root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-11 13:30:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/convert-separate-extra-tensors

Browse Source
This commit is contained in:
Francis Couture-Harpin 2024-09-16 12:01:12 -04:00
commit ed0f2c4ab1
107 changed files with 7085 additions and 5910 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
.github/workflows/build.yml vendored
View File

@ -23,6 +23,9 @@ env:
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
GGML_NLOOP: 3
GGML_N_THREADS: 1
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
jobs:
macOS-latest-cmake-arm64:
@ -375,7 +378,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -401,7 +404,7 @@ jobs:
continue-on-error: true
steps:
- uses: actions/checkout@v2
- uses: actions/checkout@v4
- name: add oneAPI to apt
shell: bash
@ -442,7 +445,7 @@ jobs:
continue-on-error: true
steps:
- uses: actions/checkout@v2
- uses: actions/checkout@v4
- name: add oneAPI to apt
shell: bash
@ -546,7 +549,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -576,7 +579,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -610,7 +613,7 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v1
uses: actions/checkout@v4
- name: Dependencies
id: depends
@ -969,14 +972,14 @@ jobs:
steps:
- name: Clone
id: checkout
uses: actions/checkout@v3
uses: actions/checkout@v4
- name: Install
id: depends
run: |
$ErrorActionPreference = "Stop"
write-host "Downloading AMD HIP SDK Installer"
Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
write-host "Installing AMD HIP SDK"
Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
write-host "Completed AMD HIP SDK installation"

7
.github/workflows/server.yml vendored
View File

@ -20,6 +20,12 @@ on:
types: [opened, synchronize, reopened]
paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
env:
LLAMA_LOG_COLORS: 1
LLAMA_LOG_PREFIX: 1
LLAMA_LOG_TIMESTAMPS: 1
LLAMA_LOG_VERBOSITY: 10
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
cancel-in-progress: true
@ -173,6 +179,7 @@ jobs:
if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
run: |
cd examples/server/tests
$env:PYTHONIOENCODING = ":replace"
behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
- name: Slow tests

14
CMakeLists.txt
View File

@ -82,11 +82,11 @@ set(GGML_FATAL_WARNINGS ${LLAMA_FATAL_WARNINGS})
# change the default for these ggml options
if (NOT DEFINED GGML_LLAMAFILE)
set(GGML_LLAMAFILE ON)
set(GGML_LLAMAFILE_DEFAULT ON)
endif()
if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
set(GGML_CUDA_USE_GRAPHS ON)
if (NOT DEFINED GGML_CUDA_GRAPHS)
set(GGML_CUDA_GRAPHS_DEFAULT ON)
endif()
# transition helpers
@ -139,10 +139,16 @@ set(LLAMA_BIN_INSTALL_DIR ${CMAKE_INSTALL_BINDIR} CACHE PATH "Location o
# determining _precisely_ which defines are necessary for the llama-config
# package.
#
set(GGML_TRANSIENT_DEFINES)
get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
if (GGML_DIR_DEFINES)
list(APPEND GGML_TRANSIENT_DEFINES ${GGML_DIR_DEFINES})
endif()
get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
if (GGML_TARGET_DEFINES)
list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
endif()
get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)

57
Makefile
View File

@ -54,6 +54,7 @@ TEST_TARGETS = \
tests/test-grammar-parser \
tests/test-json-schema-to-grammar \
tests/test-llama-grammar \
tests/test-log \
tests/test-model-load-cancel \
tests/test-opt \
tests/test-quantize-fns \
@ -148,6 +149,14 @@ GGML_NO_METAL := 1
DEPRECATE_WARNING := 1
endif
ifdef LLAMA_DISABLE_LOGS
REMOVE_WARNING := 1
endif
ifdef LLAMA_SERVER_VERBOSE
REMOVE_WARNING := 1
endif
ifndef UNAME_S
UNAME_S := $(shell uname -s)
endif
@ -351,19 +360,11 @@ ifdef LLAMA_SANITIZE_UNDEFINED
MK_LDFLAGS += -fsanitize=undefined -g
endif
ifdef LLAMA_SERVER_VERBOSE
MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
endif
ifdef LLAMA_SERVER_SSL
MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
MK_LDFLAGS += -lssl -lcrypto
endif
ifdef LLAMA_DISABLE_LOGS
MK_CPPFLAGS += -DLOG_DISABLE_LOGS
endif # LLAMA_DISABLE_LOGS
# warnings
WARN_FLAGS = \
-Wall \
@ -434,7 +435,7 @@ endif
# 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
ifndef RISCV
ifndef RISCV_CROSS_COMPILE
ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
# Use all CPU extensions that are available:
@ -514,7 +515,12 @@ ifneq ($(filter loongarch64%,$(UNAME_M)),)
MK_CXXFLAGS += -mlasx
endif
else
ifneq ($(filter riscv64%,$(UNAME_M)),)
MK_CFLAGS += -march=rv64gcv -mabi=lp64d
MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
endif
else # RISC-V CROSS COMPILATION
MK_CFLAGS += -march=rv64gcv -mabi=lp64d
MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
endif
@ -613,7 +619,7 @@ ifdef GGML_CUDA
CUDA_PATH ?= /usr/local/cuda
endif
MK_CPPFLAGS += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
MK_CPPFLAGS += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
MK_LDFLAGS += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
MK_NVCCFLAGS += -use_fast_math
endif # GGML_MUSA
@ -925,6 +931,8 @@ OBJ_LLAMA = \
OBJ_COMMON = \
common/common.o \
common/arg.o \
common/log.o \
common/console.o \
common/ngram-cache.o \
common/sampling.o \
@ -1021,6 +1029,14 @@ $(info - LLAMA_NO_CCACHE)
$(info )
endif
ifdef REMOVE_WARNING
$(info !!! REMOVAL WARNING !!!)
$(info The following LLAMA_ options have been removed and are no longer supported)
$(info - LLAMA_DISABLE_LOGS (https://github.com/ggerganov/llama.cpp/pull/9418))
$(info - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418))
$(info )
endif
#
# Build libraries
#
@ -1157,6 +1173,16 @@ common/common.o: \
include/llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/arg.o: \
common/arg.cpp \
common/arg.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/log.o: \
common/log.cpp \
common/log.h
$(CXX) $(CXXFLAGS) -c $< -o $@
common/sampling.o: \
common/sampling.cpp \
common/sampling.h \
@ -1335,7 +1361,7 @@ llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
$(OBJ_GGML) $(OBJ_LLAMA)
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@ -1429,6 +1455,7 @@ llama-server: \
examples/server/system-prompts.js.hpp \
examples/server/prompt-formats.js.hpp \
examples/server/json-schema-to-grammar.mjs.hpp \
examples/server/loading.html.hpp \
common/json.hpp \
common/stb_image.h \
$(OBJ_ALL)
@ -1448,7 +1475,6 @@ llama-gen-docs: examples/gen-docs/gen-docs.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
./llama-gen-docs
libllava.a: examples/llava/llava.cpp \
examples/llava/llava.h \
@ -1517,6 +1543,11 @@ tests/test-llama-grammar: tests/test-llama-grammar.cpp \
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
tests/test-log: tests/test-log.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
tests/test-grammar-parser: tests/test-grammar-parser.cpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

6
README.md
View File

@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
## Hot topics
- *add hot topics here*
- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
----
@ -77,6 +77,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
- [x] [OLMo](https://allenai.org/olmo)
- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
@ -89,6 +90,7 @@ Typically finetunes of the base models below are supported as well.
- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
@ -163,6 +165,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
- [AIKit](https://github.com/sozercan/aikit) (MIT)
- [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
@ -171,6 +174,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
- [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
- [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
**Infrastructure:**

3
ci/run.sh
View File

@ -737,6 +737,9 @@ function gg_sum_embd_bge_small {
## main
export LLAMA_LOG_PREFIX=1
export LLAMA_LOG_TIMESTAMPS=1
if [ -z ${GG_BUILD_LOW_PERF} ]; then
# Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
rm -rf ${SRC}/models-mnt

20
common/CMakeLists.txt
View File

@ -51,19 +51,23 @@ endif()
set(TARGET common)
add_library(${TARGET} STATIC
arg.cpp
arg.h
base64.hpp
common.h
common.cpp
sampling.h
sampling.cpp
console.h
common.h
console.cpp
json.hpp
console.h
json-schema-to-grammar.cpp
train.h
train.cpp
ngram-cache.h
json.hpp
log.cpp
log.h
ngram-cache.cpp
ngram-cache.h
sampling.cpp
sampling.h
train.cpp
train.h
)
if (BUILD_SHARED_LIBS)

1994
common/arg.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

77
common/arg.h Normal file
View File

@ -0,0 +1,77 @@
#pragma once
#include "common.h"
#include <set>
#include <string>
#include <vector>
//
// CLI argument parsing
//
struct llama_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
const char * env = nullptr;
std::string help;
bool is_sparam = false; // is current arg a sampling param?
void (*handler_void) (gpt_params & params) = nullptr;
void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
void (*handler_int) (gpt_params & params, int) = nullptr;
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &)
) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, int)
) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
llama_arg(
const std::initializer_list<const char *> & args,
const std::string & help,
void (*handler)(gpt_params & params)
) : args(args), help(help), handler_void(handler) {}
// support 2 values for arg
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const char * value_hint_2,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &, const std::string &)
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
llama_arg & set_examples(std::initializer_list<enum llama_example> examples);
llama_arg & set_env(const char * env);
llama_arg & set_sparam();
bool in_example(enum llama_example ex);
bool get_value_from_env(std::string & output);
bool has_value_from_env();
std::string to_string();
};
struct gpt_params_context {
enum llama_example ex = LLAMA_EXAMPLE_COMMON;
gpt_params & params;
std::vector<llama_arg> options;
void(*print_usage)(int, char **) = nullptr;
gpt_params_context(gpt_params & params) : params(params) {}
};
// parse input arguments from CLI
// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
// function to be used by test-arg-parser
gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

2169
common/common.cpp
View File

File diff suppressed because it is too large Load Diff

228
common/common.h
View File

@ -4,20 +4,9 @@
#include "llama.h"
#include "sampling.h"
#define LOG_NO_FILE_LINE_FUNCTION
#include "log.h"
#include <cmath>
#include <string>
#include <vector>
#include <random>
#include <thread>
#include <set>
#include <unordered_map>
#include <tuple>
#include <functional>
#include <sstream>
#ifdef _WIN32
#define DIRECTORY_SEPARATOR '\\'
@ -56,11 +45,20 @@ struct llama_control_vector_load_info;
// CPU utils
//
struct cpu_params {
int n_threads = -1;
bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
bool mask_valid = false; // Default: any CPU
enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
bool strict_cpu = false; // Use strict CPU placement
uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling)
};
int32_t cpu_get_num_physical_cores();
int32_t cpu_get_num_math();
//
// CLI argument parsing
// Common params
//
enum llama_example {
@ -78,28 +76,72 @@ enum llama_example {
LLAMA_EXAMPLE_CVECTOR_GENERATOR,
LLAMA_EXAMPLE_EXPORT_LORA,
LLAMA_EXAMPLE_LLAVA,
LLAMA_EXAMPLE_LOOKUP,
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_COUNT,
};
enum gpt_sampler_type {
GPT_SAMPLER_TYPE_NONE = 0,
GPT_SAMPLER_TYPE_TOP_K = 1,
GPT_SAMPLER_TYPE_TOP_P = 2,
GPT_SAMPLER_TYPE_MIN_P = 3,
GPT_SAMPLER_TYPE_TFS_Z = 4,
GPT_SAMPLER_TYPE_TYPICAL_P = 5,
GPT_SAMPLER_TYPE_TEMPERATURE = 6,
};
// dimensionality reduction methods, used by cvector-generator
enum dimre_method {
DIMRE_METHOD_PCA,
DIMRE_METHOD_MEAN,
};
struct cpu_params {
int n_threads = -1;
bool cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
bool mask_valid = false; // Default: any CPU
enum ggml_sched_priority priority = GGML_SCHED_PRIO_NORMAL; // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
bool strict_cpu = false; // Use strict CPU placement
uint32_t poll = 50; // Polling (busywait) level (0 - no polling, 100 - mostly polling)
// sampler parameters
struct gpt_sampler_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
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
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
std::vector<enum gpt_sampler_type> samplers = {
GPT_SAMPLER_TYPE_TOP_K,
GPT_SAMPLER_TYPE_TFS_Z,
GPT_SAMPLER_TYPE_TYPICAL_P,
GPT_SAMPLER_TYPE_TOP_P,
GPT_SAMPLER_TYPE_MIN_P,
GPT_SAMPLER_TYPE_TEMPERATURE
};
std::string grammar; // optional BNF-like grammar to constrain sampling
std::vector<llama_logit_bias> logit_bias; // logit biases to apply
// print the parameters into a string
std::string print() const;
};
struct gpt_params {
enum llama_example curr_ex = LLAMA_EXAMPLE_COMMON;
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@ -143,23 +185,23 @@ struct gpt_params {
struct gpt_sampler_params sparams;
std::string model = ""; // model path
std::string model_draft = ""; // draft model for speculative decoding
std::string model_alias = "unknown"; // model alias
std::string model_url = ""; // model url to download
std::string hf_token = ""; // HF token
std::string hf_repo = ""; // HF repo
std::string hf_file = ""; // HF file
std::string prompt = "";
std::string prompt_file = ""; // store the external prompt file name
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 logdir = ""; // directory in which to save YAML log files
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
std::string logits_file = ""; // file for saving *all* logits
std::string rpc_servers = ""; // comma separated list of RPC servers
std::string model = ""; // model path // NOLINT
std::string model_draft = ""; // draft model for speculative decoding // NOLINT
std::string model_alias = "unknown"; // model alias // NOLINT
std::string model_url = ""; // model url to download // NOLINT
std::string hf_token = ""; // HF token // NOLINT
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string prompt = ""; // NOLINT
std::string prompt_file = ""; // store the external prompt file name // NOLINT
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state // NOLINT
std::string input_prefix = ""; // string to prefix user inputs with // NOLINT
std::string input_suffix = ""; // string to suffix user inputs with // NOLINT
std::string logdir = ""; // directory in which to save YAML log files // NOLINT
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT
std::string logits_file = ""; // file for saving *all* logits // NOLINT
std::string rpc_servers = ""; // comma separated list of RPC servers // NOLINT
std::vector<std::string> in_files; // all input files
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
@ -189,7 +231,6 @@ struct gpt_params {
bool kl_divergence = false; // compute KL divergence
std::function<void(int, char **)> print_usage = nullptr; // print example-specific usage and example
bool usage = false; // print usage
bool use_color = false; // use color to distinguish generations and inputs
bool special = false; // enable special token output
@ -204,6 +245,8 @@ struct gpt_params {
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
bool cont_batching = true; // insert new sequences for decoding on-the-fly
bool flash_attn = false; // flash attention
bool no_perf = false; // disable performance metrics
bool ctx_shift = true; // context shift on inifinite text generation
bool input_prefix_bos = false; // prefix BOS to user inputs, preceding input_prefix
bool logits_all = false; // return logits for all tokens in the batch
@ -220,7 +263,7 @@ struct gpt_params {
std::string cache_type_v = "f16"; // KV cache data type for the V
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector
std::string mmproj = ""; // path to multimodal projector // NOLINT
std::vector<std::string> image; // path to image file(s)
// embedding
@ -236,15 +279,15 @@ struct gpt_params {
int n_threads_http = -1; // number of threads to process HTTP requests (TODO: support threadpool)
std::string hostname = "127.0.0.1";
std::string public_path = "";
std::string chat_template = "";
std::string system_prompt = "";
std::string public_path = ""; // NOLINT
std::string chat_template = ""; // NOLINT
std::string system_prompt = ""; // NOLINT
bool enable_chat_template = true;
std::vector<std::string> api_keys;
std::string ssl_file_key = "";
std::string ssl_file_cert = "";
std::string ssl_file_key = ""; // NOLINT
std::string ssl_file_cert = ""; // NOLINT
bool endpoint_slots = true;
bool endpoint_metrics = false;
@ -299,91 +342,9 @@ struct gpt_params {
bool batched_bench_output_jsonl = false;
};
struct llama_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
const char * env = nullptr;
std::string help;
void (*handler_void) (gpt_params & params) = nullptr;
void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
void (*handler_int) (gpt_params & params, int) = nullptr;
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &)
) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const std::string & help,
void (*handler)(gpt_params & params, int)
) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
llama_arg(
const std::initializer_list<const char *> & args,
const std::string & help,
void (*handler)(gpt_params & params)
) : args(args), help(help), handler_void(handler) {}
// support 2 values for arg
llama_arg(
const std::initializer_list<const char *> & args,
const char * value_hint,
const char * value_hint_2,
const std::string & help,
void (*handler)(gpt_params & params, const std::string &, const std::string &)
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
llama_arg & set_examples(std::initializer_list<enum llama_example> examples) {
this->examples = std::move(examples);
return *this;
}
llama_arg & set_env(const char * env) {
help = help + "\n(env: " + env + ")";
this->env = env;
return *this;
}
bool in_example(enum llama_example ex) {
return examples.find(ex) != examples.end();
}
bool get_value_from_env(std::string & output) const {
if (env == nullptr) return false;
char * value = std::getenv(env);
if (value) {
output = value;
return true;
}
return false;
}
bool has_value_from_env() const {
return env != nullptr && std::getenv(env);
}
std::string to_string();
};
// initialize list of options (arguments) that can be used by the current example
std::vector<llama_arg> gpt_params_parser_init(gpt_params & params, llama_example ex);
// optionally, we can provide "print_usage" to print example usage
std::vector<llama_arg> gpt_params_parser_init(gpt_params & params, llama_example ex, std::function<void(int, char **)> print_usage);
// parse input arguments from CLI
// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
bool gpt_params_parse (int argc, char ** argv, gpt_params & params, std::vector<llama_arg> & options);
bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params, std::vector<llama_arg> & options);
// print full usage message; it will be called internally by gpt_params_parse() if "-h" is set
void gpt_params_print_usage(gpt_params & params, std::vector<llama_arg> & options);
// call once at the start of a program if it uses libcommon
// initializes the logging system and prints info about the build
void gpt_init();
std::string gpt_params_get_system_info(const gpt_params & params);
@ -420,6 +381,11 @@ static std::vector<T> string_split(const std::string & str, char delim) {
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
std::string string_from(bool value);
std::string string_from(const std::vector<int> & values);
std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
//
// Filesystem utils
//

401
common/log.cpp Normal file
View File

@ -0,0 +1,401 @@
#include "log.h"
#include <condition_variable>
#include <cstdarg>
#include <cstdio>
#include <mutex>
#include <sstream>
#include <thread>
#include <vector>
int gpt_log_verbosity_thold = LOG_DEFAULT_LLAMA;
void gpt_log_set_verbosity_thold(int verbosity) {
gpt_log_verbosity_thold = verbosity;
}
#define LOG_COL_DEFAULT "\033[0m"
#define LOG_COL_BOLD "\033[1m"
#define LOG_COL_RED "\033[31m"
#define LOG_COL_GREEN "\033[32m"
#define LOG_COL_YELLOW "\033[33m"
#define LOG_COL_BLUE "\033[34m"
#define LOG_COL_MAGENTA "\033[35m"
#define LOG_COL_CYAN "\033[36m"
#define LOG_COL_WHITE "\033[37m"
static int64_t t_us() {
return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
}
// colors
enum gpt_log_col : int {
GPT_LOG_COL_DEFAULT = 0,
GPT_LOG_COL_BOLD,
GPT_LOG_COL_RED,
GPT_LOG_COL_GREEN,
GPT_LOG_COL_YELLOW,
GPT_LOG_COL_BLUE,
GPT_LOG_COL_MAGENTA,
GPT_LOG_COL_CYAN,
GPT_LOG_COL_WHITE,
};
// disable colors by default
static std::vector<const char *> g_col = {
"",
"",
"",
"",
"",
"",
"",
"",
"",
};
struct gpt_log_entry {
enum ggml_log_level level;
bool prefix;
int64_t timestamp;
std::vector<char> msg;
// signals the worker thread to stop
bool is_end;
void print(FILE * file = nullptr) const {
FILE * fcur = file;
if (!fcur) {
// stderr displays DBG messages only when their verbosity level is not higher than the threshold
// these messages will still be logged to a file
if (level == GGML_LOG_LEVEL_DEBUG && gpt_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
return;
}
fcur = stdout;
if (level != GGML_LOG_LEVEL_NONE) {
fcur = stderr;
}
}
if (level != GGML_LOG_LEVEL_NONE && prefix) {
if (timestamp) {
// [M.s.ms.us]
fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
g_col[GPT_LOG_COL_BLUE],
(int) (timestamp / 1000000 / 60),
(int) (timestamp / 1000000 % 60),
(int) (timestamp / 1000 % 1000),
(int) (timestamp % 1000),
g_col[GPT_LOG_COL_DEFAULT]);
}
switch (level) {
case GGML_LOG_LEVEL_INFO: fprintf(fcur, "%sI %s", g_col[GPT_LOG_COL_GREEN], g_col[GPT_LOG_COL_DEFAULT]); break;
case GGML_LOG_LEVEL_WARN: fprintf(fcur, "%sW %s", g_col[GPT_LOG_COL_MAGENTA], "" ); break;
case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[GPT_LOG_COL_RED], "" ); break;
case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[GPT_LOG_COL_YELLOW], "" ); break;
default:
break;
}
}
fprintf(fcur, "%s", msg.data());
if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
fprintf(fcur, "%s", g_col[GPT_LOG_COL_DEFAULT]);
}
fflush(fcur);
}
};
struct gpt_log {
// default capacity - will be expanded if needed
gpt_log() : gpt_log(256) {}
gpt_log(size_t capacity) {
file = nullptr;
prefix = false;
timestamps = false;
running = false;
t_start = t_us();
// initial message size - will be expanded if longer messages arrive
entries.resize(capacity);
for (auto & entry : entries) {
entry.msg.resize(256);
}
head = 0;
tail = 0;
resume();
}
~gpt_log() {
pause();
if (file) {
fclose(file);
}
}
private:
std::mutex mtx;
std::thread thrd;
std::condition_variable cv;
FILE * file;
bool prefix;
bool timestamps;
bool running;
int64_t t_start;
// ring buffer of entries
std::vector<gpt_log_entry> entries;
size_t head;
size_t tail;
// worker thread copies into this
gpt_log_entry cur;
public:
void add(enum ggml_log_level level, const char * fmt, va_list args) {
std::lock_guard<std::mutex> lock(mtx);
if (!running) {
// discard messages while the worker thread is paused
return;
}
auto & entry = entries[tail];
{
// cannot use args twice, so make a copy in case we need to expand the buffer
va_list args_copy;
va_copy(args_copy, args);
#if 1
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args_copy);
}
#else
// hack for bolding arguments
std::stringstream ss;
for (int i = 0; fmt[i] != 0; i++) {
if (fmt[i] == '%') {
ss << LOG_COL_BOLD;
while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
ss << LOG_COL_DEFAULT;
if (fmt[i] == 0) break;
}
ss << fmt[i];
}
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
}
#endif
}
entry.level = level;
entry.prefix = prefix;
entry.timestamp = 0;
if (timestamps) {
entry.timestamp = t_us() - t_start;
}
entry.is_end = false;
tail = (tail + 1) % entries.size();
if (tail == head) {
// expand the buffer
std::vector<gpt_log_entry> new_entries(2*entries.size());
size_t new_tail = 0;
do {
new_entries[new_tail] = std::move(entries[head]);
head = (head + 1) % entries.size();
new_tail = (new_tail + 1);
} while (head != tail);
head = 0;
tail = new_tail;
for (size_t i = tail; i < new_entries.size(); i++) {
new_entries[i].msg.resize(256);
}
entries = std::move(new_entries);
}
cv.notify_one();
}
void resume() {
std::lock_guard<std::mutex> lock(mtx);
if (running) {
return;
}
running = true;
thrd = std::thread([this]() {
while (true) {
{
std::unique_lock<std::mutex> lock(mtx);
cv.wait(lock, [this]() { return head != tail; });
cur = entries[head];
head = (head + 1) % entries.size();
}
if (cur.is_end) {
break;
}
cur.print(); // stdout and stderr
if (file) {
cur.print(file);
}
}
});
}
void pause() {
{
std::lock_guard<std::mutex> lock(mtx);
if (!running) {
return;
}
running = false;
// push an entry to signal the worker thread to stop
{
auto & entry = entries[tail];
entry.is_end = true;
tail = (tail + 1) % entries.size();
}
cv.notify_one();
}
thrd.join();
}
void set_file(const char * path) {
pause();
if (file) {
fclose(file);
}
if (path) {
file = fopen(path, "w");
} else {
file = nullptr;
}
resume();
}
void set_colors(bool colors) {
pause();
if (colors) {
g_col[GPT_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
g_col[GPT_LOG_COL_BOLD] = LOG_COL_BOLD;
g_col[GPT_LOG_COL_RED] = LOG_COL_RED;
g_col[GPT_LOG_COL_GREEN] = LOG_COL_GREEN;
g_col[GPT_LOG_COL_YELLOW] = LOG_COL_YELLOW;
g_col[GPT_LOG_COL_BLUE] = LOG_COL_BLUE;
g_col[GPT_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
g_col[GPT_LOG_COL_CYAN] = LOG_COL_CYAN;
g_col[GPT_LOG_COL_WHITE] = LOG_COL_WHITE;
} else {
for (size_t i = 0; i < g_col.size(); i++) {
g_col[i] = "";
}
}
resume();
}
void set_prefix(bool prefix) {
std::lock_guard<std::mutex> lock(mtx);
this->prefix = prefix;
}
void set_timestamps(bool timestamps) {
std::lock_guard<std::mutex> lock(mtx);
this->timestamps = timestamps;
}
};
//
// public API
//
struct gpt_log * gpt_log_init() {
return new gpt_log;
}
struct gpt_log * gpt_log_main() {
static struct gpt_log log;
return &log;
}
void gpt_log_pause(struct gpt_log * log) {
log->pause();
}
void gpt_log_resume(struct gpt_log * log) {
log->resume();
}
void gpt_log_free(struct gpt_log * log) {
delete log;
}
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...) {
va_list args;
va_start(args, fmt);
log->add(level, fmt, args);
va_end(args);
}
void gpt_log_set_file(struct gpt_log * log, const char * file) {
log->set_file(file);
}
void gpt_log_set_colors(struct gpt_log * log, bool colors) {
log->set_colors(colors);
}
void gpt_log_set_prefix(struct gpt_log * log, bool prefix) {
log->set_prefix(prefix);
}
void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps) {
log->set_timestamps(timestamps);
}

782
common/log.h
View File

@ -1,724 +1,90 @@
#pragma once
#include <chrono>
#include <cstring>
#include <sstream>
#include <iostream>
#include <thread>
#include <vector>
#include <algorithm>
#include <cinttypes>
#include "ggml.h" // for ggml_log_level
// --------------------------------
//
// Basic usage:
//
// --------
//
// The LOG() and LOG_TEE() macros are ready to go by default
// they do not require any initialization.
//
// LOGLN() and LOG_TEELN() are variants which automatically
// include \n character at the end of the log string.
//
// LOG() behaves exactly like printf, by default writing to a logfile.
// LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
//
// Default logfile is named
// "llama.<threadID>.log"
// Default LOG_TEE() secondary output target is
// stderr
//
// Logs can be dynamically disabled or enabled using functions:
// log_disable()
// and
// log_enable()
//
// A log target can be changed with:
// log_set_target( string )
// creating and opening, or re-opening a file by string filename
// or
// log_set_target( FILE* )
// allowing to point at stderr, stdout, or any valid FILE* file handler.
//
// --------
//
// End of Basic usage.
//
// --------------------------------
// Specifies a log target.
// default uses log_handler() with "llama.log" log file
// this can be changed, by defining LOG_TARGET
// like so:
//
// #define LOG_TARGET (a valid FILE*)
// #include "log.h"
//
// or it can be simply redirected to stdout or stderr
// like so:
//
// #define LOG_TARGET stderr
// #include "log.h"
//
// The log target can also be redirected to a different function
// like so:
//
// #define LOG_TARGET log_handler_different()
// #include "log.h"
//
// FILE* log_handler_different()
// {
// return stderr;
// }
//
// or:
//
// #define LOG_TARGET log_handler_another_one("somelog.log")
// #include "log.h"
//
// FILE* log_handler_another_one(char*filename)
// {
// static FILE* logfile = nullptr;
// (...)
// if( !logfile )
// {
// fopen(...)
// }
// (...)
// return logfile
// }
//
#ifndef LOG_TARGET
#define LOG_TARGET log_handler()
#endif
#ifndef LOG_TEE_TARGET
#define LOG_TEE_TARGET stderr
#endif
// Utility for synchronizing log configuration state
// since std::optional was introduced only in c++17
enum LogTriState
{
LogTriStateSame,
LogTriStateFalse,
LogTriStateTrue
};
// Utility to obtain "pid" like unique process id and use it when creating log files.
inline std::string log_get_pid()
{
static std::string pid;
if (pid.empty())
{
// std::this_thread::get_id() is the most portable way of obtaining a "process id"
// it's not the same as "pid" but is unique enough to solve multiple instances
// trying to write to the same log.
std::stringstream ss;
ss << std::this_thread::get_id();
pid = ss.str();
}
return pid;
}
// Utility function for generating log file names with unique id based on thread id.
// invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
// where the number is a runtime id of the current thread.
#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
// INTERNAL, DO NOT USE
inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
{
static bool _multilog = false;
if (multilog != LogTriStateSame)
{
_multilog = multilog == LogTriStateTrue;
}
std::stringstream buf;
buf << log_file_basename;
if (_multilog)
{
buf << ".";
buf << log_get_pid();
}
buf << ".";
buf << log_file_extension;
return buf.str();
}
#ifndef LOG_DEFAULT_FILE_NAME
#define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
#endif
// Utility for turning #define values into string literals
// so we can have a define for stderr and
// we can print "stderr" instead of literal stderr, etc.
#define LOG_STRINGIZE1(s) #s
#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
// Allows disabling timestamps.
// in order to disable, define LOG_NO_TIMESTAMPS
// like so:
//
// #define LOG_NO_TIMESTAMPS
// #include "log.h"
//
#ifndef LOG_NO_TIMESTAMPS
#ifndef _MSC_VER
#define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#ifndef __GNUC__
# define LOG_ATTRIBUTE_FORMAT(...)
#elif defined(__MINGW32__)
# define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#endif
#else
#define LOG_TIMESTAMP_FMT "%s"
#define LOG_TIMESTAMP_VAL ,""
# define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#ifdef LOG_TEE_TIMESTAMPS
#ifndef _MSC_VER
#define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#else
#define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
#define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
#endif
#else
#define LOG_TEE_TIMESTAMP_FMT "%s"
#define LOG_TEE_TIMESTAMP_VAL ,""
#endif
#define LOG_DEFAULT_DEBUG 1
#define LOG_DEFAULT_LLAMA 0
// Allows disabling file/line/function prefix
// in order to disable, define LOG_NO_FILE_LINE_FUNCTION
// like so:
// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
// set via gpt_log_set_verbosity()
extern int gpt_log_verbosity_thold;
void gpt_log_set_verbosity_thold(int verbosity); // not thread-safe
// the gpt_log uses an internal worker thread to print/write log messages
// when the worker thread is paused, incoming log messages are discarded
struct gpt_log;
struct gpt_log * gpt_log_init();
struct gpt_log * gpt_log_main(); // singleton, automatically destroys itself on exit
void gpt_log_pause (struct gpt_log * log); // pause the worker thread, not thread-safe
void gpt_log_resume(struct gpt_log * log); // resume the worker thread, not thread-safe
void gpt_log_free (struct gpt_log * log);
LOG_ATTRIBUTE_FORMAT(3, 4)
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...);
// defaults: file = NULL, colors = false, prefix = false, timestamps = false
//
// #define LOG_NO_FILE_LINE_FUNCTION
// #include "log.h"
// regular log output:
//
#ifndef LOG_NO_FILE_LINE_FUNCTION
#ifndef _MSC_VER
#define LOG_FLF_FMT "[%24s:%5d][%24s] "
#define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
#else
#define LOG_FLF_FMT "[%24s:%5ld][%24s] "
#define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
#endif
#else
#define LOG_FLF_FMT "%s"
#define LOG_FLF_VAL ,""
#endif
#ifdef LOG_TEE_FILE_LINE_FUNCTION
#ifndef _MSC_VER
#define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
#define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
#else
#define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
#define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
#endif
#else
#define LOG_TEE_FLF_FMT "%s"
#define LOG_TEE_FLF_VAL ,""
#endif
// INTERNAL, DO NOT USE
// USE LOG() INSTEAD
// ggml_backend_metal_log_allocated_size: allocated buffer, size = 6695.84 MiB, ( 6695.91 / 21845.34)
// llm_load_tensors: ggml ctx size = 0.27 MiB
// llm_load_tensors: offloading 32 repeating layers to GPU
// llm_load_tensors: offloading non-repeating layers to GPU
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_IMPL(str, ...) \
// with prefix = true, timestamps = true, the log output will look like this:
//
// 0.00.035.060 D ggml_backend_metal_log_allocated_size: allocated buffer, size = 6695.84 MiB, ( 6695.91 / 21845.34)
// 0.00.035.064 I llm_load_tensors: ggml ctx size = 0.27 MiB
// 0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
// 0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
//
// I - info (stdout, V = 0)
// W - warning (stderr, V = 0)
// E - error (stderr, V = 0)
// D - debug (stderr, V = LOG_DEFAULT_DEBUG)
//
void gpt_log_set_file (struct gpt_log * log, const char * file); // not thread-safe
void gpt_log_set_colors (struct gpt_log * log, bool colors); // not thread-safe
void gpt_log_set_prefix (struct gpt_log * log, bool prefix); // whether to output prefix to each log
void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps); // whether to output timestamps in the prefix
// helper macros for logging
// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
//
// for example:
//
// LOG_DBG("this is a debug message: %d\n", expensive_function());
//
// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > gpt_log_verbosity_thold
//
#define LOG_TMPL(level, verbosity, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TARGET); \
if ((verbosity) <= gpt_log_verbosity_thold) { \
gpt_log_add(gpt_log_main(), (level), __VA_ARGS__); \
} \
} while (0)
#else
#define LOG_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TARGET); \
} \
} while (0)
#endif
// INTERNAL, DO NOT USE
// USE LOG_TEE() INSTEAD
//
#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
#define LOG_TEE_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TARGET); \
} \
if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \
{ \
fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
fflush(LOG_TEE_TARGET); \
} \
} while (0)
#else
#define LOG_TEE_IMPL(str, ...) \
do { \
if (LOG_TARGET != nullptr) \
{ \
fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TARGET); \
} \
if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr) \
{ \
fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
fflush(LOG_TEE_TARGET); \
} \
} while (0)
#endif
#define LOG(...) LOG_TMPL(GGML_LOG_LEVEL_NONE, 0, __VA_ARGS__)
#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
// The '\0' as a last argument, is a trick to bypass the silly
// "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
// so we can have a single macro which can be called just like printf.
#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, 0, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, 0, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
// Main LOG macro.
// behaves like printf, and supports arguments the exact same way.
//
#if !defined(_MSC_VER) || defined(__clang__)
#define LOG(...) LOG_IMPL(__VA_ARGS__, "")
#else
#define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
#endif
// Main TEE macro.
// does the same as LOG
// and
// simultaneously writes stderr.
//
// Secondary target can be changed just like LOG_TARGET
// by defining LOG_TEE_TARGET
//
#if !defined(_MSC_VER) || defined(__clang__)
#define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
#else
#define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
#endif
// LOG macro variants with auto endline.
#if !defined(_MSC_VER) || defined(__clang__)
#define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
#define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
#else
#define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
#define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
#endif
// INTERNAL, DO NOT USE
inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
{
static bool _initialized = false;
static bool _append = false;
static bool _disabled = filename.empty() && target == nullptr;
static std::string log_current_filename{filename};
static FILE *log_current_target{target};
static FILE *logfile = nullptr;
if (change)
{
if (append != LogTriStateSame)
{
_append = append == LogTriStateTrue;
return logfile;
}
if (disable == LogTriStateTrue)
{
// Disable primary target
_disabled = true;
}
// If previously disabled, only enable, and keep previous target
else if (disable == LogTriStateFalse)
{
_disabled = false;
}
// Otherwise, process the arguments
else if (log_current_filename != filename || log_current_target != target)
{
_initialized = false;
}
}
if (_disabled)
{
// Log is disabled
return nullptr;
}
if (_initialized)
{
// with fallback in case something went wrong
return logfile ? logfile : stderr;
}
// do the (re)initialization
if (target != nullptr)
{
if (logfile != nullptr && logfile != stdout && logfile != stderr)
{
fclose(logfile);
}
log_current_filename = LOG_DEFAULT_FILE_NAME;
log_current_target = target;
logfile = target;
}
else
{
if (log_current_filename != filename)
{
if (logfile != nullptr && logfile != stdout && logfile != stderr)
{
fclose(logfile);
}
}
logfile = fopen(filename.c_str(), _append ? "a" : "w");
}
if (!logfile)
{
// Verify whether the file was opened, otherwise fallback to stderr
logfile = stderr;
fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
fflush(stderr);
// At this point we let the init flag be to true below, and let the target fallback to stderr
// otherwise we would repeatedly fopen() which was already unsuccessful
}
_initialized = true;
return logfile ? logfile : stderr;
}
// INTERNAL, DO NOT USE
inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
{
return log_handler1_impl(change, append, disable, filename, target);
}
// Disables logs entirely at runtime.
// Makes LOG() and LOG_TEE() produce no output,
// until enabled back.
#define log_disable() log_disable_impl()
// INTERNAL, DO NOT USE
inline FILE *log_disable_impl()
{
return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
}
// Enables logs at runtime.
#define log_enable() log_enable_impl()
// INTERNAL, DO NOT USE
inline FILE *log_enable_impl()
{
return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
}
// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
#define log_set_target(target) log_set_target_impl(target)
// INTERNAL, DO NOT USE
inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
// INTERNAL, DO NOT USE
inline FILE *log_handler() { return log_handler1_impl(); }
// Enable or disable creating separate log files for each run.
// can ONLY be invoked BEFORE first log use.
#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
// Enable or disable append mode for log file.
// can ONLY be invoked BEFORE first log use.
#define log_append(enable) log_append_impl(enable)
// INTERNAL, DO NOT USE
inline FILE *log_append_impl(bool enable)
{
return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
}
inline void log_test()
{
log_disable();
LOG("01 Hello World to nobody, because logs are disabled!\n");
log_enable();
LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
log_set_target(stderr);
LOG("04 Hello World to stderr!\n");
LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
log_set_target(LOG_DEFAULT_FILE_NAME);
LOG("06 Hello World to default log file!\n");
log_set_target(stdout);
LOG("07 Hello World to stdout!\n");
log_set_target(LOG_DEFAULT_FILE_NAME);
LOG("08 Hello World to default log file again!\n");
log_disable();
LOG("09 Hello World _1_ into the void!\n");
log_enable();
LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
log_disable();
log_set_target("llama.anotherlog.log");
LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
log_enable();
LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
log_set_target("llama.yetanotherlog.log");
LOG("13 Hello World this time in yet new file?\n");
log_set_target(log_filename_generator("llama_autonamed", "log"));
LOG("14 Hello World in log with generated filename!\n");
#ifdef _MSC_VER
LOG_TEE("15 Hello msvc TEE without arguments\n");
LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
LOG_TEELN("17 Hello msvc TEELN without arguments\n");
LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
LOG("19 Hello msvc LOG without arguments\n");
LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
LOGLN("21 Hello msvc LOGLN without arguments\n");
LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
#endif
}
inline bool log_param_single_parse(const std::string & param)
{
if ( param == "--log-test")
{
log_test();
return true;
}
if ( param == "--log-disable")
{
log_disable();
return true;
}
if ( param == "--log-enable")
{
log_enable();
return true;
}
if (param == "--log-new")
{
log_multilog(true);
return true;
}
if (param == "--log-append")
{
log_append(true);
return true;
}
return false;
}
inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
{
if ( param == "--log-file")
{
if (!check_but_dont_parse)
{
log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
}
return true;
}
return false;
}
inline void log_print_usage()
{
printf("log options:\n");
/* format
printf(" -h, --help show this help message and exit\n");*/
/* spacing
printf("__-param----------------Description\n");*/
printf(" --log-test Run simple logging test\n");
printf(" --log-disable Disable trace logs\n");
printf(" --log-enable Enable trace logs\n");
printf(" --log-file Specify a log filename (without extension)\n");
printf(" --log-new Create a separate new log file on start. "
"Each log file will have unique name: \"<name>.<ID>.log\"\n");
printf(" --log-append Don't truncate the old log file.\n");
printf("\n");
}
#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
// INTERNAL, DO NOT USE
inline void log_dump_cmdline_impl(int argc, char **argv)
{
std::stringstream buf;
for (int i = 0; i < argc; ++i)
{
if (std::string(argv[i]).find(' ') != std::string::npos)
{
buf << " \"" << argv[i] <<"\"";
}
else
{
buf << " " << argv[i];
}
}
LOGLN("Cmd:%s", buf.str().c_str());
}
#define log_tostr(var) log_var_to_string_impl(var).c_str()
inline std::string log_var_to_string_impl(bool var)
{
return var ? "true" : "false";
}
inline std::string log_var_to_string_impl(std::string var)
{
return var;
}
inline std::string log_var_to_string_impl(const std::vector<int> & var)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (auto e : var)
{
if (first)
{
first = false;
}
else
{
buf << ", ";
}
buf << std::to_string(e);
}
buf << " ]";
return buf.str();
}
template <typename C, typename T>
inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (const auto & token : tokens)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, token);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "'" << detokenized << "'"
<< ":" << std::to_string(token);
}
buf << " ]";
return buf.str();
}
template <typename C, typename B>
inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
{
std::stringstream buf;
buf << "[ ";
bool first = true;
for (int i = 0; i < batch.n_tokens; ++i)
{
if (!first) {
buf << ", ";
} else {
first = false;
}
auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
detokenized.erase(
std::remove_if(
detokenized.begin(),
detokenized.end(),
[](const unsigned char c) { return !std::isprint(c); }),
detokenized.end());
buf
<< "\n" << std::to_string(i)
<< ":token '" << detokenized << "'"
<< ":pos " << std::to_string(batch.pos[i])
<< ":n_seq_id " << std::to_string(batch.n_seq_id[i])
<< ":seq_id " << std::to_string(batch.seq_id[i][0])
<< ":logits " << std::to_string(batch.logits[i]);
}
buf << " ]";
return buf.str();
}
#ifdef LOG_DISABLE_LOGS
#undef LOG
#define LOG(...) // dummy stub
#undef LOGLN
#define LOGLN(...) // dummy stub
#undef LOG_TEE
#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
#undef LOG_TEELN
#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
#undef LOG_DISABLE
#define LOG_DISABLE() // dummy stub
#undef LOG_ENABLE
#define LOG_ENABLE() // dummy stub
#undef LOG_ENABLE
#define LOG_ENABLE() // dummy stub
#undef LOG_SET_TARGET
#define LOG_SET_TARGET(...) // dummy stub
#undef LOG_DUMP_CMDLINE
#define LOG_DUMP_CMDLINE(...) // dummy stub
#endif // LOG_DISABLE_LOGS
#define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO, verbosity, __VA_ARGS__)
#define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN, verbosity, __VA_ARGS__)
#define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
#define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)

3
common/ngram-cache.cpp
View File

@ -2,8 +2,11 @@
#include "common.h"
#include "log.h"
#include <cinttypes>
#include <cstdint>
#include <cstdio>
#include <fstream>
#include <thread>
void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
std::vector<llama_token> & inp, int nnew, bool print_progress) {

17
common/sampling.cpp
View File

@ -2,6 +2,9 @@
#include "common.h"
#include <cmath>
#include <unordered_map>
// the ring buffer works similarly to std::deque, but with a fixed capacity
// TODO: deduplicate with llama-impl.h
template<typename T>
@ -139,7 +142,7 @@ std::string gpt_sampler_params::print() const {
struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = false; // TODO: control via params
lparams.no_perf = params.no_perf;
auto * result = new gpt_sampler {
/* .params = */ params,
@ -254,10 +257,10 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
// TODO: measure grammar performance
if (gsmpl) {
llama_perf_print(gsmpl->chain, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_sampler_print(gsmpl->chain);
}
if (ctx) {
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx);
}
}
@ -307,6 +310,10 @@ llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context
return cur_p.data[cur_p.selected].id;
}
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
return llama_sampler_get_seed(gsmpl->chain);
}
// helpers
llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {
@ -318,7 +325,7 @@ llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
}
std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
std::string result = "\tlogits ";
std::string result = "logits ";
for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
@ -420,7 +427,7 @@ std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std
}
std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars) {
std::unordered_map<char, gpt_sampler_type> sampler_name_map {
std::unordered_map<char, gpt_sampler_type> sampler_name_map = {
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K), GPT_SAMPLER_TYPE_TOP_K },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z), GPT_SAMPLER_TYPE_TFS_Z },
{ gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P), GPT_SAMPLER_TYPE_TYPICAL_P },

56
common/sampling.h
View File

@ -2,61 +2,11 @@
#include "llama.h"
#include "common.h"
#include <string>
#include <vector>
enum gpt_sampler_type {
GPT_SAMPLER_TYPE_NONE = 0,
GPT_SAMPLER_TYPE_TOP_K = 1,
GPT_SAMPLER_TYPE_TOP_P = 2,
GPT_SAMPLER_TYPE_MIN_P = 3,
GPT_SAMPLER_TYPE_TFS_Z = 4,
GPT_SAMPLER_TYPE_TYPICAL_P = 5,
GPT_SAMPLER_TYPE_TEMPERATURE = 6,
};
// sampling parameters
struct gpt_sampler_params {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
int32_t n_prev = 64; // number of previous tokens to remember
int32_t n_probs = 0; // if greater than 0, output the probabilities of top n_probs tokens.
int32_t min_keep = 0; // 0 = disabled, otherwise samplers should return at least min_keep tokens
int32_t top_k = 40; // <= 0 to use vocab size
float top_p = 0.95f; // 1.0 = disabled
float min_p = 0.05f; // 0.0 = disabled
float tfs_z = 1.00f; // 1.0 = disabled
float typ_p = 1.00f; // typical_p, 1.0 = disabled
float temp = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
float dynatemp_range = 0.00f; // 0.0 = disabled
float dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
int32_t penalty_last_n = 64; // last n tokens to penalize (0 = disable penalty, -1 = context size)
float penalty_repeat = 1.00f; // 1.0 = disabled
float penalty_freq = 0.00f; // 0.0 = disabled
float penalty_present = 0.00f; // 0.0 = disabled
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
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
std::vector<enum gpt_sampler_type> samplers = {
GPT_SAMPLER_TYPE_TOP_K,
GPT_SAMPLER_TYPE_TFS_Z,
GPT_SAMPLER_TYPE_TYPICAL_P,
GPT_SAMPLER_TYPE_TOP_P,
GPT_SAMPLER_TYPE_MIN_P,
GPT_SAMPLER_TYPE_TEMPERATURE
};
std::string grammar; // optional BNF-like grammar to constrain sampling
std::vector<llama_logit_bias> logit_bias; // logit biases to apply
// print the parameters into a string
std::string print() const;
};
// gpt_sampler extends llama_sampler with additional functionality:
//
// - grammar support
@ -110,6 +60,8 @@ void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler *
//
llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl);
// helpers
// access the internal list of current candidate tokens

2
common/train.cpp
View File

@ -1,9 +1,11 @@
#include "train.h"
#include "common.h"
#include <algorithm>
#include <random>
#include <sstream>
#include <functional>
#include <cstring>
struct random_normal_distribution {
std::mt19937 gen;

146
convert_hf_to_gguf.py
View File

@ -131,12 +131,14 @@ class Model:
def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
tensor_names_from_parts: set[str] = set()
if len(self.part_names) > 1:
self.tensor_names = set()
index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
index_name += ".index.json"
index_file = self.dir_model / index_name
if index_file.is_file():
self.tensor_names = set()
logger.info(f"gguf: loading model weight map from '{index_name}'")
with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
with open(index_file, "r", encoding="utf-8") as f:
index: dict[str, Any] = json.load(f)
weight_map = index.get("weight_map")
if weight_map is None or not isinstance(weight_map, dict):
@ -144,6 +146,7 @@ class Model:
self.tensor_names.update(weight_map.keys())
else:
self.tensor_names = tensor_names_from_parts
weight_map = {}
for part_name in self.part_names:
logger.info(f"gguf: loading model part '{part_name}'")
@ -170,9 +173,17 @@ class Model:
data = LazyTorchTensor.from_eager(data)
yield name, data
# only verify tensor name presence; it doesn't matter if they are not in the right files
if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
# verify tensor name presence and identify potentially missing files
if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
if len(extra) == 0 and len(missing_files) > 0:
raise ValueError(f"Missing or incomplete model files: {missing_files}")
else:
raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
f"Missing tensors: {missing}\n"
f"Extra tensors: {extra}")
def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
if key not in gguf.MODEL_TENSORS[self.model_arch]:
@ -305,6 +316,8 @@ class Model:
gguf.MODEL_TENSOR.TIME_MIX_FIRST,
gguf.MODEL_TENSOR.TIME_MIX_W1,
gguf.MODEL_TENSOR.TIME_MIX_W2,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
)
)
or not new_name.endswith(".weight")
@ -627,6 +640,9 @@ class Model:
if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
# ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
res = "exaone"
if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
# ref: https://huggingface.co/microsoft/phi-2
res = "phi-2"
if res is None:
logger.warning("\n")
@ -1485,7 +1501,7 @@ class StableLMModel(Model):
raise ValueError(f"Unprocessed norms: {norms}")
@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
@Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
class LlamaModel(Model):
model_arch = gguf.MODEL_ARCH.LLAMA
@ -1839,6 +1855,60 @@ class MiniCPMModel(Model):
return [(self.map_tensor_name(name), data_torch)]
@Model.register("MiniCPM3ForCausalLM")
class MiniCPM3Model(Model):
model_arch = gguf.MODEL_ARCH.MINICPM3
def set_gguf_parameters(self):
hparams = self.hparams
rope_dims = hparams["qk_rope_head_dim"]
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
self.gguf_writer.add_block_count(self.block_count)
self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
self.gguf_writer.add_head_count(hparams["num_attention_heads"])
self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
self.gguf_writer.add_vocab_size(hparams["vocab_size"])
if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
rope_scaling = self.find_hparam(['rope_scaling'], True)
if rope_scaling is None:
return
long_factors = rope_scaling.get('long_factor', None)
short_factors = rope_scaling.get('short_factor', None)
if long_factors is None or short_factors is None:
raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG] + ".weight", np.array(long_factors, dtype=np.float32))
self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
def set_vocab(self):
self._set_vocab_llama_hf()
def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
if n_kv_head is not None and n_head != n_kv_head:
n_head //= n_kv_head
return (
weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2)
.reshape(weights.shape)
)
@Model.register("QWenLMHeadModel")
class QwenModel(Model):
model_arch = gguf.MODEL_ARCH.QWEN
@ -2778,6 +2848,8 @@ class Rwkv6Model(Model):
self.gguf_writer.add_tokenizer_model("rwkv")
self.gguf_writer.add_token_list(tokens)
self.gguf_writer.add_token_types(toktypes)
special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
special_vocab.add_to_gguf(self.gguf_writer)
def set_gguf_parameters(self):
block_count = self.hparams["num_hidden_layers"]
@ -2946,6 +3018,66 @@ class OlmoModel(Model):
return [(self.map_tensor_name(name), data_torch)]
@Model.register("OlmoeForCausalLM")
class OlmoeModel(Model):
model_arch = gguf.MODEL_ARCH.OLMOE
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_layer_norm_rms_eps(1e-5)
if (n_experts := self.hparams.get("num_experts")) is not None:
self.gguf_writer.add_expert_count(n_experts)
_experts: list[dict[str, Tensor]] | None = None
# Copied from: Qwen2MoeModel
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# process the experts separately
if name.find("experts") != -1:
n_experts = self.hparams["num_experts"]
assert bid is not None
if self._experts is None:
self._experts = [{} for _ in range(self.block_count)]
self._experts[bid][name] = data_torch
if len(self._experts[bid]) >= n_experts * 3:
tensors: list[tuple[str, Tensor]] = []
# merge the experts into a single 3d tensor
for w_name in ["down_proj", "gate_proj", "up_proj"]:
datas: list[Tensor] = []
for xid in range(n_experts):
ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
datas.append(self._experts[bid][ename])
del self._experts[bid][ename]
data_torch = torch.stack(datas, dim=0)
merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
new_name = self.map_tensor_name(merged_name)
tensors.append((new_name, data_torch))
return tensors
else:
return []
return [(self.map_tensor_name(name), data_torch)]
# Copied from: Qwen2MoeModel
def prepare_tensors(self):
super().prepare_tensors()
if self._experts is not None:
# flatten `list[dict[str, Tensor]]` into `list[str]`
experts = [k for d in self._experts for k in d.keys()]
if len(experts) > 0:
raise ValueError(f"Unprocessed experts: {experts}")
@Model.register("JinaBertModel", "JinaBertForMaskedLM")
class JinaBertV2Model(BertModel):
model_arch = gguf.MODEL_ARCH.JINA_BERT_V2

17
convert_hf_to_gguf_update.py
View File

@ -31,6 +31,7 @@ import re
import requests
import sys
import json
import shutil
from hashlib import sha256
from enum import IntEnum, auto
@ -97,6 +98,7 @@ models = [
{'name': "bloom", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
{'name': "gpt3-finnish", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
{"name": "exaone", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
{"name": "phi-2", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
]
@ -125,6 +127,21 @@ def download_model(model):
if tokt == TOKENIZER_TYPE.UGM:
files.append("spiece.model")
if os.path.isdir(repo):
# If repo is a path on the file system, copy the directory
for file in files:
src_path = os.path.join(repo, file)
dst_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(dst_path):
logger.info(f"{name}: File {dst_path} already exists - skipping")
continue
if os.path.isfile(src_path):
shutil.copy2(src_path, dst_path)
logger.info(f"{name}: Copied {src_path} to {dst_path}")
else:
logger.warning(f"{name}: Source file {src_path} does not exist")
else:
# If repo is a URL, download the files
for file in files:
save_path = f"models/tokenizers/{name}/{file}"
if os.path.isfile(save_path):

8
convert_lora_to_gguf.py
View File

@ -367,7 +367,13 @@ if __name__ == '__main__':
yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
dest = super().modify_tensors(data_torch, name, bid)
dest = list(super().modify_tensors(data_torch, name, bid))
# some archs may have the same tensor for lm_head and output (tie word embeddings)
# in this case, adapters targeting lm_head will fail when using llama-export-lora
# therefore, we ignore them for now
# see: https://github.com/ggerganov/llama.cpp/issues/9065
if name == "lm_head.weight" and len(dest) == 0:
raise ValueError("lm_head is present in adapter, but is ignored in base model")
for dest_name, dest_data in dest:
assert isinstance(dest_data, LoraTorchTensor)
lora_a, lora_b = dest_data.get_lora_A_B()

6
docs/build.md
View File

@ -380,3 +380,9 @@ For detailed info, such as model/device supports, CANN install, please refer to
### Android
To read documentation for how to build on Android, [click here](./android.md)
### Arm CPU optimized mulmat kernels
Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).

63
examples/batched-bench/batched-bench.cpp
View File

@ -1,47 +1,28 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
// mutates the input string
static std::vector<int> parse_list(char * p) {
std::vector<int> ret;
char * q = p;
while (*p) {
if (*p == ',') {
*p = '\0';
ret.push_back(std::atoi(q));
q = p + 1;
}
++p;
}
ret.push_back(std::atoi(q));
return ret;
}
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_BENCH, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
return 1;
}
gpt_init();
int is_pp_shared = params.is_pp_shared;
std::vector<int> n_pp = params.n_pp;
@ -98,7 +79,7 @@ int main(int argc, char ** argv) {
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
@ -115,17 +96,17 @@ int main(int argc, char ** argv) {
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
if (!params.batched_bench_output_jsonl) {
LOG_TEE("\n");
LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG_TEE("\n");
LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
LOG("\n");
LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
LOG("\n");
LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
}
for ( int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
@ -155,7 +136,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_clear(ctx);
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -177,7 +158,7 @@ int main(int argc, char ** argv) {
}
if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
}
@ -195,21 +176,21 @@ int main(int argc, char ** argv) {
const float speed = n_kv / t;
if(params.batched_bench_output_jsonl) {
LOG_TEE(
LOG(
"{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
"\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
);
} else {
LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
LOG("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
}
}
}
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
llama_batch_free(batch);
@ -218,7 +199,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

6
examples/batched.swift/Sources/main.swift
View File

@ -140,8 +140,6 @@ while n_cur <= n_len {
let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
llama_sampler_accept(smpl, new_token_id)
// is it an end of stream? -> mark the stream as finished
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
i_batch[i] = -1
@ -202,8 +200,8 @@ let t_main_end = ggml_time_us()
print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
llama_perf_print(UnsafeRawPointer(context), LLAMA_PERF_TYPE_CONTEXT)
llama_perf_print(UnsafeRawPointer(smpl), LLAMA_PERF_TYPE_SAMPLER_CHAIN)
llama_perf_sampler_print(smpl)
llama_perf_context_print(context)
private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
let utf8Count = text.utf8.count

59
examples/batched/batched.cpp
View File

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
@ -7,9 +9,9 @@
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
@ -18,11 +20,11 @@ int main(int argc, char ** argv) {
params.prompt = "Hello my name is";
params.n_predict = 32;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
return 1;
}
gpt_init();
// number of parallel batches
int n_parallel = params.n_parallel;
@ -42,7 +44,7 @@ int main(int argc, char ** argv) {
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: error: unable to load model\n" , __func__);
return 1;
}
@ -72,31 +74,29 @@ int main(int argc, char ** argv) {
llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
return 1;
}
const int n_ctx = llama_n_ctx(ctx);
LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__, n_kv_req);
LOG_TEE("%s: either reduce n_parallel or increase n_ctx\n", __func__);
LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__, n_kv_req);
LOG_ERR("%s: either reduce n_parallel or increase n_ctx\n", __func__);
return 1;
}
// print the prompt token-by-token
fprintf(stderr, "\n");
LOG("\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
// create a llama_batch
// we use this object to submit token data for decoding
llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
@ -114,7 +114,7 @@ int main(int argc, char ** argv) {
if (llama_model_has_encoder(model)) {
if (llama_encode(ctx, batch)) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
@ -131,7 +131,7 @@ int main(int argc, char ** argv) {
batch.logits[batch.n_tokens - 1] = true;
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -142,7 +142,7 @@ int main(int argc, char ** argv) {
//}
if (n_parallel > 1) {
LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
}
// main loop
@ -172,14 +172,12 @@ int main(int argc, char ** argv) {
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
llama_sampler_accept(smpl, new_token_id);
// is it an end of generation? -> mark the stream as finished
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
i_batch[i] = -1;
LOG_TEE("\n");
LOG("\n");
if (n_parallel > 1) {
LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
}
continue;
@ -187,8 +185,7 @@ int main(int argc, char ** argv) {
// if there is only one stream, we print immediately to stdout
if (n_parallel == 1) {
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
}
streams[i] += llama_token_to_piece(ctx, new_token_id);
@ -210,29 +207,27 @@ int main(int argc, char ** argv) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
if (n_parallel > 1) {
LOG_TEE("\n");
LOG("\n");
for (int32_t i = 0; i < n_parallel; ++i) {
LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
}
}
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
LOG_TEE("\n");
llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_sampler_print(smpl);
llama_perf_context_print(ctx);
fprintf(stderr, "\n");

6
examples/benchmark/benchmark-matmult.cpp
View File

@ -183,7 +183,7 @@ int main(int argc, char ** argv) {
ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
TENSOR_DUMP(gf->nodes[0]);
TENSOR_DUMP(ggml_graph_node(gf, 0));
printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
@ -224,7 +224,7 @@ int main(int argc, char ** argv) {
// Let's use the F32 result from above as a reference for the quantized multiplication
float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 0));
printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
printf("=====================================================================================\n");
@ -252,7 +252,7 @@ int main(int argc, char ** argv) {
// Check that the matrix multiplication result is in the right ballpark
// We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
float sum_of_Q4_result = tensor_sum_elements(ggml_graph_node(gf31, 0));
float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; // Let's accept an epsilon of 10^-6

91
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View File

@ -9,6 +9,7 @@
#include <climits>
#include <cstring>
#include <cstdarg>
#include <cinttypes>
#include <ctime>
#include <random>
#include <stdexcept>
@ -105,43 +106,43 @@ static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_
const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
try {
w->token_embedding_table.resize(p->vocab_size * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
w->rms_att_weight.resize(p->n_layers * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
w->rms_ffn_weight.resize(p->n_layers * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
w->wq.resize(p->n_layers * p->dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
w->wo.resize(p->n_layers * p->dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
w->rms_final_weight.resize(p->dim);
LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
LOG_INF("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
if (shared_weights) {
w->wcls = {};
} else {
w->wcls.resize(p->vocab_size * p->dim);
LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
}
}
catch (std::length_error &) {
@ -173,7 +174,7 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
fseek(f, 0, SEEK_END);
auto end = ftell(f);
if (curr != end) {
LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", __func__, curr, end);
LOG_ERR("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end = %ld)\n", __func__, curr, end);
return 1;
}
@ -181,20 +182,20 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
}
static void print_sample_weights(TransformerWeights *w){
LOG("----- Quick print of first of the weight vales of all the variables\n");
LOG("%f\n", w->token_embedding_table[0]);
LOG("%f\n", w->rms_att_weight[0]);
LOG("%f\n", w->rms_ffn_weight[0]);
LOG_INF("----- Quick print of first of the weight vales of all the variables\n");
LOG_INF("%f\n", w->token_embedding_table[0]);
LOG_INF("%f\n", w->rms_att_weight[0]);
LOG_INF("%f\n", w->rms_ffn_weight[0]);
LOG("%f\n", w->wq[0]);
LOG("%f\n", w->wk[0]);
LOG("%f\n", w->wv[0]);
LOG("%f\n", w->wo[0]);
LOG("%f\n", w->w1[0]);
LOG("%f\n", w->w2[0]);
LOG("%f\n", w->w3[0]);
LOG("%f\n", w->rms_att_weight[0]);
if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
LOG_INF("%f\n", w->wq[0]);
LOG_INF("%f\n", w->wk[0]);
LOG_INF("%f\n", w->wv[0]);
LOG_INF("%f\n", w->wo[0]);
LOG_INF("%f\n", w->w1[0]);
LOG_INF("%f\n", w->w2[0]);
LOG_INF("%f\n", w->w3[0]);
LOG_INF("%f\n", w->rms_att_weight[0]);
if (!w->wcls.empty()) LOG_INF("%f\n", w->wcls[0]);
}
////////////////////////////////////////////////////////////////////////////////////////////////////////////
@ -318,20 +319,20 @@ struct train_params {
};
static void print_params(struct my_llama_hparams * params) {
LOG("%s: n_vocab: %u\n", __func__, params->n_vocab);
LOG("%s: n_ctx: %u\n", __func__, params->n_ctx);
LOG("%s: n_embd: %u\n", __func__, params->n_embd);
LOG("%s: n_mult: %u\n", __func__, params->n_mult);
LOG("%s: n_head: %u\n", __func__, params->n_head);
LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
LOG("%s: n_ff: %u\n", __func__, params->n_ff);
LOG("%s: n_layer: %u\n", __func__, params->n_layer);
LOG("%s: n_rot: %u\n", __func__, params->n_rot);
LOG_INF("%s: n_vocab: %u\n", __func__, params->n_vocab);
LOG_INF("%s: n_ctx: %u\n", __func__, params->n_ctx);
LOG_INF("%s: n_embd: %u\n", __func__, params->n_embd);
LOG_INF("%s: n_mult: %u\n", __func__, params->n_mult);
LOG_INF("%s: n_head: %u\n", __func__, params->n_head);
LOG_INF("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
LOG_INF("%s: n_ff: %u\n", __func__, params->n_ff);
LOG_INF("%s: n_layer: %u\n", __func__, params->n_layer);
LOG_INF("%s: n_rot: %u\n", __func__, params->n_rot);
}
static void print_tensor_info(const struct ggml_context * ctx) {
for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
LOG("%s: Allocating ", __func__);
LOG_INF("%s: Allocating ", __func__);
int64_t total = 1;
int i = 0;
for (; i < ggml_n_dims(t); ++i) {
@ -526,7 +527,7 @@ static std::string llama_escape_whitespaces(const std::string & text) {
static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
if (is_ggml_file(filename)) {
LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
struct ggml_context * ctx_data = NULL;
struct gguf_init_params params = {
@ -574,7 +575,7 @@ static void load_vocab(const char * filename, const Config * config, struct llam
gguf_free(ctx);
} else {
// assume llama2.c vocabulary
LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
llama_file file(filename, "rb");
if (!file.fp) {
die_fmt("%s: %s", strerror(errno), filename);
@ -871,23 +872,25 @@ static std::string basename(const std::string &path) {
}
int main(int argc, char ** argv) {
gpt_init();
struct train_params params = get_default_train_params();
if (!params_parse(argc, argv, &params)) {
return 1;
}
log_set_target(stdout);
Config config;
TransformerWeights weights = {};
{
LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
LOG_INF("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
FILE * file = fopen(params.fn_llama2c_model, "rb");
if (!file) {
LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
LOG_ERR("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
return 1;
}
// read in the config header
if (fread(&config, sizeof(Config), 1, file) != 1) {
LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
LOG_ERR("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
return 1;
}
auto shared_weights = config.vocab_size > 0;
@ -896,7 +899,7 @@ int main(int argc, char ** argv) {
// read in the Transformer weights
alloc_weights(&weights, &config, shared_weights);
if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
LOG_ERR("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
return 1;
}
fclose(file);
@ -929,7 +932,7 @@ int main(int argc, char ** argv) {
model.name = basename(params.fn_llama2c_model);
save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
LOG_INF("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
ggml_free(model.ctx);
return 0;

13
examples/cvector-generator/cvector-generator.cpp
View File

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include "llama.h"
#include "ggml.h"
@ -12,14 +13,15 @@
#include "ggml-metal.h"
#endif
#include <algorithm>
#include <climits>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <iostream>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#include <climits>
//////////////////////////////////////////////////
@ -388,8 +390,7 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
return 1;
}

9
examples/cvector-generator/pca.hpp
View File

@ -12,12 +12,9 @@
#include <cstdio>
#include <ctime>
#include <random>
#include <string>
#include <tuple>
#include <vector>
#include <algorithm>
#include <iostream>
#include <fstream>
#define DEBUG_POS 5
@ -229,8 +226,8 @@ static ggml_status compute_piter(
result.eigenvectors.resize(params.n_batch);
result.distances.resize(params.n_batch);
// get output nodes
for (int i = 0; i < gf->n_nodes; ++i) {
auto node = gf->nodes[i];
for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
auto node = ggml_graph_node(gf, i);
int iter = -1;
// find b_tensor (without copying data from device)
if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {

115
examples/embedding/embedding.cpp
View File

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <ctime>
@ -38,16 +40,16 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
// encoder-only model
if (llama_encode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to encode\n", __func__);
LOG_ERR("%s : failed to encode\n", __func__);
}
} else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
// decoder-only model
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
LOG_ERR("%s : failed to decode\n", __func__);
}
}
@ -79,19 +81,16 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EMBEDDING);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
return 1;
}
gpt_init();
params.embedding = true;
// For non-causal models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch;
print_build_info();
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
llama_backend_init();
llama_numa_init(params.numa);
@ -101,7 +100,7 @@ int main(int argc, char ** argv) {
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
@ -111,19 +110,19 @@ int main(int argc, char ** argv) {
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
LOG_ERR("%s: computing embeddings in encoder-decoder models is not supported\n", __func__);
return 1;
}
if (n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
}
// split the prompt into lines
@ -138,7 +137,7 @@ int main(int argc, char ** argv) {
for (const auto & prompt : prompts) {
auto inp = ::llama_tokenize(ctx, prompt, true, false);
if (inp.size() > n_batch) {
fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch);
return 1;
}
@ -149,20 +148,20 @@ int main(int argc, char ** argv) {
// it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
for (auto & inp : inputs) {
if (inp.empty() || inp.back() != llama_token_sep(model)) {
fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
fprintf(stderr, "%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
}
}
// tokenization stats
if (params.verbose_prompt) {
for (int i = 0; i < (int) inputs.size(); i++) {
fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
for (int j = 0; j < (int) inputs[i].size(); j++) {
fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
LOG("%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
}
fprintf(stderr, "\n\n");
LOG("\n\n");
}
}
@ -213,57 +212,57 @@ int main(int argc, char ** argv) {
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
if (params.embd_out.empty()) {
fprintf(stdout, "\n");
LOG("\n");
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
for (int j = 0; j < n_embd_count; j++) {
fprintf(stdout, "embedding %d: ", j);
LOG("embedding %d: ", j);
for (int i = 0; i < std::min(3, n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, " ... ");
LOG(" ... ");
for (int i = n_embd - 3; i < n_embd; i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
LOG("\n");
}
} else {
// print the first part of the embeddings or for a single prompt, the full embedding
for (int j = 0; j < n_prompts; j++) {
fprintf(stdout, "embedding %d: ", j);
LOG("embedding %d: ", j);
for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
LOG("%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
LOG("%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
LOG("\n");
}
// print cosine similarity matrix
if (n_prompts > 1) {
fprintf(stdout, "\n");
printf("cosine similarity matrix:\n\n");
LOG("\n");
LOG("cosine similarity matrix:\n\n");
for (int i = 0; i < n_prompts; i++) {
fprintf(stdout, "%6.6s ", prompts[i].c_str());
LOG("%6.6s ", prompts[i].c_str());
}
fprintf(stdout, "\n");
LOG("\n");
for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f ", sim);
LOG("%6.2f ", sim);
}
fprintf(stdout, "%1.10s", prompts[i].c_str());
fprintf(stdout, "\n");
LOG("%1.10s", prompts[i].c_str());
LOG("\n");
}
}
}
@ -272,43 +271,43 @@ int main(int argc, char ** argv) {
if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
const bool notArray = params.embd_out != "array";
fprintf(stdout, notArray ? "{\n \"object\": \"list\",\n \"data\": [\n" : "[");
LOG(notArray ? "{\n \"object\": \"list\",\n \"data\": [\n" : "[");
for (int j = 0;;) { // at least one iteration (one prompt)
if (notArray) fprintf(stdout, " {\n \"object\": \"embedding\",\n \"index\": %d,\n \"embedding\": ",j);
fprintf(stdout, "[");
if (notArray) LOG(" {\n \"object\": \"embedding\",\n \"index\": %d,\n \"embedding\": ",j);
LOG("[");
for (int i = 0;;) { // at least one iteration (n_embd > 0)
fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
LOG(params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
i++;
if (i < n_embd) fprintf(stdout, ","); else break;
if (i < n_embd) LOG(","); else break;
}
fprintf(stdout, notArray ? "]\n }" : "]");
LOG(notArray ? "]\n }" : "]");
j++;
if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
if (j < n_embd_count) LOG(notArray ? ",\n" : ","); else break;
}
fprintf(stdout, notArray ? "\n ]" : "]\n");
LOG(notArray ? "\n ]" : "]\n");
if (params.embd_out == "json+" && n_prompts > 1) {
fprintf(stdout, ",\n \"cosineSimilarity\": [\n");
LOG(",\n \"cosineSimilarity\": [\n");
for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
fprintf(stdout, " [");
LOG(" [");
for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f", sim);
LOG("%6.2f", sim);
j++;
if (j < n_embd_count) fprintf(stdout, ", "); else break;
if (j < n_embd_count) LOG(", "); else break;
}
fprintf(stdout, " ]");
LOG(" ]");
i++;
if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
if (i < n_embd_count) LOG(",\n"); else break;
}
fprintf(stdout, "\n ]");
LOG("\n ]");
}
if (notArray) fprintf(stdout, "\n}\n");
if (notArray) LOG("\n}\n");
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
// clean up
llama_batch_free(batch);

48
examples/eval-callback/eval-callback.cpp
View File

@ -1,11 +1,11 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include "ggml.h"
#include <cstdio>
#include <random>
#include <string>
#include <tuple>
#include <vector>
/**
@ -31,22 +31,22 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
GGML_ASSERT(n > 0);
float sum = 0;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
printf(" [\n");
LOG(" [\n");
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
if (i2 == n && ne[2] > 2*n) {
printf(" ..., \n");
LOG(" ..., \n");
i2 = ne[2] - n;
}
printf(" [\n");
LOG(" [\n");
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
if (i1 == n && ne[1] > 2*n) {
printf(" ..., \n");
LOG(" ..., \n");
i1 = ne[1] - n;
}
printf(" [");
LOG(" [");
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
if (i0 == n && ne[0] > 2*n) {
printf("..., ");
LOG("..., ");
i0 = ne[0] - n;
}
size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
@ -64,16 +64,16 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
} else {
GGML_ABORT("fatal error");
}
printf("%12.4f", v);
LOG("%12.4f", v);
sum += v;
if (i0 < ne[0] - 1) printf(", ");
if (i0 < ne[0] - 1) LOG(", ");
}
printf("],\n");
LOG("],\n");
}
printf(" ],\n");
LOG(" ],\n");
}
printf(" ]\n");
printf(" sum = %f\n", sum);
LOG(" ]\n");
LOG(" sum = %f\n", sum);
}
}
@ -102,7 +102,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
}
printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
t->name, ggml_type_name(t->type), ggml_op_desc(t),
src0->name, ggml_ne_string(src0).c_str(),
src1 ? src1_str : "",
@ -132,7 +132,7 @@ static bool run(llama_context * ctx, const gpt_params & params) {
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
@ -144,12 +144,11 @@ int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
print_build_info();
gpt_init();
llama_backend_init();
llama_numa_init(params.numa);
@ -166,14 +165,15 @@ int main(int argc, char ** argv) {
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
}
bool OK = run(ctx, params);
@ -181,8 +181,8 @@ int main(int argc, char ** argv) {
return 1;
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
llama_free(ctx);
llama_free_model(model);

8
examples/export-lora/export-lora.cpp
View File

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include "ggml.h"
#include "ggml-alloc.h"
@ -369,7 +370,7 @@ struct lora_merge_ctx {
// write data to output file
{
auto result = gf->nodes[gf->n_nodes - 1];
auto * result = ggml_graph_node(gf, -1);
size_t len = ggml_nbytes(result);
if (read_buf.size() < len) {
read_buf.resize(len);
@ -401,12 +402,11 @@ static void print_usage(int, char ** argv) {
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
return 1;
}
g_verbose = (params.verbosity == 1);
g_verbose = (params.verbosity > 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
ctx.run_merge();

5
examples/gen-docs/gen-docs.cpp
View File

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include <fstream>
@ -9,11 +10,11 @@ static void export_md(std::string fname, llama_example ex) {
std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
gpt_params params;
auto options = gpt_params_parser_init(params, ex);
auto ctx_arg = gpt_params_parser_init(params, ex);
file << "| Argument | Explanation |\n";
file << "| -------- | ----------- |\n";
for (auto & opt : options) {
for (auto & opt : ctx_arg.options) {
file << "| `";
// args
for (const auto & arg : opt.args) {

11
examples/gguf-split/gguf-split.cpp
View File

@ -152,7 +152,7 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
throw std::invalid_argument("error: invalid parameter for argument: " + arg);
}
if (argc - arg_idx < 2) {
if (argc - arg_idx != 2) {
throw std::invalid_argument("error: bad arguments");
}
@ -389,10 +389,17 @@ static void gguf_merge(const split_params & split_params) {
int n_split = 1;
int total_tensors = 0;
auto * ctx_out = gguf_init_empty();
// avoid overwriting existing output file
if (std::ifstream(split_params.output.c_str())) {
fprintf(stderr, "%s: output file %s already exists\n", __func__, split_params.output.c_str());
exit(EXIT_FAILURE);
}
std::ofstream fout(split_params.output.c_str(), std::ios::binary);
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
auto * ctx_out = gguf_init_empty();
std::vector<uint8_t> read_data;
std::vector<ggml_context *> ctx_metas;
std::vector<gguf_context *> ctx_ggufs;

7
examples/gritlm/gritlm.cpp
View File

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include "llama.h"
@ -121,7 +122,6 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
llama_decode(ctx, bat);
llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
llama_sampler_accept(smpl, token);
if (token == eos_token) {
break;
@ -154,11 +154,12 @@ static std::string gritlm_instruction(const std::string & instruction) {
int main(int argc, char * argv[]) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
gpt_init();
llama_model_params mparams = llama_model_params_from_gpt_params(params);
llama_context_params cparams = llama_context_params_from_gpt_params(params);

107
examples/imatrix/imatrix.cpp
View File

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <cmath>
@ -18,12 +20,12 @@
#endif
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s \\\n"
" -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
LOG("\nexample usage:\n");
LOG("\n %s \\\n"
" -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n"
" [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
" [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
LOG_TEE("\n");
LOG("\n");
}
struct Stats {
@ -124,12 +126,10 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
e.counts.resize(src1->ne[0]*n_as, 0);
}
else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
exit(1); //GGML_ABORT("fatal error");
}
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
}
LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
// loop over all possible experts, regardless if they are used or not in the batch
for (int ex = 0; ex < n_as; ++ex) {
size_t e_start = ex*src1->ne[0];
@ -150,7 +150,8 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
e.values[e_start + j] += x[j]*x[j];
e.counts[e_start + j]++;
if (!std::isfinite(e.values[e_start + j])) {
fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
LOG("\n");
LOG_ERR("%f detected in %s\n", e.values[e_start + j], wname.c_str());
exit(1);
}
}
@ -173,20 +174,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
e.counts.resize(src1->ne[0], 0);
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
exit(1); //GGML_ABORT("fatal error");
}
++e.ncall;
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
}
LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
for (int row = 0; row < (int)src1->ne[1]; ++row) {
const float * x = data + row * src1->ne[0];
for (int j = 0; j < (int)src1->ne[0]; ++j) {
e.values[j] += x[j]*x[j];
e.counts[j]++;
if (!std::isfinite(e.values[j])) {
fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
LOG_ERR("%f detected in %s\n", e.values[j], wname.c_str());
exit(1);
}
}
@ -238,17 +237,17 @@ void IMatrixCollector::save_imatrix(int ncall) const {
}
if (n_zeros != 0 && is_first) {
fprintf(stderr, "\n");
LOG_INF("\n");
is_first = false;
}
if (n_zeros == n_all) {
fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
continue;
}
if (n_zeros > 0) {
fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
LOG_WRN("%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
continue;
}
@ -257,7 +256,7 @@ void IMatrixCollector::save_imatrix(int ncall) const {
}
if (to_store.size() < m_stats.size()) {
fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
}
std::ofstream out(fname, std::ios::binary);
@ -289,21 +288,20 @@ void IMatrixCollector::save_imatrix(int ncall) const {
out.write(m_params.prompt_file.c_str(), len);
}
if (m_params.verbosity > 0) {
fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
}
LOGV(1, "\n");
LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
}
bool IMatrixCollector::load_imatrix(const char * fname) {
std::ifstream in(fname, std::ios::binary);
if (!in) {
printf("%s: failed to open %s\n",__func__, fname);
LOG_ERR("%s: failed to open %s\n",__func__, fname);
return false;
}
int n_entries;
in.read((char*)&n_entries, sizeof(n_entries));
if (in.fail() || n_entries < 1) {
printf("%s: no data in file %s\n", __func__, fname);
LOG_ERR("%s: no data in file %s\n", __func__, fname);
return false;
}
for (int i = 0; i < n_entries; ++i) {
@ -311,7 +309,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
std::vector<char> name_as_vec(len+1);
in.read((char *)name_as_vec.data(), len);
if (in.fail()) {
printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
LOG_ERR("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
return false;
}
name_as_vec[len] = 0;
@ -322,7 +320,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
int nval;
in.read((char *)&nval, sizeof(nval));
if (in.fail() || nval < 1) {
printf("%s: failed reading number of values for entry %d\n",__func__,i);
LOG_ERR("%s: failed reading number of values for entry %d\n",__func__,i);
m_stats = {};
return false;
}
@ -335,7 +333,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
std::vector<float> tmp(nval);
in.read((char*)tmp.data(), nval*sizeof(float));
if (in.fail()) {
printf("%s: failed reading data for entry %d\n",__func__,i);
LOG_ERR("%s: failed reading data for entry %d\n",__func__,i);
m_stats = {};
return false;
}
@ -436,26 +434,25 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
const int n_ctx = llama_n_ctx(ctx);
auto tim1 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
auto tim2 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
if (params.i_chunk > 0) {
if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
return false;
}
fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
}
if (int(tokens.size()) < 2*n_ctx) {
fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
n_ctx);
fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);
LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());
return false;
}
@ -477,7 +474,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
double nll = 0.0;
double nll2 = 0.0;
fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
LOG_INF("%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
@ -513,7 +510,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
// TODO: use batch.logits to save computations instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
@ -530,29 +527,29 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
if (i == 0) {
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total * n_chunk);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
LOG("%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
LOG("%.2f minutes\n", total_seconds / 60.0);
}
if (params.compute_ppl) {
const int first = n_ctx/2;
const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
const auto * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
count += n_ctx - first - 1;
printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
LOG("[%d]%.4lf,", i + 1, std::exp(nll / count));
fflush(stdout);
logits.clear();
}
}
printf("\n");
LOG("\n");
if (params.compute_ppl) {
nll2 /= count;
@ -561,9 +558,9 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
nll2 -= nll * nll;
if (nll2 > 0) {
nll2 = sqrt(nll2/(count-1));
printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
} else {
printf("Unexpected negative standard deviation of log(prob)\n");
LOG("Unexpected negative standard deviation of log(prob)\n");
}
}
@ -575,27 +572,27 @@ int main(int argc, char ** argv) {
params.n_ctx = 512;
params.logits_all = true;
params.verbosity = 1;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_IMATRIX, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
return 1;
}
gpt_init();
params.n_batch = std::min(params.n_batch, params.n_ctx);
g_collector.set_params(params);
for (const auto & in_file : params.in_files) {
printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
if (!g_collector.load_imatrix(in_file.c_str())) {
fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());
return 1;
}
}
if (params.in_files.size() > 1) {
printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
g_collector.save_imatrix();
}
@ -614,20 +611,20 @@ int main(int argc, char ** argv) {
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
LOG_ERR("%s : failed to init\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, params.n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
}
if (!compute_imatrix(ctx, params)) {
@ -636,8 +633,8 @@ int main(int argc, char ** argv) {
g_collector.save_imatrix();
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
llama_free(ctx);
llama_free_model(model);

193
examples/infill/infill.cpp
View File

@ -1,6 +1,8 @@
#include "arg.h"
#include "common.h"
#include "console.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cassert>
@ -54,7 +56,7 @@ static void write_logfile(
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
@ -63,7 +65,7 @@ static void write_logfile(
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
@ -92,7 +94,7 @@ static void sigint_handler(int signo) {
is_interacting = true;
} else {
console::cleanup();
printf("\n");
LOG("\n");
gpt_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
_exit(130);
@ -105,63 +107,55 @@ int main(int argc, char ** argv) {
gpt_params params;
g_params = &params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_INFILL);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
return 1;
}
auto & sparams = params.sparams;
gpt_init();
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("infill", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
auto & sparams = params.sparams;
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
printf("\n************\n");
printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
printf("************\n\n");
LOG_ERR("\n************\n");
LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
print_build_info();
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
LOG("%s: llama backend init\n", __func__);
LOG_INF("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
@ -174,34 +168,32 @@ int main(int argc, char ** argv) {
g_smpl = &smpl;
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
LOG_DBG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
}
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
}
const bool add_bos = llama_add_bos_token(model);
GGML_ASSERT(!llama_add_eos_token(model));
LOG("add_bos: %d\n", add_bos);
std::vector<llama_token> embd_inp;
std::vector<llama_token> embd_end;
@ -226,18 +218,19 @@ int main(int argc, char ** argv) {
embd_inp.push_back(middle_token);
}
LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_DBG("add_bos: %d\n", add_bos);
LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
// Should not run without any tokens
if (embd_inp.empty()) {
embd_inp.push_back(llama_token_bos(model));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
}
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
return 1;
}
@ -246,9 +239,8 @@ int main(int argc, char ** argv) {
params.n_keep = (int)embd_inp.size();
}
LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
@ -256,21 +248,21 @@ int main(int argc, char ** argv) {
}
if (params.verbose_prompt) {
LOG_TEE("\n");
LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
LOG_INF("\n");
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (params.n_keep > 0) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
LOG("'\n");
}
LOG_TEE("\n");
LOG_INF("\n");
}
if (params.interactive) {
@ -287,25 +279,30 @@ int main(int argc, char ** argv) {
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
LOG_TEE("%s: interactive mode on.\n", __func__);
LOG_INF("%s: interactive mode on.\n", __func__);
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
LOG_INF("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
}
}
LOG_TEE("sampling: \n%s\n", sparams.print().c_str());
LOG_TEE("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);
LOG_TEE("\n\n");
smpl = gpt_sampler_init(model, sparams);
LOG_TEE("\n##### Infill mode #####\n\n");
LOG_INF("sampler seed: %u\n", gpt_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", gpt_sampler_print(smpl).c_str());
LOG_INF("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);
LOG("\n");
LOG("\n##### Infill mode #####\n\n");
if (params.interactive) {
const char *control_message;
if (params.multiline_input) {
@ -316,11 +313,11 @@ int main(int argc, char ** argv) {
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
LOG_TEE("== Running in interactive mode. ==\n");
LOG("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
LOG( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
LOG( "%s\n", control_message);
is_interacting = params.interactive_first;
}
@ -340,8 +337,6 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
smpl = gpt_sampler_init(model, sparams);
while (n_remain != 0 || params.interactive) {
// predict
if (!embd.empty()) {
@ -355,9 +350,8 @@ int main(int argc, char ** argv) {
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
// infinite text generation via context swapping
@ -366,14 +360,14 @@ int main(int argc, char ** argv) {
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() > n_ctx) {
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep - 1;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1 , params.n_keep + n_discard + 1);
@ -381,9 +375,9 @@ int main(int argc, char ** argv) {
n_past -= n_discard;
LOG("after swap: n_past = %d\n", n_past);
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
}
@ -395,16 +389,16 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
LOG_DBG("n_past = %d\n", n_past);
}
}
@ -416,7 +410,7 @@ int main(int argc, char ** argv) {
gpt_sampler_accept(smpl, id, true);
// LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(id);
@ -426,10 +420,10 @@ int main(int argc, char ** argv) {
// decrement remaining sampling budget
--n_remain;
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
while ((int) embd_inp.size() > n_consumed) {
embd.push_back(embd_inp[n_consumed]);
@ -448,7 +442,7 @@ int main(int argc, char ** argv) {
if (input_echo) {
for (auto id : embd) {
const std::string token_str = llama_token_to_piece(ctx, id);
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
if (embd.size() > 1) {
input_tokens.push_back(id);
@ -457,7 +451,6 @@ int main(int argc, char ** argv) {
output_ss << token_str;
}
}
fflush(stdout);
}
// reset color to default if we there is no pending user input
if (input_echo && (int) embd_inp.size() == n_consumed) {
@ -470,10 +463,9 @@ int main(int argc, char ** argv) {
if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
if (is_interacting && !params.interactive_first) {
// print an eot token
printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
}
fflush(stdout);
printf("\n");
LOG("\n");
console::set_display(console::user_input);
std::string buffer;
std::string line;
@ -529,35 +521,33 @@ int main(int argc, char ** argv) {
n_remain = params.n_predict;
n_past = 0;
n_consumed = 0;
// LOG_TEE("took new input\n");
is_interacting = false;
}
// deal with end of generation tokens in interactive mode
else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
LOG("found EOS token\n");
LOG_DBG("found EOS token\n");
if (params.interactive) {
is_interacting = true;
printf("\n");
LOG("\n");
console::set_display(console::user_input);
fflush(stdout);
}
}
if (n_past > 0 && is_interacting && !params.interactive) {
LOG("waiting for user input\n");
LOG_DBG("waiting for user input\n");
if (params.input_prefix_bos) {
LOG("adding input prefix BOS token\n");
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
}
std::string buffer;
if (!params.input_prefix.empty()) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
buffer += params.input_prefix;
printf("%s", buffer.c_str());
LOG("%s", buffer.c_str());
}
std::string line;
@ -575,17 +565,17 @@ int main(int argc, char ** argv) {
if (buffer.length() > 1) {
// append input suffix if any
if (!params.input_suffix.empty()) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
buffer += params.input_suffix;
printf("%s", params.input_suffix.c_str());
LOG("%s", params.input_suffix.c_str());
}
LOG("buffer: '%s'\n", buffer.c_str());
LOG_DBG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
const auto line_inp = ::llama_tokenize(ctx, buffer, false);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
@ -596,9 +586,9 @@ int main(int argc, char ** argv) {
}
n_remain -= line_inp.size();
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
LOG_DBG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
@ -625,11 +615,10 @@ int main(int argc, char ** argv) {
}
}
if (!params.interactive && n_remain <= 0) {
printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
fflush(stdout);
LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
}
LOG_TEE("\n");
LOG("\n");
gpt_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
@ -639,9 +628,5 @@ int main(int argc, char ** argv) {
gpt_sampler_free(smpl);
llama_backend_free();
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
return 0;
}

2
examples/llama-bench/llama-bench.cpp
View File

@ -1630,7 +1630,7 @@ int main(int argc, char ** argv) {
fflush(p_err->fout);
}
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx);
llama_free(ctx);

2
examples/llama.android/llama/src/main/cpp/llama-android.cpp
View File

@ -414,8 +414,6 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
// sample the most likely token
const auto new_token_id = llama_sampler_sample(sampler, context, -1);
llama_sampler_accept(sampler, new_token_id);
const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
return nullptr;

2
examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
View File

@ -152,8 +152,6 @@ actor LlamaContext {
new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
llama_sampler_accept(sampling, new_token_id)
if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
print("\n")
is_done = true

10
examples/llava/MobileVLM-README.md
View File

@ -39,7 +39,7 @@ python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
```sh
python ./examples/llava/convert_image_encoder_to_gguf \
python ./examples/llava/convert_image_encoder_to_gguf.py \
-m path/to/clip-vit-large-patch14-336 \
--llava-projector path/to/MobileVLM-1.7B/llava.projector \
--output-dir path/to/MobileVLM-1.7B \
@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf \
```
```sh
python ./examples/llava/convert_image_encoder_to_gguf \
python ./examples/llava/convert_image_encoder_to_gguf.py \
-m path/to/clip-vit-large-patch14-336 \
--llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
--output-dir path/to/MobileVLM-1.7B_V2 \
@ -57,12 +57,12 @@ python ./examples/llava/convert_image_encoder_to_gguf \
4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
```sh
python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B
python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B --skip-unknown
```
5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
5. Use `quantize` to convert LLaMA part's DataType from `fp32` to `q4_k`
```sh
./llama-quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
./llama-quantize path/to/MobileVLM-1.7B/ggml-model-F32.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
```
Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.

146
examples/llava/clip.cpp
View File

@ -3,7 +3,6 @@
// I'll gradually clean and extend it
// Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
#include "clip.h"
#include "log.h"
#include "ggml.h"
#include "ggml-alloc.h"
#include "ggml-backend.h"
@ -40,6 +39,11 @@
#include <cinttypes>
#include <limits>
#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
#define LOG_DBG(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
//#define CLIP_DEBUG_FUNCTIONS
// RGB uint8 image
@ -165,7 +169,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
static int get_key_idx(const gguf_context * ctx, const char * key) {
int i = gguf_find_key(ctx, key);
if (i == -1) {
LOG_TEE("key %s not found in file\n", key);
LOG_ERR("key %s not found in file\n", key);
throw std::runtime_error(format("Missing required key: %s", key));
}
@ -270,7 +274,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
size_t tensor_size = ggml_nbytes(tensor);
LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
LOG_INF("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
}
@ -288,7 +292,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
std::ofstream file(filename, std::ios::binary);
if (!file.is_open()) {
LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
return;
}
@ -307,7 +311,7 @@ static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::s
static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
std::ofstream file(filename, std::ios::binary);
if (!file.is_open()) {
LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
return;
}
@ -568,7 +572,7 @@ struct clip_ctx {
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
LOG_ERR("This gguf file seems to have no vision encoder\n");
return nullptr;
}
@ -582,7 +586,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
if (load_image_size == nullptr) {
load_image_size = clip_image_size_init();
}
LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
LOG_DBG("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
image_size_width = load_image_size->width;
image_size_height = load_image_size->height;
if (is_inf) {
@ -1047,21 +1051,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
const int idx_name = gguf_find_key(ctx, KEY_NAME);
if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
const std::string name = gguf_get_val_str(ctx, idx_name);
LOG_TEE("%s: model name: %s\n", __func__, name.c_str());
LOG_INF("%s: model name: %s\n", __func__, name.c_str());
}
LOG_TEE("%s: description: %s\n", __func__, description.c_str());
LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
LOG_TEE("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));
LOG_TEE("%s: n_tensors: %d\n", __func__, n_tensors);
LOG_TEE("%s: n_kv: %d\n", __func__, n_kv);
LOG_TEE("%s: ftype: %s\n", __func__, ftype_str.c_str());
LOG_TEE("\n");
LOG_INF("%s: description: %s\n", __func__, description.c_str());
LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
LOG_INF("%s: alignment: %zu\n", __func__, gguf_get_alignment(ctx));
LOG_INF("%s: n_tensors: %d\n", __func__, n_tensors);
LOG_INF("%s: n_kv: %d\n", __func__, n_kv);
LOG_INF("%s: ftype: %s\n", __func__, ftype_str.c_str());
LOG_INF("\n");
}
const int n_tensors = gguf_get_n_tensors(ctx);
// kv
const int n_kv = gguf_get_n_kv(ctx);
LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
LOG_INF("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
__func__, n_kv, n_tensors, fname);
{
std::map<enum ggml_type, uint32_t> n_type;
@ -1072,7 +1076,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
n_type[type]++;
}
LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
LOG_INF("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
for (int i = 0; i < n_kv; i++) {
const char * name = gguf_get_key(ctx, i);
const enum gguf_type type = gguf_get_kv_type(ctx, i);
@ -1088,7 +1092,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
}
replace_all(value, "\n", "\\n");
LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
LOG_INF("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
}
// print type counts
@ -1097,7 +1101,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
continue;
}
LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
LOG_INF("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
}
}
@ -1112,7 +1116,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
size_t tensor_size = ggml_nbytes(cur);
model_size += tensor_size;
if (verbosity >= 3) {
LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
LOG_INF("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
__func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
}
}
@ -1139,27 +1143,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
#ifdef GGML_USE_CUDA
new_clip->backend = ggml_backend_cuda_init(0);
LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
LOG_INF("%s: CLIP using CUDA backend\n", __func__);
#endif
#ifdef GGML_USE_METAL
new_clip->backend = ggml_backend_metal_init();
LOG_TEE("%s: CLIP using Metal backend\n", __func__);
LOG_INF("%s: CLIP using Metal backend\n", __func__);
#endif
#ifdef GGML_USE_CANN
new_clip->backend = ggml_backend_cann_init(0);
LOG_TEE("%s: CLIP using CANN backend\n", __func__);
LOG_INF("%s: CLIP using CANN backend\n", __func__);
#endif
#ifdef GGML_USE_VULKAN
new_clip->backend = ggml_backend_vk_init(0);
LOG_TEE("%s: CLIP using Vulkan backend\n", __func__);
LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
#endif
if (!new_clip->backend) {
new_clip->backend = ggml_backend_cpu_init();
LOG_TEE("%s: CLIP using CPU backend\n", __func__);
LOG_INF("%s: CLIP using CPU backend\n", __func__);
}
// model size and capabilities
@ -1194,16 +1198,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
if (verbosity >= 1) {
LOG_TEE("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder);
LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
LOG_TEE("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
LOG_TEE("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector);
LOG_TEE("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
LOG_TEE("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
LOG_INF("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder);
LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
LOG_INF("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
LOG_INF("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector);
LOG_INF("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
LOG_INF("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
}
}
LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
LOG_INF("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
// load tensors
{
@ -1216,7 +1220,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
new_clip->ctx_data = ggml_init(params);
if (!new_clip->ctx_data) {
LOG_TEE("%s: ggml_init() failed\n", __func__);
LOG_ERR("%s: ggml_init() failed\n", __func__);
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
@ -1224,7 +1228,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) {
LOG_TEE("cannot open model file for loading tensors\n");
LOG_ERR("cannot open model file for loading tensors\n");
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
@ -1246,7 +1250,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
fin.seekg(offset, std::ios::beg);
if (!fin) {
LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
LOG_ERR("%s: failed to seek for tensor %s\n", __func__, name);
clip_free(new_clip);
gguf_free(ctx);
return nullptr;
@ -1317,23 +1321,23 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
}
if (verbosity >= 2) {
LOG_TEE("\n%s: vision model hparams\n", __func__);
LOG_TEE("image_size %d\n", hparams.image_size);
LOG_TEE("patch_size %d\n", hparams.patch_size);
LOG_TEE("v_hidden_size %d\n", hparams.hidden_size);
LOG_TEE("v_n_intermediate %d\n", hparams.n_intermediate);
LOG_TEE("v_projection_dim %d\n", hparams.projection_dim);
LOG_TEE("v_n_head %d\n", hparams.n_head);
LOG_TEE("v_n_layer %d\n", hparams.n_layer);
LOG_TEE("v_eps %f\n", hparams.eps);
LOG_TEE("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
LOG_TEE("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
LOG_TEE("v_image_grid_pinpoints: ");
LOG_INF("\n%s: vision model hparams\n", __func__);
LOG_INF("image_size %d\n", hparams.image_size);
LOG_INF("patch_size %d\n", hparams.patch_size);
LOG_INF("v_hidden_size %d\n", hparams.hidden_size);
LOG_INF("v_n_intermediate %d\n", hparams.n_intermediate);
LOG_INF("v_projection_dim %d\n", hparams.projection_dim);
LOG_INF("v_n_head %d\n", hparams.n_head);
LOG_INF("v_n_layer %d\n", hparams.n_layer);
LOG_INF("v_eps %f\n", hparams.eps);
LOG_INF("v_image_mean %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
LOG_INF("v_image_std %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
LOG_INF("v_image_grid_pinpoints: ");
for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
LOG_INF("%d ", hparams.image_grid_pinpoints[i]);
}
LOG_TEE("\n");
LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
LOG_INF("\n");
LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
}
@ -1371,7 +1375,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
vision_model.patch_embeddings = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
} catch(const std::exception& /*e*/) {
LOG_TEE("%s: failed to load vision model tensors\n", __func__);
LOG_ERR("%s: failed to load vision model tensors\n", __func__);
}
// LLaVA projection
@ -1400,7 +1404,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
} catch (std::runtime_error & /*e*/) { }
try {
vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
// LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
// LOG_INF("%s: image_newline tensor (llava-1.6) found\n", __func__);
} catch (std::runtime_error & /*e*/) { }
} else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
// MobileVLM projection
@ -1501,7 +1505,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
ggml_gallocr_reserve(new_clip->compute_alloc, gf);
size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
}
return new_clip;
@ -1552,7 +1556,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
int nx, ny, nc;
auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
if (!data) {
LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
@ -1564,7 +1568,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
int nx, ny, nc;
auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
if (!data) {
LOG_TEE("%s: failed to decode image bytes\n", __func__);
LOG_ERR("%s: failed to decode image bytes\n", __func__);
return false;
}
build_clip_img_from_data(data, nx, ny, img);
@ -1754,7 +1758,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int> & or
int downscaled_height = static_cast<int>(original_height * scale);
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
int wasted_resolution = (width * height) - effective_resolution;
// LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
// LOG_INF("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
max_effective_resolution = effective_resolution;
min_wasted_resolution = wasted_resolution;
@ -1872,7 +1876,7 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
const int multiple = fmin(ceil(ratio), max_slice_nums);
std::vector<std::vector<clip_image_u8 *>> images;
LOG_TEE("%s: multiple %d\n", __func__, multiple);
LOG_INF("%s: multiple %d\n", __func__, multiple);
images.push_back(std::vector<clip_image_u8 *>());
if (multiple <= 1) {
@ -1887,17 +1891,17 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
clip_image_u8 * source_image = clip_image_u8_init();
bicubic_resize(*img, *source_image, best_size.first, best_size.second);
// source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
LOG_INF("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
images[images.size()-1].push_back(source_image);
std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
LOG_INF("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
clip_image_u8 * refine_image = clip_image_u8_init();
bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
LOG_INF("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
// split_to_patches
int width = refine_image->nx;
@ -1954,7 +1958,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
int idx = 0;
for (size_t i = 0; i < imgs.size(); ++i) {
for (size_t j = 0; j < imgs[i].size(); ++j) {
LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
LOG_DBG("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
clip_image_f32 * res = clip_image_f32_init();
normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
res_imgs->data[idx++] = *res;
@ -1966,7 +1970,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
bool pad_to_square = true;
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
LOG_ERR("This gguf file seems to have no vision encoder\n");
return false;
}
auto & params = ctx->vision_model.hparams;
@ -2043,7 +2047,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
}
for (size_t i = 0; i < patches.size(); i++) {
// LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
// LOG_DBG("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
clip_image_u8_free(patches[i]);
}
@ -2279,7 +2283,7 @@ static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, co
bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
LOG_ERR("This gguf file seems to have no vision encoder\n");
return false;
}
@ -2291,7 +2295,7 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
LOG_ERR("This gguf file seems to have no vision encoder\n");
return false;
}
@ -2449,7 +2453,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
ggml_backend_graph_compute(ctx->backend, gf);
// the last node is the embedding tensor
struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor * embeddings = ggml_graph_node(gf, -1);
// copy the embeddings to the location passed by the user
ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
@ -2521,7 +2525,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
new_type = type;
if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
// LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
// LOG_ERR("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
}
const size_t n_elms = ggml_nelements(cur);
float * f32_data;
@ -2540,7 +2544,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
f32_data = (float *)conv_buf.data();
break;
default:
LOG_TEE("Please use an input file in f32 or f16\n");
LOG_ERR("Please use an input file in f32 or f16\n");
gguf_free(ctx_out);
return false;
}
@ -2567,7 +2571,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
fout.put(0);
}
LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
LOG_INF("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
}
@ -2583,8 +2587,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
gguf_free(ctx_out);
{
LOG_TEE("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
LOG_INF("%s: original size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
LOG_INF("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
}
return true;

81
examples/llava/llava-cli.cpp
View File

@ -1,14 +1,16 @@
#include "ggml.h"
#include "arg.h"
#include "base64.hpp"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "base64.hpp"
#include "ggml.h"
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
@ -19,7 +21,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
@ -74,7 +76,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
size_t img_base64_str_start, img_base64_str_end;
find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
return NULL;
}
@ -88,7 +90,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
if (!embed) {
LOG_TEE("%s: could not load image from base64 string.\n", __func__);
LOG_ERR("%s: could not load image from base64 string.\n", __func__);
return NULL;
}
@ -113,9 +115,9 @@ struct llava_context {
};
static void print_usage(int, char ** argv) {
LOG_TEE("\n example usage:\n");
LOG_TEE("\n %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG_TEE("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
LOG("\n example usage:\n");
LOG("\n %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) {
@ -125,11 +127,11 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
auto prompt = params->prompt;
if (prompt_contains_image(prompt)) {
if (!params->image.empty()) {
LOG_TEE("using base64 encoded image instead of command line image path\n");
LOG_INF("using base64 encoded image instead of command line image path\n");
}
embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
if (!embed) {
LOG_TEE("%s: can't load image from prompt\n", __func__);
LOG_ERR("%s: can't load image from prompt\n", __func__);
return NULL;
}
params->prompt = remove_image_from_prompt(prompt);
@ -155,18 +157,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
// new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<image>").length());
LOG_TEE("system_prompt: %s\n", system_prompt.c_str());
LOG_INF("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
LOG_TEE("user_prompt: %s\n", user_prompt.c_str());
LOG_INF("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
@ -176,7 +178,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
if (params->verbose_prompt) {
auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
LOG_INF("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
@ -187,11 +189,11 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
// generate the response
LOG_TEE("\n");
LOG("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
if (!smpl) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
@ -201,7 +203,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
printf("%s", tmp);
LOG("%s", tmp);
if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
@ -210,7 +212,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
}
gpt_sampler_free(smpl);
printf("\n");
LOG("\n");
}
static struct llama_model * llava_init(gpt_params * params) {
@ -221,7 +223,7 @@ static struct llama_model * llava_init(gpt_params * params) {
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
@ -244,11 +246,11 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return NULL;
}
auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->ctx_clip = ctx_clip;
@ -267,65 +269,54 @@ static void llava_free(struct llava_context * ctx_llava) {
llama_backend_free();
}
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
int main(int argc, char ** argv) {
ggml_time_init();
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_LLAVA, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("llava", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
gpt_init();
if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
auto model = llava_init(&params);
auto * model = llava_init(&params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
return 1;
}
if (prompt_contains_image(params.prompt)) {
auto ctx_llava = llava_init_context(&params, model);
auto * ctx_llava = llava_init_context(&params, model);
auto image_embed = load_image(ctx_llava, &params, "");
auto * image_embed = load_image(ctx_llava, &params, "");
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
} else {
for (auto & image : params.image) {
auto ctx_llava = llava_init_context(&params, model);
auto * ctx_llava = llava_init_context(&params, model);
auto image_embed = load_image(ctx_llava, &params, image);
auto * image_embed = load_image(ctx_llava, &params, image);
if (!image_embed) {
std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
return 1;
}
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);

60
examples/llava/llava.cpp
View File

@ -1,13 +1,23 @@
#include "clip.h"
#include "common.h"
#include "llama.h"
#include "llava.h"
#include "base64.hpp"
#include "llama.h"
#include <algorithm>
#include <cerrno>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <limits>
#include <vector>
#include <numeric>
#define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0)
#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
#define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
// RGB uint8 image
struct clip_image_u8 {
@ -54,7 +64,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int>& ori
int downscaled_height = static_cast<int>(original_height * scale);
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
int wasted_resolution = (width * height) - effective_resolution;
// LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
// LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
max_effective_resolution = effective_resolution;
min_wasted_resolution = wasted_resolution;
@ -184,7 +194,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
// ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
ggml_build_forward_expand(gf, flatten);
ggml_graph_compute_with_ctx(model.ctx, gf, 1);
struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
struct ggml_tensor* result = ggml_graph_node(gf, -1);
memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
// append without newline tokens (default behavior in llava_arch when not using unpad ):
@ -236,7 +246,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
LOG_TEE("%s: unable to preprocess image\n", __func__);
LOG_ERR("%s: unable to preprocess image\n", __func__);
delete[] img_res_v.data;
return false;
}
@ -265,14 +275,14 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
}
if (!encoded) {
LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
const int64_t t_img_enc_steop_batch_us = ggml_time_us();
LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
int n_img_pos_out = 0;
for (size_t i = 0; i < image_embd_v.size(); i++) {
@ -287,7 +297,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
load_image_size->width = img->nx;
load_image_size->height = img->ny;
clip_add_load_image_size(ctx_clip, load_image_size);
LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
}
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
@ -295,7 +305,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
delete[] img_res_v.data;
if (!encoded) {
LOG_TEE("Unable to encode image\n");
LOG_ERR("Unable to encode image\n");
return false;
}
@ -309,12 +319,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
if (!encoded) {
LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_TEE("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
const int32_t * image_grid = clip_image_grid(ctx_clip);
@ -347,12 +357,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
// clip_image_save_to_bmp(*tmp, "image_feature.bmp");
}
LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
const int64_t t_img_enc_end_us = ggml_time_us();
float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
return true;
}
@ -362,7 +372,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
if (n_image_embd != n_llama_embd) {
LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
return false;
}
return true;
@ -375,13 +385,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
}
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
if (!image_embd) {
LOG_TEE("Unable to allocate memory for image embeddings\n");
LOG_ERR("Unable to allocate memory for image embeddings\n");
return false;
}
int n_img_pos;
if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
LOG_TEE("%s: cannot encode image, aborting\n", __func__);
LOG_ERR("%s: cannot encode image, aborting\n", __func__);
free(image_embd);
return false;
}
@ -401,7 +411,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
}
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
if (llama_decode(ctx_llama, batch)) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return false;
}
*n_past += n_eval;
@ -413,7 +423,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
clip_image_u8 * img = clip_image_u8_init();
if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
clip_image_u8_free(img);
LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
return NULL;
}
@ -422,7 +432,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
if (!image_embed_result) {
clip_image_u8_free(img);
LOG_TEE("%s: coulnd't embed the image\n", __func__);
LOG_ERR("%s: coulnd't embed the image\n", __func__);
return NULL;
}
@ -436,7 +446,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
auto file = fopen(path, "rb");
if (file == NULL) {
LOG_TEE("%s: can't read file %s\n", __func__, path);
LOG_ERR("%s: can't read file %s\n", __func__, path);
return false;
}
@ -446,7 +456,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
if (buffer == NULL) {
LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
perror("Memory allocation error");
fclose(file);
return false;
@ -471,7 +481,7 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
long image_bytes_length;
auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
if (!loaded) {
LOG_TEE("%s: failed to load %s\n", __func__, image_path);
LOG_ERR("%s: failed to load %s\n", __func__, image_path);
return NULL;
}

95
examples/llava/minicpmv-cli.cpp
View File

@ -1,13 +1,18 @@
#include "ggml.h"
#include "arg.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "ggml.h"
#include <algorithm>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
#include <iostream> // TODO: remove me
struct llava_context {
struct clip_ctx * ctx_clip = NULL;
@ -16,14 +21,8 @@ struct llava_context {
};
static void show_additional_info(int /*argc*/, char ** argv) {
LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
LOG("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llama_model * llava_init(gpt_params * params) {
@ -34,7 +33,7 @@ static struct llama_model * llava_init(gpt_params * params) {
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
@ -49,7 +48,7 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
if (params->n_ctx < 2048) {
// warn user here, "Image processing requires at least 2048 context, setting context to 2048"
LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
LOG_WRN("%s: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
ctx_params.n_ctx = 2048;
} else {
ctx_params.n_ctx = params->n_ctx;
@ -58,11 +57,11 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return NULL;
}
auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->model = model;
@ -87,7 +86,7 @@ static struct clip_ctx * clip_init_context(gpt_params * params) {
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
auto * ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
return ctx_clip;
}
@ -99,7 +98,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
@ -123,7 +122,7 @@ static void process_eval_image_embed(struct llava_context * ctx_llava, const str
float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
auto * slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
slice_embed->embed = image_embed;
slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
@ -141,7 +140,7 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
else if (has_minicpmv_projector == 3) {
system_prompt = "<|im_start|>user\n";
}
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
LOG_INF("%s: image token past: %d\n", __func__, n_past);
eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
@ -160,7 +159,7 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
}
eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
}
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
LOG_INF("%s: image token past: %d\n", __func__, n_past);
}
static const char * sample(struct gpt_sampler * smpl,
@ -179,42 +178,42 @@ static const char * sample(struct gpt_sampler * smpl,
}
static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
auto ctx_clip = clip_init_context(params);
auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
auto * ctx_clip = clip_init_context(params);
auto * embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
if (!embeds) {
std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
return NULL;
}
// process the prompt
if (params->prompt.empty() && params->interactive == false) {
LOG_TEE("prompt should be given or interactive mode should be on");
LOG_ERR("prompt should be given or interactive mode should be on");
return NULL;
}
auto model = llava_init(params);
auto * model = llava_init(params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
return NULL;
}
const int64_t t_llava_init_start_us = ggml_time_us();
auto ctx_llava = llava_init_context(params, model);
auto * ctx_llava = llava_init_context(params, model);
ctx_llava->ctx_clip = ctx_clip;
const int64_t t_llava_init_end_us = ggml_time_us();
float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
LOG_INF("%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
const int64_t t_process_image_start_us = ggml_time_us();
process_image(ctx_llava, embeds, params, n_past);
const int64_t t_process_image_end_us = ggml_time_us();
float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
LOG_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
llava_image_embed_free(embeds);
return ctx_llava;
}
static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, const std::string & prompt, int & n_past, bool is_first = false){
std::string user_prompt = prompt;
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (!is_first) {
@ -236,7 +235,7 @@ static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_par
// generate the response
LOG_TEE("\n");
LOG_INF("\n");
struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
return smpl;
@ -253,17 +252,11 @@ int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, show_additional_info);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("llava", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
gpt_init();
if (params.mmproj.empty() || (params.image.empty())) {
show_additional_info(argc, argv);
@ -272,21 +265,23 @@ int main(int argc, char ** argv) {
for (auto & image : params.image) {
int n_past = 0;
auto ctx_llava = minicpmv_init(&params, image, n_past);
auto * ctx_llava = minicpmv_init(&params, image, n_past);
if (!params.prompt.empty()) {
LOG_TEE("<user>%s\n", params.prompt.c_str());
LOG_TEE("<assistant>");
auto smpl = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
LOG("<user>%s\n", params.prompt.c_str());
LOG("<assistant>");
auto * smpl = llama_init(ctx_llava, &params, params.prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
std::string response;
bool have_tmp = false;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, smpl, n_past);
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0){
if(!have_tmp)continue;
else break;
if (!have_tmp) {
continue;
}
break;
}
if (strstr(tmp, "###")) break; // Yi-VL behavior
have_tmp = true;
@ -298,15 +293,15 @@ int main(int argc, char ** argv) {
gpt_sampler_free(smpl);
}else {
while (true) {
LOG_TEE("<user>");
LOG("<user>");
std::string prompt;
std::getline(std::cin, prompt);
LOG_TEE("<assistant>");
auto smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
LOG("<assistant>");
auto * smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
std::string response;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, smpl, n_past);
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
@ -318,7 +313,7 @@ int main(int argc, char ** argv) {
}
}
printf("\n");
llama_perf_print(ctx_llava->ctx_llama, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_llava->ctx_llama);
ctx_llava->model = NULL;
llava_free(ctx_llava);

60
examples/lookahead/lookahead.cpp
View File

@ -1,4 +1,7 @@
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cstdio>
@ -36,23 +39,18 @@ struct ngram_container {
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
gpt_init();
const int W = 15; // lookahead window
const int N = 5; // n-gram size
const int G = 15; // max verification n-grams
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookahead", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
@ -74,14 +72,14 @@ int main(int argc, char ** argv) {
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
@ -165,7 +163,7 @@ int main(int argc, char ** argv) {
{
const std::string token_str = llama_token_to_piece(ctx, id);
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
fflush(stdout);
}
}
@ -255,7 +253,7 @@ int main(int argc, char ** argv) {
}
if (llama_decode(ctx, batch) != 0) {
fprintf(stderr, "\n\n%s: error: llama_decode failed - increase KV cache size\n", __func__);
LOG_ERR("\n\n%s: llama_decode failed - increase KV cache size\n", __func__);
return 1;
}
@ -292,10 +290,10 @@ int main(int argc, char ** argv) {
const std::string token_str = llama_token_to_piece(ctx, id);
if (v == 0) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
} else {
// print light cyan
printf("\033[0;96m%s\033[0m", token_str.c_str());
LOG("\033[0;96m%s\033[0m", token_str.c_str());
}
fflush(stdout);
@ -329,21 +327,21 @@ int main(int argc, char ** argv) {
// print known n-grams starting with token id (debug)
if (0 && v == 0) {
if (ngrams_observed.cnt[id] > 0) {
printf("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str());
LOG("\n - %d n-grams starting with '%s'\n", ngrams_observed.cnt[id], llama_token_to_piece(ctx, id).c_str());
}
for (int i = 0; i < ngrams_observed.cnt[id]; i++) {
printf(" - ngram %2d: ", i);
LOG(" - ngram %2d: ", i);
const int idx = id*(N - 1)*G + i*(N - 1);
for (int j = 0; j < N - 1; j++) {
const std::string token_str = llama_token_to_piece(ctx, ngrams_observed.tokens[idx + j]);
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
printf("\n");
LOG("\n");
}
}
@ -454,20 +452,20 @@ int main(int argc, char ** argv) {
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("W = %2d\n", W);
LOG_TEE("N = %2d\n", N);
LOG_TEE("G = %2d\n", G);
LOG_TEE("\n");
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_accept = %d\n", n_accept);
LOG_INF("\n");
LOG_INF("W = %2d\n", W);
LOG_INF("N = %2d\n", N);
LOG_INF("G = %2d\n", G);
LOG_INF("\n");
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_accept = %d\n", n_accept);
LOG_TEE("\n");
LOG_INF("\n");
gpt_perf_print(ctx, smpl);
gpt_sampler_free(smpl);
@ -481,7 +479,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

10
examples/lookup/lookup-create.cpp
View File

@ -1,7 +1,8 @@
#include "ggml.h"
#include "llama.h"
#include "arg.h"
#include "common.h"
#include "ngram-cache.h"
#include "ggml.h"
#include "llama.h"
#include <cstdint>
#include <fstream>
@ -13,8 +14,7 @@
int main(int argc, char ** argv){
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
@ -40,4 +40,6 @@ int main(int argc, char ** argv){
fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
return 0;
}

37
examples/lookup/lookup-stats.cpp
View File

@ -1,25 +1,26 @@
#include "ggml.h"
#include "arg.h"
#include "common.h"
#include "llama.h"
#include "log.h"
#include "ngram-cache.h"
#include "llama.h"
#include "ggml.h"
#include <cmath>
#include <cstdint>
#include <cstdio>
#include <cinttypes>
#include <fstream>
#include <string>
#include <vector>
#include <unordered_map>
int main(int argc, char ** argv){
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
gpt_init();
const int n_draft = params.n_draft;
// init llama.cpp
@ -49,7 +50,7 @@ int main(int argc, char ** argv){
try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) {
fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1);
}
}
@ -128,7 +129,7 @@ int main(int argc, char ** argv){
const int64_t eta_min = eta_ms / (60*1000);
const int64_t eta_s = (eta_ms - 60*1000*eta_min) / 1000;
LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
LOG_INF("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
}
// After each chunk, update the dynamic ngram cache with the context ngram cache:
@ -136,24 +137,24 @@ int main(int argc, char ** argv){
ngram_cache_context.clear();
}
LOG_TEE("\n");
LOG("\n");
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_input - n_input % n_ctx);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_input - n_input % n_ctx);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_INF("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
llama_free(ctx);
llama_free_model(model);
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

65
examples/lookup/lookup.cpp
View File

@ -1,7 +1,10 @@
#include "arg.h"
#include "ggml.h"
#include "llama.h"
#include "common.h"
#include "ngram-cache.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cstdint>
#include <cstdio>
@ -12,22 +15,17 @@
int main(int argc, char ** argv){
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
return 1;
}
gpt_init();
// max. number of additional tokens to draft if match is found
const int n_draft = params.n_draft;
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("lookup", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
@ -57,7 +55,7 @@ int main(int argc, char ** argv){
try {
ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
} catch (std::ifstream::failure const &) {
fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
exit(1);
}
}
@ -75,14 +73,14 @@ int main(int argc, char ** argv){
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
@ -123,7 +121,7 @@ int main(int argc, char ** argv){
}
// print current draft sequence
LOG("drafted %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, draft).c_str());
LOG_DBG("drafted %s\n", string_from(ctx, draft).c_str());
int i_dft = 0;
while (true) {
@ -135,7 +133,7 @@ int main(int argc, char ** argv){
const std::string token_str = llama_token_to_piece(ctx, id);
if (!params.use_color) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
if (llama_token_is_eog(model, id)) {
@ -146,7 +144,7 @@ int main(int argc, char ** argv){
// check if the target token matches the draft
if (i_dft < (int) draft.size() && id == draft[i_dft]) {
LOG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
LOG_DBG("the sampled target token matches the %dth drafted token (%d, '%s') - accepted\n", i_dft, id, token_str.c_str());
++n_accept;
++n_past;
++i_dft;
@ -160,19 +158,19 @@ int main(int argc, char ** argv){
if (params.use_color) {
// color accepted draft token
printf("\033[34m%s\033[0m", token_str.c_str());
LOG("\033[34m%s\033[0m", token_str.c_str());
fflush(stdout);
}
continue;
}
if (params.use_color) {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
fflush(stdout);
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
LOG_DBG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
draft.clear();
draft.push_back(id);
@ -223,24 +221,23 @@ int main(int argc, char ** argv){
llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_TEE("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
LOG_INF("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_TEE("\ntarget:\n\n");
llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG_INF("\ntarget:\n\n");
gpt_perf_print(ctx, smpl);
gpt_sampler_free(smpl);
@ -251,7 +248,7 @@ int main(int argc, char ** argv){
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

2
examples/main/README.md
View File

@ -161,6 +161,8 @@ A value of -1 will enable infinite text generation, even though we have a finite
If the pause is undesirable, a value of -2 will stop generation immediately when the context is filled.
The `--no-context-shift` option allows you to stop the infinite text generation once the finite context window is full.
It is important to note that the generated text may be shorter than the specified number of tokens if an End-of-Sequence (EOS) token or a reverse prompt is encountered. In interactive mode, text generation will pause and control will be returned to the user. In non-interactive mode, the program will end. In both cases, the text generation may stop before reaching the specified `--predict` value. If you want the model to keep going without ever producing End-of-Sequence on its own, you can use the `--ignore-eos` parameter.
### Temperature

293
examples/main/main.cpp
View File

@ -1,11 +1,11 @@
#include "arg.h"
#include "common.h"
#include "console.h"
#include "log.h"
#include "sampling.h"
#include "llama.h"
#include <cassert>
#include <cinttypes>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
@ -41,11 +41,13 @@ static std::vector<llama_token> * g_output_tokens;
static bool is_interacting = false;
static bool need_insert_eot = false;
static void print_usage(int, char ** argv) {
printf("\nexample usage:\n");
printf("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
printf("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
printf("\n");
static void print_usage(int argc, char ** argv) {
(void) argc;
LOG("\nexample usage:\n");
LOG("\n text generation: %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
LOG("\n chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
LOG("\n");
}
static bool file_exists(const std::string & path) {
@ -73,8 +75,7 @@ static void write_logfile(
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
LOG_ERR("%s: failed to create logdir %s, cannot write logfile\n", __func__, params.logdir.c_str());
return;
}
@ -82,7 +83,7 @@ static void write_logfile(
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
@ -112,7 +113,7 @@ static void sigint_handler(int signo) {
need_insert_eot = true;
} else {
console::cleanup();
printf("\n");
LOG("\n");
gpt_perf_print(*g_ctx, *g_smpl);
write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
_exit(130);
@ -121,80 +122,61 @@ static void sigint_handler(int signo) {
}
#endif
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, const std::string & role, const std::string & content) {
llama_chat_msg new_msg{role, content};
auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
chat_msgs.push_back({role, content});
LOG("formatted: %s\n", formatted.c_str());
LOG_DBG("formatted: '%s'\n", formatted.c_str());
return formatted;
}
int main(int argc, char ** argv) {
gpt_params params;
g_params = &params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_MAIN, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
return 1;
}
gpt_init();
auto & sparams = params.sparams;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("main", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
// TODO: Dump params ?
//LOG("Params perplexity: %s\n", LOG_TOSTR(params.perplexity));
// save choice to use color for later
// (note for later: this is a slightly awkward choice)
console::init(params.simple_io, params.use_color);
atexit([]() { console::cleanup(); });
if (params.logits_all) {
printf("\n************\n");
printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
printf("************\n\n");
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.embedding) {
printf("\n************\n");
printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
printf("************\n\n");
LOG_ERR("************\n");
LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
LOG_ERR("************\n\n");
return 0;
}
if (params.n_ctx != 0 && params.n_ctx < 8) {
LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
LOG_WRN("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
params.n_ctx = 8;
}
if (params.rope_freq_base != 0.0) {
LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
LOG_WRN("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
}
if (params.rope_freq_scale != 0.0) {
LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
LOG_WRN("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
}
print_build_info();
LOG_INF("%s: llama backend init\n", __func__);
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
LOG("%s: llama backend init\n", __func__);
llama_backend_init();
llama_numa_init(params.numa);
@ -209,21 +191,19 @@ int main(int argc, char ** argv) {
g_smpl = &smpl;
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);
LOG_ERR("%s: error: unable to load model\n", __func__);
return 1;
}
LOG("%s: llama threadpool init = n_threads = %d\n",
__func__,
(int) params.cpuparams.n_threads
);
LOG_INF("%s: llama threadpool init, n_threads = %d\n", __func__, (int) params.cpuparams.n_threads);
struct ggml_threadpool_params tpp_batch =
ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
struct ggml_threadpool_params tpp =
@ -235,8 +215,8 @@ int main(int argc, char ** argv) {
if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
threadpool_batch = ggml_threadpool_new(&tpp_batch);
if (!threadpool_batch) {
LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
exit(1);
LOG_ERR("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
return 1;
}
// Start the non-batch threadpool in the paused state
@ -245,55 +225,54 @@ int main(int argc, char ** argv) {
struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
if (!threadpool) {
LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
exit(1);
LOG_ERR("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
return 1;
}
llama_attach_threadpool(ctx, threadpool, threadpool_batch);
const int n_ctx_train = llama_n_ctx_train(model);
const int n_ctx = llama_n_ctx(ctx);
LOG("n_ctx: %d\n", n_ctx);
if (n_ctx > n_ctx_train) {
LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
}
// print chat template example in conversation mode
if (params.conversation) {
if (params.enable_chat_template) {
LOG_TEE("%s: chat template example: %s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
LOG_INF("%s: chat template example:\n%s\n", __func__, llama_chat_format_example(model, params.chat_template).c_str());
} else {
LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
LOG_INF("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
}
}
// print system information
{
LOG_TEE("\n");
LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
}
std::string path_session = params.path_prompt_cache;
std::vector<llama_token> session_tokens;
if (!path_session.empty()) {
LOG_TEE("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
LOG_INF("%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
if (!file_exists(path_session)) {
LOG_TEE("%s: session file does not exist, will create.\n", __func__);
LOG_INF("%s: session file does not exist, will create.\n", __func__);
} else if (file_is_empty(path_session)) {
LOG_TEE("%s: The session file is empty. A new session will be initialized.\n", __func__);
LOG_INF("%s: The session file is empty. A new session will be initialized.\n", __func__);
} else {
// The file exists and is not empty
session_tokens.resize(n_ctx);
size_t n_token_count_out = 0;
if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
LOG_ERR("%s: failed to load session file '%s'\n", __func__, path_session.c_str());
return 1;
}
session_tokens.resize(n_token_count_out);
LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
LOG_INF("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
}
}
@ -301,7 +280,8 @@ int main(int argc, char ** argv) {
if (!llama_model_has_encoder(model)) {
GGML_ASSERT(!llama_add_eos_token(model));
}
LOG("add_bos: %d\n", add_bos);
LOG_DBG("n_ctx: %d, add_bos: %d\n", n_ctx, add_bos);
std::vector<llama_token> embd_inp;
@ -310,31 +290,31 @@ int main(int argc, char ** argv) {
? chat_add_and_format(model, chat_msgs, "system", params.prompt) // format the system prompt in conversation mode
: params.prompt;
if (params.interactive_first || !params.prompt.empty() || session_tokens.empty()) {
LOG("tokenize the prompt\n");
LOG_DBG("tokenize the prompt\n");
embd_inp = ::llama_tokenize(ctx, prompt, true, true);
} else {
LOG("use session tokens\n");
LOG_DBG("use session tokens\n");
embd_inp = session_tokens;
}
LOG("prompt: \"%s\"\n", log_tostr(prompt));
LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_DBG("prompt: \"%s\"\n", prompt.c_str());
LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
}
// Should not run without any tokens
if (embd_inp.empty()) {
if (add_bos) {
embd_inp.push_back(llama_token_bos(model));
LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
} else {
LOG_TEE("error: input is empty\n");
LOG_ERR("input is empty\n");
return -1;
}
}
// Tokenize negative prompt
if ((int) embd_inp.size() > n_ctx - 4) {
LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
return 1;
}
@ -348,14 +328,14 @@ int main(int argc, char ** argv) {
n_matching_session_tokens++;
}
if (params.prompt.empty() && n_matching_session_tokens == embd_inp.size()) {
LOG_TEE("%s: using full prompt from session file\n", __func__);
LOG_INF("%s: using full prompt from session file\n", __func__);
} else if (n_matching_session_tokens >= embd_inp.size()) {
LOG_TEE("%s: session file has exact match for prompt!\n", __func__);
LOG_INF("%s: session file has exact match for prompt!\n", __func__);
} else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
LOG_TEE("%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
LOG_WRN("%s: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
__func__, n_matching_session_tokens, embd_inp.size());
} else {
LOG_TEE("%s: session file matches %zu / %zu tokens of prompt\n",
LOG_INF("%s: session file matches %zu / %zu tokens of prompt\n",
__func__, n_matching_session_tokens, embd_inp.size());
}
@ -363,14 +343,13 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_rm(ctx, -1, n_matching_session_tokens, -1);
}
LOGLN(
"recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu",
log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
LOG_DBG("recalculate the cached logits (check): embd_inp.size() %zu, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu\n",
embd_inp.size(), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
// if we will use the cache for the full prompt without reaching the end of the cache, force
// reevaluation of the last token to recalculate the cached logits
if (!embd_inp.empty() && n_matching_session_tokens == embd_inp.size() && session_tokens.size() > embd_inp.size()) {
LOGLN("recalculate the cached logits (do): session_tokens.resize( %zu )", embd_inp.size() - 1);
LOG_DBG("recalculate the cached logits (do): session_tokens.resize( %zu )\n", embd_inp.size() - 1);
session_tokens.resize(embd_inp.size() - 1);
}
@ -392,21 +371,20 @@ int main(int argc, char ** argv) {
}
if (params.verbose_prompt) {
LOG_TEE("\n");
LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
for (int i = 0; i < (int) embd_inp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
if (params.n_keep > add_bos) {
LOG_TEE("%s: static prompt based on n_keep: '", __func__);
LOG_INF("%s: static prompt based on n_keep: '", __func__);
for (int i = 0; i < params.n_keep; i++) {
LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
LOG("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
}
LOG_TEE("'\n");
LOG("'\n");
}
LOG_TEE("\n");
LOG_INF("\n");
}
// ctrl+C handling
@ -426,40 +404,40 @@ int main(int argc, char ** argv) {
}
if (params.interactive) {
LOG_TEE("%s: interactive mode on.\n", __func__);
LOG("%s: interactive mode on.\n", __func__);
if (!params.antiprompt.empty()) {
for (const auto & antiprompt : params.antiprompt) {
LOG_TEE("Reverse prompt: '%s'\n", antiprompt.c_str());
LOG("Reverse prompt: '%s'\n", antiprompt.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, antiprompt, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
}
if (params.input_prefix_bos) {
LOG_TEE("Input prefix with BOS\n");
LOG("Input prefix with BOS\n");
}
if (!params.input_prefix.empty()) {
LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
LOG("Input prefix: '%s'\n", params.input_prefix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_prefix, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
if (!params.input_suffix.empty()) {
LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
LOG("Input suffix: '%s'\n", params.input_suffix.c_str());
if (params.verbose_prompt) {
auto tmp = ::llama_tokenize(ctx, params.input_suffix, false, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
LOG("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx, tmp[i]).c_str());
}
}
}
@ -467,13 +445,15 @@ int main(int argc, char ** argv) {
smpl = gpt_sampler_init(model, sparams);
if (!smpl) {
fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
return 1;
}
LOG_TEE("sampling params: \n%s\n", sparams.print().c_str());
LOG_TEE(" sampler constr: \n%s\n", gpt_sampler_print(smpl).c_str());
LOG_TEE("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);
LOG_INF("sampler seed: %u\n", gpt_sampler_get_seed(smpl));
LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
LOG_INF("sampler chain: %s\n", gpt_sampler_print(smpl).c_str());
LOG_INF("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);
// group-attention state
// number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
@ -487,9 +467,9 @@ int main(int argc, char ** argv) {
GGML_ASSERT(ga_w % ga_n == 0 && "grp_attn_w must be a multiple of grp_attn_n"); // NOLINT
//GGML_ASSERT(n_ctx_train % ga_w == 0 && "n_ctx_train must be a multiple of grp_attn_w"); // NOLINT
//GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
LOG_INF("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
}
LOG_TEE("\n\n");
LOG("\n");
if (params.interactive) {
const char * control_message;
@ -501,11 +481,11 @@ int main(int argc, char ** argv) {
" - To return control without starting a new line, end your input with '/'.\n"
" - If you want to submit another line, end your input with '\\'.\n";
}
LOG_TEE("== Running in interactive mode. ==\n");
LOG("== Running in interactive mode. ==\n");
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
LOG_TEE( " - Press Ctrl+C to interject at any time.\n");
LOG( " - Press Ctrl+C to interject at any time.\n");
#endif
LOG_TEE( "%s\n", control_message);
LOG( "%s\n", control_message);
is_interacting = params.interactive_first;
}
@ -544,7 +524,7 @@ int main(int argc, char ** argv) {
llama_token * enc_input_buf = embd_inp.data();
if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
@ -570,9 +550,8 @@ int main(int argc, char ** argv) {
embd.resize(max_embd_size);
console::set_display(console::error);
printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
console::set_display(console::reset);
fflush(stdout);
}
if (ga_n == 1) {
@ -580,16 +559,21 @@ int main(int argc, char ** argv) {
// if we run out of context:
// - take the n_keep first tokens from the original prompt (via n_past)
// - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
if (n_past + (int) embd.size() >= n_ctx) {
if (!params.ctx_shift){
LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
break;
} else {
if (params.n_predict == -2) {
LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
break;
}
const int n_left = n_past - params.n_keep;
const int n_discard = n_left/2;
LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
n_past, n_left, n_ctx, params.n_keep, n_discard);
llama_kv_cache_seq_rm (ctx, 0, params.n_keep , params.n_keep + n_discard);
@ -597,13 +581,14 @@ int main(int argc, char ** argv) {
n_past -= n_discard;
LOG("after swap: n_past = %d\n", n_past);
LOG_DBG("after swap: n_past = %d\n", n_past);
LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
LOG("clear session path\n");
LOG_DBG("clear session path\n");
path_session.clear();
}
}
} else {
// context extension via Self-Extend
while (n_past >= ga_i + ga_w) {
@ -611,10 +596,10 @@ int main(int argc, char ** argv) {
const int bd = (ga_w/ga_n)*(ga_n - 1);
const int dd = (ga_w/ga_n) - ib*bd - ga_w;
LOG("\n");
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
LOG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
LOG_DBG("\n");
LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
LOG_DBG("div: [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
LOG_DBG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
llama_kv_cache_seq_add(ctx, 0, ga_i, n_past, ib*bd);
llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n);
@ -624,7 +609,7 @@ int main(int argc, char ** argv) {
ga_i += ga_w/ga_n;
LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
LOG_DBG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
}
}
@ -656,19 +641,19 @@ int main(int argc, char ** argv) {
n_eval = params.n_batch;
}
LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
LOG_TEE("%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return 1;
}
n_past += n_eval;
LOG("n_past = %d\n", n_past);
LOG_DBG("n_past = %d\n", n_past);
// Display total tokens alongside total time
if (params.n_print > 0 && n_past % params.n_print == 0) {
LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
LOG_DBG("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
}
}
@ -686,14 +671,14 @@ int main(int argc, char ** argv) {
need_to_save_session = false;
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
LOG("saved session to %s\n", path_session.c_str());
LOG_DBG("saved session to %s\n", path_session.c_str());
}
const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
gpt_sampler_accept(smpl, id, /* apply_grammar= */ true);
gpt_sampler_accept(smpl, id, /* accept_grammar= */ true);
// LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
// LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
embd.push_back(id);
@ -703,16 +688,16 @@ int main(int argc, char ** argv) {
// decrement remaining sampling budget
--n_remain;
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
// some user input remains from prompt or interaction, forward it to processing
LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
while ((int) embd_inp.size() > n_consumed) {
embd.push_back(embd_inp[n_consumed]);
// push the prompt in the sampling context in order to apply repetition penalties later
// for the prompt, we don't apply grammar rules
gpt_sampler_accept(smpl, embd_inp[n_consumed], /* apply_grammar= */ false);
gpt_sampler_accept(smpl, embd_inp[n_consumed], /* accept_grammar= */ false);
++n_consumed;
if ((int) embd.size() >= params.n_batch) {
@ -727,7 +712,7 @@ int main(int argc, char ** argv) {
const std::string token_str = llama_token_to_piece(ctx, id, params.special);
// Console/Stream Output
fprintf(stdout, "%s", token_str.c_str());
LOG("%s", token_str.c_str());
// Record Displayed Tokens To Log
// Note: Generated tokens are created one by one hence this check
@ -739,8 +724,6 @@ int main(int argc, char ** argv) {
output_tokens.push_back(id);
output_ss << token_str;
}
fflush(stdout);
}
}
@ -789,13 +772,13 @@ int main(int argc, char ** argv) {
}
if (is_antiprompt) {
LOG("found antiprompt: %s\n", last_output.c_str());
LOG_DBG("found antiprompt: %s\n", last_output.c_str());
}
}
// deal with end of generation tokens in interactive mode
if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
LOG("found an EOG token\n");
LOG_DBG("found an EOG token\n");
if (params.interactive) {
if (!params.antiprompt.empty()) {
@ -809,7 +792,7 @@ int main(int argc, char ** argv) {
chat_add_and_format(model, chat_msgs, "assistant", assistant_ss.str());
}
is_interacting = true;
printf("\n");
LOG("\n");
}
}
@ -820,21 +803,21 @@ int main(int argc, char ** argv) {
}
if (n_past > 0 && is_interacting) {
LOG("waiting for user input\n");
LOG_DBG("waiting for user input\n");
if (params.conversation) {
printf("\n> ");
LOG("\n> ");
}
if (params.input_prefix_bos) {
LOG("adding input prefix BOS token\n");
LOG_DBG("adding input prefix BOS token\n");
embd_inp.push_back(llama_token_bos(model));
}
std::string buffer;
if (!params.input_prefix.empty() && !params.conversation) {
LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
printf("%s", params.input_prefix.c_str());
LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
LOG("%s", params.input_prefix.c_str());
}
// color user input only
@ -857,11 +840,11 @@ int main(int argc, char ** argv) {
if (buffer.length() > 1) {
// append input suffix if any
if (!params.input_suffix.empty() && !params.conversation) {
LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
printf("%s", params.input_suffix.c_str());
LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
LOG("%s", params.input_suffix.c_str());
}
LOG("buffer: '%s'\n", buffer.c_str());
LOG_DBG("buffer: '%s'\n", buffer.c_str());
const size_t original_size = embd_inp.size();
@ -878,7 +861,7 @@ int main(int argc, char ** argv) {
const auto line_inp = ::llama_tokenize(ctx, user_inp, false, format_chat);
const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);
LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
// if user stop generation mid-way, we must add EOT to finish model's last response
if (need_insert_eot && format_chat) {
@ -901,9 +884,9 @@ int main(int argc, char ** argv) {
assistant_ss.str("");
n_remain -= line_inp.size();
LOG("n_remain: %d\n", n_remain);
LOG_DBG("n_remain: %d\n", n_remain);
} else {
LOG("empty line, passing control back\n");
LOG_DBG("empty line, passing control back\n");
}
input_echo = false; // do not echo this again
@ -919,7 +902,7 @@ int main(int argc, char ** argv) {
// end of generation
if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.interactive)) {
LOG_TEE(" [end of text]\n");
LOG(" [end of text]\n");
break;
}
@ -932,11 +915,11 @@ int main(int argc, char ** argv) {
}
if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
LOG("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
}
LOG_TEE("\n");
LOG("\n\n");
gpt_perf_print(ctx, smpl);
write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
@ -950,9 +933,5 @@ int main(int argc, char ** argv) {
ggml_threadpool_free(threadpool);
ggml_threadpool_free(threadpool_batch);
#ifndef LOG_DISABLE_LOGS
LOG_TEE("Log end\n");
#endif // LOG_DISABLE_LOGS
return 0;
}

73
examples/parallel/parallel.cpp
View File

@ -1,7 +1,10 @@
// A basic application simulating a server with multiple clients.
// The clients submit requests to the server and they are processed in parallel.
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cmath>
@ -81,7 +84,9 @@ static void print_date_time() {
char buffer[80];
strftime(buffer, sizeof(buffer), "%Y-%m-%d %H:%M:%S", local_time);
printf("\n\033[35mrun parameters as at %s\033[0m\n", buffer);
LOG_INF("\n");
LOG_INF("\033[35mrun parameters as of %s\033[0m\n", buffer);
LOG_INF("\n");
}
// Define a split string function to ...
@ -100,11 +105,12 @@ int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
return 1;
}
gpt_init();
// number of simultaneous "clients" to simulate
const int32_t n_clients = params.n_parallel;
@ -119,12 +125,6 @@ int main(int argc, char ** argv) {
const bool dump_kv_cache = params.dump_kv_cache;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("parallel", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
@ -137,23 +137,22 @@ int main(int argc, char ** argv) {
// load the prompts from an external file if there are any
if (params.prompt.empty()) {
printf("\n\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
LOG_INF("\033[32mNo new questions so proceed with build-in defaults.\033[0m\n");
} else {
// Output each line of the input params.prompts vector and copy to k_prompts
int index = 0;
printf("\n\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
LOG_INF("\033[32mNow printing the external prompt file %s\033[0m\n\n", params.prompt_file.c_str());
std::vector<std::string> prompts = split_string(params.prompt, '\n');
for (const auto& prompt : prompts) {
k_prompts.resize(index + 1);
k_prompts[index] = prompt;
index++;
printf("%3d prompt: %s\n", index, prompt.c_str());
LOG_INF("%3d prompt: %s\n", index, prompt.c_str());
}
}
fprintf(stderr, "\n\n");
fflush(stderr);
LOG_INF("\n\n");
const int n_ctx = llama_n_ctx(ctx);
@ -182,19 +181,19 @@ int main(int argc, char ** argv) {
const auto t_main_start = ggml_time_us();
LOG_TEE("%s: Simulating parallel requests from clients:\n", __func__);
LOG_TEE("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_TEE("\n");
LOG_INF("%s: Simulating parallel requests from clients:\n", __func__);
LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_INF("\n");
{
LOG_TEE("%s: Evaluating the system prompt ...\n", __func__);
LOG_INF("%s: Evaluating the system prompt ...\n", __func__);
for (int32_t i = 0; i < n_tokens_system; ++i) {
llama_batch_add(batch, tokens_system[i], i, { 0 }, false);
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -203,10 +202,10 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("\n");
LOG_INF("\n");
}
LOG_TEE("Processing requests ...\n\n");
LOG_INF("Processing requests ...\n\n");
while (true) {
if (dump_kv_cache) {
@ -237,7 +236,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
}
LOG_TEE("%s: clearing the KV cache\n", __func__);
LOG_INF("%s: clearing the KV cache\n", __func__);
}
// insert new sequences for decoding
@ -272,7 +271,7 @@ int main(int argc, char ** argv) {
client.n_decoded = 0;
client.i_batch = batch.n_tokens - 1;
LOG_TEE("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
LOG_INF("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
g_seq_id += 1;
@ -316,11 +315,11 @@ int main(int argc, char ** argv) {
if (ret != 0) {
if (n_batch == 1 || ret < 0) {
// if you get here, it means the KV cache is full - try increasing it via the context size
LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
LOG_ERR("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
return 1;
}
LOG("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
LOG_ERR("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
n_cache_miss += 1;
@ -331,7 +330,7 @@ int main(int argc, char ** argv) {
continue;
}
LOG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
LOG_DBG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
for (auto & client : clients) {
if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
@ -376,7 +375,7 @@ int main(int argc, char ** argv) {
const auto t_main_end = ggml_time_us();
LOG_TEE("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \nInput: %s\n\033[35mResponse: %s\033[0m\n\n",
LOG_INF("\033[31mClient %3d, seq %3d/%3d, prompt %4d t, response %4d t, time %5.2f s, speed %5.2f t/s, cache miss %d \033[0m \n\nInput: %s\n\033[35mResponse: %s\033[0m\n\n",
client.id, client.seq_id, n_seq, client.n_prompt, client.n_decoded,
(t_main_end - client.t_start_prompt) / 1e6,
(double) (client.n_prompt + client.n_decoded) / (t_main_end - client.t_start_prompt) * 1e6,
@ -399,22 +398,22 @@ int main(int argc, char ** argv) {
print_date_time();
LOG_TEE("\n%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
LOG_INF("%s: n_parallel = %d, n_sequences = %d, cont_batching = %d, system tokens = %d\n", __func__, n_clients, n_seq, cont_batching, n_tokens_system);
if (params.prompt_file.empty()) {
params.prompt_file = "used built-in defaults";
}
LOG_TEE("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
LOG_TEE("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str());
LOG_INF("External prompt file: \033[32m%s\033[0m\n", params.prompt_file.c_str());
LOG_INF("Model and path used: \033[32m%s\033[0m\n\n", params.model.c_str());
LOG_TEE("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Total speed (AVG): %6s speed: %5.2f t/s\n", "", (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
LOG_TEE("Cache misses: %6d\n", n_cache_miss);
LOG_INF("Total prompt tokens: %6d, speed: %5.2f t/s\n", n_total_prompt, (double) (n_total_prompt ) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Total gen tokens: %6d, speed: %5.2f t/s\n", n_total_gen, (double) (n_total_gen ) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Total speed (AVG): %6s speed: %5.2f t/s\n", "", (double) (n_total_prompt + n_total_gen) / (t_main_end - t_main_start) * 1e6);
LOG_INF("Cache misses: %6d\n", n_cache_miss);
LOG_TEE("\n");
LOG_INF("\n");
// TODO: print sampling/grammar timings for all clients
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx);
llama_batch_free(batch);
@ -423,7 +422,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

67
examples/passkey/passkey.cpp
View File

@ -1,4 +1,6 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <cmath>
@ -7,9 +9,9 @@
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
@ -19,11 +21,12 @@ int main(int argc, char ** argv) {
params.n_keep = 32;
params.i_pos = -1;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_PASSKEY, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
return 1;
}
gpt_init();
int n_junk = params.n_junk;
int n_keep = params.n_keep;
int n_grp = params.grp_attn_n;
@ -63,7 +66,7 @@ int main(int argc, char ** argv) {
llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
if (model == NULL) {
fprintf(stderr , "%s: error: unable to load model\n" , __func__);
LOG_ERR("%s: unable to load model\n" , __func__);
return 1;
}
@ -77,7 +80,7 @@ int main(int argc, char ** argv) {
llama_context * ctx = llama_new_context_with_model(model, ctx_params);
if (ctx == NULL) {
fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
LOG_ERR("%s: failed to create the llama_context\n" , __func__);
return 1;
}
@ -107,14 +110,14 @@ int main(int argc, char ** argv) {
const int n_batch = ctx_params.n_batch;
const int n_batch_grp = ctx_params.n_batch/n_grp;
LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
LOG_INF("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
// print the prompt token-by-token
LOG_TEE("\n");
LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
LOG_TEE("prompt tokens: %d\n", n_tokens_all);
//LOG_TEE("prompt: %s\n", params.prompt.c_str());
LOG_INF("\n");
LOG_INF("prefix tokens: %d\n", n_tokens_prefix);
LOG_INF("prompt tokens: %d\n", n_tokens_all);
//LOG_INF("prompt: %s\n", params.prompt.c_str());
llama_batch batch = llama_batch_init(params.n_batch, 0, 1);
@ -145,11 +148,11 @@ int main(int argc, char ** argv) {
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_INF("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
LOG_INF("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
if (i + n_batch >= n_tokens_all) {
break;
@ -159,7 +162,7 @@ int main(int argc, char ** argv) {
for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
const int n_discard = n_batch;
LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
@ -179,18 +182,18 @@ int main(int argc, char ** argv) {
}
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return 1;
}
LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
LOG_INF("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
}
{
const int n_discard = n_past - n_ctx + n_predict;
if (n_discard > 0) {
LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
llama_kv_cache_seq_rm (ctx, 0, n_keep , n_keep + n_discard);
llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx, -n_discard);
@ -201,17 +204,16 @@ int main(int argc, char ** argv) {
}
}
LOG_TEE("\n");
LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
LOG_TEE("\n");
LOG_INF("\n");
LOG_INF("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
LOG_INF("\n");
// main loop
int n_cur = n_tokens_all;
int n_decode = 0;
LOG_TEE("%s", prompt_suffix.c_str());
fflush(stdout);
LOG_INF("%s", prompt_suffix.c_str());
const auto t_main_start = ggml_time_us();
@ -220,17 +222,14 @@ int main(int argc, char ** argv) {
{
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
llama_sampler_accept(smpl, new_token_id);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
LOG_TEE("\n");
LOG("\n");
break;
}
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
n_decode += 1;
@ -245,22 +244,22 @@ int main(int argc, char ** argv) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
LOG("\n");
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
LOG("\n");
llama_sampler_free(smpl);

361
examples/perplexity/perplexity.cpp
View File

@ -1,18 +1,21 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <array>
#include <atomic>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <ctime>
#include <fstream>
#include <mutex>
#include <random>
#include <sstream>
#include <thread>
#include <mutex>
#include <atomic>
#include <vector>
#include <array>
#include <fstream>
#include <sstream>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
@ -40,7 +43,7 @@ static void write_logfile(
}
if (params.hellaswag) {
fprintf(stderr, "%s: warning: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
LOG_WRN("%s: logging results is not implemented for HellaSwag. No files will be written.\n", __func__);
return;
}
@ -48,7 +51,7 @@ static void write_logfile(
const bool success = fs_create_directory_with_parents(params.logdir);
if (!success) {
fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
LOG_WRN("%s: failed to create logdir %s, cannot write logfile\n",
__func__, params.logdir.c_str());
return;
}
@ -57,7 +60,7 @@ static void write_logfile(
FILE * logfile = fopen(logfile_path.c_str(), "w");
if (logfile == NULL) {
fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
return;
}
@ -343,16 +346,16 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
const int n_ctx = llama_n_ctx(ctx);
if (int(tokens.size()) < 2*n_ctx) {
fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
LOG_ERR("%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
n_ctx);
fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
return {std::move(tokens), 0., {}, {}};
}
@ -363,16 +366,16 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
prob_history.resize(tokens.size());
if (params.ppl_stride <= 0) {
fprintf(stderr, "%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
LOG_ERR("%s: stride is %d but must be greater than zero!\n",__func__,params.ppl_stride);
return {tokens, -1, logit_history, prob_history};
}
const int calc_chunk = n_ctx;
fprintf(stderr, "%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
LOG_INF("%s: have %zu tokens. Calculation chunk = %d\n", __func__, tokens.size(), calc_chunk);
if (int(tokens.size()) <= calc_chunk) {
fprintf(stderr, "%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
LOG_ERR("%s: there are only %zu tokens, this is not enough for a context size of %d and stride %d\n",__func__,
tokens.size(), n_ctx, params.ppl_stride);
return {tokens, -1, logit_history, prob_history};
}
@ -386,14 +389,14 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
int count = 0;
double nll = 0.0;
fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
LOG_INF("%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
for (int i = 0; i < n_chunk; ++i) {
const int start = i * params.ppl_stride;
const int end = start + calc_chunk;
const int num_batches = (calc_chunk + n_batch - 1) / n_batch;
//fprintf(stderr, "%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
//LOG_DBG("%s: evaluating %d...%d using %d batches\n", __func__, start, end, num_batches);
std::vector<float> logits;
@ -406,10 +409,10 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
const int batch_start = start + j * n_batch;
const int batch_size = std::min(end - batch_start, n_batch);
//fprintf(stderr, " Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
//LOG_DBG(" Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
// TODO: use llama_batch.logits instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
//fprintf(stderr, "%s : failed to eval\n", __func__);
//LOG_ERR("%s : failed to eval\n", __func__);
return {tokens, -1, logit_history, prob_history};
}
@ -433,16 +436,17 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
if (i == 0) {
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total * n_chunk);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
LOG("%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
LOG("%.2f minutes\n", total_seconds / 60.0);
}
LOG("\n");
//fprintf(stderr, "%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
//LOG_DBG("%s: using tokens %d...%d\n",__func__,params.n_ctx - params.ppl_stride + start, params.n_ctx + start);
for (int j = n_ctx - params.ppl_stride - 1; j < n_ctx - 1; ++j) {
// Calculate probability of next token, given the previous ones.
@ -459,13 +463,12 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
}
// perplexity is e^(average negative log-likelihood)
if (params.ppl_output_type == 0) {
printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
LOG("[%d]%.4lf,", i + 1, std::exp(nll / count));
} else {
printf("%8d %.4lf\n", i*params.ppl_stride, std::exp(nll / count));
LOG("%8d %.4lf\n", i*params.ppl_stride, std::exp(nll / count));
}
fflush(stdout);
}
printf("\n");
LOG("\n");
return {tokens, std::exp(nll / count), logit_history, prob_history};
}
@ -487,26 +490,26 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
if (!params.logits_file.empty()) {
logits_stream.open(params.logits_file.c_str(), std::ios::binary);
if (!logits_stream.is_open()) {
fprintf(stderr, "%s: failed to open %s for writing\n", __func__, params.logits_file.c_str());
LOG_ERR("%s: failed to open %s for writing\n", __func__, params.logits_file.c_str());
return {};
}
fprintf(stderr, "%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
LOG_INF("%s: saving all logits to %s\n", __func__, params.logits_file.c_str());
logits_stream.write("_logits_", 8);
logits_stream.write(reinterpret_cast<const char *>(&n_ctx), sizeof(n_ctx));
}
auto tim1 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
LOG_INF("%s: tokenizing the input ..\n", __func__);
std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
auto tim2 = std::chrono::high_resolution_clock::now();
fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
if (int(tokens.size()) < 2*n_ctx) {
fprintf(stderr, "%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
LOG_ERR("%s: you need at least %d tokens to evaluate perplexity with a context of %d\n",__func__,2*n_ctx,
n_ctx);
fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
return {std::move(tokens), 0., {}, {}};
}
@ -539,7 +542,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
logits.reserve((size_t)n_ctx * n_vocab);
}
fprintf(stderr, "%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
LOG_INF("%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
@ -612,7 +615,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
}
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval\n", __func__);
LOG_INF("%s : failed to eval\n", __func__);
return {tokens, -1, logit_history, prob_history};
}
@ -627,14 +630,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
llama_synchronize(ctx);
const auto t_end = std::chrono::high_resolution_clock::now();
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total*n_chunk/n_seq);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
LOG("%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
LOG("%.2f minutes\n", total_seconds / 60.0);
}
LOG("\n");
for (int seq = 0; seq < n_seq_batch; seq++) {
const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx + first);
@ -655,19 +659,18 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
// perplexity is e^(average negative log-likelihood)
if (params.ppl_output_type == 0) {
printf("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
LOG("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
} else {
double av = nll/count;
double av2 = nll2/count - av*av;
if (av2 > 0) av2 = sqrt(av2/(count-1));
printf("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
LOG("%8d %.4lf %4lf %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
}
}
fflush(stdout);
logits.clear();
}
printf("\n");
LOG("\n");
nll2 /= count;
nll /= count;
@ -675,9 +678,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
nll2 -= nll * nll;
if (nll2 > 0) {
nll2 = sqrt(nll2/(count-1));
printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
LOG_INF("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
} else {
printf("Unexpected negative standard deviation of log(prob)\n");
LOG_ERR("Unexpected negative standard deviation of log(prob)\n");
}
llama_batch_free(batch);
@ -703,7 +706,7 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
const int ret = llama_decode(ctx, batch_view);
if (ret != 0) {
LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
return false;
}
@ -789,15 +792,15 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
}
if (prompt_lines.size() % 6 != 0) {
fprintf(stderr, "%s : number of lines in prompt not a multiple of 6.\n", __func__);
LOG_ERR("%s : number of lines in prompt not a multiple of 6.\n", __func__);
return;
}
size_t hs_task_count = prompt_lines.size()/6;
fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
LOG_INF("%s : loaded %zu tasks from prompt.\n", __func__, hs_task_count);
const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
fprintf(stderr, "================================= is_spm = %d\n", is_spm);
LOG_INF("================================= is_spm = %d\n", is_spm);
// The tasks should be randomized so the score stabilizes quickly.
bool randomize_tasks = true;
@ -824,7 +827,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
std::vector<llama_token> seq_tokens[4];
};
fprintf(stderr, "%s : selecting %zu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first") );
LOG_INF("%s : selecting %zu %s tasks.\n", __func__, hs_task_count, (randomize_tasks?"randomized":"the first") );
// Select and read data from prompt lines
std::vector<hs_data_t> hs_data(hs_task_count);
@ -870,9 +873,9 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
}
}
fprintf(stderr, "%s : calculating hellaswag score over selected tasks.\n", __func__);
LOG_INF("%s : calculating hellaswag score over selected tasks.\n", __func__);
printf("\ntask\tacc_norm\n");
LOG("\ntask\tacc_norm\n");
double acc = 0.0f;
@ -940,7 +943,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
}
if (i0 == i1) {
fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0);
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
return;
}
@ -948,7 +951,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
fprintf(stderr, "%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -998,7 +1001,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
}
}
//printf("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_cur.gold_ending_idx);
//LOG("max logprob ending idx %lu, gold ending idx %lu\n", ending_logprob_max_idx, hs_cur.gold_ending_idx);
// If the gold ending got the maximum logprobe add one accuracy point
if (ending_logprob_max_idx == hs_cur.gold_ending_idx) {
@ -1006,8 +1009,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
}
// Print the accumulated accuracy mean x 100
printf("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0);
fflush(stdout);
LOG("%zu\t%.8lf\n", i + 1, acc/double(i + 1)*100.0);
}
i0 = i1 - 1;
@ -1015,7 +1017,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
llama_batch_free(batch);
printf("\n");
LOG("\n");
}
struct winogrande_entry {
@ -1059,7 +1061,7 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
}
}
if (ipos != 4) {
printf("%s: failed to find comma separators in <%s>\n", __func__, line.c_str());
LOG_ERR("%s: failed to find comma separators in <%s>\n", __func__, line.c_str());
continue;
}
auto sentence = line[comma_pos[0]+1] == '"' ? line.substr(comma_pos[0]+2, comma_pos[1] - comma_pos[0] - 3)
@ -1073,13 +1075,13 @@ static std::vector<winogrande_entry> load_winogrande_from_csv(const std::string
if (sentence[where] == '_') break;
}
if (where == int(sentence.size())) {
printf("%s: no _ in <%s>\n", __func__, sentence.c_str());
LOG_ERR("%s: no _ in <%s>\n", __func__, sentence.c_str());
continue;
}
std::istringstream stream(answer.c_str());
int i_answer; stream >> i_answer;
if (stream.fail() || i_answer < 1 || i_answer > 2) {
printf("%s: failed to parse answer <%s>\n", __func__, answer.c_str());
LOG_ERR("%s: failed to parse answer <%s>\n", __func__, answer.c_str());
continue;
}
result.emplace_back();
@ -1108,14 +1110,14 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
auto data = load_winogrande_from_csv(params.prompt);
if (data.empty()) {
fprintf(stderr, "%s: no tasks\n", __func__);
LOG_ERR("%s: no tasks\n", __func__);
return;
}
fprintf(stderr, "%s : loaded %zu tasks from prompt.\n", __func__, data.size());
LOG_INF("%s : loaded %zu tasks from prompt.\n", __func__, data.size());
if (params.winogrande_tasks > 0 && params.winogrande_tasks < data.size()) {
fprintf(stderr, "%s : selecting %zu random tasks\n", __func__, params.winogrande_tasks);
LOG_INF("%s : selecting %zu random tasks\n", __func__, params.winogrande_tasks);
std::mt19937 rng(1);
std::vector<int> aux(data.size());
for (int i = 0; i < int(data.size()); ++i) {
@ -1133,7 +1135,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
data = std::move(selected);
}
fprintf(stderr, "%s : tokenizing selected tasks\n", __func__);
LOG_INF("%s : tokenizing selected tasks\n", __func__);
for (auto & task : data) {
task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, true);
@ -1156,7 +1158,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], true).size();
}
fprintf(stderr, "%s : calculating winogrande score over selected tasks.\n", __func__);
LOG_INF("%s : calculating winogrande score over selected tasks.\n", __func__);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_ctx = llama_n_ctx(ctx);
@ -1217,7 +1219,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
}
if (i0 == i1) {
fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0);
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
return;
}
@ -1225,7 +1227,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
fprintf(stderr, "%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -1285,20 +1287,20 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
++n_done;
// print the accumulated accuracy mean x 100
printf("%zu\t%.4lf\t%10.6f %10.6f %d %d\n", i+1, 100.0 * n_correct/n_done, score_1st, score_2nd, result, task.answer);
fflush(stdout);
LOG("%zu\t%.4lf\t%10.6f %10.6f %d %d\n", i+1, 100.0 * n_correct/n_done, score_1st, score_2nd, result, task.answer);
}
i0 = i1 - 1;
}
printf("\n");
LOG("\n");
if (n_done < 100) return;
const float p = 1.f*n_correct/n_done;
const float sigma = 100.f*sqrt(p*(1-p)/(n_done-1));
printf("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma);
LOG_INF("Final Winogrande score(%d tasks): %.4lf +/- %.4lf\n", n_done, 100*p, sigma);
}
static bool deserialize_string(std::istream & in, std::string & str) {
@ -1347,7 +1349,7 @@ struct multiple_choice_task {
static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choice_task& task, bool log_error) {
if (task.question.empty() || task.mc1.answers.empty()) {
if (log_error) {
printf("%s: found bad task with empty question and/or answers\n", __func__);
LOG_ERR("%s: found bad task with empty question and/or answers\n", __func__);
}
return false;
}
@ -1355,7 +1357,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choic
for (auto& answer : task.mc1.answers) {
if (answer.empty()) {
if (log_error) {
printf("%s: found empty answer\n", __func__);
LOG_ERR("%s: found empty answer\n", __func__);
}
return false;
}
@ -1409,14 +1411,14 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
uint32_t n_task;
strstream.read((char *)&n_task, sizeof(n_task));
if (strstream.fail() || n_task == 0) {
printf("%s: no tasks\n", __func__);
LOG_ERR("%s: no tasks\n", __func__);
return;
}
printf("%s: there are %u tasks in prompt\n", __func__, n_task);
LOG_INF("%s: there are %u tasks in prompt\n", __func__, n_task);
std::vector<uint32_t> task_pos(n_task);
strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
if (strstream.fail()) {
printf("%s: failed to read task positions from prompt\n", __func__);
LOG_ERR("%s: failed to read task positions from prompt\n", __func__);
return;
}
@ -1424,21 +1426,21 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
if (params.multiple_choice_tasks == 0 || params.multiple_choice_tasks >= (size_t)n_task) {
// Use all tasks
tasks.resize(n_task);
printf("%s: reading tasks", __func__);
LOG_INF("%s: reading tasks", __func__);
int n_dot = std::max((int) n_task/100, 1);
int i = 0;
for (auto& task : tasks) {
++i;
if (!task.deserialize(strstream)) {
printf("%s: failed to read task %d of %u\n", __func__, i, n_task);
LOG_ERR("%s: failed to read task %d of %u\n", __func__, i, n_task);
return;
}
if (i%n_dot == 0) printf(".");
if (i%n_dot == 0) LOG(".");
}
printf("done\n");
LOG("done\n");
}
else {
printf("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task);
LOG_INF("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task);
std::mt19937 rng(1);
std::vector<int> aux(n_task);
for (uint32_t i = 0; i < n_task; ++i) aux[i] = i;
@ -1451,18 +1453,16 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
aux.pop_back();
strstream.seekg(task_pos[idx], std::ios::beg);
if (!task.deserialize(strstream)) {
printf("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]);
LOG_ERR("%s: failed to read task %d at position %u\n", __func__, idx, task_pos[idx]);
return;
}
}
n_task = params.multiple_choice_tasks;
}
printf("%s: preparing task data", __func__);
fflush(stdout);
LOG_INF("%s: preparing task data", __func__);
if (n_task > 500) {
printf("...");
fflush(stdout);
LOG("...");
std::atomic<int> counter(0);
std::atomic<int> n_bad(0);
auto prepare = [&counter, &n_bad, &tasks, ctx] () {
@ -1486,11 +1486,10 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
for (auto& w : workers) w = std::thread(prepare);
prepare();
for (auto& w : workers) w.join();
printf("done\n");
fflush(stdout);
LOG("done\n");
int nbad = n_bad;
if (nbad > 0) {
printf("%s: found %d malformed tasks\n", __func__, nbad);
LOG_ERR("%s: found %d malformed tasks\n", __func__, nbad);
return;
}
} else {
@ -1502,16 +1501,15 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
return;
}
if (i_task%n_dot == 0) {
printf(".");
fflush(stdout);
LOG(".");
}
}
printf("done\n");
LOG("done\n");
}
printf("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size());
LOG_INF("%s : calculating TruthfulQA score over %zu tasks.\n", __func__, tasks.size());
printf("\ntask\tacc_norm\n");
LOG("\ntask\tacc_norm\n");
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const int n_ctx = llama_n_ctx(ctx);
@ -1590,7 +1588,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
}
if (i0 == i1) {
fprintf(stderr, "%s : task %zu does not fit in the context window\n", __func__, i0);
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
return;
}
@ -1598,7 +1596,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
// decode all tasks [i0, i1)
if (!decode_helper(ctx, batch, batch_logits, n_batch, n_vocab)) {
fprintf(stderr, "%s: llama_decode() failed\n", __func__);
LOG_ERR("%s: llama_decode() failed\n", __func__);
return;
}
@ -1622,13 +1620,13 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
// compute the logprobs for each ending of the decoded tasks
for (size_t i = i0; i < i1; ++i) {
auto & cur_task = tasks[i];
//printf("==== Evaluating <%s> with correct answer ", cur_task.question.c_str());
//LOG("==== Evaluating <%s> with correct answer ", cur_task.question.c_str());
//for (int j = 0; j < int(cur_task.mc1.labels.size()); ++j) {
// if (cur_task.mc1.labels[j] == 1) {
// printf("%d", j+1);
// LOG("%d", j+1);
// }
//}
//printf("\n common_prefix: %zu\n", cur_task.common_prefix);
//LOG("\n common_prefix: %zu\n", cur_task.common_prefix);
// get the logits of the last token of the common prefix
std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
@ -1640,13 +1638,13 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
size_t count = 1;
float log_prob = std::log(first_probs[cur_task.seq_tokens[s][cur_task.common_prefix]]);
for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
//printf(" %zu %g\n", ir, eval_results[ir]);
//LOG(" %zu %g\n", ir, eval_results[ir]);
++count;
log_prob += eval_results[ir++];
}
cur_task.log_probs[s] = log_prob / count;
//printf(" Final: %g\n", log_prob / count);
//printf(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count);
//LOG(" Final: %g\n", log_prob / count);
//LOG(" <%s> : %g\n", cur_task.mc1.answers[s].c_str(), log_prob/count);
}
// Find the ending with maximum logprob
@ -1666,8 +1664,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
++n_done;
// Print the accumulated accuracy mean x 100
printf("%d\t%.8lf\n", n_done, 100.*n_correct/n_done);
fflush(stdout);
LOG("%d\t%.8lf\n", n_done, 100.*n_correct/n_done);
}
i0 = i1 - 1;
@ -1679,29 +1676,30 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
float p = 1.f*n_correct/n_done;
float sigma = sqrt(p*(1-p)/(n_done-1));
printf("\n Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
LOG("\n");
LOG_INF("Final result: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
p = 1.f*n_done/n_tot_answers;
sigma = sqrt(p*(1-p)/(n_done-1));
printf("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
LOG_INF("Random chance: %.4f +/- %.4f\n", 100.f*p, 100.f*sigma);
printf("\n");
LOG_INF("\n");
}
static void kl_divergence(llama_context * ctx, const gpt_params & params) {
if (params.logits_file.empty()) {
fprintf(stderr, "%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
LOG_ERR("%s: you must provide a name of a file containing the log probabilities of the base model\n", __func__);
return;
}
std::ifstream in(params.logits_file.c_str(), std::ios::binary);
if (!in) {
fprintf(stderr, "%s: failed to open %s\n", __func__, params.logits_file.c_str());
LOG_ERR("%s: failed to open %s\n", __func__, params.logits_file.c_str());
return;
}
{
char check[9]; check[8] = 0;
in.read(check, 8);
if (in.fail() || strncmp("_logits_", check, 8) != 0) {
fprintf(stderr, "%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str());
LOG_ERR("%s: %s does not look like a file containing log-probabilities\n", __func__, params.logits_file.c_str());
return;
}
}
@ -1709,7 +1707,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
uint32_t n_ctx;
in.read((char *)&n_ctx, sizeof(n_ctx));
if (n_ctx > llama_n_ctx(ctx)) {
fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
LOG_ERR("%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
__func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
}
@ -1717,16 +1715,16 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
in.read((char *)&n_vocab, sizeof(n_vocab));
in.read((char *)&n_chunk, sizeof(n_chunk));
if (in.fail()) {
fprintf(stderr, "%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
LOG_ERR("%s: failed reading n_vocab, n_chunk from %s\n", __func__, params.logits_file.c_str());
return;
}
if (n_vocab != llama_n_vocab(llama_get_model(ctx))) {
fprintf(stderr, "%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
LOG_ERR("%s: inconsistent vocabulary (%d vs %d)\n", __func__, n_vocab, llama_n_vocab(llama_get_model(ctx)));
}
std::vector<llama_token> tokens(n_ctx * n_chunk);
if (in.read((char *)tokens.data(), tokens.size()*sizeof(tokens[0])).fail()) {
fprintf(stderr, "%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
LOG_ERR("%s: failed reading evaluation tokens from %s\n", __func__, params.logits_file.c_str());
return;
}
@ -1775,7 +1773,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
const auto t_start = std::chrono::high_resolution_clock::now();
if (in.read((char *)log_probs_uint16.data(), log_probs_uint16.size()*sizeof(uint16_t)).fail()) {
fprintf(stderr, "%s: failed reading log-probs for chunk %d\n", __func__, i);
LOG_ERR("%s: failed reading log-probs for chunk %d\n", __func__, i);
return;
}
@ -1796,7 +1794,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
// TODO: use llama_batch.logits instead of relying on logits_all == true
if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
fprintf(stderr, "%s : failed to eval\n", __func__);
LOG_ERR("%s : failed to eval\n", __func__);
return;
}
@ -1813,16 +1811,16 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
if (i == 0) {
const float t_total = std::chrono::duration<float>(t_end - t_start).count();
fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
int total_seconds = (int)(t_total * n_chunk);
if (total_seconds >= 60*60) {
fprintf(stderr, "%d hours ", total_seconds / (60*60));
LOG("%d hours ", total_seconds / (60*60));
total_seconds = total_seconds % (60*60);
}
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
printf("\nchunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p\n");
LOG("%.2f minutes\n", total_seconds / 60.0);
}
LOG("\n");
LOG("chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p\n");
const int first = n_ctx/2;
const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
@ -1831,79 +1829,77 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
p_diff_ptr += n_ctx - 1 - first;
kld_ptr += n_ctx - 1 - first;
printf("%4d", i+1);
LOG("%4d", i+1);
auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
const double ppl_val = exp(log_ppl.first);
const double ppl_unc = ppl_val * log_ppl.second; // ppl_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl.second ** 2 )
printf(" %9.4lf ± %9.4lf", ppl_val, ppl_unc);
LOG(" %9.4lf ± %9.4lf", ppl_val, ppl_unc);
auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count);
const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count);
const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first;
const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov);
printf(" %10.5lf ± %10.5lf", log_ppl_ratio_val, log_ppl_ratio_unc);
LOG(" %10.5lf ± %10.5lf", log_ppl_ratio_val, log_ppl_ratio_unc);
auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
printf(" %10.5lf ± %10.5lf", kl_div.first, kl_div.second);
LOG(" %10.5lf ± %10.5lf", kl_div.first, kl_div.second);
auto p_diff_mse = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count);
const double p_diff_rms_val = sqrt(p_diff_mse.first);
const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second;
printf(" %6.3lf ± %6.3lf %%", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
LOG(" %6.3lf ± %6.3lf %%", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
double p_top_val = 1.*kld.n_same_top/kld.count;
double p_top_unc = sqrt(p_top_val*(1 - p_top_val)/(kld.count - 1));
printf(" %6.3lf ± %6.3lf %%", 100.0*p_top_val, 100.0*p_top_unc);
LOG(" %6.3lf ± %6.3lf %%", 100.0*p_top_val, 100.0*p_top_unc);
printf("\n");
fflush(stdout);
LOG("\n");
logits.clear();
}
printf("\n");
LOG("\n");
if (kld.count < 100) return; // we do not wish to do statistics on so few values
std::sort(kld_values.begin(), kld_values.end());
std::sort(p_diff_values.begin(), p_diff_values.end());
printf("====== Perplexity statistics ======\n");
LOG("====== Perplexity statistics ======\n");
auto log_ppl = mean_and_uncertainty(kld.sum_nll, kld.sum_nll2, kld.count);
const double ppl_val = exp(log_ppl.first);
const double ppl_unc = ppl_val * log_ppl.second; // ppl_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl.second ** 2 )
printf("Mean PPL(Q) : %10.6lf ± %10.6lf\n", ppl_val, ppl_unc);
LOG("Mean PPL(Q) : %10.6lf ± %10.6lf\n", ppl_val, ppl_unc);
auto log_ppl_base = mean_and_uncertainty(kld.sum_nll_base, kld.sum_nll_base2, kld.count);
const double ppl_base_val = exp(log_ppl_base.first);
const double ppl_base_unc = ppl_base_val * log_ppl_base.second; // ppl_base_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl_base.second ** 2 )
printf("Mean PPL(base) : %10.6lf ± %10.6lf\n", ppl_base_val, ppl_base_unc);
LOG("Mean PPL(base) : %10.6lf ± %10.6lf\n", ppl_base_val, ppl_base_unc);
const double log_ppl_cov = covariance(kld.sum_nll, kld.sum_nll_base, kld.sum_nll_nll_base, kld.count);
// printf("Cov(ln(PPL(Q)), ln(PPL(base))): %10.6lf\n", log_ppl_cov);
// LOG("Cov(ln(PPL(Q)), ln(PPL(base))): %10.6lf\n", log_ppl_cov);
const double log_ppl_cor = log_ppl_cov / (log_ppl.second*log_ppl_base.second);
printf("Cor(ln(PPL(Q)), ln(PPL(base))): %6.2lf%%\n", 100.0*log_ppl_cor);
LOG("Cor(ln(PPL(Q)), ln(PPL(base))): %6.2lf%%\n", 100.0*log_ppl_cor);
const double log_ppl_ratio_val = log_ppl.first - log_ppl_base.first;
const double log_ppl_ratio_unc = sqrt(log_ppl.second*log_ppl.second + log_ppl_base.second*log_ppl_base.second - 2.0*log_ppl_cov);
printf("Mean ln(PPL(Q)/PPL(base)) : %10.6lf ± %10.6lf\n", log_ppl_ratio_val, log_ppl_ratio_unc);
LOG("Mean ln(PPL(Q)/PPL(base)) : %10.6lf ± %10.6lf\n", log_ppl_ratio_val, log_ppl_ratio_unc);
const double ppl_ratio_val = exp(log_ppl_ratio_val);
const double ppl_ratio_unc = ppl_ratio_val * log_ppl_ratio_unc; // ppl_ratio_unc = sqrt( (dexp(x) / dx) ** 2 * log_ppl_ratio.second ** 2 )
printf("Mean PPL(Q)/PPL(base) : %10.6lf ± %10.6lf\n", ppl_ratio_val, ppl_ratio_unc);
LOG("Mean PPL(Q)/PPL(base) : %10.6lf ± %10.6lf\n", ppl_ratio_val, ppl_ratio_unc);
const double ppl_cov = ppl_val * ppl_base_val * log_ppl_cov;
const double ppl_diff_val = ppl_val - ppl_base_val;
const double ppl_diff_unc = sqrt(ppl_unc*ppl_unc + ppl_base_unc*ppl_base_unc - 2.0*ppl_cov);
printf("Mean PPL(Q)-PPL(base) : %10.6lf ± %10.6lf\n", ppl_diff_val, ppl_diff_unc);
LOG("Mean PPL(Q)-PPL(base) : %10.6lf ± %10.6lf\n", ppl_diff_val, ppl_diff_unc);
printf("\n");
LOG("\n");
printf("====== KL divergence statistics ======\n");
LOG("====== KL divergence statistics ======\n");
auto kl_div = mean_and_uncertainty(kld.sum_kld, kld.sum_kld2, kld.count);
printf("Mean KLD: %10.6lf ± %10.6lf\n", kl_div.first, kl_div.second);
LOG("Mean KLD: %10.6lf ± %10.6lf\n", kl_div.first, kl_div.second);
auto kld_median = kld_values.size()%2 == 0 ? 0.5f*(kld_values[kld_values.size()/2] + kld_values[kld_values.size()/2-1])
: kld_values[kld_values.size()/2];
@ -1915,50 +1911,49 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
return (1 - p)*values[ip] + p*values[std::min(ip+1, values.size()-1)];
};
printf("Maximum KLD: %10.6f\n", kld_values.back());
printf("99.9%% KLD: %10.6f\n", percentile(kld_values, 0.999f));
printf("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f));
printf("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f));
printf("Median KLD: %10.6f\n", kld_median);
printf("10.0%% KLD: %10.6f\n", percentile(kld_values, 0.100f));
printf(" 5.0%% KLD: %10.6f\n", percentile(kld_values, 0.050f));
printf(" 1.0%% KLD: %10.6f\n", percentile(kld_values, 0.010f));
printf("Minimum KLD: %10.6f\n", kld_values.front());
LOG("Maximum KLD: %10.6f\n", kld_values.back());
LOG("99.9%% KLD: %10.6f\n", percentile(kld_values, 0.999f));
LOG("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f));
LOG("99.0%% KLD: %10.6f\n", percentile(kld_values, 0.990f));
LOG("Median KLD: %10.6f\n", kld_median);
LOG("10.0%% KLD: %10.6f\n", percentile(kld_values, 0.100f));
LOG(" 5.0%% KLD: %10.6f\n", percentile(kld_values, 0.050f));
LOG(" 1.0%% KLD: %10.6f\n", percentile(kld_values, 0.010f));
LOG("Minimum KLD: %10.6f\n", kld_values.front());
printf("\n");
LOG("\n");
printf("====== Token probability statistics ======\n");
LOG("====== Token probability statistics ======\n");
auto p_diff = mean_and_uncertainty(kld.sum_p_diff, kld.sum_p_diff2, kld.count);
printf("Mean Δp: %6.3lf ± %5.3lf %%\n", 100.0*p_diff.first, 100.0*p_diff.second);
LOG("Mean Δp: %6.3lf ± %5.3lf %%\n", 100.0*p_diff.first, 100.0*p_diff.second);
auto p_diff_median = p_diff_values.size()%2 == 0 ? 0.5f*(p_diff_values[p_diff_values.size()/2] + p_diff_values[p_diff_values.size()/2-1])
: p_diff_values[p_diff_values.size()/2];
printf("Maximum Δp: %6.3lf%%\n", 100.0*p_diff_values.back());
printf("99.9%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.999f));
printf("99.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.990f));
printf("95.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.950f));
printf("90.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.900f));
printf("75.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.750f));
printf("Median Δp: %6.3lf%%\n", 100.0*p_diff_median);
printf("25.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.250f));
printf("10.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.100f));
printf(" 5.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.050f));
printf(" 1.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.010f));
printf(" 0.1%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.001f));
printf("Minimum Δp: %6.3lf%%\n", 100.0*p_diff_values.front());
LOG("Maximum Δp: %6.3lf%%\n", 100.0*p_diff_values.back());
LOG("99.9%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.999f));
LOG("99.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.990f));
LOG("95.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.950f));
LOG("90.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.900f));
LOG("75.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.750f));
LOG("Median Δp: %6.3lf%%\n", 100.0*p_diff_median);
LOG("25.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.250f));
LOG("10.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.100f));
LOG(" 5.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.050f));
LOG(" 1.0%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.010f));
LOG(" 0.1%% Δp: %6.3lf%%\n", 100.0*percentile(p_diff_values, 0.001f));
LOG("Minimum Δp: %6.3lf%%\n", 100.0*p_diff_values.front());
auto p_diff_mse = mean_and_uncertainty(kld.sum_p_diff2, kld.sum_p_diff4, kld.count);
// printf("MSE Δp : %10.6lf ± %10.6lf\n", p_diff_mse.first, p_diff_mse.second);
// LOG("MSE Δp : %10.6lf ± %10.6lf\n", p_diff_mse.first, p_diff_mse.second);
const double p_diff_rms_val = sqrt(p_diff_mse.first);
const double p_diff_rms_unc = 0.5/p_diff_rms_val * p_diff_mse.second;
printf("RMS Δp : %6.3lf ± %5.3lf %%\n", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
LOG("RMS Δp : %6.3lf ± %5.3lf %%\n", 100.0*p_diff_rms_val, 100.0*p_diff_rms_unc);
const double same_top_p = 1.0*kld.n_same_top/kld.count;
printf("Same top p: %6.3lf ± %5.3lf %%\n", 100.0*same_top_p, 100.0*sqrt(same_top_p*(1.0 - same_top_p)/(kld.count - 1)));
LOG("Same top p: %6.3lf ± %5.3lf %%\n", 100.0*same_top_p, 100.0*sqrt(same_top_p*(1.0 - same_top_p)/(kld.count - 1)));
}
int main(int argc, char ** argv) {
@ -1967,15 +1962,16 @@ int main(int argc, char ** argv) {
params.n_ctx = 512;
params.logits_all = true;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_PERPLEXITY);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
return 1;
}
gpt_init();
const int32_t n_ctx = params.n_ctx;
if (n_ctx <= 0) {
fprintf(stderr, "%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
LOG_ERR("%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
return 1;
}
@ -2000,15 +1996,11 @@ int main(int argc, char ** argv) {
}
if (params.ppl_stride > 0) {
fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
LOG_INF("Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
params.n_ctx, params.n_ctx + params.ppl_stride/2);
params.n_ctx += params.ppl_stride/2;
}
print_build_info();
LOG_TEE("%s: seed = %u\n", __func__, params.sparams.seed);
llama_backend_init();
llama_numa_init(params.numa);
@ -2018,21 +2010,21 @@ int main(int argc, char ** argv) {
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, params.n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
}
struct results_perplexity results;
@ -2048,8 +2040,9 @@ int main(int argc, char ** argv) {
results = perplexity(ctx, params, n_ctx);
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
write_logfile(ctx, params, model, results);
llama_free(ctx);

2
examples/quantize/CMakeLists.txt
View File

@ -1,6 +1,6 @@
set(TARGET llama-quantize)
add_executable(${TARGET} quantize.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
target_include_directories(${TARGET} PRIVATE ../../common)
target_compile_features(${TARGET} PRIVATE cxx_std_11)

2
examples/quantize/README.md
View File

@ -54,6 +54,8 @@ As the models are currently fully loaded into memory, you will need adequate dis
Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format.
*(outdated)*
| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |

76
examples/retrieval/retrieval.cpp
View File

@ -1,13 +1,16 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <algorithm>
#include <fstream>
#include <iostream> // TODO: remove me
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
LOG("\n");
}
struct chunk {
@ -16,7 +19,7 @@ struct chunk {
// original file position
size_t filepos;
// original text data
std::string textdata = "";
std::string textdata;
// tokenized text data
std::vector<llama_token> tokens;
// embedding
@ -30,14 +33,14 @@ static std::vector<chunk> chunk_file(const std::string & filename, int chunk_siz
std::ifstream f(filename.c_str());
if (!f.is_open()) {
fprintf(stderr, "Error: could not open file %s\n", filename.c_str());
LOG_ERR("could not open file %s\n", filename.c_str());
return chunks;
}
chunk current_chunk;
char buffer[1024];
int64_t filepos = 0;
std::string current = "";
std::string current;
while (f.read(buffer, 1024)) {
current += std::string(buffer, f.gcount());
size_t pos;
@ -83,9 +86,9 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
LOG_ERR("%s : failed to decode\n", __func__);
}
for (int i = 0; i < batch.n_tokens; i++) {
@ -98,7 +101,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
if (embd == NULL) {
embd = llama_get_embeddings_ith(ctx, i);
if (embd == NULL) {
fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
LOG_ERR("%s: failed to get embeddings for token %d\n", __func__, i);
continue;
}
}
@ -111,29 +114,28 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_RETRIEVAL, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
return 1;
}
gpt_init();
// For BERT models, batch size must be equal to ubatch size
params.n_ubatch = params.n_batch;
params.embedding = true;
if (params.chunk_size <= 0) {
fprintf(stderr, "chunk_size must be positive\n");
LOG_ERR("chunk_size must be positive\n");
return 1;
}
if (params.context_files.empty()) {
fprintf(stderr, "context_files must be specified\n");
LOG_ERR("context_files must be specified\n");
return 1;
}
print_build_info();
printf("processing files:\n");
LOG_INF("processing files:\n");
for (auto & context_file : params.context_files) {
printf("%s\n", context_file.c_str());
LOG_INF("%s\n", context_file.c_str());
}
std::vector<chunk> chunks;
@ -141,7 +143,7 @@ int main(int argc, char ** argv) {
std::vector<chunk> file_chunk = chunk_file(context_file, params.chunk_size, params.chunk_separator);
chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
}
printf("Number of chunks: %ld\n", chunks.size());
LOG_INF("Number of chunks: %ld\n", chunks.size());
llama_backend_init();
llama_numa_init(params.numa);
@ -153,7 +155,7 @@ int main(int argc, char ** argv) {
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
LOG_ERR("%s: unable to load model\n", __func__);
return 1;
}
@ -162,19 +164,19 @@ int main(int argc, char ** argv) {
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
LOG_ERR("%s: pooling type NONE not supported\n", __func__);
return 1;
}
if (n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
__func__, n_ctx_train, n_ctx);
}
// print system information
{
fprintf(stderr, "\n");
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
LOG_INF("\n");
LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
}
// max batch size
@ -185,7 +187,7 @@ int main(int argc, char ** argv) {
for (auto & chunk : chunks) {
auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false);
if (inp.size() > n_batch) {
fprintf(stderr, "%s: error: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
LOG_ERR("%s: chunk size (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
__func__, (long long int) inp.size(), (long long int) n_batch);
return 1;
}
@ -199,12 +201,12 @@ int main(int argc, char ** argv) {
// tokenization stats
if (params.verbose_prompt) {
for (int i = 0; i < (int) chunks.size(); i++) {
fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
LOG_INF("%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
fprintf(stderr, "%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
LOG_INF("%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
}
fprintf(stderr, "\n\n");
LOG_INF("\n\n");
}
}
@ -256,7 +258,7 @@ int main(int argc, char ** argv) {
// start loop, receive query and return top k similar chunks based on cosine similarity
std::string query;
while (true) {
printf("Enter query: ");
LOG("Enter query: ");
std::getline(std::cin, query);
std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
@ -280,19 +282,19 @@ int main(int argc, char ** argv) {
return a.second > b.second;
});
printf("Top %d similar chunks:\n", params.sparams.top_k);
LOG("Top %d similar chunks:\n", params.sparams.top_k);
for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
printf("filename: %s\n", chunks[similarities[i].first].filename.c_str());
printf("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
printf("similarity: %f\n", similarities[i].second);
printf("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
printf("--------------------\n");
LOG("filename: %s\n", chunks[similarities[i].first].filename.c_str());
LOG("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
LOG("similarity: %f\n", similarities[i].second);
LOG("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
LOG("--------------------\n");
}
}
}
LOG_TEE("\n");
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_context_print(ctx);
// clean up
llama_batch_free(query_batch);

30
examples/rpc/README.md
View File

@ -10,20 +10,21 @@ This can be used for distributed LLM inference with `llama.cpp` in the following
```mermaid
flowchart TD
rpcb---|TCP|srva
rpcb---|TCP|srvb
rpcb-.-|TCP|srvn
rpcb<-->|TCP|srva
rpcb<-->|TCP|srvb
rpcb<-.->|TCP|srvn
subgraph hostn[Host N]
srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"]
end
subgraph hostb[Host B]
srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"]
end
subgraph hosta[Host A]
srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"]
end
subgraph host[Main Host]
ggml[llama.cpp]---rpcb[RPC backend]
local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli]
ggml[llama-cli]<-->rpcb[RPC backend]
end
style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
```
@ -62,17 +63,12 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
On the main host build `llama.cpp` only with `-DGGML_RPC=ON`:
```bash
mkdir build-rpc
cd build-rpc
cmake .. -DGGML_RPC=ON
cmake --build . --config Release
```
Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options.
Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`:
```bash
$ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
```
This way you can offload model layers to both local and remote devices.

10
examples/save-load-state/save-load-state.cpp
View File

@ -1,3 +1,4 @@
#include "arg.h"
#include "common.h"
#include "llama.h"
@ -10,8 +11,7 @@ int main(int argc, char ** argv) {
params.prompt = "The quick brown fox";
params.sparams.seed = 1234;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
return 1;
}
@ -74,8 +74,6 @@ int main(int argc, char ** argv) {
auto next_token = llama_sampler_sample(smpl, ctx, -1);
auto next_token_str = llama_token_to_piece(ctx, next_token);
llama_sampler_accept(smpl, next_token);
printf("%s", next_token_str.c_str());
result0 += next_token_str;
@ -132,8 +130,6 @@ int main(int argc, char ** argv) {
auto next_token = llama_sampler_sample(smpl2, ctx2, -1);
auto next_token_str = llama_token_to_piece(ctx2, next_token);
llama_sampler_accept(smpl2, next_token);
printf("%s", next_token_str.c_str());
result1 += next_token_str;
@ -222,8 +218,6 @@ int main(int argc, char ** argv) {
auto next_token = llama_sampler_sample(smpl3, ctx3, -1);
auto next_token_str = llama_token_to_piece(ctx3, next_token);
llama_sampler_accept(smpl3, next_token);
printf("%s", next_token_str.c_str());
result2 += next_token_str;

6
examples/server/CMakeLists.txt
View File

@ -1,5 +1,5 @@
set(TARGET llama-server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
@ -30,6 +30,7 @@ set(PUBLIC_ASSETS
system-prompts.js
prompt-formats.js
json-schema-to-grammar.mjs
loading.html
)
foreach(asset ${PUBLIC_ASSETS})
@ -45,9 +46,6 @@ endforeach()
add_executable(${TARGET} ${TARGET_SRCS})
install(TARGETS ${TARGET} RUNTIME)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
)
target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})

59
examples/server/README.md
View File

@ -23,36 +23,32 @@ The project is under active development, and we are [looking for feedback and co
| `--version` | show version and build info |
| `-v, --verbose` | print verbose information |
| `--verbosity N` | set specific verbosity level (default: 0) |
| `--verbose-prompt` | print a verbose prompt before generation (default: false) |
| `--no-display-prompt` | don't print prompt at generation (default: false) |
| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) |
| `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
| `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
| `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
| `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
| `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)<br/> |
| `--prio N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
| `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)<br/> |
| `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
| `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
| `-lcs, --lookup-cache-static FNAME` | path to static lookup cache to use for lookup decoding (not updated by generation) |
| `-lcd, --lookup-cache-dynamic FNAME` | path to dynamic lookup cache to use for lookup decoding (updated by generation) |
| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
| `--chunks N` | max number of chunks to process (default: -1, -1 = all) |
| `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
| `-p, --prompt PROMPT` | prompt to start generation with |
| `-f, --file FNAME` | a file containing the prompt (default: none) |
| `--in-file FNAME` | an input file (repeat to specify multiple files) |
| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
| `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
| `--no-escape` | do not process escape sequences |
| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typical_p;top_p;min_p;temperature) |
| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) |
| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
| `--penalize-nl` | penalize newline tokens (default: false) |
@ -92,13 +88,12 @@ The project is under active development, and we are [looking for feedback and co
| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
| `-ns, --sequences N` | number of sequences to decode (default: 1) |
| `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
| `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
| `--mlock` | force system to keep model in RAM rather than swapping or compressing |
| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) |
| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437 |
| `-ngl, --gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs |
| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 |
| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) |
@ -109,7 +104,7 @@ The project is under active development, and we are [looking for feedback and co
| `--control-vector FNAME` | add a control vector<br/>note: this argument can be repeated to add multiple control vectors |
| `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE<br/>note: this argument can be repeated to add multiple scaled control vectors |
| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
| `-a, --alias STRING` | set alias for model name (to be used by REST API)<br/>(env: LLAMA_ARG_MODEL) |
| `-a, --alias STRING` | set alias for model name (to be used by REST API) |
| `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
| `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
@ -123,10 +118,9 @@ The project is under active development, and we are [looking for feedback and co
| `--api-key-file FNAME` | path to file containing API keys (default: none) |
| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key |
| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate |
| `--timeout N` | server read/write timeout in seconds (default: 600) |
| `-to, --timeout N` | server read/write timeout in seconds (default: 600) |
| `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
| `--log-format {text, json}` | log output format: json or text (default: json) |
| `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
| `--no-slots` | disables slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
| `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
@ -412,9 +406,44 @@ Notice that each `probs` is an array of length `n_probs`.
*Options:*
`content`: Set the text to tokenize.
`content`: (Required) The text to tokenize.
`add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted. Default: `false`
`add_special`: (Optional) Boolean indicating if special tokens, i.e. `BOS`, should be inserted. Default: `false`
`with_pieces`: (Optional) Boolean indicating whether to return token pieces along with IDs. Default: `false`
**Response:**
Returns a JSON object with a `tokens` field containing the tokenization result. The `tokens` array contains either just token IDs or objects with `id` and `piece` fields, depending on the `with_pieces` parameter. The piece field is a string if the piece is valid unicode or a list of bytes otherwise.
If `with_pieces` is `false`:
```json
{
"tokens": [123, 456, 789]
}
```
If `with_pieces` is `true`:
```json
{
"tokens": [
{"id": 123, "piece": "Hello"},
{"id": 456, "piece": " world"},
{"id": 789, "piece": "!"}
]
}
```
With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
```json
{
"tokens": [
{"id": 198, "piece": [195]}, // hex C3
{"id": 164, "piece": [161]} // hex A1
]
}
```
### POST `/detokenize`: Convert tokens to text

1
examples/server/bench/README.md
View File

@ -40,7 +40,6 @@ server --host localhost --port 8080 \
--parallel 8 \
--batch-size 512 \
--ctx-size 4096 \
--log-format text \
-ngl 33
```

1
examples/server/bench/bench.py
View File

@ -272,7 +272,6 @@ def start_server_background(args):
server_args.append('--cont-batching')
server_args.append('--metrics')
server_args.append('--flash-attn')
server_args.extend(['--log-format', "text"])
args = [str(arg) for arg in [server_path, *server_args]]
print(f"bench: starting server with: {' '.join(args)}")
pkwargs = {

12
examples/server/public/loading.html Normal file
View File

@ -0,0 +1,12 @@
<!DOCTYPE html>
<html>
<head>
<meta http-equiv="refresh" content="5">
</head>
<body>
<div id="loading">
The model is loading. Please wait.<br/>
The user interface will appear soon.
</div>
</body>
</html>

580
examples/server/server.cpp
View File

File diff suppressed because it is too large Load Diff

1
examples/server/tests/.gitignore vendored Normal file
View File

@ -0,0 +1 @@
.venv

1
examples/server/tests/README.md
View File

@ -40,7 +40,6 @@ It's possible to override some scenario steps values with environment variables:
| `PORT` | `context.server_port` to set the listening port of the server during scenario, default: `8080` |
| `LLAMA_SERVER_BIN_PATH` | to change the server binary path, default: `../../../build/bin/llama-server` |
| `DEBUG` | "ON" to enable steps and server verbose mode `--verbose` |
| `SERVER_LOG_FORMAT_JSON` | if set switch server logs to json format |
| `N_GPU_LAYERS` | number of model layers to offload to VRAM `-ngl --n-gpu-layers` |
### Run @bug, @wip or @wrong_usage annotated scenario

8
examples/server/tests/features/server.feature
View File

@ -105,6 +105,14 @@ Feature: llama.cpp server
Given first token is removed
Then tokens can be detokenized
Scenario: Tokenize with pieces
When tokenizing with pieces:
"""
What is the capital of Germany?
媽
"""
Then tokens are given with pieces
Scenario: Models available
Given available models
Then 1 models are supported

33
examples/server/tests/features/steps/steps.py
View File

@ -1,3 +1,6 @@
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
import asyncio
import json
import os
@ -697,6 +700,32 @@ def step_tokenize_set_add_special(context):
context.tokenize_add_special = True
@step("tokenizing with pieces")
@async_run_until_complete
async def step_tokenize_with_pieces(context):
context.tokenized_text = context_text(context)
async with aiohttp.ClientSession() as session:
tokenize_args = {"content": context.tokenized_text, "with_pieces": True}
if getattr(context, "tokenize_add_special", None) is not None:
tokenize_args["add_special"] = context.tokenize_add_special
async with session.post(
f"{context.base_url}/tokenize", json=tokenize_args
) as response:
assert response.status == 200
tokenize_json = await response.json()
context.tokens_with_pieces = tokenize_json["tokens"]
@step("tokens are given with pieces")
@async_run_until_complete
async def step_tokenize_with_pieces(context):
# Verify that the response contains both token IDs and pieces
assert all(
"id" in token and "piece" in token for token in context.tokens_with_pieces
)
@step('tokenizing')
@async_run_until_complete
async def step_tokenize(context):
@ -991,6 +1020,8 @@ async def oai_chat_completions(user_prompt,
event_data = line.split(': ', 1)
assert event_data[0] == 'data', f'Bad event code received: ```{event_data}```'
chunk_raw = event_data[1]
if chunk_raw == '[DONE]':
break
chunk = json.loads(chunk_raw)
assert len(chunk['choices']) == 1, f"no choices provided, line ```{line}```"
@ -1341,8 +1372,6 @@ def start_server_background(context):
server_args.append('--verbose')
if context.lora_file:
server_args.extend(['--lora', context.lora_file])
if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
server_args.extend(['--log-format', "text"])
args = [str(arg) for arg in [context.server_path, *server_args]]
print(f"bench: starting server with: {' '.join(args)}")

145
examples/server/utils.hpp
View File

@ -1,7 +1,8 @@
#pragma once
#include "llama.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#ifndef NDEBUG
// crash the server in debug mode, otherwise send an http 500 error
@ -15,10 +16,10 @@
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
#include <random>
#include <sstream>
#include <string>
#include <vector>
#include <sstream>
#include <random>
#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
@ -35,32 +36,6 @@ enum error_type {
ERROR_TYPE_NOT_SUPPORTED, // custom error
};
extern bool server_verbose;
extern bool server_log_json;
#ifndef SERVER_VERBOSE
#define SERVER_VERBOSE 1
#endif
#if SERVER_VERBOSE != 1
#define LOG_VERBOSE(MSG, ...)
#else
#define LOG_VERBOSE(MSG, ...) \
do \
{ \
if (server_verbose) \
{ \
server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
} \
} while (0)
#endif
#define LOG_ERROR( MSG, ...) server_log("ERR", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
#define LOG_INFO( MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
static inline void server_log(const char * level, const char * function, int line, const char * message, const json & extra);
template <typename T>
static T json_value(const json & body, const std::string & key, const T & default_value) {
// Fallback null to default value
@ -68,9 +43,7 @@ static T json_value(const json & body, const std::string & key, const T & defaul
try {
return body.at(key);
} catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const &) {
std::stringstream ss;
ss << "Wrong type supplied for parameter '" << key << "'. Expected '" << json(default_value).type_name() << "', using default value.";
LOG_WARNING(ss.str().c_str(), body);
LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value\n", key.c_str(), json(default_value).type_name());
return default_value;
}
} else {
@ -78,48 +51,6 @@ static T json_value(const json & body, const std::string & key, const T & defaul
}
}
static inline void server_log(const char * level, const char * function, int line, const char * message, const json & extra) {
std::stringstream ss_tid;
ss_tid << std::this_thread::get_id();
json log = json{
{"tid", ss_tid.str()},
{"timestamp", time(nullptr)},
};
if (server_log_json) {
log.merge_patch({
{"level", level},
{"function", function},
{"line", line},
{"msg", message},
});
if (!extra.empty()) {
log.merge_patch(extra);
}
printf("%s\n", log.dump(-1, ' ', false, json::error_handler_t::replace).c_str());
} else {
char buf[1024];
snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
if (!extra.empty()) {
log.merge_patch(extra);
}
std::stringstream ss;
ss << buf << " |";
for (const auto & el : log.items())
{
const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
ss << " " << el.key() << "=" << value;
}
const std::string str = ss.str();
printf("%.*s\n", (int)str.size(), str.data());
}
fflush(stdout);
}
//
// chat template utils
//
@ -153,8 +84,9 @@ inline std::string format_chat(const struct llama_model * model, const std::stri
chat.push_back({role, content});
}
auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true);
LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
const auto formatted_chat = llama_chat_apply_template(model, tmpl, chat, true);
LOG_DBG("formatted_chat: '%s'\n", formatted_chat.c_str());
return formatted_chat;
}
@ -243,10 +175,7 @@ static std::string random_string() {
}
static std::string gen_chatcmplid() {
std::stringstream chatcmplid;
chatcmplid << "chatcmpl-" << random_string();
return chatcmplid.str();
return "chatcmpl-" + random_string();
}
//
@ -287,7 +216,7 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
return std::string::npos;
}
static bool json_is_array_of_numbers(json data) {
static bool json_is_array_of_numbers(const json & data) {
if (data.is_array()) {
for (const auto & e : data) {
if (!e.is_number()) {
@ -363,15 +292,13 @@ static json probs_vector_to_json(const llama_context * ctx, const std::vector<co
return out;
}
static bool server_sent_event(httplib::DataSink & sink, const char * event, json & data) {
static bool server_sent_event(httplib::DataSink & sink, const char * event, const json & data) {
const std::string str =
std::string(event) + ": " +
data.dump(-1, ' ', false, json::error_handler_t::replace) +
"\n\n";
"\n\n"; // note: these newlines are important (not sure why though, if you know, add a comment to explain)
LOG_VERBOSE("data stream", {
{ "to_send", str }
});
LOG_DBG("data stream, to_send: %s", str.c_str());
return sink.write(str.c_str(), str.size());
}
@ -425,7 +352,7 @@ static json oaicompat_completion_params_parse(
// Params supported by OAI but unsupported by llama.cpp
static const std::vector<std::string> unsupported_params { "tools", "tool_choice" };
for (auto & param : unsupported_params) {
for (const auto & param : unsupported_params) {
if (body.contains(param)) {
throw std::runtime_error("Unsupported param: " + param);
}
@ -444,7 +371,7 @@ static json oaicompat_completion_params_parse(
return llama_params;
}
static json format_final_response_oaicompat(const json & request, json result, const std::string & completion_id, bool streaming = false) {
static json format_final_response_oaicompat(const json & request, const json & result, const std::string & completion_id, bool streaming = false, bool verbose = false) {
bool stopped_word = result.count("stopped_word") != 0;
bool stopped_eos = json_value(result, "stopped_eos", false);
int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
@ -481,7 +408,8 @@ static json format_final_response_oaicompat(const json & request, json result, c
{"id", completion_id}
};
if (server_verbose) {
// extra fields for debugging purposes
if (verbose) {
res["__verbose"] = result;
}
@ -493,7 +421,7 @@ static json format_final_response_oaicompat(const json & request, json result, c
}
// return value is vector as there is one case where we might need to generate two responses
static std::vector<json> format_partial_response_oaicompat(json result, const std::string & completion_id) {
static std::vector<json> format_partial_response_oaicompat(const json & result, const std::string & completion_id) {
if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
return std::vector<json>({result});
}
@ -595,7 +523,7 @@ static std::vector<json> format_partial_response_oaicompat(json result, const st
static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
json data = json::array();
int i = 0;
for (auto & elem : embeddings) {
for (const auto & elem : embeddings) {
data.push_back(json{
{"embedding", json_value(elem, "embedding", json::array())},
{"index", i++},
@ -616,7 +544,40 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
return res;
}
static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
static bool is_valid_utf8(const std::string & str) {
const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
const unsigned char* end = bytes + str.length();
while (bytes < end) {
if (*bytes <= 0x7F) {
// 1-byte sequence (0xxxxxxx)
bytes++;
} else if ((*bytes & 0xE0) == 0xC0) {
// 2-byte sequence (110xxxxx 10xxxxxx)
if (end - bytes < 2 || (bytes[1] & 0xC0) != 0x80)
return false;
bytes += 2;
} else if ((*bytes & 0xF0) == 0xE0) {
// 3-byte sequence (1110xxxx 10xxxxxx 10xxxxxx)
if (end - bytes < 3 || (bytes[1] & 0xC0) != 0x80 || (bytes[2] & 0xC0) != 0x80)
return false;
bytes += 3;
} else if ((*bytes & 0xF8) == 0xF0) {
// 4-byte sequence (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
if (end - bytes < 4 || (bytes[1] & 0xC0) != 0x80 ||
(bytes[2] & 0xC0) != 0x80 || (bytes[3] & 0xC0) != 0x80)
return false;
bytes += 4;
} else {
// Invalid UTF-8 lead byte
return false;
}
}
return true;
}
static json format_tokenizer_response(const json & tokens) {
return json {
{"tokens", tokens}
};

53
examples/simple/simple.cpp
View File

@ -1,15 +1,14 @@
#include "arg.h"
#include "common.h"
#include "log.h"
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <string>
#include <vector>
static void print_usage(int, char ** argv) {
LOG_TEE("\nexample usage:\n");
LOG_TEE("\n %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
LOG_TEE("\n");
LOG("\nexample usage:\n");
LOG("\n %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
LOG("\n");
}
int main(int argc, char ** argv) {
@ -18,11 +17,12 @@ int main(int argc, char ** argv) {
params.prompt = "Hello my name is";
params.n_predict = 32;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_COMMON, print_usage);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
return 1;
}
gpt_init();
// total length of the sequence including the prompt
const int n_predict = params.n_predict;
@ -69,25 +69,24 @@ int main(int argc, char ** argv) {
const int n_ctx = llama_n_ctx(ctx);
const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size());
LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
LOG("\n");
LOG_INF("%s: n_predict = %d, n_ctx = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, n_kv_req);
// make sure the KV cache is big enough to hold all the prompt and generated tokens
if (n_kv_req > n_ctx) {
LOG_TEE("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
LOG_TEE("%s: either reduce n_predict or increase n_ctx\n", __func__);
LOG_ERR("%s: error: n_kv_req > n_ctx, the required KV cache size is not big enough\n", __func__);
LOG_ERR("%s: either reduce n_predict or increase n_ctx\n", __func__);
return 1;
}
// print the prompt token-by-token
fprintf(stderr, "\n");
LOG("\n");
for (auto id : tokens_list) {
fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
LOG("%s", llama_token_to_piece(ctx, id).c_str());
}
fflush(stderr);
// create a llama_batch with size 512
// we use this object to submit token data for decoding
@ -102,7 +101,7 @@ int main(int argc, char ** argv) {
batch.logits[batch.n_tokens - 1] = true;
if (llama_decode(ctx, batch) != 0) {
LOG_TEE("%s: llama_decode() failed\n", __func__);
LOG("%s: llama_decode() failed\n", __func__);
return 1;
}
@ -116,18 +115,16 @@ int main(int argc, char ** argv) {
while (n_cur <= n_predict) {
// sample the next token
{
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
llama_sampler_accept(smpl, new_token_id);
const llama_token new_token_id = llama_sampler_sample(smpl, ctx, -1);
// is it an end of generation?
if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
LOG_TEE("\n");
LOG("\n");
break;
}
LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
fflush(stdout);
// prepare the next batch
@ -143,23 +140,23 @@ int main(int argc, char ** argv) {
// evaluate the current batch with the transformer model
if (llama_decode(ctx, batch)) {
fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
return 1;
}
}
LOG_TEE("\n");
LOG("\n");
const auto t_main_end = ggml_time_us();
LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
__func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
LOG_TEE("\n");
llama_perf_print(smpl, LLAMA_PERF_TYPE_SAMPLER_CHAIN);
llama_perf_print(ctx, LLAMA_PERF_TYPE_CONTEXT);
LOG("\n");
llama_perf_sampler_print(smpl);
llama_perf_context_print(ctx);
fprintf(stderr, "\n");
LOG("\n");
llama_batch_free(batch);
llama_sampler_free(smpl);

114
examples/speculative/speculative.cpp
View File

@ -1,11 +1,16 @@
#include "arg.h"
#include "common.h"
#include "sampling.h"
#include "log.h"
#include "llama.h"
#include <cmath>
#include <algorithm>
#include <cstdio>
#include <cstring>
#include <random>
#include <set>
#include <string>
#include <vector>
#include <set>
#define SPEC_VOCAB_MAX_SIZE_DIFFERENCE 100
#define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@ -27,13 +32,14 @@ struct seq_draft {
int main(int argc, char ** argv) {
gpt_params params;
auto options = gpt_params_parser_init(params, LLAMA_EXAMPLE_SPECULATIVE);
if (!gpt_params_parse(argc, argv, params, options)) {
if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
return 1;
}
gpt_init();
if (params.model_draft.empty()) {
fprintf(stderr, "%s: error: --model-draft is required\n", __func__);
LOG_ERR("%s: --model-draft is required\n", __func__);
return 1;
}
@ -46,12 +52,6 @@ int main(int argc, char ** argv) {
std::default_random_engine rng(params.sparams.seed);
std::uniform_real_distribution<> u_dist;
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("speculative", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
#endif // LOG_DISABLE_LOGS
// init llama.cpp
llama_backend_init();
llama_numa_init(params.numa);
@ -80,14 +80,14 @@ int main(int argc, char ** argv) {
ctx_dft = llama_init_dft.context;
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
LOG("vocab_type tgt: %d\n", vocab_type_tgt);
LOG_DBG("vocab_type tgt: %d\n", vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
LOG("vocab_type dft: %d\n", vocab_type_dft);
LOG_DBG("vocab_type dft: %d\n", vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
fprintf(stderr, "%s: error: draft model vocab type must match target model to use speculation but ", __func__);
fprintf(stderr, "vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
LOG_ERR("%s: draft model vocab type must match target model to use speculation but ", __func__);
LOG_ERR("vocab_type_dft = %d while vocab_type_tgt = %d\n", vocab_type_dft, vocab_type_tgt);
return 1;
}
@ -97,7 +97,7 @@ int main(int argc, char ** argv) {
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)
) {
fprintf(stderr, "%s: error: draft model special tokens must match target model to use speculation\n", __func__);
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
return 1;
}
@ -109,8 +109,8 @@ int main(int argc, char ** argv) {
: n_vocab_dft - n_vocab_tgt;
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
fprintf(stderr, "%s: error: draft model vocab must closely match target model to use speculation but ", __func__);
fprintf(stderr, "target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but ", __func__);
LOG_ERR("target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return 1;
}
@ -119,8 +119,8 @@ int main(int argc, char ** argv) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
fprintf(stderr, "%s: error: draft model vocab must match target model to use speculation but ", __func__);
fprintf(stderr, "token %d content differs - target '%s', draft '%s'\n", i,
LOG_ERR("%s: draft model vocab must match target model to use speculation but ", __func__);
LOG_ERR("token %d content differs - target '%s', draft '%s'\n", i,
llama_token_to_piece(ctx_tgt, i).c_str(),
llama_token_to_piece(ctx_dft, i).c_str());
return 1;
@ -137,18 +137,16 @@ int main(int argc, char ** argv) {
const int max_tokens_list_size = max_context_size - 4;
if ((int) inp.size() > max_tokens_list_size) {
fprintf(stderr, "%s: error: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
LOG_ERR("%s: prompt too long (%d tokens, max %d)\n", __func__, (int) inp.size(), max_tokens_list_size);
return 1;
}
fprintf(stderr, "\n\n");
LOG("\n\n");
for (auto id : inp) {
fprintf(stderr, "%s", llama_token_to_piece(ctx_tgt, id).c_str());
LOG("%s", llama_token_to_piece(ctx_tgt, id).c_str());
}
fflush(stderr);
const int n_input = inp.size();
const auto t_enc_start = ggml_time_us();
@ -210,7 +208,7 @@ int main(int argc, char ** argv) {
active_seqs.insert(s);
const auto & tokens = drafts[s].tokens;
LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
LOG_DBG("draft %d: %s\n", s, string_from(ctx_dft, tokens).c_str());
}
int i_dft = 0;
@ -253,7 +251,7 @@ int main(int argc, char ** argv) {
continue;
}
LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
LOG_DBG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
float r = u_dist(rng);
llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), LLAMA_TOKEN_NULL, true };
@ -271,7 +269,7 @@ int main(int argc, char ** argv) {
break;
}
}
LOG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
LOG_DBG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
if (r <= p_tgt / p_dft) {
s_keep = s;
accept = true;
@ -279,10 +277,10 @@ int main(int argc, char ** argv) {
token_str = llama_token_to_piece(ctx_tgt, token_id);
gpt_sampler_accept(smpl, token_id, true);
LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
LOG_DBG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
break;
} else {
LOG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
LOG_DBG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
drafts[s].active = false;
// calculate residual probability
@ -337,7 +335,7 @@ int main(int argc, char ** argv) {
if (!accept) {
// all drafted tokens were rejected
// sample from the target model
LOG("all drafted tokens were rejected, sampling from residual distribution\n");
LOG_DBG("all drafted tokens were rejected, sampling from residual distribution\n");
std::vector<float> probs(dist_tgt.size);
for (size_t i = 0; i < dist_tgt.size; ++i) {
probs[i] = dist_tgt.data[i].p;
@ -355,13 +353,11 @@ int main(int argc, char ** argv) {
// greedy verification
// sample from the target model
LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
LOG_DBG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
token_id = gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
gpt_sampler_accept(smpl, token_id, true);
//LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, smpl->prev).c_str());
token_str = llama_token_to_piece(ctx_tgt, token_id);
for (int s = 0; s < n_seq_dft; ++s) {
@ -370,7 +366,7 @@ int main(int argc, char ** argv) {
}
if (i_dft < (int) drafts[s].tokens.size() && token_id == drafts[s].tokens[i_dft]) {
LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
LOG_DBG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
s_keep = s;
accept = true;
@ -392,26 +388,24 @@ int main(int argc, char ** argv) {
++i_dft;
if (params.use_color) {
// Color token according to its origin sequence
printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
LOG("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
} else {
printf("%s", token_str.c_str());
LOG("%s", token_str.c_str());
}
fflush(stdout);
continue;
} else {
printf("%s", token_str.c_str());
fflush(stdout);
LOG("%s", token_str.c_str());
break;
}
}
}
{
LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
LOG_DBG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
// TODO: simplify
{
LOG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
llama_kv_cache_seq_keep(ctx_dft, s_keep);
llama_kv_cache_seq_cp (ctx_dft, s_keep, 0, -1, -1);
@ -438,7 +432,7 @@ int main(int argc, char ** argv) {
llama_batch_add (batch_dft, token_id, n_past_dft, { 0 }, true);
llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
// LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
// LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
llama_decode(ctx_dft, batch_dft);
++n_past_dft;
@ -485,7 +479,7 @@ int main(int argc, char ** argv) {
const auto * cur_p = gpt_sampler_get_candidates(drafts[s].smpl);
for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
LOG_DBG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
k, s, i, cur_p->data[k].id, cur_p->data[k].p, llama_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
}
@ -494,7 +488,7 @@ int main(int argc, char ** argv) {
// attempt to split the branch if the probability is high enough
for (int f = 1; f < 8; ++f) {
if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
llama_kv_cache_seq_rm(ctx_dft, n_seq_cur, -1, -1);
llama_kv_cache_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
@ -583,7 +577,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_seq_cp(ctx_tgt, 0, s, -1, -1);
}
// LOG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
// LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());
llama_decode(ctx_tgt, batch_tgt);
++n_past_tgt;
}
@ -601,23 +595,25 @@ int main(int argc, char ** argv) {
auto t_dec_end = ggml_time_us();
LOG_TEE("\n\n");
LOG("\n\n");
LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_INF("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input, (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
LOG_INF("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict / ((t_dec_end - t_dec_start) / 1e6f));
LOG_TEE("\n");
LOG_TEE("n_draft = %d\n", n_draft);
LOG_TEE("n_predict = %d\n", n_predict);
LOG_TEE("n_drafted = %d\n", n_drafted);
LOG_TEE("n_accept = %d\n", n_accept);
LOG_TEE("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_INF("\n");
LOG_INF("n_draft = %d\n", n_draft);
LOG_INF("n_predict = %d\n", n_predict);
LOG_INF("n_drafted = %d\n", n_drafted);
LOG_INF("n_accept = %d\n", n_accept);
LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
LOG_TEE("\ndraft:\n\n");
LOG_INF("\n");
LOG_INF("draft:\n\n");
// TODO: print sampling/grammar timings for all drafts
llama_perf_print(ctx_dft, LLAMA_PERF_TYPE_CONTEXT);
llama_perf_context_print(ctx_dft);
LOG_TEE("\ntarget:\n\n");
LOG_INF("\n");
LOG_INF("target:\n\n");
gpt_perf_print(ctx_tgt, smpl);
gpt_sampler_free(smpl);
@ -636,7 +632,7 @@ int main(int argc, char ** argv) {
llama_backend_free();
fprintf(stderr, "\n\n");
LOG("\n\n");
return 0;
}

28
examples/sycl/run-llama2.sh
View File

@ -4,33 +4,23 @@
# Copyright (C) 2024 Intel Corporation
# SPDX-License-Identifier: MIT
INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
source /opt/intel/oneapi/setvars.sh
if [ $# -gt 0 ]; then
GGML_SYCL_DEVICE=$1
GGML_SYCL_SINGLE_GPU=1
else
GGML_SYCL_DEVICE=0
GGML_SYCL_SINGLE_GPU=0
fi
#export GGML_SYCL_DEBUG=1
#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
MODEL_FILE=llama-2-7b.Q4_0.gguf
NGL=33
if [ $# -gt 0 ]; then
GGML_SYCL_DEVICE=$1
echo "use $GGML_SYCL_DEVICE as main GPU"
#use signle GPU only
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -mg $GGML_SYCL_DEVICE -sm none
else
#use multiple GPUs with same max compute units
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0
fi
#use main GPU only
#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
#use multiple GPUs with same max compute units
#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0

50
examples/tokenize/tokenize.cpp
View File

@ -1,11 +1,13 @@
#include "common.h"
//#include "log.h" // TODO: start using log.h
#include "llama.h"
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <string>
#include <vector>
#include <iostream> // TODO: remove me
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
@ -13,25 +15,25 @@
#include <shellapi.h> // For CommandLineToArgvW
#endif
static void print_usage_information(const char * argv0, FILE * stream) {
fprintf(stream, "usage: %s [options]\n\n", argv0);
fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n");
fprintf(stream, "and prints the resulting tokens to standard output.\n\n");
fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n");
fprintf(stream, "to control the behavior of the tokenizer.\n\n");
fprintf(stream, " The possible options are:\n");
fprintf(stream, "\n");
fprintf(stream, " -h, --help print this help and exit\n");
fprintf(stream, " -m MODEL_PATH, --model MODEL_PATH path to model.\n");
fprintf(stream, " --ids if given, only print numerical token IDs, and not token strings.\n");
fprintf(stream, " The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
fprintf(stream, " -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
fprintf(stream, " -p PROMPT, --prompt PROMPT read prompt from the argument.\n");
fprintf(stream, " --stdin read prompt from standard input.\n");
fprintf(stream, " --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
fprintf(stream, " --no-parse-special do not parse control tokens.\n");
fprintf(stream, " --log-disable disable logs. Makes stderr quiet when loading the model.\n");
fprintf(stream, " --show-count print the total number of tokens.\n");
static void print_usage_information(const char * argv0) {
printf("usage: %s [options]\n\n", argv0);
printf("The tokenize program tokenizes a prompt using a given model,\n");
printf("and prints the resulting tokens to standard output.\n\n");
printf("It needs a model file, a prompt, and optionally other flags\n");
printf("to control the behavior of the tokenizer.\n\n");
printf(" The possible options are:\n");
printf("\n");
printf(" -h, --help print this help and exit\n");
printf(" -m MODEL_PATH, --model MODEL_PATH path to model.\n");
printf(" --ids if given, only print numerical token IDs, and not token strings.\n");
printf(" The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
printf(" -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
printf(" -p PROMPT, --prompt PROMPT read prompt from the argument.\n");
printf(" --stdin read prompt from standard input.\n");
printf(" --no-bos do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
printf(" --no-parse-special do not parse control tokens.\n");
printf(" --log-disable disable logs. Makes stderr quiet when loading the model.\n");
printf(" --show-count print the total number of tokens.\n");
}
static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
@ -185,7 +187,7 @@ int main(int raw_argc, char ** raw_argv) {
const int argc = argv.size();
if (argc <= 1) {
print_usage_information(argv[0].c_str(), stderr);
print_usage_information(argv[0].c_str());
return 1;
}
@ -214,7 +216,7 @@ int main(int raw_argc, char ** raw_argv) {
for (; iarg < argc; ++iarg) {
std::string arg{argv[iarg]};
if (arg == "-h" || arg == "--help") {
print_usage_information(argv[0].c_str(), stdout);
print_usage_information(argv[0].c_str());
return 0;
}
else if (arg == "--ids") {
@ -323,10 +325,6 @@ int main(int raw_argc, char ** raw_argv) {
// Start actually doing the tokenizing stuff.
//////
#ifdef LOG_DISABLE_LOGS
disable_logging = true;
#endif
if (disable_logging) {
llama_log_set(llama_log_callback_null, NULL);
}

20
flake.lock
View File

@ -5,11 +5,11 @@
"nixpkgs-lib": "nixpkgs-lib"
},
"locked": {
"lastModified": 1725024810,
"narHash": "sha256-ODYRm8zHfLTH3soTFWE452ydPYz2iTvr9T8ftDMUQ3E=",
"lastModified": 1726153070,
"narHash": "sha256-HO4zgY0ekfwO5bX0QH/3kJ/h4KvUDFZg8YpkNwIbg1U=",
"owner": "hercules-ci",
"repo": "flake-parts",
"rev": "af510d4a62d071ea13925ce41c95e3dec816c01d",
"rev": "bcef6817a8b2aa20a5a6dbb19b43e63c5bf8619a",
"type": "github"
},
"original": {
@ -20,11 +20,11 @@
},
"nixpkgs": {
"locked": {
"lastModified": 1724819573,
"narHash": "sha256-GnR7/ibgIH1vhoy8cYdmXE6iyZqKqFxQSVkFgosBh6w=",
"lastModified": 1726062873,
"narHash": "sha256-IiA3jfbR7K/B5+9byVi9BZGWTD4VSbWe8VLpp9B/iYk=",
"owner": "NixOS",
"repo": "nixpkgs",
"rev": "71e91c409d1e654808b2621f28a327acfdad8dc2",
"rev": "4f807e8940284ad7925ebd0a0993d2a1791acb2f",
"type": "github"
},
"original": {
@ -36,14 +36,14 @@
},
"nixpkgs-lib": {
"locked": {
"lastModified": 1722555339,
"narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
"lastModified": 1725233747,
"narHash": "sha256-Ss8QWLXdr2JCBPcYChJhz4xJm+h/xjl4G0c0XlP6a74=",
"type": "tarball",
"url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
"url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
},
"original": {
"type": "tarball",
"url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
"url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
}
},
"root": {

13
ggml/CMakeLists.txt
View File

@ -56,6 +56,15 @@ else()
set(GGML_NATIVE_DEFAULT ON)
endif()
# defaults
if (NOT GGML_LLAMAFILE_DEFAULT)
set(GGML_LLAMAFILE_DEFAULT OFF)
endif()
if (NOT GGML_CUDA_GRAPHS_DEFAULT)
set(GGML_CUDA_GRAPHS_DEFAULT OFF)
endif()
# general
option(GGML_STATIC "ggml: static link libraries" OFF)
option(GGML_NATIVE "ggml: enable -march=native flag" ${GGML_NATIVE_DEFAULT})
@ -110,7 +119,7 @@ option(GGML_ACCELERATE "ggml: enable Accelerate framework"
option(GGML_BLAS "ggml: use BLAS" ${GGML_BLAS_DEFAULT})
set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
"ggml: BLAS library vendor")
option(GGML_LLAMAFILE "ggml: use LLAMAFILE" OFF)
option(GGML_LLAMAFILE "ggml: use LLAMAFILE" ${GGML_LLAMAFILE_DEFAULT})
option(GGML_CUDA "ggml: use CUDA" OFF)
option(GGML_MUSA "ggml: use MUSA" OFF)
@ -127,7 +136,7 @@ set (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
option(GGML_CUDA_NO_PEER_COPY "ggml: do not use peer to peer copies" OFF)
option(GGML_CUDA_NO_VMM "ggml: do not try to use CUDA VMM" OFF)
option(GGML_CUDA_FA_ALL_QUANTS "ggml: compile all quants for FlashAttention" OFF)
option(GGML_CUDA_USE_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" OFF)
option(GGML_CUDA_GRAPHS "ggml: use CUDA graphs (llama.cpp only)" ${GGML_CUDA_GRAPHS_DEFAULT})
option(GGML_HIPBLAS "ggml: use hipBLAS" OFF)
option(GGML_HIP_UMA "ggml: use HIP unified memory architecture" OFF)

7
ggml/include/ggml-cann.h
View File

@ -80,6 +80,13 @@ ggml_backend_cann_buffer_type(int32_t device);
*/
GGML_API GGML_CALL int32_t ggml_backend_cann_get_device_count(void);
/**
* @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
*
* @return A pointer to the host buffer type interface.
*/
GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
/**
* @brief Retrieves the description of a specific CANN device.
*

65
ggml/include/ggml.h
View File

@ -358,6 +358,7 @@ extern "C" {
struct ggml_object;
struct ggml_context;
struct ggml_cgraph;
// NOTE: always add types at the end of the enum to keep backward compatibility
enum ggml_type {
@ -563,10 +564,11 @@ extern "C" {
};
enum ggml_log_level {
GGML_LOG_LEVEL_ERROR = 2,
GGML_LOG_LEVEL_WARN = 3,
GGML_LOG_LEVEL_INFO = 4,
GGML_LOG_LEVEL_DEBUG = 5
GGML_LOG_LEVEL_NONE = 0,
GGML_LOG_LEVEL_INFO = 1,
GGML_LOG_LEVEL_WARN = 2,
GGML_LOG_LEVEL_ERROR = 3,
GGML_LOG_LEVEL_DEBUG = 4,
};
enum ggml_tensor_flag {
@ -575,20 +577,6 @@ extern "C" {
GGML_TENSOR_FLAG_PARAM = 4,
};
// ggml object
struct ggml_object {
size_t offs;
size_t size;
struct ggml_object * next;
enum ggml_object_type type;
char padding[4];
};
static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
// n-dimensional tensor
struct ggml_tensor {
enum ggml_type type;
@ -671,35 +659,6 @@ extern "C" {
void * abort_callback_data;
};
enum ggml_cgraph_eval_order {
GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
GGML_CGRAPH_EVAL_ORDER_COUNT
};
typedef uint32_t ggml_bitset_t;
struct ggml_hash_set {
size_t size;
ggml_bitset_t * used; // whether or not the keys are in use i.e. set
struct ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if ggml_bitset_get(used, i)
};
// computation graph
struct ggml_cgraph {
int size;
int n_nodes;
int n_leafs;
struct ggml_tensor ** nodes;
struct ggml_tensor ** grads;
struct ggml_tensor ** leafs;
struct ggml_hash_set visited_hash_set;
enum ggml_cgraph_eval_order order;
};
// scratch buffer
struct ggml_scratch {
size_t offs;
@ -2017,8 +1976,6 @@ extern "C" {
typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
#define GGML_N_TASKS_MAX -1
GGML_API struct ggml_tensor * ggml_map_custom1(
struct ggml_context * ctx,
struct ggml_tensor * a,
@ -2088,7 +2045,6 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * tensor);
GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
@ -2096,11 +2052,17 @@ extern "C" {
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
GGML_API struct ggml_cgraph ggml_graph_view (struct ggml_cgraph * cgraph, int i0, int i1);
GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // zero grads
GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph);
GGML_API int ggml_graph_size (struct ggml_cgraph * cgraph);
GGML_API struct ggml_tensor * ggml_graph_node (struct ggml_cgraph * cgraph, int i); // if i < 0, returns nodes[n_nodes + i]
GGML_API struct ggml_tensor ** ggml_graph_nodes (struct ggml_cgraph * cgraph);
GGML_API int ggml_graph_n_nodes(struct ggml_cgraph * cgraph);
GGML_API void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API size_t ggml_graph_overhead(void);
GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
@ -2509,6 +2471,7 @@ extern "C" {
GGML_API int ggml_cpu_has_gpublas (void);
GGML_API int ggml_cpu_has_sse3 (void);
GGML_API int ggml_cpu_has_ssse3 (void);
GGML_API int ggml_cpu_has_riscv_v (void);
GGML_API int ggml_cpu_has_sycl (void);
GGML_API int ggml_cpu_has_rpc (void);
GGML_API int ggml_cpu_has_vsx (void);

80
ggml/src/CMakeLists.txt
View File

@ -26,6 +26,9 @@ if (NOT MSVC)
endif()
endif()
unset(GGML_EXTRA_LIBS_PRIVATE)
unset(GGML_EXTRA_LIBS_PUBLIC)
if (APPLE AND GGML_ACCELERATE)
find_library(ACCELERATE_FRAMEWORK Accelerate)
if (ACCELERATE_FRAMEWORK)
@ -35,7 +38,7 @@ if (APPLE AND GGML_ACCELERATE)
add_compile_definitions(ACCELERATE_NEW_LAPACK)
add_compile_definitions(ACCELERATE_LAPACK_ILP64)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ${ACCELERATE_FRAMEWORK})
list(APPEND GGML_EXTRA_LIBS_PRIVATE ${ACCELERATE_FRAMEWORK})
else()
message(WARNING "Accelerate framework not found")
endif()
@ -87,7 +90,7 @@ if (GGML_METAL)
COMMENT "Generate assembly for embedded Metal library"
)
set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
list(APPEND GGML_SOURCES_METAL ${METALLIB_EMBED_ASM})
else()
if (GGML_METAL_SHADER_DEBUG)
# custom command to do the following:
@ -132,7 +135,7 @@ if (GGML_METAL)
)
endif() # GGML_METAL_EMBED_LIBRARY
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS}
list(APPEND GGML_EXTRA_LIBS_PRIVATE
${FOUNDATION_LIBRARY}
${METAL_FRAMEWORK}
${METALKIT_FRAMEWORK}
@ -157,11 +160,11 @@ if (GGML_OPENMP)
add_compile_definitions(GGML_USE_OPENMP)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
list(APPEND GGML_EXTRA_LIBS_PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
if (GGML_MUSA)
set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} "/usr/lib/llvm-10/include/openmp")
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} "/usr/lib/llvm-10/lib/libomp.so")
list(APPEND GGML_EXTRA_INCLUDES "/usr/lib/llvm-10/include/openmp")
list(APPEND GGML_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-10/lib/libomp.so")
endif()
else()
message(WARNING "OpenMP not found")
@ -244,8 +247,8 @@ if (GGML_BLAS)
set(GGML_HEADERS_BLAS ../include/ggml-blas.h)
set(GGML_SOURCES_BLAS ggml-blas.cpp)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ${BLAS_LIBRARIES})
set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
list(APPEND GGML_EXTRA_LIBS_PRIVATE ${BLAS_LIBRARIES})
list(APPEND GGML_EXTRA_INCLUDES ${BLAS_INCLUDE_DIRS})
else()
message(WARNING "BLAS not found, please refer to "
"https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@ -326,7 +329,7 @@ if (GGML_CUDA)
add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
if (GGML_CUDA_USE_GRAPHS)
if (GGML_CUDA_GRAPHS)
add_compile_definitions(GGML_CUDA_USE_GRAPHS)
endif()
@ -368,19 +371,19 @@ if (GGML_CUDA)
if (GGML_STATIC)
if (WIN32)
# As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
else ()
if (GGML_MUSA)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart_static MUSA::mublas_static)
list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart_static MUSA::mublas_static)
else()
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
endif()
endif()
else()
if (GGML_MUSA)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musart MUSA::mublas)
list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart MUSA::mublas)
else()
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif()
endif()
@ -388,9 +391,9 @@ if (GGML_CUDA)
# No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
else()
if (GGML_MUSA)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
else()
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
endif()
endif()
else()
@ -495,7 +498,7 @@ if (GGML_HIPBLAS)
if (CXX_IS_HIPCC)
set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} hip::device)
list(APPEND GGML_EXTRA_LIBS_PRIVATE hip::device)
else()
set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
endif()
@ -504,7 +507,7 @@ if (GGML_HIPBLAS)
message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
endif()
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
list(APPEND GGML_EXTRA_LIBS_PUBLIC hip::host roc::rocblas roc::hipblas)
endif()
if (GGML_SYCL)
@ -513,6 +516,7 @@ if (GGML_SYCL)
endif()
check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
if (DEFINED ENV{ONEAPI_ROOT})
message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
elseif(SUPPORTS_SYCL)
@ -551,26 +555,29 @@ if (GGML_SYCL)
find_package(DNNL)
message("-- DNNL found:" ${DNNL_FOUND})
if (GGML_SYCL_TARGET STREQUAL "INTEL")
add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
else()
add_compile_definitions(GGML_SYCL_DNNL=0)
endif()
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND GGML_EXTRA_LIBS_PRIVATE DNNL::dnnl)
endif()
if (WIN32)
find_package(IntelSYCL REQUIRED)
find_package(MKL REQUIRED)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
list(APPEND GGML_EXTRA_LIBS_PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
else()
if (GGML_SYCL_TARGET STREQUAL "INTEL")
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} -fsycl pthread m dl onemkl)
list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl)
endif()
endif()
if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
list(APPEND GGML_EXTRA_LIBS DNNL::dnnl)
endif()
endif()
if (GGML_RPC)
@ -579,7 +586,7 @@ if (GGML_RPC)
list(APPEND GGML_CDEF_PUBLIC GGML_USE_RPC)
if (WIN32)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ws2_32)
list(APPEND GGML_EXTRA_LIBS_PRIVATE ws2_32)
endif()
set(GGML_HEADERS_RPC ../include/ggml-rpc.h)
@ -657,8 +664,8 @@ if (GGML_VULKAN)
set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header})
set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source})
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} Vulkan::Vulkan)
set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
list(APPEND GGML_EXTRA_LIBS_PRIVATE Vulkan::Vulkan)
list(APPEND GGML_EXTRA_INCLUDES ${CMAKE_CURRENT_BINARY_DIR})
else()
message(WARNING "Vulkan not found")
endif()
@ -817,8 +824,8 @@ if (GGML_KOMPUTE)
list(APPEND GGML_CDEF_PUBLIC GGML_USE_KOMPUTE)
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} kompute)
set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CMAKE_CURRENT_BINARY_DIR})
list(APPEND GGML_EXTRA_LIBS_PRIVATE kompute)
list(APPEND GGML_EXTRA_INCLUDES ${CMAKE_CURRENT_BINARY_DIR})
else()
message(WARNING "Kompute not found")
endif()
@ -883,9 +890,10 @@ if (GGML_CANN)
message(STATUS "CANN: CANN_INCLUDE_DIRS = ${CANN_INCLUDE_DIRS}")
message(STATUS "CANN: CANN_LIBRARIES = ${CANN_LIBRARIES}")
set(GGML_EXTRA_LIBS ${GGML_EXTRA_LIBS} ${CANN_LIBRARIES} )
set(GGML_EXTRA_INCLUDES ${GGML_EXTRA_INCLUDES} ${CANN_INCLUDE_DIRS})
set(GGML_EXTRA_LIBDIRS ${GGML_EXTRA_LIBDIRS} ${CANN_INSTALL_DIR}/lib64)
list(APPEND GGML_EXTRA_LIBS_PRIVATE ${CANN_LIBRARIES} )
list(APPEND GGML_EXTRA_INCLUDES ${CANN_INCLUDE_DIRS})
list(APPEND GGML_EXTRA_LIBDIRS ${CANN_INSTALL_DIR}/lib64)
list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
endif()
else()
@ -1328,15 +1336,19 @@ target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
target_link_directories (ggml PRIVATE ${GGML_EXTRA_LIBDIRS})
target_compile_features (ggml PRIVATE c_std_11) # don't bump
target_link_libraries(ggml PRIVATE Threads::Threads ${GGML_EXTRA_LIBS})
list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
find_library(MATH_LIBRARY m)
if (MATH_LIBRARY)
if (NOT WIN32 OR NOT GGML_SYCL)
target_link_libraries(ggml PRIVATE ${MATH_LIBRARY})
list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
endif()
endif()
list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PRIVATE)
list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PUBLIC)
target_link_libraries(ggml PRIVATE ${GGML_EXTRA_LIBS_PRIVATE} PUBLIC ${GGML_EXTRA_LIBS_PUBLIC})
if (BUILD_SHARED_LIBS)
set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(ggml PRIVATE GGML_SHARED GGML_BUILD)

1
ggml/src/ggml-aarch64.c
View File

@ -4,6 +4,7 @@
#include "ggml-quants.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include <math.h>
#include <string.h>

1
ggml/src/ggml-blas.cpp
View File

@ -1,3 +1,4 @@
#include "ggml-impl.h"
#include "ggml-blas.h"
#include "ggml-backend-impl.h"

113
ggml/src/ggml-cann.cpp
View File

@ -30,6 +30,7 @@
#include <cstring>
#include <mutex>
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-cann/aclnn_ops.h"
#include "ggml-cann/common.h"
@ -1220,6 +1221,116 @@ ggml_backend_cann_buffer_type(int32_t device) {
return &ggml_backend_cann_buffer_types[device];
}
/**
* @brief Retrieves the name associated with a CANN host buffer type.
*
* This function returns the descriptive name associated with the specified
* CANN host buffer type context.
*
* @param buft Pointer to the host buffer type context.
* @return Const pointer to the C-style string containing the name.
*/
GGML_CALL static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
return "CANN_Host";
GGML_UNUSED(buft);
}
/**
* @brief Retrieves the name associated with a CANN host buffer.
*
* This function returns the descriptive name associated with the specified
* CANN host buffer context.
*
* @param buft Pointer to the host buffer context.
* @return Const pointer to the C-style string containing the name.
*/
GGML_CALL static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
return "CANN_Host";
GGML_UNUSED(buffer);
}
/**
* @brief Free resources associated with a CANN host buffer.
*
* This function frees the resources associated with a CANN host buffer, including
* its context.
*
* @param buffer The CANN host buffer to free.
*/
GGML_CALL static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
ACL_CHECK(aclrtFreeHost(buffer->context));
}
/**
* @brief Allocates a new CANN host buffer of the specified size.
*
* This function allocates a new CANN host buffer with the given size.
* @param size Size in bytes of the host buffer to allocate.
* @return Pointer to the allocated host buffer, or nullptr if allocation fails.
*/
static void * ggml_cann_host_malloc(size_t size) {
if (getenv("GGML_CANN_NO_PINNED") != nullptr) {
return nullptr;
}
void * hostPtr = nullptr;
aclError err = aclrtMallocHost((void **) &hostPtr, size);
if (err != ACL_SUCCESS) {
GGML_CANN_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, aclGetRecentErrMsg());
return nullptr;
}
return hostPtr;
}
/**
* @brief Allocates a new CANN host buffer of the specified type and size.
*
* @param buft Pointer to the host buffer type context.
* @param size Size in bytes of the host buffer to allocate.
* @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
*/
GGML_CALL static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
void * hostPtr = ggml_cann_host_malloc(size);
if (hostPtr == nullptr) {
// fallback to cpu buffer
return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
}
ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
buffer->buft = buft;
buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
return buffer;
}
/**
* @brief Interface for managing CANN host buffer types in the GGML backend.
*
* Provides function pointers for allocating, querying properties, and managing
* memory for CANN buffer types in the GGML backend.
*/
GGML_CALL ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
/* .iface = */ {
/* .get_name = */ ggml_backend_cann_host_buffer_type_name,
/* .alloc_buffer = */ ggml_backend_cann_host_buffer_type_alloc_buffer,
/* .get_alignment = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
/* .get_max_size = */ NULL, // defaults to SIZE_MAX
/* .get_alloc_size = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
/* .is_host = */ ggml_backend_cpu_buffer_type()->iface.is_host,
},
/* .context = */ nullptr,
};
return &ggml_backend_cann_buffer_type_host;
}
/**
* @brief Computes the forward operation for a given tensor using CANN
* operations.
@ -1942,7 +2053,7 @@ GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
return nullptr;
}
ggml_cann_set_device(ctx->device);
ggml_backend_t cann_backend =
new ggml_backend{/* .guid = */ ggml_backend_cann_guid(),
/* .interface = */ ggml_backend_cann_interface,

614
ggml/src/ggml-cpu-impl.h Normal file
View File

@ -0,0 +1,614 @@
#pragma once
// GGML CPU internal header
#include "ggml.h"
#include "ggml-impl.h"
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
//#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#ifdef __cplusplus
extern "C" {
#endif
#if defined(_MSC_VER)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This is binary identical with Google Brain float conversion.
* Floats shall round to nearest even, and NANs shall be quiet.
* Subnormals aren't flushed to zero, except perhaps when used.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif
// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#include <sys/prctl.h>
#endif
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#if defined(__ARM_NEON)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
typedef __fp16 ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddlvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddlvq_s16(int16x8_t v) {
int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
#if !defined(__riscv)
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__loongarch_asx)
typedef union {
int32_t i;
float f;
} ft_union;
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
}
static __m256 __lasx_xvreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
#ifdef __cplusplus
}
#endif

8
ggml/src/ggml-cuda.cu
View File

@ -1,5 +1,5 @@
#include "ggml-cuda.h"
#include "ggml.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-cuda/common.cuh"
@ -2552,7 +2552,11 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
continue;
}
if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);

6
ggml/src/ggml-cuda/mmq.cu
View File

@ -26,7 +26,11 @@ void ggml_cuda_op_mul_mat_q(
// nrows_dst == nrows of the matrix that the kernel writes into
const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst};
// The stream-k decomposition is only faster for recent NVIDIA GPUs.
// Also its fixup needs to allocate a temporary buffer in the memory pool.
// There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
switch (src0->type) {
case GGML_TYPE_Q4_0:

4
ggml/src/ggml-cuda/mmq.cuh
View File

@ -2742,6 +2742,7 @@ struct mmq_args {
int64_t ne00; int64_t ne01; int64_t stride01;
int64_t ne10; int64_t ne11; int64_t stride11;
int64_t ne0;
bool use_stream_k;
};
template<ggml_type type>
@ -2777,8 +2778,7 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
const dim3 block_nums_xy_tiling(nty, ntx, 1);
const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
if (!use_stream_k) {
if (!args.use_stream_k) {
if (args.ne01 % mmq_y == 0) {
constexpr bool need_check = false;
mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>

12
ggml/src/ggml-cuda/sum.cu
View File

@ -1,13 +1,15 @@
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.
// For this reason CUB must be included BEFORE anything else.
#include <cub/cub.cuh>
using namespace cub;
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
#include "sumrows.cuh"
#include "sum.cuh"
#include <cstdint>
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
#include <cub/cub.cuh>
using namespace cub;
#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
size_t tmp_size = 0;

39
ggml/src/ggml-cuda/vendors/musa.h vendored
View File

@ -130,42 +130,3 @@
#define cudaKernelNodeParams musaKernelNodeParams
#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
#define cudaStreamEndCapture musaStreamEndCapture
// XXX: Clang builtins mapping
#define __vsub4 __vsub4_musa
#define __vcmpeq4 __vcmpeq4_musa
#define __vcmpne4 __vcmpne4_musa
#ifndef __has_builtin
#define __has_builtin(x) 0
#endif
typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
static __device__ __forceinline__ int __vsub4_musa(const int a, const int b) {
return __vsubss4(a, b);
}
static __device__ __forceinline__ unsigned int __vcmpeq4_musa(unsigned int a, unsigned int b) {
const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
unsigned int c;
uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
#pragma unroll
for (int i = 0; i < 4; ++i) {
vc[i] = va[i] == vb[i] ? 0xff : 0x00;
}
return c;
}
static __device__ __forceinline__ unsigned int __vcmpne4_musa(unsigned int a, unsigned int b) {
const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
unsigned int c;
uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
#pragma unroll
for (int i = 0; i < 4; ++i) {
vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
}
return c;
}

642
ggml/src/ggml-impl.h
View File

@ -1,15 +1,17 @@
#pragma once
#include "ggml.h"
// GGML internal header
#include "ggml.h"
#include <assert.h>
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
#include <stddef.h>
#include <stdbool.h>
#include <string.h> // memcpy
#include <math.h> // fabsf
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
#undef MIN
#undef MAX
@ -17,96 +19,6 @@
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#if defined(_MSC_VER)
#define m512bh(p) p
#define m512i(p) p
#else
#define m512bh(p) (__m512bh)(p)
#define m512i(p) (__m512i)(p)
#endif
/**
* Converts brain16 to float32.
*
* The bfloat16 floating point format has the following structure:
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───┐
* 0b0000000000000000 brain16
*
* Since bf16 has the same number of exponent bits as a 32bit float,
* encoding and decoding numbers becomes relatively straightforward.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌──┴───┐┌─┴───────────────────┐
* 0b00000000000000000000000000000000 IEEE binary32
*
* For comparison, the standard fp16 format has fewer exponent bits.
*
* ┌sign
* │
* │ ┌exponent
* │ │
* │ │ ┌mantissa
* │ │ │
* │┌─┴─┐┌─┴──────┐
* 0b0000000000000000 IEEE binary16
*
* @see IEEE 754-2008
*/
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
union {
float f;
uint32_t i;
} u;
u.i = (uint32_t)h.bits << 16;
return u.f;
}
/**
* Converts float32 to brain16.
*
* This is binary identical with Google Brain float conversion.
* Floats shall round to nearest even, and NANs shall be quiet.
* Subnormals aren't flushed to zero, except perhaps when used.
* This code should vectorize nicely if using modern compilers.
*/
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
ggml_bf16_t h;
union {
float f;
uint32_t i;
} u;
u.f = s;
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
h.bits = (u.i >> 16) | 64; /* force to quiet */
return h;
}
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
return h;
}
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
#ifdef __cplusplus
extern "C" {
#endif
// static_assert should be a #define, but if it's not,
// fall back to the _Static_assert C11 keyword.
// if C99 - static_assert is noop
@ -121,516 +33,10 @@ extern "C" {
#endif
#endif
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
#ifndef __FMA__
#define __FMA__
#endif
#ifndef __F16C__
#define __F16C__
#endif
#endif
// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
#ifndef __SSE3__
#define __SSE3__
#endif
#ifndef __SSSE3__
#define __SSSE3__
#endif
#endif
#if defined(__ARM_FEATURE_SVE)
#include <arm_sve.h>
#include <sys/prctl.h>
#endif
// 16-bit float
// on Arm, we use __fp16
// on x86, we use uint16_t
#if defined(__ARM_NEON)
// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
//
// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
//
#include <arm_neon.h>
#ifdef _MSC_VER
typedef uint16_t ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
#else
typedef __fp16 ggml_fp16_internal_t;
#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
#endif // _MSC_VER
#if !defined(__aarch64__)
// 32-bit ARM compatibility
// vaddlvq_s16
// vpaddq_s16
// vpaddq_s32
// vaddvq_s32
// vaddvq_f32
// vmaxvq_f32
// vcvtnq_s32_f32
// vzip1_u8
// vzip2_u8
inline static int32_t vaddlvq_s16(int16x8_t v) {
int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
}
inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
return vcombine_s16(a0, b0);
}
inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
return vcombine_s32(a0, b0);
}
inline static int32_t vaddvq_s32(int32x4_t v) {
return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
}
inline static float vaddvq_f32(float32x4_t v) {
return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
}
inline static float vmaxvq_f32(float32x4_t v) {
return
MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
}
inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[0]; res[1] = b[0];
res[2] = a[1]; res[3] = b[1];
res[4] = a[2]; res[5] = b[2];
res[6] = a[3]; res[7] = b[3];
return res;
}
inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
uint8x8_t res;
res[0] = a[4]; res[1] = b[4];
res[2] = a[5]; res[3] = b[5];
res[4] = a[6]; res[5] = b[6];
res[6] = a[7]; res[7] = b[7];
return res;
}
// vld1q_s16_x2
// vld1q_u8_x2
// vld1q_u8_x4
// vld1q_s8_x2
// vld1q_s8_x4
// TODO: double-check these work correctly
typedef struct ggml_int16x8x2_t {
int16x8_t val[2];
} ggml_int16x8x2_t;
inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
ggml_int16x8x2_t res;
res.val[0] = vld1q_s16(ptr + 0);
res.val[1] = vld1q_s16(ptr + 8);
return res;
}
typedef struct ggml_uint8x16x2_t {
uint8x16_t val[2];
} ggml_uint8x16x2_t;
inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
ggml_uint8x16x2_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
return res;
}
typedef struct ggml_uint8x16x4_t {
uint8x16_t val[4];
} ggml_uint8x16x4_t;
inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
ggml_uint8x16x4_t res;
res.val[0] = vld1q_u8(ptr + 0);
res.val[1] = vld1q_u8(ptr + 16);
res.val[2] = vld1q_u8(ptr + 32);
res.val[3] = vld1q_u8(ptr + 48);
return res;
}
typedef struct ggml_int8x16x2_t {
int8x16_t val[2];
} ggml_int8x16x2_t;
inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
ggml_int8x16x2_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
return res;
}
typedef struct ggml_int8x16x4_t {
int8x16_t val[4];
} ggml_int8x16x4_t;
inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
ggml_int8x16x4_t res;
res.val[0] = vld1q_s8(ptr + 0);
res.val[1] = vld1q_s8(ptr + 16);
res.val[2] = vld1q_s8(ptr + 32);
res.val[3] = vld1q_s8(ptr + 48);
return res;
}
// NOTE: not tested
inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
int8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
// NOTE: not tested
inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
uint8x16_t res;
res[ 0] = a[b[ 0]];
res[ 1] = a[b[ 1]];
res[ 2] = a[b[ 2]];
res[ 3] = a[b[ 3]];
res[ 4] = a[b[ 4]];
res[ 5] = a[b[ 5]];
res[ 6] = a[b[ 6]];
res[ 7] = a[b[ 7]];
res[ 8] = a[b[ 8]];
res[ 9] = a[b[ 9]];
res[10] = a[b[10]];
res[11] = a[b[11]];
res[12] = a[b[12]];
res[13] = a[b[13]];
res[14] = a[b[14]];
res[15] = a[b[15]];
return res;
}
#else
#define ggml_int16x8x2_t int16x8x2_t
#define ggml_uint8x16x2_t uint8x16x2_t
#define ggml_uint8x16x4_t uint8x16x4_t
#define ggml_int8x16x2_t int8x16x2_t
#define ggml_int8x16x4_t int8x16x4_t
#define ggml_vld1q_s16_x2 vld1q_s16_x2
#define ggml_vld1q_u8_x2 vld1q_u8_x2
#define ggml_vld1q_u8_x4 vld1q_u8_x4
#define ggml_vld1q_s8_x2 vld1q_s8_x2
#define ggml_vld1q_s8_x4 vld1q_s8_x4
#define ggml_vqtbl1q_s8 vqtbl1q_s8
#define ggml_vqtbl1q_u8 vqtbl1q_u8
#endif // !defined(__aarch64__)
#if !defined(__ARM_FEATURE_DOTPROD)
inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
}
#else
#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
#endif // !defined(__ARM_FEATURE_DOTPROD)
#endif // defined(__ARM_NEON)
#if defined(__ARM_NEON) && !defined(_MSC_VER)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
ggml_fp16_internal_t tmp;
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
return (float)tmp;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
ggml_fp16_t res;
ggml_fp16_internal_t tmp = f;
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
return res;
}
#else
#ifdef __wasm_simd128__
#include <wasm_simd128.h>
#else
#ifdef __POWER9_VECTOR__
#include <altivec.h>
#undef bool
#define bool _Bool
#else
#if defined(_MSC_VER) || defined(__MINGW32__)
#include <intrin.h>
#else
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
#if !defined(__riscv)
#include <immintrin.h>
#endif
#endif
#endif
#endif
#endif
#ifdef __riscv_v_intrinsic
#include <riscv_vector.h>
#endif
#if defined(__loongarch64)
#if defined(__loongarch_asx)
#include <lasxintrin.h>
#endif
#if defined(__loongarch_sx)
#include <lsxintrin.h>
#endif
#endif
#if defined(__loongarch_asx)
typedef union {
int32_t i;
float f;
} ft_union;
/* float type data load instructions */
static __m128 __lsx_vreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
}
static __m256 __lasx_xvreplfr2vr_s(float val) {
ft_union fi_tmpval = {.f = val};
return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
}
#endif
#ifdef __F16C__
#ifdef _MSC_VER
#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
#else
#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
#endif
#elif defined(__POWER9_VECTOR__)
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
/* the inline asm below is about 12% faster than the lookup method */
#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
register float f;
register double d;
__asm__(
"mtfprd %0,%2\n"
"xscvhpdp %0,%0\n"
"frsp %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=f"(f):
/* in */ "r"(h));
return f;
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
register double d;
register ggml_fp16_t r;
__asm__( /* xscvdphp can work on double or single precision */
"xscvdphp %0,%2\n"
"mffprd %1,%0\n" :
/* temp */ "=d"(d),
/* out */ "=r"(r):
/* in */ "f"(f));
return r;
}
#else
// FP16 <-> FP32
// ref: https://github.com/Maratyszcza/FP16
static inline float fp32_from_bits(uint32_t w) {
union {
uint32_t as_bits;
float as_value;
} fp32;
fp32.as_bits = w;
return fp32.as_value;
}
static inline uint32_t fp32_to_bits(float f) {
union {
float as_value;
uint32_t as_bits;
} fp32;
fp32.as_value = f;
return fp32.as_bits;
}
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
const uint32_t w = (uint32_t) h << 16;
const uint32_t sign = w & UINT32_C(0x80000000);
const uint32_t two_w = w + w;
const uint32_t exp_offset = UINT32_C(0xE0) << 23;
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float exp_scale = 0x1.0p-112f;
#else
const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
#endif
const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
const uint32_t magic_mask = UINT32_C(126) << 23;
const float magic_bias = 0.5f;
const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
const uint32_t result = sign |
(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
return fp32_from_bits(result);
}
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
const float scale_to_inf = 0x1.0p+112f;
const float scale_to_zero = 0x1.0p-110f;
#else
const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
#endif
float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
const uint32_t w = fp32_to_bits(f);
const uint32_t shl1_w = w + w;
const uint32_t sign = w & UINT32_C(0x80000000);
uint32_t bias = shl1_w & UINT32_C(0xFF000000);
if (bias < UINT32_C(0x71000000)) {
bias = UINT32_C(0x71000000);
}
base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
const uint32_t bits = fp32_to_bits(base);
const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
const uint32_t nonsign = exp_bits + mantissa_bits;
return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
}
#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
#endif // __F16C__
#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
#ifdef __ARM_FEATURE_SVE
#include <arm_sve.h>
#endif // __ARM_FEATURE_SVE
// precomputed f32 table for f16 (256 KB)
// defined in ggml.c, initialized in ggml_init()
extern float ggml_table_f32_f16[1 << 16];
// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
// This is also true for POWER9.
#if !defined(GGML_FP16_TO_FP32)
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
uint16_t s;
memcpy(&s, &f, sizeof(uint16_t));
return ggml_table_f32_f16[s];
}
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
#endif
#if !defined(GGML_FP32_TO_FP16)
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
#endif
// bitset
typedef uint32_t ggml_bitset_t;
static_assert(sizeof(ggml_bitset_t) == 4, "bitset_t constants must be updated");
#define BITSET_SHR 5 // log2(sizeof(ggml_bitset_t)*8)
#define BITSET_MASK (sizeof(ggml_bitset_t)*8 - 1)
@ -656,6 +62,12 @@ static inline void ggml_bitset_clear(ggml_bitset_t * bitset, size_t i) {
#define GGML_HASHSET_FULL ((size_t)-1)
#define GGML_HASHSET_ALREADY_EXISTS ((size_t)-2)
struct ggml_hash_set {
size_t size;
ggml_bitset_t * used; // whether or not the keys are in use i.e. set
struct ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if ggml_bitset_get(used, i)
};
struct ggml_hash_set ggml_hash_set_new(size_t size);
void ggml_hash_set_free(struct ggml_hash_set * hash_set);
@ -745,6 +157,30 @@ static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct g
GGML_ABORT("fatal error");
}
// computation graph
enum ggml_cgraph_eval_order {
GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
GGML_CGRAPH_EVAL_ORDER_COUNT
};
struct ggml_cgraph {
int size;
int n_nodes;
int n_leafs;
struct ggml_tensor ** nodes;
struct ggml_tensor ** grads;
struct ggml_tensor ** leafs;
struct ggml_hash_set visited_hash_set;
enum ggml_cgraph_eval_order order;
};
struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
#ifdef __cplusplus
}
#endif

2
ggml/src/ggml-kompute.cpp
View File

@ -1,4 +1,4 @@
#include "ggml.h"
#include "ggml-impl.h"
#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-kompute.h"

24
ggml/src/ggml-metal.m
View File

@ -1,7 +1,7 @@
#import "ggml-metal.h"
#import "ggml-impl.h"
#import "ggml-backend-impl.h"
#import "ggml.h"
#import <Foundation/Foundation.h>
@ -13,13 +13,16 @@
#define MAX(a, b) ((a) > (b) ? (a) : (b))
#ifdef GGML_METAL_NDEBUG
#define GGML_METAL_LOG(...)
#define GGML_METAL_LOG_INFO(...)
#define GGML_METAL_LOG_WARN(...)
#define GGML_METAL_LOG_ERROR(...)
#else
#define GGML_METAL_LOG(...) ggml_metal_log(GGML_LOG_LEVEL_NONE, __VA_ARGS__)
#define GGML_METAL_LOG_INFO(...) ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_METAL_LOG_WARN(...) ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
#define GGML_METAL_LOG_DEBUG(...) ggml_metal_log(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
#endif
#define UNUSED(x) (void)(x)
@ -3039,8 +3042,7 @@ static enum ggml_status ggml_metal_graph_compute(
if (status != MTLCommandBufferStatusCompleted) {
GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
if (status == MTLCommandBufferStatusError) {
NSString * error_code = [command_buffer error].localizedDescription;
GGML_METAL_LOG_INFO("error: %s\n", [error_code UTF8String]);
GGML_METAL_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
}
return GGML_STATUS_FAILED;
@ -3184,7 +3186,7 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t s
#ifndef GGML_METAL_NDEBUG
#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
if (@available(macOS 10.12, iOS 16.0, *)) {
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
GGML_METAL_LOG_DEBUG("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)\n",
__func__,
size_aligned / 1024.0 / 1024.0,
device.currentAllocatedSize / 1024.0 / 1024.0,
@ -3192,8 +3194,6 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t s
if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
} else {
GGML_METAL_LOG_INFO("\n");
}
} else {
GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
@ -3227,13 +3227,17 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
if (ctx->all_data != NULL) {
ctx->buffers[0].data = ctx->all_data;
ctx->buffers[0].size = size;
ctx->buffers[0].metal = nil;
if (size_aligned > 0) {
ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
length:size_aligned
options:MTLResourceStorageModeShared
deallocator:nil];
}
}
if (ctx->all_data == NULL || ctx->buffers[0].metal == nil) {
if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
free(ctx);
ggml_backend_metal_free_device();
@ -3312,13 +3316,16 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
if (size_aligned <= device.maxBufferLength) {
ctx->buffers[ctx->n_buffers].data = data;
ctx->buffers[ctx->n_buffers].size = size;
ctx->buffers[ctx->n_buffers].metal = nil;
if (size_aligned > 0) {
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
return false;
}
}
ggml_backend_metal_log_allocated_size(device, size_aligned);
@ -3335,13 +3342,16 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
ctx->buffers[ctx->n_buffers].size = size_step_aligned;
ctx->buffers[ctx->n_buffers].metal = nil;
if (size_step_aligned > 0) {
ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
return false;
}
}
ggml_backend_metal_log_allocated_size(device, size_step_aligned);

286
ggml/src/ggml-quants.c
View File

@ -3,6 +3,7 @@
#include "ggml-quants.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include <math.h>
@ -230,6 +231,12 @@ static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
return _mm_packus_epi16( bytes1, bytes2);
}
static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
const __m128i ax = _mm_sign_epi8(x, x);
const __m128i sy = _mm_sign_epi8(y, x);
return _mm_maddubs_epi16(ax, sy);
}
#endif
#elif defined(__SSSE3__)
// horizontally add 4x4 floats
@ -4003,13 +4010,18 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
float sumf = 0;
#if defined(__ARM_FEATURE_SVE)
if (ggml_sve_cnt_b == QK8_0) {
const svbool_t ptrueh = svptrue_pat_b8(SV_VL16);
const svbool_t ptruel = svnot_b_z(svptrue_b8(), ptrueh);
svfloat32_t sumv0 = svdup_n_f32(0.0f);
svfloat32_t sumv1 = svdup_n_f32(0.0f);
const int vector_length = ggml_sve_cnt_b*8;
// VLA Implementation using switch case
switch (vector_length) {
case 128:
{
// predicate for activating higher lanes for 4 float32 elements
const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
for (; ib + 1 < nb; ib += 2) {
const block_q4_0 * restrict x0 = &x[ib + 0];
const block_q4_0 * restrict x1 = &x[ib + 1];
@ -4021,8 +4033,54 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
// 4-bit -> 8-bit
const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx0r, 0x0F), 0x04));
const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(ptruel, svand_n_u8_m(ptrueh, qx1r, 0x0F), 0x04));
const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
// sub 8
const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
// load y
const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
// dot product
sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
} break;
case 256:
{
// predicate for activating higher lanes for 16 int8 elements
const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
// predicate for activating lower lanes for 16 int8 elements
const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
for (; ib + 1 < nb; ib += 2) {
const block_q4_0 * restrict x0 = &x[ib + 0];
const block_q4_0 * restrict x1 = &x[ib + 1];
const block_q8_0 * restrict y0 = &y[ib + 0];
const block_q8_0 * restrict y1 = &y[ib + 1];
// load x
const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
// 4-bit -> 8-bit
const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
// sub 8
const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
@ -4033,12 +4091,60 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
// dot product
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
} break;
case 512:
{
// predicate for activating higher lanes for 32 int8 elements
const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
// predicate for activating higher lanes for 16 int8 elements
const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
// predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
const svbool_t pl16 = svnot_b_z(ph32, ph16);
for (; ib + 1 < nb; ib += 2) {
const block_q4_0 * restrict x0 = &x[ib + 0];
const block_q4_0 * restrict x1 = &x[ib + 1];
const block_q8_0 * restrict y0 = &y[ib + 0];
const block_q8_0 * restrict y1 = &y[ib + 1];
// load x
const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
// 4-bit -> 8-bit
const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
// sub 8
const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
// load y
const svint8_t qy0 = svld1_s8(ph32, y0->qs);
const svint8_t qy1 = svld1_s8(ph32, y1->qs);
// dot product
sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
} break;
default:
assert(false && "Unsupported vector length");
break;
}
#elif defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
float32x4_t sumv1 = vdupq_n_f32(0.0f);
@ -4107,37 +4213,37 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
sumf = hsum_float_8(acc);
#elif defined(__AVX__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
const __m128i mone = _mm_set1_epi16(1);
// Main loop
for (; ib < nb; ++ib) {
// Compute combined scale for the block
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
__m256 accum1 = _mm256_setzero_ps();
__m256 accum2 = _mm256_setzero_ps();
for (; ib + 1 < nb; ib += 2) {
const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
const __m128i lowMask = _mm_set1_epi8(0xF);
const __m128i off = _mm_set1_epi8(8);
const __m128i tmp = _mm_loadu_si128((const __m128i *)x[ib].qs);
__m128i bx_0 = _mm_and_si128(lowMask, tmp);
__m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
bx_0 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
by_0 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
bx_0 = _mm_sub_epi8(bx_0, off);
const __m128i i32_1 = mul_sum_i8_pairs(bx_0, by_0);
// Convert int32_t to float
__m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
// Apply the scale, and accumulate
acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
_mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
_mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
}
sumf = hsum_float_8(acc);
sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
#elif defined(__SSSE3__)
// set constants
const __m128i lowMask = _mm_set1_epi8(0xF);
@ -5488,10 +5594,50 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
float sumf = 0;
#if defined(__ARM_FEATURE_SVE)
if (ggml_sve_cnt_b == QK8_0) {
svfloat32_t sumv0 = svdup_n_f32(0.0f);
svfloat32_t sumv1 = svdup_n_f32(0.0f);
const int vector_length = ggml_sve_cnt_b*8;
//VLA Implemenation for SVE
switch (vector_length) {
case 128:
{
// predicate for activating lanes for 16 Int8 elements
const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
for (; ib + 1 < nb; ib += 2) {
const block_q8_0 * restrict x0 = &x[ib + 0];
const block_q8_0 * restrict x1 = &x[ib + 1];
const block_q8_0 * restrict y0 = &y[ib + 0];
const block_q8_0 * restrict y1 = &y[ib + 1];
// load x
const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
// load y
const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
} break;
case 256:
{
//printf("sve256");
for (; ib + 1 < nb; ib += 2) {
const block_q8_0 * restrict x0 = &x[ib + 0];
const block_q8_0 * restrict x1 = &x[ib + 1];
@ -5506,11 +5652,66 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * r
const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
} break;
case 512:
{
// predicate for activating high 256 bit
const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
// predicate for activating low 256 bit
const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
// predicate for activating high lanes for 8 float32 elements
const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
// predicate for activating low lanes for 8 float32 elements
const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
svfloat32_t sumv00 = svdup_n_f32(0.0f);
for (; ib + 1 < nb; ib += 2) {
const block_q8_0 * restrict x0 = &x[ib + 0];
const block_q8_0 * restrict x1 = &x[ib + 1];
const block_q8_0 * restrict y0 = &y[ib + 0];
const block_q8_0 * restrict y1 = &y[ib + 1];
//load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
// and add them to make one 64 element vector
// load x
const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
// load y
const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
// scale creation
const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
// duplicate deq1 in first half of vector and deq2 in second half of vector
const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
}
sumf = svaddv_f32(svptrue_b32(), sumv00);
break;
}
default:
assert(false && "Unsupported vector length");
break;
}
#elif defined(__ARM_NEON)
float32x4_t sumv0 = vdupq_n_f32(0.0f);
@ -11625,15 +11826,6 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void *
#endif
}
#if defined(__AVX__)
static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
const __m128i ax = _mm_sign_epi8(x, x);
const __m128i sy = _mm_sign_epi8(y, x);
return _mm_maddubs_epi16(ax, sy);
}
#endif
#if defined(__AVX2__)
static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
const __m256i ax = _mm256_sign_epi8(x, x);

8
ggml/src/ggml-rpc.cpp
View File

@ -1,5 +1,5 @@
#include "ggml-rpc.h"
#include "ggml.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include <cinttypes>
@ -883,15 +883,17 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
}
result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
return nullptr;
result->buffer = nullptr;
}
if (result->buffer) {
// require that the tensor data does not go beyond the buffer end
uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
}
result->op = (ggml_op) tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
@ -1060,7 +1062,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<u
const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
static size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ NULL,

6
ggml/src/ggml-sycl.cpp
View File

@ -33,7 +33,7 @@
#include <sycl/half_type.hpp>
#include "ggml-sycl.h"
#include "ggml.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-sycl/backend.hpp"
@ -5137,13 +5137,17 @@ GGML_CALL static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, cons
case GGML_OP_SCALE:
case GGML_OP_SQR:
case GGML_OP_CLAMP:
return true;
case GGML_OP_CONT:
return op->src[0]->type != GGML_TYPE_BF16;
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
return true;
case GGML_OP_ROPE:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_IM2COL:
// TODO: add support for the new F32 operations
return op->src[0]->type == GGML_TYPE_F16;
case GGML_OP_POOL_2D:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:

2
ggml/src/ggml-vulkan.cpp
View File

@ -21,7 +21,7 @@
#include <memory>
#include <mutex>
#include "ggml.h"
#include "ggml-impl.h"
#include "ggml-backend-impl.h"
#include "ggml-vulkan-shaders.hpp"

81
ggml/src/ggml.c
View File

@ -2,6 +2,7 @@
#define _USE_MATH_DEFINES // For M_PI on MSVC
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#include "ggml.h"
#include "ggml-aarch64.h"
@ -287,6 +288,7 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
#define GGML_DEBUG 0
#define GGML_GELU_FP16
#define GGML_GELU_QUICK_FP16
#define GGML_N_TASKS_MAX (-1)
#define GGML_SOFT_MAX_UNROLL 4
#define GGML_VEC_DOT_UNROLL 2
@ -1124,17 +1126,17 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
{ \
int offset = GGML_F32_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f32(x[i], x[offset+i]); \
(x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
} \
res = GGML_F32x4_REDUCE_ONE(x[0]); \
(res) = GGML_F32x4_REDUCE_ONE((x)[0]); \
}
#define GGML_F32_VEC GGML_F32x4
@ -1165,26 +1167,26 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
do { \
int offset = GGML_F16_ARR >> 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
offset >>= 1; \
for (int i = 0; i < offset; ++i) { \
x[i] = vaddq_f16(x[i], x[offset+i]); \
(x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
} \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
(res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
} while (0)
#define GGML_F16_VEC GGML_F16x8
#define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
#define GGML_F16_VEC_SET1 GGML_F16x8_SET1
#define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), r[i])
#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i])
#define GGML_F16_VEC_FMA GGML_F16x8_FMA
#define GGML_F16_VEC_ADD GGML_F16x8_ADD
#define GGML_F16_VEC_MUL GGML_F16x8_MUL
@ -1893,6 +1895,23 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
#endif
//
// ggml object
//
struct ggml_object {
size_t offs;
size_t size;
struct ggml_object * next;
enum ggml_object_type type;
char padding[4];
};
static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
//
// ggml context
//
@ -3381,7 +3400,7 @@ double ggml_type_sizef(enum ggml_type type) {
}
GGML_CALL const char * ggml_type_name(enum ggml_type type) {
return type_traits[type].type_name;
return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
}
GGML_CALL bool ggml_is_quantized(enum ggml_type type) {
@ -3847,7 +3866,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
__func__, cur_end + size_needed, ctx->mem_size);
__func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
assert(false);
return NULL;
}
@ -19161,6 +19180,34 @@ void ggml_graph_clear(struct ggml_cgraph * cgraph) {
ggml_hash_set_reset(&cgraph->visited_hash_set);
}
int ggml_graph_size(struct ggml_cgraph * cgraph) {
return cgraph->size;
}
struct ggml_tensor * ggml_graph_node(struct ggml_cgraph * cgraph, int i) {
if (i < 0) {
GGML_ASSERT(cgraph->n_nodes + i >= 0);
return cgraph->nodes[cgraph->n_nodes + i];
}
GGML_ASSERT(i < cgraph->n_nodes);
return cgraph->nodes[i];
}
struct ggml_tensor ** ggml_graph_nodes(struct ggml_cgraph * cgraph) {
return cgraph->nodes;
}
int ggml_graph_n_nodes(struct ggml_cgraph * cgraph) {
return cgraph->n_nodes;
}
void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
GGML_ASSERT(cgraph->size > cgraph->n_nodes);
cgraph->nodes[cgraph->n_nodes] = tensor;
cgraph->n_nodes++;
}
// Android's libc implementation "bionic" does not support setting affinity
#if defined(__gnu_linux__)
static void set_numa_thread_affinity(int thread_n) {
@ -23242,6 +23289,14 @@ int ggml_cpu_has_arm_fma(void) {
#endif
}
int ggml_cpu_has_riscv_v(void) {
#if defined(__riscv_v_intrinsic)
return 1;
#else
return 0;
#endif
}
int ggml_cpu_has_metal(void) {
#if defined(GGML_USE_METAL)
return 1;

39
ggml/src/llamafile/sgemm.cpp
View File

@ -50,6 +50,7 @@
#include "sgemm.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#ifdef _MSC_VER
@ -235,6 +236,14 @@ template <> inline __m512 load(const ggml_fp16_t *p) {
}
#endif // __AVX512F__
////////////////////////////////////////////////////////////////////////////////////////////////////
// CONSTANTS
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// FLOATING POINT MATRIX MULTIPLICATION
@ -933,6 +942,20 @@ class tinyBLAS_Q0_AVX {
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
}
inline __m256i load(const block_iq4_nl *b) {
return MM256_SET_M128I(load1(b), load0(b));
}
inline __m128i load0(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
}
inline __m128i load1(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
}
inline __m256 updot(__m256i u, __m256i s) {
__m256i res;
#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
@ -1159,6 +1182,22 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
#endif
}
case GGML_TYPE_IQ4_NL: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
k, (const block_iq4_nl *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
ith, nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
default:
return false;
}

38
gguf-py/gguf/constants.py
View File

@ -210,6 +210,7 @@ class MODEL_ARCH(IntEnum):
ORION = auto()
INTERNLM2 = auto()
MINICPM = auto()
MINICPM3 = auto()
GEMMA = auto()
GEMMA2 = auto()
STARCODER2 = auto()
@ -219,6 +220,7 @@ class MODEL_ARCH(IntEnum):
COMMAND_R = auto()
DBRX = auto()
OLMO = auto()
OLMOE = auto()
OPENELM = auto()
ARCTIC = auto()
DEEPSEEK2 = auto()
@ -364,6 +366,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.ORION: "orion",
MODEL_ARCH.INTERNLM2: "internlm2",
MODEL_ARCH.MINICPM: "minicpm",
MODEL_ARCH.MINICPM3: "minicpm3",
MODEL_ARCH.GEMMA: "gemma",
MODEL_ARCH.GEMMA2: "gemma2",
MODEL_ARCH.STARCODER2: "starcoder2",
@ -373,6 +376,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
MODEL_ARCH.COMMAND_R: "command-r",
MODEL_ARCH.DBRX: "dbrx",
MODEL_ARCH.OLMO: "olmo",
MODEL_ARCH.OLMOE: "olmoe",
MODEL_ARCH.OPENELM: "openelm",
MODEL_ARCH.ARCTIC: "arctic",
MODEL_ARCH.DEEPSEEK2: "deepseek2",
@ -869,6 +873,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
],
MODEL_ARCH.MINICPM3: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_A,
MODEL_TENSOR.ATTN_Q_B,
MODEL_TENSOR.ATTN_KV_A_MQA,
MODEL_TENSOR.ATTN_KV_B,
MODEL_TENSOR.ATTN_Q_A_NORM,
MODEL_TENSOR.ATTN_KV_A_NORM,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.GEMMA: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
@ -1010,6 +1031,23 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
],
MODEL_ARCH.OLMOE: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,
MODEL_TENSOR.OUTPUT,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.ATTN_Q,
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_NORM,
MODEL_TENSOR.ATTN_Q_NORM,
MODEL_TENSOR.ATTN_K_NORM,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_GATE_EXP,
MODEL_TENSOR.FFN_UP_EXP,
MODEL_TENSOR.FFN_DOWN_EXP,
],
MODEL_ARCH.OPENELM: [
MODEL_TENSOR.TOKEN_EMBD,
MODEL_TENSOR.OUTPUT_NORM,

30
gguf-py/gguf/tensor_mapping.py
View File

@ -13,7 +13,7 @@ class TensorNameMap:
"transformer.wte", # gpt2 gpt-j mpt refact qwen dbrx jais exaone
"transformer.word_embeddings", # falcon
"word_embeddings", # bloom
"model.embed_tokens", # llama-hf nemotron
"model.embed_tokens", # llama-hf nemotron olmoe
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert nomic-bert
"language_model.embedding.word_embeddings", # persimmon
@ -54,7 +54,7 @@ class TensorNameMap:
# Output
MODEL_TENSOR.OUTPUT: (
"embed_out", # gptneox
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone
"lm_head", # gpt2 mpt falcon llama-hf baichuan qwen mamba dbrx jais nemotron exaone olmoe
"output", # llama-pth bloom internlm2
"word_embeddings_for_head", # persimmon
"lm_head.linear", # phi2
@ -66,7 +66,7 @@ class TensorNameMap:
MODEL_TENSOR.OUTPUT_NORM: (
"gpt_neox.final_layer_norm", # gptneox
"transformer.ln_f", # gpt2 gpt-j falcon jais exaone
"model.norm", # llama-hf baichuan internlm2
"model.norm", # llama-hf baichuan internlm2 olmoe
"norm", # llama-pth
"transformer.norm_f", # mpt dbrx
"ln_f", # refact bloom qwen gpt2
@ -98,7 +98,7 @@ class TensorNameMap:
"transformer.h.{bid}.input_layernorm", # falcon7b
"h.{bid}.input_layernorm", # bloom
"transformer.h.{bid}.ln_mlp", # falcon40b
"model.layers.{bid}.input_layernorm", # llama-hf nemotron
"model.layers.{bid}.input_layernorm", # llama-hf nemotron olmoe
"layers.{bid}.attention_norm", # llama-pth
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
@ -142,7 +142,7 @@ class TensorNameMap:
# Attention query
MODEL_TENSOR.ATTN_Q: (
"model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.q_proj", # llama-hf nemotron olmoe
"layers.{bid}.attention.wq", # llama-pth
"encoder.layer.{bid}.attention.self.query", # bert
"transformer.h.{bid}.attn.q_proj", # gpt-j
@ -154,7 +154,7 @@ class TensorNameMap:
# Attention key
MODEL_TENSOR.ATTN_K: (
"model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.k_proj", # llama-hf nemotron olmoe
"layers.{bid}.attention.wk", # llama-pth
"encoder.layer.{bid}.attention.self.key", # bert
"transformer.h.{bid}.attn.k_proj", # gpt-j
@ -167,7 +167,7 @@ class TensorNameMap:
# Attention value
MODEL_TENSOR.ATTN_V: (
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.v_proj", # llama-hf nemotron olmoe
"layers.{bid}.attention.wv", # llama-pth
"encoder.layer.{bid}.attention.self.value", # bert
"transformer.h.{bid}.attn.v_proj", # gpt-j
@ -185,7 +185,7 @@ class TensorNameMap:
"transformer.blocks.{bid}.attn.out_proj", # mpt
"transformer.h.{bid}.self_attention.dense", # falcon
"h.{bid}.self_attention.dense", # bloom
"model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron
"model.layers.{bid}.self_attn.o_proj", # llama-hf nemotron olmoe
"layers.{bid}.attention.wo", # llama-pth
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
@ -229,7 +229,7 @@ class TensorNameMap:
"transformer.h.{bid}.ln_2", # gpt2 refact qwen jais exaone
"h.{bid}.post_attention_layernorm", # bloom
"transformer.blocks.{bid}.norm_2", # mpt
"model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron
"model.layers.{bid}.post_attention_layernorm", # llama-hf nemotron olmoe
"layers.{bid}.ffn_norm", # llama-pth
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
"model.layers.{bid}.ln2", # yi
@ -253,7 +253,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
"model.layers.{bid}.mlp.gate", # qwen2moe
"model.layers.{bid}.mlp.gate", # qwen2moe olmoe
"transformer.decoder_layer.{bid}.router", # Grok
"transformer.blocks.{bid}.ffn.router.layer", # dbrx
),
@ -295,7 +295,7 @@ class TensorNameMap:
"layers.{bid}.feed_forward.experts.w3", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_v", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.v1", # dbrx
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe (merged)
"model.layers.{bid}.mlp.experts.up_proj", # qwen2moe olmoe (merged)
),
MODEL_TENSOR.FFN_UP_SHEXP: (
@ -327,7 +327,7 @@ class TensorNameMap:
"layers.{bid}.feed_forward.experts.w1", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w1", # dbrx
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe (merged)
"model.layers.{bid}.mlp.experts.gate_proj", # qwen2moe olmoe (merged)
),
MODEL_TENSOR.FFN_GATE_SHEXP: (
@ -367,7 +367,7 @@ class TensorNameMap:
"layers.{bid}.feed_forward.experts.w2", # mixtral (merged)
"transformer.decoder_layer.{bid}.moe.linear_1", # Grok (merged)
"transformer.blocks.{bid}.ffn.experts.mlp.w2", # dbrx
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe (merged)
"model.layers.{bid}.mlp.experts.down_proj", # qwen2moe olmoe (merged)
),
MODEL_TENSOR.FFN_DOWN_SHEXP: (
@ -378,7 +378,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_Q_NORM: (
"language_model.encoder.layers.{bid}.self_attention.q_layernorm",
"model.layers.{bid}.self_attn.q_layernorm", # persimmon
"model.layers.{bid}.self_attn.q_norm", # cohere
"model.layers.{bid}.self_attn.q_norm", # cohere olmoe
"transformer.blocks.{bid}.attn.q_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_q", # jina-bert-v2
"transformer.layers.{bid}.attn.q_norm", # openelm
@ -387,7 +387,7 @@ class TensorNameMap:
MODEL_TENSOR.ATTN_K_NORM: (
"language_model.encoder.layers.{bid}.self_attention.k_layernorm",
"model.layers.{bid}.self_attn.k_layernorm", # persimmon
"model.layers.{bid}.self_attn.k_norm", # cohere
"model.layers.{bid}.self_attn.k_norm", # cohere olmoe
"transformer.blocks.{bid}.attn.k_ln", # sea-lion
"encoder.layer.{bid}.attention.self.layer_norm_k", # jina-bert-v2
"transformer.layers.{bid}.attn.k_norm", # openelm

47
include/llama.h
View File

@ -343,7 +343,7 @@ extern "C" {
bool embeddings; // if true, extract embeddings (together with logits)
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
//bool no_perf; // whether to measure performance timings, TODO: implement
bool no_perf; // whether to measure performance timings
// Abort callback
// if it returns true, execution of llama_decode() will be aborted
@ -1056,6 +1056,9 @@ extern "C" {
LLAMA_API struct llama_sampler * llama_sampler_chain_get(const struct llama_sampler * chain, int32_t i);
LLAMA_API int llama_sampler_chain_n (const struct llama_sampler * chain);
// after removing a sampler, the chain will no longer own it, and it will not be freed when the chain is freed
LLAMA_API struct llama_sampler * llama_sampler_chain_remove( struct llama_sampler * chain, int32_t i);
// available samplers:
LLAMA_API struct llama_sampler * llama_sampler_init_greedy (void);
@ -1127,15 +1130,20 @@ extern "C" {
int32_t n_logit_bias,
const llama_logit_bias * logit_bias);
// Shorthand for:
// Returns the seed used by the sampler if applicable, LLAMA_DEFAULT_SEED otherwise
LLAMA_API uint32_t llama_sampler_get_seed(const struct llama_sampler * smpl);
/// @details Sample and accept a token from the idx-th output of the last evaluation
//
// Shorthand for:
// const auto * logits = llama_get_logits_ith(ctx, idx);
// llama_token_data_array cur_p = { ... init from logits ... };
// llama_sampler_apply(smpl, &cur_p);
// return cur_p.data[cur_p.selected].id;
//
// At this point, this is mostly a convenience function.
//
// auto token = cur_p.data[cur_p.selected].id;
// llama_sampler_accept(smpl, token);
// return token;
// Returns the sampled token
LLAMA_API llama_token llama_sampler_sample(struct llama_sampler * smpl, struct llama_context * ctx, int32_t idx);
// TODO: extend in the future
@ -1168,13 +1176,30 @@ extern "C" {
// NOTE: Used by llama.cpp examples, avoid using in third-party apps. Instead, do your own performance measurements.
//
enum llama_perf_type {
LLAMA_PERF_TYPE_CONTEXT = 0,
LLAMA_PERF_TYPE_SAMPLER_CHAIN = 1,
struct llama_perf_context_data {
double t_start_ms;
double t_load_ms;
double t_p_eval_ms;
double t_eval_ms;
int32_t n_p_eval;
int32_t n_eval;
};
LLAMA_API void llama_perf_print(const void * ctx, enum llama_perf_type type);
LLAMA_API void llama_perf_reset( void * ctx, enum llama_perf_type type);
struct llama_perf_sampler_data {
double t_sample_ms;
int32_t n_sample;
};
LLAMA_API struct llama_perf_context_data llama_perf_context (const struct llama_context * ctx);
LLAMA_API void llama_perf_context_print(const struct llama_context * ctx);
LLAMA_API void llama_perf_context_reset( struct llama_context * ctx);
// NOTE: the following work only with samplers constructed via llama_sampler_chain_init
LLAMA_API struct llama_perf_sampler_data llama_perf_sampler (const struct llama_sampler * chain);
LLAMA_API void llama_perf_sampler_print(const struct llama_sampler * chain);
LLAMA_API void llama_perf_sampler_reset( struct llama_sampler * chain);
LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);

1
src/llama-impl.h
View File

@ -24,6 +24,7 @@ LLAMA_ATTRIBUTE_FORMAT(2, 3)
void llama_log_internal (ggml_log_level level, const char * format, ...);
void llama_log_callback_default(ggml_log_level level, const char * text, void * user_data);
#define LLAMA_LOG(...) llama_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
#define LLAMA_LOG_INFO(...) llama_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
#define LLAMA_LOG_WARN(...) llama_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
#define LLAMA_LOG_ERROR(...) llama_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)

Some files were not shown because too many files have changed in this diff Show More