Merge branch 'master' into concedo_experimental

# Conflicts:
#	CMakeLists.txt
#	Makefile
#	README.md
#	ggml.c
This commit is contained in:
Concedo 2023-04-20 17:30:54 +08:00
commit 4605074245
5 changed files with 561 additions and 793 deletions

15
.gitignore vendored
View file

@ -1,11 +1,15 @@
*.o
*.a
.DS_Store
.build/
.cache/
.direnv/
.envrc
.swiftpm
.venv
.vs/
.vscode/
.DS_Store
.build/
build/
build-em/
build-debug/
@ -30,15 +34,12 @@ models/*
arm_neon.h
compile_commands.json
.envrc
.direnv/
.venv
__pycache__
.swiftpm
dist/
*.spec
zig-out/
zig-cache/
ppl-*.txt

View file

@ -1,3 +1,8 @@
default: koboldcpp koboldcpp_noavx2 koboldcpp_openblas koboldcpp_openblas_noavx2 koboldcpp_clblast
simple: koboldcpp koboldcpp_noavx2
dev: koboldcpp_openblas
ifndef UNAME_S
UNAME_S := $(shell uname -s)
endif
@ -117,6 +122,9 @@ endif
ifdef LLAMA_CUBLAS
CFLAGS += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
LDFLAGS += -lcublas_static -lculibos -lcudart_static -lcublasLt_static -lpthread -ldl -L/usr/local/cuda/lib64
OBJS += ggml-cuda.o
ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
nvcc -arch=native -c -o $@ $<
endif
ifdef LLAMA_GPROF
CFLAGS += -pg
@ -184,10 +192,6 @@ $(info I CC: $(CCV))
$(info I CXX: $(CXXV))
$(info )
default: koboldcpp koboldcpp_noavx2 koboldcpp_openblas koboldcpp_openblas_noavx2 koboldcpp_clblast
simple: koboldcpp koboldcpp_noavx2
dev: koboldcpp_openblas
#
# Build library
#
@ -234,7 +238,7 @@ gpttype_adapter.o: gpttype_adapter.cpp
clean:
rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize-stats perplexity embedding benchmark-q4_0-matmult main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_noavx2.dll koboldcpp_openblas_noavx2.dll koboldcpp_clblast.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_noavx2.so koboldcpp_openblas_noavx2.so koboldcpp_clblast.so gptj.exe gpt2.exe
main: examples/main/main.cpp ggml.o llama.o common.o
main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@echo
@echo '==== Run ./main -h for help. ===='
@ -258,32 +262,32 @@ koboldcpp_clblast: ggml_clblast.o ggml_rwkv.o ggml_v1.o expose.o common.o llama_
quantize_llama: examples/quantize/quantize.cpp ggml.o llama.o
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
quantize_gptj: ggml.o llama.o otherarch/tools/gptj_quantize.cpp
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
quantize_gpt2: ggml.o llama.o otherarch/tools/gpt2_quantize.cpp
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o
quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o
perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
vdot: pocs/vdot/vdot.cpp ggml.o
embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
libllama.so: llama.o ggml.o
vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
libllama.so: llama.o ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
#
# Tests
#
benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o
benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
./benchmark-q4_0-matmult

116
ggml-cuda.cu Normal file
View file

@ -0,0 +1,116 @@
#include <stdint.h>
#include <cuda_fp16.h>
#include "ggml-cuda.h"
typedef uint16_t ggml_fp16_t;
static_assert(sizeof(__half) == sizeof(ggml_fp16_t), "wrong fp16 size");
#define QK4_0 32
typedef struct {
float d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
float d; // delta
float m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK4_2 16
typedef struct {
__half d; // delta
uint8_t qs[QK4_2 / 2]; // nibbles / quants
} block_q4_2;
static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
const block_q4_0 * x = (const block_q4_0 *) vx;
const int i = blockIdx.x;
const float d = x[i].d;
const uint8_t * pp = x[i].qs;
for (int l = 0; l < QK4_0; l += 2) {
const uint8_t vi = pp[l/2];
const int8_t vi0 = vi & 0xf;
const int8_t vi1 = vi >> 4;
const float v0 = (vi0 - 8)*d;
const float v1 = (vi1 - 8)*d;
y[i*QK4_0 + l + 0] = v0;
y[i*QK4_0 + l + 1] = v1;
}
}
static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
const block_q4_1 * x = (const block_q4_1 *) vx;
const int i = blockIdx.x;
const float d = x[i].d;
const float m = x[i].m;
const uint8_t * pp = x[i].qs;
for (int l = 0; l < QK4_1; l += 2) {
const uint8_t vi = pp[l/2];
const int8_t vi0 = vi & 0xf;
const int8_t vi1 = vi >> 4;
const float v0 = vi0*d + m;
const float v1 = vi1*d + m;
y[i*QK4_1 + l + 0] = v0;
y[i*QK4_1 + l + 1] = v1;
}
}
static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
const block_q4_2 * x = (const block_q4_2 *) vx;
const int i = blockIdx.x;
const float d = x[i].d;
const uint8_t * pp = x[i].qs;
for (int l = 0; l < QK4_2; l += 2) {
const uint8_t vi = pp[l/2];
const int8_t vi0 = vi & 0xf;
const int8_t vi1 = vi >> 4;
const float v0 = (vi0 - 8)*d;
const float v1 = (vi1 - 8)*d;
y[i*QK4_2 + l + 0] = v0;
y[i*QK4_2 + l + 1] = v1;
}
}
extern "C" {
__host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_0;
dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
}
__host__ void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_1;
dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
}
__host__ void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
const int nb = k / QK4_2;
dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
}
}

11
ggml-cuda.h Normal file
View file

@ -0,0 +1,11 @@
#ifdef __cplusplus
extern "C" {
#endif
void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
#ifdef __cplusplus
}
#endif

1182
ggml.c

File diff suppressed because it is too large Load diff