ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-23 00:48:59 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

whisper : Metal and ggml-alloc support (#1270)

Browse Source 
* metal : init

* whisper : factor out graph builds

* whisper : allocate encoder and decoder using ggml-alloc

* whisper : ggml-alloc is now supported

* whisper : CoreML support ggml-alloc

* build : fix ggml-alloc

* ios : update submodule

* extra : update sync-ggml.sh script to also sync ggml-alloc

* ci : see if this is causing the crash

* whisper : refactor ggml-alloc init

* whisper.android : try to fix build

* whisper : initial Metal version

* ci : try to debug vmem issue

* metal : decoder works on GPU!

* metal : add multi-decoder support

* ggml : fix ggml_nbytes (probably temp solution)

* metal : run "cross" step on the GPU

* whisper : remove ggml_repeat in the encoder

* whisper : offload the Encoder to Metal

* ggml : use simpler ggml_bytes() implementation

* ggml-alloc : try to make CI happy by reducing vram to 128GB

* whisper : add whisper_allocr to wrap ggml_allocr

* whisper : factor out alloc init in a function

* cmake : update to support Metal build

* whisper : add <functional> header

* objc : fix build (no Metal yet)

* ios : add Metal support

* swiftui : fix build

* metal : speed-up KQ multiplication

* metal : sync latest llama.cpp kernels

* readme : add Metal info

* ios : update submodule

* coreml : add code to toggle Core ML config (CPU, ANE, GPU)

* bench : fix timings by running a pre-heat

* bench : start benching the decoder

* whisper : add ggml_mul_mat_pad

* bench : fix uninitialized vars

* whisper : add comment for disabling mul-mat padding

* whisper : add description of ggml_mul_mat_pad

* whisper : clean-up ggml_mul_mat_pad

* metal : remove the "concurrent" flag

* bench : variable n_past

* ios : update SPM package
This commit is contained in:
Georgi Gerganov
2023-09-15 12:18:18 +03:00
committed by GitHub
parent 3fec2119e6
commit 93935980f8
18 changed files with 1855 additions and 1252 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
examples/bench/bench.cpp
View File

@ -44,13 +44,13 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
fprintf(stderr, " -t N, --threads N [%-7d] number of threads to use during computation\n", params.n_threads);
fprintf(stderr, " -m FNAME, --model FNAME [%-7s] model path\n", params.model.c_str());
fprintf(stderr, " -w N, --what N [%-7d] what to benchmark:\n", params.what);
fprintf(stderr, " %-7s 0 - whisper encoder\n", "");
fprintf(stderr, " %-7s 0 - whisper\n", "");
fprintf(stderr, " %-7s 1 - memcpy\n", "");
fprintf(stderr, " %-7s 2 - ggml_mul_mat\n", "");
fprintf(stderr, "\n");
}
int whisper_bench_encoder(const whisper_params & params) {
int whisper_bench_full(const whisper_params & params) {
// whisper init
struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());
@ -69,12 +69,49 @@ int whisper_bench_encoder(const whisper_params & params) {
fprintf(stderr, "error: failed to set mel: %d\n", ret);
return 3;
}
// heat encoder
if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
whisper_token tokens[512];
memset(tokens, 0, sizeof(tokens));
// prompt heat
if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
// text-generation heat
if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
whisper_reset_timings(ctx);
// actual run
if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
for (int i = 0; i < 16; i++) {
if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
}
for (int i = 0; i < 256; i++) {
if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
fprintf(stderr, "error: failed to encode model: %d\n", ret);
return 4;
}
}
whisper_print_timings(ctx);
whisper_free(ctx);
@ -103,7 +140,7 @@ int main(int argc, char ** argv) {
int ret = -1;
switch (params.what) {
case 0: ret = whisper_bench_encoder(params); break;
case 0: ret = whisper_bench_full(params); break;
case 1: ret = whisper_bench_memcpy(params.n_threads); break;
case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;

2
examples/talk-llama/CMakeLists.txt
View File

@ -7,7 +7,7 @@ if (WHISPER_SDL2)
# TODO: this is temporary
# need to export ggml symbols for MSVC, but too lazy ..
add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../ggml-alloc.c ../../whisper.cpp)
target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

3
examples/whisper.android/app/src/main/jni/whisper/CMakeLists.txt
View File

@ -8,6 +8,7 @@ set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)
set(
SOURCE_FILES
${WHISPER_LIB_DIR}/ggml.c
${WHISPER_LIB_DIR}/ggml-alloc.c
${WHISPER_LIB_DIR}/whisper.cpp
${CMAKE_SOURCE_DIR}/jni.c
)
@ -20,7 +21,7 @@ function(build_library target_name)
SHARED
${SOURCE_FILES}
)
target_link_libraries(${target_name} ${LOG_LIB} android)
if (${target_name} STREQUAL "whisper_v8fp16_va")

12
examples/whisper.objc/README.md
View File

@ -28,6 +28,8 @@ This can significantly improve the performance of the transcription:
<img width="1072" alt="image" src="https://user-images.githubusercontent.com/1991296/208511239-8d7cdbd1-aa48-41b5-becd-ca288d53cc07.png">
## Core ML
If you want to enable Core ML support, you can add the `-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK` compiler flag for `whisper.cpp` in Build Phases:
<img width="1072" alt="image" src="https://github.com/ggerganov/whisper.cpp/assets/3001525/103e8f57-6eb6-490d-a60c-f6cf6c319324">
@ -35,3 +37,13 @@ If you want to enable Core ML support, you can add the `-DWHISPER_USE_COREML -DW
Then follow the [`Core ML support` section of readme](../../README.md#core-ml-support) for convert the model.
In this project, it also added `-O3 -DNDEBUG` to `Other C Flags`, but adding flags to app proj is not ideal in real world (applies to all C/C++ files), consider splitting xcodeproj in workspace in your own project.
## Metal
You can also enable Metal to make the inference run on the GPU of your device. This might or might not be more efficient
compared to Core ML depending on the model and device that you use.
To enable Metal, just add `-DGGML_USE_METAL` instead off the `-DWHISPER_USE_COREML` flag and you are ready.
This will make both the Encoder and the Decoder run on the GPU.
If you want to run the Encoder with Core ML and the Decoder with Metal then simply add both `-DWHISPER_USE_COREML -DGGML_USE_METAL` flags. That's all!

29
examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
View File

@ -7,6 +7,9 @@
objects = {
/* Begin PBXBuildFile section */
1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 184447182AB211A2007D6BFE /* ggml-alloc.c */; };
1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 1844471B2AB21655007D6BFE /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-framework Foundation -framework Metal -framework MetalKit -fno-objc-arc"; }; };
184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7A29052BDF00BD2A04 /* AppDelegate.m */; };
18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7D29052BDF00BD2A04 /* SceneDelegate.m */; };
18627C8129052BDF00BD2A04 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8029052BDF00BD2A04 /* ViewController.m */; };
@ -14,7 +17,7 @@
18627C8629052BE000BD2A04 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 18627C8529052BE000BD2A04 /* Assets.xcassets */; };
18627C8929052BE000BD2A04 /* LaunchScreen.storyboard in Resources */ = {isa = PBXBuildFile; fileRef = 18627C8729052BE000BD2A04 /* LaunchScreen.storyboard */; };
18627C8C29052BE000BD2A04 /* main.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8B29052BE000BD2A04 /* main.m */; };
18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK"; }; };
18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML"; }; };
18627C9629052C5800BD2A04 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9529052C5800BD2A04 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE"; }; };
18627C9B29052CFF00BD2A04 /* ggml-base.en.bin in Resources */ = {isa = PBXBuildFile; fileRef = 18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */; };
7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
@ -23,7 +26,24 @@
7FE3424F2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc in Resources */ = {isa = PBXBuildFile; fileRef = 7FE3424E2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc */; };
/* End PBXBuildFile section */
/* Begin PBXCopyFilesBuildPhase section */
184447202AB21B25007D6BFE /* CopyFiles */ = {
isa = PBXCopyFilesBuildPhase;
buildActionMask = 2147483647;
dstPath = "";
dstSubfolderSpec = 7;
files = (
184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */,
);
runOnlyForDeploymentPostprocessing = 0;
};
/* End PBXCopyFilesBuildPhase section */
/* Begin PBXFileReference section */
184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml-alloc.c"; sourceTree = "<group>"; };
184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml-alloc.h"; sourceTree = "<group>"; };
1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml-metal.m"; sourceTree = "<group>"; };
1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml-metal.metal"; sourceTree = "<group>"; };
18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@ -80,6 +100,10 @@
18627C7829052BDF00BD2A04 /* whisper.objc */ = {
isa = PBXGroup;
children = (
1844471D2AB2195F007D6BFE /* ggml-metal.metal */,
1844471B2AB21655007D6BFE /* ggml-metal.m */,
184447182AB211A2007D6BFE /* ggml-alloc.c */,
184447192AB211A2007D6BFE /* ggml-alloc.h */,
7FE3424E2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc */,
7FE342442A0C3FA20015A058 /* coreml */,
18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */,
@ -126,6 +150,7 @@
18627C7229052BDF00BD2A04 /* Sources */,
18627C7329052BDF00BD2A04 /* Frameworks */,
18627C7429052BDF00BD2A04 /* Resources */,
184447202AB21B25007D6BFE /* CopyFiles */,
);
buildRules = (
);
@ -194,8 +219,10 @@
18627C9629052C5800BD2A04 /* ggml.c in Sources */,
18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
18627C8C29052BE000BD2A04 /* main.m in Sources */,
18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,
7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */,
);
runOnlyForDeploymentPostprocessing = 0;

10
examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj
View File

@ -20,6 +20,7 @@
0AAC5DCC29539EB1003032C3 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DC929539EB0003032C3 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -Wno-shorten-64-to-32"; }; };
0AAC5DCE2953A05C003032C3 /* WhisperState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DCD2953A05C003032C3 /* WhisperState.swift */; };
0AAC5DD12953A394003032C3 /* LibWhisper.swift in Sources */ = {isa = PBXBuildFile; fileRef = 0AAC5DD02953A394003032C3 /* LibWhisper.swift */; };
18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 18AED47F2AB21F2B009D854F /* ggml-alloc.c */; };
/* End PBXBuildFile section */
/* Begin PBXFileReference section */
@ -41,6 +42,8 @@
0AAC5DCA29539EB0003032C3 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = ggml.h; sourceTree = "<group>"; };
0AAC5DCD2953A05C003032C3 /* WhisperState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = WhisperState.swift; sourceTree = "<group>"; };
0AAC5DD02953A394003032C3 /* LibWhisper.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibWhisper.swift; sourceTree = "<group>"; };
18AED47F2AB21F2B009D854F /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; path = "ggml-alloc.c"; sourceTree = "<group>"; };
18AED4802AB21F2B009D854F /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "ggml-alloc.h"; sourceTree = "<group>"; };
/* End PBXFileReference section */
/* Begin PBXFrameworksBuildPhase section */
@ -124,6 +127,8 @@
0AAC5DC529539E89003032C3 /* whisper.cpp */ = {
isa = PBXGroup;
children = (
18AED47F2AB21F2B009D854F /* ggml-alloc.c */,
18AED4802AB21F2B009D854F /* ggml-alloc.h */,
0AAC5DC929539EB0003032C3 /* ggml.c */,
0AAC5DCA29539EB0003032C3 /* ggml.h */,
0AAC5DC729539EB0003032C3 /* whisper.cpp */,
@ -242,6 +247,7 @@
0AA7514C2953B569001EE061 /* RiffWaveUtils.swift in Sources */,
0AAC5DCB29539EB1003032C3 /* whisper.cpp in Sources */,
0AA7514E2953D958001EE061 /* Recorder.swift in Sources */,
18AED4812AB21F2B009D854F /* ggml-alloc.c in Sources */,
);
runOnlyForDeploymentPostprocessing = 0;
};
@ -369,7 +375,7 @@
CODE_SIGN_STYLE = Automatic;
CURRENT_PROJECT_VERSION = 1;
DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
DEVELOPMENT_TEAM = 3TZ9BM962G;
DEVELOPMENT_TEAM = P8JZH34X63;
ENABLE_HARDENED_RUNTIME = YES;
ENABLE_PREVIEWS = YES;
GENERATE_INFOPLIST_FILE = YES;
@ -410,7 +416,7 @@
CODE_SIGN_STYLE = Automatic;
CURRENT_PROJECT_VERSION = 1;
DEVELOPMENT_ASSET_PATHS = "\"whisper.swiftui.demo/Supporting files/Preview Content\"";
DEVELOPMENT_TEAM = 3TZ9BM962G;
DEVELOPMENT_TEAM = P8JZH34X63;
ENABLE_HARDENED_RUNTIME = YES;
ENABLE_PREVIEWS = YES;
GENERATE_INFOPLIST_FILE = YES;