML_ASSERT(kv_self.buf.size == kv_buf_size); const size_t elt_size = ggml_element_size(kv_self.k); - ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); - ggml_cgraph gf{}; + ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true }); + ggml_cgraph * gf = ggml_new_graph(cpy_ctx); ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer); kin3d->data = (void *) inp; @@ -9540,15 +9248,16 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) { kv_head, n_embd, n_layer, elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d)); - ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d)); - ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d)); + ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d)); + ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1); ggml_free(cpy_ctx); } ctx->kv_self.head = kv_head; ctx->kv_self.size = kv_size; + ctx->kv_self.used = kv_used; ctx->kv_self.cells.resize(kv_size); @@ -9797,6 +9506,14 @@ llama_token llama_token_nl(const struct llama_model * model) { return model->vocab.linefeed_id; } +int llama_add_bos_token(const struct llama_model * model) { + return model->vocab.special_add_bos; +} + +int llama_add_eos_token(const struct llama_model * model) { + return model->vocab.special_add_eos; +} + llama_token llama_token_prefix(const struct llama_model * model) { return model->vocab.special_prefix_id; } @@ -10003,6 +9720,9 @@ const std::vector> & llama_internal void llama_log_set(ggml_log_callback log_callback, void * user_data) { g_state.log_callback = log_callback ? log_callback : llama_log_callback_default; g_state.log_callback_user_data = user_data; +#ifdef GGML_USE_METAL + ggml_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data); +#endif } static void llama_log_internal_v(ggml_log_level level, const char * format, va_list args) { diff --git a/llama.h b/llama.h index d727dbd9f..1a62058d1 100644 --- a/llama.h +++ b/llama.h @@ -106,6 +106,14 @@ extern "C" { LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file }; + enum llama_rope_scaling_type { + LLAMA_ROPE_SCALING_UNSPECIFIED = -1, + LLAMA_ROPE_SCALING_NONE = 0, + LLAMA_ROPE_SCALING_LINEAR = 1, + LLAMA_ROPE_SCALING_YARN = 2, + LLAMA_ROPE_SCALING_MAX_VALUE = LLAMA_ROPE_SCALING_YARN, + }; + typedef struct llama_token_data { llama_token id; // token id float logit; // log-odds of the token @@ -167,15 +175,21 @@ extern "C" { }; struct llama_context_params { - uint32_t seed; // RNG seed, -1 for random - uint32_t n_ctx; // text context, 0 = from model - uint32_t n_batch; // prompt processing maximum batch size - uint32_t n_threads; // number of threads to use for generation - uint32_t n_threads_batch; // number of threads to use for batch processing + uint32_t seed; // RNG seed, -1 for random + uint32_t n_ctx; // text context, 0 = from model + uint32_t n_batch; // prompt processing maximum batch size + uint32_t n_threads; // number of threads to use for generation + uint32_t n_threads_batch; // number of threads to use for batch processing + int8_t rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type` // ref: https://github.com/ggerganov/llama.cpp/pull/2054 - float rope_freq_base; // RoPE base frequency, 0 = from model - float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model + float rope_freq_base; // RoPE base frequency, 0 = from model + float rope_freq_scale; // RoPE frequency scaling factor, 0 = from model + float yarn_ext_factor; // YaRN extrapolation mix factor, NaN = from model + float yarn_attn_factor; // YaRN magnitude scaling factor + float yarn_beta_fast; // YaRN low correction dim + float yarn_beta_slow; // YaRN high correction dim + uint32_t yarn_orig_ctx; // YaRN original context size // Keep the booleans together to avoid misalignment during copy-by-value. bool mul_mat_q; // if true, use experimental mul_mat_q kernels (DEPRECATED - always true) @@ -287,6 +301,23 @@ extern "C" { // Get the model's RoPE frequency scaling factor LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model); + // Functions to access the model's GGUF metadata scalar values + // - The functions return the length of the string on success, or -1 on failure + // - The output string is always null-terminated and cleared on failure + // - GGUF array values are not supported by these functions + + // Get metadata value as a string by key name + LLAMA_API int llama_model_meta_val_str(const struct llama_model * model, const char * key, char * buf, size_t buf_size); + + // Get the number of metadata key/value pairs + LLAMA_API int llama_model_meta_count(const struct llama_model * model); + + // Get metadata key name by index + LLAMA_API int llama_model_meta_key_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); + + // Get metadata value as a string by index + LLAMA_API int llama_model_meta_val_str_by_index(const struct llama_model * model, int i, char * buf, size_t buf_size); + // Get a string describing the model type LLAMA_API int llama_model_desc(const struct llama_model * model, char * buf, size_t buf_size); @@ -330,9 +361,60 @@ extern "C" { // KV cache // - // Returns the number of tokens in the KV cache - LLAMA_API DEPRECATED(int llama_get_kv_cache_token_count(const struct llama_context * ctx), - "avoid using this, it will be removed in the future, instead - count the tokens in user code"); + // Information associated with an individual cell in the KV cache view. + struct llama_kv_cache_view_cell { + // The position for this cell. Takes KV cache shifts into account. + // May be negative if the cell is not populated. + llama_pos pos; + }; + + // An updateable view of the KV cache. + struct llama_kv_cache_view { + // Number of KV cache cells. This will be the same as the context size. + int32_t n_cells; + + // Maximum number of sequences that can exist in a cell. It's not an error + // if there are more sequences in a cell than this value, however they will + // not be visible in the view cells_sequences. + int32_t n_max_seq; + + // Number of tokens in the cache. For example, if there are two populated + // cells, the first with 1 sequence id in it and the second with 2 sequence + // ids then you'll have 3 tokens. + int32_t token_count; + + // Number of populated cache cells. + int32_t used_cells; + + // Maximum contiguous empty slots in the cache. + int32_t max_contiguous; + + // Index to the start of the max_contiguous slot range. Can be negative + // when cache is full. + int32_t max_contiguous_idx; + + // Information for an individual cell. + struct llama_kv_cache_view_cell * cells; + + // The sequences for each cell. There will be n_max_seq items per cell. + llama_seq_id * cells_sequences; + }; + + // Create an empty KV cache view. (use only for debugging purposes) + LLAMA_API struct llama_kv_cache_view llama_kv_cache_view_init(const struct llama_context * ctx, int32_t n_max_seq); + + // Free a KV cache view. (use only for debugging purposes) + LLAMA_API void llama_kv_cache_view_free(struct llama_kv_cache_view * view); + + // Update the KV cache view structure with the current state of the KV cache. (use only for debugging purposes) + LLAMA_API void llama_kv_cache_view_update(const struct llama_context * ctx, struct llama_kv_cache_view * view); + + // Returns the number of tokens in the KV cache (slow, use only for debug) + // If a KV cell has multiple sequences assigned to it, it will be counted multiple times + LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx); + + // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) + LLAMA_API int llama_get_kv_cache_used_cells(const struct llama_context * ctx); // Clear the KV cache LLAMA_API void llama_kv_cache_clear( @@ -503,6 +585,12 @@ extern "C" { LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line + // Returns -1 if unknown, 1 for true or 0 for false. + LLAMA_API int llama_add_bos_token(const struct llama_model * model); + + // Returns -1 if unknown, 1 for true or 0 for false. + LLAMA_API int llama_add_eos_token(const struct llama_model * model); + // codellama infill tokens LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle @@ -598,6 +686,13 @@ extern "C" { float p, size_t min_keep); + /// @details Minimum P sampling as described in https://github.com/ggerganov/llama.cpp/pull/3841 + LLAMA_API void llama_sample_min_p( + struct llama_context * ctx, + llama_token_data_array * candidates, + float p, + size_t min_keep); + /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. LLAMA_API void llama_sample_tail_free( struct llama_context * ctx, diff --git a/models/ggml-vocab-llama.gguf b/models/ggml-vocab-llama.gguf index 63bfaf672..549eed8c5 100644 Binary files a/models/ggml-vocab-llama.gguf and b/models/ggml-vocab-llama.gguf differ diff --git a/models/ggml-vocab-stablelm-3b-4e1t.gguf b/models/ggml-vocab-stablelm-3b-4e1t.gguf new file mode 100644 index 000000000..ebb0cdb7d Binary files /dev/null and b/models/ggml-vocab-stablelm-3b-4e1t.gguf differ diff --git a/mypy.ini b/mypy.ini index 55c168f2d..7215a05dd 100644 --- a/mypy.ini +++ b/mypy.ini @@ -3,3 +3,4 @@ strict = true allow_untyped_calls = true allow_untyped_defs = true allow_incomplete_defs = true +disable_error_code = import-untyped diff --git a/scripts/build-info.cmake b/scripts/build-info.cmake index c86ab4379..73853dfa4 100644 --- a/scripts/build-info.cmake +++ b/scripts/build-info.cmake @@ -1,5 +1,5 @@ -set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in") -set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h") +set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in") +set(OUTPUT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp") set(BUILD_NUMBER 0) set(BUILD_COMMIT "unknown") set(BUILD_COMPILER "unknown") @@ -24,15 +24,21 @@ if(Git_FOUND) WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} OUTPUT_VARIABLE HEAD OUTPUT_STRIP_TRAILING_WHITESPACE + RESULT_VARIABLE RES ) + if (RES EQUAL 0) + set(BUILD_COMMIT ${HEAD}) + endif() execute_process( COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} OUTPUT_VARIABLE COUNT OUTPUT_STRIP_TRAILING_WHITESPACE + RESULT_VARIABLE RES ) - set(BUILD_COMMIT ${HEAD}) - set(BUILD_NUMBER ${COUNT}) + if (RES EQUAL 0) + set(BUILD_NUMBER ${COUNT}) + endif() endif() if(MSVC) @@ -53,22 +59,22 @@ else() set(BUILD_TARGET ${OUT}) endif() -# Only write the header if it's changed to prevent unnecessary recompilation -if(EXISTS ${HEADER_FILE}) - file(READ ${HEADER_FILE} CONTENTS) - string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS}) +# Only write the build info if it changed +if(EXISTS ${OUTPUT_FILE}) + file(READ ${OUTPUT_FILE} CONTENTS) + string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS}) set(OLD_COMMIT ${CMAKE_MATCH_1}) - string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS}) + string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS}) set(OLD_COMPILER ${CMAKE_MATCH_1}) - string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS}) + string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS}) set(OLD_TARGET ${CMAKE_MATCH_1}) if ( NOT OLD_COMMIT STREQUAL BUILD_COMMIT OR NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR NOT OLD_TARGET STREQUAL BUILD_TARGET ) - configure_file(${TEMPLATE_FILE} ${HEADER_FILE}) + configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE}) endif() else() - configure_file(${TEMPLATE_FILE} ${HEADER_FILE}) + configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE}) endif() diff --git a/scripts/build-info.h.in b/scripts/build-info.h.in deleted file mode 100644 index e996faef0..000000000 --- a/scripts/build-info.h.in +++ /dev/null @@ -1,9 +0,0 @@ -#ifndef BUILD_INFO_H -#define BUILD_INFO_H - -#define BUILD_NUMBER @BUILD_NUMBER@ -#define BUILD_COMMIT "@BUILD_COMMIT@" -#define BUILD_COMPILER "@BUILD_COMPILER@" -#define BUILD_TARGET "@BUILD_TARGET@" - -#endif // BUILD_INFO_H diff --git a/scripts/build-info.sh b/scripts/build-info.sh index 3c8b1fb85..32682afbd 100755 --- a/scripts/build-info.sh +++ b/scripts/build-info.sh @@ -24,12 +24,7 @@ if out=$($CC -dumpmachine); then build_target=$out fi -echo "#ifndef BUILD_INFO_H" -echo "#define BUILD_INFO_H" -echo -echo "#define BUILD_NUMBER $build_number" -echo "#define BUILD_COMMIT \"$build_commit\"" -echo "#define BUILD_COMPILER \"$build_compiler\"" -echo "#define BUILD_TARGET \"$build_target\"" -echo -echo "#endif // BUILD_INFO_H" +echo "int LLAMA_BUILD_NUMBER = ${build_number};" +echo "char const *LLAMA_COMMIT = \"${build_commit}\";" +echo "char const *LLAMA_COMPILER = \"${build_compiler}\";" +echo "char const *LLAMA_BUILD_TARGET = \"${build_target}\";" diff --git a/scripts/server-llm.sh b/scripts/server-llm.sh new file mode 100644 index 000000000..7bf0929bb --- /dev/null +++ b/scripts/server-llm.sh @@ -0,0 +1,391 @@ +#!/bin/bash +# +# Helper script for deploying llama.cpp server with a single Bash command +# +# - Works on Linux and macOS +# - Supports: CPU, CUDA, Metal, OpenCL +# - Can run all GGUF models from HuggingFace +# - Can serve requests in parallel +# - Always builds latest llama.cpp from GitHub +# +# Limitations +# +# - Chat templates are poorly supported (base models recommended) +# - Might be unstable! +# +# Usage: +# ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose] +# +# --port: port number, default is 8888 +# --repo: path to a repo containing GGUF model files +# --wtype: weights type (f16, q8_0, q4_0, q4_1), default is user-input +# --backend: cpu, cuda, metal, opencl, depends on the OS +# --gpu-id: gpu id, default is 0 +# --n-parallel: number of parallel requests, default is 8 +# --n-kv: KV cache size, default is 4096 +# --verbose: verbose output +# +# Example: +# +# bash -c "$(curl -s https://ggml.ai/server-llm.sh)" +# + +set -e + +# required utils: curl, git, make +if ! command -v curl &> /dev/null; then + printf "[-] curl not found\n" + exit 1 +fi +if ! command -v git &> /dev/null; then + printf "[-] git not found\n" + exit 1 +fi +if ! command -v make &> /dev/null; then + printf "[-] make not found\n" + exit 1 +fi + +# parse arguments +port=8888 +repo="" +wtype="" +backend="cpu" + +# if macOS, use metal backend by default +if [[ "$OSTYPE" == "darwin"* ]]; then + backend="metal" +elif command -v nvcc &> /dev/null; then + backend="cuda" +fi + +gpu_id=0 +n_parallel=8 +n_kv=4096 +verbose=0 + +function print_usage { + printf "Usage:\n" + printf " ./server-llm.sh [--port] [--repo] [--wtype] [--backend] [--gpu-id] [--n-parallel] [--n-kv] [--verbose]\n\n" + printf " --port: port number, default is 8888\n" + printf " --repo: path to a repo containing GGUF model files\n" + printf " --wtype: weights type (f16, q8_0, q4_0, q4_1), default is user-input\n" + printf " --backend: cpu, cuda, metal, opencl, depends on the OS\n" + printf " --gpu-id: gpu id, default is 0\n" + printf " --n-parallel: number of parallel requests, default is 8\n" + printf " --n-kv: KV cache size, default is 4096\n" + printf " --verbose: verbose output\n\n" + printf "Example:\n\n" + printf ' bash -c "$(curl -s https://ggml.ai/server-llm.sh)"\n\n' +} + +while [[ $# -gt 0 ]]; do + key="$1" + case $key in + --port) + port="$2" + shift + shift + ;; + --repo) + repo="$2" + shift + shift + ;; + --wtype) + wtype="$2" + shift + shift + ;; + --backend) + backend="$2" + shift + shift + ;; + --gpu-id) + gpu_id="$2" + shift + shift + ;; + --n-parallel) + n_parallel="$2" + shift + shift + ;; + --n-kv) + n_kv="$2" + shift + shift + ;; + --verbose) + verbose=1 + shift + ;; + --help) + print_usage + exit 0 + ;; + *) + echo "Unknown argument: $key" + print_usage + exit 1 + ;; + esac +done + +# available weights types +wtypes=("F16" "Q8_0" "Q4_0" "Q4_1" "Q5_0" "Q5_1" "Q6_K" "Q5_K_M" "Q5_K_S" "Q4_K_M" "Q4_K_S" "Q3_K_L" "Q3_K_M" "Q3_K_S" "Q2_K") + +wfiles=() +for wt in "${wtypes[@]}"; do + wfiles+=("") +done + +# sample repos +repos=( + "https://huggingface.co/TheBloke/Llama-2-7B-GGUF" + "https://huggingface.co/TheBloke/Llama-2-13B-GGUF" + "https://huggingface.co/TheBloke/Llama-2-70B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-7B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-13B-GGUF" + "https://huggingface.co/TheBloke/CodeLlama-34B-GGUF" + "https://huggingface.co/TheBloke/Mistral-7B-v0.1-GGUF" + "https://huggingface.co/TheBloke/zephyr-7B-beta-GGUF" + "https://huggingface.co/TheBloke/OpenHermes-2-Mistral-7B-GGUF" + "https://huggingface.co/TheBloke/CausalLM-7B-GGUF" +) + +printf "\n" +printf "[I] This is a helper script for deploying llama.cpp's server on this machine.\n\n" +printf " Based on the options that follow, the script might download a model file\n" +printf " from the internet, which can be a few GBs in size. The script will also\n" +printf " build the latest llama.cpp source code from GitHub, which can be unstable.\n" +printf "\n" +printf " Upon success, an HTTP server will be started and it will serve the selected\n" +printf " model using llama.cpp for demonstration purposes.\n" +printf "\n" +printf " Please note:\n" +printf "\n" +printf " - All new data will be stored in the current folder\n" +printf " - The server will be listening on all network interfaces\n" +printf " - The server will run with default settings which are not always optimal\n" +printf " - Do not judge the quality of a model based on the results from this script\n" +printf " - Do not use this script to benchmark llama.cpp\n" +printf " - Do not use this script in production\n" +printf " - This script is only for demonstration purposes\n" +printf "\n" +printf " If you don't know what you are doing, please press Ctrl-C to abort now\n" +printf "\n" +printf " Press Enter to continue ...\n\n" + +read + +if [[ -z "$repo" ]]; then + printf "[+] No repo provided from the command line\n" + printf " Please select a number from the list below or enter an URL:\n\n" + + is=0 + for r in "${repos[@]}"; do + printf " %2d) %s\n" $is "$r" + is=$((is+1)) + done + + # ask for repo until index of sample repo is provided or an URL + while [[ -z "$repo" ]]; do + printf "\n Or choose one from: https://huggingface.co/models?sort=trending&search=gguf\n\n" + read -p "[+] Select repo: " repo + + # check if the input is a number + if [[ "$repo" =~ ^[0-9]+$ ]]; then + if [[ "$repo" -ge 0 && "$repo" -lt ${#repos[@]} ]]; then + repo="${repos[$repo]}" + else + printf "[-] Invalid repo index: %s\n" "$repo" + repo="" + fi + elif [[ "$repo" =~ ^https?:// ]]; then + repo="$repo" + else + printf "[-] Invalid repo URL: %s\n" "$repo" + repo="" + fi + done +fi + +# remove suffix +repo=$(echo "$repo" | sed -E 's/\/tree\/main$//g') + +printf "[+] Checking for GGUF model files in %s\n" "$repo" + +# find GGUF files in the source +# TODO: better logic +model_tree="${repo%/}/tree/main" +model_files=$(curl -s "$model_tree" | grep -i "\\.gguf" | sed -E 's/.*(.*)<\/span><\/a>/\1/g') + +# list all files in the provided git repo +printf "[+] Model files:\n\n" +for file in $model_files; do + # determine iw by grepping the filename with wtypes + iw=-1 + is=0 + for wt in "${wtypes[@]}"; do + # uppercase + ufile=$(echo "$file" | tr '[:lower:]' '[:upper:]') + if [[ "$ufile" =~ "$wt" ]]; then + iw=$is + break + fi + is=$((is+1)) + done + + if [[ $iw -eq -1 ]]; then + continue + fi + + wfiles[$iw]="$file" + + have=" " + if [[ -f "$file" ]]; then + have="*" + fi + + printf " %2d) %s %s\n" $iw "$have" "$file" +done + +# ask for weights type until provided and available +while [[ -z "$wtype" ]]; do + printf "\n" + read -p "[+] Select weight type: " wtype + wfile="${wfiles[$wtype]}" + + if [[ -z "$wfile" ]]; then + printf "[-] Invalid weight type: %s\n" "$wtype" + wtype="" + fi +done + +printf "[+] Selected weight type: %s (%s)\n" "$wtype" "$wfile" + +url="${repo%/}/resolve/main/$wfile" + +# check file if the model has been downloaded before +chk="$wfile.chk" + +# check if we should download the file +# - if $wfile does not exist +# - if $wfile exists but $chk does not exist +# - if $wfile exists and $chk exists but $wfile is newer than $chk +# TODO: better logic using git lfs info + +do_download=0 + +if [[ ! -f "$wfile" ]]; then + do_download=1 +elif [[ ! -f "$chk" ]]; then + do_download=1 +elif [[ "$wfile" -nt "$chk" ]]; then + do_download=1 +fi + +if [[ $do_download -eq 1 ]]; then + printf "[+] Downloading weights from %s\n" "$url" + + # download the weights file + curl -o "$wfile" -# -L "$url" + + # create a check file if successful + if [[ $? -eq 0 ]]; then + printf "[+] Creating check file %s\n" "$chk" + touch "$chk" + fi +else + printf "[+] Using cached weights %s\n" "$wfile" +fi + +# get latest llama.cpp and build + +printf "[+] Downloading latest llama.cpp\n" + +llama_cpp_dir="__llama_cpp_port_${port}__" + +if [[ -d "$llama_cpp_dir" && ! -f "$llama_cpp_dir/__ggml_script__" ]]; then + # if the dir exists and there isn't a file "__ggml_script__" in it, abort + printf "[-] Directory %s already exists\n" "$llama_cpp_dir" + printf "[-] Please remove it and try again\n" + exit 1 +elif [[ -d "$llama_cpp_dir" ]]; then + printf "[+] Directory %s already exists\n" "$llama_cpp_dir" + printf "[+] Using cached llama.cpp\n" + + cd "$llama_cpp_dir" + git reset --hard + git fetch + git checkout origin/master + + cd .. +else + printf "[+] Cloning llama.cpp\n" + + git clone https://github.com/ggerganov/llama.cpp "$llama_cpp_dir" +fi + +# mark that that the directory is made by this script +touch "$llama_cpp_dir/__ggml_script__" + +if [[ $verbose -eq 1 ]]; then + set -x +fi + +# build +cd "$llama_cpp_dir" + +make clean + +log="--silent" +if [[ $verbose -eq 1 ]]; then + log="" +fi + +if [[ "$backend" == "cuda" ]]; then + printf "[+] Building with CUDA backend\n" + LLAMA_CUBLAS=1 make -j server $log +elif [[ "$backend" == "cpu" ]]; then + printf "[+] Building with CPU backend\n" + make -j server $log +elif [[ "$backend" == "metal" ]]; then + printf "[+] Building with Metal backend\n" + make -j server $log +elif [[ "$backend" == "opencl" ]]; then + printf "[+] Building with OpenCL backend\n" + LLAMA_CLBLAST=1 make -j server $log +else + printf "[-] Unknown backend: %s\n" "$backend" + exit 1 +fi + +# run the server + +printf "[+] Running server\n" + +args="" +if [[ "$backend" == "cuda" ]]; then + export CUDA_VISIBLE_DEVICES=$gpu_id + args="-ngl 999" +elif [[ "$backend" == "cpu" ]]; then + args="-ngl 0" +elif [[ "$backend" == "metal" ]]; then + args="-ngl 999" +elif [[ "$backend" == "opencl" ]]; then + args="-ngl 999" +else + printf "[-] Unknown backend: %s\n" "$backend" + exit 1 +fi + +if [[ $verbose -eq 1 ]]; then + args="$args --verbose" +fi + +./server -m "../$wfile" --host 0.0.0.0 --port "$port" -c $n_kv -np "$n_parallel" $args + +exit 0 diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh index 4311268bd..4024531b1 100755 --- a/scripts/sync-ggml.sh +++ b/scripts/sync-ggml.sh @@ -2,14 +2,20 @@ cp -rpv ../ggml/src/ggml.c ./ggml.c cp -rpv ../ggml/src/ggml-alloc.c ./ggml-alloc.c +cp -rpv ../ggml/src/ggml-backend-impl.h ./ggml-backend-impl.h cp -rpv ../ggml/src/ggml-backend.c ./ggml-backend.c -cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h cp -rpv ../ggml/src/ggml-cuda.cu ./ggml-cuda.cu -cp -rpv ../ggml/src/ggml-opencl.h ./ggml-opencl.h -cp -rpv ../ggml/src/ggml-opencl.cpp ./ggml-opencl.cpp +cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h +cp -rpv ../ggml/src/ggml-impl.h ./ggml-impl.h cp -rpv ../ggml/src/ggml-metal.h ./ggml-metal.h cp -rpv ../ggml/src/ggml-metal.m ./ggml-metal.m cp -rpv ../ggml/src/ggml-metal.metal ./ggml-metal.metal +cp -rpv ../ggml/src/ggml-mpi.h ./ggml-mpi.h +cp -rpv ../ggml/src/ggml-mpi.c ./ggml-mpi.c +cp -rpv ../ggml/src/ggml-opencl.cpp ./ggml-opencl.cpp +cp -rpv ../ggml/src/ggml-opencl.h ./ggml-opencl.h +cp -rpv ../ggml/src/ggml-quants.c ./ggml-quants.c +cp -rpv ../ggml/src/ggml-quants.h ./ggml-quants.h cp -rpv ../ggml/include/ggml/ggml.h ./ggml.h cp -rpv ../ggml/include/ggml/ggml-alloc.h ./ggml-alloc.h cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index 6757ad1cc..c8b4bc254 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -33,9 +33,11 @@ llama_build_executable(test-tokenizer-1-bpe.cpp) llama_test_executable (test-tokenizer-1-falcon test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-falcon.gguf) llama_test_executable(test-tokenizer-1-aquila test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf) llama_test_executable(test-tokenizer-1-mpt test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-mpt.gguf) +llama_test_executable(test-tokenizer-1-stablelm-3b-4e1t test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-stablelm-3b-4e1t.gguf) llama_test_executable(test-tokenizer-1-gpt-neox test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-gpt-neox.gguf) llama_test_executable(test-tokenizer-1-refact test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-refact.gguf) llama_test_executable(test-tokenizer-1-starcoder test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf) +# llama_test_executable(test-tokenizer-1-bloom test-tokenizer-1-bpe.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bloom.gguf) # BIG llama_build_and_test_executable(test-grammar-parser.cpp) llama_build_and_test_executable(test-llama-grammar.cpp) llama_build_and_test_executable(test-grad0.cpp) # SLOW diff --git a/tests/test-grad0.cpp b/tests/test-grad0.cpp index 0a559b27a..7fe9154dd 100644 --- a/tests/test-grad0.cpp +++ b/tests/test-grad0.cpp @@ -231,9 +231,10 @@ static bool check_gradient( printf("GGML_N_THREADS = %d\n", n_threads); } - struct ggml_cgraph * gf = ggml_build_forward_ctx(ctx0, f); - struct ggml_cgraph * gb = ggml_new_graph(ctx0); - *gb = *gf; + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, GGML_DEFAULT_GRAPH_SIZE, true); + struct ggml_cgraph * gb = ggml_new_graph_custom(ctx0, GGML_DEFAULT_GRAPH_SIZE, true); + ggml_build_forward_expand(gf, f); + ggml_graph_cpy(gf, gb); ggml_build_backward_expand(ctx0, gf, gb, false); ggml_graph_compute_with_ctx(ctx0, gf, n_threads); diff --git a/tests/test-opt.cpp b/tests/test-opt.cpp index bb8af5962..2c9997fca 100644 --- a/tests/test-opt.cpp +++ b/tests/test-opt.cpp @@ -109,10 +109,11 @@ int main(void) { struct ggml_tensor * d = ggml_sub(ctx, c, ab); struct ggml_tensor * e = ggml_sum(ctx, ggml_sqr(ctx, d)); - struct ggml_cgraph ge = ggml_build_forward(e); - ggml_graph_reset(&ge); + struct ggml_cgraph * ge = ggml_new_graph_custom(ctx, GGML_DEFAULT_GRAPH_SIZE, true); + ggml_build_forward_expand(ge, e); + ggml_graph_reset(ge); - ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); + ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1); const float fe = ggml_get_f32_1d(e, 0); printf("%s: e = %.4f\n", __func__, fe); @@ -121,9 +122,9 @@ int main(void) { ggml_opt(ctx, opt_params, e); - ggml_graph_reset(&ge); + ggml_graph_reset(ge); - ggml_graph_compute_with_ctx(ctx, &ge, /*n_threads*/ 1); + ggml_graph_compute_with_ctx(ctx, ge, /*n_threads*/ 1); const float fe_opt = ggml_get_f32_1d(e, 0); printf("%s: original e = %.4f\n", __func__, fe); diff --git a/tests/test-tokenizer-0-falcon.py b/tests/test-tokenizer-0-falcon.py index cf65a3f65..4f06ec9bb 100644 --- a/tests/test-tokenizer-0-falcon.py +++ b/tests/test-tokenizer-0-falcon.py @@ -1,7 +1,5 @@ # tests with BPE tokenizer -import os -import sys import argparse from transformers import AutoTokenizer @@ -16,34 +14,34 @@ dir_tokenizer = args.dir_tokenizer tokenizer = AutoTokenizer.from_pretrained(dir_tokenizer) tests = [ - "", - " ", - " ", - " ", - "\t", - "\n", - "\t\n", - "Hello world", - " Hello world", - "Hello World", - " Hello World", - " Hello World!", - "Hello, world!", - " Hello, world!", - " this is πŸ¦™.cpp", - "w048 7tuijk dsdfhu", - "Π½Π΅Ρ‰ΠΎ Π½Π° Π‘ΡŠΠ»Π³Π°Ρ€ΡΠΊΠΈ", - "αž€αžΆαž“αŸ‹αžαŸ‚αž–αž·αžŸαŸαžŸαž’αžΆαž…αžαž›αž…αŸαž‰", - "πŸš€ (normal) πŸ˜Άβ€πŸŒ«οΈ (multiple emojis concatenated) βœ… (only emoji that has its own token)", - "Hello", - " Hello", - " Hello", - " Hello", - " Hello", - " Hello\n Hello", - "\n =", - "' era", - ] + "", + " ", + " ", + " ", + "\t", + "\n", + "\t\n", + "Hello world", + " Hello world", + "Hello World", + " Hello World", + " Hello World!", + "Hello, world!", + " Hello, world!", + " this is πŸ¦™.cpp", + "w048 7tuijk dsdfhu", + "Π½Π΅Ρ‰ΠΎ Π½Π° Π‘ΡŠΠ»Π³Π°Ρ€ΡΠΊΠΈ", + "αž€αžΆαž“αŸ‹αžαŸ‚αž–αž·αžŸαŸαžŸαž’αžΆαž…αžαž›αž…αŸαž‰", + "πŸš€ (normal) πŸ˜Άβ€πŸŒ«οΈ (multiple emojis concatenated) βœ… (only emoji that has its own token)", + "Hello", + " Hello", + " Hello", + " Hello", + " Hello", + " Hello\n Hello", + "\n =", + "' era", +] for text in tests: print('text: ', text) diff --git a/tests/test-tokenizer-0-llama.py b/tests/test-tokenizer-0-llama.py index 078f680b1..f3d4d7e3d 100644 --- a/tests/test-tokenizer-0-llama.py +++ b/tests/test-tokenizer-0-llama.py @@ -1,7 +1,5 @@ # tests with SPM tokenizer -import os -import sys import argparse from sentencepiece import SentencePieceProcessor @@ -16,32 +14,32 @@ dir_tokenizer = args.dir_tokenizer tokenizer = SentencePieceProcessor(dir_tokenizer + '/tokenizer.model') tests = [ - "", - " ", - " ", - " ", - "\t", - "\n", - "\t\n", - "Hello world", - " Hello world", - "Hello World", - " Hello World", - " Hello World!", - "Hello, world!", - " Hello, world!", - " this is πŸ¦™.cpp", - "w048 7tuijk dsdfhu", - "Π½Π΅Ρ‰ΠΎ Π½Π° Π‘ΡŠΠ»Π³Π°Ρ€ΡΠΊΠΈ", - "αž€αžΆαž“αŸ‹αžαŸ‚αž–αž·αžŸαŸαžŸαž’αžΆαž…αžαž›αž…αŸαž‰", - "πŸš€ (normal) πŸ˜Άβ€πŸŒ«οΈ (multiple emojis concatenated) βœ… (only emoji that has its own token)", - "Hello", - " Hello", - " Hello", - " Hello", - " Hello", - " Hello\n Hello", - ] + "", + " ", + " ", + " ", + "\t", + "\n", + "\t\n", + "Hello world", + " Hello world", + "Hello World", + " Hello World", + " Hello World!", + "Hello, world!", + " Hello, world!", + " this is πŸ¦™.cpp", + "w048 7tuijk dsdfhu", + "Π½Π΅Ρ‰ΠΎ Π½Π° Π‘ΡŠΠ»Π³Π°Ρ€ΡΠΊΠΈ", + "αž€αžΆαž“αŸ‹αžαŸ‚αž–αž·αžŸαŸαžŸαž’αžΆαž…αžαž›αž…αŸαž‰", + "πŸš€ (normal) πŸ˜Άβ€πŸŒ«οΈ (multiple emojis concatenated) βœ… (only emoji that has its own token)", + "Hello", + " Hello", + " Hello", + " Hello", + " Hello", + " Hello\n Hello", +] for text in tests: