haraldwolff/llama.cpp
1
0
Fork 0
Code Issues Pull requests Packages Projects Releases Wiki Activity

ggml : quantization refactoring (#3833)

Browse source 
* ggml : factor all quantization code in ggml-quants

ggml-ci

* ggml-quants : fix Zig and Swift builds + quantize tool

ggml-ci

* quantize : --pure option for disabling k-quant mixtures

---------

Co-authored-by: cebtenzzre <cebtenzzre@gmail.com>
This commit is contained in:
Georgi Gerganov 2023-10-29 18:32:28 +02:00 committed by GitHub
parent ff3bad83e2
commit d69d777c02
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
11 changed files with 2372 additions and 2385 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
CMakeLists.txt
View file

@ -94,7 +94,6 @@ option(LLAMA_CLBLAST "llama: use CLBlast"
option(LLAMA_METAL "llama: use Metal" ${LLAMA_METAL_DEFAULT})
option(LLAMA_METAL_NDEBUG "llama: disable Metal debugging" OFF)
option(LLAMA_MPI "llama: use MPI" OFF)
option(LLAMA_K_QUANTS "llama: use k-quants" ON)
option(LLAMA_QKK_64 "llama: use super-block size of 64 for k-quants" OFF)
option(LLAMA_BUILD_TESTS "llama: build tests" ${LLAMA_STANDALONE})
@ -278,13 +277,8 @@ if (LLAMA_BLAS)
endif()
endif()
if (LLAMA_K_QUANTS)
set(GGML_HEADERS_EXTRA k_quants.h)
set(GGML_SOURCES_EXTRA k_quants.c)
add_compile_definitions(GGML_USE_K_QUANTS)
if (LLAMA_QKK_64)
add_compile_definitions(GGML_QKK_64)
endif()
if (LLAMA_QKK_64)
add_compile_definitions(GGML_QKK_64)
endif()
if (LLAMA_CUBLAS)
@ -673,6 +667,8 @@ add_library(ggml OBJECT
ggml-alloc.h
ggml-backend.c
ggml-backend.h
ggml-quants.c
ggml-quants.h
${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}

18
Makefile
View file

@ -342,13 +342,9 @@ else
MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
endif
ifndef LLAMA_NO_K_QUANTS
MK_CPPFLAGS += -DGGML_USE_K_QUANTS
OBJS += k_quants.o
ifdef LLAMA_QKK_64
MK_CPPFLAGS += -DGGML_QKK_64
endif
endif
ifndef LLAMA_NO_ACCELERATE
# Mac OS - include Accelerate framework.
@ -365,7 +361,7 @@ ifdef LLAMA_MPI
MK_CPPFLAGS += -DGGML_USE_MPI
MK_CFLAGS += -Wno-cast-qual
MK_CXXFLAGS += -Wno-cast-qual
OBJS += ggml-mpi.o
OBJS += ggml-mpi.o
endif # LLAMA_MPI
ifdef LLAMA_OPENBLAS
@ -382,7 +378,7 @@ endif # LLAMA_BLIS
ifdef LLAMA_CUBLAS
MK_CPPFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
MK_LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
OBJS += ggml-cuda.o
OBJS += ggml-cuda.o
NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
ifdef LLAMA_CUDA_NVCC
NVCC = $(LLAMA_CUDA_NVCC)
@ -497,11 +493,6 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
$(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_MPI
ifndef LLAMA_NO_K_QUANTS
k_quants.o: k_quants.c k_quants.h
$(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_NO_K_QUANTS
# combine build flags with cmdline overrides
override CFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
override CXXFLAGS := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
@ -542,7 +533,10 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
$(CC) $(CFLAGS) -c $< -o $@
OBJS += ggml-alloc.o ggml-backend.o
ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
$(CC) $(CFLAGS) -c $< -o $@
OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
$(CXX) $(CXXFLAGS) -c $< -o $@

3
Package.swift
View file

@ -42,13 +42,12 @@ let package = Package(
"llama.cpp",
"ggml-alloc.c",
"ggml-backend.c",
"k_quants.c",
"ggml-quants.c",
] + additionalSources,
resources: resources,
publicHeadersPath: "spm-headers",
cSettings: [
.unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
.define("GGML_USE_K_QUANTS"),
.define("GGML_USE_ACCELERATE")
// NOTE: NEW_LAPACK will required iOS version 16.4+
// We should consider add this in the future when we drop support for iOS 14

21
build.zig
View file

@ -116,15 +116,10 @@ pub fn build(b: *std.build.Builder) !void {
var make = try Maker.init(b);
make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
try make.addFlag("-DGGML_USE_K_QUANTS");
const k_quants = make.obj("k_quants", "k_quants.c");
try make.objs.append(k_quants);
}
const ggml = make.obj("ggml", "ggml.c");
const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
const llama = make.obj("llama", "llama.cpp");
const common = make.obj("common", "common/common.cpp");
const console = make.obj("console", "common/console.cpp");
@ -133,14 +128,14 @@ pub fn build(b: *std.build.Builder) !void {
const train = make.obj("train", "common/train.cpp");
const clip = make.obj("clip", "examples/llava/clip.cpp");
_ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
_ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
_ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
_ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
_ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
_ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
_ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, console, grammar_parser });
_ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
_ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
_ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
_ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
_ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser, clip });
const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip });
if (server.target.isWindows()) {
server.linkSystemLibrary("ws2_32");
}

9
examples/quantize/quantize.cpp
View file

@ -18,7 +18,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q4_1", LLAMA_FTYPE_MOSTLY_Q4_1, " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
{ "Q5_0", LLAMA_FTYPE_MOSTLY_Q5_0, " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
{ "Q5_1", LLAMA_FTYPE_MOSTLY_Q5_1, " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
#ifdef GGML_USE_K_QUANTS
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
{ "Q3_K", LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
@ -31,7 +30,6 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q5_K_S", LLAMA_FTYPE_MOSTLY_Q5_K_S, " 4.33G, +0.0400 ppl @ LLaMA-v1-7B", },
{ "Q5_K_M", LLAMA_FTYPE_MOSTLY_Q5_K_M, " 4.45G, +0.0122 ppl @ LLaMA-v1-7B", },
{ "Q6_K", LLAMA_FTYPE_MOSTLY_Q6_K, " 5.15G, -0.0008 ppl @ LLaMA-v1-7B", },
#endif
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
@ -70,13 +68,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
}
// usage:
// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
// ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
//
[[noreturn]]
static void usage(const char * executable) {
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
printf(" --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
printf("\nAllowed quantization types:\n");
for (auto & it : QUANT_OPTIONS) {
if (it.name != "COPY") {
@ -103,6 +102,8 @@ int main(int argc, char ** argv) {
params.quantize_output_tensor = false;
} else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
params.allow_requantize = true;
} else if (strcmp(argv[arg_idx], "--pure") == 0) {
params.pure = true;
} else {
usage(argv[0]);
}

2248
k_quants.c → ggml-quants.c
View file

File diff suppressed because it is too large Load diff

103
k_quants.h → ggml-quants.h
View file

@ -1,20 +1,14 @@
#pragma once
// This is a private API for quantization and dequantization
// Should not be used directly, use ggml.h instead
#include "ggml.h"
#include <stdint.h>
#include <assert.h>
#include <stddef.h>
// Super-block size
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
@ -23,10 +17,66 @@
#endif
#endif
#define QK4_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
typedef struct {
ggml_fp16_t d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
float d; // delta
float s; // d * sum(qs[i])
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
//
// Super-block quantization structures
//
// Super-block size
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
// 2-bit quantization
// weight is represented as x = a * q + b
// 16 blocks of 16 elements each
@ -127,6 +177,13 @@ static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_
// Quantization
void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
@ -134,6 +191,13 @@ void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict
void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
@ -142,6 +206,13 @@ void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
// Dequantization
void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
@ -150,16 +221,14 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int
void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
// Dot product
void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
// Quantization with histogram collection
size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);

2301
ggml.c
View file

File diff suppressed because it is too large Load diff

7
ggml.h
View file

@ -1930,12 +1930,19 @@ extern "C" {
// quantization
//
// TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
//

34
llama.cpp
View file

@ -19,13 +19,11 @@
#ifdef GGML_USE_MPI
# include "ggml-mpi.h"
#endif
#ifdef GGML_USE_K_QUANTS
# ifndef QK_K
# ifdef GGML_QKK_64
# define QK_K 64
# else
# define QK_K 256
# endif
#ifndef QK_K
# ifdef GGML_QKK_64
# define QK_K 64
# else
# define QK_K 256
# endif
#endif
@ -8052,7 +8050,7 @@ struct no_init {
struct quantize_state_internal {
const llama_model & model;
const llama_model_quantize_params * params;
#ifdef GGML_USE_K_QUANTS
int n_attention_wv = 0;
int n_feed_forward_w2 = 0;
int i_attention_wv = 0;
@ -8060,7 +8058,7 @@ struct quantize_state_internal {
int n_k_quantized = 0;
int n_fallback = 0;
#endif
quantize_state_internal(const llama_model & model, const llama_model_quantize_params * params)
: model(model)
, params(params)
@ -8125,7 +8123,6 @@ static void llama_convert_tensor_internal(
workers.clear();
}
#ifdef GGML_USE_K_QUANTS
static ggml_type get_k_quant_type(
quantize_state_internal & qs,
ggml_type new_type, const ggml_tensor * tensor, llama_ftype ftype
@ -8237,7 +8234,6 @@ static ggml_type get_k_quant_type(
return new_type;
}
#endif
static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
ggml_type quantized_type;
@ -8252,7 +8248,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_F16: quantized_type = GGML_TYPE_F16; break;
case LLAMA_FTYPE_ALL_F32: quantized_type = GGML_TYPE_F32; break;
#ifdef GGML_USE_K_QUANTS
// K-quants
case LLAMA_FTYPE_MOSTLY_Q2_K: quantized_type = GGML_TYPE_Q2_K; break;
case LLAMA_FTYPE_MOSTLY_Q3_K_S:
@ -8263,7 +8258,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
case LLAMA_FTYPE_MOSTLY_Q5_K_S:
case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
case LLAMA_FTYPE_MOSTLY_Q6_K: quantized_type = GGML_TYPE_Q6_K; break;
#endif
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
}
@ -8304,7 +8299,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
gguf_set_val_u32(ctx_out, "general.file_type", ftype);
#ifdef GGML_USE_K_QUANTS
for (int i = 0; i < ml.n_tensors; ++i) {
struct ggml_tensor * meta = ml.get_tensor_meta(i);
@ -8322,7 +8316,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
LLAMA_LOG_WARN("%s ============ Strange model: n_attention_wv = %d, n_feed_forward_w2 = %d, hparams.n_layer = %d\n",
__func__, qs.n_attention_wv, qs.n_feed_forward_w2, model.hparams.n_layer);
}
#endif
size_t total_size_org = 0;
size_t total_size_new = 0;
@ -8387,9 +8380,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (quantize) {
new_type = quantized_type;
#ifdef GGML_USE_K_QUANTS
new_type = get_k_quant_type(qs, new_type, tensor, ftype);
#endif
if (!params->pure) {
new_type = get_k_quant_type(qs, new_type, tensor, ftype);
}
// If we've decided to quantize to the same type the tensor is already
// in then there's nothing to do.
quantize = tensor->type != new_type;
@ -8514,12 +8508,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
LLAMA_LOG_INFO("\n");
}
}
#ifdef GGML_USE_K_QUANTS
if (qs.n_fallback > 0) {
LLAMA_LOG_WARN("%s: WARNING: %d of %d tensor(s) incompatible with k-quants and required fallback quantization\n",
__func__, qs.n_fallback, qs.n_k_quantized + qs.n_fallback);
}
#endif
}
static int llama_apply_lora_from_file_internal(
@ -8844,6 +8837,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
/*.allow_requantize =*/ false,
/*.quantize_output_tensor =*/ true,
/*.only_copy =*/ false,
/*.pure =*/ false,
};
return result;

1
llama.h
View file

@ -191,6 +191,7 @@ extern "C" {
bool allow_requantize; // allow quantizing non-f32/f16 tensors
bool quantize_output_tensor; // quantize output.weight
bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
bool pure; // disable k-quant mixtures and quantize all tensors to the same type
} llama_model_quantize_params;
// grammar types