Merge remote-tracking branch 'origin/master' into server-cfg

.gitignore

@@ -61,3 +61,18 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
+
+
+pyproject.toml
+poetry.lock
+poetry.toml
+
+# Test binaries
+tests/test-double-float
+tests/test-grad0
+tests/test-opt
+tests/test-quantize-fns
+tests/test-quantize-perf
+tests/test-sampling
+tests/test-tokenizer-0
+

CMakeLists.txt

@@ -186,16 +186,7 @@ if (LLAMA_BLAS)
                 pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "Intel")
                 # all Intel* libraries share the same include path
-                pkg_check_modules(DepBLAS mkl-sdl)
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
-                if (NOT DepBLAS)
-                    if (BUILD_SHARED_LIBS)
-                        set(LINK_METHOD dynamic)
-                    else()
-                        set(LINK_METHOD static)
-                    endif()
-                    string(REGEX REPLACE ".*_" "" DATA_TYPE_MODEL ${LLAMA_BLAS_VENDOR})
-                    pkg_check_modules(DepBLAS REQUIRED mkl-${LINK_METHOD}-${DATA_TYPE_MODEL}-iomp)
-                endif()
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "NVHPC")
                 # this doesn't provide pkg-config
                 # suggest to assign BLAS_INCLUDE_DIRS on your own

Makefile

@@ -1,5 +1,8 @@
 # Define the default target now so that it is always the first target
-BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server libembdinput.so embd-input-test
+BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot train-text-from-scratch simple server embd-input-test
+
+# Binaries only useful for tests
+TEST_TARGETS = tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0
 
 default: $(BUILD_TARGETS)
 
@@ -90,6 +93,28 @@ ifeq ($(UNAME_S),Haiku)
 	CXXFLAGS += -pthread
 endif
 
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# library name prefix
+ifneq ($(_WIN32),1)
+	LIB_PRE := lib
+endif
+
+# Dynamic Shared Object extension
+ifneq ($(_WIN32),1)
+	DSO_EXT := .so
+else
+	DSO_EXT := .dll
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 ifdef LLAMA_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
@@ -102,7 +127,7 @@ endif
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
 	CFLAGS   += -march=native -mtune=native
 	CXXFLAGS += -march=native -mtune=native
@@ -168,8 +193,12 @@ ifdef LLAMA_CUBLAS
 	CXXFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
 	LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
 	OBJS      += ggml-cuda.o
-	NVCC      = nvcc
 	NVCCFLAGS = --forward-unknown-to-host-compiler
+ifdef LLAMA_CUDA_NVCC
+	NVCC = $(LLAMA_CUDA_NVCC)
+else
+	NVCC = nvcc
+endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
 else
@@ -198,19 +227,23 @@ ifdef LLAMA_CUDA_KQUANTS_ITER
 else
 	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
-
+ifdef LLAMA_CUDA_CCBIN
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
-	CFLAGS   += -DGGML_USE_CLBLAST
+
-	CXXFLAGS += -DGGML_USE_CLBLAST
+	CFLAGS   += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+	CXXFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags clblast OpenCL)
+
 	# Mac provides OpenCL as a framework
 	ifeq ($(UNAME_S),Darwin)
 		LDFLAGS += -lclblast -framework OpenCL
 	else
-		LDFLAGS += -lclblast -lOpenCL
+		LDFLAGS += $(shell pkg-config --libs clblast OpenCL)
 	endif
 	OBJS    += ggml-opencl.o
 
@@ -290,17 +323,20 @@ llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+grammar-parser.o: examples/grammar-parser.cpp examples/grammar-parser.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vf *.o *.so main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h
+	rm -vf *.o *.so *.dll main quantize quantize-stats perplexity embedding benchmark-matmult save-load-state server simple vdot train-text-from-scratch embd-input-test build-info.h $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o common.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
@@ -325,14 +361,14 @@ save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
-libembdinput.so: examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+$(LIB_PRE)embdinput$(DSO_EXT): examples/embd-input/embd-input.h examples/embd-input/embd-input-lib.cpp build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) --shared $(CXXFLAGS) $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS)
 
 
-embd-input-test: libembdinput.so examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+embd-input-test: $(LIB_PRE)embdinput$(DSO_EXT) examples/embd-input/embd-input-test.cpp build-info.h ggml.o llama.o common.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.so,$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
+	$(CXX) $(CXXFLAGS) $(filter-out %$(DSO_EXT),$(filter-out %.h,$(filter-out %.hpp,$^))) -o $@ $(LDFLAGS) -L. -lembdinput
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
@@ -349,6 +385,8 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 # Tests
 #
 
+tests: $(TEST_TARGETS)
+
 benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	./$@
@@ -356,6 +394,23 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-.PHONY: tests clean
+tests/test-double-float: tests/test-double-float.c build-info.h ggml.o llama.o common.o $(OBJS)
-tests:
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
-	bash ./tests/run-tests.sh
+
+tests/test-grad0: tests/test-grad0.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-opt: tests/test-opt.c build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)
+
+tests/test-tokenizer-0: tests/test-tokenizer-0.cpp build-info.h ggml.o llama.o common.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.txt,$^) -o $@ $(LDFLAGS)

README.md

@@ -242,6 +242,23 @@ In order to build llama.cpp you have three different options.
     zig build -Doptimize=ReleaseFast
     ```
 
+-   Using `gmake` (FreeBSD):
+
+    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
+    2. Add your user to **video** group
+    3. Install compilation dependencies.
+
+        ```bash
+        sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
+            opencl clblast openblas
+
+            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        ```
+
+    **Notes:** With this packages you can build llama.cpp with OPENBLAS and
+    CLBLAST support for use OpenCL GPU acceleration in FreeBSD. Please read
+    the instructions for use and activate this options in this document below.
+
 ### Metal Build
 
 Using Metal allows the computation to be executed on the GPU for Apple devices:
@@ -360,7 +377,7 @@ Building the program with BLAS support may lead to some performance improvements
   ```bash
   mkdir build
   cd build
-  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_lp64 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
   cmake --build . --config Release
   ```
 
@@ -384,7 +401,7 @@ Building the program with BLAS support may lead to some performance improvements
 
   | Option                  | Legal values           | Default | Description |
   |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 7.0/Turing/RTX 2000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_FORCE_DMMV   | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
   | LLAMA_CUDA_MMV_Y       | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
   | LLAMA_CUDA_DMMV_F16     | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels. Can improve performance on relatively recent GPUs. |

ci/README.md

@@ -16,5 +16,10 @@ It is a good practice, before publishing changes to execute the full CI locally
 
 ```bash
 mkdir tmp
+
+# CPU-only build
 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+
+# with CUDA support
+GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```

ci/run.sh

@@ -1,4 +1,15 @@
 #/bin/bash
+#
+# sample usage:
+#
+# mkdir tmp
+#
+# # CPU-only build
+# bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with CUDA support
+# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -101,7 +112,7 @@ function gg_run_ctest_release {
     (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    if [ -z $GG_BUILD_LOW_PERF ]; then
+    if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
         (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
@@ -154,6 +165,7 @@ function gg_run_open_llama_3b_v2 {
     model_q4_1="${path_models}/ggml-model-q4_1.bin"
     model_q5_0="${path_models}/ggml-model-q5_0.bin"
     model_q5_1="${path_models}/ggml-model-q5_1.bin"
+    model_q2_k="${path_models}/ggml-model-q2_k.bin"
     model_q3_k="${path_models}/ggml-model-q3_k.bin"
     model_q4_k="${path_models}/ggml-model-q4_k.bin"
     model_q5_k="${path_models}/ggml-model-q5_k.bin"
@@ -166,21 +178,23 @@ function gg_run_open_llama_3b_v2 {
     ./bin/quantize ${model_f16} ${model_q4_1} q4_1
     ./bin/quantize ${model_f16} ${model_q5_0} q5_0
     ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
     ./bin/quantize ${model_f16} ${model_q3_k} q3_k
     ./bin/quantize ${model_f16} ${model_q4_k} q4_k
     ./bin/quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
     (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
     (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
@@ -188,6 +202,7 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
     (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
     (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
     (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
     (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
@@ -212,6 +227,7 @@ function gg_run_open_llama_3b_v2 {
     check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
@@ -232,6 +248,133 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
     gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
     gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+}
+
+# open_llama_7b_v2
+# requires: GG_BUILD_CUDA
+
+function gg_run_open_llama_7b_v2 {
+    cd ${SRC}
+
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/tokenizer.model
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/tokenizer_config.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/special_tokens_map.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
+    gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
+
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/open-llama/7B-v2"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.bin"
+    model_q8_0="${path_models}/ggml-model-q8_0.bin"
+    model_q4_0="${path_models}/ggml-model-q4_0.bin"
+    model_q4_1="${path_models}/ggml-model-q4_1.bin"
+    model_q5_0="${path_models}/ggml-model-q5_0.bin"
+    model_q5_1="${path_models}/ggml-model-q5_1.bin"
+    model_q2_k="${path_models}/ggml-model-q2_k.bin"
+    model_q3_k="${path_models}/ggml-model-q3_k.bin"
+    model_q4_k="${path_models}/ggml-model-q4_k.bin"
+    model_q5_k="${path_models}/ggml-model-q5_k.bin"
+    model_q6_k="${path_models}/ggml-model-q6_k.bin"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    set +e
+}
+
+function gg_sum_open_llama_7b_v2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'OpenLLaMA 7B-v2:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
     gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
     gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
     gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
@@ -240,10 +383,10 @@ function gg_sum_open_llama_3b_v2 {
 
 ## main
 
-if [ -z $GG_BUILD_LOW_PERF ]; then
+if [ -z ${GG_BUILD_LOW_PERF} ]; then
     rm -rf ${SRC}/models-mnt
 
-    mnt_models=$(realpath ${MNT}/models)
+    mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
@@ -252,11 +395,15 @@ fi
 
 ret=0
 
-#test $ret -eq 0 && gg_run ctest_debug
+test $ret -eq 0 && gg_run ctest_debug
-#test $ret -eq 0 && gg_run ctest_release
+test $ret -eq 0 && gg_run ctest_release
 
-if [ -z $GG_BUILD_LOW_PERF ]; then
+if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    test $ret -eq 0 && gg_run open_llama_3b_v2
+    if [ -z ${GG_BUILD_CUDA} ]; then
+        test $ret -eq 0 && gg_run open_llama_3b_v2
+    else
+        test $ret -eq 0 && gg_run open_llama_7b_v2
+    fi
 fi
 
 exit $ret

convert.py

@@ -142,9 +142,9 @@ def find_n_mult(n_ff: int, n_embd: int) -> int:
 @dataclass
 class Params:
     n_vocab: int
-    n_embd: int
+    n_embd:  int
-    n_mult: int
+    n_mult:  int
-    n_head: int
+    n_head:  int
     n_layer: int
 
     @staticmethod
@@ -167,11 +167,11 @@ class Params:
         n_head=n_embd // 128 # guessed
 
         return Params(
-            n_vocab=n_vocab,
+            n_vocab = n_vocab,
-            n_embd=n_embd,
+            n_embd  = n_embd,
-            n_mult=256,
+            n_mult  = 256,
-            n_head=n_head,
+            n_head  = n_head,
-            n_layer=n_layer,
+            n_layer = n_layer,
         )
 
     @staticmethod
@@ -179,28 +179,53 @@ class Params:
         config = json.load(open(config_path))
 
         n_vocab = config["vocab_size"];
-        n_embd = config["hidden_size"];
+        n_embd  = config["hidden_size"];
-        n_head = config["num_attention_heads"];
+        n_head  = config["num_attention_heads"];
         n_layer = config["num_hidden_layers"];
-        n_ff = config["intermediate_size"];
+        n_ff    = config["intermediate_size"];
 
         n_mult = find_n_mult(n_ff, n_embd);
 
         return Params(
-            n_vocab=n_vocab,
+            n_vocab = n_vocab,
-            n_embd=n_embd,
+            n_embd  = n_embd,
-            n_mult=n_mult,
+            n_mult  = n_mult,
-            n_head=n_head,
+            n_head  = n_head,
-            n_layer=n_layer,
+            n_layer = n_layer,
+        )
+
+    # LLaMA v2 70B params.json
+    # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1
+    @staticmethod
+    def loadOriginalParamsJson(model: 'LazyModel', config_path: 'Path') -> 'Params':
+        config = json.load(open(config_path))
+
+        n_vocab = config["vocab_size"];
+        n_embd  = config["dim"];
+        n_head  = config["n_heads"];
+        n_layer = config["n_layers"];
+        n_mult  = config["multiple_of"];
+
+        if n_vocab == -1:
+            n_vocab = model["tok_embeddings.weight"].shape[0]
+
+        return Params(
+            n_vocab = n_vocab,
+            n_embd  = n_embd,
+            n_mult  = n_mult,
+            n_head  = n_head,
+            n_layer = n_layer,
         )
 
     @staticmethod
     def load(model_plus: 'ModelPlus') -> 'Params':
+        hf_config_path   = model_plus.paths[0].parent / "config.json"
         orig_config_path = model_plus.paths[0].parent / "params.json"
-        hf_transformer_config_path = model_plus.paths[0].parent / "config.json"
 
-        if hf_transformer_config_path.exists():
+        if hf_config_path.exists():
-            params = Params.loadHFTransformerJson(model_plus.model, hf_transformer_config_path)
+            params = Params.loadHFTransformerJson(model_plus.model, hf_config_path)
+        elif orig_config_path.exists():
+            params = Params.loadOriginalParamsJson(model_plus.model, orig_config_path)
         else:
             params = Params.guessed(model_plus.model)
 
@@ -1036,8 +1061,7 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(fname_out: Path, vocab: Vocab) -> None:
         of = OutputFile(fname_out)
-        params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0,
+        params = Params(n_vocab=vocab.vocab_size, n_embd=0, n_mult=0, n_head=1, n_layer=0)
-                        n_head=1, n_layer=0)
         of = OutputFile(fname_out)
         of.write_file_header(params, file_type=GGMLFileType.AllF32)
         of.write_vocab(vocab)

examples/CMakeLists.txt

@@ -13,6 +13,8 @@ set(TARGET common)
 add_library(${TARGET} OBJECT
     common.h
     common.cpp
+    grammar-parser.h
+    grammar-parser.cpp
     )
 
 if (BUILD_SHARED_LIBS)

examples/Miku.sh

@@ -2,21 +2,21 @@
 set -e
 
 AI_NAME="${AI_NAME:-Miku}"
-MODEL="${MODEL:-./models/gpt4all-7B/gpt4all-lora-unfiltered-quantized.bin}"
+MODEL="${MODEL:-./models/llama-2-7b-chat.ggmlv3.q4_K_M.bin}"
 USER_NAME="${USER_NAME:-Anon}"
 
 # Uncomment and adjust to the number of CPU cores you want to use.
 #N_THREAD="${N_THREAD:-4}"
+CTX_SIZE="${CTX_SIZE:-4096}"
 N_PREDICTS="${N_PREDICTS:-4096}"
 
 GEN_OPTIONS=(--batch_size 1024
---ctx_size 2048
+--ctx_size "$CTX_SIZE"
 --keep -1
 --repeat_last_n 256
 --repeat_penalty 1.17647
---temp 0.7
+--temp 0.6
---top_k 40
+--mirostat 2)
---top_p 0.5)
 
 if [ -n "$N_THREAD" ]; then
     GEN_OPTIONS+=(--threads "$N_THREAD")
@@ -24,16 +24,17 @@ fi
 
 ./main "${GEN_OPTIONS[@]}" \
     --model "$MODEL" \
+    --in-prefix " " \
+    --in-suffix "${AI_NAME}:" \
     --n_predict "$N_PREDICTS" \
     --color --interactive \
     --reverse-prompt "${USER_NAME}:" \
-    --prompt "
+    --prompt "This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
-This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the user's computer.
 ${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
 ${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct, she will ask the user for help.
 ${AI_NAME} is a very helpful AI and will help the user with anything they need. She is also very friendly and will try to make the user feel better if they are sad.
 ${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life. She will also try to make the user like her.
-The conversation is only between ${USER_NAME} and ${AI_NAME}
+The conversation is only between ${USER_NAME} and ${AI_NAME}.
 The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
 ${AI_NAME} can only communicate through text, so she can't send images or videos.
 

examples/common.cpp

@@ -117,6 +117,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                params.n_threads = std::thread::hardware_concurrency();
+            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -168,6 +171,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "-gqa" || arg == "--gqa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_gqa = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -260,12 +269,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.cfg_scale = std::stof(argv[i]);
-        } else if (arg == "--cfg-smooth-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.cfg_smooth_factor = std::stof(argv[i]);
         } else if (arg == "-b" || arg == "--batch-size") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -393,6 +396,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.antiprompt.push_back(argv[i]);
         } else if (arg == "--perplexity") {
             params.perplexity = true;
+        } else if (arg == "--perplexity-lines") {
+            params.perplexity_lines = true;
         } else if (arg == "--ignore-eos") {
             params.logit_bias[llama_token_eos()] = -INFINITY;
         } else if (arg == "--no-penalize-nl") {
@@ -433,6 +438,28 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.input_suffix = argv[i];
+        } else if (arg == "--grammar") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.grammar = argv[i];
+        } else if (arg == "--grammar-file") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(params.grammar)
+            );
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, default_params);
@@ -462,91 +489,94 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stdout, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
+    fprintf(stdout, "\n");
-    fprintf(stderr, "options:\n");
+    fprintf(stdout, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -i, --interactive     run in interactive mode\n");
+    fprintf(stdout, "  -i, --interactive     run in interactive mode\n");
-    fprintf(stderr, "  --interactive-first   run in interactive mode and wait for input right away\n");
+    fprintf(stdout, "  --interactive-first   run in interactive mode and wait for input right away\n");
-    fprintf(stderr, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
+    fprintf(stdout, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    fprintf(stderr, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
+    fprintf(stdout, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
-    fprintf(stderr, "  -r PROMPT, --reverse-prompt PROMPT\n");
+    fprintf(stdout, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stderr, "                        halt generation at PROMPT, return control in interactive mode\n");
+    fprintf(stdout, "                        halt generation at PROMPT, return control in interactive mode\n");
-    fprintf(stderr, "                        (can be specified more than once for multiple prompts).\n");
+    fprintf(stdout, "                        (can be specified more than once for multiple prompts).\n");
-    fprintf(stderr, "  --color               colorise output to distinguish prompt and user input from generations\n");
+    fprintf(stdout, "  --color               colorise output to distinguish prompt and user input from generations\n");
-    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
+    fprintf(stdout, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stdout, "  -p PROMPT, --prompt PROMPT\n");
-    fprintf(stderr, "                        prompt to start generation with (default: empty)\n");
+    fprintf(stdout, "                        prompt to start generation with (default: empty)\n");
-    fprintf(stderr, "  -e                    process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
+    fprintf(stdout, "  -e                    process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stderr, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
+    fprintf(stdout, "  --prompt-cache FNAME  file to cache prompt state for faster startup (default: none)\n");
-    fprintf(stderr, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
+    fprintf(stdout, "  --prompt-cache-all    if specified, saves user input and generations to cache as well.\n");
-    fprintf(stderr, "                        not supported with --interactive or other interactive options\n");
+    fprintf(stdout, "                        not supported with --interactive or other interactive options\n");
-    fprintf(stderr, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
+    fprintf(stdout, "  --prompt-cache-ro     if specified, uses the prompt cache but does not update it.\n");
-    fprintf(stderr, "  --random-prompt       start with a randomized prompt.\n");
+    fprintf(stdout, "  --random-prompt       start with a randomized prompt.\n");
-    fprintf(stderr, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
+    fprintf(stdout, "  --in-prefix STRING    string to prefix user inputs with (default: empty)\n");
-    fprintf(stderr, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
+    fprintf(stdout, "  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
-    fprintf(stderr, "  -f FNAME, --file FNAME\n");
+    fprintf(stdout, "  -f FNAME, --file FNAME\n");
-    fprintf(stderr, "                        prompt file to start generation.\n");
+    fprintf(stdout, "                        prompt file to start generation.\n");
-    fprintf(stderr, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
+    fprintf(stdout, "  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity)\n", params.n_predict);
-    fprintf(stderr, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
+    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stderr, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
+    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stderr, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
+    fprintf(stdout, "  -gqa N, --gqa N       grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
-    fprintf(stderr, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
+    fprintf(stdout, "  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", params.top_k);
-    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    fprintf(stdout, "  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)params.top_p);
-    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stdout, "  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)params.tfs_z);
-    fprintf(stderr, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
+    fprintf(stdout, "  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)params.typical_p);
-    fprintf(stderr, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
+    fprintf(stdout, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
-    fprintf(stderr, "  --mirostat N          use Mirostat sampling.\n");
+    fprintf(stdout, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    fprintf(stderr, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
+    fprintf(stdout, "  --presence-penalty N  repeat alpha presence penalty (default: %.1f, 0.0 = disabled)\n", (double)params.presence_penalty);
-    fprintf(stderr, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
+    fprintf(stdout, "  --frequency-penalty N repeat alpha frequency penalty (default: %.1f, 0.0 = disabled)\n", (double)params.frequency_penalty);
-    fprintf(stderr, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
+    fprintf(stdout, "  --mirostat N          use Mirostat sampling.\n");
-    fprintf(stderr, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
+    fprintf(stdout, "                        Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n");
-    fprintf(stderr, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
+    fprintf(stdout, "                        (default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)\n", params.mirostat);
-    fprintf(stderr, "                        modifies the likelihood of token appearing in the completion,\n");
+    fprintf(stdout, "  --mirostat-lr N       Mirostat learning rate, parameter eta (default: %.1f)\n", (double)params.mirostat_eta);
-    fprintf(stderr, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
+    fprintf(stdout, "  --mirostat-ent N      Mirostat target entropy, parameter tau (default: %.1f)\n", (double)params.mirostat_tau);
-    fprintf(stderr, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
+    fprintf(stdout, "  -l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS\n");
-    fprintf(stderr, "  --cfg-negative-prompt PROMPT \n");
+    fprintf(stdout, "                        modifies the likelihood of token appearing in the completion,\n");
-    fprintf(stderr, "                        negative prompt to use for guidance. (default: empty)\n");
+    fprintf(stdout, "                        i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',\n");
-    fprintf(stderr, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
+    fprintf(stdout, "                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
-    fprintf(stderr, "  --cfg-smooth-factor N smooth factor between old and new logits (default: %f, 1.0 = no smoothing)\n", params.cfg_smooth_factor);
+    fprintf(stdout, "  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
-    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stdout, "  --grammar-file FNAME  file to read grammar from\n");
-    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  --cfg-negative-prompt PROMPT \n");
-    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "                        negative prompt to use for guidance. (default: empty)\n");
-    fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
+    fprintf(stdout, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
-    fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    fprintf(stdout, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
+    fprintf(stdout, "  --no-penalize-nl      do not penalize newline token\n");
-    fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
-    fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
+    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
-    fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stderr, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
+    fprintf(stdout, "  --perplexity          compute perplexity over each ctx window of the prompt\n");
+    fprintf(stdout, "  --perplexity-lines    compute perplexity over each line of the prompt\n");
+    fprintf(stdout, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
+    fprintf(stdout, "  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     if (llama_mlock_supported()) {
-        fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported()) {
-        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    fprintf(stderr, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    fprintf(stdout, "  --numa                attempt optimizations that help on some NUMA systems\n");
-    fprintf(stderr, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    fprintf(stdout, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
-    fprintf(stderr, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
+    fprintf(stdout, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stderr, "                        number of layers to store in VRAM\n");
+    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n" );
-    fprintf(stderr, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
+    fprintf(stdout, "  -lv, --low-vram       don't allocate VRAM scratch buffer\n" );
 #endif
-    fprintf(stderr, "  --mtest               compute maximum memory usage\n");
+    fprintf(stdout, "  --mtest               compute maximum memory usage\n");
-    fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
+    fprintf(stdout, "  --export              export the computation graph to 'llama.ggml'\n");
-    fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
+    fprintf(stdout, "  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "\n");
+    fprintf(stdout, "\n");
 }
 
 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -582,18 +612,19 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
-    lparams.n_ctx        = params.n_ctx;
+    lparams.n_ctx           = params.n_ctx;
-    lparams.n_batch      = params.n_batch;
+    lparams.n_batch         = params.n_batch;
-    lparams.n_gpu_layers = params.n_gpu_layers;
+    lparams.n_gqa           = params.n_gqa;
-    lparams.main_gpu     = params.main_gpu;
+    lparams.n_gpu_layers    = params.n_gpu_layers;
-    memcpy(lparams.tensor_split, params.tensor_split, LLAMA_MAX_DEVICES*sizeof(float));
+    lparams.main_gpu        = params.main_gpu;
-    lparams.low_vram     = params.low_vram;
+    lparams.tensor_split    = params.tensor_split;
-    lparams.seed         = params.seed;
+    lparams.low_vram        = params.low_vram;
-    lparams.f16_kv       = params.memory_f16;
+    lparams.seed            = params.seed;
-    lparams.use_mmap     = params.use_mmap;
+    lparams.f16_kv          = params.memory_f16;
-    lparams.use_mlock    = params.use_mlock;
+    lparams.use_mmap        = params.use_mmap;
-    lparams.logits_all   = params.perplexity;
+    lparams.use_mlock       = params.use_mlock;
-    lparams.embedding    = params.embedding;
+    lparams.logits_all      = params.perplexity;
+    lparams.embedding       = params.embedding;
     lparams.rope_freq_base  = params.rope_freq_base;
     lparams.rope_freq_scale = params.rope_freq_scale;
 

examples/common.h

@@ -27,6 +27,7 @@ struct gpt_params {
     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_gqa                           = 1;   // grouped-query attention factor (TODO: move to hparams)
     int32_t n_keep                          = 0;   // number of tokens to keep from initial prompt
     int32_t n_chunks                        = -1;  // max number of chunks to process (-1 = unlimited)
     int32_t n_gpu_layers                    = 0;   // number of layers to store in VRAM
@@ -47,7 +48,7 @@ struct gpt_params {
     int32_t repeat_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
     float   frequency_penalty = 0.00f; // 0.0 = disabled
     float   presence_penalty  = 0.00f; // 0.0 = disabled
-    int     mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
 
@@ -55,7 +56,6 @@ struct gpt_params {
     // https://arxiv.org/abs/2306.17806
     std::string cfg_negative_prompt;       // string to help guidance
     float       cfg_scale         = 1.f;   // How strong is guidance
-    float       cfg_smooth_factor = 1.f;   // Smooth factor between old and new logits
 
     std::string model             = "models/7B/ggml-model.bin"; // model path
     std::string model_alias       = "unknown"; // model alias
@@ -63,6 +63,7 @@ struct gpt_params {
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
+    std::string grammar           = "";  // optional BNF-like grammar to constrain sampling
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
     std::string lora_adapter = "";  // lora adapter path
@@ -83,6 +84,7 @@ struct gpt_params {
     bool instruct          = false; // instruction mode (used for Alpaca models)
     bool penalize_nl       = true;  // consider newlines as a repeatable token
     bool perplexity        = false; // compute perplexity over the prompt
+    bool perplexity_lines  = false; // compute perplexity over each line of the prompt
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage

examples/embd-input/minigpt4.py

@@ -64,7 +64,7 @@ class MiniGPT4(Blip2Base):
         self.max_txt_len = max_txt_len
         self.end_sym = end_sym
         self.model = MyModel(["main", *args])
-        # system promt
+        # system prompt
         self.model.eval_string("Give the following image: <Img>ImageContent</Img>. "
            "You will be able to see the image once I provide it to you. Please answer my questions."
            "###")

examples/grammar-parser.cpp

@@ -0,0 +1,423 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from llama.cpp
+    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<llama_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<llama_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = LLAMA_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum llama_gretype type = last_sym_start < out_elements.size()
+                        ? LLAMA_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<llama_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<llama_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({LLAMA_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    bool is_char_element(llama_grammar_element elem) {
+        switch (elem.type) {
+            case LLAMA_GRETYPE_CHAR:           return true;
+            case LLAMA_GRETYPE_CHAR_NOT:       return true;
+            case LLAMA_GRETYPE_CHAR_ALT:       return true;
+            case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
+                case LLAMA_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case LLAMA_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case LLAMA_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                case LLAMA_GRETYPE_ALT:
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                case LLAMA_GRETYPE_CHAR_NOT:
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<llama_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            llama_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case LLAMA_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case LLAMA_GRETYPE_CHAR_ALT:
+                    case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, i, state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const llama_grammar_element *> parse_state::c_rules() {
+        std::vector<const llama_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

examples/grammar-parser.h

@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by llama.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "llama.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                 symbol_ids;
+        std::vector<std::vector<llama_grammar_element>> rules;
+
+        std::vector<const llama_grammar_element *> c_rules();
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}

examples/llama2-13b.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/13B/llama-2-13b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8

examples/llama2.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main -m models/available/Llama2/7B/llama-2-7b.ggmlv3.q4_0.bin \
+       --color \
+       --ctx_size 2048 \
+       -n -1 \
+       -ins -b 256 \
+       --top_k 10000 \
+       --temp 0.2 \
+       --repeat_penalty 1.1 \
+       -t 8

examples/llm.vim

@@ -0,0 +1,23 @@
+function! Llm()
+
+  let url = "http://127.0.0.1:8080/completion"
+
+  " Get the content of the current buffer
+  let buffer_content = join(getline(1, '$'), "\n")
+
+  " Create the JSON payload
+  let json_payload = {"temp":0.72,"top_k":100,"top_p":0.73,"repeat_penalty":1.100000023841858,"n_predict":10,"stream": v:false}
+  let json_payload.prompt = buffer_content
+
+  " Define the curl command
+  let curl_command = 'curl -k -s -X POST -H "Content-Type: application/json" -d @- ' . url
+  let response = system(curl_command, json_encode(json_payload))
+
+  " Extract the content field from the response
+  let content = json_decode(response).content
+
+  " Insert the content at the cursor position
+  call setline(line('.'), getline('.') . content)
+endfunction
+
+command! Llm call Llm()

examples/main/main.cpp

@@ -6,6 +6,7 @@
 #include "common.h"
 #include "llama.h"
 #include "build-info.h"
+#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -93,8 +94,8 @@ int main(int argc, char ** argv) {
     }
 
     if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified);"
+        // TODO: determine the actual max context of the model (e.g. 4096 for LLaMA v2) and use that instead of 2048
-                " you are on your own\n", __func__, params.n_ctx);
+        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified)\n", __func__, params.n_ctx);
     } else if (params.n_ctx < 8) {
         fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
@@ -139,17 +140,14 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    // determine the maximum memory usage needed to do inference for the given n_batch and n_predict parameters
+    // determine the maximum memory usage needed to do inference for the given n_batch and n_ctx parameters
     // uncomment the "used_mem" line in llama.cpp to see the results
     if (params.mem_test) {
         {
-            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos());
+            fprintf(stderr, "%s: testing memory usage for n_batch = %d, n_ctx = %d\n", __func__, params.n_batch, params.n_ctx);
-            llama_eval(ctx, tmp.data(), tmp.size(), 0, params.n_threads);
-        }
 
-        {
+            const std::vector<llama_token> tmp(params.n_batch, llama_token_bos());
-            const std::vector<llama_token> tmp = { 0, };
+            llama_eval(ctx, tmp.data(), tmp.size(), params.n_ctx, params.n_threads);
-            llama_eval(ctx, tmp.data(), tmp.size(), params.n_predict - 1, params.n_threads);
         }
 
         llama_print_timings(ctx);
@@ -340,6 +338,31 @@ int main(int argc, char ** argv) {
     fprintf(stderr, "generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
     fprintf(stderr, "\n\n");
 
+    grammar_parser::parse_state parsed_grammar;
+    llama_grammar *             grammar = NULL;
+    if (!params.grammar.empty()) {
+        parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+        // will be empty (default) if there are parse errors
+        if (parsed_grammar.rules.empty()) {
+            return 1;
+        }
+        fprintf(stderr, "%s: grammar:\n", __func__);
+        grammar_parser::print_grammar(stderr, parsed_grammar);
+        fprintf(stderr, "\n");
+
+        {
+            auto it = params.logit_bias.find(llama_token_eos());
+            if (it != params.logit_bias.end() && it->second == -INFINITY) {
+                fprintf(stderr,
+                    "%s: warning: EOS token is disabled, which will cause most grammars to fail\n", __func__);
+            }
+        }
+
+        std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+        grammar = llama_grammar_init(
+            grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
+    }
+
     // TODO: replace with ring-buffer
     std::vector<llama_token> last_n_tokens(n_ctx);
     std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
@@ -557,7 +580,7 @@ int main(int argc, char ** argv) {
                 llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
 
                 if (ctx_guidance) {
-                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale, params.cfg_smooth_factor);
+                    llama_sample_classifier_free_guidance(ctx, &candidates_p, ctx_guidance, params.cfg_scale);
                 }
 
                 // Apply penalties
@@ -573,6 +596,10 @@ int main(int argc, char ** argv) {
                     logits[llama_token_nl()] = nl_logit;
                 }
 
+                if (grammar != NULL) {
+                    llama_sample_grammar(ctx, &candidates_p, grammar);
+                }
+
                 if (temp <= 0) {
                     // Greedy sampling
                     id = llama_sample_token_greedy(ctx, &candidates_p);
@@ -598,6 +625,10 @@ int main(int argc, char ** argv) {
                 }
                 // printf("`%d`", candidates_p.size);
 
+                if (grammar != NULL) {
+                    llama_grammar_accept_token(ctx, grammar, id);
+                }
+
                 last_n_tokens.erase(last_n_tokens.begin());
                 last_n_tokens.push_back(id);
             }
@@ -728,6 +759,18 @@ int main(int argc, char ** argv) {
             }
 
             if (n_past > 0) {
+                if (is_interacting) {
+                    // reset grammar state if we're restarting generation
+                    if (grammar != NULL) {
+                        llama_grammar_free(grammar);
+
+                        std::vector<const llama_grammar_element *> grammar_rules(
+                            parsed_grammar.c_rules());
+                        grammar = llama_grammar_init(
+                            grammar_rules.data(), grammar_rules.size(),
+                            parsed_grammar.symbol_ids.at("root"));
+                    }
+                }
                 is_interacting = false;
             }
         }
@@ -759,6 +802,9 @@ int main(int argc, char ** argv) {
     llama_free(ctx);
     llama_free_model(model);
 
+    if (grammar != NULL) {
+        llama_grammar_free(grammar);
+    }
     llama_backend_free();
 
     return 0;

examples/make-ggml.py

@@ -0,0 +1,92 @@
+"""
+This script converts Hugging Face llama models to GGML and quantizes them.
+
+Usage:
+python make-ggml.py --model {model_dir_or_hf_repo_name} [--outname {output_name} (Optional)] [--outdir {output_directory} (Optional)] [--quants {quant_types} (Optional)] [--keep_fp16 (Optional)]
+
+Arguments:
+- --model: (Required) The directory of the downloaded Hugging Face model or the name of the Hugging Face model repository. If the model directory does not exist, it will be downloaded from the Hugging Face model hub.
+- --outname: (Optional) The name of the output model. If not specified, the last part of the model directory path or the Hugging Face model repo name will be used.
+- --outdir: (Optional) The directory where the output model(s) will be stored. If not specified, '../models/{outname}' will be used.
+- --quants: (Optional) The types of quantization to apply. This should be a space-separated list. The default is 'Q4_K_M Q5_K_S'.
+- --keep_fp16: (Optional) If specified, the FP16 model will not be deleted after the quantized models are created.
+
+Quant types:
+- Q4_0: small, very high quality loss - legacy, prefer using Q3_K_M
+- Q4_1: small, substantial quality loss - legacy, prefer using Q3_K_L
+- Q5_0: medium, balanced quality - legacy, prefer using Q4_K_M
+- Q5_1: medium, low quality loss - legacy, prefer using Q5_K_M
+- Q2_K: smallest, extreme quality loss - not recommended
+- Q3_K: alias for Q3_K_M
+- Q3_K_S: very small, very high quality loss
+- Q3_K_M: very small, very high quality loss
+- Q3_K_L: small, substantial quality loss
+- Q4_K: alias for Q4_K_M
+- Q4_K_S: small, significant quality loss
+- Q4_K_M: medium, balanced quality - recommended
+- Q5_K: alias for Q5_K_M
+- Q5_K_S: large, low quality loss - recommended
+- Q5_K_M: large, very low quality loss - recommended
+- Q6_K: very large, extremely low quality loss
+- Q8_0: very large, extremely low quality loss - not recommended
+- F16: extremely large, virtually no quality loss - not recommended
+- F32: absolutely huge, lossless - not recommended
+"""
+import subprocess
+subprocess.run(f"pip install huggingface-hub==0.16.4", shell=True, check=True)
+
+import argparse
+import os
+from huggingface_hub import snapshot_download
+
+def main(model, outname, outdir, quants, keep_fp16):
+    ggml_version = "v3"
+
+    if not os.path.isdir(model):
+        print(f"Model not found at {model}. Downloading...")
+        try:
+            if outname is None:
+                outname = model.split('/')[-1]
+            model = snapshot_download(repo_id=model, cache_dir='../models/hf_cache')
+        except Exception as e:
+            raise Exception(f"Could not download the model: {e}")
+
+    if outdir is None:
+        outdir = f'../models/{outname}'
+
+    if not os.path.isfile(f"{model}/config.json"):
+        raise Exception(f"Could not find config.json in {model}")
+
+    os.makedirs(outdir, exist_ok=True)
+
+    print("Building llama.cpp")
+    subprocess.run(f"cd .. && make quantize", shell=True, check=True)
+
+    fp16 = f"{outdir}/{outname}.ggml{ggml_version}.fp16.bin"
+
+    print(f"Making unquantised GGML at {fp16}")
+    if not os.path.isfile(fp16):
+        subprocess.run(f"python3 ../convert.py {model} --outtype f16 --outfile {fp16}", shell=True, check=True)
+    else:
+        print(f"Unquantised GGML already exists at: {fp16}")
+
+    print("Making quants")
+    for type in quants:
+        outfile = f"{outdir}/{outname}.ggml{ggml_version}.{type}.bin"
+        print(f"Making {type} : {outfile}")
+        subprocess.run(f"../quantize {fp16} {outfile} {type}", shell=True, check=True)
+
+    if not keep_fp16:
+        os.remove(fp16)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Convert/Quantize HF to GGML. If you have the HF model downloaded already, pass the path to the model dir. Otherwise, pass the Hugging Face model repo name. You need to be in the /examples folder for it to work.')
+    parser.add_argument('--model', required=True, help='Downloaded model dir or Hugging Face model repo name')
+    parser.add_argument('--outname', default=None, help='Output model(s) name')
+    parser.add_argument('--outdir', default=None, help='Output directory')
+    parser.add_argument('--quants', nargs='*', default=["Q4_K_M", "Q5_K_S"], help='Quant types')
+    parser.add_argument('--keep_fp16', action='store_true', help='Keep fp16 model', default=False)
+
+    args = parser.parse_args()
+
+    main(args.model, args.outname, args.outdir, args.quants, args.keep_fp16)

examples/perplexity/perplexity.cpp

@@ -4,6 +4,7 @@
 
 #include <cmath>
 #include <ctime>
+#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -120,6 +121,77 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
     printf("\n");
 }
 
+void perplexity_lines(llama_context * ctx, const gpt_params & params) {
+    // Calculates perplexity over each line of the prompt
+
+    std::vector<std::string> prompt_lines;
+    std::istringstream strstream(params.prompt);
+    std::string line;
+
+    while (std::getline(strstream,line,'\n')) {
+        prompt_lines.push_back(line);
+    }
+
+    const int n_vocab = llama_n_vocab(ctx);
+
+    int counttotal   = 0;
+    size_t n_lines = prompt_lines.size();
+
+    double nll = 0.0;
+
+    fprintf(stderr, "%s: calculating perplexity over %lu lines\n", __func__, n_lines);
+
+    printf("\nLine\tPPL line\tPPL cumulative\n");
+
+    for (size_t i = 0; i < n_lines; ++i) {
+
+        // Tokenize and insert BOS at start
+        std::vector<int> batch_embd = ::llama_tokenize(ctx, prompt_lines[i], true);
+
+        size_t batch_size  = batch_embd.size();
+
+        // Stop if line is too long
+        if( batch_size > (size_t)params.n_ctx ) {
+            fprintf(stderr, "%s : tokens in line %lu > n_ctxl\n", __func__, i);
+            return;
+        }
+
+        if (llama_eval(ctx, batch_embd.data(), batch_size, 0, params.n_threads)) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
+            return;
+        }
+
+        const auto batch_logits = llama_get_logits(ctx);
+        std::vector<float> logits;
+        logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+
+        double nllline = 0.0;
+        int countline = 0;
+
+        // Perplexity over second half of the line
+        for (size_t j = batch_size/2; j < batch_size - 1; ++j) {
+            // Calculate probability of next token, given the previous ones.
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+
+            const float prob = softmax(tok_logits)[batch_embd[ j + 1]];
+
+            nllline += -std::log(prob);
+            ++countline;
+        }
+
+        nll += nllline;
+        counttotal += countline;
+
+        // perplexity is e^(average negative log-likelihood)
+        printf("%lu\t%.8lf\t%.8lf\n", i + 1, std::exp(nllline/countline), std::exp(nll / counttotal) );
+        fflush(stdout);
+    }
+
+    printf("\n");
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -168,7 +240,11 @@ int main(int argc, char ** argv) {
                 params.n_threads, std::thread::hardware_concurrency(), llama_print_system_info());
     }
 
-    perplexity(ctx, params);
+    if (params.perplexity_lines) {
+        perplexity_lines(ctx, params);
+    } else {
+        perplexity(ctx, params);
+    }
 
     llama_print_timings(ctx);
     llama_free(ctx);

examples/server/CMakeLists.txt

@@ -7,6 +7,9 @@ target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if(TARGET BUILD_INFO)
   add_dependencies(${TARGET} BUILD_INFO)

examples/server/server.cpp

@@ -440,7 +440,7 @@ struct llama_server_context {
 
             if (cfg_enabled) {
                 llama_sample_classifier_free_guidance(
-                    ctx, &candidates_p, evaluator_guidance.ctx, params.cfg_scale, 1.0);
+                    ctx, &candidates_p, evaluator_guidance.ctx, params.cfg_scale);
             }
 
             // Apply penalties
@@ -602,47 +602,49 @@ struct llama_server_context {
     }
 };
 
-static void server_print_usage(const char * argv0, const gpt_params & params,
+static void server_print_usage(const char *argv0, const gpt_params &params,
-                               const server_params & sparams) {
+                               const server_params &sparams)
-    fprintf(stderr, "usage: %s [options]\n", argv0);
+{
-    fprintf(stderr, "\n");
+    fprintf(stdout, "usage: %s [options]\n", argv0);
-    fprintf(stderr, "options:\n");
+    fprintf(stdout, "\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stdout, "options:\n");
-    fprintf(stderr, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
+    fprintf(stdout, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stdout, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
-    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stdout, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stdout, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
-    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
+    fprintf(stdout, "  -gqa N, --gqa N       grouped-query attention factor (TEMP!!! use 8 for LLaMAv2 70B) (default: %d)\n", params.n_gqa);
-    fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stdout, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
-    fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stdout, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
-    fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
+    fprintf(stdout, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stdout, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    fprintf(stdout, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
     if (llama_mlock_supported()) {
-        fprintf(stderr, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
+        fprintf(stdout, "  --mlock               force system to keep model in RAM rather than swapping or compressing\n");
     }
     if (llama_mmap_supported()) {
-        fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
+        fprintf(stdout, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stdout, "  -ngl N, --n-gpu-layers N\n");
-    fprintf(stderr, "                        number of layers to store in VRAM\n");
+    fprintf(stdout, "                        number of layers to store in VRAM\n");
-    fprintf(stderr, "  -ts SPLIT --tensor-split SPLIT\n");
+    fprintf(stdout, "  -ts SPLIT --tensor-split SPLIT\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    fprintf(stdout, "                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
-    fprintf(stderr, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+    fprintf(stdout, "  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
-    fprintf(stderr, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
+    fprintf(stdout, "  -lv, --low-vram don't allocate VRAM scratch buffer\n");
 #endif
-    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stdout, "  -m FNAME, --model FNAME\n");
-    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stdout, "                        model path (default: %s)\n", params.model.c_str());
-    fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
+    fprintf(stdout, "  -a ALIAS, --alias ALIAS\n");
-    fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
+    fprintf(stdout, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
+    fprintf(stdout, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
-    fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
+    fprintf(stdout, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
-    fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
+    fprintf(stdout, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
-    fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
+    fprintf(stdout, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
-    fprintf(stderr, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
+    fprintf(stdout, "  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
-    fprintf(stderr, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
+    fprintf(stdout, "  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
-    fprintf(stderr, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
+    fprintf(stdout, "  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
-    fprintf(stderr, "\n");
+    fprintf(stdout, "\n");
 }
 
 static void server_params_parse(int argc, char ** argv, server_params & sparams,
@@ -700,6 +702,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "-gqa" || arg == "--gqa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_gqa = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -739,8 +747,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                         "See main README.md for information on enabling GPU BLAS support",
                         {{ "n_gpu_layers", params.n_gpu_layers }});
 #endif
-        }
+        } else if (arg == "--tensor-split" || arg == "-ts") {
-        else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -770,7 +777,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
 #else
             fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set lower vram usage.\n");
 #endif // GGML_USE_CUBLAS
-        }else if (arg == "--main-gpu" || arg == "-mg") {
+        } else if (arg == "--main-gpu" || arg == "-mg") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1434,7 +1434,7 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
     gf->perf_time_us = 0;
 
     const auto & hparams = model->hparams;
-    //const int n_ctx      = hparams.n_ctx;
+    const int n_ctx      = hparams.n_ctx;
     const int n_vocab    = hparams.n_vocab;
     const int n_embd     = hparams.n_embd;
     const int n_layer    = hparams.n_layer;
@@ -1863,10 +1863,10 @@ struct ggml_tensor * forward_batch_wo_cache_flash_attn_train(
         t12->grad = expand(gb, ggml_permute(ctx0, t15->grad, 0, 2, 3, 1));                                            assert_shape_4d(t12->grad, N, n_batch, n_embd/n_head, n_head);
         t11->grad = expand(gb, ggml_reshape_2d(ctx0, ggml_cont(ctx0, t12->grad), N*n_batch, n_embd));                 assert_shape_2d(t11->grad, N*n_batch, n_embd);
         t10->grad = expand(gb, ggml_permute(ctx0, t14->grad, 0, 2, 1, 3));                                            assert_shape_4d(t10->grad, n_embd/n_head, n_head, N, n_batch);
-        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
+        t09->grad = expand(gb, ggml_rope_back(ctx0, t10->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t09->grad, n_embd/n_head, n_head, N, n_batch);
         t08->grad = expand(gb, ggml_reshape_2d(ctx0, t09->grad, n_embd, N*n_batch));                                  assert_shape_2d(t08->grad, n_embd, N*n_batch);
         t07->grad = expand(gb, ggml_permute(ctx0, t13->grad, 0, 2, 1, 3));                                            assert_shape_4d(t07->grad, n_embd/n_head, n_head, N, n_batch);
-        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode));                            assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
+        t06->grad = expand(gb, ggml_rope_back(ctx0, t07->grad, n_past, n_rot, rope_mode, n_ctx));                     assert_shape_4d(t06->grad, n_embd/n_head, n_head, N, n_batch);
         t05->grad = expand(gb, ggml_reshape_2d(ctx0, t06->grad, n_embd, N*n_batch));                                  assert_shape_2d(t05->grad, n_embd, N*n_batch);
         t04->grad = expand(gb, ggml_add_inplace(ctx0,
                         ggml_add_inplace(ctx0,

flake.nix

@@ -6,8 +6,9 @@
   outputs = { self, nixpkgs, flake-utils }:
     flake-utils.lib.eachDefaultSystem (system:
       let
-        inherit (pkgs.stdenv) isAarch32 isAarch64 isx86_32 isx86_64 isDarwin;
+        inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
-        osSpecific = with pkgs; [ openmpi ] ++
+        buildInputs = with pkgs; [ openmpi ];
+        osSpecific = with pkgs; buildInputs ++
         (
           if isAarch64 && isDarwin then
             with pkgs.darwin.apple_sdk_11_0.frameworks; [
@@ -22,49 +23,51 @@
               CoreGraphics
               CoreVideo
             ]
-          else if isx86_32 || isx86_64 then
-            with pkgs; [ mkl ]
           else
             with pkgs; [ openblas ]
         );
         pkgs = import nixpkgs { inherit system; };
+        nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
         llama-python =
-          pkgs.python310.withPackages (ps: with ps; [ numpy sentencepiece ]);
+          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
+        postPatch = ''
+          substituteInPlace ./ggml-metal.m \
+            --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+          substituteInPlace ./*.py --replace '/usr/bin/env python' '${llama-python}/bin/python'
+        '';
+        postInstall = ''
+          mv $out/bin/main $out/bin/llama
+          mv $out/bin/server $out/bin/llama-server
+        '';
+        cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
       in {
         packages.default = pkgs.stdenv.mkDerivation {
           name = "llama.cpp";
           src = ./.;
-          postPatch = ''
+          postPatch = postPatch;
-            substituteInPlace ./ggml-metal.m \
+          nativeBuildInputs = nativeBuildInputs;
-              --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-          '';
-          nativeBuildInputs = with pkgs; [ cmake pkgconfig ];
           buildInputs = osSpecific;
-          cmakeFlags = [ "-DLLAMA_BUILD_SERVER=ON" "-DLLAMA_MPI=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ]
+          cmakeFlags = cmakeFlags
             ++ (if isAarch64 && isDarwin then [
               "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
               "-DLLAMA_METAL=ON"
-            ] else if isx86_32 || isx86_64 then [
-              "-DLLAMA_BLAS=ON"
-              "-DLLAMA_BLAS_VENDOR=Intel10_lp64"
             ] else [
               "-DLLAMA_BLAS=ON"
               "-DLLAMA_BLAS_VENDOR=OpenBLAS"
           ]);
-          installPhase = ''
+          postInstall = postInstall;
-            runHook preInstall
+          meta.mainProgram = "llama";
-
+        };
-            install -D bin/* -t $out/bin
+        packages.opencl = pkgs.stdenv.mkDerivation {
-            install -Dm644 lib*.so -t $out/lib
+          name = "llama.cpp";
-            mv $out/bin/main $out/bin/llama
+          src = ./.;
-            mv $out/bin/server $out/bin/llama-server
+          postPatch = postPatch;
-
+          nativeBuildInputs = nativeBuildInputs;
-            echo "#!${llama-python}/bin/python" > $out/bin/convert.py
+          buildInputs = with pkgs; buildInputs ++ [ clblast ];
-            cat ${./convert.py} >> $out/bin/convert.py
+          cmakeFlags = cmakeFlags ++ [
-            chmod +x $out/bin/convert.py
+            "-DLLAMA_CLBLAST=ON"
-
+          ];
-            runHook postInstall
+          postInstall = postInstall;
-          '';
           meta.mainProgram = "llama";
         };
         apps.llama-server = {
@@ -81,7 +84,7 @@
         };
         apps.default = self.apps.${system}.llama;
         devShells.default = pkgs.mkShell {
-          packages = with pkgs; [ cmake llama-python ] ++ osSpecific;
+          packages = nativeBuildInputs ++ osSpecific;
         };
       });
 }

ggml-cuda.cu

@@ -220,7 +220,7 @@ typedef struct {
 static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
 
 #define WARP_SIZE 32
-#define MATRIX_ROW_PADDING 256 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
 #define CUDA_ADD_BLOCK_SIZE 256
 #define CUDA_MUL_BLOCK_SIZE 256
@@ -935,12 +935,18 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
     uint16_t aux[4];
     const uint8_t * sc = (const uint8_t *)aux;
 
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
     float tmp = 0; // partial sum for thread in warp
 
     for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
 
-        const uint8_t * q1 = x[i].qs + q_offset;
-        const uint8_t * q2 = q1 + 64;
         const float   * y1 = yy + i*QK_K + y_offset;
         const float   * y2 = y1 + 128;
 
@@ -953,14 +959,41 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
         aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
         aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
 
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
         float4 s = {0.f, 0.f, 0.f, 0.f};
         float smin = 0;
-        for (int l = 0; l < n; ++l) {
+        for (int l = 0; l < 4; ++l) {
-            s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
-            s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
+            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
             smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
         }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
 
     }
 #else
@@ -1040,10 +1073,12 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
     uint16_t aux[4];
     const uint8_t * sc = (const uint8_t *)aux;
 
+    uint16_t q16[8];
+    const uint8_t * q4 = (const uint8_t *)q16;
+
     for (int i = ix; i < num_blocks_per_row; i += 2) {
 
         const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * ql2 = ql1 + 64;
         const uint8_t * qh  = x[i].qh + l0;
         const float   * y1  = yy + i*QK_K + y_offset;
         const float   * y2  = y1 + 128;
@@ -1059,15 +1094,25 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 
         float4 sum = {0.f, 0.f, 0.f, 0.f};
         float smin = 0;
+        const uint16_t * q1 = (const uint16_t *)ql1;
+        const uint16_t * q2 = q1 + 32;
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[8] & 0x0f0f;
+        q16[2] = (q1[0] >> 4) & 0x0f0f;
+        q16[3] = (q1[8] >> 4) & 0x0f0f;
+        q16[4] = q2[0] & 0x0f0f;
+        q16[5] = q2[8] & 0x0f0f;
+        q16[6] = (q2[0] >> 4) & 0x0f0f;
+        q16[7] = (q2[8] >> 4) & 0x0f0f;
         for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
             smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
                   + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
         }
@@ -1521,7 +1566,8 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
-    const int bq8_offset = QR4_K * (iqs / QI8_1);
+    // iqs is in 0...15. bq8_offset = 2 * (iqs/4) -> bq8_offset = 0, 2, 4, 6
+    const int bq8_offset = QR4_K * (iqs / (QI8_1/2));
 
     float sumf_d = 0.0f;
     float sumf_m = 0.0f;
@@ -1529,22 +1575,44 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     const float    d = bq4_K->d;
     const float dmin = bq4_K->dmin;
 
-    const int v = *((int *) &bq4_K->qs[sizeof(int) * iqs]);
+    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
+    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
+    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
+    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
+
+    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * (iqs%4));
+    const int v1 = q4[0];
+    const int v2 = q4[4];
+
+    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
 
     for (int i = 0; i < QR4_K; ++i) {
-        const int isc = bq8_offset + i;
-
-        uint8_t sc, m;
-        get_scale_min_k4(isc, bq4_K->scales, sc, m);
 
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        const int ui = *((int*) &bq8i->qs[sizeof(int) * (iqs % QI8_1)]);
         const float d8i = bq8i->d;
+        const int * q8 = (const int *)bq8i->qs + (iqs%4);
+        const int ui1 = q8[0];
+        const int ui2 = q8[4];
 
-        const int vi = (v >> (4*i)) & 0x0F0F0F0F;
+        const int vi1 = (v1 >> (4*i)) & 0x0F0F0F0F;
+        const int vi2 = (v2 >> (4*i)) & 0x0F0F0F0F;
 
-        sumf_d += d8i * (__dp4a(vi,         ui, 0) * sc); // SIMD dot product
+        const int dot1 = __dp4a(vi2, ui2, __dp4a(vi1, ui1, 0)); // SIMD dot product
-        sumf_m += d8i * (__dp4a(0x01010101, ui, 0) * m);  // multiply constant part of q4_K with sum of q8_1 values
+        const int dot2 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+
+        sumf_d += d8i * (dot1 * sc[i]);
+        sumf_m += d8i * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
     }
 
     return d*sumf_d - dmin*sumf_m;
@@ -1559,7 +1627,9 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
     const block_q5_K * bq5_K = (const block_q5_K *) vbq;
 
-    const int bq8_offset = QR5_K * (iqs / QI8_1);
+    const int bq8_offset = QR5_K * (iqs / (QI8_1/2));
+    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * (iqs%4));
+    const int * qh = (const int *)(bq5_K->qh + 4 * (iqs%4));
 
     float sumf_d = 0.0f;
     float sumf_m = 0.0f;
@@ -1567,28 +1637,48 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     const float    d = bq5_K->d;
     const float dmin = bq5_K->dmin;
 
-    const int vl = *((int *) &bq5_K->qs[sizeof(int) * iqs]);
+    const int vl1 = ql[0];
+    const int vl2 = ql[4];
 
-    const int vh = (*((int *) &bq5_K->qh[sizeof(int) * (iqs % (QI5_K/4))])) >> bq8_offset;
+    const int vh1 = qh[0] >> bq8_offset;
+    const int vh2 = qh[4] >> bq8_offset;
+
+    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
 
     for (int i = 0; i < QR5_K; ++i) {
-        const int isc = bq8_offset + i;
-
-        uint8_t sc, m;
-        get_scale_min_k4(isc, bq5_K->scales, sc, m);
 
         const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        const int ui = *((int*) &bq8i->qs[sizeof(int) * (iqs % QI8_1)]);
         const float d8i = bq8i->d;
+        const int * q8 = (const int *)bq8i->qs + (iqs%4);
+        const int ui1 = q8[0];
+        const int ui2 = q8[4];
 
-        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
+        const int vil1 = (vl1 >> (4*i)) & 0x0F0F0F0F;
+        const int vil2 = (vl2 >> (4*i)) & 0x0F0F0F0F;
 
-        const int vih = ((vh >> i) << 4) & 0x10101010;
+        const int vih1 = ((vh1 >> i) << 4) & 0x10101010;
+        const int vih2 = ((vh2 >> i) << 4) & 0x10101010;
 
-        const int vi = vil | vih;
+        const int vi1 = vil1 | vih1;
+        const int vi2 = vil2 | vih2;
+
+        const int dot1 = __dp4a(vi2, ui2, __dp4a(vi1, ui1, 0)); // SIMD dot product
+        const int dot2 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
+
+        sumf_d += d8i * (dot1 * sc[i]);
+        sumf_m += d8i * (dot2 * m[i]);
 
-        sumf_d += d8i * (__dp4a(vi,         ui, 0) * sc); // SIMD dot product
-        sumf_m += d8i * (__dp4a(0x01010101, ui, 0) * m);  // multiply constant part of q5_K with sum of q8_1 values
     }
 
     return d*sumf_d - dmin*sumf_m;
@@ -1745,11 +1835,15 @@ static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, cons
     }
 }
 
-static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x, const int nchannels_x) {
+static __global__ void mul_mat_p021_f16_f32(
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
+
     const half * x = (const half *) vx;
 
     const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
     const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / (nchannels_y / nchannels_x);
 
     const int nrows_y = ncols_x;
     const int nrows_dst = nrows_x;
@@ -1765,7 +1859,7 @@ static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const
         }
 
         // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel*ncols_x + col_x;
+        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
         const float xi = __half2float(x[ix]);
 
         const int row_y = col_x;
@@ -1793,12 +1887,13 @@ static __global__ void mul_mat_p021_f16_f32(const void * __restrict__ vx, const
 
 static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
     const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x) {
+    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
 
     const half * x = (const half *) vx;
 
     const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
     const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / channel_x_divisor;
 
     const int nrows_y = ncols_x;
     const int nrows_dst = nrows_x;
@@ -1815,7 +1910,7 @@ static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
             break;
         }
 
-        const int ix = channel*channel_stride_x + row_x*row_stride_x + col_x;
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
         const float xi = __half2float(x[ix]);
 
         const int row_y = col_x;
@@ -2259,7 +2354,10 @@ static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float *
     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_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI4_K, block_q4_K, vec_dot_q4_K_q8_1>
+    // Note: we use QI4_K/2 instead of QI4_K to make the dot product template require 4 groups of quants to be processed per
+    //       kernel call instead of 2. This results in a better perfmance because the cost of computing the k-quant scales
+    //       is better amortized.
+    mul_mat_vec_q<QK_K, QI4_K/2, block_q4_K, vec_dot_q4_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
@@ -2268,7 +2366,10 @@ static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float *
     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_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI5_K, block_q5_K, vec_dot_q5_K_q8_1>
+    // Note: we use QI5_K/2 instead of QI5_K to make the dot product template require 4 groups of quants to be processed per
+    //       kernel call instead of 2. This results in a better perfmance because the cost of computing the k-quant scales
+    //       is better amortized.
+    mul_mat_vec_q<QK_K, QI5_K/2, block_q5_K, vec_dot_q5_K_q8_1>
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
@@ -2324,20 +2425,23 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     }
 }
 
-static void ggml_mul_mat_p021_f16_f32_cuda(const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nchannels_x, cudaStream_t stream) {
+static void ggml_mul_mat_p021_f16_f32_cuda(
-    const dim3 block_nums(1, nrows_x, nchannels_x);
+    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
+    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
+
+    const dim3 block_nums(1, nrows_x, nchannels_y);
     const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x);
+    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
 }
 
 static void ggml_mul_mat_vec_nc_f16_f32_cuda(
     const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int channel_stride_x, cudaStream_t stream) {
+    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
 
-    const dim3 block_nums(1, nrows_x, nchannels_x);
+    const dim3 block_nums(1, nrows_x, nchannels_y);
     const dim3 block_dims(WARP_SIZE, 1, 1);
     mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x);
+        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
 }
 
 static void ggml_cpy_f32_f32_cuda(
@@ -2423,20 +2527,53 @@ static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
     int id;
     CUDA_CHECK(cudaGetDevice(&id));
-
+#ifdef DEBUG_CUDA_MALLOC
+    int nnz = 0;
+    size_t max_size = 0, tot_size = 0;
+#endif
+    size_t best_diff = 1ull << 36;
+    int ibest = -1;
     for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
         cuda_buffer& b = g_cuda_buffer_pool[id][i];
-        if (b.size >= size && b.ptr != nullptr) {
+        if (b.ptr != nullptr) {
-            void * ptr = b.ptr;
+#ifdef DEBUG_CUDA_MALLOC
-            *actual_size = b.size;
+            ++nnz;
-            b.ptr = nullptr;
+            tot_size += b.size;
-            b.size = 0;
+            if (b.size > max_size) max_size = b.size;
-            return ptr;
+#endif
+            if (b.size >= size) {
+                size_t diff = b.size - size;
+                if (diff < best_diff) {
+                    best_diff = diff;
+                    ibest = i;
+                    if (!best_diff) {
+                        void * ptr = b.ptr;
+                        *actual_size = b.size;
+                        b.ptr = nullptr;
+                        b.size = 0;
+                        return ptr;
+                    }
+                }
+            }
         }
     }
+    if (ibest >= 0) {
+        cuda_buffer& b = g_cuda_buffer_pool[id][ibest];
+        void * ptr = b.ptr;
+        *actual_size = b.size;
+        b.ptr = nullptr;
+        b.size = 0;
+        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,
+            (uint32_t)(max_size/1024/1024), (uint32_t)(tot_size/1024/1024), (uint32_t)(size/1024/1024));
+#endif
     void * ptr;
-    CUDA_CHECK(cudaMalloc((void **) &ptr, size));
+    size_t look_ahead_size = (size_t) (1.05 * size);
-    *actual_size = size;
+    look_ahead_size = 256 * ((look_ahead_size + 255)/256);
+    CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
+    *actual_size = look_ahead_size;
     return ptr;
 }
 
@@ -2464,7 +2601,9 @@ static size_t g_scratch_offset = 0;
 
 static int g_device_count = -1;
 static int g_main_device = 0;
+#ifndef GGML_CUDA_FORCE_DMMV
 static int g_compute_capabilities[GGML_CUDA_MAX_DEVICES];
+#endif
 static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 
 static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
@@ -2487,7 +2626,9 @@ void ggml_init_cublas() {
             g_tensor_split[id] = total_vram;
             total_vram += prop.totalGlobalMem;
 
+#ifndef GGML_CUDA_FORCE_DMMV
             g_compute_capabilities[id] = 100*prop.major + 10*prop.minor;
+#endif
         }
         for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
@@ -2512,6 +2653,9 @@ void ggml_init_cublas() {
 }
 
 void ggml_cuda_set_tensor_split(const float * tensor_split) {
+    if (tensor_split == nullptr) {
+        return;
+    }
     bool all_zero = true;
     for (int i = 0; i < g_device_count; ++i) {
         if (tensor_split[i] != 0.0f) {
@@ -2652,6 +2796,7 @@ inline void ggml_cuda_op_mul(
     (void) dst;
     (void) src0_ddq_i;
     (void) i02;
+    (void) i1;
 }
 
 inline void ggml_cuda_op_gelu(
@@ -2779,8 +2924,8 @@ inline void ggml_cuda_op_mul_mat_vec(
 #endif
 
     if (use_mul_mat_vec_q) {
-        int64_t padded_row_size = ne00 + MATRIX_ROW_PADDING - 1;
+        const int64_t padded_row_size = ne00 % MATRIX_ROW_PADDING == 0 ?
-        padded_row_size -= padded_row_size % MATRIX_ROW_PADDING;
+            ne00 : ne00 - ne00 % MATRIX_ROW_PADDING + MATRIX_ROW_PADDING;
         size_t as;
         void * src1_q8_1 = ggml_cuda_pool_malloc(padded_row_size*sizeof(block_q8_1)/QK8_1, &as);
         quantize_row_q8_1_cuda(src1_ddf_i, src1_q8_1, ne00, padded_row_size, cudaStream_main);
@@ -2947,13 +3092,18 @@ inline void ggml_cuda_op_rope(
     const int64_t ne00 = src0->ne[0];
     const int64_t i01_diff = i01_high - i01_low;
 
-    const int n_past = ((int32_t *) src1->data)[0];
+    const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_dims = ((int32_t *) src1->data)[1];
+    const int n_dims = ((int32_t *) dst->op_params)[1];
-    const int mode   = ((int32_t *) src1->data)[2];
+    const int mode   = ((int32_t *) dst->op_params)[2];
-    const int n_ctx  = ((int32_t *) src1->data)[3];
+    const int n_ctx  = ((int32_t *) dst->op_params)[3];
+    // RoPE alteration for extended context
 
-    const float theta_scale = powf(10000.0, -2.0f/n_dims);
+    float freq_base, freq_scale;
-    const float p = ((mode & 1) == 0 ? n_past + i02 : i02);
+    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);
+    const float p = (((mode & 1) == 0 ? n_past + i02 : i02)) * freq_scale;
 
     bool is_glm = mode & 4;
 
@@ -2966,6 +3116,7 @@ inline void ggml_cuda_op_rope(
         rope_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, p, theta_scale, cudaStream_main);
     }
 
+    (void) src1;
     (void) dst;
     (void) src0_ddq_i;
     (void) src1_ddf_i;
@@ -2984,11 +3135,12 @@ inline void ggml_cuda_op_diag_mask_inf(
     const int64_t ne01 = src0->ne[1];
     const int64_t i01_diff = i01_high - i01_low;
 
-    const int n_past = ((int32_t *) src1->data)[0];
+    const int n_past = ((int32_t *) dst->op_params)[0];
 
     // compute
     diag_mask_inf_f32_cuda(src0_ddf_i, dst_ddf_i, ne00, i01_diff, ne01, n_past, cudaStream_main);
 
+    (void) src1;
     (void) dst;
     (void) src0_ddq_i;
     (void) src1_ddf_i;
@@ -3056,6 +3208,9 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     const int64_t ne11 = use_src1 ? src1->ne[1] : 1;
     const int64_t ne12 = use_src1 ? src1->ne[2] : 1;
     const int64_t ne13 = use_src1 ? src1->ne[3] : 1;
+    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
+
+    GGML_ASSERT(ne03 == ne13);
 
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
@@ -3067,12 +3222,19 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
 
     // strides for iteration over dims 3 and 2
-    const int64_t num_iters = flatten_rows ? 1 : ne02 * ne03;
+    const int64_t num_iters_0 = ne02 >= ne12 ? ne02*ne03 : ne12*ne13;
-    const int64_t stride_mod = flatten_rows ? ne02 * ne03 : 1;
+    const int64_t num_iters = flatten_rows ? 1 : num_iters_0;
+    const int64_t stride_mod = flatten_rows ? num_iters_0 : 1;
     const int64_t src0_stride = ne00 * ne01 * stride_mod;
     const int64_t src1_stride = ne10 * ne11 * stride_mod;
     const int64_t dst_stride = ne0 * ne1 * stride_mod;
 
+    const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;
+    const int64_t i03_max = flatten_rows ? 1 : ne03;
+    const int64_t i02_max = flatten_rows ? 1 : (ne02 >= ne12 ? ne02 : ne12);
+    const int64_t i02_divisor = ne02 >= ne12 ? 1 : ne12 / ne02;
+    GGML_ASSERT(!(flatten_rows && ne02 < ne12));
+
     const size_t src0_ts = ggml_type_size(src0->type);
     const size_t src0_bs = ggml_blck_size(src0->type);
 
@@ -3089,6 +3251,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
         dst->op == GGML_OP_SCALE || dst->op == GGML_OP_DIAG_MASK_INF || dst->op == GGML_OP_ROPE);
 
     const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
+    GGML_ASSERT(!(split && ne02 < ne12));
 
     const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
 
@@ -3125,7 +3288,7 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
             row_high = id == g_device_count - 1 ? nrows0 : nrows0*g_tensor_split[id + 1];
         } else {
             row_low = 0;
-            row_high = nrows0;
+            row_high = nrows0*i02_divisor;
         }
         if (row_low == row_high) {
             continue;
@@ -3173,16 +3336,12 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
             dst_ddf[id] = (float *) ggml_cuda_pool_malloc(size_dst_ddf, &dst_asf[id]);
         }
 
-        const int64_t i03_max = flatten_rows ? 1 : ne03;
-        const int64_t i02_max = flatten_rows ? 1 : ne02;
-        const int64_t rows_per_iter = flatten_rows ? nrows0 : ne01;
-
         for (int64_t i03 = 0; i03 < i03_max; i03++) {
             const int64_t i13 = i03 % ne13;
             for (int64_t i02 = 0; i02 < i02_max; i02++) {
                 const int64_t i12 = i02 % ne12;
 
-                const int64_t i0 = i03*ne02 + i02;
+                const int64_t i0 = i03*i02_max + i02;
 
                 // i0 values that contain the lower/upper rows for a split tensor when using multiple GPUs
                 const int64_t i0_offset_low = row_low/rows_per_iter;
@@ -3216,10 +3375,10 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                 const int64_t i11 = i13*ne12 + i12;
 
                 // for split tensors the data begins at i0 == i0_offset_low
-                char  * src0_ddq_i = src0_ddq[id] + (i0 - i0_offset_low)*src0_stride*src0_ts/src0_bs;
+                char  * src0_ddq_i = src0_ddq[id] + (i0/i02_divisor - i0_offset_low)*src0_stride*src0_ts/src0_bs;
-                float * src0_ddf_i = src0_ddf[id] + (i0 - i0_offset_low)*src0_stride;
+                float * src0_ddf_i = src0_ddf[id] + (i0/i02_divisor - i0_offset_low)*src0_stride;
                 float * src1_ddf_i = src1_ddf[id] + i11*src1_stride;
-                float * dst_ddf_i  =  dst_ddf[id] + (i0 - i0_offset_low)*dst_stride;
+                float * dst_ddf_i  =  dst_ddf[id] + (i0             - i0_offset_low)*dst_stride;
 
                 // for split tensors the data pointer needs to be rounded down
                 // to the bin edge for i03, i02 bins beyond the first
@@ -3258,11 +3417,11 @@ static void ggml_cuda_op(const ggml_tensor * src0, const ggml_tensor * src1, ggm
                     }
                 }
 
-                if (!src0_on_device || !src0_is_contiguous) {
+                if ((!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
                     if (src0_is_f32) {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
+                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
                     } else {
-                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02, i01_low, i01_high, cudaStream_main));
+                        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddq_i, src0, i03, i02/i02_divisor, i01_low, i01_high, cudaStream_main));
                     }
                 }
 
@@ -3416,6 +3575,8 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     CUDA_CHECK(cudaSetDevice(g_main_device));
     cudaStream_t cudaStream_main = g_cudaStreams_main[g_main_device];
 
@@ -3428,7 +3589,7 @@ void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * sr
     struct ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
     float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
 
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, cudaStream_main);
+    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, cudaStream_main);
 }
 
 void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
@@ -3442,6 +3603,8 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
 
+    const int64_t ne12 = src1->ne[2];
+
     const int64_t nb01 = src0->nb[1];
     const int64_t nb02 = src0->nb[2];
 
@@ -3460,7 +3623,7 @@ void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1
     const int row_stride_x = nb01 / sizeof(half);
     const int channel_stride_x = nb02 / sizeof(half);
 
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, channel_stride_x, cudaStream_main);
+    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, cudaStream_main);
 }
 
 void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3601,7 +3764,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         size_t size = ggml_nbytes_split(tensor, nrows_split);
         const size_t original_size = size;
 
-        // pad last row to a multiple of 256 elements to avoid out-of-bounds memory accesses
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
         if (ne0 % MATRIX_ROW_PADDING != 0) {
             size += (MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING)
                 * ggml_type_size(tensor->type)/ggml_blck_size(tensor->type);
@@ -3617,7 +3780,7 @@ void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
         }
 
 
-        CUDA_CHECK(cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice));
+        CUDA_CHECK(cudaMemcpy(buf, buf_host, original_size, cudaMemcpyHostToDevice));
 
         extra->data_device[id] = buf;
 
@@ -3697,7 +3860,7 @@ void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bo
         char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
         size_t offset = 0;
         if (tensor->op == GGML_OP_VIEW) {
-            memcpy(&offset, tensor->src[2]->data, sizeof(size_t));
+            memcpy(&offset, tensor->op_params, sizeof(size_t));
         }
         extra = ggml_cuda_alloc_temp_tensor_extra();
         extra->data_device[g_main_device] = src0_ddc + offset;

ggml-metal.m

@@ -42,6 +42,7 @@ struct ggml_metal_context {
     id<MTLComputePipelineState> pipeline_##name
 
     GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(mul);
     GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
     GGML_METAL_DECL_KERNEL(scale);
@@ -157,6 +158,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         fprintf(stderr, "%s: loaded %-32s %16p\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name);
 
         GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(add_row);
         GGML_METAL_ADD_KERNEL(mul);
         GGML_METAL_ADD_KERNEL(mul_row);
         GGML_METAL_ADD_KERNEL(scale);
@@ -464,10 +466,16 @@ void ggml_metal_graph_compute(
                                 encoder = [command_buffer computeCommandEncoder];
                             }
 
-                            [encoder setComputePipelineState:ctx->pipeline_add];
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_add];
+                            }
                             [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(ne00) atIndex:3];
 
                             const int64_t n = ggml_nelements(dst);
 
@@ -577,7 +585,7 @@ void ggml_metal_graph_compute(
                                 encoder = [command_buffer computeCommandEncoder];
                             }
 
-                            const int n_past = ((int32_t *)(src1->data))[0];
+                            const int n_past = ((int32_t *)(dst->op_params))[0];
 
                             [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -676,8 +684,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
+                                            nth0 = 2;
-                                            nth1 = 16;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
                                         } break;
                                     case GGML_TYPE_Q3_K:
@@ -685,8 +693,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
+                                            nth0 = 2;
-                                            nth1 = 16;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
                                         } break;
                                     case GGML_TYPE_Q4_K:
@@ -694,8 +702,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
+                                            nth0 = 2;
-                                            nth1 = 16;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
                                         } break;
                                     case GGML_TYPE_Q5_K:
@@ -703,8 +711,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
+                                            nth0 = 2;
-                                            nth1 = 16;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
                                         } break;
                                     case GGML_TYPE_Q6_K:
@@ -712,8 +720,8 @@ void ggml_metal_graph_compute(
                                             GGML_ASSERT(ne02 == 1);
                                             GGML_ASSERT(ne12 == 1);
 
-                                            nth0 = 4;
+                                            nth0 = 2;
-                                            nth1 = 16;
+                                            nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
                                         } break;
                                     default:
@@ -739,16 +747,22 @@ void ggml_metal_graph_compute(
                                 [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:13];
                                 [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:14];
 
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1) {
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7) / 8, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
-                                else if (src0t == GGML_TYPE_Q2_K ||
+                                else if (src0t == GGML_TYPE_Q3_K) {
-                                         src0t == GGML_TYPE_Q3_K ||
+#ifdef GGML_QKK_64
-                                         src0t == GGML_TYPE_Q4_K ||
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                         src0t == GGML_TYPE_Q5_K ||
+#else
-                                         src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+3)/4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                    [encoder setThreadgroupMemoryLength:nth0*nth1*sizeof(float) atIndex:0];
+#endif
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3) / 4, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01+1)/2, ne11, 1) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 } else {
                                     [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
                                     [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -792,7 +806,7 @@ void ggml_metal_graph_compute(
 
                             const float eps = 1e-6f;
 
-                            const int nth = 256;
+                            const int nth = 512;
 
                             [encoder setComputePipelineState:ctx->pipeline_rms_norm];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -800,7 +814,7 @@ 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:nth/32*sizeof(float) atIndex:0];
 
                             const int64_t nrows = ggml_nrows(src0);
 
@@ -836,9 +850,10 @@ void ggml_metal_graph_compute(
 
                             GGML_ASSERT((src0t == GGML_TYPE_F32));
 
-                            const int   n_past   = ((int32_t *) src1->data)[0]; UNUSED(n_past);
+                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
-                            const int   n_head   = ((int32_t *) src1->data)[1];
+                            const int n_head = ((int32_t *) dst->op_params)[1];
-                            const float max_bias = ((float *)   src1->data)[2];
+                            float max_bias;
+                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
                             if (__builtin_popcount(n_head) != 1) {
                                 GGML_ASSERT(false && "only power-of-two n_head implemented");
@@ -876,15 +891,14 @@ void ggml_metal_graph_compute(
                                 encoder = [command_buffer computeCommandEncoder];
                             }
 
-                            const int n_dims = ((int32_t *) src1->data)[1];
+                            const int n_past = ((int32_t *) dst->op_params)[0];
-                            const int mode   = ((int32_t *) src1->data)[2];
+                            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 *)(src1->data))[0];
 
                             float freq_base;
                             float freq_scale;
-                            memcpy(&freq_base,  (int32_t *) src1->data + 4, sizeof(float));
+                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
-                            memcpy(&freq_scale, (int32_t *) src1->data + 5, sizeof(float));
+                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
 
                             [encoder setComputePipelineState:ctx->pipeline_rope];
                             [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
@@ -913,7 +927,9 @@ void ggml_metal_graph_compute(
 
                             [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                         } break;
+                    case GGML_OP_DUP:
                     case GGML_OP_CPY:
+                    case GGML_OP_CONT:
                         {
                             if (encoder == nil) {
                                 encoder = [command_buffer computeCommandEncoder];

ggml.h

@@ -199,6 +199,7 @@
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_SRC           6
 #define GGML_MAX_NAME          48
+#define GGML_MAX_OP_PARAMS     32
 #define GGML_DEFAULT_N_THREADS 4
 
 
@@ -418,6 +419,9 @@ extern "C" {
         // compute data
         enum ggml_op op;
 
+        // op params - allocated as int32_t for alignment
+        int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(uint32_t)];
+
         bool is_param;
 
         struct ggml_tensor * grad;
@@ -1128,9 +1132,9 @@ extern "C" {
             int                   n_past,
             int                   n_dims,
             int                   mode,
+            int                   n_ctx,
             float                 freq_base,
-            float                 freq_scale,
+            float                 freq_scale);
-            int                   n_ctx);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
@@ -1139,7 +1143,8 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   n_past,
             int                   n_dims,
-            int                   mode);
+            int                   mode,
+            int                   n_ctx);
 
     // alibi position embedding
     // in-place, returns view(a)

grammars/arithmetic.gbnf

@@ -0,0 +1,6 @@
+root  ::= (expr "=" ws term "\n")+
+expr  ::= term ([-+*/] term)*
+term  ::= ident | num | "(" ws expr ")" ws
+ident ::= [a-z] [a-z0-9_]* ws
+num   ::= [0-9]+ ws
+ws    ::= [ \t\n]*

grammars/chess.gbnf

@@ -0,0 +1,13 @@
+# Specifies chess moves as a list in algebraic notation, using PGN conventions
+
+# Force first move to "1. ", then any 1-2 digit number after, relying on model to follow the pattern
+root    ::= "1. " move " " move "\n" ([1-9] [0-9]? ". " move " " move "\n")+
+move    ::= (pawn | nonpawn | castle) [+#]?
+
+# piece type, optional file/rank, optional capture, dest file & rank
+nonpawn ::= [NBKQR] [a-h]? [1-8]? "x"? [a-h] [1-8]
+
+# optional file & capture, dest file & rank, optional promotion
+pawn    ::= ([a-h] "x")? [a-h] [1-8] ("=" [NBKQR])?
+
+castle  ::= "O-O" "-O"?

grammars/japanese.gbnf

@@ -0,0 +1,7 @@
+# A probably incorrect grammar for Japanese
+root        ::= jp-char+ ([ \t\n] jp-char+)*
+jp-char     ::= hiragana | katakana | punctuation | cjk
+hiragana    ::= [ぁ-ゟ]
+katakana    ::= [ァ-ヿ]
+punctuation ::= [、-〾]
+cjk         ::= [一-鿿]

grammars/json.gbnf

@@ -0,0 +1,29 @@
+# Grammar for subset of JSON - doesn't support full string or number syntax
+
+root  ::= object
+value ::= object | array | string | number | boolean | "null"
+
+object ::=
+  "{" ws (
+            string ":" ws value
+    ("," ws string ":" ws value)*
+  )? "}"
+
+array  ::=
+  "[" ws (
+            value
+    ("," ws value)*
+  )? "]"
+
+string  ::=
+  "\"" (
+    [^"\\] |
+    "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F]) # escapes
+  )* "\"" ws
+
+# Only plain integers currently
+number  ::= "-"? [0-9]+ ws
+boolean ::= ("true" | "false") ws
+
+# Optional space: by convention, applied in this grammar after literal chars when allowed
+ws ::= ([ \t\n] ws)?

grammars/list.gbnf

@@ -0,0 +1,4 @@
+root ::= item+
+
+# Excludes various line break characters
+item ::= "- " [^\r\n\x0b\x0c\x85\u2028\u2029]+ "\n"

k_quants.c

@@ -3297,8 +3297,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-
+    int8_t aux8[QK_K];
-    uint8_t aux8[QK_K];
     int16_t aux16[16];
     float   sums [8];
     memset(sums, 0, 8*sizeof(float));
@@ -3308,7 +3307,7 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
         const uint8_t * restrict q4 = x[i].qs;
         const uint8_t * restrict hm = x[i].qh;
         const  int8_t * restrict q8 = y[i].qs;
-        uint8_t * restrict a = aux8;
+        int8_t * restrict a = aux8;
         for (int l = 0; l < 32; ++l) {
             a[l+ 0] = q4[l] & 0xF;
             a[l+32] = q4[l]  >> 4;

llama.cpp

@@ -67,6 +67,7 @@ enum e_model {
     MODEL_13B,
     MODEL_30B,
     MODEL_65B,
+    MODEL_70B,
 };
 
 static const size_t kB = 1024;
@@ -98,18 +99,18 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
 }
 
 //
-// memory sizes
+// memory sizes (calculated for n_batch == 512)
 //
 
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0(int n_ctx)
 {
     static std::map<e_model, size_t> k_sizes = {
-        /* empirical scaling, still a guess */
+        { MODEL_3B,   ((size_t) n_ctx / 16ull +  92ull) * MB },
-        { MODEL_3B,   ((size_t) n_ctx / 16ull + 128ull) * MB },
+        { MODEL_7B,   ((size_t) n_ctx / 16ull + 100ull) * MB },
-        { MODEL_7B,   ((size_t) n_ctx / 16ull + 256ull) * MB },
+        { MODEL_13B,  ((size_t) n_ctx / 12ull + 120ull) * MB },
-        { MODEL_13B,  ((size_t) n_ctx / 12ull + 256ull) * MB },
+        { MODEL_30B,  ((size_t) n_ctx /  9ull + 160ull) * MB },
-        { MODEL_30B,  ((size_t) n_ctx / 10ull + 256ull) * MB },
+        { MODEL_65B,  ((size_t) n_ctx /  6ull + 256ull) * MB }, // guess
-        { MODEL_65B,  ((size_t) n_ctx /  8ull + 512ull) * MB },
+        { MODEL_70B,  ((size_t) n_ctx /  7ull + 164ull) * MB },
     };
     return k_sizes;
 }
@@ -117,38 +118,26 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0(int n_ctx)
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
 {
     static std::map<e_model, size_t> k_sizes = {
-        { MODEL_3B,    256ull * MB },
+        { MODEL_3B,  128ull * MB },
-        { MODEL_7B,    512ull * MB },
+        { MODEL_7B,  160ull * MB },
-        { MODEL_13B,   512ull * MB },
+        { MODEL_13B, 192ull * MB },
-        { MODEL_30B,   512ull * MB },
+        { MODEL_30B, 256ull * MB },
-        { MODEL_65B,  1024ull * MB },
+        { MODEL_65B, 384ull * MB }, // guess
+        { MODEL_70B, 304ull * MB },
     };
     return k_sizes;
 }
 
-// 2*n_embd*n_ctx*n_layer*sizeof(float16)
+// used to store the compute graph tensors + non-scratch data
-static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
+static const std::map<e_model, size_t> & MEM_REQ_EVAL()
 {
     static std::map<e_model, size_t> k_sizes = {
-        { MODEL_3B,    682ull * MB },
+        { MODEL_3B,   8ull * MB },
-        { MODEL_7B,   1026ull * MB },
+        { MODEL_7B,  10ull * MB },
-        { MODEL_13B,  1608ull * MB },
+        { MODEL_13B, 12ull * MB },
-        { MODEL_30B,  3124ull * MB },
+        { MODEL_30B, 16ull * MB },
-        { MODEL_65B,  5120ull * MB },
+        { MODEL_65B, 24ull * MB }, // guess
-    };
+        { MODEL_70B, 24ull * MB },
-    return k_sizes;
-}
-
-// this is mostly needed for temporary mul_mat buffers to dequantize the data
-// not actually needed if BLAS is disabled
-static const std::map<e_model, size_t> & MEM_REQ_EVAL(int n_ctx)
-{
-    static std::map<e_model, size_t> k_sizes = {
-        { MODEL_3B,  ((size_t) n_ctx / 256ull +  512ull) * MB },
-        { MODEL_7B,  ((size_t) n_ctx / 256ull +  768ull) * MB },
-        { MODEL_13B, ((size_t) n_ctx / 256ull + 1024ull) * MB },
-        { MODEL_30B, ((size_t) n_ctx / 256ull + 1280ull) * MB },
-        { MODEL_65B, ((size_t) n_ctx / 256ull + 1536ull) * MB },
     };
     return k_sizes;
 }
@@ -163,6 +152,7 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_BASE()
         { MODEL_13B,  640ull * kB },
         { MODEL_30B,  768ull * kB },
         { MODEL_65B, 1536ull * kB },
+        { MODEL_70B, 1536ull * kB }, // TODO (likely can be reduced)
     };
     return k_sizes;
 }
@@ -177,19 +167,25 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
         { MODEL_13B, 160ull },
         { MODEL_30B, 208ull },
         { MODEL_65B, 416ull },
+        { MODEL_70B, 416ull }, // TODO (likely can be reduced)
     };
     return k_sizes;
 }
 
 // default hparams (LLaMA 7B)
 struct llama_hparams {
-    uint32_t n_vocab = 32000;
+    uint32_t n_vocab   = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
+    uint32_t n_ctx     = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
+    uint32_t n_embd    = 4096;
-    uint32_t n_mult  = 256;
+    uint32_t n_mult    = 256;
-    uint32_t n_head  = 32;
+    uint32_t n_head    = 32;
-    uint32_t n_layer = 32;
+    uint32_t n_head_kv = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
+    // LLaMAv2
+    // TODO: load from model data hparams
+    float f_ffn_mult = 1.0f;
 
     float rope_freq_base  = 10000.0f;
     float rope_freq_scale = 1.0f;
@@ -197,7 +193,28 @@ struct llama_hparams {
     enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
 
     bool operator!=(const llama_hparams & other) const {
-        return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams)));
+        return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams))); // NOLINT
+    }
+
+    uint32_t n_gqa() const {
+        return n_head/n_head_kv;
+    }
+
+    uint32_t n_embd_head() const {
+        return n_embd/n_head;
+    }
+
+    uint32_t n_embd_gqa() const {
+        return n_embd/n_gqa();
+    }
+
+    size_t kv_size() const {
+        size_t result = 2ull;
+        result *= (size_t) n_embd_gqa();
+        result *= (size_t) n_ctx;
+        result *= (size_t) n_layer;
+        result *= sizeof(ggml_fp16_t);
+        return result;
     }
 };
 
@@ -499,12 +516,16 @@ struct llama_file_loader {
     }
     void read_hparams() {
         hparams.n_vocab = file.read_u32();
-        hparams.n_embd = file.read_u32();
+        hparams.n_embd  = file.read_u32();
-        hparams.n_mult = file.read_u32();
+        hparams.n_mult  = file.read_u32();
-        hparams.n_head = file.read_u32();
+        hparams.n_head  = file.read_u32();
         hparams.n_layer = file.read_u32();
-        hparams.n_rot = file.read_u32();
+        hparams.n_rot   = file.read_u32();
-        hparams.ftype = (enum llama_ftype) file.read_u32();
+        hparams.ftype   = (enum llama_ftype) file.read_u32();
+
+        // LLaMAv2
+        // TODO: read from header
+        hparams.n_head_kv = hparams.n_head;
     }
     void read_vocab() {
         vocab.id_to_token.resize(hparams.n_vocab);
@@ -555,7 +576,9 @@ struct llama_file_loader {
             }
 
             // skip to the next multiple of 32 bytes
-            file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
+            if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) {
+                file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
+            }
 
             tensor.file_off = file.tell();
             tensor.name = name;
@@ -801,7 +824,7 @@ static bool kv_cache_init(
                          ggml_type   wtype,
                                int   n_ctx,
                                int   n_gpu_layers) {
-    const int n_embd  = hparams.n_embd;
+    const int n_embd  = hparams.n_embd_gqa();
     const int n_layer = hparams.n_layer;
 
     const int64_t n_mem      = n_layer*n_ctx;
@@ -845,9 +868,10 @@ struct llama_context_params llama_context_default_params() {
         /*.seed                        =*/ LLAMA_DEFAULT_SEED,
         /*.n_ctx                       =*/ 512,
         /*.n_batch                     =*/ 512,
+        /*.n_gqa                       =*/ 1,
         /*.gpu_layers                  =*/ 0,
         /*.main_gpu                    =*/ 0,
-        /*.tensor_split                =*/ {0},
+        /*.tensor_split                =*/ nullptr,
         /*.rope_freq_base              =*/ 10000.0f,
         /*.rope_freq_scale             =*/ 1.0f,
         /*.progress_callback           =*/ nullptr,
@@ -964,6 +988,7 @@ static const char *llama_model_type_name(e_model type) {
         case MODEL_13B: return "13B";
         case MODEL_30B: return "30B";
         case MODEL_65B: return "65B";
+        case MODEL_70B: return "70B";
         default: LLAMA_ASSERT(false);
     }
 }
@@ -974,6 +999,7 @@ static void llama_model_load_internal(
         llama_vocab & vocab,
         int n_ctx,
         int n_batch,
+        int n_gqa,
         int n_gpu_layers,
         int main_gpu,
         const float * tensor_split,
@@ -995,6 +1021,7 @@ static void llama_model_load_internal(
     model.hparams = ml->file_loader->hparams;
     model.n_gpu_layers = n_gpu_layers;
     llama_file_version file_version = ml->file_loader->file_version;
+
     auto & hparams = model.hparams;
 
     {
@@ -1014,11 +1041,25 @@ static void llama_model_load_internal(
 
         hparams.n_ctx = n_ctx;
 
+        // LLaMAv2
+        // TODO: temporary until GGUF
+        LLAMA_ASSERT(hparams.n_head % n_gqa == 0);
+        hparams.n_head_kv = hparams.n_head / n_gqa;
+        if (model.type == e_model::MODEL_65B && n_gqa == 8) {
+            fprintf(stderr, "%s: warning: assuming 70B model based on GQA == %d\n", __func__, n_gqa);
+            model.type = e_model::MODEL_70B;
+            hparams.f_ffn_mult = 1.3f; // from the params.json of the 70B model
+        }
+
         hparams.rope_freq_base  = rope_freq_base;
         hparams.rope_freq_scale = rope_freq_scale;
     }
 
-    const uint32_t n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult;
+    // ref: https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/model.py#L194-L199
+    const uint32_t n_ff_raw  = 2*(4*hparams.n_embd)/3;
+    const uint32_t n_ff_mult = hparams.f_ffn_mult*n_ff_raw;
+    const uint32_t n_ff      = ((n_ff_mult + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult;
+    //const uint32_t n_ff = 28672;
 
     {
         fprintf(stderr, "%s: format     = %s\n",   __func__, llama_file_version_name(file_version));
@@ -1027,12 +1068,14 @@ static void llama_model_load_internal(
         fprintf(stderr, "%s: n_embd     = %u\n",   __func__, hparams.n_embd);
         fprintf(stderr, "%s: n_mult     = %u\n",   __func__, hparams.n_mult);
         fprintf(stderr, "%s: n_head     = %u\n",   __func__, hparams.n_head);
+        fprintf(stderr, "%s: n_head_kv  = %u\n",   __func__, hparams.n_head_kv);
         fprintf(stderr, "%s: n_layer    = %u\n",   __func__, hparams.n_layer);
-        fprintf(stderr, "%s: n_rot      = %u\n",   __func__, hparams.n_rot);
+        fprintf(stderr, "%s: n_rot      = %u\n",   __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
+        fprintf(stderr, "%s: n_gqa      = %u\n",   __func__, hparams.n_gqa());
+        fprintf(stderr, "%s: n_ff       = %u\n",   __func__, n_ff);
         fprintf(stderr, "%s: freq_base  = %.1f\n", __func__, hparams.rope_freq_base);
         fprintf(stderr, "%s: freq_scale = %g\n",   __func__, hparams.rope_freq_scale);
         fprintf(stderr, "%s: ftype      = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
-        fprintf(stderr, "%s: n_ff       = %u\n",   __func__, n_ff);
         fprintf(stderr, "%s: model size = %s\n",   __func__, llama_model_type_name(model.type));
     }
 
@@ -1067,7 +1110,7 @@ static void llama_model_load_internal(
     {
         model.buf.resize(ctx_size);
         if (use_mlock) {
-            model.mlock_buf.init(model.buf.addr);
+            model.mlock_buf.init   (model.buf.addr);
             model.mlock_buf.grow_to(model.buf.size);
         }
 
@@ -1102,9 +1145,10 @@ static void llama_model_load_internal(
     size_t vram_weights = 0;
     size_t vram_scratch = 0;
     {
-        const uint32_t n_embd  = hparams.n_embd;
+        const uint32_t n_embd     = hparams.n_embd;
-        const uint32_t n_layer = hparams.n_layer;
+        const uint32_t n_embd_gqa = hparams.n_embd_gqa();
-        const uint32_t n_vocab = hparams.n_vocab;
+        const uint32_t n_layer    = hparams.n_layer;
+        const uint32_t n_vocab    = hparams.n_vocab;
 
         ml->ggml_ctx = ctx;
 
@@ -1152,16 +1196,16 @@ static void llama_model_load_internal(
 
             layer.attention_norm = ml->get_tensor(layers_i + ".attention_norm.weight", {n_embd}, backend);
 
-            layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd}, backend_split);
+            layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd},     backend_split);
-            layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd}, backend_split);
+            layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd_gqa}, backend_split);
-            layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd}, backend_split);
+            layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd_gqa}, backend_split);
-            layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd}, backend_split);
+            layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd},     backend_split);
 
             layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd}, backend);
 
-            layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd,   n_ff},   backend_split);
+            layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd,   n_ff}, backend_split);
-            layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", {  n_ff,   n_embd}, backend_split);
+            layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", {  n_ff, n_embd}, backend_split);
-            layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd,   n_ff},   backend_split);
+            layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd,   n_ff}, backend_split);
 
             if (backend == GGML_BACKEND_GPU) {
                 vram_weights +=
@@ -1184,11 +1228,11 @@ static void llama_model_load_internal(
             mmapped_size - vram_weights + // weights in VRAM not in memory
             MEM_REQ_SCRATCH0(hparams.n_ctx).at(model.type) +
             MEM_REQ_SCRATCH1().at(model.type) +
-            MEM_REQ_EVAL(hparams.n_ctx).at(model.type);
+            MEM_REQ_EVAL().at(model.type);
 
         // this is the memory required by one llama_state
         const size_t mem_required_state =
-            scale*MEM_REQ_KV_SELF().at(model.type);
+            scale*hparams.kv_size();
 
         fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
                 mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
@@ -1229,7 +1273,7 @@ static void llama_model_load_internal(
                 fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
             } else {
                 fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
-                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+                vram_kv_cache += hparams.kv_size() / 2;
             }
         }
         if (n_gpu_layers > (int) hparams.n_layer + 2) {
@@ -1237,7 +1281,7 @@ static void llama_model_load_internal(
                 fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
             } else {
                 fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
-                vram_kv_cache += MEM_REQ_KV_SELF().at(model.type) / 2;
+                vram_kv_cache += hparams.kv_size() / 2;
             }
         }
 #elif defined(GGML_USE_CLBLAST)
@@ -1285,9 +1329,10 @@ static bool llama_model_load(
         llama_vocab & vocab,
         int n_ctx,
         int n_batch,
+        int n_gqa,
         int n_gpu_layers,
         int main_gpu,
-        float * tensor_split,
+        const float * tensor_split,
         float rope_freq_base,
         float rope_freq_scale,
         bool low_vram,
@@ -1298,7 +1343,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gpu_layers, main_gpu, tensor_split, rope_freq_base, rope_freq_scale, low_vram, memory_type,
+        llama_model_load_internal(fname, model, vocab, n_ctx, n_batch, n_gqa, n_gpu_layers, main_gpu, tensor_split, rope_freq_base, rope_freq_scale, low_vram, memory_type,
                                   use_mmap, use_mlock, vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::exception & err) {
@@ -1342,17 +1387,22 @@ static bool llama_eval_internal(
 
     LLAMA_ASSERT(!!kv_self.ctx);
 
-    const int n_embd       = hparams.n_embd;
+    const int64_t n_embd      = hparams.n_embd;
-    const int n_layer      = hparams.n_layer;
+    const int64_t n_layer     = hparams.n_layer;
-    const int n_ctx        = hparams.n_ctx;
+    const int64_t n_ctx       = hparams.n_ctx;
-    const int n_head       = hparams.n_head;
+    const int64_t n_head      = hparams.n_head;
-    const int n_vocab      = hparams.n_vocab;
+    const int64_t n_head_kv   = hparams.n_head_kv;
-    const int n_rot        = hparams.n_embd/hparams.n_head;
+    const int64_t n_embd_head = hparams.n_embd_head();
-    const int n_gpu_layers = model.n_gpu_layers;
+    const int64_t n_vocab     = hparams.n_vocab;
+    const int64_t n_embd_gqa  = hparams.n_embd_gqa();
+
+    LLAMA_ASSERT(n_embd_head == hparams.n_rot);
 
     const float freq_base  = hparams.rope_freq_base;
     const float freq_scale = hparams.rope_freq_scale;
 
+    const int n_gpu_layers = model.n_gpu_layers;
+
     auto & mem_per_token = lctx.mem_per_token;
     auto & buf_compute   = lctx.buf_compute;
 
@@ -1450,11 +1500,11 @@ static bool llama_eval_internal(
             offload_func_kq(tmpq);
             ggml_set_name(tmpq, "tmpq");
 
-            struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, 0);
+            struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpk, n_embd_head, n_head_kv, N), n_past, 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_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, 0);
+            struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, tmpq, n_embd_head, n_head, N),    n_past, n_embd_head, 0, 0, freq_base, freq_scale);
             offload_func_kq(Qcur);
             ggml_set_name(Qcur, "Qcur");
 
@@ -1466,17 +1516,17 @@ static bool llama_eval_internal(
                 offload_func_v(tmpv);
                 ggml_set_name(tmpv, "tmpv");
 
-                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd, N));
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, tmpv, n_embd_gqa, N));
                 offload_func_v(Vcur);
                 ggml_set_name(Vcur, "Vcur");
 
-                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd_gqa, (ggml_element_size(kv_self.k)*n_embd_gqa)*(il*n_ctx + n_past));
                 offload_func_kq(k);
                 ggml_set_name(k, "k");
 
-                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd_gqa,
                         (   n_ctx)*ggml_element_size(kv_self.v),
-                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd_gqa + n_past*ggml_element_size(kv_self.v));
                 offload_func_v(v);
                 ggml_set_name(v, "v");
 
@@ -1495,8 +1545,8 @@ static bool llama_eval_internal(
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
                         ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
+                            ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd_gqa, il*n_ctx*ggml_element_size(kv_self.k)*n_embd_gqa),
-                            n_embd/n_head, n_head, n_past + N),
+                            n_embd_head, n_head_kv, n_past + N),
                         0, 2, 1, 3);
             offload_func_kq(K);
             ggml_set_name(K, "K");
@@ -1506,9 +1556,9 @@ static bool llama_eval_internal(
             offload_func_kq(KQ);
             ggml_set_name(KQ, "KQ");
 
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // KQ_scaled = KQ / sqrt(n_embd_head)
             struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head));
-            ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_head)");
+            ggml_set_name(KQ_scale, "1/sqrt(n_embd_head)");
 
             // KQ_scaled shape [n_past + N, N, n_head, 1]
             struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
@@ -1528,10 +1578,10 @@ static bool llama_eval_internal(
             // split cached V into n_head heads
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
-                        n_past + N, n_embd/n_head, n_head,
+                        n_past + N, n_embd_head, n_head_kv,
                         n_ctx*ggml_element_size(kv_self.v),
-                        n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
+                        n_ctx*ggml_element_size(kv_self.v)*n_embd_head,
-                        il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
+                        n_ctx*ggml_element_size(kv_self.v)*n_embd_gqa*il);
             offload_func_v(V);
             ggml_set_name(V, "V");
 
@@ -1543,7 +1593,7 @@ static bool llama_eval_internal(
             // 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_past + N, n_embd/n_head, n_head));
+            struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd_head, n_head));
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
 #endif
 
@@ -1737,10 +1787,12 @@ static bool llama_eval_internal(
     }
 
 #if 0
-    printf("\n%s: used_mem = %.3f MB, scratch -- %.3f MB %.3f MB\n", __func__,
+    printf("\n%s: used_mem: eval ctx %.3f MB, scratch %.3f MB %.3f MB, work buf %.3f MB, n_past = %d, N = %d\n", __func__,
             ggml_used_mem(ctx0)/1024.0/1024.0,
             lctx.get_buf_max_mem(0)/1024.0/1024.0,
-            lctx.get_buf_max_mem(1)/1024.0/1024.0);
+            lctx.get_buf_max_mem(1)/1024.0/1024.0,
+            lctx.work_buffer.size()/1024.0/1024.0,
+            n_past, N);
 #endif
 
     ggml_free(ctx0);
@@ -1913,6 +1965,279 @@ static std::vector<llama_vocab::id> llama_tokenize(const llama_vocab & vocab, co
     return output;
 }
 
+//
+// grammar - internal
+//
+
+struct llama_grammar {
+    const std::vector<std::vector<llama_grammar_element>>   rules;
+    std::vector<std::vector<const llama_grammar_element *>> stacks;
+};
+
+struct llama_grammar_candidate {
+    size_t           index;
+    const uint32_t * code_points;
+};
+
+// NOTE: assumes valid utf8 (but checks for overrun)
+// adds a terminating 0 for use as pointer
+std::vector<uint32_t> decode_utf8(const char * src) {
+    static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+    const char          * pos      = src;
+    std::vector<uint32_t> code_points;
+    while (*pos != 0) {
+        uint8_t  first_byte = static_cast<uint8_t>(*pos);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = pos + len; // may overrun!
+        ++pos;
+        for ( ; pos < end && *pos != 0; ++pos) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        code_points.push_back(value);
+    }
+    code_points.push_back(0);
+    return code_points;
+}
+
+// returns true iff pos points to the end of one of the definitions of a rule
+static bool llama_grammar_is_end_of_sequence(const llama_grammar_element * pos) {
+    switch (pos->type) {
+        case LLAMA_GRETYPE_END: return true;
+        case LLAMA_GRETYPE_ALT: return true;
+        default:                return false;
+    }
+}
+
+// returns true iff chr satisfies the char range at pos (regular or inverse range)
+// asserts that pos is pointing to a char range element
+static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
+        const llama_grammar_element * pos,
+        const uint32_t                chr) {
+
+    bool found            = false;
+    bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
+    LLAMA_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT);
+
+    do {
+        if (pos[1].type == LLAMA_GRETYPE_CHAR_RNG_UPPER) {
+            // inclusive range, e.g. [a-z]
+            found = found || (pos->value <= chr && chr <= pos[1].value);
+            pos += 2;
+        } else {
+            // exact char match, e.g. [a] or "a"
+            found = found || pos->value == chr;
+            pos += 1;
+        }
+    } while (pos->type == LLAMA_GRETYPE_CHAR_ALT);
+
+    return std::make_pair(found == is_positive_char, pos);
+}
+
+// transforms a grammar pushdown stack into N possible stacks, all ending
+// at a character range (terminal element)
+static void llama_grammar_advance_stack(
+        const std::vector<std::vector<llama_grammar_element>>   & rules,
+        const std::vector<const llama_grammar_element *>        & stack,
+        std::vector<std::vector<const llama_grammar_element *>> & new_stacks) {
+
+    if (stack.empty()) {
+        new_stacks.push_back(stack);
+        return;
+    }
+
+    const llama_grammar_element * pos = stack.back();
+
+    switch (pos->type) {
+        case LLAMA_GRETYPE_RULE_REF: {
+            const size_t                  rule_id = static_cast<size_t>(pos->value);
+            const llama_grammar_element * subpos  = rules[rule_id].data();
+            do {
+                // init new stack without the top (pos)
+                std::vector<const llama_grammar_element *> new_stack(stack.begin(), stack.end() - 1);
+                if (!llama_grammar_is_end_of_sequence(pos + 1)) {
+                    // if this rule ref is followed by another element, add that to stack
+                    new_stack.push_back(pos + 1);
+                }
+                if (!llama_grammar_is_end_of_sequence(subpos)) {
+                    // if alternate is nonempty, add to stack
+                    new_stack.push_back(subpos);
+                }
+                llama_grammar_advance_stack(rules, new_stack, new_stacks);
+                while (!llama_grammar_is_end_of_sequence(subpos)) {
+                    // scan to end of alternate def
+                    subpos++;
+                }
+                if (subpos->type == LLAMA_GRETYPE_ALT) {
+                    // there's another alternate def of this rule to process
+                    subpos++;
+                } else {
+                    break;
+                }
+            } while (true);
+            break;
+        }
+        case LLAMA_GRETYPE_CHAR:
+        case LLAMA_GRETYPE_CHAR_NOT:
+            new_stacks.push_back(stack);
+            break;
+        default:
+            // end of alternate (LLAMA_GRETYPE_END, LLAMA_GRETYPE_ALT) or middle of char range
+            // (LLAMA_GRETYPE_CHAR_ALT, LLAMA_GRETYPE_CHAR_RNG_UPPER); stack should never be left on
+            // those
+            LLAMA_ASSERT(false);
+    }
+}
+
+// takes a set of possible pushdown stacks on a grammar, which are required to
+// be positioned at a character range (see `llama_grammar_advance_stack`), and
+// produces the N possible stacks if the given char is accepted at those
+// positions
+static std::vector<std::vector<const llama_grammar_element *>> llama_grammar_accept(
+        const std::vector<std::vector<llama_grammar_element>>         & rules,
+        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
+        const uint32_t                                                  chr) {
+
+    std::vector<std::vector<const llama_grammar_element *>> new_stacks;
+
+    for (const auto & stack : stacks) {
+        if (stack.empty()) {
+            continue;
+        }
+
+        auto match = llama_grammar_match_char(stack.back(), chr);
+        if (match.first) {
+            const llama_grammar_element * pos = match.second;
+
+            // update top of stack to next element, if any
+            std::vector<const llama_grammar_element *> new_stack(stack.begin(), stack.end() - 1);
+            if (!llama_grammar_is_end_of_sequence(pos)) {
+                new_stack.push_back(pos);
+            }
+            llama_grammar_advance_stack(rules, new_stack, new_stacks);
+        }
+    }
+
+    return new_stacks;
+}
+
+static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
+        const std::vector<std::vector<llama_grammar_element>>         & rules,
+        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
+        const std::vector<llama_grammar_candidate>                    & candidates);
+
+static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
+        const std::vector<std::vector<llama_grammar_element>> & rules,
+        const std::vector<const llama_grammar_element *>      & stack,
+        const std::vector<llama_grammar_candidate>            & candidates) {
+
+    std::vector<llama_grammar_candidate> rejects;
+
+    if (stack.empty()) {
+        // accept nothing; EOS is handled elsewhere
+        rejects.insert(rejects.end(), candidates.begin(), candidates.end());
+        return rejects;
+    }
+
+    const llama_grammar_element * stack_pos = stack.back();
+
+    std::vector<llama_grammar_candidate> next_candidates;
+    for (auto tok : candidates) {
+        if (llama_grammar_match_char(stack_pos, tok.code_points[0]).first) {
+            if (tok.code_points[1] != 0) {
+                next_candidates.push_back({ tok.index, tok.code_points + 1 });
+            }
+        } else {
+            rejects.push_back(tok);
+        }
+    }
+
+    auto stack_pos_after = llama_grammar_match_char(stack_pos, 0).second;
+
+    // update top of stack to next element, if any
+    std::vector<const llama_grammar_element *> stack_after(stack.begin(), stack.end() - 1);
+    if (!llama_grammar_is_end_of_sequence(stack_pos_after)) {
+        stack_after.push_back(stack_pos_after);
+    }
+    std::vector<std::vector<const llama_grammar_element *>> next_stacks;
+    llama_grammar_advance_stack(rules, stack_after, next_stacks);
+
+    auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
+    for (auto tok : next_rejects) {
+        rejects.push_back({ tok.index, tok.code_points - 1 });
+    }
+
+    return rejects;
+}
+
+static std::vector<llama_grammar_candidate> llama_grammar_reject_candidates(
+        const std::vector<std::vector<llama_grammar_element>>         & rules,
+        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
+        const std::vector<llama_grammar_candidate>                    & candidates) {
+    LLAMA_ASSERT(!stacks.empty()); // REVIEW
+
+    if (candidates.empty()) {
+        return std::vector<llama_grammar_candidate>();
+    }
+
+    auto rejects = llama_grammar_reject_candidates_for_stack(rules, stacks.front(), candidates);
+
+    for (size_t i = 1, size = stacks.size(); i < size; ++i) {
+        rejects = llama_grammar_reject_candidates_for_stack(rules, stacks[i], rejects);
+    }
+    return rejects;
+}
+
+//
+// grammar - external
+//
+
+struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index) {
+    const llama_grammar_element * pos;
+
+    // copy rule definitions into vectors
+    std::vector<std::vector<llama_grammar_element>> vec_rules(n_rules);
+    for (size_t i = 0; i < n_rules; i++) {
+        for (pos = rules[i]; pos->type != LLAMA_GRETYPE_END; pos++) {
+            vec_rules[i].push_back(*pos);
+        }
+        vec_rules[i].push_back({LLAMA_GRETYPE_END, 0});
+    }
+
+    // loop over alternates of start rule to build initial stacks
+    std::vector<std::vector<const llama_grammar_element *>> stacks;
+    pos = rules[start_rule_index];
+    do {
+        std::vector<const llama_grammar_element *> stack;
+        if (!llama_grammar_is_end_of_sequence(pos)) {
+            // if alternate is nonempty, add to stack
+            stack.push_back(pos);
+        }
+        llama_grammar_advance_stack(vec_rules, stack, stacks);
+        while (!llama_grammar_is_end_of_sequence(pos)) {
+            // scan to end of alternate def
+            pos++;
+        }
+        if (pos->type == LLAMA_GRETYPE_ALT) {
+            // there's another alternate def of this rule to process
+            pos++;
+        } else {
+            break;
+        }
+    } while (true);
+
+    return new llama_grammar{ std::move(vec_rules), std::move(stacks) };
+}
+
+void llama_grammar_free(struct llama_grammar * grammar) {
+    delete grammar;
+}
+
 //
 // sampling
 //
@@ -2198,6 +2523,47 @@ void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, l
     }
 }
 
+void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar) {
+    assert(ctx);
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    bool allow_eos = false;
+    for (const auto & stack : grammar->stacks) {
+        if (stack.empty()) {
+            allow_eos = true;
+            break;
+        }
+    }
+
+    const llama_token eos = llama_token_eos();
+
+    std::vector<std::vector<uint32_t>>   candidates_decoded;
+    std::vector<llama_grammar_candidate> candidates_grammar;
+
+    for (size_t i = 0; i < candidates->size; ++i) {
+        const llama_token id  = candidates->data[i].id;
+        const char *      str = llama_token_to_str(ctx, id);
+        if (id == eos) {
+            if (!allow_eos) {
+                candidates->data[i].logit = -INFINITY;
+            }
+        } else if (*str == 0) {
+            candidates->data[i].logit = -INFINITY;
+        } else {
+            candidates_decoded.push_back(decode_utf8(str));
+            candidates_grammar.push_back({ i, candidates_decoded.back().data() });
+        }
+    }
+
+    const auto rejects =
+        llama_grammar_reject_candidates(grammar->rules, grammar->stacks, candidates_grammar);
+    for (auto & reject : rejects) {
+        candidates->data[reject.index].logit = -INFINITY;
+    }
+
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+}
+
 static void llama_log_softmax(float * array, size_t size) {
     float max_l = *std::max_element(array, array + size);
     float sum = 0.f;
@@ -2216,8 +2582,7 @@ void llama_sample_classifier_free_guidance(
           struct llama_context * ctx,
         llama_token_data_array * candidates,
           struct llama_context * guidance_ctx,
-                         float   scale,
+                         float   scale) {
-                         float   smooth_factor) {
     int64_t t_start_sample_us = ggml_time_us();
 
     assert(ctx);
@@ -2238,16 +2603,7 @@ void llama_sample_classifier_free_guidance(
     for (int i = 0; i < n_vocab; ++i) {
         float logit_guidance = logits_guidance[i];
         float logit_base = logits_base[i];
-        logits_guidance[i] = scale * (logit_base - logit_guidance) + logit_guidance;
+        candidates->data[i].logit = scale * (logit_base - logit_guidance) + logit_guidance;
-    }
-
-    llama_log_softmax(logits_guidance, n_vocab);
-
-    for (int i = 0; i < n_vocab; ++i) {
-        float logit_base = logits_base[i];
-        float logit_guidance = logits_guidance[i];
-
-        candidates->data[i].logit = smooth_factor * logit_guidance + (1.f - smooth_factor) * logit_base;
     }
 
     if (ctx) {
@@ -2383,6 +2739,29 @@ llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_arra
     return result;
 }
 
+void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token) {
+    const int64_t t_start_sample_us = ggml_time_us();
+
+    if (token == llama_token_eos()) {
+        for (const auto & stack : grammar->stacks) {
+            if (stack.empty()) {
+                return;
+            }
+        }
+        LLAMA_ASSERT(false);
+    }
+
+    const char * str = llama_token_to_str(ctx, token);
+    // Note terminating 0 in decoded string
+    auto code_points = decode_utf8(str);
+    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+        grammar->stacks = llama_grammar_accept(grammar->rules, grammar->stacks, *it);
+    }
+    LLAMA_ASSERT(!grammar->stacks.empty());
+
+    ctx->t_sample_us += ggml_time_us() - t_start_sample_us;
+}
+
 //
 // quantization
 //
@@ -2456,8 +2835,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_0: quantized_type = GGML_TYPE_Q5_0; break;
         case LLAMA_FTYPE_MOSTLY_Q5_1: quantized_type = GGML_TYPE_Q5_1; break;
         case LLAMA_FTYPE_MOSTLY_Q8_0: quantized_type = GGML_TYPE_Q8_0; break;
-        case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break;
+        case LLAMA_FTYPE_MOSTLY_F16:  quantized_type = GGML_TYPE_F16;  break;
-        case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break;
+        case LLAMA_FTYPE_ALL_F32:     quantized_type = GGML_TYPE_F32;  break;
 
 #ifdef GGML_USE_K_QUANTS
         // K-quants
@@ -2541,16 +2920,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         } else {
             new_type = quantized_type;
 #ifdef GGML_USE_K_QUANTS
-            bool convert_incompatible_tensor = false;
-            if (quantized_type == GGML_TYPE_Q2_K || quantized_type == GGML_TYPE_Q3_K || quantized_type == GGML_TYPE_Q4_K ||
-                quantized_type == GGML_TYPE_Q5_K || quantized_type == GGML_TYPE_Q6_K) {
-                int nx = tensor.ne.at(0);
-                int ny = tensor.ne.at(1);
-                if (nx % QK_K != 0 || ny % QK_K != 0) {
-                    fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
-                    convert_incompatible_tensor = true;
-                }
-            }
             if (tensor.name == "output.weight") {
                 int nx = tensor.ne.at(0);
                 int ny = tensor.ne.at(1);
@@ -2576,6 +2945,16 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 if      (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q4_K;
                 else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
             }
+            bool convert_incompatible_tensor = false;
+            if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
+                new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
+                int nx = tensor.ne.at(0);
+                int ny = tensor.ne.at(1);
+                if (nx % QK_K != 0 || ny % QK_K != 0) {
+                    fprintf(stderr, "\n\nTensor sizes %d x %d are not divisible by %d, required for k-quants.\n",nx,ny,QK_K);
+                    convert_incompatible_tensor = true;
+                }
+            }
             if (convert_incompatible_tensor) {
                 if (tensor.name == "output.weight") {
                     new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
@@ -2602,7 +2981,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 f32_data = (float *) f32_conv_buf.addr;
             }
 
-            printf("quantizing .. ");
+            printf("quantizing to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
             work.resize(nelements * 4); // upper bound on size
@@ -2705,7 +3084,7 @@ struct llama_model * llama_load_model_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gpu_layers,
+    if (!llama_model_load(path_model, *model, model->vocab, params.n_ctx, params.n_batch, params.n_gqa, params.n_gpu_layers,
                 params.main_gpu, params.tensor_split, params.rope_freq_base, params.rope_freq_scale,params.low_vram,
                 memory_type, params.use_mmap, params.use_mlock, params.vocab_only, params.progress_callback,
                 params.progress_callback_user_data)) {
@@ -2783,7 +3162,7 @@ struct llama_context * llama_new_context_with_model(
             ctx->embedding.resize(hparams.n_embd);
         }
 
-        ctx->buf_compute.resize(MEM_REQ_EVAL(hparams.n_ctx).at(ctx->model.type));
+        ctx->buf_compute.resize(MEM_REQ_EVAL().at(ctx->model.type));
 
         ctx->buf_scratch[0].resize(MEM_REQ_SCRATCH0(hparams.n_ctx).at(ctx->model.type));
         ctx->buf_scratch[1].resize(MEM_REQ_SCRATCH1().at(ctx->model.type));
@@ -2807,7 +3186,7 @@ struct llama_context * llama_new_context_with_model(
 
         const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
 
-        printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+        fprintf(stderr, "%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
 
 #define LLAMA_METAL_CHECK_BUF(result)                                          \
     if (!(result)) {                                                           \

llama.h

@@ -83,12 +83,14 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
    struct llama_context_params {
-        uint32_t seed;                         // RNG seed, -1 for random
+        uint32_t seed;         // RNG seed, -1 for random
-        int32_t  n_ctx;                        // text context
+        int32_t  n_ctx;        // text context
-        int32_t  n_batch;                      // prompt processing batch size
+        int32_t  n_batch;      // prompt processing batch size
-        int32_t  n_gpu_layers;                 // number of layers to store in VRAM
+        int32_t  n_gqa;        // grouped-query attention (TEMP - will be moved to model hparams)
-        int32_t  main_gpu;                     // the GPU that is used for scratch and small tensors
+        int32_t  n_gpu_layers; // number of layers to store in VRAM
-        float tensor_split[LLAMA_MAX_DEVICES]; // how to split layers across multiple GPUs
+        int32_t  main_gpu;     // the GPU that is used for scratch and small tensors
+
+        const float * tensor_split; // how to split layers across multiple GPUs (size: LLAMA_MAX_DEVICES)
 
         // ref: https://github.com/ggerganov/llama.cpp/pull/2054
         float    rope_freq_base;  // RoPE base frequency
@@ -139,6 +141,40 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
     } llama_model_quantize_params;
 
+    // grammar types
+    struct llama_grammar;
+
+    // grammar element type
+    enum llama_gretype {
+        // end of rule definition
+        LLAMA_GRETYPE_END            = 0,
+
+        // start of alternate definition for rule
+        LLAMA_GRETYPE_ALT            = 1,
+
+        // non-terminal element: reference to rule
+        LLAMA_GRETYPE_RULE_REF       = 2,
+
+        // terminal element: character (code point)
+        LLAMA_GRETYPE_CHAR           = 3,
+
+        // inverse char(s) ([^a], [^a-b] [^abc])
+        LLAMA_GRETYPE_CHAR_NOT       = 4,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or LLAMA_GRETYPE_CHAR_ALT to
+        // be an inclusive range ([a-z])
+        LLAMA_GRETYPE_CHAR_RNG_UPPER = 5,
+
+        // modifies a preceding LLAMA_GRETYPE_CHAR or
+        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
+        LLAMA_GRETYPE_CHAR_ALT       = 6,
+    };
+
+    typedef struct llama_grammar_element {
+        enum llama_gretype type;
+        uint32_t           value; // Unicode code point or rule ID
+    } llama_grammar_element;
+
     // performance timing information
     struct llama_timings {
         double t_start_ms;
@@ -331,6 +367,15 @@ extern "C" {
     LLAMA_API llama_token llama_token_eos();  // end-of-sentence
     LLAMA_API llama_token llama_token_nl();   // next-line
 
+    // Grammar
+    //
+    LLAMA_API struct llama_grammar * llama_grammar_init(
+            const llama_grammar_element ** rules,
+                                 size_t    n_rules,
+                                 size_t    start_rule_index);
+
+    LLAMA_API void llama_grammar_free(struct llama_grammar * grammar);
+
     // Sampling functions
 
     /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
@@ -343,13 +388,11 @@ extern "C" {
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, the logits must be directly extracted from the original generation context without being sorted.
     /// @params guidance_ctx A separate context from the same model. Other than a negative prompt at the beginning, it should have all generated and user input tokens copied from the main context.
     /// @params scale Guidance strength. 1.0f means no guidance. Higher values mean stronger guidance.
-    /// @params smooth_factor Smooth factor between guidance logits and original logits. 1.0f means only use guidance logits. 0.0f means only original logits.
     LLAMA_API void llama_sample_classifier_free_guidance(
               struct llama_context * ctx,
             llama_token_data_array * candidates,
               struct llama_context * guidance_ctx,
-                             float   scale,
+                             float   scale);
-                             float   smooth_factor);
 
     /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
     LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
@@ -367,6 +410,9 @@ extern "C" {
     LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
     LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
 
+    /// @details Apply constraints from grammar
+    LLAMA_API void llama_sample_grammar(struct llama_context * ctx, llama_token_data_array * candidates, const struct llama_grammar * grammar);
+
     /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
     /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
     /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
@@ -388,6 +434,9 @@ extern "C" {
     /// @details Randomly selects a token from the candidates based on their probabilities.
     LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
 
+    /// @details Accepts the sampled token into the grammar
+    LLAMA_API void llama_grammar_accept_token(struct llama_context * ctx, struct llama_grammar * grammar, llama_token token);
+
     // Performance information
     LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
     LLAMA_API void llama_print_timings(struct llama_context * ctx);