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

cmake : respect LLAMA_QKK_64 option

Browse source
This commit is contained in:
Georgi Gerganov 2023-12-31 13:34:07 +02:00
parent 049a32fffa
commit a8b9bb4566
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 11 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

9
CMakeLists.txt
View file

@ -182,10 +182,15 @@ if (LLAMA_METAL)
# note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
# disabling fast math is needed in order to pass tests/test-backend-ops # disabling fast math is needed in order to pass tests/test-backend-ops
# note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1 # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
set(XC_FLAGS -fno-fast-math -fno-inline -g)
if (LLAMA_QKK_64)
set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
endif()
add_custom_command( add_custom_command(
OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
COMMAND xcrun -sdk macosx metal -fno-fast-math -fno-inline -g -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
DEPENDS ggml-metal.metal DEPENDS ggml-metal.metal
COMMENT "Compiling Metal kernels" COMMENT "Compiling Metal kernels"
) )

8
tests/test-backend-ops.cpp
View file

@ -15,19 +15,18 @@
#include <thread> #include <thread>
#include <vector> #include <vector>
static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) { static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float max = 1.0f) {
size_t size = ggml_nelements(tensor); size_t size = ggml_nelements(tensor);
std::vector<float> data(size); std::vector<float> data(size);
#if 0 #if 0
std::default_random_engine generator(rd()); static std::default_random_engine generator(1234);
std::uniform_real_distribution<float> distribution(min, max); std::uniform_real_distribution<float> distribution(min, max);
for (size_t i = 0; i < size; i++) { for (size_t i = 0; i < size; i++) {
data[i] = distribution(generator); data[i] = distribution(generator);
} }
#endif #else
auto init_thread = [&](size_t start, size_t end) { auto init_thread = [&](size_t start, size_t end) {
std::random_device rd; std::random_device rd;
std::default_random_engine generator(rd()); std::default_random_engine generator(rd());
@ -49,6 +48,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
for (auto & t : threads) { for (auto & t : threads) {
t.join(); t.join();
} }
#endif
if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) { if (tensor->type == GGML_TYPE_F32 || tensor->type == GGML_TYPE_I32) {
ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float)); ggml_backend_tensor_set(tensor, data.data(), 0, size * sizeof(float));
@ -437,7 +437,7 @@ struct test_case {
double err = nmse(f1.data(), f2.data(), f1.size()); double err = nmse(f1.data(), f2.data(), f1.size());
if (err > ud->max_err) { if (err > ud->max_err) {
printf("[%s] NMSE = %f ", ggml_op_desc(t1), err); printf("[%s] NMSE = %f ", ggml_op_desc(t1), err);
//for (int i = 0; i < f1.size(); i++) { //for (int i = 0; i < (int) f1.size(); i++) {
// printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]); // printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
//} //}
//printf("\n"); //printf("\n");