Changed package root to /opt/whisper.cpp

* Add windows build with cuBLAS

* Remove error causing lines for cuBLAS on Windows

.github/workflows/build.yml

@@ -236,6 +236,61 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
+  windows-cublas:
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [x64]
+        cublas: [ON]
+        sdl2: [ON]
+        include:
+          - arch: x64
+            s2arc: x64
+          - sdl2: ON
+            s2ver: 2.26.0
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v1
+
+      - name: Install CUDA Toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.10
+
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
+          7z x sdl2.zip
+          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
+
+      - name: Configure
+        run: >
+          cmake -S . -B ./build -A ${{ matrix.arch }}
+          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+          -DWHISPER_CUBLAS=1
+
+      - name: Build
+        run: |
+          cd ./build
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
+
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+
+      - name: Upload binaries
+        if: matrix.sdl2 == 'ON'
+        uses: actions/upload-artifact@v1
+        with:
+          name: whisper-cublas-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
+
   emscripten:
     runs-on: ubuntu-latest
 
@@ -265,3 +320,44 @@ jobs:
           popd
           emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           make
+
+  ios:
+    runs-on: macos-latest
+
+    strategy:
+      matrix:
+        build: [Release]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Configure
+        run: cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
+
+      - name: Build objc example
+        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
+
+      - name: Build swiftui example
+        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
+
+  android:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v1
+
+      - name: Install Java
+        uses: actions/setup-java@v3
+        with:
+          distribution: zulu
+          java-version: 17
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v2
+
+      - name: Build
+        run: |
+          cd examples/whisper.android
+          ./gradlew assembleRelease --no-daemon

.github/workflows/release-deb.yml

@@ -0,0 +1,68 @@
+name: release-deb
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  build:
+    runs-on: ubuntu-20.04
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Configure
+        run: |
+          set -x -e
+          VERSION=$(echo $GITHUB_REF | cut --delimiter=/ -f 3)
+          ID="whisper-cpp-small_${VERSION}_amd64"
+          
+          echo "PKG_VERSION=$VERSION"     >> $GITHUB_ENV
+          echo "PKG_ID=$ID"               >> $GITHUB_ENV
+
+      - name: Install deps
+        run: |
+          sudo apt install -y --no-install-recommends intel-mkl
+
+      - name: Build
+        run: |
+          cmake -S . -B build-mkl \
+            -DCMAKE_BUILD_TYPE=Release\
+            -DBUILD_SHARED_LIBS=0\
+            -DWHISPER_BLAS=1\
+            -DWHISPER_BLAS_VENDOR=Intel10_64lp
+          cd build-mkl
+          make
+          cd ..
+
+      - name: Create package tree
+        env:
+          GITHUB_REPO: ${{ github.repository }}
+        run: |
+          export ROOT=$PKG_ID/opt/whisper.cpp
+          mkdir -p $ROOT/bin
+          mkdir -p $ROOT/share
+          mkdir -p $PKG_ID/DEBIAN
+          
+          cp build-mkl/bin/main $ROOT/bin/whisper 
+          cp -r contrib/debian/control $PKG_ID/DEBIAN/
+          
+          echo "Version: $PKG_VERSION"          >> $PKG_ID/DEBIAN/control
+          echo "Vcs-Git: $GITHUB_REPO"          >> $PKG_ID/DEBIAN/control
+          echo "Vcs-Git-Commit: $GITHUB_SHA"    >> $PKG_ID/DEBIAN/control
+          
+          models/download-ggml-model.sh small
+          build-mkl/bin/quantize models/ggml-small.bin \
+            $ROOT/share/ggml-small-q5_1.bin q5_1
+
+      - name: Create deb package
+        run: |
+          mkdir artifacts
+          dpkg-deb --build --root-owner-group $PKG_ID
+
+      - name: Upload Release Asset
+        uses: xresloader/upload-to-github-release@v1
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          release_id: ${{ github.event.release.id }}
+          file: ${{ env.PKG_ID }}.deb

.gitignore

@@ -1,8 +1,11 @@
 *.o
 *.a
 .cache/
+.coreml/
+.test/
 .vs/
 .vscode/
+.idea/
 .DS_Store
 
 build/
@@ -10,14 +13,19 @@ build-em/
 build-debug/
 build-release/
 build-static/
+build-cublas/
+build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/
+cmake-build-debug/
 
 /main
 /stream
 /command
 /talk
+/talk-llama
 /bench
+/quantize
 
 arm_neon.h
 sync.sh
@@ -31,3 +39,7 @@ examples/whisper.objc/whisper.objc.xcodeproj/xcuserdata/
 examples/whisper.objc/whisper.objc.xcodeproj/project.xcworkspace/xcuserdata
 
 extra/bench-gg.txt
+
+models/*.mlmodel
+models/*.mlmodelc
+models/*.mlpackage

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required (VERSION 3.0)
 
-project(whisper.cpp VERSION 1.2.0)
+project(whisper.cpp VERSION 1.4.1)
 
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -35,30 +35,38 @@ endif()
 
 # options
 
-option(BUILD_SHARED_LIBS               "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
+option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})
 
-option(WHISPER_ALL_WARNINGS            "whisper: enable all compiler warnings"                   ON)
-option(WHISPER_ALL_WARNINGS_3RD_PARTY  "whisper: enable all compiler warnings in 3rd party libs" OFF)
+option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
+option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
 
-option(WHISPER_SANITIZE_THREAD         "whisper: enable thread sanitizer"    OFF)
-option(WHISPER_SANITIZE_ADDRESS        "whisper: enable address sanitizer"   OFF)
-option(WHISPER_SANITIZE_UNDEFINED      "whisper: enable undefined sanitizer" OFF)
+option(WHISPER_SANITIZE_THREAD        "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS       "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED     "whisper: enable undefined sanitizer" OFF)
 
-option(WHISPER_BUILD_TESTS             "whisper: build tests"    ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_EXAMPLES          "whisper: build examples" ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_TESTS            "whisper: build tests"    ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDALONE})
 
-option(WHISPER_SUPPORT_SDL2            "whisper: support for libSDL2" OFF)
+option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)
+
+option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)
 
 if (APPLE)
-    option(WHISPER_NO_ACCELERATE       "whisper: disable Accelerate framework" OFF)
-    option(WHISPER_NO_AVX              "whisper: disable AVX" OFF)
-    option(WHISPER_NO_AVX2             "whisper: disable AVX2" OFF)
-    option(WHISPER_NO_FMA              "whisper: disable FMA" OFF)
+    option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
+    option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
+    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
 else()
-    option(WHISPER_SUPPORT_OPENBLAS    "whisper: support for OpenBLAS" OFF)
+    option(WHISPER_BLAS                  "whisper: use BLAS libraries"  OFF)
+    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor" Generic)
+    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"     OFF)
+    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"  OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"         OFF)
 endif()
 
-option(WHISPER_PERF                    "whisper: enable perf timings" OFF)
+option(WHISPER_PERF "whisper: enable perf timings" OFF)
 
 # sanitizers
 
@@ -86,30 +94,104 @@ endif()
 
 find_package(Threads REQUIRED)
 
-# on APPLE - include Accelerate framework
-if (APPLE AND NOT WHISPER_NO_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
+# on APPLE
+if (APPLE)
+    # include Accelerate framework
+    if (NOT WHISPER_NO_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
 
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
-    else()
-        message(WARNING "Accelerate framework not found")
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
+        else()
+            message(WARNING "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (WHISPER_COREML)
+        find_library(FOUNDATION_FRAMEWORK Foundation)
+        find_library(COREML_FRAMEWORK CoreML)
+
+        if (COREML_FRAMEWORK)
+            message(STATUS "CoreML framework found")
+
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+        else()
+            message(WARNING "CoreML framework not found")
+        endif()
+
+        if (WHISPER_COREML_ALLOW_FALLBACK)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML_ALLOW_FALLBACK)
+        endif()
     endif()
 endif()
 
-if (WHISPER_SUPPORT_OPENBLAS)
-    find_library(OPENBLAS_LIB
-        NAMES openblas libopenblas
-        )
-    if (OPENBLAS_LIB)
-        message(STATUS "OpenBLAS found")
+if (WHISPER_OPENBLAS)
+    set(WHISPER_BLAS_VENDOR "OpenBLAS")
+    set(WHISPER_BLAS ON)
+endif()
 
-        set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${OPENBLAS_LIB})
-        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+if (WHISPER_BLAS)
+    if (WHISPER_STATIC)
+        set(BLA_STATIC 1)
     else()
-        message(WARNING "OpenBLAS not found")
+        set(BLA_STATIC 0)
+    endif ()
+    set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
+    set(BLA_SIZEOF_INTEGER 8)
+    find_package(BLAS)
+
+    if(BLAS_FOUND)
+        message(STATUS "BLAS compatible library found")
+        message(STATUS "Libraries ${BLAS_LIBRARIES}")
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+
+        include_directories(${BLAS_INCLUDE_DIRS})
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+    else()
+        message(WARNING "BLAS library was not found")
+    endif()
+endif ()
+
+if (WHISPER_CUBLAS)
+    cmake_minimum_required(VERSION 3.17)
+
+    find_package(CUDAToolkit)
+
+    if (CUDAToolkit_FOUND)
+        message(STATUS "cuBLAS found")
+        set(CMAKE_CUDA_COMPILER /usr/local/cuda/bin/nvcc)
+        enable_language(CUDA)
+
+        set(GGML_CUDA_SOURCES ggml-cuda.cu ggml-cuda.h)
+
+        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)
+        else()
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        endif()
+
+    else()
+        message(WARNING "cuBLAS not found")
+    endif()
+endif()
+
+if (WHISPER_CLBLAST)
+    find_package(CLBlast)
+    if (CLBlast_FOUND)
+        message(STATUS "CLBlast found")
+
+        set(GGML_OPENCL_SOURCES ggml-opencl.c ggml-opencl.h)
+
+        add_compile_definitions(GGML_USE_CLBLAST)
+
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
+    else()
+        message(WARNING "CLBlast not found")
     endif()
 endif()
 
@@ -172,7 +254,9 @@ else()
             if(NOT WHISPER_NO_FMA)
                 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
             endif()
-            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
         endif()
     endif()
 endif()
@@ -181,6 +265,33 @@ if (WHISPER_PERF)
     set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
 
+#
+# whisper.coreml - Core ML support
+#
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        )
+endif()
+
 #
 # whisper - this is the main library of the project
 #
@@ -190,6 +301,8 @@ set(TARGET whisper)
 add_library(${TARGET}
     ggml.h
     ggml.c
+    ${GGML_CUDA_SOURCES}
+    ${GGML_OPENCL_SOURCES}
     whisper.h
     whisper.cpp
     )
@@ -200,6 +313,10 @@ target_include_directories(${TARGET} PUBLIC
     .
     )
 
+if (WHISPER_COREML)
+    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
+endif()
+
 if (MSVC)
     target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
 
@@ -215,7 +332,19 @@ if (BUILD_SHARED_LIBS)
 
     target_compile_definitions(${TARGET} PUBLIC
         WHISPER_SHARED
+        GGML_SHARED
         )
+
+    target_compile_definitions(${TARGET} PRIVATE
+        WHISPER_BUILD
+        GGML_BUILD
+        )
+endif()
+
+if (GGML_CUDA_SOURCES)
+    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
+    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES OFF)
+    set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
 endif()
 
 if (EMSCRIPTEN)

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2022 Georgi Gerganov
+Copyright (c) 2023 Georgi Gerganov
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -1,3 +1,5 @@
+default: main bench quantize
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -30,10 +32,16 @@ endif
 # Compile flags
 #
 
-CFLAGS   = -I.              -O3 -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =
 
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq ($(wildcard /usr/include/musl/*),)
+	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+endif
+
 # OS specific
 # TODO: support Windows
 ifeq ($(UNAME_S),Linux)
@@ -71,10 +79,6 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 			CFLAGS += -mavx2
 		endif
 	else ifeq ($(UNAME_S),Linux)
-		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
 		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@@ -86,16 +90,17 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
+
+			AVX1_M := $(shell grep "avx " /proc/cpuinfo)
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
 		endif
 		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
 		ifneq (,$(findstring sse3,$(SSE3_M)))
 			CFLAGS += -msse3
 		endif
 	else ifeq ($(UNAME_S),Haiku)
-		AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
 		AVX2_M := $(shell sysinfo -cpu | grep "AVX2 ")
 		ifneq (,$(findstring avx2,$(AVX2_M)))
 			CFLAGS += -mavx2
@@ -107,6 +112,11 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
 		F16C_M := $(shell sysinfo -cpu | grep "F16C ")
 		ifneq (,$(findstring f16c,$(F16C_M)))
 			CFLAGS += -mf16c
+
+			AVX1_M := $(shell sysinfo -cpu | grep "AVX ")
+			ifneq (,$(findstring avx,$(AVX1_M)))
+				CFLAGS += -mavx
+			endif
 		endif
 	else
 		CFLAGS += -mfma -mf16c -mavx -mavx2
@@ -115,6 +125,7 @@ endif
 ifeq ($(UNAME_M),amd64)
 	CFLAGS += -mavx -mavx2 -mfma -mf16c
 endif
+
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
 	ifneq (,$(findstring POWER9,$(POWER9_M)))
@@ -125,6 +136,7 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 		CXXFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
 	endif
 endif
+
 ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
@@ -132,27 +144,71 @@ ifndef WHISPER_NO_ACCELERATE
 		LDFLAGS += -framework Accelerate
 	endif
 endif
+
+ifdef WHISPER_COREML
+	CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS  += -framework Foundation -framework CoreML
+
+ifdef WHISPER_COREML_ALLOW_FALLBACK
+	CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
+endif
+endif
+
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
 	LDFLAGS += -lopenblas
 endif
+
+ifdef WHISPER_CUBLAS
+	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
+	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
+	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	WHISPER_OBJ += ggml-cuda.o
+	NVCC        = nvcc
+	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=native
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	$(NVCC) $(NVCCFLAGS) $(CXXFLAGS) -Wno-pedantic -c $< -o $@
+endif
+
+ifdef WHISPER_CLBLAST
+	CFLAGS 		+= -DGGML_USE_CLBLAST
+	LDFLAGS	 	+= -lclblast -lOpenCL
+	WHISPER_OBJ	+= ggml-opencl.o
+	
+ggml-opencl.o: ggml-opencl.c ggml-opencl.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif
+
 ifdef WHISPER_GPROF
 	CFLAGS   += -pg
 	CXXFLAGS += -pg
 endif
+
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	CFLAGS   += -mcpu=native
+	CXXFLAGS += -mcpu=native
 endif
+
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, 2, 3
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	# 32-bit Raspberry Pi 1, 2, 3
+	CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
 endif
+
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 4
-	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	# 32-bit ARM, for example on Armbian or possibly raspbian
+	#CFLAGS   += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	#CXXFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+
+	# 64-bit ARM on 32-bit OS, use these (TODO: auto-detect 64-bit)
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
+
 ifneq ($(filter armv8%,$(UNAME_M)),)
 	# Raspberry Pi 4
-	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
+	CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
+	CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
 endif
 
 #
@@ -170,26 +226,36 @@ $(info I CC:       $(CCV))
 $(info I CXX:      $(CXXV))
 $(info )
 
-default: main
-
 #
 # Build library
 #
 
-ggml.o: ggml.c ggml.h
-	$(CC)  $(CFLAGS)   -c ggml.c -o ggml.o
+ggml.o: ggml.c ggml.h ggml-cuda.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
 
-whisper.o: whisper.cpp whisper.h
-	$(CXX) $(CXXFLAGS) -c whisper.cpp -o whisper.o
+whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-libwhisper.a: ggml.o whisper.o
-	$(AR) rcs libwhisper.a ggml.o whisper.o
+ifndef WHISPER_COREML
+WHISPER_OBJ += whisper.o
+else
+whisper-encoder.o: coreml/whisper-encoder.mm coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -c coreml/whisper-encoder.mm -o whisper-encoder.o
 
-libwhisper.so: ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o whisper.o $(LDFLAGS)
+whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-impl.h
+	$(CXX) -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
+
+WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
+endif
+
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)
+
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
 
 clean:
-	rm -f *.o main stream command talk bench libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize libwhisper.a libwhisper.so
 
 #
 # Examples
@@ -197,24 +263,30 @@ clean:
 
 CC_SDL=`sdl2-config --cflags --libs`
 
-SRC_COMMON = examples/common.cpp
+SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp
 
-main: examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o whisper.o -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h
 
-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o stream $(CC_SDL) $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
 
-command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o command $(CC_SDL) $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)
 
-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o whisper.o -o talk $(CC_SDL) $(LDFLAGS)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
 
-bench: examples/bench/bench.cpp ggml.o whisper.o
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o whisper.o -o bench $(LDFLAGS)
+command: examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
 
 #
 # Audio samples

README.md

@@ -1,21 +1,26 @@
 # whisper.cpp
 
+![whisper.cpp](https://user-images.githubusercontent.com/1991296/235238348-05d0f6a4-da44-4900-a1de-d0707e75b763.jpeg)
+
 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![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: [v1.2.0](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Beta: [v1.4.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.1) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [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:
 
 - Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via Arm Neon and Accelerate framework
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
+- [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
 - Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
 - Runs on the CPU
+- [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)
 
 Supported platforms:
@@ -58,7 +63,9 @@ the Accelerate framework utilizes the special-purpose AMX coprocessor available
 
 ## Quick start
 
-First, download one of the Whisper models converted in [ggml format](models). For example:
+First clone the repository.
+
+Then, download one of the Whisper models converted in [ggml format](models). For example:
 
 ```bash
 bash ./models/download-ggml-model.sh base.en
@@ -223,10 +230,116 @@ make large
 | medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
 | large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
 
+## Quantization
+
+`whisper.cpp` supports integer quantization of the Whisper `ggml` models.
+Quantized models require less memory and disk space and depending on the hardware can be processed more efficiently.
+
+Here are the steps for creating and using a quantized model:
+
+```bash
+# quantize a model with Q5_0 method
+make quantize
+./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
+
+# run the examples as usual, specifying the quantized model file
+./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
+```
+
+## Core ML support
+
+On Apple Silicon devices, the Encoder inference can be executed on the Apple Neural Engine (ANE) via Core ML. This can result in significant
+speed-up - more than x3 faster compared with CPU-only execution. Here are the instructions for generating a Core ML model and using it with `whisper.cpp`:
+
+- Install Python dependencies needed for the creation of the Core ML model:
+
+  ```bash
+  pip install ane_transformers
+  pip install openai-whisper
+  pip install coremltools
+  ```
+
+- Generate a Core ML model. For example, to generate a `base.en` model, use:
+
+  ```bash
+  ./models/generate-coreml-model.sh base.en
+  ```
+
+  This will generate the folder `models/ggml-base.en-encoder.mlmodelc`
+
+- Build `whisper.cpp` with Core ML support:
+
+  ```bash
+  # using Makefile
+  make clean
+  WHISPER_COREML=1 make -j
+
+  # using CMake
+  cd build
+  cmake -DWHISPER_COREML=1 ..
+  ```
+
+- Run the examples as usual. For example:
+
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+
+  ...
+
+  whisper_init_state: loading Core ML model from 'models/ggml-base.en-encoder.mlmodelc'
+  whisper_init_state: first run on a device may take a while ...
+  whisper_init_state: Core ML model loaded
+
+  system_info: n_threads = 4 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | VSX = 0 | COREML = 1 |
+
+  ...
+  ```
+
+  The first run on a device is slow, since the ANE service compiles the Core ML model to some device-specific format.
+  Next runs are faster.
+
+For more information about the Core ML implementation please refer to PR [#566](https://github.com/ggerganov/whisper.cpp/pull/566).
+
+## NVIDIA GPU support via cuBLAS
+
+With NVIDIA cards, the Encoder processing can be offloaded to the GPU to a large extend through cuBLAS.
+First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads
+
+Now build `whisper.cpp` with cuBLAS support:
+
+```
+make clean
+WHISPER_CUBLAS=1 make -j
+```
+
+## OpenCL GPU support via CLBlast
+
+For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APU's or low end devices for up to ~2x speedup.
+
+First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
+
+Now build `whisper.cpp` with CLBlast support:
+
+```
+Makefile:
+cd whisper.cpp
+make clean
+WHISPER_CLBLAST=1 make -j
+
+CMake:
+cd whisper.cpp ; mkdir build ; cd build
+cmake -DWHISPER_CLBLAST=ON  ..
+make clean
+make -j
+cp bin/* ../ 
+```
+
+
+Run all the examples as usual.
+
 ## Limitations
 
 - Inference only
-- No GPU support (yet)
 
 ## Another example
 
@@ -313,7 +426,7 @@ whisper_print_timings:    total time = 32733.52 ms
 ## Real-time audio input example
 
 This is a naive example of performing real-time inference on audio from your microphone.
-The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continously.
+The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```java
@@ -328,6 +441,10 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:
 
+```java
+./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
+```
+
 <img width="965" alt="image" src="https://user-images.githubusercontent.com/1991296/197356445-311c8643-9397-4e5e-b46e-0b4b4daa2530.png">
 
 ## Controlling the length of the generated text segments (experimental)
@@ -367,7 +484,7 @@ system_info: n_threads = 4 / 10 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1
 
 main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
 
-[00:00:00.000 --> 00:00:00.320]  
+[00:00:00.000 --> 00:00:00.320]
 [00:00:00.320 --> 00:00:00.370]   And
 [00:00:00.370 --> 00:00:00.690]   so
 [00:00:00.690 --> 00:00:00.850]   my
@@ -433,6 +550,19 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a
 
 ---
 
+## Video comparison of different models
+
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:
+
+```java
+./extra/bench-wts.sh samples/jfk.wav
+ffplay ./samples/jfk.wav.all.mp4
+```
+
+https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8c3b-eb9f9c39d6fc.mp4
+
+---
+
 ## Benchmarks
 
 In order to have an objective comparison of the performance of the inference across different system configurations,
@@ -453,7 +583,7 @@ The original models are converted to a custom binary format. This allows to pack
 You can download the converted models using the [models/download-ggml-model.sh](models/download-ggml-model.sh) script
 or manually from here:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
@@ -463,13 +593,19 @@ in [models](models).
 
 - [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
 - [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+  - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
 - [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
 - [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
 - [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
 - [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
   - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
   - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
-- [ ] Python: soon | [WIP](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
+  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
 
 ## Examples
 
@@ -483,6 +619,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
 | [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
 | [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
 | [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
 | [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
 | [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |

bindings/go/examples/go-model-download/main.go

@@ -17,9 +17,9 @@ import (
 // CONSTANTS
 
 const (
-	srcUrl  = "https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main" // The location of the models
-	srcExt  = ".bin"                                                               // Filename extension
-	bufSize = 1024 * 64                                                            // Size of the buffer used for downloading the model
+	srcUrl  = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main" // The location of the models
+	srcExt  = ".bin"                                                      // Filename extension
+	bufSize = 1024 * 64                                                   // Size of the buffer used for downloading the model
 )
 
 var (

bindings/go/params.go

@@ -105,6 +105,10 @@ func (p *Params) SetMaxSegmentLength(n int) {
 	p.max_len = C.int(n)
 }
 
+func (p *Params) SetTokenTimestamps(b bool) {
+	p.token_timestamps = toBool(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)

bindings/go/pkg/whisper/context.go

@@ -111,6 +111,11 @@ func (context *context) SetMaxSegmentLength(n uint) {
 	context.params.SetMaxSegmentLength(int(n))
 }
 
+// Set token timestamps flag
+func (context *context) SetTokenTimestamps(b bool) {
+	context.params.SetTokenTimestamps(b)
+}
+
 // Set max tokens per segment (0 = no limit)
 func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
@@ -280,10 +285,14 @@ func toSegment(ctx *whisper.Context, n int) Segment {
 func toTokens(ctx *whisper.Context, n int) []Token {
 	result := make([]Token, ctx.Whisper_full_n_tokens(n))
 	for i := 0; i < len(result); i++ {
+		data := ctx.Whisper_full_get_token_data(n, i)
+
 		result[i] = Token{
-			Id:   int(ctx.Whisper_full_get_token_id(n, i)),
-			Text: strings.TrimSpace(ctx.Whisper_full_get_token_text(n, i)),
-			P:    ctx.Whisper_full_get_token_p(n, i),
+			Id:    int(ctx.Whisper_full_get_token_id(n, i)),
+			Text:  ctx.Whisper_full_get_token_text(n, i),
+			P:     ctx.Whisper_full_get_token_p(n, i),
+			Start: time.Duration(data.T0()) * time.Millisecond * 10,
+			End:   time.Duration(data.T1()) * time.Millisecond * 10,
 		}
 	}
 	return result

bindings/go/pkg/whisper/interface.go

@@ -41,6 +41,7 @@ type Context interface {
 	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)
 
 	// Process mono audio data and return any errors.
@@ -85,7 +86,8 @@ type Segment struct {
 
 // Token is a text or special token
 type Token struct {
-	Id   int
-	Text string
-	P    float32
+	Id         int
+	Text       string
+	P          float32
+	Start, End time.Duration
 }

bindings/go/pkg/whisper/model.go

@@ -94,6 +94,7 @@ func (model *model) NewContext() (Context, error) {
 	params.SetPrintRealtime(false)
 	params.SetPrintTimestamps(false)
 	params.SetThreads(runtime.NumCPU())
+	params.SetNoContext(true)
 
 	// Return new context
 	return newContext(model, params)

bindings/go/whisper.go

@@ -20,7 +20,7 @@ extern bool callEncoderBegin(void* user_data);
 // Text segment callback
 // Called on every newly generated text segment
 // Use the whisper_full_...() functions to obtain the text segments
-static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void* user_data) {
+static void whisper_new_segment_cb(struct whisper_context* ctx, struct whisper_state* state, int n_new, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         callNewSegment(user_data, n_new);
     }
@@ -29,7 +29,7 @@ static void whisper_new_segment_cb(struct whisper_context* ctx, int n_new, void*
 // Encoder begin callback
 // If not NULL, called before the encoder starts
 // If it returns false, the computation is aborted
-static bool whisper_encoder_begin_cb(struct whisper_context* ctx, void* user_data) {
+static bool whisper_encoder_begin_cb(struct whisper_context* ctx, struct whisper_state* state, void* user_data) {
     if(user_data != NULL && ctx != NULL) {
         return callEncoderBegin(user_data);
     }
@@ -356,7 +356,7 @@ func (ctx *Context) Whisper_full_get_token_id(segment int, token int) Token {
 
 // Get token data for the specified token in the specified segment.
 // This contains probabilities, timestamps, etc.
-func (ctx *Context) whisper_full_get_token_data(segment int, token int) TokenData {
+func (ctx *Context) Whisper_full_get_token_data(segment int, token int) TokenData {
 	return TokenData(C.whisper_full_get_token_data((*C.struct_whisper_context)(ctx), C.int(segment), C.int(token)))
 }
 
@@ -407,3 +407,11 @@ func callEncoderBegin(user_data unsafe.Pointer) C.bool {
 	}
 	return true
 }
+
+func (t TokenData) T0() int64 {
+	return int64(t.t0)
+}
+
+func (t TokenData) T1() int64 {
+	return int64(t.t1)
+}

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.2.0",
+  "version": "1.4.1",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/ext/ruby_whisper.cpp

@@ -199,7 +199,7 @@ static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
   {
     static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+    rwp->params.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
       bool is_aborted = *(bool*)user_data;
       return !is_aborted;
     };

contrib/debian/control

@@ -0,0 +1,5 @@
+Package: whisper-small-cpp
+Architecture: amd64
+Maintainer: Alexey Kharlamov <alexey@kharlamov.biz>
+Description: Whisper Speech to Text Converter
+Depends: libc6 (>= 2.2.1), intel-mkl

coreml/whisper-decoder-impl.h

@@ -0,0 +1,146 @@
+//
+// whisper-decoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implInput : NSObject<MLFeatureProvider>
+
+/// token_data as 1 by 1 matrix of 32-bit integers
+@property (readwrite, nonatomic, strong) MLMultiArray * token_data;
+
+/// audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * audio_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_implOutput : NSObject<MLFeatureProvider>
+
+/// var_1346 as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * var_1346;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_decoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_decoder_implInput to predict from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param token_data as 1 by 1 matrix of 32-bit integers:
+    @param audio_data as 1 × 384 × 1 × 1500 4-dimensional array of floats:
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_decoder_implOutput
+*/
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_decoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the predictions as NSArray<whisper_decoder_implOutput *>
+*/
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-decoder-impl.m

@@ -0,0 +1,201 @@
+//
+// whisper-decoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-decoder-impl.h"
+
+@implementation whisper_decoder_implInput
+
+- (instancetype)initWithToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data {
+    self = [super init];
+    if (self) {
+        _token_data = token_data;
+        _audio_data = audio_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"token_data", @"audio_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"token_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.token_data];
+    }
+    if ([featureName isEqualToString:@"audio_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.audio_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_implOutput
+
+- (instancetype)initWithVar_1346:(MLMultiArray *)var_1346 {
+    self = [super init];
+    if (self) {
+        _var_1346 = var_1346;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"var_1346"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"var_1346"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.var_1346];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_decoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_decoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-decoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_decoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_decoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_decoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_decoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_decoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_decoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_decoder_impl *typedModel = [[whisper_decoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromFeatures:(whisper_decoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[outFeatures featureValueForName:@"var_1346"].multiArrayValue];
+}
+
+- (nullable whisper_decoder_implOutput *)predictionFromToken_data:(MLMultiArray *)token_data audio_data:(MLMultiArray *)audio_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_decoder_implInput *input_ = [[whisper_decoder_implInput alloc] initWithToken_data:token_data audio_data:audio_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_decoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_decoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_decoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_decoder_implOutput * result = [[whisper_decoder_implOutput alloc] initWithVar_1346:(MLMultiArray *)[resultProvider featureValueForName:@"var_1346"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder-impl.h

@@ -0,0 +1,142 @@
+//
+// whisper-encoder-impl.h
+//
+// This file was automatically generated and should not be edited.
+//
+
+#import <Foundation/Foundation.h>
+#import <CoreML/CoreML.h>
+#include <stdint.h>
+#include <os/log.h>
+
+NS_ASSUME_NONNULL_BEGIN
+
+
+/// Model Prediction Input Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implInput : NSObject<MLFeatureProvider>
+
+/// logmel_data as 1 × 80 × 3000 3-dimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * logmel_data;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Model Prediction Output Type
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_implOutput : NSObject<MLFeatureProvider>
+
+/// output as multidimensional array of floats
+@property (readwrite, nonatomic, strong) MLMultiArray * output;
+- (instancetype)init NS_UNAVAILABLE;
+- (instancetype)initWithOutput:(MLMultiArray *)output NS_DESIGNATED_INITIALIZER;
+
+@end
+
+
+/// Class for model loading and prediction
+API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((visibility("hidden")))
+@interface whisper_encoder_impl : NSObject
+@property (readonly, nonatomic, nullable) MLModel * model;
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle;
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model NS_DESIGNATED_INITIALIZER;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init;
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the standard interface
+    @param input an instance of whisper_encoder_implInput to predict from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Make a prediction using the convenience interface
+    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the prediction as whisper_encoder_implOutput
+*/
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+
+/**
+    Batch prediction
+    @param inputArray array of whisper_encoder_implInput instances to obtain predictions from
+    @param options prediction options
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+    @return the predictions as NSArray<whisper_encoder_implOutput *>
+*/
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error;
+@end
+
+NS_ASSUME_NONNULL_END

coreml/whisper-encoder-impl.m

@@ -0,0 +1,197 @@
+//
+// whisper-encoder-impl.m
+//
+// This file was automatically generated and should not be edited.
+//
+
+#if !__has_feature(objc_arc)
+#error This file must be compiled with automatic reference counting enabled (-fobjc-arc)
+#endif
+
+#import "whisper-encoder-impl.h"
+
+@implementation whisper_encoder_implInput
+
+- (instancetype)initWithLogmel_data:(MLMultiArray *)logmel_data {
+    self = [super init];
+    if (self) {
+        _logmel_data = logmel_data;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"logmel_data"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"logmel_data"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.logmel_data];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_implOutput
+
+- (instancetype)initWithOutput:(MLMultiArray *)output {
+    self = [super init];
+    if (self) {
+        _output = output;
+    }
+    return self;
+}
+
+- (NSSet<NSString *> *)featureNames {
+    return [NSSet setWithArray:@[@"output"]];
+}
+
+- (nullable MLFeatureValue *)featureValueForName:(NSString *)featureName {
+    if ([featureName isEqualToString:@"output"]) {
+        return [MLFeatureValue featureValueWithMultiArray:self.output];
+    }
+    return nil;
+}
+
+@end
+
+@implementation whisper_encoder_impl
+
+
+/**
+    URL of the underlying .mlmodelc directory.
+*/
++ (nullable NSURL *)URLOfModelInThisBundle {
+    NSString *assetPath = [[NSBundle bundleForClass:[self class]] pathForResource:@"whisper_encoder_impl" ofType:@"mlmodelc"];
+    if (nil == assetPath) { os_log_error(OS_LOG_DEFAULT, "Could not load whisper-encoder-impl.mlmodelc in the bundle resource"); return nil; }
+    return [NSURL fileURLWithPath:assetPath];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from an existing MLModel object.
+
+    Usually the application does not use this initializer unless it makes a subclass of whisper_encoder_impl.
+    Such application may want to use `-[MLModel initWithContentsOfURL:configuration:error:]` and `+URLOfModelInThisBundle` to create a MLModel object to pass-in.
+*/
+- (instancetype)initWithMLModel:(MLModel *)model {
+    self = [super init];
+    if (!self) { return nil; }
+    _model = model;
+    if (_model == nil) { return nil; }
+    return self;
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+*/
+- (nullable instancetype)init {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle error:nil];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance with the model in this bundle.
+
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithConfiguration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self initWithContentsOfURL:(NSURL * _Nonnull)self.class.URLOfModelInThisBundle configuration:configuration error:error];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Initialize whisper_encoder_impl instance from the model URL.
+
+    @param modelURL URL to the .mlmodelc directory for whisper_encoder_impl.
+    @param configuration The model configuration object
+    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
+*/
+- (nullable instancetype)initWithContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    MLModel *model = [MLModel modelWithContentsOfURL:modelURL configuration:configuration error:error];
+    if (model == nil) { return nil; }
+    return [self initWithMLModel:model];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with configuration.
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadWithConfiguration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [self loadContentsOfURL:(NSURL * _Nonnull)[self URLOfModelInThisBundle]
+              configuration:configuration
+          completionHandler:handler];
+}
+
+
+/**
+    Construct whisper_encoder_impl instance asynchronously with URL of .mlmodelc directory and optional configuration.
+
+    Model loading may take time when the model content is not immediately available (e.g. encrypted model). Use this factory method especially when the caller is on the main thread.
+
+    @param modelURL The model URL.
+    @param configuration The model configuration
+    @param handler When the model load completes successfully or unsuccessfully, the completion handler is invoked with a valid whisper_encoder_impl instance or NSError object.
+*/
++ (void)loadContentsOfURL:(NSURL *)modelURL configuration:(MLModelConfiguration *)configuration completionHandler:(void (^)(whisper_encoder_impl * _Nullable model, NSError * _Nullable error))handler {
+    [MLModel loadContentsOfURL:modelURL
+                 configuration:configuration
+             completionHandler:^(MLModel *model, NSError *error) {
+        if (model != nil) {
+            whisper_encoder_impl *typedModel = [[whisper_encoder_impl alloc] initWithMLModel:model];
+            handler(typedModel, nil);
+        } else {
+            handler(nil, error);
+        }
+    }];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    return [self predictionFromFeatures:input options:[[MLPredictionOptions alloc] init] error:error];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromFeatures:(whisper_encoder_implInput *)input options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLFeatureProvider> outFeatures = [self.model predictionFromFeatures:input options:options error:error];
+    if (!outFeatures) { return nil; }
+    return [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[outFeatures featureValueForName:@"output"].multiArrayValue];
+}
+
+- (nullable whisper_encoder_implOutput *)predictionFromLogmel_data:(MLMultiArray *)logmel_data error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    whisper_encoder_implInput *input_ = [[whisper_encoder_implInput alloc] initWithLogmel_data:logmel_data];
+    return [self predictionFromFeatures:input_ error:error];
+}
+
+- (nullable NSArray<whisper_encoder_implOutput *> *)predictionsFromInputs:(NSArray<whisper_encoder_implInput*> *)inputArray options:(MLPredictionOptions *)options error:(NSError * _Nullable __autoreleasing * _Nullable)error {
+    id<MLBatchProvider> inBatch = [[MLArrayBatchProvider alloc] initWithFeatureProviderArray:inputArray];
+    id<MLBatchProvider> outBatch = [self.model predictionsFromBatch:inBatch options:options error:error];
+    if (!outBatch) { return nil; }
+    NSMutableArray<whisper_encoder_implOutput*> *results = [NSMutableArray arrayWithCapacity:(NSUInteger)outBatch.count];
+    for (NSInteger i = 0; i < outBatch.count; i++) {
+        id<MLFeatureProvider> resultProvider = [outBatch featuresAtIndex:i];
+        whisper_encoder_implOutput * result = [[whisper_encoder_implOutput alloc] initWithOutput:(MLMultiArray *)[resultProvider featureValueForName:@"output"].multiArrayValue];
+        [results addObject:result];
+    }
+    return results;
+}
+
+@end

coreml/whisper-encoder.h

@@ -0,0 +1,22 @@
+// Wrapper of the Core ML Whisper Encoder model
+//
+// Code is derived from the work of Github user @wangchou
+// ref: https://github.com/wangchou/callCoreMLFromCpp
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context;
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model);
+void whisper_coreml_free(struct whisper_coreml_context * ctx);
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out);
+
+#if __cplusplus
+}
+#endif

coreml/whisper-encoder.mm

@@ -0,0 +1,67 @@
+#import "coreml/whisper-encoder.h"
+#import "coreml/whisper-encoder-impl.h"
+
+#import <CoreML/CoreML.h>
+
+#include <stdlib.h>
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_coreml_context {
+    const void * data;
+};
+
+struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {
+    NSString * path_model_str = [[NSString alloc] initWithUTF8String:path_model];
+
+    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];
+
+    const void * data = CFBridgingRetain([[whisper_encoder_impl alloc] initWithContentsOfURL:url_model error:nil]);
+
+    if (data == NULL) {
+        return NULL;
+    }
+
+    whisper_coreml_context * ctx = new whisper_coreml_context;
+
+    ctx->data = data;
+
+    return ctx;
+}
+
+void whisper_coreml_free(struct whisper_coreml_context * ctx) {
+    CFRelease(ctx->data);
+    delete ctx;
+}
+
+void whisper_coreml_encode(
+        const whisper_coreml_context * ctx,
+                               float * mel,
+                               float * out) {
+    MLMultiArray * inMultiArray = [
+        [MLMultiArray alloc] initWithDataPointer: mel
+                                           shape: @[@1, @80, @3000]
+                                        dataType: MLMultiArrayDataTypeFloat32
+                                         strides: @[@(240000), @(3000), @1]
+                                     deallocator: nil
+                                           error: nil
+    ];
+
+    whisper_encoder_implOutput * outCoreML = [(__bridge id) ctx->data predictionFromLogmel_data:inMultiArray error:nil];
+
+    MLMultiArray * outMA = outCoreML.output;
+
+    //NSArray<NSNumber *> * shape = outMA.shape;
+    //NSArray<NSNumber *> * strides = outMA.strides;
+
+    //printf("shape:   %ld %ld %ld %ld\n", [shape[0] longValue], [shape[1] longValue], [shape[2] longValue], [shape[3] longValue]);
+    //printf("strides: %ld %ld %ld %ld\n", [strides[0] longValue], [strides[1] longValue], [strides[2] longValue], [strides[3] longValue]);
+
+    memcpy(out, outMA.dataPointer, outMA.count * sizeof(float));
+}
+
+#if __cplusplus
+}
+#endif

examples/CMakeLists.txt

@@ -4,7 +4,7 @@ find_package(Threads REQUIRED)
 
 # third-party
 
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # SDL2
     find_package(SDL2 REQUIRED)
 
@@ -21,13 +21,17 @@ set(TARGET common)
 add_library(${TARGET} STATIC
     common.h
     common.cpp
+    common-ggml.h
+    common-ggml.cpp
     )
 
 include(DefaultTargetOptions)
 
+target_link_libraries(${TARGET} PRIVATE whisper)
+
 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # common-sdl
 
     set(TARGET common-sdl)
@@ -62,5 +66,7 @@ else()
     add_subdirectory(stream)
     add_subdirectory(command)
     add_subdirectory(bench)
+    add_subdirectory(quantize)
     add_subdirectory(talk)
+    add_subdirectory(talk-llama)
 endif()

examples/addon.node/__test__/whisper.spec.js

@@ -1,15 +1,23 @@
-const path = require('path');
-const { whisper } = require(path.join(__dirname, '../../../build/Release/whisper-addon'));
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
 
 const whisperParamsMock = {
-    language: 'en',
-    model: path.join(__dirname, '../../../models/ggml-base.en.bin'),
-    fname_inp: path.join(__dirname, '../../../samples/jfk.wav'),
+  language: "en",
+  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
+  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
 };
 
 describe("Run whisper.node", () => {
+    test("it should receive a non-empty value", async () => {
+        let result = await whisperAsync(whisperParamsMock);
 
-    test("it should receive a non-empty value", () => {
-        expect(whisper(whisperParamsMock).length).toBeGreaterThan(0);
-    });
+        expect(result.length).toBeGreaterThan(0);
+    }, 10000);
 });
+

examples/addon.node/addon.cpp

@@ -72,7 +72,7 @@ int timestamp_to_sample(int64_t t, int n_samples) {
     return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
 }
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -160,22 +160,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
         const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -243,8 +227,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            wparams.initial_prompt   = params.prompt.c_str();
 
             whisper_print_user_data user_data = { &params, &pcmf32s };
 
@@ -260,7 +243,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -292,51 +275,64 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
     return 0;
 }
 
-Napi::Object whisper(const Napi::CallbackInfo& info) {
-    Napi::Env env = info.Env();
-    if (info.Length() <= 0 || !info[0].IsObject()) {
-        Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
-    }
-    whisper_params params;
-    std::vector<std::vector<std::string>> result;
+class Worker : public Napi::AsyncWorker {
+ public:
+  Worker(Napi::Function& callback, whisper_params params)
+      : Napi::AsyncWorker(callback), params(params) {}
 
-    Napi::Object whisper_params = info[0].As<Napi::Object>();
-    std::string language = whisper_params.Get("language").As<Napi::String>();
-    std::string model = whisper_params.Get("model").As<Napi::String>();
-    std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
-
-    params.language = language;
-    params.model = model;
-    params.fname_inp.emplace_back(input);
-
-    // run model
+  void Execute() override {
     run(params, result);
+  }
 
-    fprintf(stderr, "RESULT:\n");
-    for (auto sentence:result) {
-        fprintf(stderr, "t0: %s, t1: %s, content: %s \n",
-                sentence[0].c_str(), sentence[1].c_str(), sentence[2].c_str());
-    }
-
-    Napi::Object res = Napi::Array::New(env, result.size());
+  void OnOK() override {
+    Napi::HandleScope scope(Env());
+    Napi::Object res = Napi::Array::New(Env(), result.size());
     for (uint64_t i = 0; i < result.size(); ++i) {
-        Napi::Object tmp = Napi::Array::New(env, 3);
-        for (uint64_t j = 0; j < 3; ++j) {
-            tmp[j] = Napi::String::New(env, result[i][j]);
-        }
-        res[i] = tmp;
+      Napi::Object tmp = Napi::Array::New(Env(), 3);
+      for (uint64_t j = 0; j < 3; ++j) {
+        tmp[j] = Napi::String::New(Env(), result[i][j]);
+      }
+      res[i] = tmp;
     }
+    Callback().Call({Env().Null(), res});
+  }
 
-    return res;
+ private:
+  whisper_params params;
+  std::vector<std::vector<std::string>> result;
+};
+
+
+
+Napi::Value whisper(const Napi::CallbackInfo& info) {
+  Napi::Env env = info.Env();
+  if (info.Length() <= 0 || !info[0].IsObject()) {
+    Napi::TypeError::New(env, "object expected").ThrowAsJavaScriptException();
+  }
+  whisper_params params;
+
+  Napi::Object whisper_params = info[0].As<Napi::Object>();
+  std::string language = whisper_params.Get("language").As<Napi::String>();
+  std::string model = whisper_params.Get("model").As<Napi::String>();
+  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
+
+  params.language = language;
+  params.model = model;
+  params.fname_inp.emplace_back(input);
+
+  Napi::Function callback = info[1].As<Napi::Function>();
+  Worker* worker = new Worker(callback, params);
+  worker->Queue();
+  return env.Undefined();
 }
 
 
 Napi::Object Init(Napi::Env env, Napi::Object exports) {
-    exports.Set(
-            Napi::String::New(env, "whisper"),
-            Napi::Function::New(env, whisper)
-    );
-    return exports;
+  exports.Set(
+      Napi::String::New(env, "whisper"),
+      Napi::Function::New(env, whisper)
+  );
+  return exports;
 }
 
 NODE_API_MODULE(whisper, Init);

examples/addon.node/index.js

@@ -1,27 +1,36 @@
-const path = require('path');
-const { whisper } = require(path.join(__dirname, '../../build/Release/whisper-addon'));
+const path = require("path");
+const { whisper } = require(path.join(
+  __dirname,
+  "../../build/Release/whisper-addon"
+));
+const { promisify } = require("util");
+
+const whisperAsync = promisify(whisper);
 
 const whisperParams = {
-    language: 'en',
-    model: path.join(__dirname, '../../models/ggml-base.en.bin'),
-    fname_inp: '',
+  language: "en",
+  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
+  fname_inp: "../../samples/jfk.wav",
 };
 
 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
-    arguments.reduce((pre, item) => {
-        if (item.startsWith("--")) {
-            return [...pre, item.slice(2).split("=")];
-        }
-        return pre;
-    }, []),
+  arguments.reduce((pre, item) => {
+    if (item.startsWith("--")) {
+      return [...pre, item.slice(2).split("=")];
+    }
+    return pre;
+  }, [])
 );
 
 for (const key in params) {
-    if (whisperParams.hasOwnProperty(key)) {
-        whisperParams[key] = params[key];
-    }
+  if (whisperParams.hasOwnProperty(key)) {
+    whisperParams[key] = params[key];
+  }
 }
 
-console.log('whisperParams =', whisperParams);
-console.log(whisper(whisperParams));
+console.log("whisperParams =", whisperParams);
+
+whisperAsync(whisperParams).then((result) => {
+  console.log(`Result from whisper: ${result}`);
+});

examples/bench.wasm/CMakeLists.txt

@@ -31,9 +31,9 @@ endif()
 set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
     -s USE_PTHREADS=1 \
-    -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1024MB \
-    -s TOTAL_MEMORY=1024MB \
+    -s PTHREAD_POOL_SIZE_STRICT=0 \
+    -s INITIAL_MEMORY=2000MB \
+    -s TOTAL_MEMORY=2000MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/bench.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use and click the "Bench" button.<br>
@@ -44,11 +53,18 @@
 
             <div id="model-whisper">
                 Whisper model: <span id="model-whisper-status"></span>
-                <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
-                <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
-                <span id="fetch-whisper-progress"></span>
-
+                <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
+                <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
                 <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
+                <button id="fetch-whisper-small-en-q5_1"  onclick="loadWhisper('small-en-q5_1')">small.en (Q5_1, 182 MB)</button>
+                <button id="fetch-whisper-medium-en-q5_0" onclick="loadWhisper('medium-en-q5_0')">medium.en (Q5_0, 515 MB)</button>
+                <button id="fetch-whisper-large-q5_0"     onclick="loadWhisper('large-q5_0')">large (Q5_0, 1030 MB)</button>
+                <span id="fetch-whisper-progress"></span>
             </div>
 
             <br>
@@ -160,6 +176,14 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
                 document.getElementById('whisper-file'         ).style.display = 'none';
                 document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
             }
@@ -168,19 +192,42 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+                    'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
+                    'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
+                    'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
+                    'large-q5_0':    'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+                    'small.en': 466,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
+                    'small-en-q5_1':  182,
+                    'medium-en-q5_0': 515,
+                    'large-q5_0':     1030,
                 };
 
                 let url     = urls[model];
                 let dst     = 'whisper.bin';
                 let size_mb = sizes[model];
 
-                document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
-                document.getElementById('fetch-whisper-base-en').style.display = 'none';
+                document.getElementById('fetch-whisper-tiny-en').style.display  = 'none';
+                document.getElementById('fetch-whisper-base-en').style.display  = 'none';
+                document.getElementById('fetch-whisper-small-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
+                document.getElementById('whisper-file'        ).style.display = 'none';
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -190,9 +237,18 @@
 
                 cbCancel = function() {
                     var el;
-                    el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
+                    el = document.getElementById('fetch-whisper-tiny-en');  if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en');  if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en-q5_1' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-medium-en-q5_0'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-large-q5_0'    ); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('whisper-file'        ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
                 };
 
                 loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);

examples/command.wasm/emscripten.cpp

@@ -28,31 +28,6 @@ std::string g_transcribed   = "";
 
 std::vector<float> g_pcmf32;
 
-// compute similarity between two strings using Levenshtein distance
-static float similarity(const std::string & s0, const std::string & s1) {
-    const size_t len0 = s0.size() + 1;
-    const size_t len1 = s1.size() + 1;
-
-    std::vector<int> col(len1, 0);
-    std::vector<int> prevCol(len1, 0);
-
-    for (size_t i = 0; i < len1; i++) {
-        prevCol[i] = i;
-    }
-
-    for (size_t i = 0; i < len0; i++) {
-        col[0] = i;
-        for (size_t j = 1; j < len1; j++) {
-            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (s0[i - 1] == s1[j - 1] ? 0 : 1));
-        }
-        col.swap(prevCol);
-    }
-
-    const float dist = prevCol[len1 - 1];
-
-    return 1.0f - (dist / std::max(s0.size(), s1.size()));
-}
-
 void command_set_status(const std::string & status) {
     std::lock_guard<std::mutex> lock(g_mutex);
     g_status = status;

examples/command.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use, click the "Start" button and follow the instructions.
@@ -45,6 +54,10 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <!--
@@ -162,11 +175,17 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
                 };
 
                 let url     = urls[model];
@@ -177,6 +196,10 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
+
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -188,6 +211,10 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
+
                     el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
                 };
 

examples/command/CMakeLists.txt

@@ -1,4 +1,4 @@
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # command
     set(TARGET command)
     add_executable(${TARGET} command.cpp)

examples/command/command.cpp

@@ -163,31 +163,6 @@ std::string transcribe(whisper_context * ctx, const whisper_params & params, con
     return result;
 }
 
-// compute similarity between two strings using Levenshtein distance
-float similarity(const std::string & s0, const std::string & s1) {
-    const size_t len0 = s0.size() + 1;
-    const size_t len1 = s1.size() + 1;
-
-    std::vector<int> col(len1, 0);
-    std::vector<int> prevCol(len1, 0);
-
-    for (size_t i = 0; i < len1; i++) {
-        prevCol[i] = i;
-    }
-
-    for (size_t i = 0; i < len0; i++) {
-        col[0] = i;
-        for (size_t j = 1; j < len1; j++) {
-            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (s0[i - 1] == s1[j - 1] ? 0 : 1));
-        }
-        col.swap(prevCol);
-    }
-
-    const float dist = prevCol[len1 - 1];
-
-    return 1.0f - (dist / std::max(s0.size(), s1.size()));
-}
-
 std::vector<std::string> read_allowed_commands(const std::string & fname) {
     std::vector<std::string> allowed_commands;
 

examples/common-ggml.cpp

@@ -0,0 +1,241 @@
+#include "common-ggml.h"
+
+#include <regex>
+#include <map>
+
+static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
+    {"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
+    {"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
+    {"q4_2", GGML_FTYPE_MOSTLY_Q4_2},
+    {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
+    {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
+    {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+};
+
+void ggml_print_ftypes(FILE * fp) {
+    for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
+        fprintf(fp, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+    }
+}
+
+enum ggml_ftype ggml_parse_ftype(const char * str) {
+    enum ggml_ftype ftype;
+    if (str[0] == 'q') {
+        const auto it = GGML_FTYPE_MAP.find(str);
+        if (it == GGML_FTYPE_MAP.end()) {
+            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
+            return GGML_FTYPE_UNKNOWN;
+        }
+        ftype = it->second;
+    } else {
+        ftype = (enum ggml_ftype) atoi(str);
+    }
+
+    return ftype;
+}
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip) {
+
+    ggml_type qtype = GGML_TYPE_F32;
+
+    switch (ftype) {
+        case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
+        case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
+        case GGML_FTYPE_MOSTLY_Q4_2: qtype = GGML_TYPE_Q4_2; break;
+        case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
+        case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
+        case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
+        case GGML_FTYPE_UNKNOWN:
+        case GGML_FTYPE_ALL_F32:
+        case GGML_FTYPE_MOSTLY_F16:
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+                {
+                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
+                    return false;
+                }
+    };
+
+    if (!ggml_is_quantized(qtype)) {
+        fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
+        return false;
+    }
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector<float> work;
+
+    std::vector<uint8_t>     data_u8;
+    std::vector<ggml_fp16_t> data_f16;
+    std::vector<float>       data_f32;
+
+    std::vector<int64_t> hist_all(1 << 4, 0);
+
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ttype;
+
+        finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+        finp.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+        if (finp.eof()) {
+            break;
+        }
+
+        int32_t nelements = 1;
+        int32_t ne[4] = { 1, 1, 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            nelements *= ne[i];
+        }
+
+        std::string name(length, 0);
+        finp.read (&name[0], length);
+
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
+
+        bool quantize = false;
+
+        // check if we should quantize this tensor
+        for (const auto & s : to_quant) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // check if we should skip this tensor
+        for (const auto & s : to_skip) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = false;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (n_dims == 2);
+
+        if (quantize) {
+            if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
+                fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                return false;
+            }
+
+            if (ttype == GGML_TYPE_F16) {
+                data_f16.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                data_f32.resize(nelements);
+                for (int i = 0; i < nelements; ++i) {
+                    data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                }
+            } else {
+                data_f32.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+            }
+
+            ttype = qtype;
+        } else {
+            const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+            data_u8.resize(nelements*bpe);
+            finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+        }
+
+        fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+        fout.write(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+        for (int i = 0; i < n_dims; ++i) {
+            fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+        fout.write(&name[0], length);
+
+        if (quantize) {
+            work.resize(nelements); // for quantization
+
+            size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
+            switch ((ggml_type) ttype) {
+                case GGML_TYPE_Q4_0:
+                    {
+                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_1:
+                    {
+                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_2:
+                    {
+                        cur_size = ggml_quantize_q4_2(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_0:
+                    {
+                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_1:
+                    {
+                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q8_0:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_I8:
+                case GGML_TYPE_I16:
+                case GGML_TYPE_I32:
+                case GGML_TYPE_Q8_1:
+                case GGML_TYPE_COUNT:
+                    {
+                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                        return false;
+                    }
+            }
+
+            fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+            total_size_new += cur_size;
+
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
+        } else {
+            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+            fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+            total_size_new += data_u8.size();
+        }
+
+        total_size_org += nelements * sizeof(float);
+    }
+
+    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
+
+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
+    return true;
+}

examples/common-ggml.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <fstream>
+#include <vector>
+#include <string>
+
+enum ggml_ftype ggml_parse_ftype(const char * str);
+
+void ggml_print_ftypes(FILE * fp = stderr);
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip);

examples/common.cpp

@@ -6,12 +6,86 @@
 #include "dr_wav.h"
 
 #include <cmath>
+#include <fstream>
 #include <regex>
 
 #ifndef M_PI
 #define M_PI 3.14159265358979323846
 #endif
 
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-p" || arg == "--prompt") {
+            params.prompt = argv[++i];
+        } else if (arg == "-n" || arg == "--n_predict") {
+            params.n_predict = std::stoi(argv[++i]);
+        } else if (arg == "--top_k") {
+            params.top_k = std::stoi(argv[++i]);
+        } else if (arg == "--top_p") {
+            params.top_p = std::stof(argv[++i]);
+        } else if (arg == "--temp") {
+            params.temp = std::stof(argv[++i]);
+        } else if (arg == "-b" || arg == "--batch_size") {
+            params.n_batch = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string gpt_random_prompt(std::mt19937 & rng) {
+    const int r = rng() % 10;
+    switch (r) {
+        case 0: return "So";
+        case 1: return "Once upon a time";
+        case 2: return "When";
+        case 3: return "The";
+        case 4: return "After";
+        case 5: return "If";
+        case 6: return "import";
+        case 7: return "He";
+        case 8: return "She";
+        case 9: return "They";
+        default: return "To";
+    }
+
+    return "The";
+}
+
 std::string trim(const std::string & s) {
     std::regex e("^\\s+|\\s+$");
     return std::regex_replace(s, e, "");
@@ -27,6 +101,251 @@ std::string replace(const std::string & s, const std::string & from, const std::
     return result;
 }
 
+std::map<std::string, int32_t> json_parse(const std::string & fname) {
+    std::map<std::string, int32_t> result;
+
+    // read file into string
+    std::string json;
+    {
+        std::ifstream ifs(fname);
+        if (!ifs) {
+            fprintf(stderr, "Failed to open %s\n", fname.c_str());
+            exit(1);
+        }
+
+        json = std::string((std::istreambuf_iterator<char>(ifs)),
+                (std::istreambuf_iterator<char>()));
+    }
+
+    if (json[0] != '{') {
+        return result;
+    }
+
+    // parse json
+    {
+        bool has_key  = false;
+        bool in_token = false;
+
+        std::string str_key = "";
+        std::string str_val = "";
+
+        int n = json.size();
+        for (int i = 1; i < n; ++i) {
+            if (!in_token) {
+                if (json[i] == ' ') continue;
+                if (json[i] == '"') {
+                    in_token = true;
+                    continue;
+                }
+            } else {
+                if (json[i] == '\\' && i+1 < n) {
+                    if (has_key == false) {
+                        str_key += json[i];
+                    } else {
+                        str_val += json[i];
+                    }
+                    ++i;
+                } else if (json[i] == '"') {
+                    if (has_key == false) {
+                        has_key = true;
+                        ++i;
+                        while (json[i] == ' ') ++i;
+                        ++i; // :
+                        while (json[i] == ' ') ++i;
+                        if (json[i] != '\"') {
+                            while (json[i] != ',' && json[i] != '}') {
+                                str_val += json[i++];
+                            }
+                            has_key = false;
+                        } else {
+                            in_token = true;
+                            continue;
+                        }
+                    } else {
+                        has_key = false;
+                    }
+
+                    str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
+                    str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
+                    str_key = ::replace(str_key, "\\\"",    "\""); // \\\"   -> "
+
+                    try {
+                        result[str_key] = std::stoi(str_val);
+                    } catch (...) {
+                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
+
+                    }
+                    str_key = "";
+                    str_val = "";
+                    in_token = false;
+                    continue;
+                }
+                if (has_key == false) {
+                    str_key += json[i];
+                } else {
+                    str_val += json[i];
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+
+        std::regex re(pat);
+        std::smatch m;
+
+        while (std::regex_search(str, m, re)) {
+            for (auto x : m) {
+                words.push_back(x);
+            }
+            str = m.suffix();
+        }
+    }
+
+    // find the longest tokens that form the words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        if (word.size() == 0) continue;
+
+        int i = 0;
+        int n = word.size();
+        while (i < n) {
+            int j = n;
+            while (j > i) {
+                auto it = vocab.token_to_id.find(word.substr(i, j-i));
+                if (it != vocab.token_to_id.end()) {
+                    tokens.push_back(it->second);
+                    i = j;
+                    break;
+                }
+                --j;
+            }
+            if (i == n) {
+                break;
+            }
+            if (j == i) {
+                auto sub = word.substr(i, 1);
+                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
+                    tokens.push_back(vocab.token_to_id.at(sub));
+                } else {
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
+                }
+                ++i;
+            }
+        }
+    }
+
+    return tokens;
+}
+
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
+    printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
+
+    vocab.token_to_id = ::json_parse(fname);
+
+    for (const auto & kv : vocab.token_to_id) {
+        vocab.id_to_token[kv.second] = kv.first;
+    }
+
+    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
+
+    // print the vocabulary
+    //for (auto kv : vocab.token_to_id) {
+    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
+    //}
+
+    return true;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[i]*scale, i));
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) probs.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
     std::vector<uint8_t> wav_data; // used for pipe input from stdin
@@ -160,3 +479,27 @@ bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float
 
     return true;
 }
+
+float similarity(const std::string & s0, const std::string & s1) {
+    const size_t len0 = s0.size() + 1;
+    const size_t len1 = s1.size() + 1;
+
+    std::vector<int> col(len1, 0);
+    std::vector<int> prevCol(len1, 0);
+
+    for (size_t i = 0; i < len1; i++) {
+        prevCol[i] = i;
+    }
+
+    for (size_t i = 0; i < len0; i++) {
+        col[0] = i;
+        for (size_t j = 1; j < len1; j++) {
+            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (i > 0 && s0[i - 1] == s1[j - 1] ? 0 : 1));
+        }
+        col.swap(prevCol);
+    }
+
+    const float dist = prevCol[len1 - 1];
+
+    return 1.0f - (dist / std::max(s0.size(), s1.size()));
+}

examples/common.h

@@ -1,10 +1,44 @@
+// Various helper functions and utilities
+
 #pragma once
 
-// needs to match WHISPER_SAMPLE_RATE
+#include <string>
+#include <map>
+#include <vector>
+#include <random>
+#include <thread>
+
 #define COMMON_SAMPLE_RATE 16000
 
-#include <vector>
-#include <string>
+//
+// CLI argument parsing
+//
+
+struct gpt_params {
+    int32_t seed      = -1; // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict = 200; // new tokens to predict
+
+    // sampling parameters
+    int32_t top_k = 40;
+    float   top_p = 0.9f;
+    float   temp  = 0.9f;
+
+    int32_t n_batch = 8; // batch size for prompt processing
+
+    std::string model = "models/gpt-2-117M/ggml-model.bin"; // model path
+    std::string prompt;
+};
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+//
+// Vocab utils
+//
 
 std::string trim(const std::string & s);
 
@@ -13,6 +47,52 @@ std::string replace(
         const std::string & from,
         const std::string & to);
 
+struct gpt_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+};
+
+// poor-man's JSON parsing
+std::map<std::string, int32_t> json_parse(const std::string & fname);
+
+// split text into tokens
+//
+// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
+//
+// Regex (Python):
+// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
+//
+// Regex (C++):
+// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
+//
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
+
+// load the tokens from encoder.json
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
+
+// sample next token given probabilities for each embedding
+//
+//   - consider only the top K tokens
+//   - from them, consider only the top tokens with cumulative probability > P
+//
+// TODO: not sure if this implementation is correct
+// TODO: temperature is not implemented
+//
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng);
+
+//
+// Audio utils
+//
+
 // Read WAV audio file and store the PCM data into pcmf32
 // The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
 // If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
@@ -38,3 +118,5 @@ bool vad_simple(
         float freq_thold,
         bool  verbose);
 
+// compute similarity between two strings using Levenshtein distance
+float similarity(const std::string & s0, const std::string & s1);

examples/helpers.js

@@ -145,7 +145,15 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
                             var db = event.target.result;
                             var tx = db.transaction(['models'], 'readwrite');
                             var os = tx.objectStore('models');
-                            var rq = os.put(data, url);
+
+                            var rq = null;
+                            try {
+                                var rq = os.put(data, url);
+                            } catch (e) {
+                                cbPrint('loadRemote: failed to store "' + url + '" in the IndexedDB: \n' + e);
+                                cbCancel();
+                                return;
+                            }
 
                             rq.onsuccess = function (event) {
                                 cbPrint('loadRemote: "' + url + '" stored in the IndexedDB');
@@ -180,7 +188,6 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
 
     rq.onabort = function (event) {
         cbPrint('loadRemote: failed to open IndexedDB: abort');
-
+        cbCancel();
     };
 }
-

examples/main/README.md

@@ -31,6 +31,7 @@ options:
   -osrt,     --output-srt        [false  ] output result in a srt file
   -owts,     --output-words      [false  ] output script for generating karaoke video
   -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors

examples/main/main.cpp

@@ -8,6 +8,7 @@
 #include <string>
 #include <thread>
 #include <vector>
+#include <cstring>
 
 // Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
 // Lowest is red, middle is yellow, highest is green.
@@ -56,7 +57,7 @@ struct whisper_params {
     int32_t duration_ms  =  0;
     int32_t max_context  = -1;
     int32_t max_len      =  0;
-    int32_t best_of      =  5;
+    int32_t best_of      =  2;
     int32_t beam_size    = -1;
 
     float word_thold    =  0.01f;
@@ -65,6 +66,7 @@ struct whisper_params {
 
     bool speed_up       = false;
     bool translate      = false;
+    bool detect_language= false;
     bool diarize        = false;
     bool split_on_word  = false;
     bool no_fallback    = false;
@@ -73,6 +75,8 @@ struct whisper_params {
     bool output_srt     = false;
     bool output_wts     = false;
     bool output_csv     = false;
+    bool output_jsn     = false;
+    bool output_lrc     = false;
     bool print_special  = false;
     bool print_colors   = false;
     bool print_progress = false;
@@ -80,6 +84,7 @@ struct whisper_params {
 
     std::string language = "en";
     std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
     std::string model    = "models/ggml-base.en.bin";
 
     std::vector<std::string> fname_inp = {};
@@ -127,13 +132,17 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-ovtt" || arg == "--output-vtt")     { params.output_vtt     = true; }
         else if (arg == "-osrt" || arg == "--output-srt")     { params.output_srt     = true; }
         else if (arg == "-owts" || arg == "--output-words")   { params.output_wts     = true; }
+        else if (arg == "-olrc" || arg == "--output-lrc")     { params.output_lrc     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")      { params.font_path      = argv[++i]; }
         else if (arg == "-ocsv" || arg == "--output-csv")     { params.output_csv     = true; }
+        else if (arg == "-oj"   || arg == "--output-json")    { params.output_jsn     = true; }
         else if (arg == "-of"   || arg == "--output-file")    { params.fname_out.emplace_back(argv[++i]); }
         else if (arg == "-ps"   || arg == "--print-special")  { params.print_special  = true; }
         else if (arg == "-pc"   || arg == "--print-colors")   { params.print_colors   = true; }
         else if (arg == "-pp"   || arg == "--print-progress") { params.print_progress = true; }
         else if (arg == "-nt"   || arg == "--no-timestamps")  { params.no_timestamps  = true; }
         else if (arg == "-l"    || arg == "--language")       { params.language       = argv[++i]; }
+        else if (arg == "-dl"   || arg == "--detect-language"){ params.detect_language= true; }
         else if (                  arg == "--prompt")         { params.prompt         = argv[++i]; }
         else if (arg == "-m"    || arg == "--model")          { params.model          = argv[++i]; }
         else if (arg == "-f"    || arg == "--file")           { params.fname_inp.emplace_back(argv[++i]); }
@@ -173,14 +182,18 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  -otxt,     --output-txt        [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
     fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
     fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
+    fprintf(stderr, "  -olrc,     --output-lrc        [%-7s] output result in a lrc file\n",                    params.output_lrc ? "true" : "false");
     fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
     fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
+    fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
     fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
     fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
     fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
     fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
     fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "false" : "true");
     fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
     fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
     fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
     fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
@@ -193,7 +206,7 @@ struct whisper_print_user_data {
     const std::vector<std::vector<float>> * pcmf32s;
 };
 
-void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
     const auto & params  = *((whisper_print_user_data *) user_data)->params;
     const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
 
@@ -201,8 +214,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, int n_new, voi
 
     std::string speaker = "";
 
-    int64_t t0;
-    int64_t t1;
+    int64_t t0 = 0;
+    int64_t t1 = 0;
 
     // print the last n_new segments
     const int s0 = n_segments - n_new;
@@ -342,6 +355,37 @@ bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_
     return true;
 }
 
+char *escape_double_quotes_and_backslashes(const char *str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped[pos++] = '\\';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
 bool output_csv(struct whisper_context * ctx, const char * fname) {
     std::ofstream fout(fname);
     if (!fout.is_open()) {
@@ -352,28 +396,160 @@ bool output_csv(struct whisper_context * ctx, const char * fname) {
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
     const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,text\n";
     for (int i = 0; i < n_segments; ++i) {
         const char * text = whisper_full_get_segment_text(ctx, i);
         const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
         const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        char * text_escaped = escape_double_quotes_and_backslashes(text);
 
         //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
-        fout << 10 * t0 << ", " << 10 * t1 << ", \"" << text    << "\"\n";
+        fout << 10 * t0 << "," << 10 * t1 << ",\"" << text_escaped    << "\"\n";
     }
 
     return true;
 }
 
+bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params) {
+    std::ofstream fout(fname);
+    int indent = 0;
+
+    auto doindent = [&]() {
+        for (int i = 0; i < indent; i++) fout << "\t";
+    };
+
+    auto start_arr = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": [\n";
+        indent++;
+    };
+
+    auto end_arr = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "},\n");
+    };
+
+    auto start_obj = [&](const char *name = nullptr) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end = false) {
+        indent--;
+        doindent();
+        fout << (end ? "}\n" : "},\n");
+    };
+
+    auto start_value = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": ";
+    };
+
+    auto value_s = [&](const char *name, const char *val, bool end = false) {
+        start_value(name);
+        char * val_escaped = escape_double_quotes_and_backslashes(val);
+        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
+        free(val_escaped);
+    };
+
+    auto end_value = [&](bool end = false) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    auto value_i = [&](const char *name, const int64_t val, bool end = false) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end = false) {
+        start_value(name);
+        fout << (val ? "true" : "false");
+        end_value(end);
+    };
+
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+    start_obj();
+        value_s("systeminfo", whisper_print_system_info());
+        start_obj("model");
+            value_s("type", whisper_model_type_readable(ctx));
+            value_b("multilingual", whisper_is_multilingual(ctx));
+            value_i("vocab", whisper_model_n_vocab(ctx));
+            start_obj("audio");
+                value_i("ctx", whisper_model_n_audio_ctx(ctx));
+                value_i("state", whisper_model_n_audio_state(ctx));
+                value_i("head", whisper_model_n_audio_head(ctx));
+                value_i("layer", whisper_model_n_audio_layer(ctx), true);
+            end_obj();
+            start_obj("text");
+                value_i("ctx", whisper_model_n_text_ctx(ctx));
+                value_i("state", whisper_model_n_text_state(ctx));
+                value_i("head", whisper_model_n_text_head(ctx));
+                value_i("layer", whisper_model_n_text_layer(ctx), true);
+            end_obj();
+            value_i("mels", whisper_model_n_mels(ctx));
+            value_i("ftype", whisper_model_ftype(ctx), true);
+        end_obj();
+        start_obj("params");
+            value_s("model", params.model.c_str());
+            value_s("language", params.language.c_str());
+            value_b("translate", params.translate, true);
+        end_obj();
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        end_obj();
+        start_arr("transcription");
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+                start_obj();
+                    start_obj("timestamps");
+                        value_s("from", to_timestamp(t0, true).c_str());
+                        value_s("to", to_timestamp(t1, true).c_str(), true);
+                    end_obj();
+                    start_obj("offsets");
+                        value_i("from", t0 * 10);
+                        value_i("to", t1 * 10, true);
+                    end_obj();
+                    value_s("text", text, true);
+                end_obj(i == (n_segments - 1));
+            }
+
+        end_arr(true);
+    end_obj(true);
+    return true;
+}
+
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & /*params*/, float t_sec) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec) {
     std::ofstream fout(fname);
 
     fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
 
-    // TODO: become parameter
-    static const char * font = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
 
     fout << "#!/bin/bash" << "\n";
     fout << "\n";
@@ -476,6 +652,39 @@ bool output_wts(struct whisper_context * ctx, const char * fname, const char * f
     return true;
 }
 
+bool output_lrc(struct whisper_context * ctx, const char * fname) {
+
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    fout << "[by:whisper.cpp]\n";
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t = whisper_full_get_segment_t0(ctx, i);
+
+        int64_t msec = t * 10;
+        int64_t min = msec / (1000 * 60);
+        msec = msec - min * (1000 * 60);
+        int64_t sec = msec / 1000;
+        msec = msec - sec * 1000;
+
+        char buf[16];
+        snprintf(buf, sizeof(buf), "%02d:%02d.%02d", (int) min, (int) sec, (int) ( msec / 10));
+        std::string timestamp_lrc = std::string(buf);
+
+        fout <<  '[' << timestamp_lrc << ']' << text << "\n";
+    }
+
+    return true;
+}
+
 int main(int argc, char ** argv) {
     whisper_params params;
 
@@ -504,22 +713,6 @@ int main(int argc, char ** argv) {
         return 3;
     }
 
-    // initial prompt
-    std::vector<whisper_token> prompt_tokens;
-
-    if (!params.prompt.empty()) {
-        prompt_tokens.resize(1024);
-        prompt_tokens.resize(whisper_tokenize(ctx, params.prompt.c_str(), prompt_tokens.data(), prompt_tokens.size()));
-
-        fprintf(stderr, "\n");
-        fprintf(stderr, "initial prompt: '%s'\n", params.prompt.c_str());
-        fprintf(stderr, "initial tokens: [ ");
-        for (int i = 0; i < (int) prompt_tokens.size(); ++i) {
-            fprintf(stderr, "%d ", prompt_tokens[i]);
-        }
-        fprintf(stderr, "]\n");
-    }
-
     for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
         const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@@ -549,6 +742,9 @@ int main(int argc, char ** argv) {
                     fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                 }
             }
+            if (params.detect_language) {
+                params.language = "auto";
+            }
             fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
                     __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                     params.n_threads, params.n_processors,
@@ -571,6 +767,7 @@ int main(int argc, char ** argv) {
             wparams.print_special    = params.print_special;
             wparams.translate        = params.translate;
             wparams.language         = params.language.c_str();
+            wparams.detect_language  = params.detect_language;
             wparams.n_threads        = params.n_threads;
             wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
             wparams.offset_ms        = params.offset_t_ms;
@@ -583,8 +780,7 @@ int main(int argc, char ** argv) {
 
             wparams.speed_up         = params.speed_up;
 
-            wparams.prompt_tokens     = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
-            wparams.prompt_n_tokens   = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+            wparams.initial_prompt   = params.prompt.c_str();
 
             wparams.greedy.best_of        = params.best_of;
             wparams.beam_search.beam_size = params.beam_size;
@@ -607,7 +803,7 @@ int main(int argc, char ** argv) {
             {
                 static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
 
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, void * user_data) {
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
                     bool is_aborted = *(bool*)user_data;
                     return !is_aborted;
                 };
@@ -618,8 +814,6 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                 return 10;
             }
-
-            whisper_full_cluster_segments(ctx);
         }
 
         // output stuff
@@ -655,6 +849,18 @@ int main(int argc, char ** argv) {
                 const auto fname_csv = fname_out + ".csv";
                 output_csv(ctx, fname_csv.c_str());
             }
+
+            // output to JSON file
+            if (params.output_jsn) {
+                const auto fname_jsn = fname_out + ".json";
+                output_json(ctx, fname_jsn.c_str(), params);
+            }
+
+            // output to LRC file
+            if (params.output_lrc) {
+                const auto fname_lrc = fname_out + ".lrc";
+                output_lrc(ctx, fname_lrc.c_str());
+            }
         }
     }
 

examples/quantize/CMakeLists.txt

@@ -0,0 +1,6 @@
+set(TARGET quantize)
+add_executable(${TARGET} quantize.cpp)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

examples/quantize/README.md

@@ -0,0 +1,3 @@
+# quantize
+
+Tool for integer quantization of Whisper `ggml` model files

examples/quantize/quantize.cpp

@@ -0,0 +1,215 @@
+#include "ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (Whisper tiny)
+struct whisper_hparams {
+    int32_t n_vocab       = 51864;
+    int32_t n_audio_ctx   = 1500;
+    int32_t n_audio_state = 384;
+    int32_t n_audio_head  = 6;
+    int32_t n_audio_layer = 4;
+    int32_t n_text_ctx    = 448;
+    int32_t n_text_state  = 384;
+    int32_t n_text_head   = 6;
+    int32_t n_text_layer  = 4;
+    int32_t n_mels        = 80;
+    int32_t f16           = 1;
+};
+
+struct whisper_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+// quantize a model
+bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    whisper_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        finp.read((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        finp.read((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        finp.read((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        finp.read((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        finp.read((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        finp.read((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        finp.read((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        finp.read((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        finp.read((char *) &hparams.f16,           sizeof(hparams.f16));
+
+        fprintf(stderr, "%s: n_vocab       = %d\n", __func__, hparams.n_vocab);
+        fprintf(stderr, "%s: n_audio_ctx   = %d\n", __func__, hparams.n_audio_ctx);
+        fprintf(stderr, "%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
+        fprintf(stderr, "%s: n_audio_head  = %d\n", __func__, hparams.n_audio_head);
+        fprintf(stderr, "%s: n_audio_layer = %d\n", __func__, hparams.n_audio_layer);
+        fprintf(stderr, "%s: n_text_ctx    = %d\n", __func__, hparams.n_text_ctx);
+        fprintf(stderr, "%s: n_text_state  = %d\n", __func__, hparams.n_text_state);
+        fprintf(stderr, "%s: n_text_head   = %d\n", __func__, hparams.n_text_head);
+        fprintf(stderr, "%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
+        fprintf(stderr, "%s: n_mels        = %d\n", __func__, hparams.n_mels);
+        fprintf(stderr, "%s: f16           = %d\n", __func__, hparams.f16);
+
+        fout.write((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        fout.write((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        fout.write((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        fout.write((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        fout.write((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        fout.write((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        fout.write((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        fout.write((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        fout.write((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        fout.write((char *) &ftype,                 sizeof(hparams.f16));
+    }
+
+    // load mel filters
+    {
+        whisper_filters filters;
+
+        finp.read ((char *) &filters.n_mel, sizeof(filters.n_mel));
+        fout.write((char *) &filters.n_mel, sizeof(filters.n_mel));
+        finp.read ((char *) &filters.n_fft, sizeof(filters.n_fft));
+        fout.write((char *) &filters.n_fft, sizeof(filters.n_fft));
+
+        filters.data.resize(filters.n_mel * filters.n_fft);
+        finp.read ((char *) filters.data.data(), filters.data.size() * sizeof(float));
+        fout.write((char *) filters.data.data(), filters.data.size() * sizeof(float));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        //if (n_vocab != hparams.n_vocab) {
+        //    fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+        //            __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+        //    return false;
+        //}
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to not be quantized
+    const std::vector<std::string> to_skip = {
+        //"encoder.*",
+        "encoder.conv1.bias",
+        "encoder.conv2.bias",
+        "encoder.positional_embedding",
+        "decoder.positional_embedding",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, { ".*" }, to_skip)) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!whisper_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/stream.wasm/index-tmpl.html

@@ -35,6 +35,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the model you would like to use, click the "Start" button and start speaking
@@ -45,6 +54,10 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <!--
@@ -162,11 +175,17 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
                 };
 
                 let url     = urls[model];
@@ -177,6 +196,10 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
+
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -188,6 +211,10 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
+
                     el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
                 };
 

examples/stream/CMakeLists.txt

@@ -1,4 +1,4 @@
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # stream
     set(TARGET stream)
     add_executable(${TARGET} stream.cpp)

examples/stream/stream.cpp

@@ -43,6 +43,7 @@ struct whisper_params {
 
     bool speed_up      = false;
     bool translate     = false;
+    bool no_fallback   = false;
     bool print_special = false;
     bool no_context    = true;
     bool no_timestamps = false;
@@ -73,6 +74,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
         else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
         else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
         else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
         else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
         else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
         else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
@@ -94,22 +96,23 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "\n");
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",             params.step_ms);
-    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                params.length_ms);
-    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",      params.keep_ms);
-    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
-    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
-    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
-    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
-    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
-    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
-    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",           params.no_context ? "false" : "true");
-    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                  params.model.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "            --step N        [%-7d] audio step size in milliseconds\n",                params.step_ms);
+    fprintf(stderr, "            --length N      [%-7d] audio length in milliseconds\n",                   params.length_ms);
+    fprintf(stderr, "            --keep N        [%-7d] audio to keep from previous step in ms\n",         params.keep_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                              params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",       params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -kc,      --keep-context  [%-7s] keep context between audio chunks\n",              params.no_context ? "false" : "true");
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
+    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
     fprintf(stderr, "\n");
 }
 
@@ -148,7 +151,7 @@ int main(int argc, char ** argv) {
 
     // whisper init
 
-    if (whisper_lang_id(params.language.c_str()) == -1) {
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
         fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
         whisper_print_usage(argc, argv, params);
         exit(0);
@@ -288,7 +291,6 @@ int main(int argc, char ** argv) {
             wparams.print_realtime   = false;
             wparams.print_timestamps = !params.no_timestamps;
             wparams.translate        = params.translate;
-            wparams.no_context       = true;
             wparams.single_segment   = !use_vad;
             wparams.max_tokens       = params.max_tokens;
             wparams.language         = params.language.c_str();
@@ -298,7 +300,8 @@ int main(int argc, char ** argv) {
             wparams.speed_up         = params.speed_up;
 
             // disable temperature fallback
-            wparams.temperature_inc  = -1.0f;
+            //wparams.temperature_inc  = -1.0f;
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
 
             wparams.prompt_tokens    = params.no_context ? nullptr : prompt_tokens.data();
             wparams.prompt_n_tokens  = params.no_context ? 0       : prompt_tokens.size();
@@ -380,6 +383,7 @@ int main(int argc, char ** argv) {
                     }
                 }
             }
+            fflush(stdout);
         }
     }
 

examples/talk-llama/.gitignore

@@ -0,0 +1 @@
+audio.mp3

examples/talk-llama/CMakeLists.txt

@@ -0,0 +1,16 @@
+if (WHISPER_SDL2)
+    # talk-llama
+    set(TARGET talk-llama)
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+    include(DefaultTargetOptions)
+endif ()

examples/talk-llama/README.md

@@ -0,0 +1,50 @@
+# talk-llama
+
+Talk with an LLaMA AI in your terminal
+
+[Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)
+
+## Building
+
+The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2 on Linux
+sudo apt-get install libsdl2-dev
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+# Build the "talk-llama" executable
+make talk-llama
+
+# Run it
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+- The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
+- The `-ml` argument specifies the LLaMA model that you would like to use. Read the instructions in https://github.com/ggerganov/llama.cpp for information about how to obtain a `ggml` compatible LLaMA model
+
+## Session
+
+The `talk-llama` tool supports session management to enable more coherent and continuous conversations. By maintaining context from previous interactions, it can better understand and respond to user requests in a more natural way.
+
+To enable session support, use the `--session FILE` command line option when running the program. The `talk-llama` model state will be saved to the specified file after each interaction. If the file does not exist, it will be created. If the file exists, the model state will be loaded from it, allowing you to resume a previous session.
+
+This feature is especially helpful for maintaining context in long conversations or when interacting with the AI assistant across multiple sessions. It ensures that the assistant remembers the previous interactions and can provide more relevant and contextual responses.
+
+Example usage:
+
+```bash
+./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
+```
+
+## TTS
+
+For best experience, this example needs a TTS tool to convert the generated text responses to voice.
+You can use any TTS engine that you would like - simply edit the [speak.sh](speak.sh) script to your needs.
+By default, it is configured to use MacOS's `say`, but you can use whatever you wish.
+
+## Discussion
+
+If you have any feedback, please let "us" know in the following discussion: https://github.com/ggerganov/whisper.cpp/discussions/672?converting=1

examples/talk-llama/eleven-labs.py

@@ -0,0 +1,23 @@
+import sys
+import importlib.util
+
+api_key = "" #Write your https://beta.elevenlabs.io api key here
+if not api_key:
+    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk-llama/eleven-labs.py")
+    sys.exit()
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import ElevenLabs
+eleven = ElevenLabs(api_key)
+
+# Get a Voice object, by name or UUID
+voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = voice.generate(str(sys.argv[2:]))
+
+# Save the TTS to a file
+audio.save("audio") 

examples/talk-llama/llama-util.h

@@ -0,0 +1,433 @@
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw format("failed to open %s: %s", fname, std::strerror(errno));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            throw format("read error: %s", strerror(errno));
+        }
+        if (ret != 1) {
+            throw std::string("unexpectedly reached end of file");
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            throw format("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+#ifdef __linux__
+        flags |= MAP_POPULATE;
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw format("mmap failed: %s", strerror(errno));
+        }
+
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            if (madvise(addr, file->size, MADV_WILLNEED)) {
+                fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+
+        if (hMapping == NULL) {
+            throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
+        }
+
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
+        }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *) {
+        throw std::string("mmap not supported");
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * addr) {
+        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
+        this->addr = addr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            struct rlimit lock_limit;
+            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
+                suggest = false;
+            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
+                suggest = false;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * addr, size_t size) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(addr, size)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                        size, this->size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = size + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * addr, size_t size) {
+        if (!VirtualUnlock(addr, size)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    void raw_lock(const void * addr, size_t size) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+    }
+
+    void raw_unlock(const void * addr, size_t size) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    void resize(size_t size) {
+        delete[] addr;
+        addr = new uint8_t[size];
+        this->size = size;
+    }
+
+    ~llama_buffer() {
+        delete[] addr;
+    }
+};
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    void resize(size_t size) {
+        if (addr) {
+            ggml_cuda_host_free(addr);
+        }
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        this->size = size;
+    }
+
+    ~llama_ctx_buffer() {
+        if (addr) {
+            ggml_cuda_host_free(addr);
+        }
+    }
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
+#endif

examples/talk-llama/llama.h

@@ -0,0 +1,258 @@
+#ifndef LLAMA_H
+#define LLAMA_H
+
+#include <stddef.h>
+#include <stdint.h>
+#include <stdbool.h>
+
+#ifdef LLAMA_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef LLAMA_BUILD
+#            define LLAMA_API __declspec(dllexport)
+#        else
+#            define LLAMA_API __declspec(dllimport)
+#        endif
+#    else
+#        define LLAMA_API __attribute__ ((visibility ("default")))
+#    endif
+#else
+#    define LLAMA_API
+#endif
+
+#define LLAMA_FILE_VERSION           1
+#define LLAMA_FILE_MAGIC             'ggjt'
+#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
+#define LLAMA_SESSION_MAGIC          'ggsn'
+#define LLAMA_SESSION_VERSION        0
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+    //
+    // C interface
+    //
+    // TODO: show sample usage
+    //
+
+    struct llama_context;
+
+    typedef int llama_token;
+
+    typedef struct llama_token_data {
+        llama_token id;  // token id
+        float logit; // log-odds of the token
+        float p;     // probability of the token
+    } llama_token_data;
+
+    typedef struct llama_token_data_array {
+        llama_token_data * data;
+        size_t size;
+        bool sorted;
+    } llama_token_data_array;
+
+    typedef void (*llama_progress_callback)(float progress, void *ctx);
+
+    struct llama_context_params {
+        int n_ctx;   // text context
+        int n_parts; // -1 for default
+        int seed;    // RNG seed, 0 for random
+
+        bool f16_kv;     // use fp16 for KV cache
+        bool logits_all; // the llama_eval() call computes all logits, not just the last one
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
+        bool embedding;  // embedding mode only
+
+        // called with a progress value between 0 and 1, pass NULL to disable
+        llama_progress_callback progress_callback;
+        // context pointer passed to the progress callback
+        void * progress_callback_user_data;
+    };
+
+    // model file types
+    enum llama_ftype {
+        LLAMA_FTYPE_ALL_F32     = 0,
+        LLAMA_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
+        // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+    };
+
+    LLAMA_API struct llama_context_params llama_context_default_params();
+
+    LLAMA_API bool llama_mmap_supported();
+    LLAMA_API bool llama_mlock_supported();
+
+    // Various functions for loading a ggml llama model.
+    // Allocate (almost) all memory needed for the model.
+    // Return NULL on failure
+    LLAMA_API struct llama_context * llama_init_from_file(
+                             const char * path_model,
+            struct llama_context_params   params);
+
+    // Frees all allocated memory
+    LLAMA_API void llama_free(struct llama_context * ctx);
+
+    // TODO: not great API - very likely to change
+    // Returns 0 on success
+    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
+    LLAMA_API int llama_model_quantize(
+            const char * fname_inp,
+            const char * fname_out,
+      enum llama_ftype   ftype,
+            int          nthread);
+
+    // Apply a LoRA adapter to a loaded model
+    // path_base_model is the path to a higher quality model to use as a base for
+    // the layers modified by the adapter. Can be NULL to use the current loaded model.
+    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
+    // will be applied on top of the previous one
+    // Returns 0 on success
+    LLAMA_API int llama_apply_lora_from_file(
+            struct llama_context * ctx,
+                      const char * path_lora,
+                      const char * path_base_model,
+                             int   n_threads);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+
+    // Sets the current rng seed.
+    LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
+
+    // Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
+    LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
+
+    // Copies the state to the specified destination address.
+    // Destination needs to have allocated enough memory.
+    // Returns the number of bytes copied
+    LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest);
+
+    // Set the state reading from the specified address
+    // Returns the number of bytes read
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
+
+    // Save/load session file
+    LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
+    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    // Run the llama inference to obtain the logits and probabilities for the next token.
+    // tokens + n_tokens is the provided batch of new tokens to process
+    // n_past is the number of tokens to use from previous eval calls
+    // Returns 0 on success
+    LLAMA_API int llama_eval(
+            struct llama_context * ctx,
+               const llama_token * tokens,
+                             int   n_tokens,
+                             int   n_past,
+                             int   n_threads);
+
+    // Convert the provided text into tokens.
+    // The tokens pointer must be large enough to hold the resulting tokens.
+    // Returns the number of tokens on success, no more than n_max_tokens
+    // Returns a negative number on failure - the number of tokens that would have been returned
+    // TODO: not sure if correct
+    LLAMA_API int llama_tokenize(
+            struct llama_context * ctx,
+                      const char * text,
+                     llama_token * tokens,
+                             int   n_max_tokens,
+                            bool   add_bos);
+
+    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+
+    // Token logits obtained from the last call to llama_eval()
+    // The logits for the last token are stored in the last row
+    // Can be mutated in order to change the probabilities of the next token
+    // Rows: n_tokens
+    // Cols: n_vocab
+    LLAMA_API float * llama_get_logits(struct llama_context * ctx);
+
+    // Get the embeddings for the input
+    // shape: [n_embd] (1-dimensional)
+    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
+
+    // Token Id -> String. Uses the vocabulary in the provided context
+    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+
+    // Special tokens
+    LLAMA_API llama_token llama_token_bos();
+    LLAMA_API llama_token llama_token_eos();
+    LLAMA_API llama_token llama_token_nl();
+
+    // Sampling functions
+
+    /// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
+    LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float penalty);
+
+    /// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
+    LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
+
+    /// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
+    LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep = 1);
+
+    /// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
+    LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+
+    /// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
+    LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep = 1);
+
+    /// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
+    LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
+    LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
+
+    /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
+
+    /// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
+    /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
+    /// @param tau  The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
+    /// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
+    /// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
+    LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
+
+    /// @details Selects the token with the highest probability.
+    LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    /// @details Randomly selects a token from the candidates based on their probabilities.
+    LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
+
+    // Performance information
+    LLAMA_API void llama_print_timings(struct llama_context * ctx);
+    LLAMA_API void llama_reset_timings(struct llama_context * ctx);
+
+    // Print system information
+    LLAMA_API const char * llama_print_system_info(void);
+
+#ifdef __cplusplus
+}
+#endif
+
+// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
+#ifdef LLAMA_API_INTERNAL
+
+#include <vector>
+#include <string>
+struct ggml_tensor;
+
+std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
+
+#endif
+
+#endif // LLAMA_H

examples/talk-llama/prompts/talk-alpaca.txt

@@ -0,0 +1,23 @@
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+
+Write a text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+### Response:
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4}

examples/talk-llama/speak.sh

@@ -0,0 +1,21 @@
+#!/bin/bash
+
+# Usage:
+#  speak.sh <voice_id> <text-to-speak>
+
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# for Mac
+say "$2"
+
+# Eleven Labs
+# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk-llama/eleven-labs.py 
+#
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1

examples/talk-llama/talk-llama.cpp

@@ -0,0 +1,673 @@
+// Talk with AI
+//
+
+#include "common.h"
+#include "common-sdl.h"
+#include "whisper.h"
+#include "llama.h"
+
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <regex>
+
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);
+
+    return res;
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t voice_ms   = 10000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    int32_t n_parts_llama = -1;
+
+    float vad_thold    = 0.6f;
+    float freq_thold   = 100.0f;
+
+    bool speed_up      = false;
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool verbose_prompt = false;
+
+    std::string person      = "Georgi";
+    std::string language    = "en";
+    std::string model_wsp   = "models/ggml-base.en.bin";
+    std::string model_llama = "models/ggml-llama-7B.bin";
+    std::string speak       = "./examples/talk-llama/speak.sh";
+    std::string prompt      = "";
+    std::string fname_out;
+    std::string path_session = "";       // path to file for saving/loading model eval state
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "--n-parts-llama")                  { params.n_parts_llama = std::stoi(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "--session")                        { params.path_session  = argv[++i];}
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "--prompt-file")                    {
+            std::ifstream file(argv[++i]);
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  --n-parts-llama N         [%-7d] num parts in llama model file\n",               params.n_parts_llama);
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
+    fprintf(stderr, "  --session FNAME       file to cache model state in (may be large!) (default: none)\n");
+    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string transcribe(
+        whisper_context * ctx,
+        const whisper_params & params,
+        const std::vector<float> & pcmf32,
+        const std::string prompt_text,
+        float & prob,
+        int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    prob = 0.0f;
+    t_ms = 0;
+
+    std::vector<whisper_token> prompt_tokens;
+
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+    prompt_tokens.resize(1024);
+    prompt_tokens.resize(whisper_tokenize(ctx, prompt_text.c_str(), prompt_tokens.data(), prompt_tokens.size()));
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.prompt_tokens    = prompt_tokens.empty() ? nullptr : prompt_tokens.data();
+    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();
+
+    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    int prob_n = 0;
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n_tokens; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            prob += token.p;
+            ++prob_n;
+        }
+    }
+
+    if (prob_n > 0) {
+        prob /= prob_n;
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";
+
+const std::string k_prompt_llama = R"(Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
+{1} is helpful, kind, honest, friendly, good at writing and never fails to answer {0}’s requests immediately and with details and precision.
+There are no annotations like (30 seconds passed...) or (to himself), just what {0} and {1} say aloud to each other.
+The transcript only includes text, it does not include markup like HTML and Markdown.
+{1} responds with short and concise answers.
+
+{0}{4} Hello, {1}!
+{1}{4} Hello {0}! How may I help you today?
+{0}{4} What time is it?
+{1}{4} It is {2} o'clock.
+{0}{4} What year is it?
+{1}{4} We are in {3}.
+{0}{4} What is a cat?
+{1}{4} A cat is a domestic species of small carnivorous mammal. It is the only domesticated species in the family Felidae.
+{0}{4} Name a color.
+{1}{4} Blue
+{0}{4})";
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());
+
+    // llama init
+
+    auto lparams = llama_context_default_params();
+
+    // tune these to your liking
+    lparams.n_ctx      = 2048;
+    lparams.seed       = 1;
+    lparams.f16_kv     = true;
+    lparams.n_parts    = params.n_parts_llama;
+
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+
+        if (!whisper_is_multilingual(ctx_wsp)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    int n_iter = 0;
+
+    bool is_running  = true;
+    bool force_speak = false;
+
+    float prob0 = 0.0f;
+
+    const std::string chat_symb = ":";
+    const std::string bot_name  = "LLaMA";
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);
+
+    // construct the initial prompt for LLaMA inference
+    std::string prompt_llama = params.prompt.empty() ? k_prompt_llama : params.prompt;
+
+    // need to have leading ' '
+    prompt_llama.insert(0, 1, ' ');
+
+    prompt_llama = ::replace(prompt_llama, "{0}", params.person);
+    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);
+
+    {
+        // get time string
+        std::string time_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%H:%M", now);
+            time_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{2}", time_str);
+    }
+
+    {
+        // get year string
+        std::string year_str;
+        {
+            time_t t = time(0);
+            struct tm * now = localtime(&t);
+            char buf[128];
+            strftime(buf, sizeof(buf), "%Y", now);
+            year_str = buf;
+        }
+        prompt_llama = ::replace(prompt_llama, "{3}", year_str);
+    }
+
+    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);
+
+    // init session
+    std::string path_session = params.path_session;
+    std::vector<llama_token> session_tokens;
+    auto embd_inp = ::llama_tokenize(ctx_llama, prompt_llama, true);
+
+    if (!path_session.empty()) {
+        fprintf(stderr, "%s: attempting to load saved session from %s\n", __func__, path_session.c_str());
+
+        // fopen to check for existing session
+        FILE * fp = std::fopen(path_session.c_str(), "rb");
+        if (fp != NULL) {
+            std::fclose(fp);
+
+            session_tokens.resize(lparams.n_ctx);
+            size_t n_token_count_out = 0;
+            if (!llama_load_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
+                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
+                return 1;
+            }
+            session_tokens.resize(n_token_count_out);
+            for (size_t i = 0; i < session_tokens.size(); i++) {
+                embd_inp[i] = session_tokens[i];
+            }
+
+            fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
+        } else {
+            fprintf(stderr, "%s: session file does not exist, will create\n", __func__);
+        }
+    }
+
+    // evaluate the initial prompt
+
+    printf("\n");
+    printf("%s : initializing - please wait ...\n", __func__);
+
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return 1;
+    }
+
+    if (params.verbose_prompt) {
+        fprintf(stdout, "\n");
+        fprintf(stdout, "%s", prompt_llama.c_str());
+        fflush(stdout);
+    }
+
+     // debug message about similarity of saved session, if applicable
+    size_t n_matching_session_tokens = 0;
+    if (session_tokens.size()) {
+        for (llama_token id : session_tokens) {
+            if (n_matching_session_tokens >= embd_inp.size() || id != embd_inp[n_matching_session_tokens]) {
+                break;
+            }
+            n_matching_session_tokens++;
+        }
+        if (n_matching_session_tokens >= embd_inp.size()) {
+            fprintf(stderr, "%s: session file has exact match for prompt!\n", __func__);
+        } else if (n_matching_session_tokens < (embd_inp.size() / 2)) {
+            fprintf(stderr, "%s: warning: session file has low similarity to prompt (%zu / %zu tokens); will mostly be reevaluated\n",
+                __func__, n_matching_session_tokens, embd_inp.size());
+        } else {
+            fprintf(stderr, "%s: session file matches %zu / %zu tokens of prompt\n",
+                __func__, n_matching_session_tokens, embd_inp.size());
+        }
+    }
+
+    // HACK - because session saving incurs a non-negligible delay, for now skip re-saving session
+    // if we loaded a session with at least 75% similarity. It's currently just used to speed up the
+    // initial prompt so it doesn't need to be an exact match.
+    bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < (embd_inp.size() * 3 / 4);
+
+    printf("%s : done! start speaking in the microphone\n", __func__);
+    printf("\n");
+    printf("%s%s", params.person.c_str(), chat_symb.c_str());
+    fflush(stdout);
+
+    // clear audio buffer
+    audio.clear();
+
+    // text inference variables
+    const int voice_id = 2;
+    const int n_keep   = embd_inp.size();
+    const int n_ctx    = llama_n_ctx(ctx_llama);
+
+    int n_past = n_keep;
+    int n_prev = 64; // TODO arg
+    int n_session_consumed = !path_session.empty() && session_tokens.size() > 0 ? session_tokens.size() : 0;
+
+    std::vector<llama_token> embd;
+
+    // reverse prompts for detecting when it's time to stop speaking
+    std::vector<std::string> antiprompts = {
+        params.person + chat_symb,
+    };
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        if (!is_running) {
+            break;
+        }
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        int64_t t_ms = 0;
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1250, params.vad_thold, params.freq_thold, params.print_energy) || force_speak) {
+                //fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                audio.get(params.voice_ms, pcmf32_cur);
+
+                std::string text_heard;
+
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\[.*?\\]");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove text between brackets using regex
+                {
+                    std::regex re("\\(.*?\\)");
+                    text_heard = std::regex_replace(text_heard, re, "");
+                }
+
+                // remove all characters, except for letters, numbers, punctuation and ':', '\'', '-', ' '
+                text_heard = std::regex_replace(text_heard, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+
+                // take first line
+                text_heard = text_heard.substr(0, text_heard.find_first_of('\n'));
+
+                // remove leading and trailing whitespace
+                text_heard = std::regex_replace(text_heard, std::regex("^\\s+"), "");
+                text_heard = std::regex_replace(text_heard, std::regex("\\s+$"), "");
+
+                const std::vector<llama_token> tokens = llama_tokenize(ctx_llama, text_heard.c_str(), false);
+
+                if (text_heard.empty() || tokens.empty() || force_speak) {
+                    //fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
+                    audio.clear();
+
+                    continue;
+                }
+
+                force_speak = false;
+
+                text_heard.insert(0, 1, ' ');
+                text_heard += "\n" + bot_name + chat_symb;
+                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
+                fflush(stdout);
+
+                embd = ::llama_tokenize(ctx_llama, text_heard, false);
+
+                // Append the new input tokens to the session_tokens vector
+                if (!path_session.empty()) {
+                    session_tokens.insert(session_tokens.end(), tokens.begin(), tokens.end());
+                }
+
+                // text inference
+                bool done = false;
+                std::string text_to_speak;
+                while (true) {
+                    // predict
+                    if (embd.size() > 0) {
+                        if (n_past + (int) embd.size() > n_ctx) {
+                            n_past = n_keep;
+
+                            // insert n_left/2 tokens at the start of embd from last_n_tokens
+                            embd.insert(embd.begin(), embd_inp.begin() + embd_inp.size() - n_prev, embd_inp.end());
+                            // stop saving session if we run out of context
+                            path_session = "";
+                            //printf("\n---\n");
+                            //printf("resetting: '");
+                            //for (int i = 0; i < (int) embd.size(); i++) {
+                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
+                            //}
+                            //printf("'\n");
+                            //printf("\n---\n");
+                        }
+
+                        // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
+                        // REVIEW
+                        if (n_session_consumed < (int) session_tokens.size()) {
+                            size_t i = 0;
+                            for ( ; i < embd.size(); i++) {
+                                if (embd[i] != session_tokens[n_session_consumed]) {
+                                    session_tokens.resize(n_session_consumed);
+                                    break;
+                                }
+
+                                n_past++;
+                                n_session_consumed++;
+
+                                if (n_session_consumed >= (int) session_tokens.size()) {
+                                    i++;
+                                    break;
+                                }
+                            }
+                            if (i > 0) {
+                                embd.erase(embd.begin(), embd.begin() + i);
+                            }
+                        }
+
+                        if (embd.size() > 0 && !path_session.empty()) {
+                            session_tokens.insert(session_tokens.end(), embd.begin(), embd.end());
+                            n_session_consumed = session_tokens.size();
+                        }
+
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                            fprintf(stderr, "%s : failed to eval\n", __func__);
+                            return 1;
+                        }
+                    }
+
+
+                    embd_inp.insert(embd_inp.end(), embd.begin(), embd.end());
+                    n_past += embd.size();
+                    
+                    embd.clear();
+
+                    if (done) break;
+
+                    {
+                        // out of user input, sample next token
+                        const float top_k          = 5;
+                        const float top_p          = 0.80f;
+                        const float temp           = 0.30f;
+                        const float repeat_penalty = 1.1764f;
+
+                        const int repeat_last_n    = 256;
+
+                        if (!path_session.empty() && need_to_save_session) {
+                            need_to_save_session = false;
+                            llama_save_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.size());
+                        } 
+
+                        llama_token id = 0;
+
+                        {
+                            auto logits = llama_get_logits(ctx_llama);
+                            auto n_vocab = llama_n_vocab(ctx_llama);
+
+                            logits[llama_token_eos()] = 0;
+
+                            std::vector<llama_token_data> candidates;
+                            candidates.reserve(n_vocab);
+                            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                                candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+                            }
+
+                            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+                            // apply repeat penalty
+                            const float nl_logit = logits[llama_token_nl()];
+
+                            llama_sample_repetition_penalty(ctx_llama, &candidates_p,
+                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
+                                    repeat_last_n, repeat_penalty);
+
+                            logits[llama_token_nl()] = nl_logit;
+
+                            if (temp <= 0) {
+                                // Greedy sampling
+                                id = llama_sample_token_greedy(ctx_llama, &candidates_p);
+                            } else {
+                                // Temperature sampling
+                                llama_sample_top_k(ctx_llama, &candidates_p, top_k);
+                                llama_sample_top_p(ctx_llama, &candidates_p, top_p);
+                                llama_sample_temperature(ctx_llama, &candidates_p, temp);
+                                id = llama_sample_token(ctx_llama, &candidates_p);
+                            }
+                        }
+
+                        if (id != llama_token_eos()) {
+                            // add it to the context
+                            embd.push_back(id);
+
+                            text_to_speak += llama_token_to_str(ctx_llama, id);
+
+                            printf("%s", llama_token_to_str(ctx_llama, id));
+                        }
+                    }
+
+                    {
+                        std::string last_output;
+                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
+                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
+                        }
+                        last_output += llama_token_to_str(ctx_llama, embd[0]);
+
+                        for (std::string & antiprompt : antiprompts) {
+                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
+                                done = true;
+                                text_to_speak = ::replace(text_to_speak, antiprompt, "");
+                                fflush(stdout);
+                                need_to_save_session = true;
+                                break;
+                            }
+                        }
+                    }
+
+                    is_running = sdl_poll_events();
+
+                    if (!is_running) {
+                        break;
+                    }
+                }
+
+                text_to_speak = ::replace(text_to_speak, "\"", "");
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());
+
+                audio.clear();
+
+                ++n_iter;
+            }
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx_wsp);
+    whisper_free(ctx_wsp);
+
+    llama_print_timings(ctx_llama);
+    llama_free(ctx_llama);
+
+    return 0;
+}

examples/talk.wasm/CMakeLists.txt

@@ -13,6 +13,7 @@ include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE
     whisper
+    common
     )
 
 unset(EXTRA_FLAGS)

examples/talk.wasm/gpt-2.cpp

@@ -1,4 +1,6 @@
 #include "ggml.h"
+#include "common-ggml.h"
+
 #include "gpt-2.h"
 
 #include <cmath>
@@ -14,150 +16,6 @@
 
 /////////////////////// GPT-2 BEGIN /////////////////////////
 
-//
-// Vocab utils
-//
-
-std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
-    std::vector<std::string> words;
-
-    // first split the text into words
-    {
-        std::string str = text;
-        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
-
-        std::regex re(pat);
-        std::smatch m;
-
-        while (std::regex_search(str, m, re)) {
-            for (auto x : m) {
-                words.push_back(x);
-            }
-            str = m.suffix();
-        }
-    }
-
-    // find the longest tokens that form the words:
-    std::vector<gpt_vocab::id> tokens;
-    for (const auto & word : words) {
-        if (word.size() == 0) continue;
-
-        int i = 0;
-        int n = word.size();
-        while (i < n) {
-            int j = n;
-            while (j > i) {
-                auto it = vocab.token_to_id.find(word.substr(i, j-i));
-                if (it != vocab.token_to_id.end()) {
-                    tokens.push_back(it->second);
-                    i = j;
-                    break;
-                }
-                --j;
-            }
-            if (i == n) {
-                break;
-            }
-            if (j == i) {
-                auto sub = word.substr(i, 1);
-                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
-                    tokens.push_back(vocab.token_to_id.at(sub));
-                } else {
-                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
-                }
-                ++i;
-            }
-        }
-    }
-
-    return tokens;
-}
-
-gpt_vocab::id gpt_sample_top_k_top_p(
-        const gpt_vocab & vocab,
-        const float * logits,
-        int    top_k,
-        double top_p,
-        double temp,
-        std::mt19937 & rng) {
-    int n_logits = vocab.id_to_token.size();
-
-    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
-    logits_id.reserve(n_logits);
-
-    for (int i = 0; i < n_logits; i++) {
-        logits_id.push_back(std::make_pair(logits[i], i));
-    }
-
-    // find the top K tokens
-    std::partial_sort(
-            logits_id.begin(),
-            logits_id.begin() + top_k, logits_id.end(),
-            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
-        return a.first > b.first;
-    });
-
-    logits_id.resize(top_k);
-
-    // normalize
-    {
-        double sum = 0.0f;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            sum += logits_id[i].first;
-        }
-
-        sum = 1.0/sum;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            logits_id[i].first *= sum;
-        }
-    }
-
-    if (top_p < 1.0f) {
-        {
-            double cumsum = 0.0f;
-            for (int i = 0; i < top_k; i++) {
-                cumsum += logits_id[i].first;
-                if (cumsum >= top_p) {
-                    logits_id.resize(i+1);
-                    break;
-                }
-            }
-        }
-
-        // normalize again
-        {
-            double sum = 0.0f;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                sum += logits_id[i].first;
-            }
-
-            sum = 1.0/sum;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                logits_id[i].first *= sum;
-            }
-        }
-    }
-
-    //printf("\n");
-    //for (int i = 0; i < (int)logits_id.size(); i++) {
-    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
-    //}
-    //exit(0);
-
-    // sample from the obtained distribution
-    std::vector<double> probs;
-    probs.reserve(logits_id.size());
-
-    for (int i = 0; i < (int) logits_id.size(); i++) {
-        probs.push_back(logits_id[i].first);
-    }
-
-    std::discrete_distribution<> dist(probs.begin(), probs.end());
-    int idx = dist(rng);
-
-    return logits_id[idx].second;
-}
-
 // default hparams (GPT-2 117M)
 struct gpt2_hparams {
     int32_t n_vocab = 50257;
@@ -165,7 +23,7 @@ struct gpt2_hparams {
     int32_t n_embd  = 768;
     int32_t n_head  = 12;
     int32_t n_layer = 12;
-    int32_t f16     = 1;
+    int32_t ftype   = 1;
 };
 
 struct gpt2_layer {
@@ -187,7 +45,7 @@ struct gpt2_layer {
     struct ggml_tensor * c_mlp_fc_w;
     struct ggml_tensor * c_mlp_fc_b;
 
-    struct ggml_tensor * c_mlp_proj_w_trans; // transposed for efficiency
+    struct ggml_tensor * c_mlp_proj_w;
     struct ggml_tensor * c_mlp_proj_b;
 };
 
@@ -198,8 +56,9 @@ struct gpt2_model {
     struct ggml_tensor * ln_f_g;
     struct ggml_tensor * ln_f_b;
 
-    struct ggml_tensor * wte; // position embedding
-    struct ggml_tensor * wpe; //    token embedding
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
 
     std::vector<gpt2_layer> layers;
 
@@ -241,14 +100,14 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
         fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
         fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
 
         printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
         printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
         printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
         printf("%s: n_head  = %d\n", __func__, hparams.n_head);
         printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
-        printf("%s: f16     = %d\n", __func__, hparams.f16);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
     }
 
     // load vocab
@@ -275,9 +134,14 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
     // in order to save memory and also to speed up the computation
-    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
 
     auto & ctx = model.ctx;
 
@@ -291,32 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
 
-        ctx_size += n_vocab*n_embd*ggml_type_size(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
 
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_size(GGML_TYPE_F32));   // c_attn_proj_b
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -325,9 +190,11 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
     // create the ggml context
     {
-        struct ggml_init_params params;
-        params.mem_size   = ctx_size;
-        params.mem_buffer = NULL;
+        struct ggml_init_params params = {
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+            .no_alloc   = false,
+        };
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -350,36 +217,38 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
         model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        model.wte = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
-        model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
 
         // map by name
         model.tensors["model/ln_f/g"] = model.ln_f_g;
         model.tensors["model/ln_f/b"] = model.ln_f_b;
 
-        model.tensors["model/wte"] = model.wte;
-        model.tensors["model/wpe"] = model.wpe;
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
 
         for (int i = 0; i < n_layer; ++i) {
             auto & layer = model.layers[i];
 
-            layer.ln_1_g             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_1_b             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.ln_2_g             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_2_b             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_attn_attn_w      = ggml_new_tensor_2d(ctx, wtype,         3*n_embd, n_embd);
-            layer.c_attn_attn_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
 
-            layer.c_attn_proj_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
-            layer.c_attn_proj_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_mlp_fc_w         = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
-            layer.c_mlp_fc_b         = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
 
-            layer.c_mlp_proj_w_trans = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
-            layer.c_mlp_proj_b       = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
             // map by name
             model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
@@ -397,7 +266,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
 
-            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w_trans;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
         }
     }
@@ -425,14 +294,16 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
     {
         size_t total_size = 0;
 
+        bool has_lm_head = false;
+
         while (true) {
             int32_t n_dims;
             int32_t length;
-            int32_t ftype;
+            int32_t ttype;
 
             fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
             fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
 
             if (fin.eof()) {
                 break;
@@ -461,13 +332,18 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
                 fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
                 return false;
             }
 
-            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
 
-            if (nelements*bpe != ggml_nbytes(tensor)) {
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
@@ -475,7 +351,15 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
             fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
 
-            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
             total_size += ggml_nbytes(tensor);
         }
 
@@ -493,7 +377,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 //   - n_threads: number of threads to use
 //   - n_past:    the context size so far
 //   - embd_inp:  the embeddings of the tokens in the context
-//   - embd_w:    the predicted probabilities of the next token
+//   - embd_w:    the predicted logits for the next token
 //
 bool gpt2_eval(
         const gpt2_model & model,
@@ -512,12 +396,12 @@ bool gpt2_eval(
     const int n_head  = hparams.n_head;
     const int n_vocab = hparams.n_vocab;
 
-    static size_t buf_size = 640u*1024*1024;
+    static size_t buf_size = 512u*1024*1024;
     static void * buf = malloc(buf_size);
 
     if (mem_per_token > 0 && mem_per_token*N > buf_size) {
         const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
-        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate
         buf_size = buf_size_new;
@@ -528,13 +412,14 @@ bool gpt2_eval(
         }
     }
 
-    struct ggml_init_params params;
-    params.mem_size   = buf_size;
-    params.mem_buffer = buf;
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
-
-    struct ggml_cgraph gf = { };
+    struct ggml_cgraph gf = {};
     gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@@ -578,7 +463,7 @@ bool gpt2_eval(
         // [2304, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_attn_w),
+                    model.layers[il].c_attn_attn_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -654,11 +539,13 @@ bool gpt2_eval(
             // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
             // [n_past + N, 64, 12]
             struct ggml_tensor * V_trans =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        1, 2, 0, 3);
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
 
             // KQV = transpose(V) * KQ_soft_max
             // [64, N, 12]
@@ -685,7 +572,7 @@ bool gpt2_eval(
         // [768, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
+                    model.layers[il].c_attn_proj_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -722,7 +609,7 @@ bool gpt2_eval(
             // cur = fc_w*cur + fc_b
             // [3072, N]
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
+                    model.layers[il].c_mlp_fc_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -742,7 +629,7 @@ bool gpt2_eval(
             // cur = proj_w*cur + proj_b
             // [768, N]
             cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_proj_w_trans,
+                    model.layers[il].c_mlp_proj_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -769,12 +656,12 @@ bool gpt2_eval(
     }
 
     // inpL = WTE * inpL
-    // [ 768, 50257] - model.wte
+    // [ 768, 50257] - model.lm_head
     // [ 768, N]     - inpL
-    inpL = ggml_mul_mat(ctx0, model.wte, inpL);
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
 
     // logits -> probs
-    inpL = ggml_soft_max(ctx0, inpL);
+    //inpL = ggml_soft_max(ctx0, inpL);
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
@@ -788,7 +675,7 @@ bool gpt2_eval(
     //embd_w.resize(n_vocab*N);
     //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
 
-    // return result for just the last token
+    // return result just for the last token
     embd_w.resize(n_vocab);
     memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
 
@@ -825,7 +712,7 @@ Me too.
     int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
 
     // sampling parameters
-    int32_t top_k = 40;
+    int32_t top_k = 5;
     float   top_p = 0.9f;
     float   temp  = 1.0f;
 };
@@ -833,14 +720,15 @@ Me too.
 struct gpt2_context * gpt2_init(const char * path_model) {
     gpt2_context * ctx = new gpt2_context;
 
-    ctx->rng = std::mt19937(time(NULL));
+    ctx->rng = std::mt19937(time(nullptr));
 
     // load the model
     {
         const int64_t t_start_us = ggml_time_us();
 
         if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
-            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
+            delete ctx;
             return nullptr;
         }
 
@@ -884,9 +772,9 @@ std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens)
 
     std::string result;
 
-    for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
+    for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
         // predict
-        if (embd.size() > 0) {
+        if (!embd.empty()) {
             if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
                 printf("gpt-2: failed to generate text\n");
                 return "";
@@ -913,10 +801,7 @@ std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens)
         result += ctx->vocab.id_to_token[embd[0]];
 
         // end of text token
-        if (embd.back() == 50256 ||
-            ctx->vocab.id_to_token[embd.back()] == "." ||
-            ctx->vocab.id_to_token[embd.back()] == "!" ||
-            ctx->vocab.id_to_token[embd.back()] == "?") {
+        if (embd.back() == 50256) {
             break;
         }
     }

examples/talk.wasm/gpt-2.h

@@ -2,18 +2,12 @@
 
 // TODO: Change to C-style API and move to ./examples for easy reuse.
 
+#include "common.h"
+
 #include <vector>
 #include <map>
 #include <string>
 
-struct gpt_vocab {
-    using id    = int32_t;
-    using token = std::string;
-
-    std::map<token, id> token_to_id;
-    std::map<id, token> id_to_token;
-};
-
 struct gpt2_context;
 
 struct gpt2_context * gpt2_init(const char * path_model);

examples/talk.wasm/index-tmpl.html

@@ -44,6 +44,15 @@
 
             <br><br>
 
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <br><br>
+
             <hr>
 
             Select the models you would like to use and click the "Start" button to begin the conversation
@@ -54,6 +63,10 @@
                 Whisper model: <span id="model-whisper-status"></span>
                 <button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
                 <span id="fetch-whisper-progress"></span>
 
                 <!--
@@ -266,11 +279,17 @@
                 let urls = {
                     'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
                     'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
                 };
 
                 let sizes = {
                     'tiny.en': 75,
                     'base.en': 142,
+
+                    'tiny-en-q5_1':   31,
+                    'base-en-q5_1':   57,
                 };
 
                 let url     = urls[model];
@@ -281,6 +300,10 @@
 
                 document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
                 document.getElementById('fetch-whisper-base-en').style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
+
                 document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
 
                 cbProgress = function(p) {
@@ -292,6 +315,10 @@
                     var el;
                     el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
+
                     el = document.getElementById('model-whisper-status');  if (el) el.innerHTML = '';
                 };
 

examples/talk/.gitignore

@@ -1 +1 @@
-eleven-labs.py
+audio.mp3

examples/talk/CMakeLists.txt

@@ -1,16 +1,8 @@
-if (WHISPER_SUPPORT_SDL2)
+if (WHISPER_SDL2)
     # talk
     set(TARGET talk)
-    #add_executable(${TARGET} talk.cpp gpt-2.cpp)
-    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
-    #target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
-
-    # TODO: this is temporary
-    #       need to export ggml symbols for MSVC, but too lazy ..
-    add_executable(${TARGET} talk.cpp gpt-2.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
+    add_executable(${TARGET} talk.cpp gpt-2.cpp)
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 
     include(DefaultTargetOptions)
-
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
-    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
 endif ()

examples/talk/README.md

@@ -31,7 +31,7 @@ To run this, you will need a ggml GPT-2 model: [instructions](https://github.com
 Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:
 
 ```
-wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/datasets/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
+wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
 ```
 
 ## TTS

examples/talk/eleven-labs.py

@@ -0,0 +1,23 @@
+import sys
+import importlib.util
+
+api_key = "" #Write your https://beta.elevenlabs.io api key here
+if not api_key:
+    print("To use elevenlabs you have to register to https://beta.elevenlabs.io and add your elevenlabs api key to examples/talk/eleven-labs.py")
+    sys.exit()
+
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
+    sys.exit()
+
+from elevenlabs import ElevenLabs
+eleven = ElevenLabs(api_key)
+
+# Get a Voice object, by name or UUID
+voice = eleven.voices["Arnold"] #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
+audio = voice.generate(str(sys.argv[2:]))
+
+# Save the TTS to a file
+audio.save("audio") 

examples/talk/gpt-2.cpp

@@ -1,4 +1,6 @@
 #include "ggml.h"
+#include "common-ggml.h"
+
 #include "gpt-2.h"
 
 #include <cmath>
@@ -14,150 +16,6 @@
 
 /////////////////////// GPT-2 BEGIN /////////////////////////
 
-//
-// Vocab utils
-//
-
-std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
-    std::vector<std::string> words;
-
-    // first split the text into words
-    {
-        std::string str = text;
-        std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
-
-        std::regex re(pat);
-        std::smatch m;
-
-        while (std::regex_search(str, m, re)) {
-            for (auto x : m) {
-                words.push_back(x);
-            }
-            str = m.suffix();
-        }
-    }
-
-    // find the longest tokens that form the words:
-    std::vector<gpt_vocab::id> tokens;
-    for (const auto & word : words) {
-        if (word.empty()) continue;
-
-        int i = 0;
-        int n = word.size();
-        while (i < n) {
-            int j = n;
-            while (j > i) {
-                auto it = vocab.token_to_id.find(word.substr(i, j-i));
-                if (it != vocab.token_to_id.end()) {
-                    tokens.push_back(it->second);
-                    i = j;
-                    break;
-                }
-                --j;
-            }
-            if (i == n) {
-                break;
-            }
-            if (j == i) {
-                auto sub = word.substr(i, 1);
-                if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
-                    tokens.push_back(vocab.token_to_id.at(sub));
-                } else {
-                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
-                }
-                ++i;
-            }
-        }
-    }
-
-    return tokens;
-}
-
-gpt_vocab::id gpt_sample_top_k_top_p(
-        const gpt_vocab & vocab,
-        const float * logits,
-        int    top_k,
-        double top_p,
-        double /*temp*/,
-        std::mt19937 & rng) {
-    int n_logits = vocab.id_to_token.size();
-
-    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
-    logits_id.reserve(n_logits);
-
-    for (int i = 0; i < n_logits; i++) {
-        logits_id.emplace_back(logits[i], i);
-    }
-
-    // find the top K tokens
-    std::partial_sort(
-            logits_id.begin(),
-            logits_id.begin() + top_k, logits_id.end(),
-            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
-        return a.first > b.first;
-    });
-
-    logits_id.resize(top_k);
-
-    // normalize
-    {
-        double sum = 0.0f;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            sum += logits_id[i].first;
-        }
-
-        sum = 1.0/sum;
-        for (int i = 0; i < (int)logits_id.size(); i++) {
-            logits_id[i].first *= sum;
-        }
-    }
-
-    if (top_p < 1.0f) {
-        {
-            double cumsum = 0.0f;
-            for (int i = 0; i < top_k; i++) {
-                cumsum += logits_id[i].first;
-                if (cumsum >= top_p) {
-                    logits_id.resize(i+1);
-                    break;
-                }
-            }
-        }
-
-        // normalize again
-        {
-            double sum = 0.0f;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                sum += logits_id[i].first;
-            }
-
-            sum = 1.0/sum;
-            for (int i = 0; i < (int)logits_id.size(); i++) {
-                logits_id[i].first *= sum;
-            }
-        }
-    }
-
-    //printf("\n");
-    //for (int i = 0; i < (int) logits_id.size(); i++) {
-    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
-    //}
-    //exit(0);
-
-    // sample from the obtained distribution
-    std::vector<double> probs;
-    probs.reserve(logits_id.size());
-
-    for (int i = 0; i < (int) logits_id.size(); i++) {
-        probs.push_back(logits_id[i].first);
-    }
-
-    std::discrete_distribution<> dist(probs.begin(), probs.end());
-    int idx = dist(rng);
-
-    return logits_id[idx].second;
-}
-
 // default hparams (GPT-2 117M)
 struct gpt2_hparams {
     int32_t n_vocab = 50257;
@@ -165,7 +23,7 @@ struct gpt2_hparams {
     int32_t n_embd  = 768;
     int32_t n_head  = 12;
     int32_t n_layer = 12;
-    int32_t f16     = 1;
+    int32_t ftype   = 1;
 };
 
 struct gpt2_layer {
@@ -187,7 +45,7 @@ struct gpt2_layer {
     struct ggml_tensor * c_mlp_fc_w;
     struct ggml_tensor * c_mlp_fc_b;
 
-    struct ggml_tensor * c_mlp_proj_w_trans; // transposed for efficiency
+    struct ggml_tensor * c_mlp_proj_w;
     struct ggml_tensor * c_mlp_proj_b;
 };
 
@@ -198,8 +56,9 @@ struct gpt2_model {
     struct ggml_tensor * ln_f_g;
     struct ggml_tensor * ln_f_b;
 
-    struct ggml_tensor * wte; // position embedding
-    struct ggml_tensor * wpe; //    token embedding
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
 
     std::vector<gpt2_layer> layers;
 
@@ -241,14 +100,14 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
         fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
         fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
-        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
 
         printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
         printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
         printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
         printf("%s: n_head  = %d\n", __func__, hparams.n_head);
         printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
-        printf("%s: f16     = %d\n", __func__, hparams.f16);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
     }
 
     // load vocab
@@ -268,16 +127,21 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             fin.read((char *) &len, sizeof(len));
 
             word.resize(len);
-            fin.read((char *) &word[0], len);
+            fin.read((char *) word.data(), len);
 
             vocab.token_to_id[word] = i;
             vocab.id_to_token[i] = word;
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
     // in order to save memory and also to speed up the computation
-    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
 
     auto & ctx = model.ctx;
 
@@ -291,32 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
 
-        ctx_size += n_vocab*n_embd*ggml_type_size(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
 
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_size(GGML_TYPE_F32));   // c_attn_proj_b
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -325,9 +190,11 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
     // create the ggml context
     {
-        struct ggml_init_params params;
-        params.mem_size   = ctx_size;
-        params.mem_buffer = nullptr;
+        struct ggml_init_params params = {
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+            .no_alloc   = false,
+        };
 
         model.ctx = ggml_init(params);
         if (!model.ctx) {
@@ -350,36 +217,38 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
         model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        model.wte = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
-        model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
 
         // map by name
         model.tensors["model/ln_f/g"] = model.ln_f_g;
         model.tensors["model/ln_f/b"] = model.ln_f_b;
 
-        model.tensors["model/wte"] = model.wte;
-        model.tensors["model/wpe"] = model.wpe;
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
 
         for (int i = 0; i < n_layer; ++i) {
             auto & layer = model.layers[i];
 
-            layer.ln_1_g             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_1_b             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.ln_2_g             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
-            layer.ln_2_b             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_attn_attn_w      = ggml_new_tensor_2d(ctx, wtype,         3*n_embd, n_embd);
-            layer.c_attn_attn_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
 
-            layer.c_attn_proj_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
-            layer.c_attn_proj_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_mlp_fc_w         = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
-            layer.c_mlp_fc_b         = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
 
-            layer.c_mlp_proj_w_trans = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
-            layer.c_mlp_proj_b       = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
             // map by name
             model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
@@ -397,7 +266,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
 
-            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w_trans;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
             model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
         }
     }
@@ -425,14 +294,16 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
     {
         size_t total_size = 0;
 
+        bool has_lm_head = false;
+
         while (true) {
             int32_t n_dims;
             int32_t length;
-            int32_t ftype;
+            int32_t ttype;
 
             fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
             fin.read(reinterpret_cast<char *>(&length), sizeof(length));
-            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
 
             if (fin.eof()) {
                 break;
@@ -448,7 +319,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             std::string name(length, 0);
             fin.read(&name[0], length);
 
-            if (model.tensors.find(name) == model.tensors.end()) {
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
                 fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
                 return false;
             }
@@ -461,13 +332,18 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
                 fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
-                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
                 return false;
             }
 
-            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
 
-            if (nelements*bpe != ggml_nbytes(tensor)) {
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
@@ -475,7 +351,15 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
             fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
 
-            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
             total_size += ggml_nbytes(tensor);
         }
 
@@ -493,7 +377,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 //   - n_threads: number of threads to use
 //   - n_past:    the context size so far
 //   - embd_inp:  the embeddings of the tokens in the context
-//   - embd_w:    the predicted probabilities of the next token
+//   - embd_w:    the predicted logits for the next token
 //
 bool gpt2_eval(
         const gpt2_model & model,
@@ -512,12 +396,12 @@ bool gpt2_eval(
     const int n_head  = hparams.n_head;
     const int n_vocab = hparams.n_vocab;
 
-    static size_t buf_size = 5640ull*1024*1024;
+    static size_t buf_size = 512u*1024*1024;
     static void * buf = malloc(buf_size);
 
     if (mem_per_token > 0 && mem_per_token*N > buf_size) {
         const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
-        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate
         buf_size = buf_size_new;
@@ -528,13 +412,14 @@ bool gpt2_eval(
         }
     }
 
-    struct ggml_init_params params;
-    params.mem_size   = buf_size;
-    params.mem_buffer = buf;
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
 
     struct ggml_context * ctx0 = ggml_init(params);
-
-    struct ggml_cgraph gf = { };
+    struct ggml_cgraph gf = {};
     gf.n_threads = n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@@ -578,7 +463,7 @@ bool gpt2_eval(
         // [2304, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_attn_w),
+                    model.layers[il].c_attn_attn_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -654,11 +539,13 @@ bool gpt2_eval(
             // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
             // [n_past + N, 64, 12]
             struct ggml_tensor * V_trans =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        1, 2, 0, 3);
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
 
             // KQV = transpose(V) * KQ_soft_max
             // [64, N, 12]
@@ -685,7 +572,7 @@ bool gpt2_eval(
         // [768, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
+                    model.layers[il].c_attn_proj_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -722,7 +609,7 @@ bool gpt2_eval(
             // cur = fc_w*cur + fc_b
             // [3072, N]
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
+                    model.layers[il].c_mlp_fc_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -742,7 +629,7 @@ bool gpt2_eval(
             // cur = proj_w*cur + proj_b
             // [768, N]
             cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_proj_w_trans,
+                    model.layers[il].c_mlp_proj_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -769,12 +656,12 @@ bool gpt2_eval(
     }
 
     // inpL = WTE * inpL
-    // [ 768, 50257] - model.wte
+    // [ 768, 50257] - model.lm_head
     // [ 768, N]     - inpL
-    inpL = ggml_mul_mat(ctx0, model.wte, inpL);
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
 
     // logits -> probs
-    inpL = ggml_soft_max(ctx0, inpL);
+    //inpL = ggml_soft_max(ctx0, inpL);
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
@@ -788,7 +675,7 @@ bool gpt2_eval(
     //embd_w.resize(n_vocab*N);
     //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
 
-    // return result for just the last token
+    // return result just for the last token
     embd_w.resize(n_vocab);
     memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
 

examples/talk/gpt-2.h

@@ -2,18 +2,12 @@
 
 // TODO: Change to C-style API and move to ./examples for easy reuse.
 
+#include "common.h"
+
 #include <vector>
 #include <map>
 #include <string>
 
-struct gpt_vocab {
-    using id    = int32_t;
-    using token = std::string;
-
-    std::map<token, id> token_to_id;
-    std::map<id, token> id_to_token;
-};
-
 struct gpt2_context;
 
 struct gpt2_context * gpt2_init(const char * path_model);

examples/talk/speak.sh

@@ -7,9 +7,13 @@
 # Mac OS: brew install espeak
 # Linux: apt-get install espeak
 #
-espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"
+
+# Mac OS "say" command
+say "$2"
 
 # Eleven Labs
+# To use it, install the elevenlabs module from pip (pip install elevenlabs), register to https://beta.elevenlabs.io to get an api key and paste it in /examples/talk/eleven-labs.py 
 #
 #wd=$(dirname $0)
 #script=$wd/eleven-labs.py

examples/whisper.android/README.md

@@ -9,4 +9,6 @@ To use:
 5. Select the "release" active build variant, and use Android Studio to run and deploy to your device.
 [^1]: I recommend the tiny or base models for running on an Android device.
 
-<img width="300" alt="image" src="https://user-images.githubusercontent.com/1991296/208154256-82d972dc-221b-48c4-bfcb-36ce68602f93.png">
+(PS: Do not move this android project folder individually to other folders, because this android project folder depends on the files of the whole project.)
+
+<img width="300" alt="image" src="https://user-images.githubusercontent.com/1670775/221613663-a17bf770-27ef-45ab-9a46-a5f99ba65d2a.jpg">

examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreen.kt

@@ -2,6 +2,7 @@ package com.whispercppdemo.ui.main
 
 import androidx.compose.foundation.layout.*
 import androidx.compose.foundation.rememberScrollState
+import androidx.compose.foundation.text.selection.SelectionContainer
 import androidx.compose.foundation.verticalScroll
 import androidx.compose.material3.*
 import androidx.compose.runtime.Composable
@@ -19,6 +20,7 @@ fun MainScreen(viewModel: MainScreenViewModel) {
         canTranscribe = viewModel.canTranscribe,
         isRecording = viewModel.isRecording,
         messageLog = viewModel.dataLog,
+        onBenchmarkTapped = viewModel::benchmark,
         onTranscribeSampleTapped = viewModel::transcribeSample,
         onRecordTapped = viewModel::toggleRecord
     )
@@ -30,6 +32,7 @@ private fun MainScreen(
     canTranscribe: Boolean,
     isRecording: Boolean,
     messageLog: String,
+    onBenchmarkTapped: () -> Unit,
     onTranscribeSampleTapped: () -> Unit,
     onRecordTapped: () -> Unit
 ) {
@@ -45,8 +48,11 @@ private fun MainScreen(
                 .padding(innerPadding)
                 .padding(16.dp)
         ) {
-            Row(horizontalArrangement = Arrangement.SpaceBetween) {
-                TranscribeSampleButton(enabled = canTranscribe, onClick = onTranscribeSampleTapped)
+            Column(verticalArrangement = Arrangement.SpaceBetween) {
+                Row(horizontalArrangement = Arrangement.SpaceBetween, modifier = Modifier.fillMaxWidth()) {
+                    BenchmarkButton(enabled = canTranscribe, onClick = onBenchmarkTapped)
+                    TranscribeSampleButton(enabled = canTranscribe, onClick = onTranscribeSampleTapped)
+                }
                 RecordButton(
                     enabled = canTranscribe,
                     isRecording = isRecording,
@@ -60,7 +66,16 @@ private fun MainScreen(
 
 @Composable
 private fun MessageLog(log: String) {
-    Text(modifier = Modifier.verticalScroll(rememberScrollState()), text = log)
+    SelectionContainer() {
+        Text(modifier = Modifier.verticalScroll(rememberScrollState()), text = log)
+    }
+}
+
+@Composable
+private fun BenchmarkButton(enabled: Boolean, onClick: () -> Unit) {
+    Button(onClick = onClick, enabled = enabled) {
+        Text("Benchmark")
+    }
 }
 
 @Composable

examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt

@@ -41,10 +41,15 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
 
     init {
         viewModelScope.launch {
+            printSystemInfo()
             loadData()
         }
     }
 
+    private suspend fun printSystemInfo() {
+        printMessage(String.format("System Info: %s\n", WhisperContext.getSystemInfo()));
+    }
+
     private suspend fun loadData() {
         printMessage("Loading data...\n")
         try {
@@ -81,10 +86,29 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         //whisperContext = WhisperContext.createContextFromFile(firstModel.absolutePath)
     }
 
+    fun benchmark() = viewModelScope.launch {
+        runBenchmark(6)
+    }
+
     fun transcribeSample() = viewModelScope.launch {
         transcribeAudio(getFirstSample())
     }
 
+    private suspend fun runBenchmark(nthreads: Int) {
+        if (!canTranscribe) {
+            return
+        }
+
+        canTranscribe = false
+
+        printMessage("Running benchmark. This will take minutes...\n")
+        whisperContext?.benchMemory(nthreads)?.let{ printMessage(it) }
+        printMessage("\n")
+        whisperContext?.benchGgmlMulMat(nthreads)?.let{ printMessage(it) }
+
+        canTranscribe = true
+    }
+
     private suspend fun getFirstSample(): File = withContext(Dispatchers.IO) {
         samplesPath.listFiles()!!.first()
     }
@@ -114,11 +138,14 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
         canTranscribe = false
 
         try {
-            printMessage("Reading wave samples...\n")
+            printMessage("Reading wave samples... ")
             val data = readAudioSamples(file)
+            printMessage("${data.size / (16000 / 1000)} ms\n")
             printMessage("Transcribing data...\n")
+            val start = System.currentTimeMillis()
             val text = whisperContext?.transcribeData(data)
-            printMessage("Done: $text\n")
+            val elapsed = System.currentTimeMillis() - start
+            printMessage("Done ($elapsed ms): $text\n")
         } catch (e: Exception) {
             Log.w(LOG_TAG, e)
             printMessage("${e.localizedMessage}\n")

examples/whisper.android/app/src/main/java/com/whispercppdemo/whisper/LibWhisper.kt

@@ -27,6 +27,14 @@ class WhisperContext private constructor(private var ptr: Long) {
         }
     }
 
+    suspend fun benchMemory(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchMemcpy(nthreads)
+    }
+
+    suspend fun benchGgmlMulMat(nthreads: Int): String = withContext(scope.coroutineContext) {
+        return@withContext WhisperLib.benchGgmlMulMat(nthreads)
+    }
+
     suspend fun release() = withContext(scope.coroutineContext) {
         if (ptr != 0L) {
             WhisperLib.freeContext(ptr)
@@ -66,6 +74,10 @@ class WhisperContext private constructor(private var ptr: Long) {
             }
             return WhisperContext(ptr)
         }
+
+        fun getSystemInfo(): String {
+            return WhisperLib.getSystemInfo()
+        }
     }
 }
 
@@ -74,6 +86,7 @@ private class WhisperLib {
         init {
             Log.d(LOG_TAG, "Primary ABI: ${Build.SUPPORTED_ABIS[0]}")
             var loadVfpv4 = false
+            var loadV8fp16 = false
             if (isArmEabiV7a()) {
                 // armeabi-v7a needs runtime detection support
                 val cpuInfo = cpuInfo()
@@ -84,11 +97,24 @@ private class WhisperLib {
                         loadVfpv4 = true
                     }
                 }
+            } else if (isArmEabiV8a()) {
+                // ARMv8.2a needs runtime detection support
+                val cpuInfo = cpuInfo()
+                cpuInfo?.let {
+                    Log.d(LOG_TAG, "CPU info: $cpuInfo")
+                    if (cpuInfo.contains("fphp")) {
+                        Log.d(LOG_TAG, "CPU supports fp16 arithmetic")
+                        loadV8fp16 = true
+                    }
+                }
             }
 
             if (loadVfpv4) {
                 Log.d(LOG_TAG, "Loading libwhisper_vfpv4.so")
                 System.loadLibrary("whisper_vfpv4")
+            } else if (loadV8fp16) {
+                Log.d(LOG_TAG, "Loading libwhisper_v8fp16_va.so")
+                System.loadLibrary("whisper_v8fp16_va")
             } else {
                 Log.d(LOG_TAG, "Loading libwhisper.so")
                 System.loadLibrary("whisper")
@@ -103,6 +129,9 @@ private class WhisperLib {
         external fun fullTranscribe(contextPtr: Long, audioData: FloatArray)
         external fun getTextSegmentCount(contextPtr: Long): Int
         external fun getTextSegment(contextPtr: Long, index: Int): String
+        external fun getSystemInfo(): String
+        external fun benchMemcpy(nthread: Int): String
+        external fun benchGgmlMulMat(nthread: Int): String
     }
 }
 
@@ -110,6 +139,10 @@ private fun isArmEabiV7a(): Boolean {
     return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
 
+private fun isArmEabiV8a(): Boolean {
+    return Build.SUPPORTED_ABIS[0].equals("arm64-v8a")
+}
+
 private fun cpuInfo(): String? {
     return try {
         File("/proc/cpuinfo").inputStream().bufferedReader().use {

examples/whisper.android/app/src/main/jni/whisper/Android.mk

@@ -12,4 +12,15 @@ ifeq ($(TARGET_ARCH_ABI),armeabi-v7a)
 	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
 	LOCAL_CFLAGS += -mfpu=neon-vfpv4
 	include $(BUILD_SHARED_LIBRARY)
-endif
+endif
+
+ifeq ($(TARGET_ARCH_ABI),arm64-v8a)
+	include $(CLEAR_VARS)
+	LOCAL_MODULE    := libwhisper_v8fp16_va
+	include $(LOCAL_PATH)/Whisper.mk
+	# Allow building NEON FMA code.
+	# https://android.googlesource.com/platform/ndk/+/master/sources/android/cpufeatures/cpu-features.h
+	LOCAL_CFLAGS += -march=armv8.2-a+fp16
+	include $(BUILD_SHARED_LIBRARY)
+endif
+

examples/whisper.android/app/src/main/jni/whisper/jni.c

@@ -6,6 +6,7 @@
 #include <sys/sysinfo.h>
 #include <string.h>
 #include "whisper.h"
+#include "ggml.h"
 
 #define UNUSED(x) (void)(x)
 #define TAG "JNI"
@@ -213,4 +214,30 @@ Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getTextSegment(
     const char *text = whisper_full_get_segment_text(context, index);
     jstring string = (*env)->NewStringUTF(env, text);
     return string;
-}
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_getSystemInfo(
+        JNIEnv *env, jobject thiz
+) {
+    UNUSED(thiz);
+    const char *sysinfo = whisper_print_system_info();
+    jstring string = (*env)->NewStringUTF(env, sysinfo);
+    return string;
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchMemcpy(JNIEnv *env, jobject thiz,
+                                                                      jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_memcpy = whisper_bench_memcpy_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_memcpy);
+}
+
+JNIEXPORT jstring JNICALL
+Java_com_whispercppdemo_whisper_WhisperLib_00024Companion_benchGgmlMulMat(JNIEnv *env, jobject thiz,
+                                                                          jint n_threads) {
+    UNUSED(thiz);
+    const char *bench_ggml_mul_mat = whisper_bench_ggml_mul_mat_str(n_threads);
+    jstring string = (*env)->NewStringUTF(env, bench_ggml_mul_mat);
+}

examples/whisper.objc/README.md

@@ -24,3 +24,5 @@ Also, don't forget to add the `-DGGML_USE_ACCELERATE` compiler flag in Build Pha
 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">
+
+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.

examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj

@@ -296,6 +296,10 @@
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
+				OTHER_CFLAGS = (
+					"-O3",
+					"-DNDEBUG",
+				);
 				SDKROOT = iphoneos;
 				VALIDATE_PRODUCT = YES;
 			};

examples/whisper.swiftui/README.md

@@ -1,14 +1,18 @@
 A sample SwiftUI app using [whisper.cpp](https://github.com/ggerganov/whisper.cpp/) to do voice-to-text transcriptions.
 See also: [whisper.objc](https://github.com/ggerganov/whisper.cpp/tree/master/examples/whisper.objc).
 
-To use:
+**Usage**:
 
 1. Select a model from the [whisper.cpp repository](https://github.com/ggerganov/whisper.cpp/tree/master/models).[^1]
-2. Add the model to "whisper.swiftui.demo/Resources/models" via Xcode.
+2. Add the model to `whisper.swiftui.demo/Resources/models` **via Xcode**.
 3. Select a sample audio file (for example, [jfk.wav](https://github.com/ggerganov/whisper.cpp/raw/master/samples/jfk.wav)).
-4. Add the model to "whisper.swiftui.demo/Resources/samples" via Xcode.
-5. Select the "release" build configuration under "Run", then deploy and run to your device.
+4. Add the sample audio file to `whisper.swiftui.demo/Resources/samples` **via Xcode**.
+5. Select the "Release" [^2] build configuration under "Run", then deploy and run to your device.
+
+**Note:** Pay attention to the folder path: `whisper.swiftui.demo/Resources/models` is the appropriate directory to place resources whilst `whisper.swiftui.demo/Models` is related to actual code.
 
 [^1]: I recommend the tiny, base or small models for running on an iOS device.
 
+[^2]: The `Release` build can boost performance of transcription. 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.
+
 ![image](https://user-images.githubusercontent.com/1991296/212539216-0aef65e4-f882-480a-8358-0f816838fd52.png)

examples/whisper.swiftui/whisper.swiftui.xcodeproj/project.pbxproj

@@ -430,6 +430,10 @@
 				LLVM_LTO = YES;
 				MACOSX_DEPLOYMENT_TARGET = 13.0;
 				MARKETING_VERSION = 1.0;
+				OTHER_CFLAGS = (
+					"-O3",
+					"-DNDEBUG",
+				);
 				PRODUCT_BUNDLE_IDENTIFIER = com.whispercppdemo.WhisperCppDemo;
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SDKROOT = auto;

examples/whisper.wasm/CMakeLists.txt

@@ -31,9 +31,9 @@ endif()
 set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
     --bind \
     -s USE_PTHREADS=1 \
-    -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1500MB \
-    -s TOTAL_MEMORY=1500MB \
+    -s PTHREAD_POOL_SIZE_STRICT=0 \
+    -s INITIAL_MEMORY=2000MB \
+    -s TOTAL_MEMORY=2000MB \
     -s FORCE_FILESYSTEM=1 \
     -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
     ${EXTRA_FLAGS} \

examples/whisper.wasm/README.md

@@ -37,6 +37,6 @@ emcmake cmake ..
 make -j
 
 # copy the produced page to your HTTP path
-cp bin/whisper.wasm/*       /path/to/html/
-cp bin/libwhisper.worker.js /path/to/html/
+cp bin/whisper.wasm/*    /path/to/html/
+cp bin/libmain.worker.js /path/to/html/
 ```

examples/whisper.wasm/emscripten.cpp

@@ -10,6 +10,12 @@ std::thread g_worker;
 
 std::vector<struct whisper_context *> g_contexts(4, nullptr);
 
+static inline int mpow2(int n) {
+    int p = 1;
+    while (p <= n) p *= 2;
+    return p/2;
+}
+
 EMSCRIPTEN_BINDINGS(whisper) {
     emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
         if (g_worker.joinable()) {
@@ -43,7 +49,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
         }
     }));
 
-    emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, bool translate) {
+    emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, int nthreads, bool translate) {
         if (g_worker.joinable()) {
             g_worker.join();
         }
@@ -66,7 +72,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
         params.print_special    = false;
         params.translate        = translate;
         params.language         = whisper_is_multilingual(g_contexts[index]) ? lang.c_str() : "en";
-        params.n_threads        = std::min(8, (int) std::thread::hardware_concurrency());
+        params.n_threads        = std::min(nthreads, std::min(16, mpow2(std::thread::hardware_concurrency())));
         params.offset_ms        = 0;
 
         std::vector<float> pcmf32;

examples/whisper.wasm/index-tmpl.html

@@ -40,21 +40,42 @@
 
             Note that the computation is quite heavy and may take a few seconds to complete.<br>
             The transcription results will be displayed in the text area below.<br><br>
-            <b>Important: your browser must support WASM SIMD instructions for this to work.</b>
+            <b>Important:</b>
+                <ul>
+                    <li>your browser must support WASM SIMD instructions for this to work</li>
+                    <li>Firefox cannot load files larger than 256 MB - use Chrome instead</li>
+                </ul>
 
-            <br><br><hr>
+            <b>More examples:</b>
+                <a href="https://whisper.ggerganov.com/">main</a> |
+                <a href="https://whisper.ggerganov.com/bench">bench</a> |
+                <a href="https://whisper.ggerganov.com/stream">stream</a> |
+                <a href="https://whisper.ggerganov.com/command">command</a> |
+                <a href="https://whisper.ggerganov.com/talk">talk</a> |
+
+            <hr>
 
             <div id="model">
-                Whisper model: <span id="model-whisper-status"></span>
+                Whisper models: <span id="model-whisper-status"></span><br><br>
                 <button id="fetch-whisper-tiny-en"  onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
                 <button id="fetch-whisper-tiny"     onclick="loadWhisper('tiny')">tiny (75 MB)</button>
                 <button id="fetch-whisper-base-en"  onclick="loadWhisper('base.en')">base.en (142 MB)</button>
                 <button id="fetch-whisper-base"     onclick="loadWhisper('base')">base (142 MB)</button>
                 <button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
                 <button id="fetch-whisper-small"    onclick="loadWhisper('small')">small (466 MB)</button>
-                <span id="fetch-whisper-progress"></span>
-
                 <input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
+                <br><br>
+                Quantized models:<br><br>
+                <button id="fetch-whisper-tiny-en-q5_1"   onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-tiny-q5_1"      onclick="loadWhisper('tiny-q5_1')">tiny (Q5_1, 31 MB)</button>
+                <button id="fetch-whisper-base-en-q5_1"   onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
+                <button id="fetch-whisper-base-q5_1"      onclick="loadWhisper('base-q5_1')">base (Q5_1, 57 MB)</button>
+                <button id="fetch-whisper-small-en-q5_1"  onclick="loadWhisper('small-en-q5_1')">small.en (Q5_1, 182 MB)</button>
+                <button id="fetch-whisper-small-q5_1"     onclick="loadWhisper('small-q5_1')">small (Q5_1, 182 MB)</button><br>
+                <button id="fetch-whisper-medium-en-q5_0" onclick="loadWhisper('medium-en-q5_0')">medium.en (Q5_0, 515 MB)</button>
+                <button id="fetch-whisper-medium-q5_0"    onclick="loadWhisper('medium-q5_0')">medium (Q5_0, 515 MB)</button>
+                <button id="fetch-whisper-large-q5_0"     onclick="loadWhisper('large-q5_0')">large (Q5_0, 1030 MB)</button>
+                <span id="fetch-whisper-progress"></span>
             </div>
 
             <br>
@@ -161,6 +182,12 @@
                             <option value="yi">Yiddish</option>
                         </select>
                     </td>
+                    <!-- Slider to select number of threads between 1 and 16 -->
+                    <td>
+                        Threads:
+                        <input type="range" id="threads" name="threads" min="1" max="16" value="8" onchange="changeThreads(this.value)" />
+                        <span id="threads-value">8</span>
+                    </td>
                     <td>
                         <button onclick="onProcess(false);">Transcribe</button>
                     </td>
@@ -263,11 +290,13 @@
 
                 Module.FS_createDataFile("/", fname, buf, true, true);
 
-                model_whisper = fname;
+                //model_whisper = fname;
 
                 document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
 
                 printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
+
+                document.getElementById('model').innerHTML = 'Model fetched: ' + model_whisper;
             }
 
             function loadFile(event, fname) {
@@ -292,6 +321,17 @@
                 document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-base'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-small'   ).style.display = 'none';
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-tiny-q5_1'     ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-q5_1'     ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-q5_1'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-medium-q5_0'   ).style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
                 document.getElementById('whisper-file'          ).style.display = 'none';
                 document.getElementById('model-whisper-status'  ).innerHTML = 'loaded model: ' + file.name;
             }
@@ -304,6 +344,16 @@
                     'base':     'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
                     'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
                     'small':    'https://whisper.ggerganov.com/ggml-model-whisper-small.bin',
+
+                    'tiny-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
+                    'tiny-q5_1':     'https://whisper.ggerganov.com/ggml-model-whisper-tiny-q5_1.bin',
+                    'base-en-q5_1':  'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
+                    'base-q5_1':     'https://whisper.ggerganov.com/ggml-model-whisper-base-q5_1.bin',
+                    'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
+                    'small-q5_1':    'https://whisper.ggerganov.com/ggml-model-whisper-small-q5_1.bin',
+                    'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
+                    'medium-q5_0':   'https://whisper.ggerganov.com/ggml-model-whisper-medium-q5_0.bin',
+                    'large-q5_0':    'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
                 };
 
                 let sizes = {
@@ -313,6 +363,16 @@
                     'base':     142,
                     'small.en': 466,
                     'small':    466,
+
+                    'tiny-en-q5_1':   31,
+                    'tiny-q5_1':      31,
+                    'base-en-q5_1':   57,
+                    'base-q5_1':      57,
+                    'small-en-q5_1':  182,
+                    'small-q5_1':     182,
+                    'medium-en-q5_0': 515,
+                    'medium-q5_0':    515,
+                    'large-q5_0':     1030,
                 };
 
                 let url     = urls[model];
@@ -327,8 +387,19 @@
                 document.getElementById('fetch-whisper-tiny'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-base'    ).style.display = 'none';
                 document.getElementById('fetch-whisper-small'   ).style.display = 'none';
-                document.getElementById('whisper-file'          ).style.display = 'none';
-                document.getElementById('model-whisper-status'  ).innerHTML = 'loading model: ' + model;
+
+                document.getElementById('fetch-whisper-tiny-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-tiny-q5_1'     ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-en-q5_1'  ).style.display = 'none';
+                document.getElementById('fetch-whisper-base-q5_1'     ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
+                document.getElementById('fetch-whisper-small-q5_1'    ).style.display = 'none';
+                document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
+                document.getElementById('fetch-whisper-medium-q5_0'   ).style.display = 'none';
+                document.getElementById('fetch-whisper-large-q5_0'    ).style.display = 'none';
+
+                document.getElementById('whisper-file'        ).style.display = 'none';
+                document.getElementById('model-whisper-status').innerHTML = 'loading model: ' + model;
 
                 cbProgress = function(p) {
                     let el = document.getElementById('fetch-whisper-progress');
@@ -337,14 +408,26 @@
 
                 cbCancel = function() {
                     var el;
+
                     el = document.getElementById('fetch-whisper-tiny-en' ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base-en' ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-tiny'    ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-base'    ); if (el) el.style.display = 'inline-block';
                     el = document.getElementById('fetch-whisper-small'   ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('whisper-file'          ); if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-whisper-status'  ); if (el) el.innerHTML = '';
+
+                    el = document.getElementById('fetch-whisper-tiny-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-tiny-q5_1'     ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-en-q5_1'  ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-base-q5_1'     ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-en-q5_1' ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-small-q5_1'    ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-medium-en-q5_0'); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-medium-q5_0'   ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('fetch-whisper-large-q5_0'    ); if (el) el.style.display = 'inline-block';
+
+                    el = document.getElementById('whisper-file'        ); if (el) el.style.display = 'inline-block';
+                    el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
                 };
 
                 loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
@@ -354,7 +437,8 @@
             // audio file
             //
 
-            const kMaxAudio_s = 120;
+            const kMaxAudio_s = 30*60;
+            const kMaxRecording_s = 2*60;
             const kSampleRate = 16000;
 
             window.AudioContext = window.AudioContext || window.webkitAudioContext;
@@ -423,7 +507,7 @@
                 doRecording = false;
             }
 
-            // record up to kMaxAudio_s seconds of audio from the microphone
+            // record up to kMaxRecording_s seconds of audio from the microphone
             // check if doRecording is false every 1000 ms and stop recording if so
             // update progress information
             function startRecording() {
@@ -479,9 +563,9 @@
                                         printTextarea('js: audio recorded, size: ' + audio.length);
 
                                         // truncate to first 30 seconds
-                                        if (audio.length > kMaxAudio_s*kSampleRate) {
-                                            audio = audio.slice(0, kMaxAudio_s*kSampleRate);
-                                            printTextarea('js: truncated audio to first ' + kMaxAudio_s + ' seconds');
+                                        if (audio.length > kMaxRecording_s*kSampleRate) {
+                                            audio = audio.slice(0, kMaxRecording_s*kSampleRate);
+                                            printTextarea('js: truncated audio to first ' + kMaxRecording_s + ' seconds');
                                         }
                                         setAudio(audio);
                                     });
@@ -509,24 +593,31 @@
                         });
                     }
 
-                    document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxAudio_s) + '%';
-                    document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxAudio_s).toFixed(0) + '%';
+                    document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxRecording_s) + '%';
+                    document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxRecording_s).toFixed(0) + '%';
                 }, 1000);
 
                 printTextarea('js: recording ...');
 
                 setTimeout(function() {
                     if (doRecording) {
-                        printTextarea('js: recording stopped after ' + kMaxAudio_s + ' seconds');
+                        printTextarea('js: recording stopped after ' + kMaxRecording_s + ' seconds');
                         stopRecording();
                     }
-                }, kMaxAudio_s*1000);
+                }, kMaxRecording_s*1000);
             }
 
             //
             // transcribe
             //
 
+            var nthreads = 8;
+
+            function changeThreads(value) {
+                nthreads = value;
+                document.getElementById('threads-value').innerHTML = nthreads;
+            }
+
             function onProcess(translate) {
                 if (!instance) {
                     instance = Module.init('whisper.bin');
@@ -553,7 +644,7 @@
                     printTextarea('');
 
                     setTimeout(function() {
-                        var ret = Module.full_default(instance, audio, document.getElementById('language').value, translate);
+                        var ret = Module.full_default(instance, audio, document.getElementById('language').value, nthreads, translate);
                         console.log('js: full_default returned: ' + ret);
                         if (ret) {
                             printTextarea("js: whisper returned: " + ret);

extra/bench-all.sh

@@ -2,7 +2,7 @@
 
 # Helper script to run the bench tool on all models and print the results in share-able format
 
-printf "Usage: ./bench.sh [n_threads]\n"
+printf "Usage: ./bench.sh [n_threads] [encoder-only]\n"
 
 if [ -z "$1" ]; then
     n_threads=4
@@ -10,24 +10,39 @@ else
     n_threads=$1
 fi
 
-models=( "tiny" "base" "small" "medium" "large" )
+encoder_only=0
+if [ -z "$2" ]; then
+    encoder_only=0
+else
+    encoder_only=$2
+fi
 
-printf "\n"
-printf "Running memcpy benchmark with 1 thread\n"
-printf "\n"
+models=(                                               \
+      "tiny"   "tiny-q5_0"   "tiny-q5_1"   "tiny-q8_0" \
+      "base"   "base-q5_0"   "base-q5_1"   "base-q8_0" \
+     "small"  "small-q5_0"  "small-q5_1"  "small-q8_0" \
+    "medium" "medium-q5_0" "medium-q5_1" "medium-q8_0" \
+     "large"  "large-q5_0"  "large-q5_1"  "large-q8_0" \
+)
 
-./bench -w 1 -t 1 2>&1
+if [ "$encoder_only" -eq 0 ]; then
+    printf "\n"
+    printf "Running memcpy benchmark\n"
+    printf "\n"
 
-printf "\n"
-printf "Running ggml_mul_mat benchmark with $n_threads threads\n"
-printf "\n"
+    ./bench -w 1 -t $n_threads 2>&1
 
-./bench -w 2 -t $n_threads 2>&1
+    printf "\n"
+    printf "Running ggml_mul_mat benchmark with $n_threads threads\n"
+    printf "\n"
 
-printf "\n"
-printf "Running benchmark for all models\n"
-printf "This can take a while!\n"
-printf "\n"
+    ./bench -w 2 -t $n_threads 2>&1
+
+    printf "\n"
+    printf "Running benchmark for all models\n"
+    printf "This can take a while!\n"
+    printf "\n"
+fi
 
 printf "| CPU | OS | Config | Model | Th | Load | Enc. | Commit |\n"
 printf "| --- | -- | ------ | ----- | -- | ---- | ---- | ------ |\n"
@@ -39,6 +54,7 @@ for model in "${models[@]}"; do
     # actual run
     # store stderr output in a variable in order to parse it later
     output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
+    ret=$?
 
     # parse the output:
     load_time=$(echo "$output" | grep "load time" | awk '{print $5}')
@@ -64,7 +80,13 @@ for model in "${models[@]}"; do
         config="$config BLAS"
     fi
 
+    if [[ $system_info == *"COREML = 1"* ]]; then
+        config="$config COREML"
+    fi
+
     commit=$(git rev-parse --short HEAD)
 
-    printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
+    if [ $ret -eq 0 ]; then
+        printf "| <todo> | <todo> | $config | $model | $n_threads | $load_time | $encode_time | $commit |\n"
+    fi
 done

extra/bench-wts.sh

@@ -0,0 +1,70 @@
+# Benchmark word-level timestamps for different models
+#
+# This script takes two arguments
+# - an audio file
+# - [optional] path to a font file
+
+# I'm using "/usr/share/fonts/truetype/freefont/FreeMono.ttf" on Ubuntu
+
+if [ -z "$1" ]; then
+    echo "Usage: $0 <audio file> [font file]"
+    exit 1
+fi
+
+#TODO: Make this a command line parameter
+#models="base small large"
+#models="tiny.en tiny base.en base small.en small medium.en medium large-v1 large"
+models="tiny.en base.en small.en medium.en large"
+
+DURATION=$(ffprobe -i $1 -show_entries format=duration -v quiet -of csv="p=0")
+DURATION=$(printf "%.2f" $DURATION)
+echo "Input file duration: ${DURATION}s"
+
+for model in $models; do
+    echo "Running $model"
+    COMMAND="./main -m models/ggml-$model.bin -owts -f $1 -of $1.$model"
+
+    if [ ! -z "$2" ]; then
+        COMMAND="$COMMAND -fp $2"
+    fi
+    #TODO: Surface errors better
+    # TIMEFMT is for zsh, TIMEFORMAT is for bash
+    EXECTIME=$({ TIMEFMT="%E";TIMEFORMAT=%E; time $COMMAND >/dev/null 2>&1; } 2>&1)
+
+    # Slightly different formats between zsh and bash
+    if [ "${EXECTIME: -1}" == "s" ]; then
+        EXECTIME=${EXECTIME::-1}
+    fi
+
+    RATIO=$(echo "$DURATION / $EXECTIME" | bc -l)
+    RATIO=$(printf "%.2f" $RATIO)
+
+    echo "Execution time: ${EXECTIME}s (${RATIO}x realtime)"
+
+    # If the file already exists, delete it
+    if [ -f $1.mp4 ]; then
+        rm $1.mp4
+    fi
+
+    bash $1.$model.wts >/dev/null 2>&1
+    mv $1.mp4 $1.$model.mp4
+
+    ffmpeg -y -f lavfi -i color=c=black:s=1200x50:d=$DURATION -vf "drawtext=fontfile=$2:fontsize=36:x=10:y=(h-text_h)/2:text='ggml-$model - ${EXECTIME}s (${RATIO}x realtime)':fontcolor=lightgrey" $1.$model.info.mp4 >/dev/null 2>&1
+done
+
+COMMAND="ffmpeg -y"
+for model in $models; do
+    COMMAND="$COMMAND -i $1.$model.info.mp4 -i $1.$model.mp4"
+done
+COMMAND="$COMMAND -filter_complex \""
+COUNT=0
+for model in $models; do
+    COMMAND="$COMMAND[${COUNT}:v][$(($COUNT+1)):v]"
+    COUNT=$((COUNT+2))
+done
+COMMAND="$COMMAND vstack=inputs=${COUNT}[v]\" -map \"[v]\" -map 1:a $1.all.mp4 >/dev/null 2>&1"
+
+echo $COMMAND
+
+# Run the command
+eval $COMMAND

extra/quantize-all.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+
+printf "Usage: $0 <upload>"
+
+if [ $# -ne 1 ]; then
+    printf "\nError: Invalid number of arguments\n"
+    exit 1
+fi
+
+qtype0="q5_0"
+qtype1="q5_1"
+upload="$1"
+
+cd `dirname $0`
+cd ../
+
+./quantize ./models/ggml-tiny.en.bin   ./models/ggml-tiny.en-${qtype1}.bin ${qtype1}
+./quantize ./models/ggml-tiny.bin      ./models/ggml-tiny-${qtype1}.bin    ${qtype1}
+
+./quantize ./models/ggml-base.en.bin   ./models/ggml-base.en-${qtype1}.bin ${qtype1}
+./quantize ./models/ggml-base.bin      ./models/ggml-base-${qtype1}.bin    ${qtype1}
+
+./quantize ./models/ggml-small.en.bin  ./models/ggml-small.en-${qtype1}.bin ${qtype1}
+./quantize ./models/ggml-small.bin     ./models/ggml-small-${qtype1}.bin    ${qtype1}
+
+./quantize ./models/ggml-medium.en.bin ./models/ggml-medium.en-${qtype0}.bin ${qtype0}
+./quantize ./models/ggml-medium.bin    ./models/ggml-medium-${qtype0}.bin    ${qtype0}
+
+./quantize ./models/ggml-large.bin     ./models/ggml-large-${qtype0}.bin ${qtype0}
+
+if [ "$upload" == "1" ]; then
+    scp ./models/ggml-tiny.en-${qtype1}.bin   root@linode0:/mnt/Data/ggml/ggml-model-whisper-tiny.en-${qtype1}.bin
+    scp ./models/ggml-tiny-${qtype1}.bin      root@linode0:/mnt/Data/ggml/ggml-model-whisper-tiny-${qtype1}.bin
+
+    scp ./models/ggml-base.en-${qtype1}.bin   root@linode0:/mnt/Data/ggml/ggml-model-whisper-base.en-${qtype1}.bin
+    scp ./models/ggml-base-${qtype1}.bin      root@linode0:/mnt/Data/ggml/ggml-model-whisper-base-${qtype1}.bin
+
+    scp ./models/ggml-small.en-${qtype1}.bin  root@linode0:/mnt/Data/ggml/ggml-model-whisper-small.en-${qtype1}.bin
+    scp ./models/ggml-small-${qtype1}.bin     root@linode0:/mnt/Data/ggml/ggml-model-whisper-small-${qtype1}.bin
+
+    scp ./models/ggml-medium.en-${qtype0}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-medium.en-${qtype0}.bin
+    scp ./models/ggml-medium-${qtype0}.bin    root@linode0:/mnt/Data/ggml/ggml-model-whisper-medium-${qtype0}.bin
+
+    scp ./models/ggml-large-${qtype0}.bin     root@linode0:/mnt/Data/ggml/ggml-model-whisper-large-${qtype0}.bin
+fi

extra/sync-ggml.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+
+cp -rpv ../ggml/src/ggml.c               ./ggml.c
+cp -rpv ../ggml/src/ggml-cuda.h          ./ggml-cuda.h
+cp -rpv ../ggml/src/ggml-cuda.cu         ./ggml-cuda.cu
+cp -rpv ../ggml/src/ggml-opencl.h        ./ggml-opencl.h
+cp -rpv ../ggml/src/ggml-opencl.c        ./ggml-opencl.c
+cp -rpv ../ggml/include/ggml/ggml.h      ./ggml.h
+cp -rpv ../ggml/examples/common.h        ./examples/common.h
+cp -rpv ../ggml/examples/common.cpp      ./examples/common.cpp
+cp -rpv ../ggml/examples/common-ggml.h   ./examples/common-ggml.h
+cp -rpv ../ggml/examples/common-ggml.cpp ./examples/common-ggml.cpp

ggml-cuda.cu

@@ -0,0 +1,716 @@
+#include <cstddef>
+#include <cstdint>
+#include <stdint.h>
+#include <stdio.h>
+#include <atomic>
+
+#include <cuda_runtime.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#include "ggml-cuda.h"
+#include "ggml.h"
+
+static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+#define CUDA_CHECK(err)                                                                 \
+    do {                                                                                \
+        cudaError_t err_ = (err);                                                       \
+        if (err_ != cudaSuccess) {                                                      \
+            fprintf(stderr, "CUDA error %d at %s:%d: %s\n", err_, __FILE__, __LINE__,   \
+                cudaGetErrorString(err_));                                              \
+            exit(1);                                                                    \
+        }                                                                               \
+    } while (0)
+
+#define CUBLAS_CHECK(err)                                                               \
+    do {                                                                                \
+        cublasStatus_t err_ = (err);                                                    \
+        if (err_ != CUBLAS_STATUS_SUCCESS) {                                            \
+            fprintf(stderr, "cuBLAS error %d at %s:%d\n", err_, __FILE__, __LINE__);    \
+            exit(1);                                                                    \
+        }                                                                               \
+    } while (0)
+
+typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
+
+#define QK4_0 32
+typedef struct {
+    float   d;              // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;              // delta
+    float   m;              // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK4_2 16
+typedef struct {
+    half  d;                // delta
+    uint8_t qs[QK4_2 / 2];  // nibbles / quants
+} block_q4_2;
+static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    half d;                 // delta
+    uint8_t qh[4];          // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2];  // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    half d;                 // delta
+    half m;                 // min
+    uint8_t qh[4];          // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2];  // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK8_0];      // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+
+static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_0; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_0 + l + 0] = v0;
+        y[i*QK4_0 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_1; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_1 + l + 0] = v0;
+        y[i*QK4_1 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
+    const block_q4_2 * x = (const block_q4_2 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_2; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = (vi0 - 8)*d;
+        const float v1 = (vi1 - 8)*d;
+
+        y[i*QK4_2 + l + 0] = v0;
+        y[i*QK4_2 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q5_0(const void * vx, float * y) {
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const uint8_t * pp = x[i].qs;
+
+    uint32_t qh;
+    memcpy(&qh, x[i].qh, sizeof(qh));
+
+    for (int l = 0; l < QK5_0; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
+        const int8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+
+        const int8_t vi0 = ((vi & 0xf) | vh0);
+        const int8_t vi1 = ((vi >>  4) | vh1);
+
+        const float v0 = (vi0 - 16)*d;
+        const float v1 = (vi1 - 16)*d;
+
+        y[i*QK5_0 + l + 0] = v0;
+        y[i*QK5_0 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q5_1(const void * vx, float * y) {
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    uint32_t qh;
+    memcpy(&qh, x[i].qh, sizeof(qh));
+
+    for (int l = 0; l < QK5_1; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
+        const int8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+
+        const int8_t vi0 = (vi & 0xf) | vh0;
+        const int8_t vi1 = (vi >>  4) | vh1;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK5_1 + l + 0] = v0;
+        y[i*QK5_1 + l + 1] = v1;
+    }
+}
+
+static __global__ void dequantize_block_q8_0(const void * vx, float * y) {
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+
+    const int8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK8_0; l++) {
+        const int8_t vi = pp[l];
+
+        y[i*QK8_0 + l] = vi*d;
+    }
+}
+
+static void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK4_0;
+    dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK4_1;
+    dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK4_2;
+    dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK5_0;
+    dequantize_block_q5_0<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q5_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK5_1;
+    dequantize_block_q5_1<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+    const int nb = k / QK8_0;
+    dequantize_block_q8_0<<<nb, 1, 0, stream>>>(vx, y);
+}
+
+// TODO: optimize
+static __global__ void convert_fp16_to_fp32(const void * vx, float * y) {
+    const half * x = (const half *) vx;
+
+    const int i = blockIdx.x;
+
+    y[i] = __half2float(x[i]);
+}
+
+static void convert_fp16_to_fp32_cuda(const void * x, float * y, int k, cudaStream_t stream) {
+    convert_fp16_to_fp32<<<k, 1, 0, stream>>>(x, y);
+}
+
+static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q4_2:
+            return dequantize_row_q4_2_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_row_q5_0_cuda;
+        case GGML_TYPE_Q5_1:
+            return dequantize_row_q5_1_cuda;
+        case GGML_TYPE_Q8_0:
+            return dequantize_row_q8_0_cuda;
+        case GGML_TYPE_F16:
+            return convert_fp16_to_fp32_cuda;
+        default:
+            return nullptr;
+    }
+}
+
+// buffer pool for cuda
+#define MAX_CUDA_BUFFERS 16
+
+struct scoped_spin_lock {
+    std::atomic_flag& lock;
+    scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
+        while (lock.test_and_set(std::memory_order_acquire)) {
+            ; // spin
+        }
+    }
+    ~scoped_spin_lock() {
+        lock.clear(std::memory_order_release);
+    }
+    scoped_spin_lock(const scoped_spin_lock&) = delete;
+    scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
+};
+
+struct cuda_buffer {
+    void * ptr = nullptr;
+    size_t size = 0;
+};
+
+static cuda_buffer g_cuda_buffer_pool[MAX_CUDA_BUFFERS];
+static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;
+
+static void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size) {
+    scoped_spin_lock lock(g_cuda_pool_lock);
+
+    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
+        cuda_buffer& b = g_cuda_buffer_pool[i];
+        if (b.size >= size && b.ptr != nullptr) {
+            void * ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+    }
+    void * ptr;
+    CUDA_CHECK(cudaMalloc((void **) &ptr, size));
+    *actual_size = size;
+    return ptr;
+}
+
+static void ggml_cuda_pool_free(void * ptr, size_t size) {
+    scoped_spin_lock lock(g_cuda_pool_lock);
+
+    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
+        cuda_buffer& b = g_cuda_buffer_pool[i];
+        if (b.ptr == nullptr) {
+            b.ptr = ptr;
+            b.size = size;
+            return;
+        }
+    }
+    fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
+    CUDA_CHECK(cudaFree(ptr));
+}
+
+#define GGML_CUDA_MAX_STREAMS 8
+#define GGML_CUDA_MAX_EVENTS 64
+static cublasHandle_t g_cublasH = nullptr;
+static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_STREAMS] = { nullptr };
+static cudaStream_t g_cudaStreams2[GGML_CUDA_MAX_STREAMS] = { nullptr };
+static cudaEvent_t g_cudaEvents[GGML_CUDA_MAX_EVENTS] = { nullptr };
+
+void ggml_init_cublas() {
+    if (g_cublasH == nullptr) {
+        // create streams
+        for (int i = 0; i < GGML_CUDA_MAX_STREAMS; ++i) {
+            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[i], cudaStreamNonBlocking));
+            CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams2[i], cudaStreamNonBlocking));
+        }
+        // create events
+        for (int i = 0; i < GGML_CUDA_MAX_EVENTS; ++i) {
+            CUDA_CHECK(cudaEventCreateWithFlags(&g_cudaEvents[i], cudaEventDisableTiming));
+        }
+
+        // create cublas handle
+        CUBLAS_CHECK(cublasCreate(&g_cublasH));
+        CUBLAS_CHECK(cublasSetMathMode(g_cublasH, CUBLAS_TF32_TENSOR_OP_MATH));
+
+        // configure logging to stdout
+        // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
+    }
+}
+
+void * ggml_cuda_host_malloc(size_t size) {
+    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * ptr = nullptr;
+    cudaError_t err = cudaMallocHost((void **) &ptr, size);
+    if (err != cudaSuccess) {
+        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
+            size/1024.0/1024.0, cudaGetErrorString(err));
+        return nullptr;
+    }
+
+    return ptr;
+}
+
+void ggml_cuda_host_free(void * ptr) {
+    CUDA_CHECK(cudaFreeHost(ptr));
+}
+
+static cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cudaStream_t stream) {
+    const uint64_t ne0 = src->ne[0];
+    const uint64_t ne1 = src->ne[1];
+    const uint64_t nb0 = src->nb[0];
+    const uint64_t nb1 = src->nb[1];
+    const uint64_t nb2 = src->nb[2];
+    const uint64_t nb3 = src->nb[3];
+    const enum ggml_type type = src->type;
+    const size_t ts = ggml_type_size(type);
+    const size_t bs = ggml_blck_size(type);
+
+    const void * x = (const void *) ((const char *) src->data + i2*nb2 + i3*nb3);
+    if (nb0 == ts && nb1 == ts*ne0/bs) {
+        return cudaMemcpyAsync(dst, x, ne1*nb1, cudaMemcpyHostToDevice, stream);
+    } else if (nb0 == ts) {
+        return cudaMemcpy2DAsync(dst, ts*ne0/bs, x, nb1, ts*ne0/bs, ne1, cudaMemcpyHostToDevice, stream);
+    } else {
+        for (uint64_t i1 = 0; i1 < ne1; i1++) {
+            const void * rx = (const void *) ((const char *) x + i1*nb1);
+            void * rd = (void *) ((char *) dst + i1*ts*ne0/bs);
+            // pretend the row is a matrix with cols=1
+            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyHostToDevice, stream);
+            if (r != cudaSuccess) return r;
+        }
+        return cudaSuccess;
+    }
+}
+
+static void ggml_cuda_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+    const int x_ne = ne01 * ne00;
+    const int y_ne = ne11 * ne10;
+    const int d_ne = ne11 * ne01;
+    const int n_mm = ne03 * ne02;
+
+    size_t x_size, y_size, d_size;
+    float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
+    float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size);
+    float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            int i = i03*ne02 + i02;
+            cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
+
+            float * c_X = d_X + i * x_ne;
+            float * c_Y = d_Y + i * y_ne;
+            float * c_D = d_D + i * d_ne;
+
+            // copy data to device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream));
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
+
+            // compute
+            CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
+            CUBLAS_CHECK(
+                cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
+                        ne01, ne11, ne10,
+                        &alpha, c_X, ne00,
+                                c_Y, ne10,
+                        &beta,  c_D, ne01));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
+        }
+    }
+
+    CUDA_CHECK(cudaDeviceSynchronize());
+    ggml_cuda_pool_free(d_X, x_size);
+    ggml_cuda_pool_free(d_Y, y_size);
+    ggml_cuda_pool_free(d_D, d_size);
+}
+
+static void ggml_cuda_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t /* wsize */) {
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+
+    const int nb10 = src1->nb[0];
+    const int nb11 = src1->nb[1];
+    const int nb12 = src1->nb[2];
+    const int nb13 = src1->nb[3];
+
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+    const int x_ne = ne01 * ne00;
+    const int y_ne = ne11 * ne10;
+    const int d_ne = ne11 * ne01;
+    const int n_mm = ne03 * ne02;
+
+    size_t x_size, y_size, d_size;
+    half  * d_X =  (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * x_ne, &x_size);
+    half  * d_Y =  (half *) ggml_cuda_pool_malloc(n_mm * sizeof(half) * y_ne, &y_size);
+    float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
+
+    bool src1_cont_rows = nb10 == sizeof(float);
+    bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            int i = i03*ne02 + i02;
+            cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
+
+            half  * c_X = d_X + i * x_ne;
+            half  * c_Y = d_Y + i * y_ne;
+            float * c_D = d_D + i * d_ne;
+
+            // copy src0 to device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X, src0, i03, i02, cudaStream));
+
+            // convert src1 to fp16
+            // TODO: use multiple threads
+            ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata + (ne11 * ne10) * (i03 * ne02 + i02);
+            char * src1i = (char *) src1->data + i03*nb13 + i02*nb12;
+            if (src1_cont_rows) {
+                if (src1_cont_cols) {
+                    ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
+                }
+                else {
+                    for (int64_t i01 = 0; i01 < ne11; i01++) {
+                        ggml_fp32_to_fp16_row((float *) (src1i + i01*nb11), tmp + i01*ne10, ne10);
+                    }
+                }
+            }
+            else {
+                for (int64_t i01 = 0; i01 < ne11; i01++) {
+                    for (int64_t i00 = 0; i00 < ne10; i00++) {
+                        // very slow due to no inlining
+                        tmp[i01*ne10 + i00] = ggml_fp32_to_fp16(*(float *) (src1i + i01*nb11 + i00*nb10));
+                    }
+                }
+            }
+
+            // copy src1 to device
+            CUDA_CHECK(cudaMemcpyAsync(c_Y, tmp, sizeof(half) * y_ne, cudaMemcpyHostToDevice, cudaStream));
+
+            // compute
+            CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
+            CUBLAS_CHECK(
+                cublasGemmEx(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
+                        ne01, ne11, ne10,
+                        &alpha, c_X, CUDA_R_16F, ne00,
+                                c_Y, CUDA_R_16F, ne10,
+                        &beta,  c_D, CUDA_R_32F, ne01,
+                        CUBLAS_COMPUTE_32F_FAST_16F,
+                        CUBLAS_GEMM_DEFAULT));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
+        }
+    }
+
+    CUDA_CHECK(cudaDeviceSynchronize());
+    ggml_cuda_pool_free(d_X, x_size);
+    ggml_cuda_pool_free(d_Y, y_size);
+    ggml_cuda_pool_free(d_D, d_size);
+}
+
+static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+    const ggml_type type = src0->type;
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+    const int x_ne = ne01 * ne00;
+    const int y_ne = ne11 * ne10;
+    const int d_ne = ne11 * ne01;
+    const int n_mm = ne03 * ne02;
+    const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
+
+    size_t x_size, y_size, d_size, q_size;
+    float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
+    float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size);
+    float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
+    char  * d_Q = (char  *) ggml_cuda_pool_malloc(n_mm * q_sz, &q_size);
+
+    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(type);
+    GGML_ASSERT(to_fp32_cuda != nullptr);
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            int i = i03*ne02 + i02;
+            cudaStream_t cudaStream = g_cudaStreams[i % GGML_CUDA_MAX_STREAMS];
+            cudaStream_t cudaStream2 = g_cudaStreams2[i % GGML_CUDA_MAX_STREAMS];
+            cudaEvent_t  cudaEvent = g_cudaEvents[i % GGML_CUDA_MAX_EVENTS];
+
+            float * c_X = d_X + i * x_ne;
+            float * c_Y = d_Y + i * y_ne;
+            float * c_D = d_D + i * d_ne;
+            char  * c_Q = d_Q + i * q_sz;
+
+            // copy src0 and convert to fp32 on device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2));
+            to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2);
+            CUDA_CHECK(cudaGetLastError());
+            CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
+
+            // copy src1 to device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
+
+            // wait for conversion
+            CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
+
+            // compute
+            CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
+            CUBLAS_CHECK(
+                cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
+                        ne01, ne11, ne10,
+                        &alpha, c_X, ne00,
+                                c_Y, ne10,
+                        &beta,  c_D, ne01));
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CUDA_CHECK(cudaMemcpyAsync(d, c_D, sizeof(float) * d_ne, cudaMemcpyDeviceToHost, cudaStream));
+        }
+    }
+
+    CUDA_CHECK(cudaDeviceSynchronize());
+    ggml_cuda_pool_free(d_X, x_size);
+    ggml_cuda_pool_free(d_Y, y_size);
+    ggml_cuda_pool_free(d_D, d_size);
+    ggml_cuda_pool_free(d_Q, q_size);
+}
+
+bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
+        src1->type == GGML_TYPE_F32 &&
+        dst->type == GGML_TYPE_F32 &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+
+        return true;
+    }
+
+    return false;
+}
+
+bool ggml_cuda_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
+    size_t src0_sz = ggml_nbytes(src0);
+    size_t src1_sz = ggml_nbytes(src1);
+
+    // mul_mat_q: src0 is converted to fp32 on device
+    size_t mul_mat_q_transfer = src0_sz + src1_sz;
+
+    // mul_mat_f16: src1 is converted to fp16 on cpu
+    size_t mul_mat_f16_transfer = src0_sz + sizeof(half) * ggml_nelements(src1);
+
+    // choose the smaller one to transfer to the device
+    // TODO: this is not always the best choice due to the overhead of converting to fp16
+    return mul_mat_f16_transfer < mul_mat_q_transfer;
+}
+
+void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
+    GGML_ASSERT(ggml_cuda_can_mul_mat(src0, src1, dst));
+
+    if (src0->type == GGML_TYPE_F32) {
+        ggml_cuda_mul_mat_f32(src0, src1, dst);
+    }
+    else if (src0->type == GGML_TYPE_F16) {
+        if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) {
+            ggml_cuda_mul_mat_f16(src0, src1, dst, wdata, wsize);
+        }
+        else {
+            ggml_cuda_mul_mat_q_f32(src0, src1, dst);
+        }
+    }
+    else if (ggml_is_quantized(src0->type)) {
+        ggml_cuda_mul_mat_q_f32(src0, src1, dst);
+    }
+    else {
+        GGML_ASSERT(false);
+    }
+}
+
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    if (ggml_cuda_mul_mat_use_f16(src0, src1, dst)) {
+        return ggml_nelements(src1) * sizeof(ggml_fp16_t);
+    }
+    else {
+        return 0;
+    }
+}

ggml-cuda.h

@@ -0,0 +1,19 @@
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void   ggml_init_cublas(void);
+
+bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
+void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
+
+// TODO: export these with GGML_API
+void * ggml_cuda_host_malloc(size_t size);
+void   ggml_cuda_host_free(void * ptr);
+
+#ifdef  __cplusplus
+}
+#endif

ggml-opencl.c

@@ -0,0 +1,398 @@
+#include "ggml-opencl.h"
+
+#define CL_TARGET_OPENCL_VERSION 110
+#include <clblast_c.h>
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+
+#include "ggml.h"
+
+#define MULTILINE_QUOTE(...) #__VA_ARGS__
+const char * clblast_dequant = MULTILINE_QUOTE(
+
+struct block_q4_0
+{
+    float d;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+struct block_q4_1
+{
+    float d;
+    float m;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+    const float m = blocks[i].m;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = (vi & 0xf) * d + m;
+    result[index + 1] = (vi >> 4) * d + m;
+}
+
+struct block_q4_2
+{
+    ushort d;
+    uchar qs[8];
+};
+
+__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 16;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*16 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+
+struct block_q5_0
+{
+    float d;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_0(__global struct block_q5_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = (((vi & 0xf) | vh0) - 16)*d;
+    result[index + 1] = (((vi >>  4) | vh1) - 16)*d;
+}
+
+struct block_q5_1
+{
+    ushort d;
+    ushort m;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_1(__global struct block_q5_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+    const float m = vload_half(0, (__global half*) &blocks[i].m);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = ((vi & 0xf) | vh0)*d + m;
+    result[index + 1] = ((vi >>  4) | vh1)*d + m;
+}
+
+struct block_q8_0
+{
+    float d;
+    char qs[32];
+};
+
+__kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    result[i*32 + l] = blocks[i].qs[l] * blocks[i].d;
+}
+
+);
+
+#define CL_CHECK(err, name)                                                                     \
+    do {                                                                                        \
+        cl_int err_ = (err);                                                                    \
+        if (err_ != CL_SUCCESS) {                                                               \
+            fprintf(stderr, "OpenCL %s error %d at %s:%d\n", name, err_, __FILE__, __LINE__);   \
+            exit(1);                                                                            \
+        }                                                                                       \
+    } while (0)
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+
+typedef struct {
+    float d;                // delta
+    uint32_t qh;          // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2];  // nibbles / quants
+} cl_block_q5_0;
+
+static cl_platform_id platform;
+static cl_device_id device;
+static cl_context context;
+static cl_command_queue queue;
+static cl_program program;
+static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q5_0, kernel_q5_1, kernel_q8_0;
+static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
+static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
+
+static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
+    cl_program p;
+    char *program_log;
+    size_t program_size, log_size;
+    int err;
+
+    program_size = strlen(program_buffer);
+
+    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
+    if(err < 0) {
+        fprintf(stderr, "OpenCL error creating program");
+        exit(1);
+    }
+
+    err = clBuildProgram(p, 0, NULL, NULL, NULL, NULL);
+    if(err < 0) {
+
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
+        program_log = (char*) malloc(log_size + 1);
+        program_log[log_size] = '\0';
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
+        printf("%s\n", program_log);
+        free(program_log);
+        exit(1);
+    }
+
+    return p;
+}
+
+void ggml_cl_init(void) {
+    cl_int err = 0;
+    char * GGML_CLBLAST_PLATFORM = getenv("GGML_CLBLAST_PLATFORM");
+    char * GGML_CLBLAST_DEVICE = getenv("GGML_CLBLAST_DEVICE");
+    int plat_num = (GGML_CLBLAST_PLATFORM == NULL ? 0 : atoi(GGML_CLBLAST_PLATFORM));
+    int dev_num = (GGML_CLBLAST_DEVICE == NULL ? 0 : atoi(GGML_CLBLAST_DEVICE));
+    printf("\nInitializing CLBlast (First Run)...");
+    printf("\nAttempting to use: Platform=%d, Device=%d (If invalid, program will crash)\n",plat_num,dev_num);
+    cl_uint num_platforms;
+    clGetPlatformIDs(0, NULL, &num_platforms);
+    cl_platform_id* platforms = (cl_platform_id*)malloc(num_platforms*sizeof(cl_platform_id));
+    clGetPlatformIDs(num_platforms, platforms, NULL);
+    platform = platforms[plat_num];
+    char platform_buffer[1024];
+    clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(platform_buffer), &platform_buffer, NULL);
+    cl_uint num_devices;
+    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, 0, NULL, &num_devices);
+    cl_device_id* devices = (cl_device_id*)malloc(num_devices*sizeof(cl_device_id));
+    clGetDeviceIDs(platform, CL_DEVICE_TYPE_ALL, num_devices, devices, NULL);
+    device = devices[dev_num];
+    char device_buffer[1024];
+    clGetDeviceInfo(device, CL_DEVICE_NAME, sizeof(device_buffer), &device_buffer, NULL);
+    printf("Using Platform: %s Device: %s\n", platform_buffer, device_buffer);
+    context = clCreateContext(NULL, 1, &device, NULL, NULL, &err);
+    CL_CHECK(err, "clCreateContext");
+    queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err);
+    CL_CHECK(err, "clCreateCommandQueue");
+
+    free(platforms);
+    free(devices);
+
+    program = build_program_from_source(context, device, clblast_dequant);
+
+    // Prepare dequantize kernels
+    kernel_q4_0 = clCreateKernel(program, "dequantize_row_q4_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q4_1 = clCreateKernel(program, "dequantize_row_q4_1", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q4_2 = clCreateKernel(program, "dequantize_row_q4_2", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+}
+
+static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
+    if (req_size <= *cur_size) {
+        return;
+    }
+
+    // Reallocate buffer with enough space
+    if (*cur_size > 0) {
+        clReleaseMemObject(*buf);
+    }
+    cl_int err;
+    *buf = clCreateBuffer(context, flags, req_size, NULL, &err);
+    *cur_size = req_size;
+    CL_CHECK(err, "clCreateBuffer");
+}
+
+void ggml_cl_sgemm_wrapper(
+        const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b,
+        const int m, const int n, const int k,
+        const float alpha, const void *host_a, const int lda,
+        const float *host_b, const int ldb, const float beta,
+        float *host_c, const int ldc, const int btype) {
+    cl_int err = 0;
+
+    cl_kernel kernel;
+    size_t global = n * k, local, size_qb;
+    bool dequant;
+    cl_block_q5_0* cl_host_b;
+
+    switch (btype) {
+    case GGML_TYPE_F32:
+        dequant = false;
+        break;
+    case GGML_TYPE_Q4_0:
+        dequant = true;
+        kernel = kernel_q4_0;
+        local = 16;
+        size_qb = global * (sizeof(float) + local) / 32;
+        break;
+    case GGML_TYPE_Q4_1:
+        dequant = true;
+        kernel = kernel_q4_1;
+        local = 16;
+        size_qb = global * (sizeof(float) * 2 + local) / 32;
+        break;
+    case GGML_TYPE_Q4_2:
+        dequant = true;
+        kernel = kernel_q4_2;
+        local = 8;
+        size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
+        break;
+    case GGML_TYPE_Q5_0:
+        dequant = true;
+        kernel = kernel_q5_0;
+        local = 16;
+        // For some reason OpenCL seems to be incapable of working with structs of size 22.
+        // 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU...
+        // TODO Find the reason, fix and remove workaround.
+        const block_q5_0* b = (const block_q5_0*) host_b;
+        cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32);
+        for (size_t i = 0; i < global / 32; i++) {
+            cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d);
+            memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t));
+            memcpy(&cl_host_b[i].qs, b[i].qs, QK5_0 / 2);
+        }
+        host_b = (const float*) cl_host_b;
+        size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q5_1:
+        dequant = true;
+        kernel = kernel_q5_1;
+        local = 16;
+        size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q8_0:
+        dequant = true;
+        kernel = kernel_q8_0;
+        local = 32;
+        size_qb = global * (sizeof(float) + local) / 32;
+        break;
+    default:
+        fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
+        abort();
+    }
+
+    const size_t size_a =  m * k * sizeof(float);
+    const size_t size_b =  n * k * sizeof(float);
+    const size_t size_c =  m * n * sizeof(float);
+
+    // Prepare buffers
+    ggml_cl_malloc(size_a, &cl_size_a, CL_MEM_READ_ONLY, &cl_buffer_a);
+    if (dequant) {
+        ggml_cl_malloc(size_qb, &cl_size_qb, CL_MEM_READ_ONLY, &cl_buffer_qb);
+    }
+    ggml_cl_malloc(size_b, &cl_size_b, CL_MEM_READ_WRITE, &cl_buffer_b);
+    ggml_cl_malloc(size_c, &cl_size_c, CL_MEM_WRITE_ONLY, &cl_buffer_c);
+
+    cl_event ev_a, ev_qb, ev_b;
+
+    if (dequant) {
+        err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
+        err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
+        CL_CHECK(err, "clSetKernelArg");
+        err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        CL_CHECK(err, "clEnqueueWriteBuffer qb");
+    } else {
+        err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        CL_CHECK(err, "clEnqueueWriteBuffer b");
+    }
+
+    err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
+    CL_CHECK(err, "clEnqueueWriteBuffer a");
+    if (dequant) {
+        err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b);
+        CL_CHECK(err, "clEnqueueNDRangeKernel");
+        clReleaseEvent(ev_qb);
+    }
+    clWaitForEvents(1, &ev_a);
+    clWaitForEvents(1, &ev_b);
+    clReleaseEvent(ev_a);
+    clReleaseEvent(ev_b);
+
+    cl_event ev_sgemm;
+    CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order,
+                                            (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
+                                            m, n, k,
+                                            alpha,
+                                            cl_buffer_a, 0, lda,
+                                            cl_buffer_b, 0, ldb,
+                                            beta,
+                                            cl_buffer_c, 0, ldc,
+                                            &queue, &ev_sgemm);
+
+    if (status != CLBlastSuccess) {
+        fprintf(stderr, "Error: CLBlast SGEMM %d\n", status);
+        abort();
+    }
+
+    cl_event ev_c;
+    clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c);
+
+    // Wait for completion
+    clWaitForEvents(1, &ev_c);
+    clReleaseEvent(ev_sgemm);
+    clReleaseEvent(ev_c);
+    if (btype == GGML_TYPE_Q5_0) {
+        free((void*) cl_host_b);
+    }
+}

ggml-opencl.h

@@ -0,0 +1,24 @@
+#pragma once
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+void ggml_cl_init(void);
+
+enum ggml_blas_order {
+    GGML_BLAS_ORDER_ROW_MAJOR = 101,
+    GGML_BLAS_ORDER_COLUMN_MAJOR = 102,
+};
+
+enum ggml_blas_op {
+    GGML_BLAS_OP_N = 111,
+    GGML_BLAS_OP_T = 112,
+    GGML_BLAS_OP_C = 113,
+};
+
+void ggml_cl_sgemm_wrapper(const enum ggml_blas_order order, const enum ggml_blas_op trans_a, const enum ggml_blas_op trans_b, const int m, const int n, const int k, const float alpha, const void *host_a, const int lda, const float *host_b, const int ldb, const float beta, float *host_c, const int ldc, const int btype);
+
+#ifdef  __cplusplus
+}
+#endif

models/README.md

@@ -6,7 +6,7 @@ using the [convert-pt-to-ggml.py](convert-pt-to-ggml.py) script. You can either
 the `ggml` files yourself using the conversion script, or you can use the [download-ggml-model.sh](download-ggml-model.sh)
 script to download the already converted models. Currently, they are hosted on the following locations:
 
-- https://huggingface.co/datasets/ggerganov/whisper.cpp
+- https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com
 
 Sample usage:
@@ -23,7 +23,7 @@ You can now use it like this:
 
 A third option to obtain the model files is to download them from Hugging Face:
 
-https://huggingface.co/datasets/ggerganov/whisper.cpp/tree/main
+https://huggingface.co/ggerganov/whisper.cpp/tree/main
 
 ## Available models
 

models/convert-h5-to-ggml.py

@@ -23,6 +23,7 @@ import json
 import code
 import torch
 import numpy as np
+from pathlib import Path
 
 from transformers import WhisperForConditionalGeneration
 
@@ -75,16 +76,13 @@ if len(sys.argv) < 4:
     print("Usage: convert-h5-to-ggml.py dir_model path-to-whisper-repo dir-output [use-f32]\n")
     sys.exit(1)
 
-dir_model   = sys.argv[1]
-dir_whisper = sys.argv[2]
-dir_out     = sys.argv[3]
+dir_model   = Path(sys.argv[1])
+dir_whisper = Path(sys.argv[2])
+dir_out     = Path(sys.argv[3])
 
-with open(dir_model + "/vocab.json", "r") as f:
-    encoder = json.load(f)
-with open(dir_model + "/added_tokens.json", "r") as f:
-    encoder_added = json.load(f)
-with open(dir_model + "/config.json", "r") as f:
-    hparams = json.load(f)
+encoder = json.load((dir_model / "vocab.json").open("r", encoding="utf8"))
+encoder_added = json.load((dir_model / "added_tokens.json").open( "r", encoding="utf8"))
+hparams = json.load((dir_model / "config.json").open("r", encoding="utf8") )
 
 model = WhisperForConditionalGeneration.from_pretrained(dir_model)
 
@@ -96,16 +94,15 @@ with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as
 
 dir_tokenizer = dir_model
 
-fname_out = dir_out + "/ggml-model.bin"
+fname_out = dir_out / "ggml-model.bin"
 
-with open(dir_tokenizer + "/vocab.json", "r", encoding="utf8") as f:
-    tokens = json.load(f)
+tokens = json.load(open(dir_tokenizer / "vocab.json", "r", encoding="utf8"))
 
 # use 16-bit or 32-bit floats
 use_f16 = True
 if len(sys.argv) > 4:
     use_f16 = False
-    fname_out = dir_out + "/ggml-model-f32.bin"
+    fname_out = dir_out / "ggml-model-f32.bin"
 
 fout = open(fname_out, "wb")
 
@@ -171,10 +168,9 @@ for name in list_vars.keys():
     data = data.astype(np.float16)
 
     # reshape conv bias from [n] to [n, 1]
-    if name == "encoder.conv1.bias" or \
-       name == "encoder.conv2.bias":
+    if name in ["encoder.conv1.bias", "encoder.conv2.bias"]:
         data = data.reshape(data.shape[0], 1)
-        print("  Reshaped variable: " + name + " to shape: ", data.shape)
+        print("  Reshaped variable: " , name , " to shape: ", data.shape)
 
     n_dims = len(data.shape)
     print(name, n_dims, data.shape)
@@ -182,7 +178,7 @@ for name in list_vars.keys():
     # looks like the whisper models are in f16 by default
     # so we need to convert the small tensors to f32 until we fully support f16 in ggml
     # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype = 1;
+    ftype = 1
     if use_f16:
         if n_dims < 2 or \
                 name == "encoder.conv1.bias"   or \
@@ -197,16 +193,16 @@ for name in list_vars.keys():
         ftype = 0
 
     # header
-    str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    str_ = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str_), ftype))
     for i in range(n_dims):
         fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str);
+    fout.write(str_)
 
     # data
     data.tofile(fout)
 
 fout.close()
 
-print("Done. Output file: " + fname_out)
+print("Done. Output file: " , fname_out)
 print("")

models/convert-pt-to-ggml.py

@@ -39,7 +39,8 @@ import json
 import code
 import torch
 import numpy as np
-
+import base64
+from pathlib import Path
 #from transformers import GPTJForCausalLM
 #from transformers import GPT2TokenizerFast
 
@@ -193,17 +194,17 @@ if len(sys.argv) < 4:
     print("Usage: convert-pt-to-ggml.py model.pt path-to-whisper-repo dir-output [use-f32]\n")
     sys.exit(1)
 
-fname_inp   = sys.argv[1]
-dir_whisper = sys.argv[2]
-dir_out     = sys.argv[3]
+fname_inp   = Path(sys.argv[1])
+dir_whisper = Path(sys.argv[2])
+dir_out     = Path(sys.argv[3])
 
 # try to load PyTorch binary data
 try:
     model_bytes = open(fname_inp, "rb").read()
     with io.BytesIO(model_bytes) as fp:
         checkpoint = torch.load(fp, map_location="cpu")
-except:
-    print("Error: failed to load PyTorch model file: %s" % fname_inp)
+except Exception:
+    print("Error: failed to load PyTorch model file:" , fname_inp)
     sys.exit(1)
 
 hparams = checkpoint["dims"]
@@ -217,33 +218,29 @@ list_vars = checkpoint["model_state_dict"]
 
 # load mel filters
 n_mels = hparams["n_mels"]
-with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as f:
+with np.load(dir_whisper / "whisper" / "assets" / "mel_filters.npz") as f:
     filters = torch.from_numpy(f[f"mel_{n_mels}"])
     #print (filters)
 
 #code.interact(local=locals())
 
 multilingual = hparams["n_vocab"] == 51865
-dir_tokenizer = os.path.join(dir_whisper, "whisper/assets", multilingual and "multilingual" or "gpt2")
-
-#tokenizer = build_tokenizer(dir_whisper, multilingual and "multilingual" or "gpt2")
-#print(tokenizer)
-#print(tokenizer.name_or_path)
-#print(len(tokenizer.additional_special_tokens))
+tokenizer = dir_whisper / "whisper" / "assets" / (multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
 
 # output in the same directory as the model
-fname_out = dir_out + "/ggml-model.bin"
+fname_out = dir_out / "ggml-model.bin"
 
-with open(dir_tokenizer + "/vocab.json", "r", encoding="utf8") as f:
-    tokens = json.load(f)
+with open(tokenizer, "rb") as f:
+    contents = f.read()
+    tokens = {base64.b64decode(token): int(rank) for token, rank in (line.split() for line in contents.splitlines() if line)}
 
 # use 16-bit or 32-bit floats
 use_f16 = True
 if len(sys.argv) > 4:
     use_f16 = False
-    fname_out = dir_out + "/ggml-model-f32.bin"
+    fname_out = dir_out / "ggml-model-f32.bin"
 
-fout = open(fname_out, "wb")
+fout = fname_out.open("wb")
 
 fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
 fout.write(struct.pack("i", hparams["n_vocab"]))
@@ -271,26 +268,24 @@ byte_decoder = {v:k for k, v in byte_encoder.items()}
 fout.write(struct.pack("i", len(tokens)))
 
 for key in tokens:
-    text = bytearray([byte_decoder[c] for c in key])
-    fout.write(struct.pack("i", len(text)))
-    fout.write(text)
+    fout.write(struct.pack("i", len(key)))
+    fout.write(key)
 
 for name in list_vars.keys():
     data = list_vars[name].squeeze().numpy()
-    print("Processing variable: " + name + " with shape: ", data.shape)
+    print("Processing variable: " , name ,  " with shape: ", data.shape)
 
     # reshape conv bias from [n] to [n, 1]
-    if name == "encoder.conv1.bias" or \
-       name == "encoder.conv2.bias":
+    if name in ["encoder.conv1.bias", "encoder.conv2.bias"]:
         data = data.reshape(data.shape[0], 1)
-        print("  Reshaped variable: " + name + " to shape: ", data.shape)
+        print(f"  Reshaped variable: {name} to shape: ", data.shape)
 
-    n_dims = len(data.shape);
+    n_dims = len(data.shape)
 
     # looks like the whisper models are in f16 by default
     # so we need to convert the small tensors to f32 until we fully support f16 in ggml
     # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype = 1;
+    ftype = 1
     if use_f16:
         if n_dims < 2 or \
                 name == "encoder.conv1.bias"   or \
@@ -311,16 +306,16 @@ for name in list_vars.keys():
     #        data = data.transpose()
 
     # header
-    str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    str_ = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str_), ftype))
     for i in range(n_dims):
         fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
-    fout.write(str);
+    fout.write(str_)
 
     # data
     data.tofile(fout)
 
 fout.close()
 
-print("Done. Output file: " + fname_out)
+print("Done. Output file: " , fname_out)
 print("")

models/convert-whisper-to-coreml.py

@@ -0,0 +1,331 @@
+import argparse
+import torch
+import torch.nn.functional as F
+import coremltools as ct
+
+from torch import Tensor
+from torch import nn
+from typing import Dict
+from typing import Optional
+from ane_transformers.reference.layer_norm import LayerNormANE as LayerNormANEBase
+from coremltools.models.neural_network.quantization_utils import quantize_weights
+from whisper.model import Whisper, AudioEncoder, TextDecoder, ResidualAttentionBlock, MultiHeadAttention, ModelDimensions
+from whisper import load_model
+
+# Use for changing dim of input in encoder and decoder embeddings
+def linear_to_conv2d_map(state_dict, prefix, local_metadata, strict,
+                         missing_keys, unexpected_keys, error_msgs):
+    """
+    Unsqueeze twice to map nn.Linear weights to nn.Conv2d weights
+    """
+    for k in state_dict:
+        is_attention = all(substr in k for substr in ['attn', '.weight'])
+        is_mlp = any(k.endswith(s) for s in ['mlp.0.weight', 'mlp.2.weight'])
+
+        if (is_attention or is_mlp) and len(state_dict[k].shape) == 2:
+            state_dict[k] = state_dict[k][:, :, None, None]
+
+
+def correct_for_bias_scale_order_inversion(state_dict, prefix, local_metadata,
+                                           strict, missing_keys,
+                                           unexpected_keys, error_msgs):
+    state_dict[prefix + 'bias'] = state_dict[prefix + 'bias'] / state_dict[prefix + 'weight']
+    return state_dict
+
+class LayerNormANE(LayerNormANEBase):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._register_load_state_dict_pre_hook(
+            correct_for_bias_scale_order_inversion)
+
+class MultiHeadAttentionANE(MultiHeadAttention):
+    def __init__(self, n_state: int, n_head: int):
+        super().__init__(n_state, n_head)
+        self.query =  nn.Conv2d(n_state, n_state, kernel_size=1)
+        self.key = nn.Conv2d(n_state, n_state, kernel_size=1, bias=False)
+        self.value = nn.Conv2d(n_state, n_state, kernel_size=1)
+        self.out = nn.Conv2d(n_state, n_state, kernel_size=1)
+
+    def forward(self,
+                x: Tensor,
+                xa: Optional[Tensor] = None,
+                mask: Optional[Tensor] = None,
+                kv_cache: Optional[dict] = None):
+
+        q = self.query(x)
+
+        if kv_cache is None or xa is None or self.key not in kv_cache:
+            # hooks, if installed (i.e. kv_cache is not None), will prepend the cached kv tensors;
+            # otherwise, perform key/value projections for self- or cross-attention as usual.
+            k = self.key(x if xa is None else xa)
+            v = self.value(x if xa is None else xa)
+
+        else:
+            # for cross-attention, calculate keys and values once and reuse in subsequent calls.
+            k = kv_cache[self.key]
+            v = kv_cache[self.value]
+
+        wv, qk = self.qkv_attention_ane(q, k, v, mask)
+
+        return self.out(wv), qk
+
+    def qkv_attention_ane(self, q: Tensor, k: Tensor, v: Tensor, mask: Optional[Tensor] = None):
+
+        _, dim, _, seqlen = q.size()
+
+        dim_per_head = dim // self.n_head
+
+        scale = float(dim_per_head)**-0.5
+
+        q = q * scale
+
+        mh_q = q.split(dim_per_head, dim=1)
+        mh_k = k.transpose(1,3).split(dim_per_head, dim=3)
+        mh_v = v.split(dim_per_head, dim=1)
+
+        mh_qk = [
+            torch.einsum('bchq,bkhc->bkhq', [qi, ki])
+            for qi, ki in zip(mh_q, mh_k)
+        ]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
+
+        if mask is not None:
+            for head_idx in range(self.n_head):
+                mh_qk[head_idx] = mh_qk[head_idx] + mask[:, :seqlen, :, :seqlen]
+
+        attn_weights = [aw.softmax(dim=1) for aw in mh_qk]  # (batch_size, max_seq_length, 1, max_seq_length) * n_heads
+        attn = [torch.einsum('bkhq,bchk->bchq', wi, vi) for wi, vi in zip(attn_weights, mh_v)]  # (batch_size, dim_per_head, 1, max_seq_length) * n_heads
+        attn = torch.cat(attn, dim=1)  # (batch_size, dim, 1, max_seq_length)
+
+        return attn, torch.cat(mh_qk, dim=1).float().detach()
+
+
+class ResidualAttentionBlockANE(ResidualAttentionBlock):
+    def __init__(self, n_state: int, n_head: int, cross_attention: bool = False):
+        super().__init__(n_state, n_head, cross_attention)
+        self.attn =  MultiHeadAttentionANE(n_state, n_head)
+        self.attn_ln = LayerNormANE(n_state)
+        self.cross_attn =  MultiHeadAttentionANE(n_state, n_head) if cross_attention else None
+        self.cross_attn_ln =  LayerNormANE(n_state) if cross_attention else None
+
+        n_mlp = n_state * 4
+        self.mlp =  nn.Sequential(
+            nn.Conv2d(n_state, n_mlp, kernel_size=1),
+            nn.GELU(),
+            nn.Conv2d(n_mlp, n_state, kernel_size=1)
+        )
+        self.mlp_ln = LayerNormANE(n_state)
+
+
+class AudioEncoderANE(AudioEncoder):
+    def __init__(self, n_mels: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
+        super().__init__(n_mels, n_ctx, n_state, n_head, n_layer)
+
+        self.blocks = nn.ModuleList(
+            [ResidualAttentionBlockANE(n_state, n_head) for _ in range(n_layer)]
+        )
+        self.ln_post = LayerNormANE(n_state)
+
+    def forward(self, x: Tensor):
+        """
+        x : torch.Tensor, shape = (batch_size, n_mels, n_ctx)
+            the mel spectrogram of the audio
+        """
+        x = F.gelu(self.conv1(x))
+        x = F.gelu(self.conv2(x))
+
+        assert x.shape[1:] == self.positional_embedding.shape[::-1], "incorrect audio shape"
+
+        # Add positional embedding and add dummy dim for ANE
+        x = (x + self.positional_embedding.transpose(0,1)).to(x.dtype).unsqueeze(2)
+
+        for block in self.blocks:
+            x = block(x)
+
+        x = self.ln_post(x)
+
+        # """
+        # TODO:
+        # I think we need to transpose the result here to make it fit whisper.cpp memory order.
+        # However, even doing this, the results are still wrong. Kind of less wrong compared to
+        # not transposing, but still wrong.
+
+        # Also, I don't know why the original OpenAI implementation does not need to transpose
+
+        # transpose to (batch_size, n_ctx, n_state)
+        # x : torch.Tensor, shape = (batch_size, n_state, 1, n_ctx)
+
+        # """
+        # x = x.transpose(1,3)
+
+        return x
+
+class TextDecoderANE(TextDecoder):
+
+    def __init__(self, n_vocab: int, n_ctx: int, n_state: int, n_head: int, n_layer: int):
+        super().__init__(n_vocab, n_ctx, n_state, n_head, n_layer)
+
+        self.blocks= nn.ModuleList(
+            [ResidualAttentionBlockANE(n_state, n_head, cross_attention=True) for _ in range(n_layer)]
+        )
+        self.ln= LayerNormANE(n_state)
+
+    def forward(self, x: Tensor, xa: Tensor, kv_cache: Optional[dict] = None):
+        """
+        x : torch.LongTensor, shape = (batch_size, <= n_ctx)
+            the text tokens
+        xa : torch.Tensor, shape = (batch_size, n_mels, n_audio_ctx)
+            the encoded audio features to be attended on
+        """
+        offset = next(iter(kv_cache.values())).shape[3] if kv_cache else 0
+        x = self.token_embedding(x) + self.positional_embedding[offset : offset + x.shape[-1]]
+        x = x.to(xa.dtype)
+
+        # Reformat for ANE
+        mask = self.mask[None, None, :, :].permute(0,3,1,2)
+        x = x.transpose(1,2).unsqueeze(2)
+
+        for block in self.blocks:
+            x = block(x, xa, mask=mask, kv_cache=kv_cache)
+
+        x = self.ln(x)
+
+        # Reformat back from ANE
+        x = x.permute(0,2,3,1).squeeze(0)
+
+        # ANE can only load tensors with dim size of at most 16,384 - whisper uses 51,864 (en) or 51,865 (multi-lang) tokens so we need to compute in chunks
+        if self.token_embedding.weight.shape[0] == 51865:
+            # split in 11 chunks - 4715 each
+            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//11, dim=0)
+            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
+        else:
+            # split in 12 chunks - 4322 each
+            assert(self.token_embedding.weight.shape[0] == 51864)
+            splits = self.token_embedding.weight.split(self.token_embedding.weight.shape[0]//12, dim=0)
+            logits = torch.cat([torch.einsum('bid,jd->bij', x, split) for split in splits]).view(*x.shape[:2], -1)
+
+        return logits
+
+class WhisperANE(Whisper):
+    def __init__(self, dims: ModelDimensions):
+        super().__init__(dims)
+
+        self.encoder = AudioEncoderANE(
+            self.dims.n_mels,
+            self.dims.n_audio_ctx,
+            self.dims.n_audio_state,
+            self.dims.n_audio_head,
+            self.dims.n_audio_layer,
+        )
+        self.decoder = TextDecoderANE(
+            self.dims.n_vocab,
+            self.dims.n_text_ctx,
+            self.dims.n_text_state,
+            self.dims.n_text_head,
+            self.dims.n_text_layer,
+        )
+
+        self._register_load_state_dict_pre_hook(linear_to_conv2d_map)
+
+    def forward(self, mel: torch.Tensor, tokens: torch.Tensor) -> Dict[str, torch.Tensor]:
+        return self.decoder(tokens, self.encoder(mel))
+
+    def install_kv_cache_hooks(self, cache: Optional[dict] = None):
+        cache = {**cache} if cache is not None else {}
+        hooks = []
+
+        def save_to_cache(module, _, output):
+            if module not in cache or output.shape[3] > self.decoder.positional_embedding.shape[0]:
+                cache[module] = output  # save as-is, for the first token or cross attention
+            else:
+                cache[module] = torch.cat([cache[module], output], dim=3).detach()
+            return cache[module]
+
+        def install_hooks(layer: nn.Module):
+            if isinstance(layer, MultiHeadAttentionANE):
+                hooks.append(layer.key.register_forward_hook(save_to_cache))
+                hooks.append(layer.value.register_forward_hook(save_to_cache))
+
+        self.decoder.apply(install_hooks)
+        return cache, hooks
+
+def convert_encoder(hparams, model, quantize=False):
+    model.eval()
+
+    input_shape = (1, 80, 3000)
+    input_data = torch.randn(input_shape)
+    traced_model = torch.jit.trace(model, input_data)
+
+    model = ct.convert(
+        traced_model,
+        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
+        inputs=[ct.TensorType(name="logmel_data", shape=input_shape)],
+        outputs=[ct.TensorType(name="output")],
+        compute_units=ct.ComputeUnit.ALL
+    )
+
+    if quantize:
+        model = quantize_weights(model, nbits=16)
+
+    return model
+
+def convert_decoder(hparams, model, quantize=False):
+    model.eval()
+
+    tokens_shape = (1, 1)
+    audio_shape = (1, hparams.n_audio_state, 1, 1500)
+
+    audio_data = torch.randn(audio_shape)
+    token_data = torch.randint(50257, tokens_shape).long()
+    traced_model = torch.jit.trace(model, (token_data, audio_data))
+
+    model = ct.convert(
+        traced_model,
+        convert_to=None if quantize else "mlprogram", # convert will fail if weights are quantized, not sure why
+        inputs=[
+            ct.TensorType(name="token_data", shape=tokens_shape, dtype=int),
+            ct.TensorType(name="audio_data", shape=audio_shape)
+        ]
+    )
+
+    if quantize:
+        model = quantize_weights(model, nbits=16)
+
+    return model
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, help="model to convert (e.g. tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large)", required=True)
+    parser.add_argument("--encoder-only", type=bool, help="only convert encoder", default=False)
+    parser.add_argument("--quantize",     type=bool, help="quantize weights to F16", default=False)
+    parser.add_argument("--optimize-ane", type=bool, help="optimize for ANE execution (currently broken)", default=False)
+    args = parser.parse_args()
+
+    if args.model not in ["tiny", "tiny.en", "base", "base.en", "small", "small.en", "medium", "medium.en", "large"]:
+        raise ValueError("Invalid model name")
+
+    whisper = load_model(args.model).cpu()
+    hparams = whisper.dims
+    print(hparams)
+
+    if args.optimize_ane:
+        whisperANE = WhisperANE(hparams).eval()
+        whisperANE.load_state_dict(whisper.state_dict())
+
+        encoder = whisperANE.encoder
+        decoder = whisperANE.decoder
+    else:
+        encoder = whisper.encoder
+        decoder = whisper.decoder
+
+    # Convert encoder
+    encoder = convert_encoder(hparams, encoder, quantize=args.quantize)
+    encoder.save(f"models/coreml-encoder-{args.model}.mlpackage")
+
+    if args.encoder_only is False:
+        # Convert decoder
+        decoder = convert_decoder(hparams, decoder, quantize=args.quantize)
+        decoder.save(f"models/coreml-decoder-{args.model}.mlpackage")
+
+    print("done converting")

models/download-coreml-model.sh

@@ -0,0 +1,82 @@
+#!/bin/bash
+
+# This script downloads Whisper model files that have already been converted to Core ML format.
+# This way you don't have to convert them yourself.
+
+src="https://huggingface.co/datasets/ggerganov/whisper.cpp-coreml"
+pfx="resolve/main/ggml"
+
+# get the path of this script
+function get_script_path() {
+    if [ -x "$(command -v realpath)" ]; then
+        echo "$(dirname $(realpath $0))"
+    else
+        local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
+        echo "$ret"
+    fi
+}
+
+models_path="$(get_script_path)"
+
+# Whisper models
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download Core ML model
+
+printf "Downloading Core ML model $model from '$src' ...\n"
+
+cd $models_path
+
+if [ -f "ggml-$model.mlmodel" ]; then
+    printf "Model $model already exists. Skipping download.\n"
+    exit 0
+fi
+
+if [ -x "$(command -v wget)" ]; then
+    wget --quiet --show-progress -O ggml-$model.mlmodel $src/$pfx-$model.mlmodel
+elif [ -x "$(command -v curl)" ]; then
+    curl -L --output ggml-$model.mlmodel $src/$pfx-$model.mlmodel
+else
+    printf "Either wget or curl is required to download models.\n"
+    exit 1
+fi
+
+
+if [ $? -ne 0 ]; then
+    printf "Failed to download Core ML model $model \n"
+    printf "Please try again later or download the original Whisper model files and convert them yourself.\n"
+    exit 1
+fi
+
+printf "Done! Model '$model' saved in 'models/ggml-$model.mlmodel'\n"
+printf "Run the following command to compile it:\n\n"
+printf "  $ xcrun coremlc compile ./models/ggml-$model.mlmodel ./models\n\n"
+printf "You can now use it like this:\n\n"
+printf "  $ ./main -m models/ggml-$model.bin -f samples/jfk.wav\n"
+printf "\n"

models/download-ggml-model.cmd

@@ -40,7 +40,7 @@ if exist "ggml-%model%.bin" (
   goto :eof
 )
 
-PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/datasets/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
+PowerShell -NoProfile -ExecutionPolicy Bypass -Command "Invoke-WebRequest -Uri https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-%model%.bin -OutFile ggml-%model%.bin"
 
 if %ERRORLEVEL% neq 0 (
   echo Failed to download ggml model %model%

models/download-ggml-model.sh

@@ -6,13 +6,13 @@
 #src="https://ggml.ggerganov.com"
 #pfx="ggml-model-whisper"
 
-src="https://huggingface.co/datasets/ggerganov/whisper.cpp"
+src="https://huggingface.co/ggerganov/whisper.cpp"
 pfx="resolve/main/ggml"
 
 # get the path of this script
 function get_script_path() {
     if [ -x "$(command -v realpath)" ]; then
-        echo "$(dirname $(realpath $0))"
+        echo "$(dirname "$(realpath "$0")")"
     else
         local ret="$(cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P)"
         echo "$ret"

models/generate-coreml-interface.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+#
+# This generates:
+#   - coreml/whisper-encoder-impl.h and coreml/whisper-encoder-impl.m
+#   - coreml/whisper-decoder-impl.h and coreml/whisper-decoder-impl.m
+#
+
+wd=$(dirname "$0")
+cd "$wd/../"
+
+python3 models/convert-whisper-to-coreml.py --model tiny.en
+
+mv -v models/coreml-encoder-tiny.en.mlpackage models/whisper-encoder-impl.mlpackage
+xcrun coremlc generate models/whisper-encoder-impl.mlpackage coreml/
+mv coreml/whisper_encoder_impl.h coreml/whisper-encoder-impl.h
+mv coreml/whisper_encoder_impl.m coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.m/whisper-encoder-impl.m/g' coreml/whisper-encoder-impl.m
+sed -i '' 's/whisper_encoder_impl\.h/whisper-encoder-impl.h/g' coreml/whisper-encoder-impl.h
+
+mv -v models/coreml-decoder-tiny.en.mlpackage models/whisper-decoder-impl.mlpackage
+xcrun coremlc generate models/whisper-decoder-impl.mlpackage coreml/
+mv coreml/whisper_decoder_impl.h coreml/whisper-decoder-impl.h
+mv coreml/whisper_decoder_impl.m coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.m/whisper-decoder-impl.m/g' coreml/whisper-decoder-impl.m
+sed -i '' 's/whisper_decoder_impl\.h/whisper-decoder-impl.h/g' coreml/whisper-decoder-impl.h
+
+rm -rfv models/whisper-encoder-impl.mlpackage models/whisper-decoder-impl.mlpackage