Skip to content

Commit 05c55e5

Browse files
Djip007Djip007
committed
try to refactor include and simd
- correct local CI.
1 parent f35ea1c commit 05c55e5

File tree

6 files changed

+66
-47
lines changed

6 files changed

+66
-47
lines changed

‎Makefile

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -556,6 +556,7 @@ ifndef GGML_NO_OPENMP_SIMD
556556
MK_CPPFLAGS += -DGGML_USE_OPENMP_SIMD
557557
MK_CFLAGS += -fopenmp-simd
558558
MK_CXXFLAGS += -fopenmp-simd
559+
# -openmp:experimental pour MSVC?
559560
endif # GGML_NO_OPENMP_SIMD
560561

561562
ifdef GGML_OPENBLAS
@@ -1085,6 +1086,7 @@ ggml/src/ggml-aarch64.o: \
10851086
ggml/src/ggml-common.h
10861087
$(CC) $(CFLAGS) -c $< -o $@
10871088

1089+
#$(CXX) $(CXXFLAGS) -std=c++17 -fopt-info-vec-missed -c $< -o $@
10881090
ggml/src/ggml-fp8.o: \
10891091
ggml/src/ggml-fp8.cpp \
10901092
ggml/src/ggml-fp8.h \

‎ggml/src/CMakeLists.txt

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -154,7 +154,7 @@ if (GGML_MUSA)
154154
endif()
155155

156156
if (GGML_OPENMP)
157-
set(OpenMP_RUNTIME_MSVC "experimental")
157+
# set(OpenMP_RUNTIME_MSVC "experimental")
158158
find_package(OpenMP)
159159
if (OpenMP_FOUND)
160160
message(STATUS "OpenMP found")

‎ggml/src/ggml-common.h

Lines changed: 18 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -427,8 +427,24 @@ typedef struct {
427427
} block_iq4_xs;
428428
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
429429

430-
// the fp8 types.
431-
#include "ggml-fp8.h"
430+
// fp8 support
431+
// - fp8 simple type
432+
typedef struct { uint8_t bits; } ggml_e5m2_t;
433+
typedef struct { uint8_t bits; } ggml_e4m3_t;
434+
typedef struct { uint8_t bits; } ggml_e3m4_t;
435+
436+
// - fp8 with bloc delta => 8.125 bpw
437+
typedef struct {
438+
float d; // delta
439+
uint8_t qs[QK_K];
440+
} block_e4m3_q;
441+
static_assert(sizeof(block_e4m3_q) == sizeof(float) + QK_K, "wrong block_e4m3_q block size/padding");
442+
443+
typedef struct {
444+
float d; // delta
445+
uint8_t qs[QK_K];
446+
} block_e3m4_q;
447+
static_assert(sizeof(block_e3m4_q) == sizeof(float) + QK_K, "wrong block_e3m4_q block size/padding");
432448

433449
#endif // GGML_COMMON_DECL
434450
#endif // GGML_COMMON_DECL

‎ggml/src/ggml-fp8.cpp

Lines changed: 42 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -1,8 +1,10 @@
1-
// #include <iostream>
2-
#include <cstdint>
31
#include <cassert>
42
#include <algorithm>
53

4+
#define GGML_COMMON_DECL_CPP
5+
#include "ggml-common.h"
6+
#include "ggml.h"
7+
68
#include "ggml-fp8.h"
79

810
/*
@@ -20,6 +22,11 @@ cmake --build build --config Release
2022
./build/bin/llama-cli -c 1024 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf -p "[INST]bonjour a tu un nom. je ne sais pas comment t'appeler. Si tu n'en as pas je peux t'appeler TINTIN[/INST]" -s 42
2123
./build/bin/llama-perplexity --kl-divergence-base ~/LLM/Mistral-Nemo-Instruct-2407.BF16.kld --kl-divergence -s 31337 -m ~/LLM/Mistral-Nemo-Instruct-2407.E3M4_Q.gguf
2224
25+
# la CI local:
26+
rm -rf tmp
27+
mkdir tmp
28+
bash ./ci/run.sh ./tmp/results ./tmp/mnt
29+
2330
*/
2431

2532
template<int N> constexpr float EXP2() {
@@ -45,7 +52,9 @@ struct FP8 {
4552
static constexpr float MIN() { return EXP2<-M>()*EXP2<2-EXP_I2<_E-1>()>(); }
4653
//=============================================
4754

55+
#ifdef GGML_USE_OPENMP_SIMD
4856
#pragma omp declare simd
57+
#endif
4958
void operator=(float value) {
5059
union {
5160
float f;
@@ -67,7 +76,9 @@ struct FP8 {
6776
}
6877
}
6978

79+
#ifdef GGML_USE_OPENMP_SIMD
7080
#pragma omp declare simd
81+
#endif
7182
operator float () const {
7283
union {
7384
float f;
@@ -84,23 +95,21 @@ struct FP8 {
8495
}
8596
};
8697

87-
// block_e4m3_q
88-
//typedef struct {
89-
// float d; // delta
90-
// ggml_e4m3 qs[QK_K];
91-
//} block_e4m3_q;
92-
9398
template<int E>
9499
static inline void conv(const FP8<E>* x, float* y, int64_t size) {
100+
#ifdef GGML_USE_OPENMP_SIMD
95101
#pragma omp simd
102+
#endif
96103
for (int64_t i=0; i<size; i++) {
97104
y[i] = (float) x[i];
98105
}
99106
}
100107

101108
template<int E>
102109
static inline void conv(const float* x, FP8<E>* y, int64_t size) {
110+
#ifdef GGML_USE_OPENMP_SIMD
103111
#pragma omp simd
112+
#endif
104113
for (int64_t i=0; i<size; i++) {
105114
y[i] = x[i];
106115
}
@@ -109,7 +118,9 @@ static inline void conv(const float* x, FP8<E>* y, int64_t size) {
109118
template<int E>
110119
static inline float dot(const FP8<E>* x, const float* y, int64_t size) {
111120
float z = 0;
121+
#ifdef GGML_USE_OPENMP_SIMD
112122
#pragma omp simd reduction(+:z)
123+
#endif
113124
for (int64_t i=0; i<size; i++) {
114125
z += ((float)x[i])*y[i];
115126
}
@@ -126,7 +137,9 @@ template <int E, int QK>
126137
static inline void conv(const bloc_fp8<E, QK>* x, float* y, int64_t size) {
127138
const auto qk_size = size / QK;
128139
for (int64_t q=0; q<qk_size; ++q) {
140+
#ifdef GGML_USE_OPENMP_SIMD
129141
#pragma omp simd
142+
#endif
130143
for (int64_t i=0; i<QK; i++) {
131144
y[q*QK+i] = ((float) x[q].qs[i])*(x[q]).d;
132145
}
@@ -138,13 +151,18 @@ static inline void conv(const float* x, bloc_fp8<E, QK>* y, int64_t size) {
138151
const auto qk_size = size / QK;
139152
for (int64_t q=0; q<qk_size; ++q) {
140153
float m = 0;
154+
// @ voir si c'est lui qui pose probleme et si c'est sur toutes les target
155+
#ifdef GGML_USE_OPENMP_SIMD
141156
#pragma omp simd reduction(max:m)
157+
#endif
142158
for (int64_t i=0; i<QK; i++) {
143159
m = std::max(std::abs(x[q*QK+i]),m);
144160
}
145161
const float D = FP8<E>::MAX()/m;
146162
y[q].d = m/FP8<E>::MAX();
163+
#ifdef GGML_USE_OPENMP_SIMD
147164
#pragma omp simd
165+
#endif
148166
for (int64_t i=0; i<QK; i++) {
149167
y[q].qs[i] = x[q*QK+i]*D;
150168
}
@@ -157,7 +175,9 @@ static inline float dot(const bloc_fp8<E, QK>* x, const float* y, int64_t size)
157175
const auto qk_size = size / QK;
158176
for (int64_t q=0; q<qk_size; ++q) {
159177
float z0 = 0;
178+
#ifdef GGML_USE_OPENMP_SIMD
160179
#pragma omp simd reduction(+:z0)
180+
#endif
161181
for (int64_t i=0; i<QK; i++) {
162182
z0 += ((float)x[q].qs[i])*y[q*QK+i];
163183
}
@@ -192,29 +212,29 @@ void ggml_fp32_to_e4m3_row_ref(const float * GGML_RESTRICT x, ggml_e4m3_t * GGML
192212
}
193213

194214
void dequantize_row_e4m3_q(const block_e4m3_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
195-
assert(k % FP8_QK == 0);
196-
conv(reinterpret_cast<const bloc_fp8<4, FP8_QK>*>(x), y, k);
215+
assert(k % QK_K == 0);
216+
conv(reinterpret_cast<const bloc_fp8<4, QK_K>*>(x), y, k);
197217
}
198218
void quantize_row_e4m3_q(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k) {
199-
assert(k % FP8_QK == 0);
200-
conv(x, reinterpret_cast<bloc_fp8<4, FP8_QK>*>(y), k);
219+
assert(k % QK_K == 0);
220+
conv(x, reinterpret_cast<bloc_fp8<4, QK_K>*>(y), k);
201221
}
202222
void quantize_row_e4m3_q_ref(const float * GGML_RESTRICT x, block_e4m3_q * GGML_RESTRICT y, int64_t k) {
203-
assert(k % FP8_QK == 0);
204-
conv(x, reinterpret_cast<bloc_fp8<4, FP8_QK>*>(y), k);
223+
assert(k % QK_K == 0);
224+
conv(x, reinterpret_cast<bloc_fp8<4, QK_K>*>(y), k);
205225
}
206226

207227
void dequantize_row_e3m4_q(const block_e3m4_q * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
208-
assert(k % FP8_QK == 0);
209-
conv(reinterpret_cast<const bloc_fp8<3, FP8_QK>*>(x), y, k);
228+
assert(k % QK_K == 0);
229+
conv(reinterpret_cast<const bloc_fp8<3, QK_K>*>(x), y, k);
210230
}
211231
void quantize_row_e3m4_q(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k) {
212-
assert(k % FP8_QK == 0);
213-
conv(x, reinterpret_cast<bloc_fp8<3, FP8_QK>*>(y), k);
232+
assert(k % QK_K == 0);
233+
conv(x, reinterpret_cast<bloc_fp8<3, QK_K>*>(y), k);
214234
}
215235
void quantize_row_e3m4_q_ref(const float * GGML_RESTRICT x, block_e3m4_q * GGML_RESTRICT y, int64_t k) {
216-
assert(k % FP8_QK == 0);
217-
conv(x, reinterpret_cast<bloc_fp8<3, FP8_QK>*>(y), k);
236+
assert(k % QK_K == 0);
237+
conv(x, reinterpret_cast<bloc_fp8<3, QK_K>*>(y), k);
218238
}
219239

220240
// the dot product for FP8 weight
@@ -242,7 +262,7 @@ void ggml_vec_dot_e4m3_q(int n, float * GGML_RESTRICT s, size_t bs, const block_
242262
GGML_UNUSED(bx);
243263
GGML_UNUSED(by);
244264
GGML_UNUSED(bs);
245-
*s = dot(reinterpret_cast<const bloc_fp8<4, FP8_QK>*>(vx), vy, n);
265+
*s = dot(reinterpret_cast<const bloc_fp8<4, QK_K>*>(vx), vy, n);
246266
}
247267

248268
void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_e3m4_q * GGML_RESTRICT vx, size_t bx, const float * GGML_RESTRICT vy, size_t by, int nrc) {
@@ -251,5 +271,5 @@ void ggml_vec_dot_e3m4_q(int n, float * GGML_RESTRICT s, size_t bs, const block_
251271
GGML_UNUSED(bx);
252272
GGML_UNUSED(by);
253273
GGML_UNUSED(bs);
254-
*s = dot(reinterpret_cast<const bloc_fp8<3, FP8_QK>*>(vx), vy, n);
274+
*s = dot(reinterpret_cast<const bloc_fp8<3, QK_K>*>(vx), vy, n);
255275
}

‎ggml/src/ggml-fp8.h

Lines changed: 2 additions & 21 deletions
Original file line numberDiff line numberDiff line change
@@ -1,28 +1,9 @@
1-
#pragma once
2-
#include "ggml.h"
3-
1+
// this is more a .inc.
42
#ifdef __cplusplus
53
extern "C" {
64
#endif
75

8-
#define FP8_QK 256
9-
10-
typedef struct { uint8_t bits; } ggml_e5m2_t;
11-
typedef struct { uint8_t bits; } ggml_e4m3_t;
12-
typedef struct { uint8_t bits; } ggml_e3m4_t;
13-
14-
// fp8 with bloc delta => 8.125 bpw
15-
typedef struct {
16-
float d; // delta
17-
uint8_t qs[FP8_QK];
18-
} block_e4m3_q;
19-
static_assert(sizeof(block_e4m3_q) == sizeof(float) + FP8_QK, "wrong block_e4m3_q block size/padding");
20-
21-
typedef struct {
22-
float d; // delta
23-
uint8_t qs[FP8_QK];
24-
} block_e3m4_q;
25-
static_assert(sizeof(block_e3m4_q) == sizeof(float) + FP8_QK, "wrong block_e3m4_q block size/padding");
6+
// Note: types are define in ggml-common.h
267

278
GGML_API void ggml_e5m2_to_fp32_row(const ggml_e5m2_t * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
289
GGML_API void ggml_fp32_to_e5m2_row(const float * GGML_RESTRICT x, ggml_e5m2_t * GGML_RESTRICT y, int64_t k);

‎scripts/hf.sh

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -26,7 +26,7 @@ function has_cmd {
2626
}
2727

2828
if has_cmd wget; then
29-
cmd="wget -q --show-progress -c -O %s/%s %s"
29+
cmd="wget -q -c -O %s/%s %s"
3030
elif has_cmd curl; then
3131
cmd="curl -C - -f --output-dir %s -o %s -L %s"
3232
else

0 commit comments

Comments
 (0)