Merge remote-tracking branch 'ggerganov/master'

* ggerganov/master: (86 commits)
  server : fix building and simplify lib deps on Windows (ggerganov#1772)
  talk-llama : sync llama.cpp
  talk-llama : llama.cpp
  sync : ggml
  metal : correctly set SIMD support flags on iOS (llama/4923)
  2-bit quantizations (llama/4897)
  scripts : sync-ggml-am.sh add option to skip commits
  talk-llama : sync llama.cpp
  sync : ggml
  examples : adapt to metal API
  ggml: cache sin/cos for RoPE (llama/4908)
  metal : remove old API (llama/4919)
  metal : disable log for loaded kernels (llama/4794)
  gguf : fix potential infinite for-loop (llama/4600)
  metal : refactor kernel loading code (llama/4794)
  CUDA: faster q8_0 -> f16 dequantization (llama/4895)
  talk-llama : add optional CLI arg to set the bot name (ggerganov#1764)
  examples : add python example for transcription (ggerganov#1744)
  whisper : load the model into multiple buffers of max size 1GB (ggerganov#1763)
  talk-llama : sync llama.cpp
  ...

.devops/main-cuda.Dockerfile

@@ -20,6 +20,10 @@ RUN apt-get update && \
     apt-get install -y build-essential \
     && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
+# Ref: https://stackoverflow.com/a/53464012
+ENV CUDA_MAIN_VERSION=12.3
+ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
+
 COPY .. .
 RUN make
 

.github/workflows/build.yml

@@ -223,9 +223,12 @@ jobs:
           - arch: Win32
             obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
             s2arc: x86
+            clblast: OFF
           - arch: x64
             obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
             s2arc: x64
+            clblast: ON
+            clver: 1.6.1
           - sdl2: ON
             s2ver: 2.28.5
 
@@ -252,13 +255,26 @@ jobs:
           7z x sdl2.zip
           echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
 
+      - name: Install OpenCL
+        if: matrix.clblast == 'ON'
+        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
+
+      - name: Fetch CLBlast and set CLBlast_DIR
+        if: matrix.clblast == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
+          7z x clblast.zip
+          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
+          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
+
       - name: Configure
         run: >
           cmake -S . -B ./build -A ${{ matrix.arch }}
           -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           -DWHISPER_OPENBLAS=${{ matrix.blas }}
           -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
           -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CLBLAST=${{ matrix.clblast }}
 
       - name: Build
         run: |
@@ -273,11 +289,15 @@ jobs:
         if: matrix.sdl2 == 'ON'
         run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
 
+      - name: Copy clblast.dll
+        if: matrix.clblast == 'ON'
+        run: copy "$env:CLBlast_DIR/../../clblast.dll" build/bin/${{ matrix.build }}
+
       - name: Upload binaries
         if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
         uses: actions/upload-artifact@v1
         with:
-          name: whisper-blas-bin-${{ matrix.arch }}
+          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
   windows-cublas:

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.5)
 
-project(whisper.cpp VERSION 1.5.2)
+project(whisper.cpp VERSION 1.5.4)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -223,11 +223,17 @@ if (WHISPER_CUBLAS)
         add_compile_definitions(GGML_USE_CUBLAS)
 
         if (WHISPER_STATIC)
-            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
         else()
             set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
     else()
         message(FATAL_ERROR "cuBLAS not found")
     endif()
@@ -343,8 +349,8 @@ else()
         endif()
     else()
         if (EMSCRIPTEN)
-            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
         else()
             if(NOT WHISPER_NO_AVX)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")

Makefile

@@ -99,6 +99,16 @@ ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$
 	CXXFLAGS += -pthread
 endif
 
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -206,7 +216,7 @@ ifdef WHISPER_CUBLAS
 
 	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
+	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
 	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
@@ -360,7 +370,7 @@ quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
 	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
 server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS) $(LWINSOCK2)
 
 stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
 	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)

Package.swift

@@ -13,9 +13,13 @@ let package = Package(
     products: [
         .library(name: "whisper", targets: ["whisper"]),
     ],
+    dependencies: [
+        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
+    ],
     targets: [
         .target(
             name: "whisper",
+            dependencies: ["ggml"],
             path: ".",
             exclude: [
                "bindings",
@@ -32,14 +36,8 @@ let package = Package(
                "Makefile"
             ],
             sources: [
-                "ggml.c",
                 "whisper.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m"
             ],
-            resources: [.process("ggml-metal.metal")],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),

README.md

@@ -6,7 +6,7 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.5.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 

bindings/go/params.go

@@ -123,6 +123,11 @@ func (p *Params) SetAudioCtx(n int) {
 	p.audio_ctx = C.int(n)
 }
 
+// Set initial prompt
+func (p *Params) SetInitialPrompt(prompt string) {
+	p.initial_prompt = C.CString(prompt)
+}
+
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS
 
@@ -147,6 +152,7 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
 	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
+	str += fmt.Sprintf(" initial_prompt=%s", C.GoString(p.initial_prompt))
 	if p.translate {
 		str += " translate"
 	}

bindings/go/pkg/whisper/context.go

@@ -130,6 +130,11 @@ func (context *context) SetAudioCtx(n uint) {
 	context.params.SetAudioCtx(int(n))
 }
 
+// Set initial prompt
+func (context *context) SetInitialPrompt(prompt string) {
+	context.params.SetInitialPrompt(prompt)
+}
+
 // ResetTimings resets the mode timings. Should be called before processing
 func (context *context) ResetTimings() {
 	context.model.ctx.Whisper_reset_timings()

bindings/go/pkg/whisper/interface.go

@@ -38,17 +38,18 @@ type Context interface {
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language
 
-	SetOffset(time.Duration)      // Set offset
-	SetDuration(time.Duration)    // Set duration
-	SetThreads(uint)              // Set number of threads to use
-	SetSpeedup(bool)              // Set speedup flag
-	SetSplitOnWord(bool)          // Set split on word flag
-	SetTokenThreshold(float32)    // Set timestamp token probability threshold
-	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
-	SetMaxSegmentLength(uint)     // Set max segment length in characters
-	SetTokenTimestamps(bool)      // Set token timestamps flag
-	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)
-	SetAudioCtx(uint)             // Set audio encoder context
+	SetOffset(time.Duration)        // Set offset
+	SetDuration(time.Duration)      // Set duration
+	SetThreads(uint)                // Set number of threads to use
+	SetSpeedup(bool)                // Set speedup flag
+	SetSplitOnWord(bool)            // Set split on word flag
+	SetTokenThreshold(float32)      // Set timestamp token probability threshold
+	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
+	SetMaxSegmentLength(uint)       // Set max segment length in characters
+	SetTokenTimestamps(bool)        // Set token timestamps flag
+	SetMaxTokensPerSegment(uint)    // Set max tokens per segment (0 = no limit)
+	SetAudioCtx(uint)               // Set audio encoder context
+	SetInitialPrompt(prompt string) // Set initial prompt
 
 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.5.2",
+  "version": "1.5.4",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/ggml-backend-impl.h

@@ -70,7 +70,7 @@ extern "C" {
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
         // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
         bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);

bindings/ruby/ext/ggml-backend.c

@@ -156,8 +156,8 @@ void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_
     backend->iface.graph_plan_compute(backend, plan);
 }
 
-void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    backend->iface.graph_compute(backend, cgraph);
+bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_compute(backend, cgraph);
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {

bindings/ruby/ext/ggml-backend.h

@@ -52,7 +52,7 @@ extern "C" {
 
     GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
     GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends

coreml/whisper-encoder.mm

@@ -24,9 +24,9 @@
 
     // select which device to run the Core ML model on
     MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
-    config.computeUnits = MLComputeUnitsCPUAndGPU;
+    // config.computeUnits = MLComputeUnitsCPUAndGPU;
     //config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
-    //config.computeUnits = MLComputeUnitsAll;
+    config.computeUnits = MLComputeUnitsAll;
 
     const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model configuration:config error:nil]);
 

examples/CMakeLists.txt

@@ -15,6 +15,10 @@ if (WHISPER_SDL2)
     message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()
 
+if (WHISPER_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
 # common
 
 set(TARGET common)

examples/common-ggml.cpp

@@ -62,6 +62,8 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:
+        case GGML_FTYPE_MOSTLY_IQ2_XS:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -182,7 +184,7 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_Q5_K:
                 case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements, hist_cur.data());
+                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], hist_cur.data(), nullptr);
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
@@ -191,6 +193,8 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_I32:
                 case GGML_TYPE_Q8_1:
                 case GGML_TYPE_Q8_K:
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ2_XS:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));