root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-24 10:24:35 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

metal : opt-in compile flag for BF16 (#10218)
Some checks are pending
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-cuda.Dockerfile platforms:linux/amd64 tag:full-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full-musa.Dockerfile platforms:linux/amd64 tag:full-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/full.Dockerfile platforms:linux/amd64,linux/arm64 tag:full]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-cuda.Dockerfile platforms:linux/amd64 tag:light-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-intel.Dockerfile platforms:linux/amd64 tag:light-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli-musa.Dockerfile platforms:linux/amd64 tag:light-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-cli.Dockerfile platforms:linux/amd64,linux/arm64 tag:light]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-cuda.Dockerfile platforms:linux/amd64 tag:server-cuda]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-intel.Dockerfile platforms:linux/amd64 tag:server-intel]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server-musa.Dockerfile platforms:linux/amd64 tag:server-musa]) (push) Waiting to run
Details
Publish Docker image / Push Docker image to Docker Hub (map[dockerfile:.devops/llama-server.Dockerfile platforms:linux/amd64,linux/arm64 tag:server]) (push) Waiting to run
Details
Nix CI / nix-eval (macos-latest) (push) Waiting to run
Details
Nix CI / nix-eval (ubuntu-latest) (push) Waiting to run
Details
Nix CI / nix-build (macos-latest) (push) Waiting to run
Details
Nix CI / nix-build (ubuntu-latest) (push) Waiting to run
Details
flake8 Lint / Lint (push) Waiting to run
Details

Browse Source 
* metal : opt-in compile flag for BF16

ggml-ci

* ci : use BF16

ggml-ci

* swift : switch back to v12

* metal : has_float -> use_float

ggml-ci

* metal : fix BF16 check in MSL

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-11-08 21:59:46 +02:00 committed by GitHub
parent 695ad752b2
commit ec450d3bbf
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 77 additions and 42 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
.github/workflows/build.yml vendored
View File

@ -55,7 +55,13 @@ jobs:
sysctl -a
mkdir build
cd build
cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
cmake .. \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -113,7 +119,12 @@ jobs:
sysctl -a
# Metal is disabled due to intermittent failures with Github runners not having a GPU:
# https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
cmake -B build \
-DLLAMA_FATAL_WARNINGS=ON \
-DLLAMA_CURL=ON \
-DGGML_METAL=OFF \
-DGGML_RPC=ON \
-DBUILD_SHARED_LIBS=OFF
cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
- name: Test
@ -569,6 +580,7 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \
@ -599,6 +611,7 @@ jobs:
mkdir build
cd build
cmake -G Xcode .. \
-DGGML_METAL_USE_BF16=ON \
-DGGML_METAL_EMBED_LIBRARY=ON \
-DLLAMA_BUILD_EXAMPLES=OFF \
-DLLAMA_BUILD_TESTS=OFF \

4
Makefile
View File

@ -878,6 +878,10 @@ ifdef GGML_METAL
MK_CPPFLAGS += -DGGML_USE_METAL
MK_LDFLAGS += -framework Foundation -framework Metal -framework MetalKit
OBJ_GGML += ggml/src/ggml-metal.o
ifdef GGML_METAL_USE_BF16
MK_CPPFLAGS += -DGGML_METAL_USE_BF16
endif # GGML_METAL_USE_BF16
ifdef GGML_METAL_NDEBUG
MK_CPPFLAGS += -DGGML_METAL_NDEBUG
endif

2
ci/run.sh
View File

@ -39,7 +39,7 @@ SRC=`pwd`
CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
if [ ! -z ${GG_BUILD_METAL} ]; then
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
fi
if [ ! -z ${GG_BUILD_CUDA} ]; then

1
ggml/CMakeLists.txt
View File

@ -153,6 +153,7 @@ option(GGML_VULKAN_VALIDATE "ggml: enable Vulkan validation"
option(GGML_VULKAN_RUN_TESTS "ggml: run Vulkan tests" OFF)
option(GGML_KOMPUTE "ggml: use Kompute" OFF)
option(GGML_METAL "ggml: use Metal" ${GGML_METAL_DEFAULT})
option(GGML_METAL_USE_BF16 "ggml: use bfloat if available" OFF)
option(GGML_METAL_NDEBUG "ggml: disable Metal debugging" OFF)
option(GGML_METAL_SHADER_DEBUG "ggml: compile Metal with -fno-fast-math" OFF)
option(GGML_METAL_EMBED_LIBRARY "ggml: embed Metal library" ${GGML_METAL})

4
ggml/src/CMakeLists.txt
View File

@ -58,6 +58,10 @@ if (GGML_METAL)
add_compile_definitions(GGML_METAL_NDEBUG)
endif()
if (GGML_METAL_USE_BF16)
add_compile_definitions(GGML_METAL_USE_BF16)
endif()
# copy ggml-common.h and ggml-metal.metal to bin directory
configure_file(ggml-common.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h COPYONLY)
configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)

59
ggml/src/ggml-metal.m
View File

@ -39,6 +39,7 @@ static struct ggml_backend_metal_device_context {
bool has_simdgroup_reduction;
bool has_simdgroup_mm;
bool has_bfloat;
bool use_bfloat;
char name[128];
} g_ggml_ctx_dev_main = {
@ -47,6 +48,7 @@ static struct ggml_backend_metal_device_context {
/*.has_simdgroup_reduction =*/ false,
/*.has_simdgroup_mm =*/ false,
/*.has_bfloat =*/ false,
/*.use_bfloat =*/ false,
/*.name =*/ "",
};
@ -65,6 +67,12 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
ctx->has_bfloat = [ctx->mtl_device supportsFamily:MTLGPUFamilyMetal3_GGML];
ctx->has_bfloat |= [ctx->mtl_device supportsFamily:MTLGPUFamilyApple6];
#if defined(GGML_METAL_USE_BF16)
ctx->use_bfloat = ctx->has_bfloat;
#else
ctx->use_bfloat = false;
#endif
strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
}
@ -504,6 +512,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
// dictionary of preprocessor macros
NSMutableDictionary * prep = [NSMutableDictionary dictionary];
if (ctx_dev->use_bfloat) {
[prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"];
}
MTLCompileOptions * options = [MTLCompileOptions new];
options.preprocessorMacros = prep;
@ -556,7 +568,8 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_LOG_INFO("%s: simdgroup reduction = %s\n", __func__, ctx_dev->has_simdgroup_reduction ? "true" : "false");
GGML_LOG_INFO("%s: simdgroup matrix mul. = %s\n", __func__, ctx_dev->has_simdgroup_mm ? "true" : "false");
GGML_LOG_INFO("%s: bfloat = %s\n", __func__, ctx_dev->has_bfloat ? "true" : "false");
GGML_LOG_INFO("%s: has bfloat = %s\n", __func__, ctx_dev->has_bfloat ? "true" : "false");
GGML_LOG_INFO("%s: use bfloat = %s\n", __func__, ctx_dev->use_bfloat ? "true" : "false");
GGML_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx_dev->mtl_device.hasUnifiedMemory ? "true" : "false");
ctx->capture_next_compute = false;
@ -608,7 +621,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
const bool has_simdgroup_mm = ctx_dev->has_simdgroup_mm;
const bool has_simdgroup_reduction = ctx_dev->has_simdgroup_reduction;
const bool has_bfloat = ctx_dev->has_bfloat;
const bool use_bfloat = ctx_dev->use_bfloat;
// simd_sum and simd_max requires MTLGPUFamilyApple7
@ -644,7 +657,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8, diag_mask_inf_8, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32, get_rows_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16, get_rows_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_BF16, get_rows_bf16, use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0, get_rows_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1, get_rows_q4_1, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0, get_rows_q5_0, true);
@ -671,10 +684,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32, mul_mv_bf16_f32, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW, mul_mv_bf16_f32_1row, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4, mul_mv_bf16_f32_l4, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16, mul_mv_bf16_bf16, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW, mul_mv_f16_f32_1row, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4, mul_mv_f16_f32_l4, has_simdgroup_reduction);
@ -703,7 +716,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW, mul_mv_id_f16_f32_1row, has_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4, mul_mv_id_f16_f32_l4, has_simdgroup_reduction);
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_BF16_F32, mul_mv_id_bf16_f32, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32, mul_mv_id_q4_0_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32, mul_mv_id_q4_1_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32, mul_mv_id_q5_0_f32, has_simdgroup_reduction);
@ -725,7 +738,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32, mul_mm_bf16_f32, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32, mul_mm_q4_1_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32, mul_mm_q5_0_f32, has_simdgroup_mm);
@ -747,7 +760,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_BF16_F32, mul_mm_id_bf16_f32, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32, mul_mm_id_q4_1_f32, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32, mul_mm_id_q5_0_f32, has_simdgroup_mm);
@ -788,12 +801,12 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112, flash_attn_ext_f16_h112, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128, flash_attn_ext_f16_h128, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256, flash_attn_ext_f16_h256, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64, flash_attn_ext_bf16_h64, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80, flash_attn_ext_bf16_h80, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96, flash_attn_ext_bf16_h96, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112, flash_attn_ext_bf16_h112, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128, flash_attn_ext_bf16_h128, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256, flash_attn_ext_bf16_h256, has_simdgroup_mm && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H64, flash_attn_ext_bf16_h64, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H80, flash_attn_ext_bf16_h80, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H96, flash_attn_ext_bf16_h96, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H112, flash_attn_ext_bf16_h112, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H128, flash_attn_ext_bf16_h128, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_BF16_H256, flash_attn_ext_bf16_h256, has_simdgroup_mm && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H64, flash_attn_ext_q4_0_h64, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H80, flash_attn_ext_q4_0_h80, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q4_0_H96, flash_attn_ext_q4_0_h96, has_simdgroup_mm);
@ -825,14 +838,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H128, flash_attn_ext_q8_0_h128, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_Q8_0_H256, flash_attn_ext_q8_0_h256, has_simdgroup_mm);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128, flash_attn_ext_vec_f16_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, flash_attn_ext_vec_bf16_h128, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H128, flash_attn_ext_vec_bf16_h128, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H128, flash_attn_ext_vec_q4_0_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H128, flash_attn_ext_vec_q4_1_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H128, flash_attn_ext_vec_q5_0_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H128, flash_attn_ext_vec_q5_1_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H128, flash_attn_ext_vec_q8_0_h128, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, flash_attn_ext_vec_f16_h256, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256, flash_attn_ext_vec_bf16_h256, has_simdgroup_reduction && has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_BF16_H256, flash_attn_ext_vec_bf16_h256, has_simdgroup_reduction && use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_0_H256, flash_attn_ext_vec_q4_0_h256, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q4_1_H256, flash_attn_ext_vec_q4_1_h256, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256, flash_attn_ext_vec_q5_0_h256, has_simdgroup_reduction);
@ -840,11 +853,11 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256, flash_attn_ext_vec_q8_0_h256, has_simdgroup_reduction);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32, cpy_f32_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16, cpy_f32_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16, cpy_f32_bf16, use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32, cpy_f16_f32, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16, cpy_f16_f16, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, has_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_F32, cpy_bf16_f32, use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_BF16_BF16, cpy_bf16_bf16, use_bfloat);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0, cpy_f32_q8_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0, cpy_f32_q4_0, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1, cpy_f32_q4_1, true);
@ -936,9 +949,9 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_context * ctx_dev, const struct ggml_tensor * op) {
const bool has_simdgroup_mm = ctx_dev->has_simdgroup_mm;
const bool has_simdgroup_reduction = ctx_dev->has_simdgroup_reduction;
const bool has_bfloat = ctx_dev->has_bfloat;
const bool use_bfloat = ctx_dev->use_bfloat;
if (!has_bfloat) {
if (!use_bfloat) {
for (size_t i = 0, n = 3; i < n; ++i) {
if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
return false;

32
ggml/src/ggml-metal.metal
View File

@ -18,11 +18,11 @@ using namespace metal;
// .../usr/bin/metal -dM -E -c ggml/src/ggml-metal.metal
// .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal.metal
//
#if __METAL_VERSION__ < 310
#define GGML_METAL_NO_BFLOAT
#if __METAL_VERSION__ < 310 && defined(GGML_METAL_USE_BF16)
#undef GGML_METAL_USE_BF16
#endif
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
typedef matrix<bfloat, 4, 4> bfloat4x4;
#endif
@ -41,7 +41,7 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
reg = (type4x4)(*src);
}
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template <typename type4x4>
void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
reg = (type4x4)(*src);
@ -2082,7 +2082,7 @@ typedef decltype(kernel_mul_mv<half, half4, half, half4>) mul_mv_t;
template [[host_name("kernel_mul_mv_f32_f32")]] kernel mul_mv_t kernel_mul_mv<float, float4, float, float4>;
template [[host_name("kernel_mul_mv_f16_f32")]] kernel mul_mv_t kernel_mul_mv<half, half4, float, float4>;
template [[host_name("kernel_mul_mv_f16_f16")]] kernel mul_mv_t kernel_mul_mv<half, half4, half, half4>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mv_bf16_f32")]] kernel mul_mv_t kernel_mul_mv<bfloat, bfloat4, float, float4>;
template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<bfloat, bfloat4, bfloat, bfloat4>;
#endif
@ -2155,7 +2155,7 @@ kernel void kernel_mul_mv_1row(
typedef decltype(kernel_mul_mv_1row<half, half4>) mul_mv_1row_t;
template [[host_name("kernel_mul_mv_f16_f32_1row")]] kernel mul_mv_1row_t kernel_mul_mv_1row<half, half4>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mv_bf16_f32_1row")]] kernel mul_mv_1row_t kernel_mul_mv_1row<bfloat, bfloat4>;
#endif
@ -2217,7 +2217,7 @@ kernel void kernel_mul_mv_l4(
typedef decltype(kernel_mul_mv_l4<half, half4>) mul_mv_l4_t;
template [[host_name("kernel_mul_mv_f16_f32_l4")]] kernel mul_mv_l4_t kernel_mul_mv_l4<half, half4>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mv_bf16_f32_l4")]] kernel mul_mv_l4_t kernel_mul_mv_l4<bfloat, bfloat4>;
#endif
@ -3249,7 +3249,7 @@ template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_
template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 128>;
template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 256>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_bf16_h64" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 64>;
template [[host_name("kernel_flash_attn_ext_bf16_h80" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 80>;
template [[host_name("kernel_flash_attn_ext_bf16_h96" )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 96>;
@ -3648,7 +3648,7 @@ kernel void kernel_flash_attn_ext_vec(
typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64>) flash_attn_ext_vec_t;
template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 128>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 128>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 128>;
@ -3658,7 +3658,7 @@ template [[host_name("kernel_flash_attn_ext_vec_q5_1_h128")]] kernel flash_attn_
template [[host_name("kernel_flash_attn_ext_vec_q8_0_h128")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 128>;
template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 256>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_flash_attn_ext_vec_bf16_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, bfloat4x4, 1, dequantize_bf16, bfloat4x4, 1, dequantize_bf16, 256>;
#endif
template [[host_name("kernel_flash_attn_ext_vec_q4_0_h256")]] kernel flash_attn_ext_vec_t kernel_flash_attn_ext_vec<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 256>;
@ -3715,12 +3715,12 @@ typedef decltype(kernel_cpy<float, float>) kernel_cpy_t;
template [[host_name("kernel_cpy_f32_f32")]] kernel kernel_cpy_t kernel_cpy<float, float>;
template [[host_name("kernel_cpy_f32_f16")]] kernel kernel_cpy_t kernel_cpy<float, half>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_cpy_f32_bf16")]] kernel kernel_cpy_t kernel_cpy<float, bfloat>;
#endif
template [[host_name("kernel_cpy_f16_f32")]] kernel kernel_cpy_t kernel_cpy<half, float>;
template [[host_name("kernel_cpy_f16_f16")]] kernel kernel_cpy_t kernel_cpy<half, half>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_cpy_bf16_f32")]] kernel kernel_cpy_t kernel_cpy<bfloat, float>;
template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bfloat, bfloat>;
#endif
@ -6628,7 +6628,7 @@ typedef decltype(kernel_get_rows_f<float>) get_rows_f_t;
template [[host_name("kernel_get_rows_f32")]] kernel get_rows_f_t kernel_get_rows_f<float>;
template [[host_name("kernel_get_rows_f16")]] kernel get_rows_f_t kernel_get_rows_f<half>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_get_rows_bf16")]] kernel get_rows_f_t kernel_get_rows_f<bfloat>;
#endif
@ -6662,7 +6662,7 @@ typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, de
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, half4x4, 1, dequantize_f16>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mm_bf16_f32")]] kernel mat_mm_t kernel_mul_mm<bfloat, bfloat4x4, simdgroup_bfloat8x8, bfloat4x4, 1, dequantize_bf16>;
#endif
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<half, half4x4, simdgroup_half8x8, block_q4_0, 2, dequantize_q4_0>;
@ -6693,7 +6693,7 @@ typedef decltype(kernel_mul_mm_id<float4x4, 1, dequantize_f32>) mat_mm_id_t;
template [[host_name("kernel_mul_mm_id_f32_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_id_f16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<half4x4, 1, dequantize_f16>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mm_id_bf16_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<bfloat4x4, 1, dequantize_bf16>;
#endif
template [[host_name("kernel_mul_mm_id_q4_0_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0, 2, dequantize_q4_0>;
@ -6919,7 +6919,7 @@ typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<float, float4, float
template [[host_name("kernel_mul_mv_id_f32_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<float, float4, float, float4>>>;
template [[host_name("kernel_mul_mv_id_f16_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<half, half4, float, float4>>>;
#if !defined(GGML_METAL_NO_BFLOAT)
#if defined(GGML_METAL_USE_BF16)
template [[host_name("kernel_mul_mv_id_bf16_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<bfloat, bfloat4, float, float4>>>;
#endif
template [[host_name("kernel_mul_mv_id_q8_0_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl>>;