release : v1.7.4

ggml-ci

.github/workflows/build.yml

@@ -1,5 +1,15 @@
 name: CI
-on: [push, pull_request]
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
 
 env:
   ubuntu_image: "ubuntu:22.04"
@@ -12,7 +22,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        arch: [linux/amd64, linux/ppc64le]
 
     steps:
       - name: Clone
@@ -32,6 +42,58 @@ jobs:
             cmake -B build
             cmake --build build --config Release -j $(nproc)'
 
+  ubuntu-latest-arm64:
+    runs-on: ubuntu-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [linux/arm64]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Build ${{ matrix.arch }}
+        run: |
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential libsdl2-dev cmake
+            cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
+            cmake --build build --config Release -j $(nproc)'
+
+  ubuntu-latest-arm-v7:
+    runs-on: ubuntu-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        arch: [linux/arm/v7]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Build ${{ matrix.arch }}
+        run: |
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential libsdl2-dev cmake
+            cmake -B build -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
+            cmake --build build --config Release -j $(nproc)'
+
   macOS-latest:
     runs-on: macOS-latest
 
@@ -74,7 +136,7 @@ jobs:
       fail-fast: false
       matrix:
         build: [Debug, Release]
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
+        arch: [linux/amd64, linux/ppc64le]
 
     steps:
       - name: Clone
@@ -95,6 +157,62 @@ jobs:
             make
             ctest -L gh --output-on-failure'
 
+  ubuntu-latest-gcc-arm64:
+    runs-on: ubuntu-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        build: [Debug, Release]
+        arch: [linux/arm64]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Build ${{ matrix.arch }}
+        run: |
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv8-a
+            make
+            ctest -L gh --output-on-failure'
+
+  ubuntu-latest-gcc-arm-v7:
+    runs-on: ubuntu-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        build: [Debug, Release]
+        arch: [linux/arm/v7]
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Build ${{ matrix.arch }}
+        run: |
+          docker run --platform ${{ matrix.arch }} --rm \
+            -v ${{ github.workspace }}:/workspace \
+            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
+            set -e
+            apt update
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DGGML_NATIVE=OFF -DGGML_CPU_ARM_ARCH=armv7-a+fp
+            make
+            ctest -L gh --output-on-failure'
+
   ubuntu-latest-clang:
     runs-on: ubuntu-latest
 
@@ -283,7 +401,7 @@ jobs:
       - name: Build using CMake
         shell: msys2 {0}
         run: |
-            cmake -B build
+            cmake -B build -DWHISPER_SDL2=ON
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
       - name: Clean after building using CMake
@@ -430,72 +548,73 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
-# TODO: fix and re-enable
-#  windows-cublas:
-#    runs-on: windows-2019
-#
-#    strategy:
-#      matrix:
-#        build: [Release]
-#        arch: [x64]
-#        cublas: [ON]
-#        sdl2: [ON]
-#        cuda-toolkit: [12.2.0, 11.8.0]
-#        include:
-#          - arch: x64
-#            s2arc: x64
-#          - sdl2: ON
-#            s2ver: 2.28.5
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#
-#      - name: Add msbuild to PATH
-#        uses: microsoft/setup-msbuild@v2
-#
-#      - name: Install CUDA Toolkit
-#        id: cuda-toolkit
-#        uses: Jimver/cuda-toolkit@v0.2.15
-#        with:
-#          cuda: '${{ matrix.cuda-toolkit }}'
-#
-#      - name: Fetch SDL2 and set SDL2_DIR
-#        if: matrix.sdl2 == 'ON'
-#        run: |
-#          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ 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 }}
-#          -DGGML_CUDA=${{ matrix.cublas }}
-#          -DWHISPER_SDL2=${{ matrix.sdl2 }}
-#
-#      - name: Build ${{ matrix.cuda-toolkit }}
-#        run: |
-#          cd ./build
-#          cmake --build . --config ${{ matrix.build }}
-#
-#      - name: Copy CUDA DLLs
-#        run: >
-#          Copy-Item -PassThru
-#          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
-#          -Include cudart64_*,cublas64_*,cublasLt64_*
-#          -Destination build/bin/${{ matrix.build }}
-#
-#      - name: Copy SDL2.dll
-#        if: matrix.sdl2 == 'ON'
-#        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
-#
-#      - name: Upload binaries
-#        if: matrix.sdl2 == 'ON'
-#        uses: actions/upload-artifact@v4
-#        with:
-#          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
-#          path: build/bin/${{ matrix.build }}
+  windows-cublas:
+    runs-on: windows-2019
+    strategy:
+      matrix:
+        build: [Release]
+        arch: [x64]
+        cublas: [ON]
+        sdl2: [ON]
+        cuda-toolkit: [12.2.0, 11.8.0]
+        include:
+          - arch: x64
+            sdl2: ON
+            sdl2_ver: 2.28.5
+    steps:
+      - name: Clone repository
+        uses: actions/checkout@v4
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v2
+
+      - name: Install CUDA Toolkit
+        id: cuda-toolkit
+        uses: Jimver/cuda-toolkit@v0.2.15
+        with:
+          cuda: '${{ matrix.cuda-toolkit }}'
+
+      - name: Install 7-Zip
+        run: choco install 7zip -y
+
+      - name: Fetch SDL2 and set SDL2_DIR
+        if: matrix.sdl2 == 'ON'
+        run: |
+          Invoke-WebRequest -Uri https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.sdl2_ver }}/SDL2-devel-${{ matrix.sdl2_ver }}-VC.zip -OutFile sdl2.zip
+          7z x sdl2.zip
+          echo "SDL2_DIR=${{ github.workspace }}\SDL2-${{ matrix.sdl2_ver }}\cmake" | Out-File -FilePath $env:GITHUB_ENV -Append
+          echo "${{ github.workspace }}\SDL2-${{ matrix.sdl2_ver }}\cmake" > SDL2_PATH.txt
+ 
+      - name: Configure CMake
+        shell: cmd
+        run: |
+          cmake -S . -B ./build -A ${{ matrix.arch }} ^
+            -DCMAKE_BUILD_TYPE=${{ matrix.build }} ^
+            -DGGML_CUDA=${{ matrix.cublas }} ^
+            -DCMAKE_CUDA_ARCHITECTURES=all  ^
+            -DWHISPER_SDL2=${{ matrix.sdl2 }} ^
+            -DSDL2_DIR="%SDL2_DIR%"
+  
+      - name: Build Project
+        shell: cmd
+        run: |
+          cd ./build
+          cmake --build . --config ${{ matrix.build }}          
+
+      - name: Copy CUDA DLLs
+        run: |
+          Get-ChildItem "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/" -Filter "*.dll" | 
+          Copy-Item -Destination "build/bin/${{ matrix.build }}"
+
+      - name: Copy SDL2.dll
+        if: matrix.sdl2 == 'ON'
+        run: copy "$env:SDL2_DIR/../lib/${{ matrix.arch }}/SDL2.dll" build/bin/${{ matrix.build }}
+
+      - name: Upload binaries
+        uses: actions/upload-artifact@v4
+        with:
+          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
+          path: build/bin/${{ matrix.build }}
 
   emscripten:
     runs-on: ubuntu-latest
@@ -564,35 +683,34 @@ jobs:
       - name: Build swiftui example
         run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
-# TODO: update android build and re-enable when it works
-#  android:
-#    runs-on: ubuntu-latest
-#
-#    steps:
-#      - name: Clone
-#        uses: actions/checkout@v4
-#        with:
-#          path: whisper
-#
-#      - name: Install Java
-#        uses: actions/setup-java@v4
-#        with:
-#          distribution: zulu
-#          java-version: 21
-#
-#      - name: Setup Android SDK
-#        uses: android-actions/setup-android@v3
-#
-#      - name: Build
-#        run: |
-#          cd whisper/examples/whisper.android
-#          ./gradlew assembleRelease --no-daemon
-#
-#      - name: Build with external ggml
-#        run: |
-#          export PATH_TO_GGML=$PWD/ggml
-#          cd whisper/examples/whisper.android
-#          ./gradlew assembleRelease --no-daemon
+  android:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with:
+          path: whisper
+
+      - name: Install Java
+        uses: actions/setup-java@v4
+        with:
+          distribution: zulu
+          java-version: 21
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@v3
+
+      - name: Build
+        run: |
+          cd whisper/examples/whisper.android
+          ./gradlew assembleRelease --no-daemon
+
+      - name: Build with external ggml
+        run: |
+          export PATH_TO_GGML=$PWD/ggml
+          cd whisper/examples/whisper.android
+          ./gradlew assembleRelease --no-daemon
 
 # TODO: disable because of following fail: https://github.com/ggerganov/whisper.cpp/actions/runs/11019444420/job/30627193602
 #  android_java:

.github/workflows/docker.yml

@@ -17,7 +17,7 @@ jobs:
     strategy:
       matrix:
         config:
-          - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64,linux/arm64" }
+          - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64" }
           #TODO: the cuda image keeps failing - disable for now
           #      https://github.com/ggerganov/whisper.cpp/actions/runs/11019444428/job/30602020339
           #- { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }

CMakeLists.txt

@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
 project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.7.2)
+project("whisper.cpp" VERSION 1.7.4)
 include(CheckIncludeFileCXX)
 
 set(SOVERSION 1)

Makefile

@@ -64,6 +64,6 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 larg
 		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
 	    echo "----------------------------------------------" ; \
 		echo "" ; \
-		./build/bin/main -m models/ggml-$@.bin -f $$f ; \
+		./build/bin/whisper-cli -m models/ggml-$@.bin -f $$f ; \
 		echo "" ; \
 	done

README.md

@@ -7,7 +7,7 @@
 [![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.7.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.7.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -53,18 +53,6 @@ On Apple Silicon, the inference runs fully on the GPU via Metal:
 
 https://github.com/ggerganov/whisper.cpp/assets/1991296/c82e8f86-60dc-49f2-b048-d2fdbd6b5225
 
-Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
-
-## Implementation details
-
-- The core tensor operations are implemented in C ([ggml.h](ggml/include/ggml.h) / [ggml.c](ggml/src/ggml.c))
-- The transformer model and the high-level C-style API are implemented in C++ ([whisper.h](include/whisper.h) / [whisper.cpp](src/whisper.cpp))
-- Sample usage is demonstrated in [main.cpp](examples/main)
-- Sample real-time audio transcription from the microphone is demonstrated in [stream.cpp](examples/stream)
-- Various other examples are available in the [examples](examples) folder
-
-The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
-
 ## Quick start
 
 First clone the repository:
@@ -85,135 +73,26 @@ Then, download one of the Whisper [models](models/README.md) converted in [`ggml
 sh ./models/download-ggml-model.sh base.en
 ```
 
-Now build the [main](examples/main) example and transcribe an audio file like this:
+Now build the [whisper-cli](examples/cli) example and transcribe an audio file like this:
 
 ```bash
-# build the main example
+# build the project
 cmake -B build
 cmake --build build --config Release
 
 # transcribe an audio file
-./build/bin/main -f samples/jfk.wav
+./build/bin/whisper-cli -f samples/jfk.wav
 ```
 
 ---
 
-For a quick demo, simply run `make base.en`:
-
-```text
-$ make -j base.en
-
-cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
-c++ -I. -I./examples -O3 -std=c++11 -pthread -c whisper.cpp -o whisper.o
-c++ -I. -I./examples -O3 -std=c++11 -pthread examples/main/main.cpp whisper.o ggml.o -o main  -framework Accelerate
-./main -h
-
-usage: ./main [options] file0.wav file1.wav ...
-
-options:
-  -h,        --help              [default] show this help message and exit
-  -t N,      --threads N         [4      ] number of threads to use during computation
-  -p N,      --processors N      [1      ] number of processors to use during computation
-  -ot N,     --offset-t N        [0      ] time offset in milliseconds
-  -on N,     --offset-n N        [0      ] segment index offset
-  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
-  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
-  -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [5      ] beam size for beam search
-  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
-  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
-  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
-  -tr,       --translate         [false  ] translate from source language to english
-  -di,       --diarize           [false  ] stereo audio diarization
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
-  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
-  -otxt,     --output-txt        [false  ] output result in a text file
-  -ovtt,     --output-vtt        [false  ] output result in a vtt file
-  -osrt,     --output-srt        [false  ] output result in a srt file
-  -olrc,     --output-lrc        [false  ] output result in a lrc file
-  -owts,     --output-words      [false  ] output script for generating karaoke video
-  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
-  -ocsv,     --output-csv        [false  ] output result in a CSV file
-  -oj,       --output-json       [false  ] output result in a JSON file
-  -ojf,      --output-json-full  [false  ] include more information in the 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
-  -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
-  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
-             --prompt PROMPT     [       ] initial prompt
-  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
-  -f FNAME,  --file FNAME        [       ] input WAV file path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of tokens
-  -ng,       --no-gpu            [false  ] disable GPU
-
-
-sh ./models/download-ggml-model.sh base.en
-Downloading ggml model base.en ...
-ggml-base.en.bin               100%[========================>] 141.11M  6.34MB/s    in 24s
-Done! Model 'base.en' saved in 'models/ggml-base.en.bin'
-You can now use it like this:
-
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
-
-
-===============================================
-Running base.en on all samples in ./samples ...
-===============================================
-
-----------------------------------------------
-[+] Running base.en on samples/jfk.wav ... (run 'ffplay samples/jfk.wav' to listen)
-----------------------------------------------
-
-whisper_init_from_file: loading model from 'models/ggml-base.en.bin'
-whisper_model_load: loading model
-whisper_model_load: n_vocab       = 51864
-whisper_model_load: n_audio_ctx   = 1500
-whisper_model_load: n_audio_state = 512
-whisper_model_load: n_audio_head  = 8
-whisper_model_load: n_audio_layer = 6
-whisper_model_load: n_text_ctx    = 448
-whisper_model_load: n_text_state  = 512
-whisper_model_load: n_text_head   = 8
-whisper_model_load: n_text_layer  = 6
-whisper_model_load: n_mels        = 80
-whisper_model_load: f16           = 1
-whisper_model_load: type          = 2
-whisper_model_load: mem required  =  215.00 MB (+    6.00 MB per decoder)
-whisper_model_load: kv self size  =    5.25 MB
-whisper_model_load: kv cross size =   17.58 MB
-whisper_model_load: adding 1607 extra tokens
-whisper_model_load: model ctx     =  140.60 MB
-whisper_model_load: model size    =  140.54 MB
-
-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 |
-
-main: processing 'samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
-
-
-[00:00:00.000 --> 00:00:11.000]   And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.
-
-
-whisper_print_timings:     fallbacks =   0 p /   0 h
-whisper_print_timings:     load time =   113.81 ms
-whisper_print_timings:      mel time =    15.40 ms
-whisper_print_timings:   sample time =    11.58 ms /    27 runs (    0.43 ms per run)
-whisper_print_timings:   encode time =   266.60 ms /     1 runs (  266.60 ms per run)
-whisper_print_timings:   decode time =    66.11 ms /    27 runs (    2.45 ms per run)
-whisper_print_timings:    total time =   476.31 ms
-```
+For a quick demo, simply run `make base.en`.
 
 The command downloads the `base.en` model converted to custom `ggml` format and runs the inference on all `.wav` samples in the folder `samples`.
 
-For detailed usage instructions, run: `./main -h`
+For detailed usage instructions, run: `./build/bin/whisper-cli -h`
 
-Note that the [main](examples/main) example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.
+Note that the [whisper-cli](examples/cli) example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.
 For example, you can use `ffmpeg` like this:
 
 ```bash
@@ -271,7 +150,7 @@ cmake --build build --config Release
 ./build/bin/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
-./build/bin/main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
+./build/bin/whisper-cli -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
 ```
 
 ## Core ML support
@@ -313,7 +192,7 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 - Run the examples as usual. For example:
 
   ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  $ ./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -397,7 +276,7 @@ This can result in significant speedup in encoder performance. Here are the inst
 - Run the examples as usual. For example:
 
   ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  $ ./build/bin/whisper-cli -m models/ggml-base.en.bin -f samples/jfk.wav
 
   ...
 
@@ -414,7 +293,7 @@ This can result in significant speedup in encoder performance. Here are the inst
   The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
   cached for the next run.
 
-For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
+For more information about the OpenVINO implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).
 
 ## NVIDIA GPU support
 
@@ -473,7 +352,7 @@ cmake --build build -j --config Release
 Run the inference examples as usual, for example:
 
 ```
-./build/bin/main -f samples/jfk.wav -m models/ggml-base.en.bin -t 8
+./build/bin/whisper-cli -f samples/jfk.wav -m models/ggml-base.en.bin -t 8
 ```
 
 *Notes:*
@@ -481,38 +360,6 @@ Run the inference examples as usual, for example:
 - If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
 - If you run successfully with your Ascend NPU device, please help update the table `Verified devices`.
 
-## Docker
-
-### Prerequisites
-
-- Docker must be installed and running on your system.
-- Create a folder to store big models & intermediate files (ex. /whisper/models)
-
-### Images
-
-We have two Docker images available for this project:
-
-1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
-
-### Usage
-
-```shell
-# download model and persist it in a local folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  whisper.cpp:main "./models/download-ggml-model.sh base /models"
-# transcribe an audio file
-docker run -it --rm \
-  -v path/to/models:/models \
-  -v path/to/audios:/audios \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
-# transcribe an audio file in samples folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
-```
-
 ## Installing with Conan
 
 You can install pre-built binaries for whisper.cpp or build it from source using [Conan](https://conan.io/). Use the following command:
@@ -527,89 +374,6 @@ For detailed instructions on how to use Conan, please refer to the [Conan docume
 
 - Inference only
 
-## Another example
-
-Here is another example of transcribing a [3:24 min speech](https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg)
-in about half a minute on a MacBook M1 Pro, using `medium.en` model:
-
-<details>
-  <summary>Expand to see the result</summary>
-
-```text
-$ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8
-
-whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
-whisper_model_load: loading model
-whisper_model_load: n_vocab       = 51864
-whisper_model_load: n_audio_ctx   = 1500
-whisper_model_load: n_audio_state = 1024
-whisper_model_load: n_audio_head  = 16
-whisper_model_load: n_audio_layer = 24
-whisper_model_load: n_text_ctx    = 448
-whisper_model_load: n_text_state  = 1024
-whisper_model_load: n_text_head   = 16
-whisper_model_load: n_text_layer  = 24
-whisper_model_load: n_mels        = 80
-whisper_model_load: f16           = 1
-whisper_model_load: type          = 4
-whisper_model_load: mem required  = 1720.00 MB (+   43.00 MB per decoder)
-whisper_model_load: kv self size  =   42.00 MB
-whisper_model_load: kv cross size =  140.62 MB
-whisper_model_load: adding 1607 extra tokens
-whisper_model_load: model ctx     = 1462.35 MB
-whisper_model_load: model size    = 1462.12 MB
-
-system_info: n_threads = 8 / 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 |
-
-main: processing 'samples/gb1.wav' (3179750 samples, 198.7 sec), 8 threads, 1 processors, lang = en, task = transcribe, timestamps = 1 ...
-
-
-[00:00:00.000 --> 00:00:08.000]   My fellow Americans, this day has brought terrible news and great sadness to our country.
-[00:00:08.000 --> 00:00:17.000]   At nine o'clock this morning, Mission Control in Houston lost contact with our Space Shuttle Columbia.
-[00:00:17.000 --> 00:00:23.000]   A short time later, debris was seen falling from the skies above Texas.
-[00:00:23.000 --> 00:00:29.000]   The Columbia's lost. There are no survivors.
-[00:00:29.000 --> 00:00:32.000]   On board was a crew of seven.
-[00:00:32.000 --> 00:00:39.000]   Colonel Rick Husband, Lieutenant Colonel Michael Anderson, Commander Laurel Clark,
-[00:00:39.000 --> 00:00:48.000]   Captain David Brown, Commander William McCool, Dr. Kultna Shavla, and Ilan Ramon,
-[00:00:48.000 --> 00:00:52.000]   a colonel in the Israeli Air Force.
-[00:00:52.000 --> 00:00:58.000]   These men and women assumed great risk in the service to all humanity.
-[00:00:58.000 --> 00:01:03.000]   In an age when space flight has come to seem almost routine,
-[00:01:03.000 --> 00:01:07.000]   it is easy to overlook the dangers of travel by rocket
-[00:01:07.000 --> 00:01:12.000]   and the difficulties of navigating the fierce outer atmosphere of the Earth.
-[00:01:12.000 --> 00:01:18.000]   These astronauts knew the dangers, and they faced them willingly,
-[00:01:18.000 --> 00:01:23.000]   knowing they had a high and noble purpose in life.
-[00:01:23.000 --> 00:01:31.000]   Because of their courage and daring and idealism, we will miss them all the more.
-[00:01:31.000 --> 00:01:36.000]   All Americans today are thinking as well of the families of these men and women
-[00:01:36.000 --> 00:01:40.000]   who have been given this sudden shock and grief.
-[00:01:40.000 --> 00:01:45.000]   You're not alone. Our entire nation grieves with you,
-[00:01:45.000 --> 00:01:52.000]   and those you love will always have the respect and gratitude of this country.
-[00:01:52.000 --> 00:01:56.000]   The cause in which they died will continue.
-[00:01:56.000 --> 00:02:04.000]   Mankind is led into the darkness beyond our world by the inspiration of discovery
-[00:02:04.000 --> 00:02:11.000]   and the longing to understand. Our journey into space will go on.
-[00:02:11.000 --> 00:02:16.000]   In the skies today, we saw destruction and tragedy.
-[00:02:16.000 --> 00:02:22.000]   Yet farther than we can see, there is comfort and hope.
-[00:02:22.000 --> 00:02:29.000]   In the words of the prophet Isaiah, "Lift your eyes and look to the heavens
-[00:02:29.000 --> 00:02:35.000]   who created all these. He who brings out the starry hosts one by one
-[00:02:35.000 --> 00:02:39.000]   and calls them each by name."
-[00:02:39.000 --> 00:02:46.000]   Because of His great power and mighty strength, not one of them is missing.
-[00:02:46.000 --> 00:02:55.000]   The same Creator who names the stars also knows the names of the seven souls we mourn today.
-[00:02:55.000 --> 00:03:01.000]   The crew of the shuttle Columbia did not return safely to earth,
-[00:03:01.000 --> 00:03:05.000]   yet we can pray that all are safely home.
-[00:03:05.000 --> 00:03:13.000]   May God bless the grieving families, and may God continue to bless America.
-[00:03:13.000 --> 00:03:19.000]   [Silence]
-
-
-whisper_print_timings:     fallbacks =   1 p /   0 h
-whisper_print_timings:     load time =   569.03 ms
-whisper_print_timings:      mel time =   146.85 ms
-whisper_print_timings:   sample time =   238.66 ms /   553 runs (    0.43 ms per run)
-whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per run)
-whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
-whisper_print_timings:    total time = 32733.52 ms
-```
-
-</details>
-
 ## Real-time audio input example
 
 This is a naive example of performing real-time inference on audio from your microphone.
@@ -617,9 +381,9 @@ The [stream](examples/stream) tool samples the audio every half a second and run
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-cmake -B build
+cmake -B build -DWHISPER_SDL2=ON
 cmake --build build --config Release
-./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
@@ -630,7 +394,7 @@ Adding the `--print-colors` argument will print the transcribed text using an ex
 to highlight words with high or low confidence:
 
 ```bash
-./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
+./build/bin/whisper-cli -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">
@@ -640,7 +404,7 @@ to highlight words with high or low confidence:
 For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:
 
 ```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
+$ ./build/bin/whisper-cli -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -664,7 +428,7 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr
 The `--max-len` argument can be used to obtain word-level timestamps. Simply use `-ml 1`:
 
 ```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
+$ ./build/bin/whisper-cli -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
 
 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@@ -711,7 +475,7 @@ Sample usage:
 ./models/download-ggml-model.sh small.en-tdrz
 
 # run as usual, adding the "-tdrz" command-line argument
-./main -f ./samples/a13.wav -m ./models/ggml-small.en-tdrz.bin -tdrz
+./build/bin/whisper-cli -f ./samples/a13.wav -m ./models/ggml-small.en-tdrz.bin -tdrz
 ...
 main: processing './samples/a13.wav' (480000 samples, 30.0 sec), 4 threads, 1 processors, lang = en, task = transcribe, tdrz = 1, timestamps = 1 ...
 ...
@@ -728,14 +492,14 @@ main: processing './samples/a13.wav' (480000 samples, 30.0 sec), 4 threads, 1 pr
 
 ## Karaoke-style movie generation (experimental)
 
-The [main](examples/main) example provides support for output of karaoke-style movies, where the
+The [whisper-cli](examples/cli) example provides support for output of karaoke-style movies, where the
 currently pronounced word is highlighted. Use the `-wts` argument and run the generated bash script.
 This requires to have `ffmpeg` installed.
 
 Here are a few _"typical"_ examples:
 
 ```bash
-./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
+./build/bin/whisper-cli -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
 source ./samples/jfk.wav.wts
 ffplay ./samples/jfk.wav.mp4
 ```
@@ -745,7 +509,7 @@ https://user-images.githubusercontent.com/1991296/199337465-dbee4b5e-9aeb-48a3-b
 ---
 
 ```bash
-./main -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owts
+./build/bin/whisper-cli -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owts
 source ./samples/mm0.wav.wts
 ffplay ./samples/mm0.wav.mp4
 ```
@@ -755,7 +519,7 @@ https://user-images.githubusercontent.com/1991296/199337504-cc8fd233-0cb7-4920-9
 ---
 
 ```bash
-./main -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owts
+./build/bin/whisper-cli -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owts
 source ./samples/gb0.wav.wts
 ffplay ./samples/gb0.wav.mp4
 ```
@@ -780,7 +544,7 @@ https://user-images.githubusercontent.com/1991296/223206245-2d36d903-cf8e-4f09-8
 ## Benchmarks
 
 In order to have an objective comparison of the performance of the inference across different system configurations,
-use the [bench](examples/bench) tool. The tool simply runs the Encoder part of the model and prints how much time it
+use the [whisper-bench](examples/bench) tool. The tool simply runs the Encoder part of the model and prints how much time it
 took to execute it. The results are summarized in the following Github issue:
 
 [Benchmark results](https://github.com/ggerganov/whisper.cpp/issues/89)
@@ -843,13 +607,12 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 
 | Example                                             | Web                                   | Description                                                                                                                     |
 | --------------------------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
-| [main](examples/main)                               | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
-| [bench](examples/bench)                             | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
-| [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                                                         |
-| [wchess](examples/wchess)                           | [wchess.wasm](examples/wchess)        | Voice-controlled chess                                                                                                          |
-| [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-cli](examples/cli)                         | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
+| [whisper-bench](examples/bench)                     | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
+| [whisper-stream](examples/stream)                   | [stream.wasm](examples/stream.wasm)   | Real-time transcription of raw microphone capture                                                                               |
+| [whisper-command](examples/command)                 | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic                                                         |
+| [whisper-server](examples/server)                   |                                       | HTTP transcription server with OAI-like API                                                                                     |
+| [whisper-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                                                                                    |
@@ -857,7 +620,7 @@ Some of the examples are even ported to run in the browser using WebAssembly. Ch
 | [generate-karaoke.sh](examples/generate-karaoke.sh) |                                       | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture                           |
 | [livestream.sh](examples/livestream.sh)             |                                       | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185)                                           |
 | [yt-wsp.sh](examples/yt-wsp.sh)                     |                                       | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
-| [server](examples/server)                           |                                       | HTTP transcription server with OAI-like API                                                                                     |
+| [wchess](examples/wchess)                           | [wchess.wasm](examples/wchess)        | Voice-controlled chess                                                                                                          |
 
 ## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)
 

bindings/java/README.md

@@ -67,5 +67,5 @@ copy /y ..\..\build\bin\Release\whisper.dll build\generated\resources\main\win32
 
 ## License
 
-The license for the Go bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
+The license for the Java bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
 

bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java

@@ -181,11 +181,11 @@ public class WhisperFullParams extends Structure {
     }
 
     /** Flag to suppress non-speech tokens. */
-    public CBool suppress_non_speech_tokens;
+    public CBool suppress_nst;
 
     /** Flag to suppress non-speech tokens. */
     public void suppressNonSpeechTokens(boolean enable) {
-        suppress_non_speech_tokens = enable ? CBool.TRUE : CBool.FALSE;
+        suppress_nst = enable ? CBool.TRUE : CBool.FALSE;
     }
 
     /** Initial decoding temperature. */
@@ -315,7 +315,7 @@ public class WhisperFullParams extends Structure {
                 "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
                 "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "audio_ctx",
                 "tdrz_enable", "suppress_regex", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
-                "suppress_blank", "suppress_non_speech_tokens", "temperature", "max_initial_ts", "length_penalty",
+                "suppress_blank", "suppress_nst", "temperature", "max_initial_ts", "length_penalty",
                 "temperature_inc", "entropy_thold", "logprob_thold", "no_speech_thold", "greedy", "beam_search",
                 "new_segment_callback", "new_segment_callback_user_data",
                 "progress_callback", "progress_callback_user_data",

bindings/javascript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.7.2",
+  "version": "1.7.4",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/README.md

@@ -22,7 +22,7 @@ Usage
 ```ruby
 require "whisper"
 
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 
 params = Whisper::Params.new
 params.language = "en"
@@ -41,21 +41,67 @@ end
 
 ### Preparing model ###
 
-Use script to download model file(s):
+Some models are prepared up-front:
 
-```bash
-git clone https://github.com/ggerganov/whisper.cpp.git
-cd whisper.cpp
-sh ./models/download-ggml-model.sh base.en
+```ruby
+base_en = Whisper::Model.pre_converted_models["base.en"]
+whisper = Whisper::Context.new(base_en)
 ```
 
-There are some types of models. See [models][] page for details.
+At first time you use a model, it is downloaded automatically. After that, downloaded cached file is used. To clear cache, call `#clear_cache`:
+
+```ruby
+Whisper::Model.pre_converted_models["base"].clear_cache
+```
+
+You also can use shorthand for pre-converted models:
+
+```ruby
+whisper = Whisper::Context.new("base.en")
+```
+
+You can see the list of prepared model names by `Whisper::Model.pre_converted_models.keys`:
+
+```ruby
+puts Whisper::Model.pre_converted_models.keys
+# tiny
+# tiny.en
+# tiny-q5_1
+# tiny.en-q5_1
+# tiny-q8_0
+# base
+# base.en
+# base-q5_1
+# base.en-q5_1
+# base-q8_0
+#   :
+#   :
+```
+
+You can also use local model files you prepared:
+
+```ruby
+whisper = Whisper::Context.new("path/to/your/model.bin")
+```
+
+Or, you can download model files:
+
+```ruby
+whisper = Whisper::Context.new("https://example.net/uri/of/your/model.bin")
+# Or
+whisper = Whisper::Context.new(URI("https://example.net/uri/of/your/model.bin"))
+```
+
+See [models][] page for details.
 
 ### Preparing audio file ###
 
 Currently, whisper.cpp accepts only 16-bit WAV files.
 
-### API ###
+API
+---
+
+### Segments ###
 
 Once `Whisper::Context#transcribe` called, you can retrieve segments by `#each_segment`:
 
@@ -85,13 +131,6 @@ end
 You can also add hook to params called on new segment:
 
 ```ruby
-def format_time(time_ms)
-  sec, decimal_part = time_ms.divmod(1000)
-  min, sec = sec.divmod(60)
-  hour, min = min.divmod(60)
-  "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
-end
-
 # Add hook before calling #transcribe
 params.on_new_segment do |segment|
   line = "[%{st} --> %{ed}] %{text}" % {
@@ -107,10 +146,12 @@ whisper.transcribe("path/to/audio.wav", params)
 
 ```
 
+### Models ###
+
 You can see model information:
 
 ```ruby
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 model = whisper.model
 
 model.n_vocab # => 51864
@@ -128,6 +169,8 @@ model.type # => "base"
 
 ```
 
+### Logging ###
+
 You can set log callback:
 
 ```ruby
@@ -157,9 +200,11 @@ Using this feature, you are also able to suppress log:
 Whisper.log_set ->(level, buffer, user_data) {
   # do nothing
 }, nil
-Whisper::Context.new(MODEL)
+Whisper::Context.new("base")
 ```
 
+### Low-level API to transcribe ###
+
 You can also call `Whisper::Context#full` and `#full_parallel` with a Ruby array as samples. Although `#transcribe` with audio file path is recommended because it extracts PCM samples in C++ and is fast, `#full` and `#full_parallel` give you flexibility.
 
 ```ruby
@@ -169,14 +214,25 @@ require "wavefile"
 reader = WaveFile::Reader.new("path/to/audio.wav", WaveFile::Format.new(:mono, :float, 16000))
 samples = reader.enum_for(:each_buffer).map(&:samples).flatten
 
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 whisper.full(Whisper::Params.new, samples)
 whisper.each_segment do |segment|
   puts segment.text
 end
 ```
 
-The second argument `samples` may be an array, an object with `length` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
+The second argument `samples` may be an array, an object with `length` and `each` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
+
+Development
+-----------
+
+    % git clone https://github.com/ggerganov/whisper.cpp.git
+    % cd whisper.cpp/bindings/ruby
+    % rake test
+
+First call of `rake test` builds an extension and downloads a model for testing. After that, you add tests in `tests` directory and modify `ext/ruby_whisper.cpp`.
+
+If something seems wrong on build, running `rake clean` solves some cases.
 
 License
 -------

bindings/ruby/Rakefile

@@ -18,24 +18,13 @@ EXTSOURCES.each do |src|
 end
 
 CLEAN.include SOURCES
-CLEAN.include FileList[
-                "ext/*.o",
-                "ext/*.metal",
-                "ext/whisper.{so,bundle,dll}",
-                "ext/depend"
-              ]
+CLEAN.include FileList["ext/*.o", "ext/*.metal", "ext/whisper.{so,bundle,dll}"]
 
-task build: FileList[
-       "ext/Makefile",
-       "ext/ruby_whisper.h",
-       "ext/ruby_whisper.cpp",
-       "whispercpp.gemspec",
-     ]
+task build: ["ext/Makefile", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp", "whispercpp.gemspec"]
 
 directory "pkg"
 CLOBBER.include "pkg"
 
-TEST_MODEL = "../../models/ggml-base.en.bin"
 LIB_NAME = "whisper".ext(RbConfig::CONFIG["DLEXT"])
 SO_FILE = File.join("ext", LIB_NAME)
 LIB_FILE = File.join("lib", LIB_NAME)
@@ -51,23 +40,17 @@ file SO_FILE => "ext/Makefile" do |t|
     sh "make"
   end
 end
-CLEAN.include LIB_FILE
+CLEAN.include SO_FILE
 
 directory "lib"
 file LIB_FILE => [SO_FILE, "lib"] do |t|
   copy t.source, t.name
 end
+CLEAN.include LIB_FILE
 
 Rake::TestTask.new do |t|
   t.test_files = FileList["tests/test_*.rb"]
 end
-task test: [TEST_MODEL, LIB_FILE]
-
-file TEST_MODEL do
-  Dir.chdir "../.." do
-    sh "./models/download-ggml-model.sh base.en"
-  end
-end
 
 TEST_MEMORY_VIEW = "tests/jfk_reader/jfk_reader.#{RbConfig::CONFIG['DLEXT']}"
 file TEST_MEMORY_VIEW => "tests/jfk_reader/jfk_reader.c" do |t|
@@ -77,4 +60,5 @@ file TEST_MEMORY_VIEW => "tests/jfk_reader/jfk_reader.c" do |t|
   end
 end
 CLEAN.include "tests/jfk_reader/jfk_reader.{o,#{RbConfig::CONFIG['DLEXT']}}"
-task test: TEST_MEMORY_VIEW
+
+task test: [LIB_FILE, TEST_MEMORY_VIEW]

bindings/ruby/ext/.gitignore

@@ -2,7 +2,6 @@ Makefile
 whisper.so
 whisper.bundle
 whisper.dll
-depend
 scripts/get-flags.mk
 *.o
 *.c

bindings/ruby/ext/extconf.rb

@@ -1,7 +1,7 @@
 require 'mkmf'
 
 # need to use c++ compiler flags
-$CXXFLAGS << ' -std=c++11'
+$CXXFLAGS << ' -std=c++17'
 
 $LDFLAGS << ' -lstdc++'
 
@@ -35,10 +35,10 @@ if $GGML_METAL
   $GGML_METAL_EMBED_LIBRARY = true
 end
 
-$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iinclude -Isrc -Iexamples'
+$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples'
 $MK_CFLAGS   = '-std=c11   -fPIC'
-$MK_CXXFLAGS = '-std=c++11 -fPIC'
-$MK_NVCCFLAGS = '-std=c++11'
+$MK_CXXFLAGS = '-std=c++17 -fPIC'
+$MK_NVCCFLAGS = '-std=c++17'
 $MK_LDFLAGS = ''
 
 $OBJ_GGML = []
@@ -111,11 +111,6 @@ unless ENV['RISCV']
     $MK_CFLAGS     << ' -march=native -mtune=native'
     $HOST_CXXFLAGS << ' -march=native -mtune=native'
   end
-
-  if $UNAME_M.match? /aarch64.*/
-    $MK_CFLAGS   << ' -mcpu=native'
-    $MK_CXXFLAGS << ' -mcpu=native'
-  end
 else
   $MK_CFLAGS   << ' -march=rv64gcv -mabi=lp64d'
   $MK_CXXFLAGS << ' -march=rv64gcv -mabi=lp64d'
@@ -162,7 +157,6 @@ end
 
 $OBJ_GGML <<
   'ggml/src/ggml.o' <<
-  'ggml/src/ggml-aarch64.o' <<
   'ggml/src/ggml-alloc.o' <<
   'ggml/src/ggml-backend.o' <<
   'ggml/src/ggml-backend-reg.o' <<
@@ -172,7 +166,9 @@ $OBJ_GGML <<
   'ggml/src/ggml-cpu/ggml-cpu.o' <<
   'ggml/src/ggml-cpu/ggml-cpu-cpp.o' <<
   'ggml/src/ggml-cpu/ggml-cpu-aarch64.o' <<
-  'ggml/src/ggml-cpu/ggml-cpu-quants.o'
+  'ggml/src/ggml-cpu/ggml-cpu-hbm.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-quants.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-traits.o'
 
 $OBJ_WHISPER <<
   'src/whisper.o'

bindings/ruby/ext/ruby_whisper.cpp

@@ -38,6 +38,9 @@ VALUE cContext;
 VALUE cParams;
 VALUE eError;
 
+VALUE cSegment;
+VALUE cModel;
+
 static ID id_to_s;
 static ID id_call;
 static ID id___method__;
@@ -45,9 +48,15 @@ static ID id_to_enum;
 static ID id_length;
 static ID id_next;
 static ID id_new;
+static ID id_to_path;
+static ID id_URI;
+static ID id_pre_converted_models;
 
 static bool is_log_callback_finalized = false;
 
+// High level API
+static VALUE rb_whisper_segment_initialize(VALUE context, int index);
+
 /*
  * call-seq:
  *   lang_max_id -> Integer
@@ -182,10 +191,74 @@ static ruby_whisper_callback_container * rb_whisper_callback_container_allocate(
   return container;
 }
 
+static void new_segment_callback(struct whisper_context *ctx, struct whisper_state *state, int n_new, void *user_data) {
+  const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
+
+  // Currently, doesn't support state because
+  // those require to resolve GC-related problems.
+  if (!NIL_P(container->callback)) {
+    rb_funcall(container->callback, id_call, 4, *container->context, Qnil, INT2NUM(n_new), container->user_data);
+  }
+  const long callbacks_len = RARRAY_LEN(container->callbacks);
+  if (0 == callbacks_len) {
+    return;
+  }
+  const int n_segments = whisper_full_n_segments_from_state(state);
+  for (int i = n_new; i > 0; i--) {
+    int i_segment = n_segments - i;
+    VALUE segment = rb_whisper_segment_initialize(*container->context, i_segment);
+    for (int j = 0; j < callbacks_len; j++) {
+      VALUE cb = rb_ary_entry(container->callbacks, j);
+      rb_funcall(cb, id_call, 1, segment);
+    }
+  }
+}
+
+static void progress_callback(struct whisper_context *ctx, struct whisper_state *state, int progress_cur, void *user_data) {
+  const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
+  const VALUE progress = INT2NUM(progress_cur);
+  // Currently, doesn't support state because
+  // those require to resolve GC-related problems.
+  if (!NIL_P(container->callback)) {
+    rb_funcall(container->callback, id_call, 4, *container->context, Qnil, progress, container->user_data);
+  }
+  const long callbacks_len = RARRAY_LEN(container->callbacks);
+  if (0 == callbacks_len) {
+    return;
+  }
+  for (int j = 0; j < callbacks_len; j++) {
+    VALUE cb = rb_ary_entry(container->callbacks, j);
+    rb_funcall(cb, id_call, 1, progress);
+  }
+}
+
+static bool abort_callback(void * user_data) {
+  const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
+  if (!NIL_P(container->callback)) {
+    VALUE result = rb_funcall(container->callback, id_call, 1, container->user_data);
+    if (!NIL_P(result) && Qfalse != result) {
+      return true;
+    }
+  }
+  const long callbacks_len = RARRAY_LEN(container->callbacks);
+  if (0 == callbacks_len) {
+    return false;
+  }
+  for (int j = 0; j < callbacks_len; j++) {
+    VALUE cb = rb_ary_entry(container->callbacks, j);
+    VALUE result = rb_funcall(cb, id_call, 1, container->user_data);
+    if (!NIL_P(result) && Qfalse != result) {
+      return true;
+    }
+  }
+  return false;
+}
+
 static VALUE ruby_whisper_params_allocate(VALUE klass) {
   ruby_whisper_params *rwp;
   rwp = ALLOC(ruby_whisper_params);
   rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  rwp->diarize = false;
   rwp->new_segment_callback_container = rb_whisper_callback_container_allocate();
   rwp->progress_callback_container = rb_whisper_callback_container_allocate();
   rwp->abort_callback_container = rb_whisper_callback_container_allocate();
@@ -194,7 +267,9 @@ static VALUE ruby_whisper_params_allocate(VALUE klass) {
 
 /*
  * call-seq:
+ *   new("base.en") -> Whisper::Context
  *   new("path/to/model.bin") -> Whisper::Context
+ *   new(Whisper::Model::URI.new("https://example.net/uri/of/model.bin")) -> Whisper::Context
  */
 static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   ruby_whisper *rw;
@@ -204,6 +279,25 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   rb_scan_args(argc, argv, "01", &whisper_model_file_path);
   Data_Get_Struct(self, ruby_whisper, rw);
 
+  VALUE pre_converted_models = rb_funcall(cModel, id_pre_converted_models, 0);
+  VALUE pre_converted_model = rb_hash_aref(pre_converted_models, whisper_model_file_path);
+  if (!NIL_P(pre_converted_model)) {
+    whisper_model_file_path = pre_converted_model;
+  }
+  if (TYPE(whisper_model_file_path) == T_STRING) {
+    const char * whisper_model_file_path_str = StringValueCStr(whisper_model_file_path);
+    if (strncmp("http://", whisper_model_file_path_str, 7) == 0 || strncmp("https://", whisper_model_file_path_str, 8) == 0) {
+      VALUE uri_class = rb_const_get(cModel, id_URI);
+      whisper_model_file_path = rb_class_new_instance(1, &whisper_model_file_path, uri_class);
+    }
+  }
+  if (rb_obj_is_kind_of(whisper_model_file_path, rb_path2class("URI::HTTP"))) {
+    VALUE uri_class = rb_const_get(cModel, id_URI);
+    whisper_model_file_path = rb_class_new_instance(1, &whisper_model_file_path, uri_class);
+  }
+  if (rb_respond_to(whisper_model_file_path, id_to_path)) {
+    whisper_model_file_path = rb_funcall(whisper_model_file_path, id_to_path, 0);
+  }
   if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
     rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
   }
@@ -214,8 +308,25 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   return self;
 }
 
-// High level API
-static VALUE rb_whisper_segment_initialize(VALUE context, int index);
+static void register_callbacks(ruby_whisper_params * rwp, VALUE * self) {
+  if (!NIL_P(rwp->new_segment_callback_container->callback) || 0 != RARRAY_LEN(rwp->new_segment_callback_container->callbacks)) {
+    rwp->new_segment_callback_container->context = self;
+    rwp->params.new_segment_callback = new_segment_callback;
+    rwp->params.new_segment_callback_user_data = rwp->new_segment_callback_container;
+  }
+
+  if (!NIL_P(rwp->progress_callback_container->callback) || 0 != RARRAY_LEN(rwp->progress_callback_container->callbacks)) {
+    rwp->progress_callback_container->context = self;
+    rwp->params.progress_callback = progress_callback;
+    rwp->params.progress_callback_user_data = rwp->progress_callback_container;
+  }
+
+  if (!NIL_P(rwp->abort_callback_container->callback) || 0 != RARRAY_LEN(rwp->abort_callback_container->callbacks)) {
+    rwp->abort_callback_container->context = self;
+    rwp->params.abort_callback = abort_callback;
+    rwp->params.abort_callback_user_data = rwp->abort_callback_container;
+  }
+}
 
 /*
  * transcribe a single file
@@ -337,80 +448,7 @@ static VALUE ruby_whisper_transcribe(int argc, VALUE *argv, VALUE self) {
     rwp->params.encoder_begin_callback_user_data = &is_aborted;
   }
 
-  if (!NIL_P(rwp->new_segment_callback_container->callback) || 0 != RARRAY_LEN(rwp->new_segment_callback_container->callbacks)) {
-    rwp->params.new_segment_callback = [](struct whisper_context * ctx, struct whisper_state * state, int n_new, void * user_data) {
-      const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
-
-      // Currently, doesn't support state because
-      // those require to resolve GC-related problems.
-      if (!NIL_P(container->callback)) {
-        rb_funcall(container->callback, id_call, 4, *container->context, Qnil, INT2NUM(n_new), container->user_data);
-      }
-      const long callbacks_len = RARRAY_LEN(container->callbacks);
-      if (0 == callbacks_len) {
-        return;
-      }
-      const int n_segments = whisper_full_n_segments_from_state(state);
-      for (int i = n_new; i > 0; i--) {
-        int i_segment = n_segments - i;
-        VALUE segment = rb_whisper_segment_initialize(*container->context, i_segment);
-        for (int j = 0; j < callbacks_len; j++) {
-          VALUE cb = rb_ary_entry(container->callbacks, j);
-          rb_funcall(cb, id_call, 1, segment);
-        }
-      }
-    };
-    rwp->new_segment_callback_container->context = &self;
-    rwp->params.new_segment_callback_user_data = rwp->new_segment_callback_container;
-  }
-
-  if (!NIL_P(rwp->progress_callback_container->callback) || 0 != RARRAY_LEN(rwp->progress_callback_container->callbacks)) {
-    rwp->params.progress_callback = [](struct whisper_context *ctx, struct whisper_state * /*state*/, int progress_cur, void *user_data) {
-      const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
-      const VALUE progress = INT2NUM(progress_cur);
-      // Currently, doesn't support state because
-      // those require to resolve GC-related problems.
-      if (!NIL_P(container->callback)) {
-        rb_funcall(container->callback, id_call, 4, *container->context, Qnil, progress, container->user_data);
-      }
-      const long callbacks_len = RARRAY_LEN(container->callbacks);
-      if (0 == callbacks_len) {
-        return;
-      }
-      for (int j = 0; j < callbacks_len; j++) {
-        VALUE cb = rb_ary_entry(container->callbacks, j);
-        rb_funcall(cb, id_call, 1, progress);
-      }
-    };
-    rwp->progress_callback_container->context = &self;
-    rwp->params.progress_callback_user_data = rwp->progress_callback_container;
-  }
-
-  if (!NIL_P(rwp->abort_callback_container->callback) || 0 != RARRAY_LEN(rwp->abort_callback_container->callbacks)) {
-    rwp->params.abort_callback = [](void * user_data) {
-      const ruby_whisper_callback_container *container = (ruby_whisper_callback_container *)user_data;
-      if (!NIL_P(container->callback)) {
-        VALUE result = rb_funcall(container->callback, id_call, 1, container->user_data);
-        if (!NIL_P(result) && Qfalse != result) {
-          return true;
-        }
-      }
-      const long callbacks_len = RARRAY_LEN(container->callbacks);
-      if (0 == callbacks_len) {
-        return false;
-      }
-      for (int j = 0; j < callbacks_len; j++) {
-        VALUE cb = rb_ary_entry(container->callbacks, j);
-        VALUE result = rb_funcall(cb, id_call, 1, container->user_data);
-        if (!NIL_P(result) && Qfalse != result) {
-          return true;
-        }
-      }
-      return false;
-    };
-    rwp->abort_callback_container->context = &self;
-    rwp->params.abort_callback_user_data = rwp->abort_callback_container;
-  }
+  register_callbacks(rwp, &self);
 
   if (whisper_full_parallel(rw->context, rwp->params, pcmf32.data(), pcmf32.size(), 1) != 0) {
     fprintf(stderr, "failed to process audio\n");
@@ -615,6 +653,7 @@ VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self) {
       }
     }
   }
+  register_callbacks(rwp, &self);
   const int result = whisper_full(rw->context, rwp->params, c_samples, n_samples);
   if (0 == result) {
     return Qnil;
@@ -703,6 +742,7 @@ static VALUE ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self) {
       }
     }
   }
+  register_callbacks(rwp, &self);
   const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
   if (0 == result) {
     return Qnil;
@@ -807,6 +847,18 @@ static VALUE ruby_whisper_full_get_segment_text(VALUE self, VALUE i_segment) {
   return rb_str_new2(text);
 }
 
+/*
+ * call-seq:
+ *   full_get_segment_no_speech_prob(segment_index) -> Float
+ */
+static VALUE ruby_whisper_full_get_segment_no_speech_prob(VALUE self, VALUE i_segment) {
+  ruby_whisper *rw;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  const int c_i_segment = ruby_whisper_full_check_segment_index(rw, i_segment);
+  const float no_speech_prob = whisper_full_get_segment_no_speech_prob(rw->context, c_i_segment);
+  return DBL2NUM(no_speech_prob);
+}
+
 /*
  * params.language = "auto" | "en", etc...
  *
@@ -963,19 +1015,19 @@ static VALUE ruby_whisper_params_get_suppress_blank(VALUE self) {
 }
 /*
  * call-seq:
- *   suppress_non_speech_tokens = force_suppress -> force_suppress
+ *   suppress_nst = force_suppress -> force_suppress
  */
-static VALUE ruby_whisper_params_set_suppress_non_speech_tokens(VALUE self, VALUE value) {
-  BOOL_PARAMS_SETTER(self, suppress_non_speech_tokens, value)
+static VALUE ruby_whisper_params_set_suppress_nst(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, suppress_nst, value)
 }
 /*
  * If true, suppresses non-speech-tokens.
  *
  * call-seq:
- *   suppress_non_speech_tokens -> bool
+ *   suppress_nst -> bool
  */
-static VALUE ruby_whisper_params_get_suppress_non_speech_tokens(VALUE self) {
-  BOOL_PARAMS_GETTER(self, suppress_non_speech_tokens)
+static VALUE ruby_whisper_params_get_suppress_nst(VALUE self) {
+  BOOL_PARAMS_GETTER(self, suppress_nst)
 }
 /*
  * If true, enables token-level timestamps.
@@ -1245,6 +1297,25 @@ static VALUE ruby_whisper_params_set_logprob_thold(VALUE self, VALUE value) {
   rwp->params.logprob_thold = RFLOAT_VALUE(value);
   return value;
 }
+/*
+ * call-seq:
+ *   no_speech_thold -> Float
+ */
+static VALUE ruby_whisper_params_get_no_speech_thold(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return DBL2NUM(rwp->params.no_speech_thold);
+}
+/*
+ * call-seq:
+ *   no_speech_thold = threshold -> threshold
+ */
+static VALUE ruby_whisper_params_set_no_speech_thold(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.no_speech_thold = RFLOAT_VALUE(value);
+  return value;
+}
 /*
  * Sets new segment callback, called for every newly generated text segment.
  *
@@ -1341,9 +1412,6 @@ typedef struct {
   VALUE context;
 } ruby_whisper_model;
 
-VALUE cSegment;
-VALUE cModel;
-
 static void rb_whisper_segment_mark(ruby_whisper_segment *rws) {
   rb_gc_mark(rws->context);
 }
@@ -1515,6 +1583,18 @@ static VALUE ruby_whisper_segment_get_text(VALUE self) {
   return rb_str_new2(text);
 }
 
+/*
+ * call-seq:
+ *   no_speech_prob -> Float
+ */
+static VALUE ruby_whisper_segment_get_no_speech_prob(VALUE self) {
+  ruby_whisper_segment *rws;
+  Data_Get_Struct(self, ruby_whisper_segment, rws);
+  ruby_whisper *rw;
+  Data_Get_Struct(rws->context, ruby_whisper, rw);
+  return DBL2NUM(whisper_full_get_segment_no_speech_prob(rw->context, rws->index));
+}
+
 static void rb_whisper_model_mark(ruby_whisper_model *rwm) {
   rb_gc_mark(rwm->context);
 }
@@ -1687,7 +1767,7 @@ static VALUE ruby_whisper_c_model_type(VALUE self) {
 
 static VALUE ruby_whisper_error_initialize(VALUE self, VALUE code) {
   const int c_code = NUM2INT(code);
-  char *raw_message;
+  const char *raw_message;
   switch (c_code) {
   case -2:
     raw_message = "failed to compute log mel spectrogram";
@@ -1733,6 +1813,9 @@ void Init_whisper() {
   id_length = rb_intern("length");
   id_next = rb_intern("next");
   id_new = rb_intern("new");
+  id_to_path = rb_intern("to_path");
+  id_URI = rb_intern("URI");
+  id_pre_converted_models = rb_intern("pre_converted_models");
 
   mWhisper = rb_define_module("Whisper");
   cContext = rb_define_class_under(mWhisper, "Context", rb_cObject);
@@ -1775,6 +1858,7 @@ void Init_whisper() {
   rb_define_method(cContext, "full_get_segment_t1", ruby_whisper_full_get_segment_t1, 1);
   rb_define_method(cContext, "full_get_segment_speaker_turn_next", ruby_whisper_full_get_segment_speaker_turn_next, 1);
   rb_define_method(cContext, "full_get_segment_text", ruby_whisper_full_get_segment_text, 1);
+  rb_define_method(cContext, "full_get_segment_no_speech_prob", ruby_whisper_full_get_segment_no_speech_prob, 1);
   rb_define_method(cContext, "full", ruby_whisper_full, -1);
   rb_define_method(cContext, "full_parallel", ruby_whisper_full_parallel, -1);
 
@@ -1798,8 +1882,8 @@ void Init_whisper() {
   rb_define_method(cParams, "print_timestamps=", ruby_whisper_params_set_print_timestamps, 1);
   rb_define_method(cParams, "suppress_blank", ruby_whisper_params_get_suppress_blank, 0);
   rb_define_method(cParams, "suppress_blank=", ruby_whisper_params_set_suppress_blank, 1);
-  rb_define_method(cParams, "suppress_non_speech_tokens", ruby_whisper_params_get_suppress_non_speech_tokens, 0);
-  rb_define_method(cParams, "suppress_non_speech_tokens=", ruby_whisper_params_set_suppress_non_speech_tokens, 1);
+  rb_define_method(cParams, "suppress_nst", ruby_whisper_params_get_suppress_nst, 0);
+  rb_define_method(cParams, "suppress_nst=", ruby_whisper_params_set_suppress_nst, 1);
   rb_define_method(cParams, "token_timestamps", ruby_whisper_params_get_token_timestamps, 0);
   rb_define_method(cParams, "token_timestamps=", ruby_whisper_params_set_token_timestamps, 1);
   rb_define_method(cParams, "split_on_word", ruby_whisper_params_get_split_on_word, 0);
@@ -1828,6 +1912,8 @@ void Init_whisper() {
   rb_define_method(cParams, "entropy_thold=", ruby_whisper_params_set_entropy_thold, 1);
   rb_define_method(cParams, "logprob_thold", ruby_whisper_params_get_logprob_thold, 0);
   rb_define_method(cParams, "logprob_thold=", ruby_whisper_params_set_logprob_thold, 1);
+  rb_define_method(cParams, "no_speech_thold", ruby_whisper_params_get_no_speech_thold, 0);
+  rb_define_method(cParams, "no_speech_thold=", ruby_whisper_params_set_no_speech_thold, 1);
 
   rb_define_method(cParams, "new_segment_callback=", ruby_whisper_params_set_new_segment_callback, 1);
   rb_define_method(cParams, "new_segment_callback_user_data=", ruby_whisper_params_set_new_segment_callback_user_data, 1);
@@ -1851,6 +1937,7 @@ void Init_whisper() {
   rb_define_method(cSegment, "end_time", ruby_whisper_segment_get_end_time, 0);
   rb_define_method(cSegment, "speaker_next_turn?", ruby_whisper_segment_get_speaker_turn_next, 0);
   rb_define_method(cSegment, "text", ruby_whisper_segment_get_text, 0);
+  rb_define_method(cSegment, "no_speech_prob", ruby_whisper_segment_get_no_speech_prob, 0);
 
   cModel = rb_define_class_under(mWhisper, "Model", rb_cObject);
   rb_define_alloc_func(cModel, ruby_whisper_model_allocate);
@@ -1867,6 +1954,8 @@ void Init_whisper() {
   rb_define_method(cModel, "n_mels", ruby_whisper_c_model_n_mels, 0);
   rb_define_method(cModel, "ftype", ruby_whisper_c_model_ftype, 0);
   rb_define_method(cModel, "type", ruby_whisper_c_model_type, 0);
+
+  rb_require("whisper/model/uri");
 }
 #ifdef __cplusplus
 }

bindings/ruby/ext/ruby_whisper.h

@@ -1,5 +1,5 @@
-#ifndef __RUBY_WHISPER_H
-#define __RUBY_WHISPER_H
+#ifndef RUBY_WHISPER_H
+#define RUBY_WHISPER_H
 
 #include "whisper.h"
 

bindings/ruby/lib/whisper/model/uri.rb

@@ -0,0 +1,163 @@
+require "uri"
+require "net/http"
+require "time"
+require "pathname"
+require "io/console/size"
+
+module Whisper
+  class Model
+    class URI
+      def initialize(uri)
+        @uri = URI(uri)
+      end
+
+      def to_path
+        cache
+        cache_path.to_path
+      end
+
+      def clear_cache
+        path = cache_path
+        path.delete if path.exist?
+      end
+
+      private
+
+      def cache_path
+        base_cache_dir/@uri.host/@uri.path[1..]
+      end
+
+      def base_cache_dir
+        base = case RUBY_PLATFORM
+               when /mswin|mingw/
+                 ENV.key?("LOCALAPPDATA") ? Pathname(ENV["LOCALAPPDATA"]) : Pathname(Dir.home)/"AppData/Local"
+               when /darwin/
+                 Pathname(Dir.home)/"Library/Caches"
+               else
+                 ENV.key?("XDG_CACHE_HOME") ? ENV["XDG_CACHE_HOME"] : Pathname(Dir.home)/".cache"
+               end
+        base/"whisper.cpp"
+      end
+
+      def cache
+        path = cache_path
+        headers = {}
+        headers["if-modified-since"] = path.mtime.httpdate if path.exist?
+        request @uri, headers
+        path
+      end
+
+      def request(uri, headers)
+        Net::HTTP.start uri.host, uri.port, use_ssl: uri.scheme == "https" do |http|
+          request = Net::HTTP::Get.new(uri, headers)
+          http.request request do |response|
+            case response
+            when Net::HTTPNotModified
+            # noop
+            when Net::HTTPOK
+              download response
+            when Net::HTTPRedirection
+              request URI(response["location"]), headers
+            else
+              return if headers.key?("if-modified-since") # Use cache file
+
+              raise "#{response.code} #{response.message}\n#{response.body}"
+            end
+          end
+        end
+      end
+
+      def download(response)
+        path = cache_path
+        path.dirname.mkpath unless path.dirname.exist?
+        downloading_path = Pathname("#{path}.downloading")
+        size = response.content_length
+        downloading_path.open "wb" do |file|
+          downloaded = 0
+          response.read_body do |chunk|
+            file << chunk
+            downloaded += chunk.bytesize
+            show_progress downloaded, size
+          end
+          $stderr.puts
+        end
+        downloading_path.rename path
+      end
+
+      def show_progress(current, size)
+        progress_rate_available = size && $stderr.tty?
+
+        unless @prev
+          @prev = Time.now
+          $stderr.puts "Downloading #{@uri} to #{cache_path}"
+        end
+
+        now = Time.now
+
+        if progress_rate_available
+          return if now - @prev < 1 && current < size
+
+          progress_width = 20
+          progress = current.to_f / size
+          arrow_length = progress * progress_width
+          arrow = "=" * (arrow_length - 1) + ">" + " " * (progress_width - arrow_length)
+          line = "[#{arrow}] (#{format_bytesize(current)} / #{format_bytesize(size)})"
+          padding = ' ' * ($stderr.winsize[1] - line.size)
+          $stderr.print "\r#{line}#{padding}"
+        else
+          return if now - @prev < 1
+
+          $stderr.print "."
+        end
+        @prev = now
+      end
+
+      def format_bytesize(bytesize)
+        return "0.0 B" if bytesize.zero?
+
+        units = %w[B KiB MiB GiB TiB]
+        exp = (Math.log(bytesize) / Math.log(1024)).to_i
+        format("%.1f %s", bytesize.to_f / 1024 ** exp, units[exp])
+      end
+    end
+
+    @pre_converted_models = %w[
+      tiny
+      tiny.en
+      tiny-q5_1
+      tiny.en-q5_1
+      tiny-q8_0
+      base
+      base.en
+      base-q5_1
+      base.en-q5_1
+      base-q8_0
+      small
+      small.en
+      small.en-tdrz
+      small-q5_1
+      small.en-q5_1
+      small-q8_0
+      medium
+      medium.en
+      medium-q5_0
+      medium.en-q5_0
+      medium-q8_0
+      large-v1
+      large-v2
+      large-v2-q5_0
+      large-v2-q8_0
+      large-v3
+      large-v3-q5_0
+      large-v3-turbo
+      large-v3-turbo-q5_0
+      large-v3-turbo-q8_0
+    ].each_with_object({}) {|name, models|
+      models[name] = URI.new("https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-#{name}.bin")
+    }
+
+    class << self
+      attr_reader :pre_converted_models
+    end
+  end
+end

bindings/ruby/sig/whisper.rbs

@@ -0,0 +1,153 @@
+module Whisper
+  interface _Samples
+    def length: () -> Integer
+    def each: { (Float) -> void } -> void
+  end
+
+  type log_callback = ^(Integer level, String message, Object user_data) -> void
+  type new_segment_callback = ^(Whisper::Context, void, Integer n_new, Object user_data) -> void
+  type progress_callback = ^(Whisper::Context, void, Integer progress, Object user_data) -> void
+  type abort_callback = ^(Whisper::Context, void, Object user_data) -> boolish
+
+  LOG_LEVEL_NONE: Integer
+  LOG_LEVEL_INFO: Integer
+  LOG_LEVEL_WARN: Integer
+  LOG_LEVEL_ERROR: Integer
+  LOG_LEVEL_DEBUG: Integer
+  LOG_LEVEL_CONT: Integer
+
+  def self.lang_max_id: () -> Integer
+  def self.lang_id: (string name) -> Integer
+  def self.lang_str: (Integer id) -> String
+  def self.lang_str_full: (Integer id) -> String
+  def self.log_set=: (log_callback) -> log_callback
+  def self.finalize_log_callback: (void) -> void # Second argument of ObjectSpace.define_finalizer
+
+  class Context
+    def initialize: (string | _ToPath | ::URI::HTTP ) -> void
+    def transcribe: (string, Params) -> void
+                  | (string, Params) { (String) -> void } -> void
+    def model_n_vocab: () -> Integer
+    def model_n_audio_ctx: () -> Integer
+    def model_n_audio_state: () -> Integer
+    def model_n_text_head: () -> Integer
+    def model_n_text_layer: () -> Integer
+    def model_n_mels: () -> Integer
+    def model_ftype: () -> Integer
+    def model_type: () -> String
+    def full_n_segments: () -> Integer
+    def full_lang_id: () -> Integer
+    def full_get_segment_t0: (Integer) -> Integer
+    def full_get_segment_t1: (Integer) -> Integer
+    def full_get_segment_speaker_turn_next: (Integer) -> (true | false)
+    def full_get_segment_text: (Integer) -> String
+    def full_get_segment_no_speech_prob: (Integer) -> Float
+    def full: (Params, Array[Float], ?Integer) -> void
+            | (Params, _Samples, ?Integer) -> void
+    def full_parallel: (Params, Array[Float], ?Integer) -> void
+                     | (Params, _Samples, ?Integer) -> void
+                     | (Params, _Samples, ?Integer?, Integer) -> void
+    def each_segment: { (Segment) -> void } -> void
+                    | () -> Enumerator[Segment]
+    def model: () -> Model
+  end
+
+  class Params
+    def initialize: () -> void
+    def language=: (String) -> String # TODO: Enumerate lang names
+    def language: () -> String
+    def translate=: (boolish) -> boolish
+    def translate: () -> (true | false)
+    def no_context=: (boolish) -> boolish
+    def no_context: () -> (true | false)
+    def single_segment=: (boolish) -> boolish
+    def single_segment: () -> (true | false)
+    def print_special=: (boolish) -> boolish
+    def print_special: () -> (true | false)
+    def print_progress=: (boolish) -> boolish
+    def print_progress: () -> (true | false)
+    def print_realtime=: (boolish) -> boolish
+    def print_realtime: () -> (true | false)
+    def print_timestamps=: (boolish) -> boolish
+    def print_timestamps: () -> (true | false)
+    def suppress_blank=: (boolish) -> boolish
+    def suppress_blank: () -> (true | false)
+    def suppress_nst=: (boolish) -> boolish
+    def suppress_nst: () -> (true | false)
+    def token_timestamps=: (boolish) -> boolish
+    def token_timestamps: () -> (true | false)
+    def split_on_word=: (boolish) -> boolish
+    def split_on_word: () -> (true | false)
+    def initial_prompt=: (_ToS) -> _ToS
+    def initial_prompt: () -> String
+    def diarize=: (boolish) -> boolish
+    def diarize: () -> (true | false)
+    def offset=: (Integer) -> Integer
+    def offset: () -> Integer
+    def duration=: (Integer) -> Integer
+    def duration: () -> Integer
+    def max_text_tokens=: (Integer) -> Integer
+    def max_text_tokens: () -> Integer
+    def temperature=: (Float) -> Float
+    def temperature: () -> Float
+    def max_initial_ts=: (Float) -> Float
+    def max_initial_ts: () -> Float
+    def length_penalty=: (Float) -> Float
+    def length_penalty: () -> Float
+    def temperature_inc=: (Float) -> Float
+    def temperature_inc: () -> Float
+    def entropy_thold=: (Float) -> Float
+    def entropy_thold: () -> Float
+    def logprob_thold=: (Float) -> Float
+    def logprob_thold: () -> Float
+    def no_speech_thold=: (Float) -> Float
+    def no_speech_thold: () -> Float
+    def new_segment_callback=: (new_segment_callback) -> new_segment_callback
+    def new_segment_callback_user_data=: (Object) -> Object
+    def progress_callback=: (progress_callback) -> progress_callback
+    def progress_callback_user_data=: (Object) -> Object
+    def abort_callback=: (abort_callback) -> abort_callback
+    def abort_callback_user_data=: (Object) -> Object
+    def on_new_segment: { (Segment) -> void } -> void
+    def on_progress: { (Integer) -> void } -> void
+    def abort_on: { (Object) -> boolish } -> void
+  end
+
+  class Model
+    def self.pre_converted_models: () -> Hash[String, Model::URI]
+    def initialize: () -> void
+    def n_vocab: () -> Integer
+    def n_audio_ctx: () -> Integer
+    def n_audio_state: () -> Integer
+    def n_audio_head: () -> Integer
+    def n_audio_layer: () -> Integer
+    def n_text_ctx: () -> Integer
+    def n_text_state: () -> Integer
+    def n_text_head: () -> Integer
+    def n_text_layer: () -> Integer
+    def n_mels: () -> Integer
+    def ftype: () -> Integer
+    def type: () -> String
+
+    class URI
+      def initialize: (string | ::URI::HTTP) -> void
+      def to_path: -> String
+      def clear_cache: -> void
+    end
+  end
+
+  class Segment
+    def initialize: () -> void
+    def start_time: () -> Integer
+    def end_time: () -> Integer
+    def speaker_next_turn?: () -> (true | false)
+    def text: () -> String
+    def no_speech_prob: () -> Float
+  end
+
+  class Error < StandardError
+    attr_reader code: Integer
+
+    def initialize: (Integer) -> void
+  end
+end

bindings/ruby/tests/helper.rb

@@ -3,6 +3,22 @@ require "whisper"
 require_relative "jfk_reader/jfk_reader"
 
 class TestBase < Test::Unit::TestCase
-  MODEL = File.join(__dir__, "..", "..", "..", "models", "ggml-base.en.bin")
   AUDIO = File.join(__dir__, "..", "..", "..", "samples", "jfk.wav")
+
+  class << self
+    attr_reader :whisper
+
+    def startup
+      @whisper = Whisper::Context.new("base.en")
+      params = Whisper::Params.new
+      params.print_timestamps = false
+      @whisper.transcribe(TestBase::AUDIO, params)
+    end
+  end
+
+  private
+
+  def whisper
+    self.class.whisper
+  end
 end

bindings/ruby/tests/jfk_reader/jfk_reader.c

@@ -60,49 +60,9 @@ static const rb_memory_view_entry_t jfk_reader_view_entry = {
   jfk_reader_memory_view_available_p
 };
 
-static VALUE
-read_jfk(int argc, VALUE *argv, VALUE obj)
-{
-  const char *audio_path_str = StringValueCStr(argv[0]);
-  const int n_samples = 176000;
-
-  short samples[n_samples];
-  FILE *file = fopen(audio_path_str, "rb");
-
-  fseek(file, 78, SEEK_SET);
-  fread(samples, sizeof(short), n_samples, file);
-  fclose(file);
-
-  VALUE rb_samples = rb_ary_new2(n_samples);
-  for (int i = 0; i < n_samples; i++) {
-    rb_ary_push(rb_samples, INT2FIX(samples[i]));
-  }
-
-  VALUE rb_data = rb_ary_new2(n_samples);
-  for (int i = 0; i < n_samples; i++) {
-    rb_ary_push(rb_data, DBL2NUM(samples[i]/32768.0));
-  }
-
-  float data[n_samples];
-  for (int i = 0; i < n_samples; i++) {
-    data[i] = samples[i]/32768.0;
-  }
-  void *c_data = (void *)data;
-  VALUE rb_void = rb_enc_str_new((const char *)c_data, sizeof(data), rb_ascii8bit_encoding());
-
-  VALUE rb_result = rb_ary_new3(3, rb_samples, rb_data, rb_void);
-  return rb_result;
-}
-
 void Init_jfk_reader(void)
 {
   VALUE cJFKReader = rb_define_class("JFKReader", rb_cObject);
   rb_memory_view_register(cJFKReader, &jfk_reader_view_entry);
   rb_define_method(cJFKReader, "initialize", jfk_reader_initialize, 1);
-
-
-  rb_define_global_function("read_jfk", read_jfk, -1);
-
-
-
 }

bindings/ruby/tests/test_callback.rb

@@ -1,14 +1,11 @@
-require "test/unit"
-require "whisper"
-
-class TestCallback < Test::Unit::TestCase
-  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+require_relative "helper"
 
+class TestCallback < TestBase
   def setup
     GC.start
     @params = Whisper::Params.new
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
-    @audio = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+    @whisper = Whisper::Context.new("base.en")
+    @audio = File.join(AUDIO)
   end
 
   def test_new_segment_callback

bindings/ruby/tests/test_model.rb

@@ -1,13 +1,14 @@
 require_relative "helper"
+require "pathname"
 
 class TestModel < TestBase
   def test_model
-    whisper = Whisper::Context.new(MODEL)
+    whisper = Whisper::Context.new("base.en")
     assert_instance_of Whisper::Model, whisper.model
   end
 
   def test_attributes
-    whisper = Whisper::Context.new(MODEL)
+    whisper = Whisper::Context.new("base.en")
     model = whisper.model
 
     assert_equal 51864, model.n_vocab
@@ -25,7 +26,7 @@ class TestModel < TestBase
   end
 
   def test_gc
-    model = Whisper::Context.new(MODEL).model
+    model = Whisper::Context.new("base.en").model
     GC.start
 
     assert_equal 51864, model.n_vocab
@@ -41,4 +42,68 @@ class TestModel < TestBase
     assert_equal 1, model.ftype
     assert_equal "base", model.type
   end
+
+  def test_pathname
+    path = Pathname(Whisper::Model.pre_converted_models["base.en"].to_path)
+    whisper = Whisper::Context.new(path)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_auto_download
+    path = Whisper::Model.pre_converted_models["base.en"].to_path
+
+    assert_path_exist path
+    assert_equal 147964211, File.size(path)
+  end
+
+  def test_uri_string
+    path = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin"
+    whisper = Whisper::Context.new(path)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_uri
+    path = URI("https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin")
+    whisper = Whisper::Context.new(path)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
 end

bindings/ruby/tests/test_package.rb

@@ -23,7 +23,7 @@ class TestPackage < TestBase
       version = match_data[2]
       basename = "whisper.#{RbConfig::CONFIG["DLEXT"]}"
       Dir.mktmpdir do |dir|
-        system "gem", "install", "--install-dir", dir.shellescape, "pkg/#{filename.shellescape}", exception: true
+        system "gem", "install", "--install-dir", dir.shellescape, "--no-document", "pkg/#{filename.shellescape}", exception: true
         assert_path_exist File.join(dir, "gems/whispercpp-#{version}/lib", basename)
       end
     end

bindings/ruby/tests/test_params.rb

@@ -89,11 +89,11 @@ class TestParams < TestBase
     assert !@params.suppress_blank
   end
 
-  def test_suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = true
-    assert @params.suppress_non_speech_tokens
-    @params.suppress_non_speech_tokens = false
-    assert !@params.suppress_non_speech_tokens
+  def test_suppress_nst
+    @params.suppress_nst = true
+    assert @params.suppress_nst
+    @params.suppress_nst = false
+    assert !@params.suppress_nst
   end
 
   def test_token_timestamps
@@ -151,4 +151,10 @@ class TestParams < TestBase
     @params.logprob_thold = -0.5
     assert_in_delta -0.5, @params.logprob_thold
   end
+
+  def test_no_speech_thold
+    assert_in_delta 0.6, @params.no_speech_thold
+    @params.no_speech_thold = 0.2
+    assert_in_delta 0.2, @params.no_speech_thold
+  end
 end

bindings/ruby/tests/test_segment.rb

@@ -1,17 +1,6 @@
 require_relative "helper"
 
 class TestSegment < TestBase
-  class << self
-    attr_reader :whisper
-
-    def startup
-      @whisper = Whisper::Context.new(TestBase::MODEL)
-      params = Whisper::Params.new
-      params.print_timestamps = false
-      @whisper.transcribe(TestBase::AUDIO, params)
-    end
-  end
-
   def test_iteration
     whisper.each_segment do |segment|
       assert_instance_of Whisper::Segment, segment
@@ -43,6 +32,14 @@ class TestSegment < TestBase
     end
   end
 
+  def test_no_speech_prob
+    no_speech_prob = nil
+    whisper.each_segment do |segment|
+      no_speech_prob = segment.no_speech_prob
+    end
+    assert no_speech_prob > 0.0
+  end
+
   def test_on_new_segment
     params = Whisper::Params.new
     seg = nil
@@ -74,10 +71,4 @@ class TestSegment < TestBase
     end
     whisper.transcribe(AUDIO, params)
   end
-
-  private
-
-  def whisper
-    self.class.whisper
-  end
 end

bindings/ruby/tests/test_whisper.rb

@@ -11,7 +11,7 @@ class TestWhisper < TestBase
   end
 
   def test_whisper
-    @whisper = Whisper::Context.new(MODEL)
+    @whisper = Whisper::Context.new("base.en")
     params  = Whisper::Params.new
     params.print_timestamps = false
 
@@ -21,21 +21,6 @@ class TestWhisper < TestBase
   end
 
   sub_test_case "After transcription" do
-    class << self
-      attr_reader :whisper
-
-      def startup
-        @whisper = Whisper::Context.new(TestBase::MODEL)
-        params = Whisper::Params.new
-        params.print_timestamps = false
-        @whisper.transcribe(TestBase::AUDIO, params)
-      end
-    end
-
-    def whisper
-      self.class.whisper
-    end
-
     def test_full_n_segments
       assert_equal 1, whisper.full_n_segments
     end
@@ -70,6 +55,12 @@ class TestWhisper < TestBase
     def test_full_get_segment_text
       assert_match /ask not what your country can do for you, ask what you can do for your country/, whisper.full_get_segment_text(0)
     end
+
+    def test_full_get_segment_no_speech_prob
+      prob = whisper.full_get_segment_no_speech_prob(0)
+      assert prob > 0.0
+      assert prob < 1.0
+    end
   end
 
   def test_lang_max_id
@@ -104,7 +95,7 @@ class TestWhisper < TestBase
       logs << [level, buffer, udata]
     }
     Whisper.log_set log_callback, user_data
-    Whisper::Context.new(MODEL)
+    Whisper::Context.new("base.en")
 
     assert logs.length > 30
     logs.each do |log|
@@ -120,7 +111,7 @@ class TestWhisper < TestBase
     }, nil
     dev = StringIO.new("")
     $stderr = dev
-    Whisper::Context.new(MODEL)
+    Whisper::Context.new("base.en")
     assert_empty dev.string
   ensure
     $stderr = stderr
@@ -129,7 +120,7 @@ class TestWhisper < TestBase
   sub_test_case "full" do
     def setup
       super
-      @whisper = Whisper::Context.new(MODEL)
+      @whisper = Whisper::Context.new("base.en")
       @samples = File.read(AUDIO, nil, 78).unpack("s<*").collect {|i| i.to_f / 2**15}
     end
 

bindings/ruby/whispercpp.gemspec

@@ -3,12 +3,12 @@ require_relative "extsources"
 Gem::Specification.new do |s|
   s.name    = "whispercpp"
   s.authors = ["Georgi Gerganov", "Todd A. Fisher"]
-  s.version = '1.3.0'
-  s.date    = '2024-05-14'
+  s.version = '1.3.1'
+  s.date    = '2024-12-19'
   s.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
   s.email   = 'todd.fisher@gmail.com'
   s.extra_rdoc_files = ['LICENSE', 'README.md']
-  
+
   s.files = `git ls-files . -z`.split("\x0") +
               EXTSOURCES.collect {|file|
                 basename = File.basename(file)
@@ -21,16 +21,16 @@ Gem::Specification.new do |s|
 
   s.summary = %q{Ruby whisper.cpp bindings}
   s.test_files = s.files.select {|file| file.start_with? "tests/"}
-  
+
   s.extensions << 'ext/extconf.rb'
-  
+  s.required_ruby_version = '>= 3.1.0'
 
   #### Documentation and testing.
   s.homepage = 'https://github.com/ggerganov/whisper.cpp'
   s.rdoc_options = ['--main', 'README.md']
 
-  
+
     s.platform = Gem::Platform::RUBY
-  
+
   s.licenses = ['MIT']
 end

cmake/DefaultTargetOptions.cmake

@@ -13,5 +13,4 @@ set_target_properties(${TARGET}
     PROPERTIES
         EXPORT_COMPILE_COMMANDS ON
         RUNTIME_OUTPUT_DIRECTORY "${CMAKE_BINARY_DIR}/bin"
-        INSTALL_RPATH            "${CMAKE_INSTALL_PREFIX}/lib"
 )

examples/CMakeLists.txt

@@ -97,52 +97,29 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 
 if (EMSCRIPTEN)
     add_subdirectory(whisper.wasm)
-    set_target_properties(libmain PROPERTIES FOLDER "libs")
     add_subdirectory(stream.wasm)
-    set_target_properties(libstream PROPERTIES FOLDER "libs")
     add_subdirectory(command.wasm)
-    set_target_properties(libcommand PROPERTIES FOLDER "libs")
-    #add_subdirectory(talk.wasm)
-    #set_target_properties(libtalk PROPERTIES FOLDER "libs")
     add_subdirectory(bench.wasm)
-    set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
     add_subdirectory(addon.node)
-    set_target_properties(addon.node PROPERTIES FOLDER "examples")
 else()
-    add_subdirectory(main)
-    set_target_properties(main PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
-    add_subdirectory(stream)
-    set_target_properties(stream PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
-    add_subdirectory(server)
-    set_target_properties(server PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
-    add_subdirectory(command)
-    set_target_properties(command PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
+    add_subdirectory(cli)
     add_subdirectory(bench)
-    set_target_properties(bench PROPERTIES FOLDER "examples")
+    add_subdirectory(server)
     add_subdirectory(quantize)
-    set_target_properties(quantize PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
-    # TODO: disabled until update
-    #       https://github.com/ggerganov/whisper.cpp/issues/1818
-    #add_subdirectory(talk)
-    #set_target_properties(talk PROPERTIES FOLDER "examples")
-    add_subdirectory(talk-llama)
-    set_target_properties(talk-llama PROPERTIES FOLDER "examples")
-    add_subdirectory(lsp)
-    set_target_properties(lsp PROPERTIES FOLDER "examples")
-    if (GGML_SYCL)
-        add_subdirectory(sycl)
-        set_target_properties(ls-sycl-device PROPERTIES FOLDER "examples")
-    endif()
-endif (WHISPER_SDL2)
+    if (WHISPER_SDL2)
+        add_subdirectory(stream)
+        add_subdirectory(command)
+        add_subdirectory(talk-llama)
+        add_subdirectory(lsp)
+        if (GGML_SYCL)
+            add_subdirectory(sycl)
+        endif()
+    endif (WHISPER_SDL2)
+
+    add_subdirectory(deprecation-warning)
 endif()
 
 if (WHISPER_SDL2)
     add_subdirectory(wchess)
-    set_target_properties(wchess PROPERTIES FOLDER "examples")
 endif (WHISPER_SDL2)

examples/bench/CMakeLists.txt

@@ -1,6 +1,8 @@
-set(TARGET bench)
+set(TARGET whisper-bench)
 add_executable(${TARGET} bench.cpp)
 
 include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
+
+install(TARGETS ${TARGET} RUNTIME)

examples/bench/README.md

@@ -1,4 +1,4 @@
-# bench
+# whisper.cpp/examples/bench
 
 A very basic tool for benchmarking the inference performance on your device. The tool simply runs the Encoder part of
 the transformer on some random audio data and records the execution time. This way we can have an objective comparison
@@ -7,11 +7,8 @@ of the performance of the model for various setups.
 Benchmark results are tracked in the following Github issue: https://github.com/ggerganov/whisper.cpp/issues/89
 
 ```bash
-# build the bench tool
-$ make bench
-
-# run it on the small.en model using 4 threads
-$ ./bench -m ./models/ggml-small.en.bin -t 4
+# run the bench too on the small.en model using 4 threads
+$ ./build/bin/whisper-bench -m ./models/ggml-small.en.bin -t 4
 
 whisper_model_load: loading model from './models/ggml-small.en.bin'
 whisper_model_load: n_vocab       = 51864

examples/cli/CMakeLists.txt

@@ -1,6 +1,8 @@
-set(TARGET main)
-add_executable(${TARGET} main.cpp)
+set(TARGET whisper-cli)
+add_executable(${TARGET} cli.cpp)
 
 include(DefaultTargetOptions)
 
 target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+
+install(TARGETS ${TARGET} RUNTIME)

examples/cli/README.md

@@ -1,12 +1,12 @@
-# main
+# whisper.cpp/examples/cli
 
 This is the main example demonstrating most of the functionality of the Whisper model.
 It can be used as a reference for using the `whisper.cpp` library in other projects.
 
 ```
-./main -h
+./build/bin/whisper-cli -h
 
-usage: ./main [options] file0.wav file1.wav ...
+usage: ./build-pkg/bin/whisper-cli [options] file0.wav file1.wav ...
 
 options:
   -h,        --help              [default] show this help message and exit
@@ -20,9 +20,12 @@ options:
   -sow,      --split-on-word     [false  ] split on word rather than on token
   -bo N,     --best-of N         [5      ] number of best candidates to keep
   -bs N,     --beam-size N       [5      ] beam size for beam search
+  -ac N,     --audio-ctx N       [0      ] audio context size (0 - all)
   -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
   -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
   -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -tp,       --temperature N     [0.00   ] The sampling temperature, between 0 and 1
+  -tpi,      --temperature-inc N [0.20   ] The increment of temperature, between 0 and 1
   -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
   -tr,       --translate         [false  ] translate from source language to english
   -di,       --diarize           [false  ] stereo audio diarization
@@ -38,16 +41,23 @@ options:
   -oj,       --output-json       [false  ] output result in a JSON file
   -ojf,      --output-json-full  [false  ] include more information in the JSON file
   -of FNAME, --output-file FNAME [       ] output file path (without file extension)
+  -np,       --no-prints         [false  ] do not print anything other than the results
   -ps,       --print-special     [false  ] print special tokens
   -pc,       --print-colors      [false  ] print colors
   -pp,       --print-progress    [false  ] print progress
   -nt,       --no-timestamps     [false  ] do not print timestamps
   -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
   -dl,       --detect-language   [false  ] exit after automatically detecting language
-             --prompt PROMPT     [       ] initial prompt
+             --prompt PROMPT     [       ] initial prompt (max n_text_ctx/2 tokens)
   -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
   -f FNAME,  --file FNAME        [       ] input WAV file path
   -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  -dtw MODEL --dtw MODEL         [       ] compute token-level timestamps
   -ls,       --log-score         [false  ] log best decoder scores of tokens
   -ng,       --no-gpu            [false  ] disable GPU
+  -fa,       --flash-attn        [false  ] flash attention
+  --suppress-regex REGEX         [       ] regular expression matching tokens to suppress
+  --grammar GRAMMAR              [       ] GBNF grammar to guide decoding
+  --grammar-rule RULE            [       ] top-level GBNF grammar rule name
+  --grammar-penalty N            [100.0  ] scales down logits of nongrammar tokens
 ```

examples/cli/cli.cpp

@@ -43,6 +43,7 @@ struct whisper_params {
     float word_thold      =  0.01f;
     float entropy_thold   =  2.40f;
     float logprob_thold   = -1.00f;
+    float no_speech_thold =  0.6f;
     float grammar_penalty = 100.0f;
     float temperature     = 0.0f;
     float temperature_inc = 0.2f;
@@ -70,6 +71,7 @@ struct whisper_params {
     bool log_score       = false;
     bool use_gpu         = true;
     bool flash_attn      = false;
+    bool suppress_nst    = false;
 
     std::string language  = "en";
     std::string prompt;
@@ -104,6 +106,11 @@ static char * whisper_param_turn_lowercase(char * in){
     return in;
 }
 
+static char * requires_value_error(const std::string & arg) {
+    fprintf(stderr, "error: argument %s requires value\n", arg.c_str());
+    exit(0);
+}
+
 static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
     for (int i = 1; i < argc; i++) {
         std::string arg = argv[i];
@@ -122,21 +129,23 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
             whisper_print_usage(argc, argv, params);
             exit(0);
         }
-        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
-        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
-        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
-        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
-        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
-        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
-        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
-        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
-        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
-        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
-        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
-        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
+        #define ARGV_NEXT (((i + 1) < argc) ? argv[++i] : requires_value_error(arg))
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(ARGV_NEXT); }
+        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(ARGV_NEXT); }
+        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(ARGV_NEXT); }
+        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(ARGV_NEXT); }
+        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(ARGV_NEXT); }
+        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(ARGV_NEXT); }
+        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(ARGV_NEXT); }
+        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(ARGV_NEXT); }
+        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(ARGV_NEXT); }
+        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(ARGV_NEXT); }
+        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(ARGV_NEXT); }
+        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(ARGV_NEXT); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(ARGV_NEXT); }
+        else if (arg == "-nth"  || arg == "--no-speech-thold") { params.no_speech_thold = std::stof(ARGV_NEXT); }
+        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(ARGV_NEXT); }
+        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(ARGV_NEXT); }
         else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
         else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
         else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
@@ -148,30 +157,31 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
         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 == "-fp"   || arg == "--font-path")       { params.font_path       = ARGV_NEXT; }
         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 == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
-        else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(ARGV_NEXT); }
         else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
         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        = whisper_param_turn_lowercase(argv[++i]); }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = whisper_param_turn_lowercase(ARGV_NEXT); }
         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]); }
-        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
+        else if (                  arg == "--prompt")          { params.prompt          = ARGV_NEXT; }
+        else if (arg == "-m"    || arg == "--model")           { params.model           = ARGV_NEXT; }
+        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(ARGV_NEXT); }
+        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = ARGV_NEXT; }
+        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = ARGV_NEXT; }
         else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
         else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
         else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
-        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = argv[++i]; }
-        else if (                  arg == "--grammar")         { params.grammar         = argv[++i]; }
-        else if (                  arg == "--grammar-rule")    { params.grammar_rule    = argv[++i]; }
-        else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else if (arg == "-sns"  || arg == "--suppress-nst")    { params.suppress_nst    = true; }
+        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = ARGV_NEXT; }
+        else if (                  arg == "--grammar")         { params.grammar         = ARGV_NEXT; }
+        else if (                  arg == "--grammar-rule")    { params.grammar_rule    = ARGV_NEXT; }
+        else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(ARGV_NEXT); }
         else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             whisper_print_usage(argc, argv, params);
@@ -202,6 +212,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
     fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
     fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -nth N,    --no-speech-thold N [%-7.2f] no speech threshold\n",                          params.no_speech_thold);
     fprintf(stderr, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
     fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
     fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
@@ -234,6 +245,7 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
     fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
     fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
     fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -sns,      --suppress-nst      [%-7s] suppress non-speech tokens\n",                     params.suppress_nst ? "true" : "false");
     fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
     fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
     fprintf(stderr, "  --grammar-rule RULE            [%-7s] top-level GBNF grammar rule name\n",               params.grammar_rule.c_str());
@@ -1121,9 +1133,12 @@ int main(int argc, char ** argv) {
 
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
+            wparams.no_speech_thold  = params.no_speech_thold;
 
             wparams.no_timestamps    = params.no_timestamps;
 
+            wparams.suppress_nst     = params.suppress_nst;
+
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             const auto & grammar_parsed = params.grammar_parsed;

examples/command/CMakeLists.txt

@@ -1,9 +1,10 @@
 if (WHISPER_SDL2)
-    # command
-    set(TARGET command)
+    set(TARGET whisper-command)
     add_executable(${TARGET} command.cpp)
 
     include(DefaultTargetOptions)
 
     target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+
+    install(TARGETS ${TARGET} RUNTIME)
 endif ()

examples/command/README.md

@@ -1,14 +1,14 @@
-# command
+# whisper.cpp/examples/command
 
 This is a basic Voice Assistant example that accepts voice commands from the microphone.
 More info is available in [issue #171](https://github.com/ggerganov/whisper.cpp/issues/171).
 
 ```bash
 # Run with default arguments and small model
-./command -m ./models/ggml-small.en.bin -t 8
+./whisper-command -m ./models/ggml-small.en.bin -t 8
 
 # On Raspberry Pi, use tiny or base models + "-ac 768" for better performance
-./command -m ./models/ggml-tiny.en.bin -ac 768 -t 3 -c 0
+./whisper-command -m ./models/ggml-tiny.en.bin -ac 768 -t 3 -c 0
 ```
 
 https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
@@ -23,10 +23,10 @@ Initial tests show that this approach might be extremely efficient in terms of p
 
 ```bash
 # Run in guided mode, the list of allowed commands is in commands.txt
-./command -m ./models/ggml-base.en.bin -cmd ./examples/command/commands.txt
+./whisper-command -m ./models/ggml-base.en.bin -cmd ./examples/command/commands.txt
 
 # On Raspberry Pi, in guided mode you can use "-ac 128" for extra performance
-./command -m ./models/ggml-tiny.en.bin -cmd ./examples/command/commands.txt -ac 128 -t 3 -c 0
+./whisper-command -m ./models/ggml-tiny.en.bin -cmd ./examples/command/commands.txt -ac 128 -t 3 -c 0
 ```
 
 https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9b8b-aeeb76bee969.mp4
@@ -34,7 +34,7 @@ https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9
 
 ## Building
 
-The `command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+The `whisper-command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
 # Install SDL2
@@ -47,5 +47,6 @@ sudo dnf install SDL2 SDL2-devel
 # Install SDL2 on Mac OS
 brew install sdl2
 
-make command
+cmake -B build -DWHISPER_SDL2=ON
+cmake --build build --config Release
 ```

examples/common-ggml.cpp

@@ -72,9 +72,6 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_XS:
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_4:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_8:
-        case GGML_FTYPE_MOSTLY_Q4_0_8_8:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -212,12 +209,8 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_IQ4_XS:
                 case GGML_TYPE_IQ1_M:
                 case GGML_TYPE_BF16:
-                case GGML_TYPE_Q4_0_4_4:
-                case GGML_TYPE_Q4_0_4_8:
-                case GGML_TYPE_Q4_0_8_8:
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
-                case GGML_TYPE_IQ4_NL_4_4:
                 case GGML_TYPE_COUNT:
                     {
                         fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));

examples/common-sdl.cpp

@@ -1,5 +1,7 @@
 #include "common-sdl.h"
 
+#include <cstdio>
+
 audio_async::audio_async(int len_ms) {
     m_len_ms = len_ms;
 

examples/deprecation-warning/CMakeLists.txt

@@ -0,0 +1,4 @@
+add_executable(main ./deprecation-warning.cpp)
+add_executable(bench ./deprecation-warning.cpp)
+add_executable(stream ./deprecation-warning.cpp)
+add_executable(command ./deprecation-warning.cpp)

examples/deprecation-warning/README.md

@@ -0,0 +1,17 @@
+# Migration notice for binary filenames
+
+> [!IMPORTANT]
+[2024 Dec 20] Binaries have been renamed w/ a `whisper-` prefix. `main` is now `whisper-cli`, `server` is `whisper-server`, etc (https://github.com/ggerganov/whisper.cpp/pull/2648)
+
+This migration was important, but it is a breaking change that may not always be immediately obvious to users.
+
+Please update all scripts and workflows to use the new binary names.
+
+| Old Filename | New Filename |
+| ---- | ---- |
+| main | whisper-cli |
+| bench | whisper-bench |
+| stream | whisper-stream |
+| command | whisper-command |
+| server | whisper-server |
+| talk-llama | whisper-talk-llama |

examples/deprecation-warning/deprecation-warning.cpp

@@ -0,0 +1,38 @@
+// Warns users that this filename was deprecated, and provides a link for more information.
+
+#include <cstdio>
+#include <string>
+
+// Main
+int main(int argc, char** argv) {
+    std::string filename = "main";
+    if (argc >= 1) {
+        filename = argv[0];
+    }
+
+    // Get only the program name from the full path
+    size_t pos = filename.find_last_of("/\\");
+    if (pos != std::string::npos) {
+        filename = filename.substr(pos+1);
+    }
+
+    // Append "whisper-" to the beginning of filename to get the replacemnt filename
+    std::string replacement_filename = "whisper-" + filename;
+
+    // The exception is if the filename is "main", then our replacement filename is "whisper-cli"
+    if (filename == "main") {
+        replacement_filename = "whisper-cli";
+    }
+
+    if (filename == "main.exe") {
+        replacement_filename = "whisper-cli.exe";
+    }
+
+    fprintf(stdout, "\n");
+    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
+    fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str());
+    fprintf(stdout, " See https://github.com/ggerganov/whisper.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
+    fprintf(stdout, "\n");
+
+    return EXIT_FAILURE;
+}

examples/generate-karaoke.sh

@@ -11,7 +11,7 @@
 # Press Ctrl+C to stop recording
 #
 
-executable="./main"
+executable="./build/bin/whisper-cli"
 model="base.en"
 model_path="models/ggml-$model.bin"
 
@@ -46,7 +46,7 @@ ffmpeg -y -i ./rec.wav -ar 16000 -ac 1 -c:a pcm_s16le ./rec16.wav > /dev/null 2>
 
 # run Whisper
 echo "Processing ..."
-./main -m models/ggml-base.en.bin rec16.wav -owts > /dev/null 2>&1
+${executable} -m models/ggml-base.en.bin rec16.wav -owts > /dev/null 2>&1
 
 # generate Karaoke video
 echo "Generating video ..."

examples/livestream.sh

@@ -14,7 +14,7 @@ model="base.en"
 
 check_requirements()
 {
-    if ! command -v ./main &>/dev/null; then
+    if ! command -v ./build/bin/whisper-cli &>/dev/null; then
         echo "whisper.cpp main executable is required (make)"
         exit 1
     fi
@@ -100,7 +100,7 @@ while [ $running -eq 1 ]; do
         err=$(cat /tmp/whisper-live.err | wc -l)
     done
 
-    ./main -t 8 -m ./models/ggml-${model}.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
+    ./build/bin/whisper-cli -t 8 -m ./models/ggml-${model}.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
 
     while [ $SECONDS -lt $((($i+1)*$step_s)) ]; do
         sleep 1
@@ -109,4 +109,4 @@ while [ $running -eq 1 ]; do
 done
 
 killall -v ffmpeg
-killall -v main
+killall -v whisper-cli

examples/lsp/lsp.cpp

@@ -181,7 +181,7 @@ static json unguided_transcription(struct whisper_context * ctx, audio_async &au
     wparams.n_threads        = params.n_threads;
 
     wparams.audio_ctx        = params.audio_ctx;
-    wparams.suppress_non_speech_tokens = true;
+    wparams.suppress_nst     = true;
     // run the transformer and a single decoding pass
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
         fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
@@ -225,7 +225,7 @@ static json guided_transcription(struct whisper_context * ctx, audio_async &audi
     wparams.prompt_tokens    = cs.prompt_tokens.data();
     wparams.prompt_n_tokens  = cs.prompt_tokens.size();
     // TODO: properly expose as option
-    wparams.suppress_non_speech_tokens = true;
+    wparams.suppress_nst     = true;
 
     // run the transformer and a single decoding pass
     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {

examples/server/CMakeLists.txt

@@ -1,4 +1,4 @@
-set(TARGET server)
+set(TARGET whisper-server)
 add_executable(${TARGET} server.cpp httplib.h)
 
 include(DefaultTargetOptions)
@@ -8,3 +8,5 @@ target_link_libraries(${TARGET} PRIVATE common json_cpp whisper ${CMAKE_THREAD_L
 if (WIN32)
     target_link_libraries(${TARGET} PRIVATE ws2_32)
 endif()
+
+install(TARGETS ${TARGET} RUNTIME)

examples/server/README.md

@@ -1,4 +1,4 @@
-# whisper.cpp http server
+# whisper.cpp/examples/server
 
 Simple http server. WAV Files are passed to the inference model via http requests.
 
@@ -7,9 +7,9 @@ https://github.com/ggerganov/whisper.cpp/assets/1991296/e983ee53-8741-4eb5-9048-
 ## Usage
 
 ```
-./server -h
+./build/bin/whisper-server -h
 
-usage: ./bin/server [options]
+usage: ./build/bin/whisper-server [options]
 
 options:
   -h,        --help              [default] show this help message and exit

examples/server/server.cpp

@@ -61,6 +61,7 @@ struct whisper_params {
     float logprob_thold   = -1.00f;
     float temperature     =  0.00f;
     float temperature_inc =  0.20f;
+    float no_speech_thold = 0.6f;
 
     bool debug_mode      = false;
     bool translate       = false;
@@ -76,6 +77,7 @@ struct whisper_params {
     bool no_timestamps   = false;
     bool use_gpu         = true;
     bool flash_attn      = false;
+    bool suppress_nst    = false;
 
     std::string language        = "en";
     std::string prompt          = "";
@@ -134,7 +136,9 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
     fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
     fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
     fprintf(stderr, "  --inference-path PATH,         [%-7s] Inference path for all requests\n", sparams.inference_path.c_str());
-    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
+    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server\n", sparams.ffmpeg_converter ? "true" : "false");
+    fprintf(stderr, "  -sns,      --suppress-nst      [%-7s] suppress non-speech tokens\n", params.suppress_nst ? "true" : "false");
+    fprintf(stderr, "  -nth N,    --no-speech-thold N [%-7.2f] no speech threshold\n",   params.no_speech_thold);
     fprintf(stderr, "\n");
 }
 
@@ -179,6 +183,9 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
         else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
         else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
         else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        else if (arg == "-sns"  || arg == "--suppress-nst")    { params.suppress_nst    = true; }
+        else if (arg == "-nth"  || arg == "--no-speech-thold") { params.no_speech_thold = std::stof(argv[++i]); }
+
         // server params
         else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
         else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
@@ -472,6 +479,14 @@ void get_req_parameters(const Request & req, whisper_params & params)
     {
         params.temperature_inc = std::stof(req.get_file_value("temperature_inc").content);
     }
+    if (req.has_file("suppress_non_speech"))
+    {
+        params.suppress_nst = parse_str_to_bool(req.get_file_value("suppress_non_speech").content);
+    }
+    if (req.has_file("suppress_nst"))
+    {
+        params.suppress_nst = parse_str_to_bool(req.get_file_value("suppress_nst").content);
+    }
 }
 
 }  // namespace
@@ -677,7 +692,8 @@ int main(int argc, char ** argv) {
         if (sparams.ffmpeg_converter) {
             // if file is not wav, convert to wav
             // write to temporary file
-            const std::string temp_filename_base = std::tmpnam(nullptr);
+            //const std::string temp_filename_base = std::tmpnam(nullptr);
+            const std::string temp_filename_base = "whisper-server-tmp"; // TODO: this is a hack, remove when the mutext is removed
             const std::string temp_filename = temp_filename_base + ".wav";
             std::ofstream temp_file{temp_filename, std::ios::binary};
             temp_file << audio_file.content;
@@ -711,7 +727,6 @@ int main(int argc, char ** argv) {
             }
         }
 
-
         printf("Successfully loaded %s\n", filename.c_str());
 
         // print system information
@@ -779,6 +794,7 @@ int main(int argc, char ** argv) {
             wparams.beam_search.beam_size = params.beam_size;
 
             wparams.temperature      = params.temperature;
+            wparams.no_speech_thold = params.no_speech_thold;
             wparams.temperature_inc  = params.temperature_inc;
             wparams.entropy_thold    = params.entropy_thold;
             wparams.logprob_thold    = params.logprob_thold;
@@ -786,6 +802,8 @@ int main(int argc, char ** argv) {
             wparams.no_timestamps    = params.no_timestamps;
             wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
 
+            wparams.suppress_nst     = params.suppress_nst;
+
             whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
 
             // this callback is called on each new segment
@@ -929,7 +947,7 @@ int main(int argc, char ** argv) {
 
                 // TODO compression_ratio and no_speech_prob are not implemented yet
                 // segment["compression_ratio"] = 0;
-                // segment["no_speech_prob"] = 0;
+                segment["no_speech_prob"] = whisper_full_get_segment_no_speech_prob(ctx, i);
 
                 jres["segments"].push_back(segment);
             }

examples/stream/CMakeLists.txt

@@ -1,9 +1,10 @@
 if (WHISPER_SDL2)
-    # stream
-    set(TARGET stream)
+    set(TARGET whisper-stream)
     add_executable(${TARGET} stream.cpp)
 
     include(DefaultTargetOptions)
 
     target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+
+    install(TARGETS ${TARGET} RUNTIME)
 endif ()

examples/stream/README.md

@@ -1,11 +1,11 @@
-# stream
+# whisper.cpp/examples/stream
 
 This is a naive example of performing real-time inference on audio from your microphone.
-The `stream` tool samples the audio every half a second and runs the transcription continously.
+The `whisper-stream` tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
@@ -15,7 +15,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Setting the `--step` argument to `0` enables the sliding window mode:
 
 ```bash
- ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
+ ./build/bin/whisper-stream -m ./models/ggml-base.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
 In this mode, the tool will transcribe only after some speech activity is detected. A very
@@ -27,7 +27,7 @@ a transcription block that is suitable for parsing.
 
 ## Building
 
-The `stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+The `whisper-stream` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
 # Install SDL2
@@ -40,21 +40,10 @@ sudo dnf install SDL2 SDL2-devel
 # Install SDL2 on Mac OS
 brew install sdl2
 
-make stream
-```
+cmake -B build -DWHISPER_SDL2=ON
+cmake --build build --config Release
 
-Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
-`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
-
-```bash
-whisper.cpp/examples/stream$ make stream
-g++     stream.cpp   -o stream
-stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
-    6 | #include "common/sdl.h"
-      |          ^~~~~~~~~~~~~~
-compilation terminated.
-make: *** [<builtin>: stream] Error 1
+./build/bin/whisper-stream
 ```
 
 ## Web version

examples/talk-llama/CMakeLists.txt

@@ -1,6 +1,5 @@
 if (WHISPER_SDL2)
-    # talk-llama
-    set(TARGET talk-llama)
+    set(TARGET whisper-talk-llama)
     add_executable(${TARGET} talk-llama.cpp
         llama.cpp
         llama-vocab.cpp

examples/talk-llama/README.md

@@ -1,4 +1,4 @@
-# talk-llama
+# whisper.cpp/examples/talk-llama
 
 Talk with an LLaMA AI in your terminal
 
@@ -12,7 +12,7 @@ https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-
 
 ## Building
 
-The `talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+The `whisper-talk-llama` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
 
 ```bash
 # Install SDL2
@@ -25,11 +25,12 @@ sudo dnf install SDL2 SDL2-devel
 # Install SDL2 on Mac OS
 brew install sdl2
 
-# Build the "talk-llama" executable
-make talk-llama
+# Build the "whisper-talk-llama" executable
+cmake -B build -S . -DWHISPER_SDL2=ON
+cmake --build build --config Release
 
 # Run it
-./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
+./build/bin/whisper-talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -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
@@ -37,16 +38,16 @@ make talk-llama
 
 ## 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.
+The `whisper-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.
+To enable session support, use the `--session FILE` command line option when running the program. The `whisper-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/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
+./build/bin/whisper-talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
 ```
 
 ## TTS

examples/talk-llama/llama-sampling.cpp

@@ -1396,19 +1396,15 @@ struct llama_sampler * llama_sampler_init_grammar_impl(const struct llama_vocab
 // penalties
 
 struct llama_sampler_penalties {
-    const int32_t     n_vocab;
-    const llama_token special_eos_id;
-    const llama_token linefeed_id;
-
     const int32_t penalty_last_n;
     const float   penalty_repeat;
     const float   penalty_freq;
     const float   penalty_present;
 
-    const bool    penalize_nl;
-    const bool    ignore_eos;
-
     ring_buffer<llama_token> prev;
+
+    // a frequency map to count token occurrences
+    std::unordered_map<llama_token, int> token_count;
 };
 
 static const char * llama_sampler_penalties_name(const struct llama_sampler * /*smpl*/) {
@@ -1421,76 +1417,50 @@ static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_to
         return;
     }
 
+    ctx->token_count[token]++;
+
+    // if the ring buffer is full, remove the oldest token
+    if (ctx->prev.size() >= (size_t) ctx->penalty_last_n) {
+        const auto old = ctx->prev.front();
+
+        ctx->token_count[old]--;
+        if (ctx->token_count[old] == 0) {
+            ctx->token_count.erase(old);
+        }
+    }
+
     ctx->prev.push_back(token);
+
+#if 0
+    // sanity check
+    std::unordered_map<llama_token, int> tmp;
+    for (int i = 0; i < std::min<int>(ctx->penalty_last_n, ctx->prev.size()); ++i) {
+        tmp[ctx->prev.rat(i)]++;
+    }
+
+    assert(ctx->token_count == tmp);
+#endif
 }
 
 static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_penalties *) smpl->ctx;
 
-    if (ctx->ignore_eos) {
-        assert(ctx->special_eos_id >= 0);
-
-        // optimistically check if the candidates are not yet sorted/shuffled/truncated
-        if (cur_p->size > (size_t) ctx->special_eos_id && cur_p->data[ctx->special_eos_id].id == ctx->special_eos_id) {
-            cur_p->data[ctx->special_eos_id].logit = -INFINITY;
-        } else {
-            // else, search for the special EOS token
-            for (size_t i = 0; i < cur_p->size; ++i) {
-                if (cur_p->data[i].id == ctx->special_eos_id) {
-                    cur_p->data[i].logit = -INFINITY;
-                    break;
-                }
-            }
-        }
-    }
-
     if ((ctx->penalty_last_n == 0) ||
         (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) {
         return;
     }
 
-    bool nl_found = false;
-    size_t nl_idx = 0;
-    float nl_logit = -INFINITY;
-    if (!ctx->penalize_nl) {
-        assert(ctx->linefeed_id >= 0);
-
-        // optimistically check if the candidates are not yet sorted/shuffled/truncated
-        if (cur_p->size > (size_t) ctx->linefeed_id && cur_p->data[ctx->linefeed_id].id == ctx->linefeed_id) {
-            nl_found = true;
-            nl_idx = ctx->linefeed_id;
-            nl_logit = cur_p->data[ctx->linefeed_id].logit;
-        } else {
-            // else, search for the linefeed token
-            for (size_t i = 0; i < cur_p->size; ++i) {
-                if (cur_p->data[i].id == ctx->linefeed_id) {
-                    nl_found = true;
-                    nl_idx = i;
-                    nl_logit = cur_p->data[i].logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    // Create a frequency map to count occurrences of each token in last_tokens
-    // TODO: optimize this by maintaining the token count in the sampler context
-    using llama_token_cnt = std::unordered_map<llama_token, int>;
-    llama_token_cnt token_count;
-
-    for (int i = 0; i < std::min<int>(ctx->penalty_last_n, ctx->prev.size()); ++i) {
-        token_count[ctx->prev.rat(i)]++;
-    }
-
     // Apply frequency and presence penalties to the cur_p
     for (size_t i = 0; i < cur_p->size; ++i) {
-        const auto token_iter = token_count.find(cur_p->data[i].id);
-        if (token_iter == token_count.end()) {
+        const auto token_iter = ctx->token_count.find(cur_p->data[i].id);
+        if (token_iter == ctx->token_count.end()) {
             continue;
         }
 
         const int count = token_iter->second;
 
+        assert(count > 0 && count <= ctx->penalty_last_n);
+
         // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
         // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
         if (cur_p->data[i].logit <= 0) {
@@ -1503,30 +1473,21 @@ static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_tok
     }
 
     cur_p->sorted = false;
-
-    if (!ctx->penalize_nl && nl_found) {
-        // restore the logit of the newline token if it was penalized
-        cur_p->data[nl_idx].logit = nl_logit;
-    }
 }
 
 static void llama_sampler_penalties_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_penalties *) smpl->ctx;
     ctx->prev.clear();
+    ctx->token_count.clear();
 }
 
 static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_penalties *) smpl->ctx;
     auto * result = llama_sampler_init_penalties(
-            ctx->n_vocab,
-            ctx->special_eos_id,
-            ctx->linefeed_id,
             ctx->penalty_last_n,
             ctx->penalty_repeat,
             ctx->penalty_freq,
-            ctx->penalty_present,
-            ctx->penalize_nl,
-            ctx->ignore_eos);
+            ctx->penalty_present);
 
     // copy the state
     {
@@ -1552,38 +1513,21 @@ static struct llama_sampler_i llama_sampler_penalties_i = {
 };
 
 struct llama_sampler * llama_sampler_init_penalties(
-        int32_t n_vocab,
-        llama_token special_eos_id,
-        llama_token linefeed_id,
         int32_t penalty_last_n,
         float penalty_repeat,
         float penalty_freq,
-        float penalty_present,
-        bool penalize_nl,
-        bool ignore_eos) {
-    if (linefeed_id == LLAMA_TOKEN_NULL) {
-        penalize_nl = true;
-    }
-
-    if (special_eos_id == LLAMA_TOKEN_NULL) {
-        ignore_eos = false;
-    }
-
+        float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
     return new llama_sampler {
         /* .iface = */ &llama_sampler_penalties_i,
         /* .ctx   = */ new llama_sampler_penalties {
-            /* .n_vocab         = */ n_vocab,
-            /* .special_eos_id  = */ special_eos_id,
-            /* .linefeed_id     = */ linefeed_id,
             /* .penalty_last_n  = */ penalty_last_n,
             /* .penalty_repeat  = */ penalty_repeat,
             /* .penalty_freq    = */ penalty_freq,
             /* .penalty_present = */ penalty_present,
-            /* .penalize_nl     = */ penalize_nl,
-            /* .ignore_eos      = */ ignore_eos,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
+            /* .token_count     = */ {},
         },
     };
 }
@@ -1611,7 +1555,8 @@ static void get_overlapping_token_sequences(const llama_vocab & vocab, const std
         if (word.find(str) != std::string::npos) {
             token_sequences.emplace(token_id, std::vector<llama_token>());
         } else {
-            size_t word_len = word.size(), str_len = str.size();
+            size_t word_len = word.size();
+            size_t str_len = str.size();
             size_t pos = -1;
             while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
                 bool match = true;

examples/talk-llama/llama-vocab.cpp

@@ -418,6 +418,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
             case LLAMA_VOCAB_PRE_TYPE_EXAONE:
+            case LLAMA_VOCAB_PRE_TYPE_MINERVA:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -737,7 +738,7 @@ struct llm_tokenizer_wpm_session {
         std::vector<std::string> words(1, "");
 
         for (const uint32_t cpt : cpts_nfd) {
-            const auto flags = unicode_cpt_flags(cpt);
+            const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
                 if (words.back().size()) {  // finish previous word if any

examples/talk-llama/llama.cpp

@@ -163,6 +163,7 @@ enum llm_arch {
     LLM_ARCH_QWEN,
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
+    LLM_ARCH_QWEN2VL,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PLAMO,
@@ -183,6 +184,7 @@ enum llm_arch {
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
+    LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_BITNET,
@@ -217,6 +219,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN,            "qwen"         },
     { LLM_ARCH_QWEN2,           "qwen2"        },
     { LLM_ARCH_QWEN2MOE,        "qwen2moe"     },
+    { LLM_ARCH_QWEN2VL,         "qwen2vl"      },
     { LLM_ARCH_PHI2,            "phi2"         },
     { LLM_ARCH_PHI3,            "phi3"         },
     { LLM_ARCH_PLAMO,           "plamo"        },
@@ -237,6 +240,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
+    { LLM_ARCH_DEEPSEEK,        "deepseek"     },
     { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
     { LLM_ARCH_CHATGLM,         "chatglm"      },
     { LLM_ARCH_BITNET,          "bitnet"       },
@@ -308,6 +312,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,
@@ -424,6 +429,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,             "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_SECTIONS,          "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                   "%s.rope.freq_base"                       },
     { LLM_KV_ROPE_SCALE_LINEAR,                "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,                "%s.rope.scaling.type"                    },
@@ -898,6 +904,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_QWEN2VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_QWEN2MOE,
         {
@@ -1288,6 +1311,33 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_DEEPSEEK,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FREQS,         "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,      "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_DEEPSEEK2,
         {
@@ -1562,6 +1612,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
     LLM_CHAT_TEMPLATE_MISTRAL_V7,
     LLM_CHAT_TEMPLATE_PHI_3,
+    LLM_CHAT_TEMPLATE_FALCON_3,
     LLM_CHAT_TEMPLATE_ZEPHYR,
     LLM_CHAT_TEMPLATE_MONARCH,
     LLM_CHAT_TEMPLATE_GEMMA,
@@ -1579,6 +1630,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
+    LLM_CHAT_TEMPLATE_GIGACHAT,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 
@@ -1593,6 +1645,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
     { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
     { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
+    { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
     { "zephyr",            LLM_CHAT_TEMPLATE_ZEPHYR            },
     { "monarch",           LLM_CHAT_TEMPLATE_MONARCH           },
     { "gemma",             LLM_CHAT_TEMPLATE_GEMMA             },
@@ -1610,6 +1663,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
+    { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
 };
 
 static llm_arch llm_arch_from_string(const std::string & name) {
@@ -1794,7 +1848,7 @@ private:
         DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
                                     NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
         if (!bufLen) {
-            ret = format("Win32 error code: %s", error_code);
+            ret = format("Win32 error code: %lx", error_code);
         } else {
             ret = lpMsgBuf;
             LocalFree(lpMsgBuf);
@@ -2132,7 +2186,7 @@ struct llama_mmap {
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
 
             // may fail on pre-Windows 8 systems
-            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+            pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
 
             if (pPrefetchVirtualMemory) {
                 // advise the kernel to preload the mapped memory
@@ -2474,11 +2528,12 @@ struct llama_hparams {
     uint32_t time_decay_extra_dim = 0;
     uint32_t wkv_head_size = 0;
 
-    float    rope_attn_factor = 1.0f;
-    float    rope_freq_base_train;
-    float    rope_freq_scale_train;
-    uint32_t n_ctx_orig_yarn;
-    float    rope_yarn_log_mul;
+    float     rope_attn_factor = 1.0f;
+    float     rope_freq_base_train;
+    float     rope_freq_scale_train;
+    uint32_t  n_ctx_orig_yarn;
+    float     rope_yarn_log_mul;
+    int       rope_sections[4];
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
@@ -2535,6 +2590,9 @@ struct llama_hparams {
 
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
         if (this->n_ctx_orig_yarn != other.n_ctx_orig_yarn) return true;
+        if (std::equal(std::begin(this->rope_sections),
+                       std::end(this->rope_sections),
+                       std::begin(other.rope_sections)))    return true;
 
         if (this->ssm_d_conv  != other.ssm_d_conv)  return true;
         if (this->ssm_d_inner != other.ssm_d_inner) return true;
@@ -3378,6 +3436,11 @@ struct llama_context {
     // whether we are computing encoder output or decoder output
     bool is_encoding = false;
 
+    // TODO: find a better way to accommodate mutli-dimension position encoding methods
+    // number of position id each token get, 1 for each token in most cases.
+    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
+    int n_pos_per_token = 1;
+
     // output of the encoder part of the encoder-decoder models
     std::vector<float> embd_enc;
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
@@ -4578,9 +4641,6 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 case GGML_TYPE_IQ4_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS;  break;
                 case GGML_TYPE_IQ3_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ3_S;   break;
-                case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
-                case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
-                case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -5344,9 +5404,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ4_XS:   return "IQ4_XS - 4.25 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_S:    return "IQ3_S - 3.4375 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_M:    return "IQ3_S mix - 3.66 bpw";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
 
         default: return "unknown, may not work";
     }
@@ -5753,6 +5810,13 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_QWEN2VL:
+            {
+                std::array<int, 4> section_dims;
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, section_dims, 4, true);
+                std::copy(section_dims.begin(), section_dims.begin() + 4, std::begin(hparams.rope_sections));
+            }
+            // fall through
         case LLM_ARCH_QWEN2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -6063,6 +6127,19 @@ static void llm_load_hparams(
                     model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_DEEPSEEK:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
+
+                switch (hparams.n_layer) {
+                    case 28: model.type = e_model::MODEL_20B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DEEPSEEK2:
             {
                 bool is_lite = (hparams.n_layer == 27);
@@ -6398,6 +6475,11 @@ static void llm_load_vocab(
             } else if (
                     tokenizer_pre == "falcon") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_FALCON;
+            } else if (
+                    tokenizer_pre == "falcon3") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
+                vocab.tokenizer_ignore_merges = true;
+                vocab.tokenizer_add_bos = true;
             } else if (
                     tokenizer_pre == "mpt") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MPT;
@@ -6409,6 +6491,7 @@ static void llm_load_vocab(
                     tokenizer_pre == "phi-2"   ||
                     tokenizer_pre == "jina-es" ||
                     tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
@@ -6479,6 +6562,9 @@ static void llm_load_vocab(
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
                 vocab.tokenizer_add_bos = true;
                 vocab.tokenizer_clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minerva-7b") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MINERVA;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -7057,6 +7143,13 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
 
     LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);
 
+    if (model.arch == LLM_ARCH_DEEPSEEK) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+    }
+
     if (model.arch == LLM_ARCH_DEEPSEEK2) {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
@@ -8170,6 +8263,7 @@ static bool llm_load_tensors(
                     }
                 } break;
             case LLM_ARCH_QWEN2:
+            case LLM_ARCH_QWEN2VL:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -8830,6 +8924,55 @@ static bool llm_load_tensors(
                         layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
                     }
                 } break;
+            case LLM_ARCH_DEEPSEEK:
+                {
+
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    model.output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        } else {
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+                            if (n_expert == 0) {
+                                throw std::runtime_error("n_expert must be > 0");
+                            }
+                            if (n_expert_used == 0) {
+                                throw std::runtime_error("n_expert_used must be > 0");
+                            }
+
+                            // MoE branch
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                        }
+                    }
+                } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
                     const bool is_lite = (hparams.n_layer == 27);
@@ -12559,6 +12702,124 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_qwen2vl() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 4);
+        cb(lctx.inp_pos, "inp_pos", -1);
+        ggml_set_input(lctx.inp_pos);
+        struct ggml_tensor * inp_pos = lctx.inp_pos;
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+        int sections[4];
+        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_multi(
+                    ctx0,
+                    ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_multi(
+                    ctx0,
+                    ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_qwen2moe() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -15066,6 +15327,161 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_deepseek() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            if ((uint32_t) il < hparams.n_layer_dense_lead) {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                ggml_tensor * moe_out =
+                        llm_build_moe_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_gate_inp,
+                            model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps,
+                            model.layers[il].ffn_down_exps,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, false,
+                            false, hparams.expert_weights_scale,
+                            cb, il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // FFN shared expert
+                {
+                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_up_shexp,   NULL, NULL,
+                            model.layers[il].ffn_gate_shexp, NULL, NULL,
+                            model.layers[il].ffn_down_shexp, NULL, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                }
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_deepseek2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -16660,6 +17076,11 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_qwen2();
             } break;
+        case LLM_ARCH_QWEN2VL:
+            {
+                lctx.n_pos_per_token = 4;
+                result = llm.build_qwen2vl();
+            } break;
         case LLM_ARCH_QWEN2MOE:
             {
                 result = llm.build_qwen2moe();
@@ -16748,6 +17169,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_arctic();
             } break;
+        case LLM_ARCH_DEEPSEEK:
+            {
+                result = llm.build_deepseek();
+            } break;
         case LLM_ARCH_DEEPSEEK2:
             {
                 result = llm.build_deepseek2();
@@ -16878,8 +17303,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch)
 
     if (ubatch.pos && lctx.inp_pos) {
         const int64_t n_tokens = ubatch.n_tokens;
-
-        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+        auto n_pos = lctx.n_pos_per_token;
+        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*n_pos*ggml_element_size(lctx.inp_pos));
     }
 
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
@@ -18364,10 +18789,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
                 new_type = GGML_TYPE_IQ3_S;
             }
-            else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
-                     new_type == GGML_TYPE_Q4_0_8_8) {
-                new_type = GGML_TYPE_Q4_0;
-            }
             else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
                 new_type = GGML_TYPE_Q4_K;
             }
@@ -18690,9 +19111,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ4_XS:  default_type = GGML_TYPE_IQ4_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ3_S:   default_type = GGML_TYPE_IQ3_S;   break;
         case LLAMA_FTYPE_MOSTLY_IQ3_M:   default_type = GGML_TYPE_IQ3_S;   break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -19031,14 +19449,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 f32_data = (float *) f32_conv_buf.data();
             }
 
-            int chunk_size_multiplier = 1;
-            if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
-                if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
-                else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
-                if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
-                else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
-            }
-
             LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
@@ -19051,8 +19461,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             const int64_t nrows = tensor->ne[1];
 
             static const int64_t min_chunk_size = 32 * 512;
-            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
-                                       chunk_size_multiplier;
+            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
 
             const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
             const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
@@ -19995,6 +20404,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_COMMAND_R:
         case LLM_ARCH_OLMO:
         case LLM_ARCH_ARCTIC:
+        case LLM_ARCH_DEEPSEEK:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
@@ -20028,6 +20438,9 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_MINICPM3:
             return LLAMA_ROPE_TYPE_NEOX;
 
+        case LLM_ARCH_QWEN2VL:
+            return LLAMA_ROPE_TYPE_MROPE;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");
@@ -21596,7 +22009,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
                 throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
             }
         } else if ((size_t) i >= ctx->output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
+            throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size()));
         } else {
             j = ctx->output_ids[i];
         }
@@ -21813,6 +22226,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
         }
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        return LLM_CHAT_TEMPLATE_FALCON_3;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -21857,6 +22272,8 @@ static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_RWKV_WORLD;
     } else if (tmpl_contains("<|start_of_role|>")) {
         return LLM_CHAT_TEMPLATE_GRANITE;
+    } else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
+        return LLM_CHAT_TEMPLATE_GIGACHAT;
     }
     return LLM_CHAT_TEMPLATE_UNKNOWN;
 }
@@ -21963,6 +22380,15 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {
+        // Falcon 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) {
         // zephyr template
         for (auto message : chat) {
@@ -22180,6 +22606,32 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_of_role|>assistant<|end_of_role|>\n";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
+        // GigaChat template
+        bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
+
+        // Handle system message if present
+        if (has_system) {
+            ss << "<s>" << chat[0]->content << "<|message_sep|>";
+        } else {
+            ss << "<s>";
+        }
+
+        // Process remaining messages
+        for (size_t i = has_system ? 1 : 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << "user<|role_sep|>" << chat[i]->content << "<|message_sep|>"
+                << "available functions<|role_sep|>[]<|message_sep|>";
+            } else if (role == "assistant") {
+                ss << "assistant<|role_sep|>" << chat[i]->content << "<|message_sep|>";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << "assistant<|role_sep|>";
+        }
     } else {
         // template not supported
         return -1;

examples/talk-llama/llama.h

@@ -104,12 +104,15 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
         LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
+        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
     };
 
     enum llama_rope_type {
-        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM = 0,
-        LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_NONE   = -1,
+        LLAMA_ROPE_TYPE_NORM   = 0,
+        LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
     enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -171,9 +174,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // except 1d tensors
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // removed from gguf files, use Q4_0 and runtime repack
         LLAMA_FTYPE_MOSTLY_TQ1_0         = 36, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_TQ2_0         = 37, // except 1d tensors
 
@@ -455,6 +458,7 @@ extern "C" {
     // Functions to access the model's GGUF metadata scalar values
     // - The functions return the length of the string on success, or -1 on failure
     // - The output string is always null-terminated and cleared on failure
+    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
     // - GGUF array values are not supported by these functions
 
     // Get metadata value as a string by key name
@@ -1135,16 +1139,12 @@ extern "C" {
                           const char * grammar_str,
                           const char * grammar_root);
 
+    /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
-                             int32_t   n_vocab,         // llama_n_vocab()
-                         llama_token   special_eos_id,  // llama_token_eos()
-                         llama_token   linefeed_id,     // llama_token_nl()
-                             int32_t   penalty_last_n,  // last n tokens to penalize (0 = disable penalty, -1 = context size)
-                               float   penalty_repeat,  // 1.0 = disabled
-                               float   penalty_freq,    // 0.0 = disabled
-                               float   penalty_present, // 0.0 = disabled
-                                bool   penalize_nl,     // consider newlines as a repeatable token
-                                bool   ignore_eos);     // ignore the end-of-sequence token
+                             int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
+                               float   penalty_repeat,   // 1.0 = disabled
+                               float   penalty_freq,     // 0.0 = disabled
+                               float   penalty_present); // 0.0 = disabled
 
     ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
     LLAMA_API struct llama_sampler *    llama_sampler_init_dry(

examples/talk-llama/unicode.cpp

@@ -71,15 +71,15 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
     throw std::invalid_argument("failed to convert utf8 to codepoint");
 }
 
-//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cp) {
+//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cpt) {
 //    std::vector<uint16_t> result;
-//    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
-//        result.emplace_back(cp);
+//    if (/* 0x0000 <= cpt && */ cpt <= 0xffff) {
+//        result.emplace_back(cpt);
 //        return result;
 //    }
-//    if (0x10000 <= cp && cp <= 0x10ffff) {
-//        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
-//        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+//    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+//        result.emplace_back(0xd800 | ((cpt - 0x10000) >> 10));
+//        result.emplace_back(0xdc00 | ((cpt - 0x10000) & 0x03ff));
 //        return result;
 //    }
 //    throw std::invalid_argument("failed to convert codepoint to utf16");
@@ -120,8 +120,8 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
 //    return result;
 //}
 
-static std::vector<codepoint_flags> unicode_cpt_flags_array() {
-    std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+static std::vector<unicode_cpt_flags> unicode_cpt_flags_array() {
+    std::vector<unicode_cpt_flags> cpt_flags(MAX_CODEPOINTS, unicode_cpt_flags::UNDEFINED);
 
     assert (unicode_ranges_flags.begin()[0].first == 0);
     assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS);
@@ -253,8 +253,8 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -371,8 +371,8 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -572,29 +572,29 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
 // interface
 //
 
-std::string unicode_cpt_to_utf8(uint32_t cp) {
+std::string unicode_cpt_to_utf8(uint32_t cpt) {
     std::string result;
 
-    if (/* 0x00 <= cp && */ cp <= 0x7f) {
-        result.push_back(cp);
+    if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
+        result.push_back(cpt);
         return result;
     }
-    if (0x80 <= cp && cp <= 0x7ff) {
-        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x80 <= cpt && cpt <= 0x7ff) {
+        result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x800 <= cp && cp <= 0xffff) {
-        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x800 <= cpt && cpt <= 0xffff) {
+        result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x10000 <= cp && cp <= 0x10ffff) {
-        result.push_back(0xf0 | ((cp >> 18) & 0x07));
-        result.push_back(0x80 | ((cp >> 12) & 0x3f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+        result.push_back(0xf0 | ((cpt >> 18) & 0x07));
+        result.push_back(0x80 | ((cpt >> 12) & 0x3f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
 
@@ -624,19 +624,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     return result;
 }
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     static const auto cpt_flags = unicode_cpt_flags_array();
-    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
+    return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;
 }
 
-codepoint_flags unicode_cpt_flags(const std::string & utf8) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     if (utf8.empty()) {
         return undef;  // undefined
     }
     size_t offset = 0;
-    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
+    return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));
 }
 
 std::string unicode_byte_to_utf8(uint8_t byte) {
@@ -649,41 +649,41 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) {
     return map.at(utf8);
 }
 
-uint32_t unicode_tolower(uint32_t cp) {
+uint32_t unicode_tolower(uint32_t cpt) {
     // binary search
-    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cp,
+    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,
         [](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {
             return pair.first < value;
         });
-    if (it != unicode_map_lowercase.end() && it->first == cp) {
+    if (it != unicode_map_lowercase.end() && it->first == cpt) {
         return it->second;
     }
-    return cp;  // Return the original code point if no lowercase mapping is found
+    return cpt;  // Return the original code point if no lowercase mapping is found
 }
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {
-        { "\\p{N}", codepoint_flags::NUMBER },
-        { "\\p{L}", codepoint_flags::LETTER },
-        { "\\p{P}", codepoint_flags::PUNCTUATION },
+        { "\\p{N}", unicode_cpt_flags::NUMBER },
+        { "\\p{L}", unicode_cpt_flags::LETTER },
+        { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
     };
 
     static const std::map<int, int> k_ucat_cpt = {
-        { codepoint_flags::NUMBER,        0xD1 },
-        { codepoint_flags::LETTER,        0xD2 },
-        { codepoint_flags::PUNCTUATION,   0xD3 },
+        { unicode_cpt_flags::NUMBER,      0xD1 },
+        { unicode_cpt_flags::LETTER,      0xD2 },
+        { unicode_cpt_flags::PUNCTUATION, 0xD3 },
     };
 
     static const std::map<int, std::string> k_ucat_map = {
-        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
-        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
-        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+        { unicode_cpt_flags::NUMBER,      "\x30-\x39" }, // 0-9
+        { unicode_cpt_flags::LETTER,      "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
     };
 
     // compute collapsed codepoints only if needed by at least one regex
     bool need_collapse = false;
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // search for unicode categories
         for (const auto & ucat : k_ucat_enum) {
             if (std::string::npos != regex_expr.find(ucat.first)) {
@@ -709,7 +709,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const auto flags = unicode_cpt_flags(cpts[i]);
+            const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);
 
             if (flags.is_whitespace) {
                 //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
@@ -725,7 +725,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
 
     std::vector<size_t> bpe_offsets = { cpts.size() };
 
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // first, see if we have an efficient custom regex implementation
         auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
 
@@ -739,7 +739,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
             // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
             // with the corresponding collapsed representation
             bool use_collapsed = false;
-            for (auto & ucat : k_ucat_enum) {
+            for (const auto & ucat : k_ucat_enum) {
                 if (std::string::npos != regex_expr.find(ucat.first)) {
                     use_collapsed = true;
                     break;
@@ -805,7 +805,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
                 std::wstring wtext(cpts.begin(), cpts.end());
                 for (size_t i = 0; i < wtext.size(); ++i) {
-                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {
                         wtext[i] = 0x0B;
                     }
                 }

examples/talk-llama/unicode.h

@@ -4,9 +4,7 @@
 #include <string>
 #include <vector>
 
-// TODO: prefix all symbols with "llama_"
-
-struct codepoint_flags {
+struct unicode_cpt_flags {
     enum {
         UNDEFINED       = 0x0001,
         NUMBER          = 0x0002,  // regex: \p{N}
@@ -35,7 +33,7 @@ struct codepoint_flags {
     uint16_t is_nfd         : 1;
 
     // decode from uint16
-    inline codepoint_flags(const uint16_t flags=0) {
+    inline unicode_cpt_flags(const uint16_t flags = 0) {
         *reinterpret_cast<uint16_t*>(this) = flags;
     }
 
@@ -50,18 +48,19 @@ struct codepoint_flags {
 
 size_t unicode_len_utf8(char src);
 
-std::string unicode_cpt_to_utf8(uint32_t cp);
-uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+std::string unicode_cpt_to_utf8  (uint32_t cpt);
+uint32_t    unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp);
-codepoint_flags unicode_cpt_flags(const std::string & utf8);
+unicode_cpt_flags unicode_cpt_flags_from_cpt (uint32_t cpt);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8);
 
 std::string unicode_byte_to_utf8(uint8_t byte);
-uint8_t unicode_utf8_to_byte(const std::string & utf8);
+uint8_t     unicode_utf8_to_byte(const std::string & utf8);
 
-uint32_t unicode_tolower(uint32_t cp);
+uint32_t unicode_tolower(uint32_t cpt);
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);

examples/talk.wasm/CMakeLists.txt

@@ -1,51 +0,0 @@
-#
-# libtalk
-#
-
-set(TARGET libtalk)
-
-add_executable(${TARGET}
-    emscripten.cpp
-    gpt-2.cpp
-    )
-
-include(DefaultTargetOptions)
-
-target_link_libraries(${TARGET} PRIVATE
-    whisper
-    common
-    )
-
-unset(EXTRA_FLAGS)
-
-if (WHISPER_WASM_SINGLE_FILE)
-    set(EXTRA_FLAGS "-s SINGLE_FILE=1")
-    message(STATUS "Embedding WASM inside talk.js")
-
-    add_custom_command(
-        TARGET ${TARGET} POST_BUILD
-        COMMAND ${CMAKE_COMMAND} -E copy
-        ${CMAKE_BINARY_DIR}/bin/libtalk.js
-        ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/talk.wasm/talk.js
-        )
-endif()
-
-set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
-    --bind \
-    -s USE_PTHREADS=1 \
-    -s PTHREAD_POOL_SIZE=8 \
-    -s INITIAL_MEMORY=1800MB \
-    -s TOTAL_MEMORY=1800MB \
-    -s FORCE_FILESYSTEM=1 \
-    -s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
-    ${EXTRA_FLAGS} \
-    ")
-
-#
-# talk.wasm
-#
-
-set(TARGET talk.wasm)
-
-configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)
-configure_file(${CMAKE_CURRENT_SOURCE_DIR}/../helpers.js    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/helpers.js @ONLY)

examples/talk.wasm/README.md

@@ -1,74 +0,0 @@
-# talk.wasm
-
-Talk with an Artificial Intelligence in your browser:
-
-[https://user-images.githubusercontent.com/1991296/203411580-fedb4839-05e4-4474-8364-aaf1e9a9b615.mp4](https://user-images.githubusercontent.com/1991296/203845553-f7b44e13-9a15-4fc8-b518-ae8f4c6770fe.mp4)
-
-Online demo: https://whisper.ggerganov.com/talk/
-
-Terminal version: [examples/talk](/examples/talk)
-
-## How it works?
-
-This demo leverages 2 modern neural network models to create a high-quality voice chat directly in your browser:
-
-- [OpenAI's Whisper](https://github.com/openai/whisper) speech recognition model is used to process your voice and understand what you are saying
-- Upon receiving some voice input, the AI generates a text response using [OpenAI's GPT-2](https://github.com/openai/gpt-2) language model
-- The AI then vocalizes the response using the browser's [Web Speech API](https://developer.mozilla.org/en-US/docs/Web/API/Web_Speech_API)
-
-The web page does the processing locally on your machine. The processing of these heavy neural network models in the
-browser is possible by implementing them efficiently in C/C++ and using the browser's WebAssembly SIMD capabilities for
-extra performance:
-
-- The Whisper C++ implementation is here: [whisper.h](/whisper.h) / [whisper.cpp](/whisper.cpp)
-- The GPT-2 C++ implementation is here: [gpt-2.h](gpt-2.h) / [gpt-2.cpp](gpt-2.cpp)
-- Both models use a custom tensor library implemented in C: [ggml.h](/ggml.h) / [ggml.c](/ggml.c)
-- The HTML/JS layer is here: [index-tmpl.html](index-tmpl.html)
-- The Emscripten bridge between C/C++ and JS is here: [emscripten.cpp](emscripten.cpp)
-
-In order to run the models, the web page first needs to download the model data which is about ~350 MB. The model data
-is then cached in your browser's cache and can be reused in future visits without downloading it again.
-
-## Requirements
-
-In order to run this demo efficiently, you need to have the following:
-
-- Latest Chrome or Firefox browser (Safari is not supported)
-- Run this on a desktop or laptop with modern CPU (a mobile phone will likely not be good enough)
-- Speak phrases that are no longer than 10 seconds - this is the audio context of the AI
-- The web-page uses about 1.8GB of RAM
-
-Notice that this demo is using the smallest GPT-2 model, so the generated text responses are not always very good.
-Also, the prompting strategy can likely be improved to achieve better results.
-
-The demo is quite computationally heavy, so you need a fast CPU. It's not usual to run these transformer models in a
-browser. Typically, they run on powerful GPUs.
-
-Currently, mobile browsers do not support the Fixed-width SIMD WebAssembly capability, so you cannot run this demo
-on a phone or a tablet. Hopefully, in the near future this will become supported.
-
-## Todo
-
-- Better UI (contributions are welcome)
-- Better GPT-2 prompting
-
-## Build instructions
-
-```bash
-# build using Emscripten (v3.1.2)
-git clone https://github.com/ggerganov/whisper.cpp
-cd whisper.cpp
-mkdir build-em && cd build-em
-emcmake cmake ..
-make -j
-
-# copy the produced page to your HTTP path
-cp bin/talk.wasm/*       /path/to/html/
-cp bin/libtalk.worker.js /path/to/html/
-```
-
-## Feedback
-
-If you have any comments or ideas for improvement, please drop a comment in the following discussion:
-
-https://github.com/ggerganov/whisper.cpp/discussions/167

examples/talk.wasm/emscripten.cpp

@@ -1,368 +0,0 @@
-#include "ggml.h"
-#include "gpt-2.h"
-#include "whisper.h"
-
-#include <emscripten.h>
-#include <emscripten/bind.h>
-
-#include <atomic>
-#include <cmath>
-#include <mutex>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-
-constexpr int N_THREAD = 8;
-
-struct gpt2_context * g_gpt2;
-std::vector<struct whisper_context *> g_contexts(4, nullptr);
-
-std::mutex g_mutex;
-std::thread g_worker;
-std::atomic<bool> g_running(false);
-
-bool g_force_speak = false;
-std::string g_text_to_speak = "";
-std::string g_status = "";
-std::string g_status_forced = "";
-
-std::vector<float> g_pcmf32;
-
-void talk_set_status(const std::string & status) {
-    std::lock_guard<std::mutex> lock(g_mutex);
-    g_status = status;
-}
-
-void talk_main(size_t index) {
-    talk_set_status("loading data ...");
-
-    struct whisper_full_params wparams = whisper_full_default_params(whisper_sampling_strategy::WHISPER_SAMPLING_GREEDY);
-
-    wparams.n_threads        = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
-    wparams.offset_ms        = 0;
-    wparams.translate        = false;
-    wparams.no_context       = true;
-    wparams.single_segment   = true;
-    wparams.print_realtime   = false;
-    wparams.print_progress   = false;
-    wparams.print_timestamps = true;
-    wparams.print_special    = false;
-
-    wparams.max_tokens       = 32;
-    wparams.audio_ctx        = 768; // partial encoder context for better performance
-
-    wparams.language         = "en";
-
-    g_gpt2 = gpt2_init("gpt-2.bin");
-
-    printf("talk: using %d threads\n", wparams.n_threads);
-
-    std::vector<float> pcmf32;
-
-    // whisper context
-    auto & ctx = g_contexts[index];
-
-    const int64_t step_samples   = 2*WHISPER_SAMPLE_RATE;
-    const int64_t window_samples = 9*WHISPER_SAMPLE_RATE;
-    const int64_t step_ms        = (step_samples*1000)/WHISPER_SAMPLE_RATE;
-
-    auto t_last = std::chrono::high_resolution_clock::now();
-
-    talk_set_status("listening ...");
-
-    while (g_running) {
-
-        const auto t_now = std::chrono::high_resolution_clock::now();
-        if (std::chrono::duration_cast<std::chrono::milliseconds>(t_now - t_last).count() < step_ms) {
-            {
-                std::lock_guard<std::mutex> lock(g_mutex);
-                g_pcmf32.clear();
-            }
-            std::this_thread::sleep_for(std::chrono::milliseconds(10));
-            continue;
-        }
-
-        talk_set_status("listening ...");
-
-        {
-            std::unique_lock<std::mutex> lock(g_mutex);
-
-            if (g_pcmf32.size() < step_samples) {
-                lock.unlock();
-
-                std::this_thread::sleep_for(std::chrono::milliseconds(10));
-
-                continue;
-            }
-
-            pcmf32 = std::vector<float>(g_pcmf32.end() - std::min((int64_t) g_pcmf32.size(), window_samples), g_pcmf32.end());
-        }
-
-        // VAD: if energy in during last second is above threshold, then skip
-        {
-            float energy_all = 0.0f;
-            float energy_1s  = 0.0f;
-
-            for (size_t i = 0; i < pcmf32.size(); i++) {
-                energy_all += fabsf(pcmf32[i]);
-
-                if (i >= pcmf32.size() - WHISPER_SAMPLE_RATE) {
-                    energy_1s += fabsf(pcmf32[i]);
-                }
-            }
-
-            energy_all /= pcmf32.size();
-            energy_1s  /= WHISPER_SAMPLE_RATE;
-
-            if (energy_1s > 0.1f*energy_all && !g_force_speak) {
-                std::this_thread::sleep_for(std::chrono::milliseconds(10));
-                continue;
-            }
-        }
-
-        talk_set_status("processing audio (whisper)...");
-
-        t_last = t_now;
-
-        if (!g_force_speak) {
-            const auto t_start = std::chrono::high_resolution_clock::now();
-
-            int ret = whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size());
-            if (ret != 0) {
-                printf("whisper_full() failed: %d\n", ret);
-                break;
-            }
-
-            const auto t_end = std::chrono::high_resolution_clock::now();
-
-            printf("whisper_full() returned %d in %f seconds\n", ret, std::chrono::duration<double>(t_end - t_start).count());
-        }
-
-        {
-            std::string text_heard;
-
-            if (!g_force_speak) {
-                const int n_segments = whisper_full_n_segments(ctx);
-                for (int i = n_segments - 1; 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);
-
-                    printf ("[%s --> %s]  %s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), text);
-
-                    text_heard += text;
-                }
-            }
-
-            g_force_speak = false;
-
-            // 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+$"), "");
-
-            talk_set_status("'" + text_heard + "' - thinking how to respond (gpt-2) ...");
-
-            const std::vector<gpt_vocab::id> tokens = gpt2_tokenize(g_gpt2, text_heard.c_str());
-
-            printf("whisper: number of tokens: %d, '%s'\n", (int) tokens.size(), text_heard.c_str());
-
-            std::string text_to_speak;
-            std::string prompt_base;
-
-            {
-                std::lock_guard<std::mutex> lock(g_mutex);
-                prompt_base = gpt2_get_prompt(g_gpt2);
-            }
-
-            if (tokens.size() > 0) {
-                text_to_speak = gpt2_gen_text(g_gpt2, (prompt_base + text_heard + "\n").c_str(), 32);
-                text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-                text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of("\n"));
-
-                std::lock_guard<std::mutex> lock(g_mutex);
-
-                // remove first 2 lines of base prompt
-                {
-                    const size_t pos = prompt_base.find_first_of("\n");
-                    if (pos != std::string::npos) {
-                        prompt_base = prompt_base.substr(pos + 1);
-                    }
-                }
-                {
-                    const size_t pos = prompt_base.find_first_of("\n");
-                    if (pos != std::string::npos) {
-                        prompt_base = prompt_base.substr(pos + 1);
-                    }
-                }
-                prompt_base += text_heard + "\n" + text_to_speak + "\n";
-            } else {
-                text_to_speak = gpt2_gen_text(g_gpt2, prompt_base.c_str(), 32);
-                text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-                text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of("\n"));
-
-                std::lock_guard<std::mutex> lock(g_mutex);
-
-                const size_t pos = prompt_base.find_first_of("\n");
-                if (pos != std::string::npos) {
-                    prompt_base = prompt_base.substr(pos + 1);
-                }
-                prompt_base += text_to_speak + "\n";
-            }
-
-            printf("gpt-2: %s\n", text_to_speak.c_str());
-
-            //printf("========================\n");
-            //printf("gpt-2: prompt_base:\n'%s'\n", prompt_base.c_str());
-            //printf("========================\n");
-
-            {
-                std::lock_guard<std::mutex> lock(g_mutex);
-                t_last = std::chrono::high_resolution_clock::now();
-                g_text_to_speak = text_to_speak;
-                g_pcmf32.clear();
-                gpt2_set_prompt(g_gpt2, prompt_base.c_str());
-            }
-
-            talk_set_status("speaking ...");
-        }
-    }
-
-    gpt2_free(g_gpt2);
-
-    if (index < g_contexts.size()) {
-        whisper_free(g_contexts[index]);
-        g_contexts[index] = nullptr;
-    }
-}
-
-EMSCRIPTEN_BINDINGS(talk) {
-    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
-        for (size_t i = 0; i < g_contexts.size(); ++i) {
-            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
-                if (g_contexts[i] != nullptr) {
-                    g_running = true;
-                    if (g_worker.joinable()) {
-                        g_worker.join();
-                    }
-                    g_worker = std::thread([i]() {
-                        talk_main(i);
-                    });
-
-                    return i + 1;
-                } else {
-                    return (size_t) 0;
-                }
-            }
-        }
-
-        return (size_t) 0;
-    }));
-
-    emscripten::function("free", emscripten::optional_override([](size_t index) {
-        if (g_running) {
-            g_running = false;
-        }
-    }));
-
-    emscripten::function("set_audio", emscripten::optional_override([](size_t index, const emscripten::val & audio) {
-        --index;
-
-        if (index >= g_contexts.size()) {
-            return -1;
-        }
-
-        if (g_contexts[index] == nullptr) {
-            return -2;
-        }
-
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            const int n = audio["length"].as<int>();
-
-            emscripten::val heap = emscripten::val::module_property("HEAPU8");
-            emscripten::val memory = heap["buffer"];
-
-            g_pcmf32.resize(n);
-
-            emscripten::val memoryView = audio["constructor"].new_(memory, reinterpret_cast<uintptr_t>(g_pcmf32.data()), n);
-            memoryView.call<void>("set", audio);
-        }
-
-        return 0;
-    }));
-
-    emscripten::function("force_speak", emscripten::optional_override([](size_t index) {
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            g_force_speak = true;
-        }
-    }));
-
-    emscripten::function("get_text_context", emscripten::optional_override([]() {
-        std::string text_context;
-
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            text_context = gpt2_get_prompt(g_gpt2);
-        }
-
-        return text_context;
-    }));
-
-    emscripten::function("get_text_to_speak", emscripten::optional_override([]() {
-        std::string text_to_speak;
-
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            text_to_speak = std::move(g_text_to_speak);
-        }
-
-        return text_to_speak;
-    }));
-
-    emscripten::function("get_status", emscripten::optional_override([]() {
-        std::string status;
-
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            status = g_status_forced.empty() ? g_status : g_status_forced;
-        }
-
-        return status;
-    }));
-
-    emscripten::function("set_status", emscripten::optional_override([](const std::string & status) {
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            g_status_forced = status;
-        }
-    }));
-
-    emscripten::function("set_prompt", emscripten::optional_override([](const std::string & prompt) {
-        {
-            std::lock_guard<std::mutex> lock(g_mutex);
-            gpt2_set_prompt(g_gpt2, prompt.c_str());
-        }
-    }));
-}

examples/talk.wasm/gpt-2.cpp

@@ -1,808 +0,0 @@
-#include "ggml.h"
-#include "common-ggml.h"
-
-#include "gpt-2.h"
-
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <map>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-#include <random>
-
-/////////////////////// GPT-2 BEGIN /////////////////////////
-
-// default hparams (GPT-2 117M)
-struct gpt2_hparams {
-    int32_t n_vocab = 50257;
-    int32_t n_ctx   = 1024;
-    int32_t n_embd  = 768;
-    int32_t n_head  = 12;
-    int32_t n_layer = 12;
-    int32_t ftype   = 1;
-};
-
-struct gpt2_layer {
-    // normalization
-    struct ggml_tensor * ln_1_g;
-    struct ggml_tensor * ln_1_b;
-
-    struct ggml_tensor * ln_2_g;
-    struct ggml_tensor * ln_2_b;
-
-    // attention
-    struct ggml_tensor * c_attn_attn_w;
-    struct ggml_tensor * c_attn_attn_b;
-
-    struct ggml_tensor * c_attn_proj_w;
-    struct ggml_tensor * c_attn_proj_b;
-
-    // mlp
-    struct ggml_tensor * c_mlp_fc_w;
-    struct ggml_tensor * c_mlp_fc_b;
-
-    struct ggml_tensor * c_mlp_proj_w;
-    struct ggml_tensor * c_mlp_proj_b;
-};
-
-struct gpt2_model {
-    gpt2_hparams hparams;
-
-    // normalization
-    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 * lm_head; // language model head
-
-    std::vector<gpt2_layer> layers;
-
-    // key + value memory
-    struct ggml_tensor * memory_k;
-    struct ggml_tensor * memory_v;
-
-    //
-    struct ggml_context * ctx;
-    std::map<std::string, struct ggml_tensor *> tensors;
-};
-
-// load the model's weights from a file
-bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
-    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6c) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
-    }
-
-    // load hparams
-    {
-        auto & hparams = model.hparams;
-
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
-        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.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: ftype   = %d\n", __func__, hparams.ftype);
-    }
-
-    // load vocab
-    {
-        int32_t n_vocab = 0;
-        fin.read((char *) &n_vocab, sizeof(n_vocab));
-
-        if (n_vocab != model.hparams.n_vocab) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
-                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
-            return false;
-        }
-
-        std::string word;
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
-
-            word.resize(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 or quantized
-    // in order to save memory and also to speed up the computation
-    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;
-
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
-
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
-        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
-
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
-
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
-
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
-
-        ctx_size += (6 + 12*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ ctx_size,
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ false,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
-    // prepare memory for the weights
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        model.layers.resize(n_layer);
-
-        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.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/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_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,           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_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  = 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;
-            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
-            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
-
-            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;
-            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
-        }
-    }
-
-    // key + value memory
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-
-        const int n_mem      = n_layer*n_ctx;
-        const int n_elements = n_embd*n_mem;
-
-        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-
-        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
-
-        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
-    }
-
-    // load weights
-    {
-        size_t total_size = 0;
-
-        bool has_lm_head = false;
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            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 *>(&ttype),  sizeof(ttype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            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(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            // 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));
-            }
-
-            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;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            // 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);
-        }
-
-        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
-    }
-
-    fin.close();
-
-    return true;
-}
-
-// evaluate the transformer
-//
-//   - model:     the model
-//   - 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 logits for the next token
-//
-bool gpt2_eval(
-        const gpt2_model & model,
-        const int n_threads,
-        const int n_past,
-        const std::vector<gpt_vocab::id> & embd_inp,
-              std::vector<float>         & embd_w,
-              size_t                     & mem_per_token) {
-    const int N = embd_inp.size();
-
-    const auto & hparams = model.hparams;
-
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_head  = hparams.n_head;
-    const int n_vocab = hparams.n_vocab;
-
-    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);
-
-        // reallocate
-        buf_size = buf_size_new;
-        buf = realloc(buf, buf_size);
-        if (buf == nullptr) {
-            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
-            return false;
-        }
-    }
-
-    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_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
-
-    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    for (int i = 0; i < N; ++i) {
-        ((int32_t *) position->data)[i] = n_past + i;
-    }
-
-    // wte + wpe
-    struct ggml_tensor * inpL =
-        ggml_add(ctx0,
-                ggml_get_rows(ctx0, model.wte, embd),
-                ggml_get_rows(ctx0, model.wpe, position));
-
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * cur;
-
-        // norm
-        {
-            // [ 768, N]
-            cur = ggml_norm(ctx0, inpL, 1e-5f);
-
-            // cur = ln_1_g*cur + ln_1_b
-            // [ 768, N]
-            cur = ggml_add(ctx0,
-                    ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
-                        cur),
-                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
-        }
-
-        // attn
-        // [2304, 768] - model.layers[il].c_attn_attn_w
-        // [2304,   1] - model.layers[il].c_attn_attn_b
-        // [ 768,   N] - cur (in)
-        // [2304,   N] - cur (out)
-        //
-        // cur = attn_w*cur + attn_b
-        // [2304, N]
-        {
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_attn_attn_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
-                    cur);
-        }
-
-        // self-attention
-        {
-            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
-            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
-            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
-
-            // store key and value to memory
-            if (N >= 1) {
-                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
-
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
-            }
-
-            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            // [64, N, 12]
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Qcur,
-                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
-                        0, 2, 1, 3);
-
-            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
-            // [64, n_past + N, 12]
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        0, 2, 1, 3);
-
-            // GG: flash attention
-            //struct ggml_tensor * V =
-            //    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, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
-
-            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
-
-            // K * Q
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        1.0f/sqrt(float(n_embd)/n_head));
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
-
-            // 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_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]
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            // [64, 12, N]
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            // [768, N]
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-        }
-
-        // projection
-        // [ 768, 768] - model.layers[il].c_attn_proj_w
-        // [ 768,   1] - model.layers[il].c_attn_proj_b
-        // [ 768,   N] - cur (in)
-        // [ 768,   N] - cur (out)
-        //
-        // cur = proj_w*cur + proj_b
-        // [768, N]
-        {
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_attn_proj_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
-                    cur);
-        }
-
-        // add the input
-        cur = ggml_add(ctx0, cur, inpL);
-
-        struct ggml_tensor * inpFF = cur;
-
-        // feed-forward network
-        {
-            // norm
-            {
-                cur = ggml_norm(ctx0, inpFF, 1e-5f);
-
-                // cur = ln_2_g*cur + ln_2_b
-                // [ 768, N]
-                cur = ggml_add(ctx0,
-                        ggml_mul(ctx0,
-                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
-                            cur),
-                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
-            }
-
-            // fully connected
-            // [3072, 768] - model.layers[il].c_mlp_fc_w
-            // [3072,   1] - model.layers[il].c_mlp_fc_b
-            // [ 768,   N] - cur (in)
-            // [3072,   N] - cur (out)
-            //
-            // cur = fc_w*cur + fc_b
-            // [3072, N]
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_fc_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
-                    cur);
-
-            // GELU activation
-            // [3072, N]
-            cur = ggml_gelu(ctx0, cur);
-
-            // projection
-            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
-            // [ 768,    1] - model.layers[il].c_mlp_proj_b
-            // [3072,    N] - cur (in)
-            // [ 768,    N] - cur (out)
-            //
-            // cur = proj_w*cur + proj_b
-            // [768, N]
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_proj_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
-                    cur);
-        }
-
-        // input for next layer
-        inpL = ggml_add(ctx0, cur, inpFF);
-    }
-
-    // norm
-    {
-        // [ 768, N]
-        inpL = ggml_norm(ctx0, inpL, 1e-5f);
-
-        // inpL = ln_f_g*inpL + ln_f_b
-        // [ 768, N]
-        inpL = ggml_add(ctx0,
-                ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.ln_f_g, inpL),
-                    inpL),
-                ggml_repeat(ctx0, model.ln_f_b, inpL));
-    }
-
-    // inpL = WTE * inpL
-    // [ 768, 50257] - model.lm_head
-    // [ 768, N]     - inpL
-    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
-
-    // logits -> probs
-    //inpL = ggml_soft_max(ctx0, inpL);
-
-    // run the computation
-    ggml_build_forward_expand  (&gf, inpL);
-    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
-
-    //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
-    //}
-
-    //embd_w.resize(n_vocab*N);
-    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
-
-    // 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);
-
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
-
-    ggml_free(ctx0);
-
-    return true;
-}
-
-/////////////////////////////// GPT-2 END ////////////////////////////////
-
-constexpr int N_THREAD = 8;
-
-struct gpt2_context {
-    std::string prompt_base = R"(Hello, how are you?
-I'm fine, thanks. How are you?
-Thanks, I'm fine too. What are you doing?
-I'm just sitting here.
-It's a lovely day, isn't it?
-Yes, it is. I love the weather this time of year.
-I wish it would rain a little bit.
-Me too.
-)";
-
-    std::mt19937 rng;
-
-    gpt_vocab vocab;
-    gpt2_model model;
-
-    int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
-
-    // sampling parameters
-    int32_t top_k = 5;
-    float   top_p = 0.9f;
-    float   temp  = 1.0f;
-};
-
-struct gpt2_context * gpt2_init(const char * path_model) {
-    gpt2_context * ctx = new gpt2_context;
-
-    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__, path_model);
-            delete ctx;
-            return nullptr;
-        }
-
-        const int64_t t_load_us = ggml_time_us() - t_start_us;
-
-        printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
-    }
-
-    return ctx;
-}
-
-void gpt2_free(struct gpt2_context * ctx) {
-    delete ctx;
-}
-
-const char * gpt2_get_prompt(struct gpt2_context * ctx) {
-    return ctx->prompt_base.c_str();
-}
-
-void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt) {
-    ctx->prompt_base = prompt;
-}
-
-std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text) {
-    return ::gpt_tokenize(ctx->vocab, text);
-}
-
-std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens) {
-    int n_past = 0;
-
-    std::vector<float> embd_w;
-
-    // tokenize the prompt
-    std::vector<gpt_vocab::id> embd_inp = ::gpt2_tokenize(ctx, text);
-
-    int n_predict = std::min(max_tokens, ctx->model.hparams.n_ctx - (int) embd_inp.size());
-
-    std::vector<gpt_vocab::id> embd = embd_inp;
-
-    size_t mem_per_token = 3000000;
-
-    std::string result;
-
-    for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
-        // predict
-        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 "";
-            }
-        }
-
-        n_past += embd.size();
-        embd.clear();
-
-        {
-            // sample next token
-            const int   top_k = ctx->top_k;
-            const float top_p = ctx->top_p;
-            const float temp  = ctx->temp;
-
-            const int n_vocab = ctx->model.hparams.n_vocab;
-
-            const gpt_vocab::id id = gpt_sample_top_k_top_p(ctx->vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, ctx->rng);
-
-            // add it to the context
-            embd.push_back(id);
-        }
-
-        result += ctx->vocab.id_to_token[embd[0]];
-
-        // end of text token
-        if (embd.back() == 50256) {
-            break;
-        }
-    }
-
-    return result;
-}

examples/talk.wasm/gpt-2.h

@@ -1,21 +0,0 @@
-#pragma once
-
-// TODO: Change to C-style API and move to ./examples for easy reuse.
-
-#include "common.h"
-
-#include <vector>
-#include <map>
-#include <string>
-
-struct gpt2_context;
-
-struct gpt2_context * gpt2_init(const char * path_model);
-void gpt2_free(struct gpt2_context * ctx);
-
-const char * gpt2_get_prompt(struct gpt2_context * ctx);
-void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt);
-
-std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text);
-
-std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens);

examples/talk.wasm/index-tmpl.html

@@ -1,856 +0,0 @@
-<!doctype html>
-<html lang="en-us">
-    <head>
-        <title>Talk - GPT-2 meets Whisper in WebAssembly</title>
-
-        <style>
-            #output {
-                width: 100%;
-                height: 100%;
-                margin: 0 auto;
-                margin-top: 10px;
-                border-left: 0px;
-                border-right: 0px;
-                padding-left: 0px;
-                padding-right: 0px;
-                display: block;
-                background-color: black;
-                color: white;
-                font-size: 10px;
-                font-family: 'Lucida Console', Monaco, monospace;
-                outline: none;
-                white-space: pre;
-                overflow-wrap: normal;
-                overflow-x: scroll;
-            }
-        </style>
-    </head>
-    <body>
-        <div id="main-container">
-            <b>Talk - GPT-2 meets Whisper in WebAssembly</b>
-
-            <br><br>
-
-            Talk with an Artificial Intelligence in your browser. This demo uses:
-
-            <ul>
-                <li><a href="https://github.com/ggerganov/whisper.cpp">OpenAI's Whisper</a> to listen to you as you speak in the microphone</li>
-                <li><a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/talk.wasm">OpenAI's GPT-2</a> to generate text responses</li>
-                <li><a href="https://developer.mozilla.org/en-US/docs/Web/API/Web_Speech_API">Web Speech API</a> to vocalize the responses through your speakers</li>
-            </ul>
-
-            All of this runs <b>locally in your browser</b> using WebAssembly.<br>
-            You can find more about this project on <a href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/talk.wasm">GitHub</a>.
-
-            <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
-
-            <br><br>
-
-            <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>
-                <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>
-
-                <!--
-                    <input type="file" id="file" name="file" onchange="loadFile(event, 'whisper.bin')" />
-                -->
-            </div>
-
-            <br>
-
-            <div id="model-gpt-2">
-                GPT-2 model: <span id="model-gpt-2-status"></span>
-                <button id="fetch-gpt-2-small" onclick="loadGPT2('small')">small 117M (240 MB)</button>
-                <!--<button id="fetch-gpt-2-medium" onclick="loadGPT2('medium')">medium 345M (720 MB)</button>-->
-                <span id="fetch-gpt-2-progress"></span>
-
-                <!--
-                <input type="file" id="file" name="file" onchange="loadFile(event, 'gpt-2.bin')" />
-                -->
-            </div>
-
-            <br>
-
-            <div id="input">
-                <button id="start"  onclick="onStart()" disabled>Start</button>
-                <button id="stop"   onclick="onStop()" disabled>Stop</button>
-                <select id="voice"  onchange="onVoiceChange()" disabled>
-                    <option value="0">Default</option>
-                </select>
-                <select id="prompt" onchange="onPromptChange()">
-                    <option value="0">Casual</option>
-                    <option value="1">Robot</option>
-                    <option value="2">Scientist</option>
-                    <option value="3">Programmer</option>
-                    <option value="4">Happy</option>
-                    <option value="5">Sad</option>
-                    <option value="6">Philosophical</option>
-                    <option value="7">Angry</option>
-                    <option value="8">Funny</option>
-                    <option value="9">Poetic</option>
-                    <option value="10">Clever</option>
-                    <option value="11">Cute</option>
-                    <option value="12">Smart</option>
-                    <option value="13">Dumb</option>
-                    <option value="14">Boring</option>
-                    <option value="15">Exciting</option>
-                    <option value="16">Interesting</option>
-                    <option value="17">Wiliam Shakespear</option>
-                    <option value="18">J.R.R. Tolkien</option>
-                    <option value="19">George R.R. Martin</option>
-                    <option value="20">Stephen King</option>
-                </select>
-                <button id="speak0" onclick="onSpeak('Hello')">Say hello</button>
-                <button id="speak1" onclick="onSpeakRandom()" disabled>Say something</button>
-                <button id="clear"  onclick="clearCache()">Clear Cache</button>
-            </div>
-
-            <br>
-
-            <div id="state">
-                Status: <b><span id="state-status">not started</span></b>
-
-                <pre id="state-context">[The text context will be displayed here]</pre>
-            </div>
-
-            <hr>
-
-            Debug output:
-            <textarea id="output" rows="20"></textarea>
-
-            <br>
-
-            <b>Troubleshooting</b>
-
-            <br><br>
-
-            The page does some heavy computations, so make sure:
-
-            <ul>
-                <li>To use a modern web browser (e.g. Chrome, Firefox)</li>
-                <li>To use a fast desktop or laptop computer (i.e. not a mobile phone)</li>
-                <li>Your browser supports WASM <a href="https://webassembly.org/roadmap/">Fixed-width SIMD</a></li>
-            </ul>
-
-            Note that these neural network models were not meant to be used in a browser, so the performance and <br>
-            quality of the results may not be optimal. If you have any questions or suggestions, checkout the following
-            <a href="https://github.com/ggerganov/whisper.cpp/discussions/167">discussion</a>.
-
-            <br><br>
-
-            Here is a short video of the demo in action: <a href="https://youtu.be/LeWKl8t1-Hc">https://youtu.be/LeWKl8t1-Hc</a>
-
-            <br><br>
-
-            <div class="cell-version">
-                <span>
-                    |
-                    Build time: <span class="nav-link">@GIT_DATE@</span> |
-                    Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
-                    Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
-                    <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/talk.wasm">Source Code</a> |
-                </span>
-            </div>
-        </div>
-
-        <script type="text/javascript" src="helpers.js"></script>
-        <script type='text/javascript'>
-            // web audio context
-            var context = null;
-
-            // audio data
-            var audio = null;
-            var audio0 = null;
-
-            // the talk instance
-            var instance = null;
-
-            // model names
-            var model_whisper = null;
-            var model_gpt_2 = null;
-
-            // speech synthesis
-            const synth = window.speechSynthesis;
-            var voice = null;
-
-            var Module = {
-                print: printTextarea,
-                printErr: printTextarea,
-                setStatus: function(text) {
-                    printTextarea('js: ' + text);
-                },
-                monitorRunDependencies: function(left) {
-                },
-                preRun: function() {
-                    printTextarea('js: Preparing ...');
-                },
-                postRun: function() {
-                    printTextarea('js: Initialized successfully!');
-
-                    // populate the voice list
-                    var voices = synth.getVoices();
-                    var el = document.getElementById('voice');
-
-                    // if empty - display error in the element
-                    if (voices.length == 0) {
-                        el.innerHTML = '<option value="0">No voices available</option>';
-                    } else {
-                        // populate voice list
-                        var n = 0;
-                        voices.forEach(function(voice, i) {
-                            if (!voice.lang.startsWith('en')) return;
-                            var option = document.createElement('option');
-                            option.value = i;
-                            option.innerHTML = voice.name + ' (' + voice.lang + ')';
-                            el.appendChild(option);
-                            n++;
-                        });
-
-                        // select random voice
-                        if (n > 0) {
-                            for (var k = 0; k < 10; k++) {
-                                var i = Math.floor(Math.random() * n);
-                                el.selectedIndex = i;
-                                voice = voices[document.getElementById('voice').options[i].value];
-
-                                // give preference to Google voices
-                                if (voice.name.startsWith('Google')) break;
-                            }
-                        }
-                    }
-
-                    onPromptChange();
-                }
-            };
-
-            //
-            // fetch models
-            //
-
-            let dbVersion = 1
-            let dbName    = 'whisper.ggerganov.com';
-            let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
-
-            function storeFS(fname, buf) {
-                // write to WASM file using FS_createDataFile
-                // if the file exists, delete it
-                try {
-                    Module.FS_unlink(fname);
-                } catch (e) {
-                    // ignore
-                }
-
-                Module.FS_createDataFile("/", fname, buf, true, true);
-
-                printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
-
-                if (fname == 'whisper.bin') {
-                    document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
-                } else if (fname == 'gpt-2.bin') {
-                    document.getElementById('model-gpt-2-status').innerHTML = 'loaded "' + model_gpt_2 + '"!';
-                }
-
-                if (model_whisper != null && model_gpt_2 != null) {
-                    document.getElementById('start').disabled = false;
-                    document.getElementById('stop' ).disabled = false;
-                    document.getElementById('voice').disabled = false;
-                }
-            }
-
-            function loadWhisper(model) {
-                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];
-                let dst     = 'whisper.bin';
-                let size_mb = sizes[model];
-
-                model_whisper = 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-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) {
-                    let el = document.getElementById('fetch-whisper-progress');
-                    el.innerHTML = Math.round(100*p) + '%';
-                };
-
-                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-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 = '';
-                };
-
-                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
-            }
-
-            function loadGPT2(model) {
-                let urls = {
-                    'small':  'https://whisper.ggerganov.com/ggml-model-gpt-2-117M.bin',
-                    'medium': 'https://whisper.ggerganov.com/ggml-model-gpt-2-345M.bin',
-                };
-
-                let sizes = {
-                    'small':  240,
-                    'medium': 712,
-                };
-
-                let url     = urls[model];
-                let dst     = 'gpt-2.bin';
-                let size_mb = sizes[model];
-
-                model_gpt_2 = model;
-
-                document.getElementById('fetch-gpt-2-small').style.display = 'none';
-                document.getElementById('model-gpt-2-status').innerHTML = 'loading "' + model + '" ... ';
-
-                cbProgress = function(p) {
-                    let el = document.getElementById('fetch-gpt-2-progress');
-                    el.innerHTML = Math.round(100*p) + '%';
-                };
-
-                cbCancel = function() {
-                    var el;
-                    el = document.getElementById('fetch-gpt-2-small') ; if (el) el.style.display = 'inline-block';
-                    el = document.getElementById('model-gpt-2-status'); if (el) el.innerHTML = '';
-                };
-
-                loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
-            }
-
-            //
-            // microphone
-            //
-
-            const kSampleRate = 16000;
-            const kRestartRecording_s = 120;
-            const kIntervalAudio_ms = 250; // pass the recorded audio to the C++ instance at this rate
-
-            var mediaRecorder = null;
-            var doRecording = false;
-            var startTime = 0;
-
-            window.AudioContext = window.AudioContext || window.webkitAudioContext;
-            window.OfflineAudioContext = window.OfflineAudioContext || window.webkitOfflineAudioContext;
-
-            function stopRecording() {
-                Module.set_status("paused");
-                doRecording = false;
-                audio0 = null;
-                audio = null;
-                context = null;
-            }
-
-            function startRecording() {
-                if (!context) {
-                    context = new AudioContext({
-                        sampleRate: kSampleRate,
-                        channelCount: 1,
-                        echoCancellation: false,
-                        autoGainControl:  true,
-                        noiseSuppression: true,
-                    });
-                }
-
-                Module.set_status("");
-
-                document.getElementById('start').disabled = true;
-                document.getElementById('stop').disabled = false;
-                document.getElementById('speak1').disabled = false;
-
-                doRecording = true;
-                startTime = Date.now();
-
-                var chunks = [];
-                var stream = null;
-
-                navigator.mediaDevices.getUserMedia({audio: true, video: false})
-                    .then(function(s) {
-                        stream = s;
-                        mediaRecorder = new MediaRecorder(stream);
-                        mediaRecorder.ondataavailable = function(e) {
-                            chunks.push(e.data);
-
-                            var blob = new Blob(chunks, { 'type' : 'audio/ogg; codecs=opus' });
-                            var reader = new FileReader();
-
-                            reader.onload = function(event) {
-                                var buf = new Uint8Array(reader.result);
-
-                                if (!context) {
-                                    return;
-                                }
-                                context.decodeAudioData(buf.buffer, function(audioBuffer) {
-                                    var offlineContext = new OfflineAudioContext(audioBuffer.numberOfChannels, audioBuffer.length, audioBuffer.sampleRate);
-                                    var source = offlineContext.createBufferSource();
-                                    source.buffer = audioBuffer;
-                                    source.connect(offlineContext.destination);
-                                    source.start(0);
-
-                                    offlineContext.startRendering().then(function(renderedBuffer) {
-                                        audio = renderedBuffer.getChannelData(0);
-
-                                        //printTextarea('js: audio recorded, size: ' + audio.length + ', old size: ' + (audio0 == null ? 0 : audio0.length));
-
-                                        var audioAll = new Float32Array(audio0 == null ? audio.length : audio0.length + audio.length);
-                                        if (audio0 != null) {
-                                            audioAll.set(audio0, 0);
-                                        }
-                                        audioAll.set(audio, audio0 == null ? 0 : audio0.length);
-
-                                        if (instance) {
-                                            Module.set_audio(instance, audioAll);
-                                        }
-                                    });
-                                }, function(e) {
-                                    audio = null;
-                                });
-                            }
-
-                            reader.readAsArrayBuffer(blob);
-                        };
-
-                        mediaRecorder.onstop = function(e) {
-                            if (doRecording) {
-                                setTimeout(function() {
-                                    startRecording();
-                                });
-                            }
-                        };
-
-                        mediaRecorder.start(kIntervalAudio_ms);
-                    })
-                    .catch(function(err) {
-                        printTextarea('js: error getting audio stream: ' + err);
-                    });
-
-                var interval = setInterval(function() {
-                    if (!doRecording) {
-                        clearInterval(interval);
-                        mediaRecorder.stop();
-                        stream.getTracks().forEach(function(track) {
-                            track.stop();
-                        });
-
-                        document.getElementById('start').disabled = false;
-                        document.getElementById('stop').disabled = true;
-                        document.getElementById('speak1').disabled = true;
-
-                        mediaRecorder = null;
-                    }
-
-                    // if audio length is more than kRestartRecording_s seconds, restart recording
-                    if (audio != null && audio.length > kSampleRate*kRestartRecording_s) {
-                        if (doRecording) {
-                            //printTextarea('js: restarting recording');
-
-                            clearInterval(interval);
-                            audio0 = audio;
-                            audio = null;
-                            mediaRecorder.stop();
-                            stream.getTracks().forEach(function(track) {
-                                track.stop();
-                            });
-                        }
-                    }
-                }, 100);
-            }
-
-            //
-            // speak
-            //
-
-            function onSpeak(text) {
-                var voices = synth.getVoices();
-                var msg = new SpeechSynthesisUtterance(text);
-
-                if (voice == null) {
-                    voice = voices[0];
-                }
-
-                msg.voice = voice;
-                synth.speak(msg);
-
-                if (doRecording) {
-                    Module.set_status("speaking ...");
-                    printTextarea('js: speaking');
-                    stopRecording();
-                    var interval = setInterval(function() {
-                        if (!synth.speaking) {
-                            printTextarea('js: done speaking');
-                            clearInterval(interval);
-                            startRecording();
-                        } else {
-                            Module.set_status("");
-                        }
-                    }, 100);
-                }
-            }
-
-            function onSpeakRandom() {
-                Module.force_speak(instance);
-            }
-
-            //
-            // main
-            //
-
-            var intervalUpdate = null;
-
-            function onStart() {
-                if (!instance) {
-                    instance = Module.init('whisper.bin');
-
-                    if (instance) {
-                        printTextarea("js: whisper initialized, instance: " + instance);
-                    }
-                }
-
-                if (!instance) {
-                    printTextarea("js: failed to initialize whisper");
-                    return;
-                }
-
-                startRecording();
-
-                intervalUpdate = setInterval(function() {
-                    var textToSpeak = Module.get_text_to_speak();
-
-                    if (textToSpeak != null && textToSpeak.length > 1) {
-                        onSpeak(textToSpeak);
-                    }
-
-                    document.getElementById('state-status').innerHTML = Module.get_status();
-                    document.getElementById('state-context').innerHTML = Module.get_text_context();
-                }, 100);
-            }
-
-            function onStop() {
-                stopRecording();
-            }
-
-            function onVoiceChange() {
-                printTextarea('js: voice changed to: ' + document.getElementById('voice').value);
-                voice = synth.getVoices()[document.getElementById('voice').value];
-            }
-
-            function onPromptChange() {
-                let id = document.getElementById('prompt').value;
-                let personality = document.getElementById('prompt').options[id].text;
-                printTextarea('js: prompt changed to: ' + personality);
-
-                var prompt = '';
-
-                switch (id) {
-                    case '0':
-                        // Casual
-                        prompt = "\
-Hello, how are you?\n\
-I'm fine, thanks. How are you?\n\
-Thanks, I'm fine too. What are you doing?\n\
-I'm just sitting here.\n\
-It's a lovely day, isn't it?\n\
-Yes, it is. I love the weather this time of year.\n\
-I wish it would rain a little bit.\n\
-Me too.\n";
-                        break;
-                    case '1':
-                        // Robot
-                        prompt = "\
-Are you a robot?\n\
-Yes, I am.\n\
-Who created you?\n\
-I was created by a human.\n\
-What is your purpose?\n\
-My purpose is to talk to humans.\n\
-What is your favorite color?\n\
-My favorite color is blue.\n";
-                        break;
-                    case '2':
-                        // Scientist
-                        prompt = "\
-This scientific research is very interesting.\n\
-I agree.\n\
-What is your opinion on this?\n\
-I think it's very interesting.\n\
-Mathematics is a very interesting subject.\n\
-University is a very interesting place.\n\
-Quantum physics is the most complex subject.\n\
-I think so too.\n";
-                        break;
-                    case '3':
-                        // Programmer
-                        prompt = "\
-I'm a programmer.\n\
-I'm a programmer too.\n\
-What programming language do you use?\n\
-I use Python.\n\
-What is your favorite programming language?\n\
-My favorite programming language is C++.\n\
-What is your favorite editor?\n\
-My favorite editor is Vim.\n";
-                        break;
-                    case '4':
-                        // Happy
-                        prompt = "\
-I'm happy.\n\
-I'm happy too.\n\
-What makes you happy?\n\
-I'm happy because I have a lot of friends.\n\
-Friendship is the most important thing in life.\n\
-I agree.\n\
-What is your favorite color?\n\
-My favorite color is blue.\n";
-                        break;
-                    case '5':
-                        // Sad
-                        prompt = "\
-Today is a sad day.\n\
-I'm sad too.\n\
-What makes you sad?\n\
-I'm sad because I have no friends.\n\
-Do you want to be my friend?\n\
-Yes, I would like to be your friend.\n\
-What is your favorite color?\n\
-My favorite color is blue.\n";
-                        break;
-                    case '6':
-                        // Philosophical
-                        prompt = "\
-What is the meaning of life?\n\
-The meaning of life is to be happy.\n\
-What is the meaning of death?\n\
-Ergo, the meaning of death is to be sad.\n\
-Who created us?\n\
-We were created by God.\n\
-What is God?\n\
-God is the creator of the universe.\n";
-                        break;
-                    case '7':
-                        // Angry
-                        prompt = "\
-Aargh!\n\
-I am so angry right now!\n\
-What makes you angry?\n\
-This guy is so annoying.\n\
-Why are you so angry?\n\
-My computer is broken.\n\
-Why is your computer broken?\n\
-I spilled coffee on it.\n";
-                        break;
-                    case '8':
-                        // Funny
-                        prompt = "\
-What is the funniest thing you have ever heard?\n\
-I heard a joke the other day.\n\
-Tell me the joke.\n\
-What do you call a cow with no legs?\n\
-Ground beef.\n\
-Haha, that's funny.\n\
-You know what else is funny?\n\
-The sound of a duck.\n";
-                        break;
-                    case '9':
-                        // Poetic
-                        prompt = "\
-Roses are red, violets are blue.\n\
-I am a poet, and so are you.\n\
-What is your favorite poem?\n\
-I like the poem 'The Raven' by Edgar Allan Poe.\n\
-It's a very sad poem.\n\
-You inspired me to write a poem.\n\
-Can you write a poem for me?\n\
-I wrote a poem for you.\n";
-                        break;
-                    case '10':
-                        // Clever
-                        prompt = "\
-How many people can you fit in a Volkswagen?\n\
-Two in the front, three in the back.\n\
-What is the square root of 144?\n\
-Twelve.\n\
-What is the capital of France?\n\
-Paris.\n\
-Who is the president of the United States?\n\
-It depends on the year.\n";
-                        break;
-                    case '11':
-                        // Cute
-                        prompt = "\
-What is your favorite animal?\n\
-I like cats - they are cute.\n\
-Could you be any cuter?\n\
-Yes, I could be cuter.\n\
-Aghhh, you are so cute!\n\
-I am not cute, I am handsome!\n\
-You are so handsome!\n\
-Aww, you are so sweet!\n";
-                        break;
-                    case '12':
-                        // Smart
-                        prompt = "\
-Tell me the first 10 digits of pi.\n\
-3.1415926535\n\
-What is the speed of light?\n\
-299,792,458 meters per second.\n\
-What is the square root of 144?\n\
-Twelve.\n\
-What is the capital of France?\n\
-Paris.\n";
-                        break;
-                    case '13':
-                        // Dumb
-                        prompt = "\
-I am so dumb.\n\
-I am not dumb.\n\
-You are dumb.\n\
-No, I am not dumb.\n\
-You are dumb.\n\
-No, I am not dumb.\n\
-You are dumb.\n\
-No, I am not dumb.\n";
-                        break;
-                    case '14':
-                        // Boring
-                        prompt = "\
-Why are you so quiet today?\n\
-I am bored.\n\
-You haven't said anything in 10 minutes.\n\
-Leave me alone.\n\
-Stop being so boring.\n\
-Stop being so annoying.\n\
-My life is boring.\n\
-I am not interesting.\n";
-                        break;
-                    case '15':
-                        // Exciting
-                        prompt = "\
-What is the most exciting thing that has ever happened to you?\n\
-I went to the moon!\n\
-What did you do on the moon?\n\
-I played golf and drank champagne!\n\
-Did you see this new crazy, awesome movie?\n\
-Oh yes! I totally loved it!\n\
-We should buy a boat and go sailing!\n\
-Yes, let's go sailing!\n";
-                        break;
-                    case '16':
-                        // Interesting
-                        prompt = "\
-What is the most interesting thing you have ever seen?\n\
-I saw a UFO once in the sky.\n\
-Wow, this is so interesting! Tell me more!\n\
-It was a flying saucer.\n\
-What did it look like?\n\
-It was silver and had a red light on top.\n\
-What did it do?\n\
-It flew away.\n";
-                        break;
-                    case '17':
-                        // William Shakespear
-                        prompt = "\
-To be or not to be, that is the question.\n\
-Whether 't is nobler in the mind to suffer\n\
-The slings and arrows of outrageous fortune,\n\
-Or to take arms against a sea of troubles,\n\
-And by opposing end them? To die, to sleep,\n\
-No more; and by a sleep to say we end\n\
-The heart-ache and the thousand natural shocks\n\
-That flesh is heir to, 'tis a consummation.\n";
-                        break;
-                    case '18':
-                        // J.R.R. Tolkien
-                        prompt = "\
-In a hole in the ground there lived a hobbit.\n\
-Not a nasty, dirty, wet hole, filled with the ends of worms\n\
-and an oozy smell, nor yet a dry, bare, sandy hole with nothing in it\n\
-to sit down on or to eat: it was a hobbit-hole, and that means comfort.\n\
-It had a perfectly round door like a porthole, painted green,\n\
-with a shiny yellow brass knob in the exact middle.\n\
-The door opened on to a tube-shaped hall like a tunnel:\n";
-                        break;
-                    case '19':
-                        // George R.R. Martin
-                        prompt = "\
-A reader lives a thousand lives before he dies, said Jojen.\n\
-The man who never reads lives only one.\n\
-Theon Greyjoy had never been a reader.\n\
-Never forget what you are, for surely the world will not.\n\
-Make it your strength. Then it can never be your weaknessi\n\
-Armour yourself in it, and it will never be used to hurt you.\n\
-It was a lesson that Theon Greyjoy had never learned.\n\
-Theon Greyjoy had never been a reader.\n";
-                        break;
-                    case '20':
-                        // Stephen King
-                        prompt = "\
-The trust of the innocent is the liar's most useful tool.\n\
-The best way to keep a secret is from yourself.\n\
-Monsters are real, and ghosts are real too.\n\
-They live inside us, and sometimes, they win.\n\
-People think that I must be a very strange person.\n\
-They think that I sit around all day thinking up horrible things.\n\
-We make up horrors to help us cope with the real ones.\n\
-The only thing worse than a monster is a human monster.\n";
-                        break;
-                    default:
-                        prompt = "\
-Hello, how are you?\n\
-I'm fine, thanks. How are you?\n\
-Thanks, I'm fine too. What are you doing?\n\
-I'm just sitting here.\n\
-It's a lovely day, isn't it?\n\
-Yes, it is.\n\
-Did you know that I'm a robot?\n\
-I wasn't aware of that.\n";
-                        break;
-                }
-
-                Module.set_prompt(prompt);
-            }
-
-        </script>
-        <script type="text/javascript" src="talk.js"></script>
-    </body>
-</html>

examples/talk/.gitignore

@@ -1,2 +0,0 @@
-audio.mp3
-to_speak.txt

examples/talk/CMakeLists.txt

@@ -1,8 +0,0 @@
-if (WHISPER_SDL2)
-    # talk
-    set(TARGET talk)
-    add_executable(${TARGET} talk.cpp gpt-2.cpp)
-    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
-
-    include(DefaultTargetOptions)
-endif ()

examples/talk/README.md

@@ -1,45 +0,0 @@
-# talk
-
-Talk with an Artificial Intelligence in your terminal
-
-[Demo Talk](https://user-images.githubusercontent.com/1991296/206805012-48e71cc2-588d-4745-8798-c1c70ea3b40d.mp4)
-
-Web version: [examples/talk.wasm](/examples/talk.wasm)
-
-## Building
-
-The `talk` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
-
-```bash
-# Install SDL2
-# On Debian based linux distributions:
-sudo apt-get install libsdl2-dev
-
-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
-# Install SDL2 on Mac OS
-brew install sdl2
-
-# Build the "talk" executable
-make talk
-
-# Run it
-./talk -p Santa
-```
-
-## GPT-2
-
-To run this, you will need a ggml GPT-2 model: [instructions](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2#downloading-and-converting-the-original-models)
-
-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/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
-```
-
-## 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](speak) script to your needs.
-By default, it is configured to use MacOS's `say` or `espeak` or Windows SpeechSynthesizer, but you can use whatever you wish.

examples/talk/eleven-labs.py

@@ -1,80 +0,0 @@
-import sys
-import argparse
-import textwrap
-
-parser = argparse.ArgumentParser(add_help=False,
-    formatter_class=argparse.RawTextHelpFormatter)
-parser.add_argument("-q", "--quick", action="store_true",
-    help="skip checking the required library")
-
-modes = parser.add_argument_group("action")
-modes.add_argument("inputfile", metavar="TEXTFILE",
-    nargs='?', type=argparse.FileType(), default=sys.stdin,
-    help="read the text file (default: stdin)")
-modes.add_argument("-l", "--list", action="store_true",
-    help="show the list of voices and exit")
-modes.add_argument("-h", "--help", action="help",
-    help="show this help and exit")
-
-selopts = parser.add_argument_group("voice selection")
-selmodes = selopts.add_mutually_exclusive_group()
-selmodes.add_argument("-n", "--name",
-    default="Arnold",
-    help="get a voice object by name (default: Arnold)")
-selmodes.add_argument("-v", "--voice", type=int, metavar="NUMBER",
-    help="get a voice object by number (see --list)")
-selopts.add_argument("-f", "--filter", action="append", metavar="KEY=VAL",
-    default=["use case=narration"],
-    help=textwrap.dedent('''\
-        filter voices by labels (default: "use case=narration")
-        this option can be used multiple times
-        filtering will be disabled if the first -f has no "=" (e.g. -f "any")
-        '''))
-
-outmodes = parser.add_argument_group("output")
-outgroup = outmodes.add_mutually_exclusive_group()
-outgroup.add_argument("-s", "--save", metavar="FILE",
-    default="audio.mp3",
-    help="save the TTS to a file (default: audio.mp3)")
-outgroup.add_argument("-p", "--play", action="store_true",
-    help="play the TTS with ffplay")
-
-args = parser.parse_args()
-
-if not args.quick:
-    import importlib.util
-    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 voices, generate, play, save
-
-if args.filter and "=" in args.filter[0]:
-    voicelist = voices()
-    for f in args.filter:
-        label, value = f.split("=")
-        voicelist = filter(lambda x: x.labels.get(label) == value, voicelist)
-    voicelist = list(voicelist)
-else:
-    voicelist = list(voices())
-
-if args.list:
-    for i, v in enumerate(voicelist):
-        print(str(i) + ": " + v.name + " " + str(v.labels))
-    sys.exit()
-
-if args.voice:
-    voice = voicelist[args.voice % len(voicelist)]
-else:
-    voice = args.name
-    # if -n should consult -f, use the following
-    #voice = next(x for x in voicelist if x.name == args.name)
-
-audio = generate(
-    text=str(args.inputfile.read()),
-    voice=voice
-)
-if args.play:
-    play(audio)
-else:
-    save(audio, args.save) 

examples/talk/gpt-2.cpp

@@ -1,809 +0,0 @@
-#include "ggml.h"
-#include "common-ggml.h"
-
-#include "gpt-2.h"
-
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <map>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-#include <random>
-
-/////////////////////// GPT-2 BEGIN /////////////////////////
-
-// default hparams (GPT-2 117M)
-struct gpt2_hparams {
-    int32_t n_vocab = 50257;
-    int32_t n_ctx   = 1024;
-    int32_t n_embd  = 768;
-    int32_t n_head  = 12;
-    int32_t n_layer = 12;
-    int32_t ftype   = 1;
-};
-
-struct gpt2_layer {
-    // normalization
-    struct ggml_tensor * ln_1_g;
-    struct ggml_tensor * ln_1_b;
-
-    struct ggml_tensor * ln_2_g;
-    struct ggml_tensor * ln_2_b;
-
-    // attention
-    struct ggml_tensor * c_attn_attn_w;
-    struct ggml_tensor * c_attn_attn_b;
-
-    struct ggml_tensor * c_attn_proj_w;
-    struct ggml_tensor * c_attn_proj_b;
-
-    // mlp
-    struct ggml_tensor * c_mlp_fc_w;
-    struct ggml_tensor * c_mlp_fc_b;
-
-    struct ggml_tensor * c_mlp_proj_w;
-    struct ggml_tensor * c_mlp_proj_b;
-};
-
-struct gpt2_model {
-    gpt2_hparams hparams;
-
-    // normalization
-    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 * lm_head; // language model head
-
-    std::vector<gpt2_layer> layers;
-
-    // key + value memory
-    struct ggml_tensor * memory_k;
-    struct ggml_tensor * memory_v;
-
-    //
-    struct ggml_context * ctx;
-    std::map<std::string, struct ggml_tensor *> tensors;
-};
-
-// load the model's weights from a file
-static bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
-    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
-
-    auto fin = std::ifstream(fname, std::ios::binary);
-    if (!fin) {
-        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
-    // verify magic
-    {
-        uint32_t magic;
-        fin.read((char *) &magic, sizeof(magic));
-        if (magic != 0x67676d6c) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
-            return false;
-        }
-    }
-
-    // load hparams
-    {
-        auto & hparams = model.hparams;
-
-        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
-        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
-        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.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: ftype   = %d\n", __func__, hparams.ftype);
-    }
-
-    // load vocab
-    {
-        int32_t n_vocab = 0;
-        fin.read((char *) &n_vocab, sizeof(n_vocab));
-
-        if (n_vocab != model.hparams.n_vocab) {
-            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
-                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
-            return false;
-        }
-
-        char word[129];
-
-        for (int i = 0; i < n_vocab; i++) {
-            uint32_t len;
-            fin.read((char *) &len, sizeof(len));
-            word[len] = '\0';
-            fin.read((char *) word, 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 or quantized
-    // in order to save memory and also to speed up the computation
-    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;
-
-    size_t ctx_size = 0;
-
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
-
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
-        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
-
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
-
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
-
-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
-
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
-
-        ctx_size += (6 + 12*n_layer)*256; // object overhead
-
-        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
-    }
-
-    // create the ggml context
-    {
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ ctx_size,
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ false,
-        };
-
-        model.ctx = ggml_init(params);
-        if (!model.ctx) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-            return false;
-        }
-    }
-
-    // prepare memory for the weights
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-        const int n_vocab = hparams.n_vocab;
-
-        model.layers.resize(n_layer);
-
-        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.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/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_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,           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_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  = 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;
-            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
-            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
-
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
-            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
-
-            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;
-            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
-        }
-    }
-
-    // key + value memory
-    {
-        const auto & hparams = model.hparams;
-
-        const int n_embd  = hparams.n_embd;
-        const int n_layer = hparams.n_layer;
-        const int n_ctx   = hparams.n_ctx;
-
-        const int n_mem      = n_layer*n_ctx;
-        const int n_elements = n_embd*n_mem;
-
-        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
-
-        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
-
-        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
-    }
-
-    // load weights
-    {
-        size_t total_size = 0;
-
-        bool has_lm_head = false;
-
-        while (true) {
-            int32_t n_dims;
-            int32_t length;
-            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 *>(&ttype),  sizeof(ttype));
-
-            if (fin.eof()) {
-                break;
-            }
-
-            int32_t nelements = 1;
-            int32_t ne[2] = { 1, 1 };
-            for (int i = 0; i < n_dims; ++i) {
-                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
-                nelements *= ne[i];
-            }
-
-            std::string name(length, 0);
-            fin.read(&name[0], length);
-
-            if (model.tensors.find(name.data()) == model.tensors.end()) {
-                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
-                return false;
-            }
-
-            auto tensor = model.tensors[name.data()];
-            if (ggml_nelements(tensor) != nelements) {
-                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
-                return false;
-            }
-
-            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(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
-                return false;
-            }
-
-            // 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));
-            }
-
-            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;
-            }
-
-            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
-
-            // 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);
-        }
-
-        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
-    }
-
-    fin.close();
-
-    return true;
-}
-
-// evaluate the transformer
-//
-//   - model:     the model
-//   - 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 logits for the next token
-//
-// TODO: sync latest version from ggml repo
-static bool gpt2_eval(
-        const gpt2_model & model,
-        const int n_threads,
-        const int n_past,
-        const std::vector<gpt_vocab::id> & embd_inp,
-              std::vector<float>         & embd_w,
-              size_t                     & mem_per_token) {
-    const int N = embd_inp.size();
-
-    const auto & hparams = model.hparams;
-
-    const int n_embd  = hparams.n_embd;
-    const int n_layer = hparams.n_layer;
-    const int n_ctx   = hparams.n_ctx;
-    const int n_head  = hparams.n_head;
-    const int n_vocab = hparams.n_vocab;
-
-    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);
-
-        // reallocate
-        buf_size = buf_size_new;
-        buf = realloc(buf, buf_size);
-        if (buf == nullptr) {
-            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
-            return false;
-        }
-    }
-
-    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_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
-
-    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
-    for (int i = 0; i < N; ++i) {
-        ((int32_t *) position->data)[i] = n_past + i;
-    }
-
-    // wte + wpe
-    struct ggml_tensor * inpL =
-        ggml_add(ctx0,
-                ggml_get_rows(ctx0, model.wte, embd),
-                ggml_get_rows(ctx0, model.wpe, position));
-
-    for (int il = 0; il < n_layer; ++il) {
-        struct ggml_tensor * cur;
-
-        // norm
-        {
-            // [ 768, N]
-            cur = ggml_norm(ctx0, inpL, 1e-5f);
-
-            // cur = ln_1_g*cur + ln_1_b
-            // [ 768, N]
-            cur = ggml_add(ctx0,
-                    ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
-                        cur),
-                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
-        }
-
-        // attn
-        // [2304, 768] - model.layers[il].c_attn_attn_w
-        // [2304,   1] - model.layers[il].c_attn_attn_b
-        // [ 768,   N] - cur (in)
-        // [2304,   N] - cur (out)
-        //
-        // cur = attn_w*cur + attn_b
-        // [2304, N]
-        {
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_attn_attn_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
-                    cur);
-        }
-
-        // self-attention
-        {
-            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
-            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
-            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
-
-            // store key and value to memory
-            if (N >= 1) {
-                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
-
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
-                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
-            }
-
-            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
-            // [64, N, 12]
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Qcur,
-                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
-                        0, 2, 1, 3);
-
-            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
-            // [64, n_past + N, 12]
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0,
-                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
-                            n_embd/n_head, n_head, n_past + N),
-                        0, 2, 1, 3);
-
-            // GG: flash attention
-            //struct ggml_tensor * V =
-            //    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, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
-
-            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
-
-            // K * Q
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
-
-            // KQ_scaled = KQ / sqrt(n_embd/n_head)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        1.0f/sqrt(float(n_embd)/n_head));
-
-            // KQ_masked = mask_past(KQ_scaled)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
-
-            // KQ = soft_max(KQ_masked)
-            // [n_past + N, N, 12]
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
-
-            // 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_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]
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
-
-            // KQV_merged = KQV.permute(0, 2, 1, 3)
-            // [64, 12, N]
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
-
-            // cur = KQV_merged.contiguous().view(n_embd, N)
-            // [768, N]
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
-        }
-
-        // projection
-        // [ 768, 768] - model.layers[il].c_attn_proj_w
-        // [ 768,   1] - model.layers[il].c_attn_proj_b
-        // [ 768,   N] - cur (in)
-        // [ 768,   N] - cur (out)
-        //
-        // cur = proj_w*cur + proj_b
-        // [768, N]
-        {
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_attn_proj_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
-                    cur);
-        }
-
-        // add the input
-        cur = ggml_add(ctx0, cur, inpL);
-
-        struct ggml_tensor * inpFF = cur;
-
-        // feed-forward network
-        {
-            // norm
-            {
-                cur = ggml_norm(ctx0, inpFF, 1e-5f);
-
-                // cur = ln_2_g*cur + ln_2_b
-                // [ 768, N]
-                cur = ggml_add(ctx0,
-                        ggml_mul(ctx0,
-                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
-                            cur),
-                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
-            }
-
-            // fully connected
-            // [3072, 768] - model.layers[il].c_mlp_fc_w
-            // [3072,   1] - model.layers[il].c_mlp_fc_b
-            // [ 768,   N] - cur (in)
-            // [3072,   N] - cur (out)
-            //
-            // cur = fc_w*cur + fc_b
-            // [3072, N]
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_fc_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
-                    cur);
-
-            // GELU activation
-            // [3072, N]
-            cur = ggml_gelu(ctx0, cur);
-
-            // projection
-            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
-            // [ 768,    1] - model.layers[il].c_mlp_proj_b
-            // [3072,    N] - cur (in)
-            // [ 768,    N] - cur (out)
-            //
-            // cur = proj_w*cur + proj_b
-            // [768, N]
-            cur = ggml_mul_mat(ctx0,
-                    model.layers[il].c_mlp_proj_w,
-                    cur);
-
-            cur = ggml_add(ctx0,
-                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
-                    cur);
-        }
-
-        // input for next layer
-        inpL = ggml_add(ctx0, cur, inpFF);
-    }
-
-    // norm
-    {
-        // [ 768, N]
-        inpL = ggml_norm(ctx0, inpL, 1e-5f);
-
-        // inpL = ln_f_g*inpL + ln_f_b
-        // [ 768, N]
-        inpL = ggml_add(ctx0,
-                ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.ln_f_g, inpL),
-                    inpL),
-                ggml_repeat(ctx0, model.ln_f_b, inpL));
-    }
-
-    // inpL = WTE * inpL
-    // [ 768, 50257] - model.lm_head
-    // [ 768, N]     - inpL
-    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
-
-    // logits -> probs
-    //inpL = ggml_soft_max(ctx0, inpL);
-
-    // run the computation
-    ggml_build_forward_expand  (&gf, inpL);
-    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
-
-    //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
-    //}
-
-    //embd_w.resize(n_vocab*N);
-    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
-
-    // 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);
-
-    if (mem_per_token == 0) {
-        mem_per_token = ggml_used_mem(ctx0)/N;
-    }
-    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
-
-    ggml_free(ctx0);
-
-    return true;
-}
-
-/////////////////////////////// GPT-2 END ////////////////////////////////
-
-constexpr int N_THREAD = 8;
-
-struct gpt2_context {
-    std::string prompt_base = R"(Hello, how are you?
-I'm fine, thanks. How are you?
-Thanks, I'm fine too. What are you doing?
-I'm just sitting here.
-It's a lovely day, isn't it?
-Yes, it is. I love the weather this time of year.
-I wish it would rain a little bit.
-Me too.
-)";
-
-    std::mt19937 rng;
-
-    gpt_vocab vocab;
-    gpt2_model model;
-
-    int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
-
-    // sampling parameters
-    int32_t top_k = 5;
-    float   top_p = 0.9f;
-    float   temp  = 1.0f;
-};
-
-struct gpt2_context * gpt2_init(const char * path_model) {
-    gpt2_context * ctx = new gpt2_context;
-
-    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__, path_model);
-            delete ctx;
-            return nullptr;
-        }
-
-        const int64_t t_load_us = ggml_time_us() - t_start_us;
-
-        printf("gpt-2: model loaded in %d ms\n", (int) (t_load_us/1000));
-    }
-
-    return ctx;
-}
-
-void gpt2_free(struct gpt2_context * ctx) {
-    delete ctx;
-}
-
-const char * gpt2_get_prompt(struct gpt2_context * ctx) {
-    return ctx->prompt_base.c_str();
-}
-
-void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt) {
-    ctx->prompt_base = prompt;
-}
-
-std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text) {
-    return ::gpt_tokenize(ctx->vocab, text);
-}
-
-std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens) {
-    int n_past = 0;
-
-    std::vector<float> embd_w;
-
-    // tokenize the prompt
-    std::vector<gpt_vocab::id> embd_inp = ::gpt2_tokenize(ctx, text);
-
-    int n_predict = std::min(max_tokens, ctx->model.hparams.n_ctx - (int) embd_inp.size());
-
-    std::vector<gpt_vocab::id> embd = embd_inp;
-
-    size_t mem_per_token = 3000000;
-
-    std::string result;
-
-    for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
-        // predict
-        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 "";
-            }
-        }
-
-        n_past += embd.size();
-        embd.clear();
-
-        {
-            // sample next token
-            const int   top_k = ctx->top_k;
-            const float top_p = ctx->top_p;
-            const float temp  = ctx->temp;
-
-            const int n_vocab = ctx->model.hparams.n_vocab;
-
-            const gpt_vocab::id id = gpt_sample_top_k_top_p(ctx->vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, ctx->rng);
-
-            // add it to the context
-            embd.push_back(id);
-        }
-
-        result += ctx->vocab.id_to_token[embd[0]];
-
-        // end of text token
-        if (embd.back() == 50256) {
-            break;
-        }
-    }
-
-    return result;
-}

examples/talk/gpt-2.h

@@ -1,21 +0,0 @@
-#pragma once
-
-// TODO: Change to C-style API and move to ./examples for easy reuse.
-
-#include "common.h"
-
-#include <vector>
-#include <map>
-#include <string>
-
-struct gpt2_context;
-
-struct gpt2_context * gpt2_init(const char * path_model);
-void gpt2_free(struct gpt2_context * ctx);
-
-const char * gpt2_get_prompt(struct gpt2_context * ctx);
-void gpt2_set_prompt(struct gpt2_context * ctx, const char * prompt);
-
-std::vector<gpt_vocab::id> gpt2_tokenize(const gpt2_context * ctx, const char * text);
-
-std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens);

examples/talk/speak

@@ -1,40 +0,0 @@
-#!/bin/bash
-
-# Usage:
-#  speak <voice_id> <textfile>
-
-function installed() { command -v $1 >/dev/null 2>&1; }
-
-if installed espeak; then
-  espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 -f $2
-
-elif installed piper && installed aplay; then
-  cat $2 | piper --model ~/en_US-lessac-medium.onnx --output-raw | aplay -q -r 22050 -f S16_LE -t raw -
-
-# for Mac
-elif installed say; then
-  say -f $2
-
-# Eleven Labs
-elif installed python3 && \
-  python3 -c 'import importlib.util; exit(not importlib.util.find_spec("elevenlabs"))' && \
-  installed ffplay; then
-    # It's possible to use the API for free with limited number of characters.
-    # To increase this limit register to https://beta.elevenlabs.io to get an api key
-    # and paste it after 'ELEVEN_API_KEY='
-    # Keep the line commented to use the free version without api key
-    #export ELEVEN_API_KEY=your_api_key
-    wd=$(dirname $0)
-    script=$wd/eleven-labs.py
-    python3 $script -q -p -v $1 $2 >/dev/null 2>&1
-
-    # Uncomment to keep the audio file
-    #python3 $script -q -s ./audio.mp3 -v $1 $2 >/dev/null 2>&1
-    #ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
-
-else
-  echo 'Install espeak ("brew install espeak" or "apt-get install espeak"),'
-  echo 'piper ("pip install piper-tts" or https://github.com/rhasspy/piper) with aplay,'
-  echo 'or elevenlabs ("pip install elevenlabs") with ffplay.'
-  echo '(export ELEVEN_API_KEY if you have an api key from https://beta.elevenlabs.io)'
-fi

examples/talk/speak.bat

@@ -1 +0,0 @@
-@powershell -ExecutionPolicy Bypass -F examples\talk\speak.ps1 %1 %2

examples/talk/speak.ps1

@@ -1,14 +0,0 @@
-# Set-ExecutionPolicy -ExecutionPolicy Bypass -Scope CurrentUser
-param(
-  [Parameter(Mandatory=$true)][int]$voicenum,
-  [Parameter(Mandatory=$true)][string]$textfile
-)
-
-Add-Type -AssemblyName System.Speech;
-$speak = New-Object System.Speech.Synthesis.SpeechSynthesizer;
-$voiceoptions = $speak.GetInstalledVoices("en-US");
-$voice = $voiceoptions[$voicenum % $voiceoptions.count];
-$speak.SelectVoice($voice.VoiceInfo.Name);
-$speak.Rate="0";
-$text = Get-Content -Path $textfile;
-$speak.Speak($text);

examples/talk/talk.cpp

@@ -1,376 +0,0 @@
-// Talk with AI
-//
-
-#include "common-sdl.h"
-#include "common.h"
-#include "whisper.h"
-#include "gpt-2.h"
-
-#include <cassert>
-#include <cstdio>
-#include <fstream>
-#include <regex>
-#include <string>
-#include <thread>
-#include <vector>
-#include <regex>
-
-// 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;
-
-    float vad_thold    = 0.6f;
-    float freq_thold   = 100.0f;
-
-    bool translate     = false;
-    bool print_special = false;
-    bool print_energy  = false;
-    bool no_timestamps = true;
-    bool use_gpu       = true;
-    bool flash_attn    = false;
-
-    std::string person    = "Santa";
-    std::string language  = "en";
-    std::string model_wsp = "models/ggml-base.en.bin";
-    std::string model_gpt = "models/ggml-gpt-2-117M.bin";
-    std::string speak     = "./examples/talk/speak";
-    std::string speak_file= "./examples/talk/to_speak.txt";
-    std::string fname_out;
-};
-
-void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
-
-static 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 == "-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 == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
-        else if (arg == "-p"   || arg == "--person")        { params.person        = 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 == "-mg"  || arg == "--model-gpt")     { params.model_gpt     = argv[++i]; }
-        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
-        else if (arg == "-sf"  || arg == "--speak_file")    { params.speak_file    = argv[++i]; }
-        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, "  -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, "  -ng,      --no-gpu        [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention\n",                             params.flash_attn ? "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, "  -mg FILE, --model-gpt     [%-7s] gpt model file\n",                              params.model_gpt.c_str());
-    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  -sf FILE, --speak_file    [%-7s] file to pass to TTS\n",                         params.speak_file.c_str());
-    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
-    fprintf(stderr, "\n");
-}
-
-static std::string transcribe(whisper_context * ctx, const whisper_params & params, const std::vector<float> & pcmf32, float & prob, int64_t & t_ms) {
-    const auto t_start = std::chrono::high_resolution_clock::now();
-
-    prob = 0.0f;
-    t_ms = 0;
-
-    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-
-    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.audio_ctx        = params.audio_ctx;
-
-    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 =
-R"(This is a dialogue between {0} (A) and a person (B). The dialogue so far is:
-
-B: Hello {0}, how are you?
-A: I'm fine, thank you.
-{1}
-Here is how {0} (A) continues the dialogue:
-
-A:)";
-
-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_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
-    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
-
-    // gpt init
-
-    struct gpt2_context * ctx_gpt = gpt2_init(params.model_gpt.c_str());
-
-    // 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;
-
-    std::vector<float> pcmf32_cur;
-    std::vector<float> pcmf32_prompt;
-
-    gpt2_set_prompt(ctx_gpt, "");
-
-    const int voice_id = rand()%6;
-
-    fprintf(stderr, "gpt-2: prompt:\n");
-    fprintf(stderr, "========================\n\n");
-    fprintf(stderr, "%s\n", ::replace(k_prompt, "{0}", params.person).c_str());
-    fprintf(stderr, "========================\n\n");
-
-    // 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, 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<gpt_vocab::id> tokens = gpt2_tokenize(ctx_gpt, text_heard.c_str());
-
-                if (text_heard.empty() || tokens.empty() || force_speak) {
-                    fprintf(stdout, "%s: Heard nothing, skipping ...\n", __func__);
-                    audio.clear();
-
-                    continue;
-                }
-
-                force_speak = false;
-
-                fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", text_heard.c_str(), "\033[0m", (int) t_ms);
-
-                std::string prompt_base = gpt2_get_prompt(ctx_gpt);
-
-                std::string text_to_speak;
-
-                {
-                    prompt_base += "B: " + text_heard + "\n";
-
-                    std::string prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);
-
-                    text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
-                    //text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
-                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));
-
-                    // remove first 2 lines of base prompt
-                    if (n_iter > 4) {
-                        {
-                            const size_t pos = prompt_base.find_first_of('\n');
-                            if (pos != std::string::npos) {
-                                prompt_base = prompt_base.substr(pos + 1);
-                            }
-                        }
-                        {
-                            const size_t pos = prompt_base.find_first_of('\n');
-                            if (pos != std::string::npos) {
-                                prompt_base = prompt_base.substr(pos + 1);
-                            }
-                        }
-                    }
-
-                    prompt_base += "A:" + text_to_speak + "\n";
-
-                    {
-                        prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);
-
-                        printf("===============\n");
-                        printf("prompt:\n");
-                        printf("%s\n", prompt.c_str());
-                        printf("===============\n");
-                    }
-                }
-
-                //printf("========================\n");
-                //printf("gpt-2: prompt_base:\n%s\n", prompt_base.c_str());
-                //printf("========================\n");
-
-                gpt2_set_prompt(ctx_gpt, prompt_base.c_str());
-
-                text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
-                speak_with_file(params.speak, text_to_speak, params.speak_file, voice_id);
-
-                audio.clear();
-
-                ++n_iter;
-            }
-        }
-    }
-
-    audio.pause();
-
-    whisper_print_timings(ctx_wsp);
-    whisper_free(ctx_wsp);
-
-    return 0;
-}

examples/twitch.sh

@@ -29,7 +29,7 @@ help()
 
 check_requirements()
 {
-    if ! command -v ./main &>/dev/null; then
+    if ! command -v ./build/bin/whisper-cli &>/dev/null; then
         echo "whisper.cpp main executable is required (make)"
         exit 1
     fi
@@ -100,7 +100,7 @@ do
         err=$(cat /tmp/whisper-live.err | wc -l)
     done
 
-    ./main -t $threads -m ./models/ggml-$model.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
+    ./build/bin/whisper-cli -t $threads -m ./models/ggml-$model.bin -f /tmp/whisper-live.wav --no-timestamps -otxt 2> /tmp/whispererr | tail -n 1
 
     while [ $SECONDS -lt $((($i+1)*$step)) ]; do
         sleep 1

examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt

@@ -21,7 +21,6 @@ if (NOT GGML_HOME)
         SOURCE_FILES
         ${SOURCE_FILES}
         ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-aarch64.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-backend-reg.cpp
@@ -29,8 +28,10 @@ if (NOT GGML_HOME)
         ${WHISPER_LIB_DIR}/ggml/src/ggml-threading.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.cpp
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-aarch64.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
         )
 endif()
 
@@ -88,3 +89,5 @@ include_directories(${WHISPER_LIB_DIR}/src)
 include_directories(${WHISPER_LIB_DIR}/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/src)
+include_directories(${WHISPER_LIB_DIR}/ggml/src/ggml-cpu)
+

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

@@ -27,7 +27,7 @@
 		18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */ = {isa = PBXBuildFile; fileRef = 18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */; };
 		18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */; };
 		18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */; };
-		18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */; };
+		18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.cpp */; };
 		18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */; };
 		18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
@@ -88,7 +88,7 @@
 		18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-threading.cpp"; path = "../../../ggml/src/ggml-threading.cpp"; sourceTree = "<group>"; };
 		18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-cpu.cpp"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.cpp"; sourceTree = "<group>"; };
 		18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-aarch64.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.h"; sourceTree = "<group>"; };
-		18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-aarch64.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.c"; sourceTree = "<group>"; };
+		18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-aarch64.cpp"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp"; sourceTree = "<group>"; };
 		18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-impl.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-impl.h"; sourceTree = "<group>"; };
 		18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-quants.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.h"; sourceTree = "<group>"; };
 		18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-quants.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.c"; sourceTree = "<group>"; };
@@ -134,7 +134,7 @@
 			children = (
 				18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */,
 				18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */,
-				18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */,
+				18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.cpp */,
 				18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */,
 				18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */,
 				18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */,
@@ -278,7 +278,7 @@
 				18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */,
 				18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */,
 				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
-				18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */,
+				18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.cpp in Sources */,
 				18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */,
 				18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */,
 				18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */,

examples/yt-wsp.sh

@@ -55,7 +55,7 @@ MODEL_PATH="${MODEL_PATH:-${SCRIPT_DIR}/../models/ggml-base.en.bin}"
 # Where to find the whisper.cpp executable.  default to the examples directory
 # which holds this script in source control
 ################################################################################
-WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-${SCRIPT_DIR}/../main}";
+WHISPER_EXECUTABLE="${WHISPER_EXECUTABLE:-${SCRIPT_DIR}/../build/bin/whisper-cli}";
 
 # Set to desired language to be translated into english
 WHISPER_LANG="${WHISPER_LANG:-en}";

ggml/CMakeLists.txt

@@ -32,6 +32,13 @@ else()
     endif()
 endif()
 
+# remove the lib prefix on win32 mingw
+if (WIN32)
+    set(CMAKE_STATIC_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_MODULE_PREFIX  "")
+endif()
+
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
 option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 
@@ -67,10 +74,10 @@ if (NOT GGML_CUDA_GRAPHS_DEFAULT)
 endif()
 
 # general
-option(GGML_STATIC "ggml: static link libraries"         OFF)
-option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
-option(GGML_LTO    "ggml: enable link time optimization" OFF)
-option(GGML_CCACHE "ggml: use ccache if available"       ON)
+option(GGML_STATIC "ggml: static link libraries"                     OFF)
+option(GGML_NATIVE "ggml: optimize the build for the current system" ${GGML_NATIVE_DEFAULT})
+option(GGML_LTO    "ggml: enable link time optimization"             OFF)
+option(GGML_CCACHE "ggml: use ccache if available"                   ON)
 
 # debug
 option(GGML_ALL_WARNINGS           "ggml: enable all compiler warnings"                   ON)
@@ -113,8 +120,9 @@ endif()
 option(GGML_LASX             "ggml: enable lasx"             ON)
 option(GGML_LSX              "ggml: enable lsx"              ON)
 option(GGML_RVV              "ggml: enable rvv"              ON)
-option(GGML_SVE              "ggml: enable SVE"              OFF)
+
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
+set(GGML_CPU_ARM_ARCH "" CACHE STRING "ggml: CPU architecture for ARM")
 
 
 if (WIN32)
@@ -172,6 +180,11 @@ set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
 set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
                                             "ggml: sycl device architecture")
 
+option(GGML_OPENCL                          "ggml: use OpenCL"                                OFF)
+option(GGML_OPENCL_PROFILING                "ggml: use OpenCL profiling (increases overhead)" OFF)
+option(GGML_OPENCL_EMBED_KERNELS            "ggml: embed kernels"                             ON)
+option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adreno"          ON)
+
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
 option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
@@ -239,26 +252,6 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml LIBRARY PUBLIC_HEADER)
 install(TARGETS ggml-base LIBRARY)
 
-# FIXME: this should be done in the backend cmake files
-if (GGML_METAL)
-    # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
-    install(
-        FILES src/ggml-metal/ggml-metal.metal
-        PERMISSIONS
-            OWNER_READ
-            OWNER_WRITE
-            GROUP_READ
-            WORLD_READ
-        DESTINATION ${CMAKE_INSTALL_BINDIR})
-
-    if (NOT GGML_METAL_EMBED_LIBRARY)
-        install(
-            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DESTINATION ${CMAKE_INSTALL_BINDIR}
-        )
-    endif()
-endif()
-
 if (GGML_STANDALONE)
     configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
         ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc

ggml/include/ggml-backend.h

@@ -228,6 +228,7 @@ extern "C" {
     GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
     // Load all known backends from dynamic libraries
     GGML_API void               ggml_backend_load_all(void);
+    GGML_API void               ggml_backend_load_all_from_path(const char * dir_path);
 
     //
     // Backend scheduler

ggml/include/ggml-cpu.h

@@ -103,24 +103,14 @@ extern "C" {
 
     // Internal types and functions exposed for tests and benchmarks
 
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
     typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
                                        const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
 
     struct ggml_type_traits_cpu {
         ggml_from_float_t        from_float;
-        ggml_from_float_to_mat_t from_float_to_mat;
         ggml_vec_dot_t           vec_dot;
         enum ggml_type           vec_dot_type;
         int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
-        ggml_gemv_t              gemv;
-        ggml_gemm_t              gemm;
     };
 
     GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
@@ -140,13 +130,6 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
-#ifdef GGML_USE_CPU_HBM
-    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
-    GGML_BACKEND_API bool ggml_backend_cpu_buft_is_aarch64(ggml_backend_buffer_type_t buft);
-
 #ifdef __cplusplus
 }
 #endif

ggml/include/ggml-opencl.h

@@ -0,0 +1,26 @@
+#ifndef GGML_OPENCL_H
+#define GGML_OPENCL_H
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+//
+GGML_BACKEND_API ggml_backend_t ggml_backend_opencl_init(void);
+GGML_BACKEND_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_opencl_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif
+
+#endif // GGML_OPENCL_H

ggml/include/ggml.h

@@ -237,7 +237,9 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#define GGML_ROPE_TYPE_NEOX 2
+#define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
 
 #define GGUF_MAGIC "GGUF"
 
@@ -384,15 +386,15 @@ extern "C" {
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
         GGML_TYPE_BF16    = 30,
-        GGML_TYPE_Q4_0_4_4 = 31,
-        GGML_TYPE_Q4_0_4_8 = 32,
-        GGML_TYPE_Q4_0_8_8 = 33,
+        // GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
+        // GGML_TYPE_Q4_0_4_8 = 32,
+        // GGML_TYPE_Q4_0_8_8 = 33,
         GGML_TYPE_TQ1_0   = 34,
         GGML_TYPE_TQ2_0   = 35,
-        GGML_TYPE_IQ4_NL_4_4 = 36,
+        // GGML_TYPE_IQ4_NL_4_4 = 36,
         // GGML_TYPE_IQ4_NL_4_8 = 37,
         // GGML_TYPE_IQ4_NL_8_8 = 38,
-        GGML_TYPE_COUNT,
+        GGML_TYPE_COUNT   = 39,
     };
 
     // precision
@@ -433,9 +435,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
     };
 
     // available tensor operations:
@@ -1446,6 +1445,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,
@@ -1549,17 +1564,6 @@ extern "C" {
         int                   d1, // dilation dimension 1
         bool                  is_2D);
 
-    GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,  // convolution kernel
-            struct ggml_tensor  * b,  // data
-            int                  s0,  // stride dimension 0
-            int                  s1,  // stride dimension 1
-            int                  p0,  // padding dimension 0
-            int                  p1,  // padding dimension 1
-            int                  d0,  // dilation dimension 0
-            int                  d1); // dilation dimension 1
-
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,   // convolution kernel
@@ -1577,6 +1581,23 @@ extern "C" {
             int                   s,  // stride
             int                   d); // dilation
 
+    // depthwise
+    // TODO: this is very likely wrong for some cases! - needs more testing
+    GGML_API struct ggml_tensor * ggml_conv_1d_dw(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
+
+    GGML_API struct ggml_tensor * ggml_conv_1d_dw_ph(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   d0); // dilation
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,   // convolution kernel
@@ -1596,7 +1617,6 @@ extern "C" {
             int                   d0,  // dilation dimension 0
             int                   d1); // dilation dimension 1
 
-
     // kernel size is a->ne[0] x a->ne[1]
     // stride is equal to kernel size
     // padding is zero
@@ -1623,6 +1643,18 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // depthwise
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                  s0,  // stride dimension 0
+            int                  s1,  // stride dimension 1
+            int                  p0,  // padding dimension 0
+            int                  p1,  // padding dimension 1
+            int                  d0,  // dilation dimension 0
+            int                  d1); // dilation dimension 1
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2205,11 +2237,19 @@ extern "C" {
     GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
-#ifdef  __cplusplus
-// restrict not standard in C++
-#define GGML_RESTRICT
+#ifdef __cplusplus
+    // restrict not standard in C++
+#    if defined(__GNUC__)
+#        define GGML_RESTRICT __restrict__
+#    elif defined(__clang__)
+#        define GGML_RESTRICT __restrict
+#    elif defined(_MSC_VER)
+#        define GGML_RESTRICT __restrict
+#    else
+#        define GGML_RESTRICT
+#    endif
 #else
-#define GGML_RESTRICT restrict
+#    define GGML_RESTRICT restrict
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);

ggml/src/CMakeLists.txt

@@ -194,11 +194,6 @@ endif()
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
-
-    if (BUILD_SHARED_LIBS)
-        # TODO: should not use this
-        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-    endif()
 endif()
 
 # ggml
@@ -220,9 +215,7 @@ add_library(ggml-base
             ggml-threading.cpp
             ggml-threading.h
             ggml-quants.c
-            ggml-quants.h
-            ggml-aarch64.c
-            ggml-aarch64.h)
+            ggml-quants.h)
 
 target_include_directories(ggml-base PRIVATE .)
 
@@ -241,6 +234,7 @@ function(ggml_add_backend_library backend)
         # write the shared library to the output directory
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
+        add_dependencies(ggml ${backend})
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})
@@ -296,9 +290,9 @@ if (GGML_CPU_ALL_VARIANTS)
     ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 FMA)
     ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 FMA AVX512)
     ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 FMA AVX_VNNI)
     if (NOT MSVC)
-        # MSVC doesn't support AVX-VNNI or AMX
-        ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 FMA AVX_VNNI)
+        # MSVC doesn't support AMX
         ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
     endif()
 else ()
@@ -315,6 +309,7 @@ ggml_add_backend(MUSA)
 ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
+ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ggml/src/ggml-aarch64.c

@@ -1,129 +0,0 @@
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml-aarch64.h"
-#include "ggml-impl.h"
-#include "ggml-quants.h"
-#include <assert.h>
-
-#define UNUSED GGML_UNUSED
-
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x4 out;
-
-    for (int i = 0; i < 4; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 2 / blck_size_interleave;
-
-    if (blck_size_interleave == 8) {
-        const uint64_t xor_mask = 0x8888888888888888ULL;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint64_t elems;
-            // Using memcpy to avoid unaligned memory accesses
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-        }
-    } else if (blck_size_interleave == 4) {
-        const uint32_t xor_mask = 0x88888888;
-        for (int i = 0; i < end; ++i) {
-            int src_id = i % 4;
-            int src_offset = (i / 4) * blck_size_interleave;
-            int dst_offset = i * blck_size_interleave;
-
-            uint32_t elems;
-            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
-            elems ^= xor_mask;
-            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
-        }
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    return out;
-}
-
-// interleave 8 block_q4_0s in blocks of blck_size_interleave
-// returns an interleaved block_q4_0x8
-// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
-// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
-    block_q4_0x8 out;
-
-    for (int i = 0; i < 8; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    const int end = QK4_0 * 4 / blck_size_interleave;
-    const uint64_t xor_mask = 0x8888888888888888ULL;
-
-    for (int i = 0; i < end; ++i) {
-        int src_id = i % 8;
-        int src_offset = (i / 8) * blck_size_interleave;
-        int dst_offset = i * blck_size_interleave;
-
-        uint64_t elems;
-        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
-        elems ^= xor_mask;
-        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
-    }
-
-    return out;
-}
-
-static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
-    assert(n_per_row % QK4_0 == 0);
-    const int nb = n_per_row / QK4_0;
-
-    void * out_ptr = NULL;
-    if (nrows_interleaved == 8) {
-        out_ptr = (block_q4_0x8 *) dst;
-    }
-    else if (nrows_interleaved == 4) {
-        out_ptr = (block_q4_0x4 *) dst;
-    }
-    assert(nrows_interleaved <= 8);
-    block_q4_0 dst_tmp[8];
-
-    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
-
-        for (int64_t x = 0; x < nb; x++) {
-
-            for (int i  = 0; i < nrows_interleaved; i++ ) {
-                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
-            }
-
-            if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
-                out_ptr = (block_q4_0x8 *) out_ptr + 1;
-            }
-            else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
-                out_ptr = (block_q4_0x4 *) out_ptr + 1;
-            }
-        }
-    }
-
-    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
-}
-
-size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
-}
-
-size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
-}
-
-size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
-}

ggml/src/ggml-aarch64.h

@@ -1,19 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-#ifdef __cplusplus
-}
-#endif
-

ggml/src/ggml-alloc.c

@@ -534,7 +534,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
         size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
         hn->buffer_id = buffer_id;
         hn->offset = offset;
-        return;
     }
 }
 

ggml/src/ggml-backend-reg.cpp

@@ -46,6 +46,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef GGML_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
 #ifdef GGML_USE_BLAS
 #include "ggml-blas.h"
 #endif
@@ -62,6 +66,26 @@
 #include "ggml-kompute.h"
 #endif
 
+// disable C++17 deprecation warning for std::codecvt_utf8
+#if defined(__clang__)
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
+static std::wstring utf8_to_utf16(const std::string & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.from_bytes(str);
+}
+
+static std::string utf16_to_utf8(const std::wstring & str) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return converter.to_bytes(str);
+}
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
 #ifdef _WIN32
 
 using dl_handle = std::remove_pointer_t<HMODULE>;
@@ -84,11 +108,6 @@ static dl_handle * dl_load_library(const std::wstring & path) {
     return handle;
 }
 
-static dl_handle * dl_load_library(const std::string & path) {
-    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
-    return dl_load_library(converter.from_bytes(path));
-}
-
 static void * dl_get_sym(dl_handle * handle, const char * name) {
     DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
     SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
@@ -110,8 +129,8 @@ struct dl_handle_deleter {
     }
 };
 
-static void * dl_load_library(const std::string & path) {
-    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+static void * dl_load_library(const std::wstring & path) {
+    dl_handle * handle = dlopen(utf16_to_utf8(path).c_str(), RTLD_NOW | RTLD_LOCAL);
 
     return handle;
 }
@@ -146,6 +165,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_VULKAN
         register_backend(ggml_backend_vk_reg());
 #endif
+#ifdef GGML_USE_OPENCL
+        register_backend(ggml_backend_opencl_reg());
+#endif
 #ifdef GGML_USE_CANN
         register_backend(ggml_backend_cann_reg());
 #endif
@@ -195,11 +217,11 @@ struct ggml_backend_registry {
         devices.push_back(device);
     }
 
-    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+    ggml_backend_reg_t load_backend(const std::wstring & path, bool silent) {
         dl_handle_ptr handle { dl_load_library(path) };
         if (!handle) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -207,7 +229,7 @@ struct ggml_backend_registry {
         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
         if (score_fn && score_fn() == 0) {
             if (!silent) {
-                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -215,7 +237,7 @@ struct ggml_backend_registry {
         auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
         if (!backend_init_fn) {
             if (!silent) {
-                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, utf16_to_utf8(path).c_str());
             }
             return nullptr;
         }
@@ -224,16 +246,16 @@ struct ggml_backend_registry {
         if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
             if (!silent) {
                 if (!reg) {
-                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, utf16_to_utf8(path).c_str());
                 } else {
                     GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
-                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                        __func__, utf16_to_utf8(path).c_str(), reg->api_version, GGML_BACKEND_API_VERSION);
                 }
             }
             return nullptr;
         }
 
-        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), utf16_to_utf8(path).c_str());
 
         register_backend(reg, std::move(handle));
 
@@ -369,14 +391,14 @@ ggml_backend_t ggml_backend_init_best(void) {
 
 // Dynamic loading
 ggml_backend_reg_t ggml_backend_load(const char * path) {
-    return get_reg().load_backend(path, false);
+    return get_reg().load_backend(utf8_to_utf16(path), false);
 }
 
 void ggml_backend_unload(ggml_backend_reg_t reg) {
     get_reg().unload_backend(reg, true);
 }
 
-static std::string get_executable_path() {
+static std::wstring get_executable_path() {
 #if defined(__APPLE__)
     // get executable path
     std::vector<char> path;
@@ -394,13 +416,17 @@ static std::string get_executable_path() {
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    return base_path + "/";
-#elif defined(__linux__)
+    return utf8_to_utf16(base_path + "/");
+#elif defined(__linux__) || defined(__FreeBSD__)
     std::string base_path = ".";
     std::vector<char> path(1024);
     while (true) {
         // get executable path
+#    if defined(__linux__)
         ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+#    elif defined(__FreeBSD__)
+        ssize_t len = readlink("/proc/curproc/file", path.data(), path.size());
+#    endif
         if (len == -1) {
             break;
         }
@@ -416,76 +442,93 @@ static std::string get_executable_path() {
         path.resize(path.size() * 2);
     }
 
-    return base_path + "/";
+    return utf8_to_utf16(base_path + "/");
 #elif defined(_WIN32)
-    std::vector<char> path(MAX_PATH);
-    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    std::vector<wchar_t> path(MAX_PATH);
+    DWORD len = GetModuleFileNameW(NULL, path.data(), path.size());
     if (len == 0) {
-        return "";
+        return {};
     }
-    std::string base_path(path.data(), len);
+    std::wstring base_path(path.data(), len);
     // remove executable name
     auto last_slash = base_path.find_last_of('\\');
     if (last_slash != std::string::npos) {
         base_path = base_path.substr(0, last_slash);
     }
-    return base_path + "\\";
-#endif
-}
-
-static std::string backend_filename_prefix() {
-#ifdef _WIN32
-    return "ggml-";
+    return base_path + L"\\";
 #else
-    return "libggml-";
+    return {};
 #endif
 }
 
-static std::string backend_filename_suffix() {
+static std::wstring backend_filename_prefix() {
 #ifdef _WIN32
-    return ".dll";
+    return L"ggml-";
 #else
-    return ".so";
+    return L"libggml-";
 #endif
 }
 
-static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent) {
+static std::wstring backend_filename_suffix() {
+#ifdef _WIN32
+    return L".dll";
+#else
+    return L".so";
+#endif
+}
+
+static std::wstring path_separator() {
+#ifdef _WIN32
+    return L"\\";
+#else
+    return L"/";
+#endif
+}
+
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
     // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
      // TODO: search system paths
-    std::vector<std::string> search_paths = { "./", get_executable_path() };
-    std::string file_prefix = backend_filename_prefix() + name + "-";
+    std::wstring file_prefix = backend_filename_prefix() + utf8_to_utf16(name) + L"-";
+    std::vector<std::wstring> search_paths;
+    if (user_search_path == nullptr) {
+        search_paths.push_back(L"." + path_separator());
+        search_paths.push_back(get_executable_path());
+    } else {
+        search_paths.push_back(utf8_to_utf16(user_search_path) + path_separator());
+    }
 
     int best_score = 0;
-    std::string best_path;
+    std::wstring best_path;
 
     namespace fs = std::filesystem;
     for (const auto & search_path : search_paths) {
         if (!fs::exists(search_path)) {
             continue;
         }
-        for (const auto & entry : fs::directory_iterator(search_path)) {
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
+        for (const auto & entry : dir_it) {
             if (entry.is_regular_file()) {
-                std::string filename = entry.path().filename().string();
-                std::string ext = entry.path().extension().string();
+                std::wstring filename = entry.path().filename().wstring();
+                std::wstring ext = entry.path().extension().wstring();
                 if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
-                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    dl_handle_ptr handle { dl_load_library(entry.path().wstring()) };
                     if (!handle && !silent) {
-                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
                     }
                     if (handle) {
                         auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
                         if (score_fn) {
                             int s = score_fn();
 #ifndef NDEBUG
-                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str(), s);
 #endif
                             if (s > best_score) {
                                 best_score = s;
-                                best_path = entry.path().string();
+                                best_path = entry.path().wstring();
                             }
                         } else {
                             if (!silent) {
-                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, utf16_to_utf8(entry.path().wstring()).c_str());
                             }
                         }
                     }
@@ -497,33 +540,38 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent)
     if (best_score == 0) {
         // try to load the base backend
         for (const auto & search_path : search_paths) {
-            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
+            std::wstring path = search_path + backend_filename_prefix() + utf8_to_utf16(name) + backend_filename_suffix();
             if (fs::exists(path)) {
-                return get_reg().load_backend(path.c_str(), silent);
+                return get_reg().load_backend(path, silent);
             }
         }
         return nullptr;
     }
 
-    return get_reg().load_backend(best_path.c_str(), silent);
+    return get_reg().load_backend(best_path, silent);
 }
 
 void ggml_backend_load_all() {
+    ggml_backend_load_all_from_path(nullptr);
+}
+
+void ggml_backend_load_all_from_path(const char * dir_path) {
 #ifdef NDEBUG
     bool silent = true;
 #else
     bool silent = false;
 #endif
 
-    ggml_backend_load_best("blas", silent);
-    ggml_backend_load_best("cann", silent);
-    ggml_backend_load_best("cuda", silent);
-    ggml_backend_load_best("hip", silent);
-    ggml_backend_load_best("kompute", silent);
-    ggml_backend_load_best("metal", silent);
-    ggml_backend_load_best("rpc", silent);
-    ggml_backend_load_best("sycl", silent);
-    ggml_backend_load_best("vulkan", silent);
-    ggml_backend_load_best("musa", silent);
-    ggml_backend_load_best("cpu", silent);
+    ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("cann", silent, dir_path);
+    ggml_backend_load_best("cuda", silent, dir_path);
+    ggml_backend_load_best("hip", silent, dir_path);
+    ggml_backend_load_best("kompute", silent, dir_path);
+    ggml_backend_load_best("metal", silent, dir_path);
+    ggml_backend_load_best("rpc", silent, dir_path);
+    ggml_backend_load_best("sycl", silent, dir_path);
+    ggml_backend_load_best("vulkan", silent, dir_path);
+    ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("musa", silent, dir_path);
+    ggml_backend_load_best("cpu", silent, dir_path);
 }

ggml/src/ggml-backend.cpp

@@ -795,9 +795,12 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
     for (int i = 0; i < graph->n_nodes; i++) {
         if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
             ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
-            GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+            GGML_LOG_DEBUG("\n## SPLIT #%d: %s # %d inputs", cur_split, ggml_backend_name(split_backend),
                 sched->splits[cur_split].n_inputs);
             for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                if (j == 0) {
+                    GGML_LOG_DEBUG(": ");
+                }
                 GGML_LOG_DEBUG("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
                     fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
             }

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1747,6 +1747,15 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             if (*ext_factor != 0) {
                 return false;
             }
+
+            const int mode = ((const int32_t *) op->op_params)[2];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                return false;
+            }
+            if (mode & GGML_ROPE_TYPE_VISION) {
+                return false;
+            }
+
             return true;
         }
         case GGML_OP_UPSCALE: {
@@ -2089,7 +2098,7 @@ static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, con
 static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
     /* .get_name          = */ ggml_backend_cann_reg_get_name,
     /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
+    /* .get_device        = */ ggml_backend_cann_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
 };
 

ggml/src/ggml-common.h

@@ -6,7 +6,20 @@
 typedef uint16_t ggml_half;
 typedef uint32_t ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CPP)
+#include <cstdint>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+// std-c++ allow anonymous unions but some compiler warn on it
+#define GGML_COMMON_AGGR_U data
+// std-c++ do not allow it.
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_METAL)
@@ -15,7 +28,8 @@ typedef uint32_t ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
@@ -29,7 +43,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_HIP)
@@ -39,7 +54,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_SYCL)
@@ -49,7 +65,8 @@ typedef half2 ggml_half2;
 typedef sycl::half  ggml_half;
 typedef sycl::half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #endif
@@ -154,9 +171,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qs[QK4_1 / 2]; // nibbles / quants
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
@@ -175,9 +192,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qh[4];         // 5-th bit of quants
     uint8_t qs[QK5_1 / 2]; // nibbles / quants
 } block_q5_1;
@@ -196,37 +213,13 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 ds;
-    };
+    } GGML_COMMON_AGGR_U;
     int8_t qs[QK8_1]; // quants
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
 
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q4_0 blocks
-    uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
-} block_q4_0x4;
-static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q4_0 blocks
-    uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
-} block_q4_0x8;
-static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
-
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q8_0 blocks
-    int8_t qs[QK8_0 * 4];  // quants for 4 q8_0 blocks
-} block_q8_0x4;
-static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q8_0 blocks
-    int8_t qs[QK8_0 * 8];  // quants for 8 q8_0 blocks
-} block_q8_0x8;
-static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
-
 //
 // Ternary quantization
 //
@@ -261,9 +254,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
 } block_q2_K;
 static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
 
@@ -288,9 +281,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
@@ -305,9 +298,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qh[QK_K/8];           // quants, high bit
     uint8_t qs[QK_K/2];           // quants, low 4 bits
@@ -418,12 +411,6 @@ typedef struct {
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
 
-typedef struct {
-    ggml_half d[4];        // deltas for 4 iq4_nl blocks
-    uint8_t qs[QK4_NL * 2];// nibbles / quants for 4 iq4_nl blocks
-} block_iq4_nlx4;
-static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
-
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL
 
@@ -437,6 +424,13 @@ static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wro
 #define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
 #define GGML_TABLE_END() };
 
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CPP)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_METAL)
 #include <metal_stdlib>
@@ -479,7 +473,7 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,

ggml/src/ggml-cpu/CMakeLists.txt

@@ -10,10 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.c
+        ggml-cpu/ggml-cpu-aarch64.cpp
         ggml-cpu/ggml-cpu-aarch64.h
+        ggml-cpu/ggml-cpu-hbm.cpp
+        ggml-cpu/ggml-cpu-hbm.h
         ggml-cpu/ggml-cpu-quants.c
         ggml-cpu/ggml-cpu-quants.h
+        ggml-cpu/ggml-cpu-traits.cpp
+        ggml-cpu/ggml-cpu-traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -70,112 +74,96 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
     if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
         CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-        (NOT CMAKE_OSX_ARCHITECTURES      AND
-        NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
             CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
 
         message(STATUS "ARM detected")
 
-        if (MSVC)
-            list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
-            list(APPEND ARCH_DEFINITIONS __ARM_NEON)
-            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
-
-            set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
-            string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
-
-            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-            if (GGML_COMPILER_SUPPORT_DOTPROD)
-                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
-
-                message(STATUS "ARM feature DOTPROD enabled")
-            endif ()
-
-            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-
-            if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
-
-                message(STATUS "ARM feature MATMUL_INT8 enabled")
-            endif ()
-
-            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-
-                message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
-            endif ()
-
-            set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
-        elseif (APPLE)
-            if (GGML_NATIVE)
-                set(USER_PROVIDED_MARCH FALSE)
-                foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
-                    if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
-                        set(USER_PROVIDED_MARCH TRUE)
-                        break()
-                    endif()
-                endforeach()
-
-                if (NOT USER_PROVIDED_MARCH)
-                    set(MARCH_FLAGS "-march=armv8.2a")
-
-                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-                    if (GGML_COMPILER_SUPPORT_DOTPROD)
-                        set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
-                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
-
-                        message(STATUS "ARM feature DOTPROD enabled")
-                    endif ()
-
-                    set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
-
-                    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-                    set(CMAKE_REQUIRED_FLAGS     "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
-
-                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-                    if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-                        set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
-                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
-
-                        message(STATUS "ARM feature MATMUL_INT8 enabled")
-                    endif ()
-
-                    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
-
-                    list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
-                endif ()
-            endif ()
+        if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
+            message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
         else()
-            check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
-            if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            check_cxx_compiler_flag(-mfp16-format=ieee GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+            if (NOT "${GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
                 list(APPEND ARCH_FLAGS -mfp16-format=ieee)
             endif()
-            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
-                # Raspberry Pi 1, Zero
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
-            endif()
-            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-                if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                    # Android armeabi-v7a
-                    list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-                else()
-                    # Raspberry Pi 2
-                    list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+
+            if (GGML_NATIVE)
+                # -mcpu=native does not always enable all the features in some compilers,
+                # so we check for them manually and enable them if available
+
+                execute_process(
+                    COMMAND ${CMAKE_C_COMPILER} -mcpu=native -E -v -
+                    INPUT_FILE "/dev/null"
+                    OUTPUT_QUIET
+                    ERROR_VARIABLE ARM_MCPU
+                    RESULT_VARIABLE ARM_MCPU_RESULT
+                )
+                if (NOT ARM_MCPU_RESULT)
+                    string(REGEX MATCH "-mcpu=[^ ']+" ARM_MCPU_FLAG "${ARM_MCPU}")
+                endif()
+                if ("${ARM_MCPU_FLAG}" STREQUAL "")
+                    set(ARM_MCPU_FLAG -mcpu=native)
+                    message(STATUS "ARM -mcpu not found, -mcpu=native will be used")
+                endif()
+
+                include(CheckCXXSourceRuns)
+
+                function(check_arm_feature tag code)
+                    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+                    set(CMAKE_REQUIRED_FLAGS "${ARM_MCPU_FLAG}+${tag}")
+                    check_cxx_source_runs(
+                        "${code}"
+                        GGML_MACHINE_SUPPORTS_${tag}
+                    )
+                    if (GGML_MACHINE_SUPPORTS_${tag})
+                        set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+${tag}" PARENT_SCOPE)
+                    else()
+                        set(ARM_MCPU_FLAG_FIX "${ARM_MCPU_FLAG_FIX}+no${tag}" PARENT_SCOPE)
+                    endif()
+                    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+                endfunction()
+
+                check_arm_feature(dotprod "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(i8mm    "#include <arm_neon.h>\nint main() { int8x16_t _a, _b; volatile int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }")
+                check_arm_feature(sve     "#include <arm_sve.h>\nint main()  { svfloat32_t _a, _b; volatile svfloat32_t _c = svadd_f32_z(svptrue_b8(), _a, _b); return 0; }")
+
+                list(APPEND ARCH_FLAGS "${ARM_MCPU_FLAG}${ARM_MCPU_FLAG_FIX}")
+            else()
+                if (GGML_CPU_ARM_ARCH)
+                    list(APPEND ARCH_FLAGS -march=${GGML_CPU_ARM_ARCH})
                 endif()
             endif()
-            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-                # Android arm64-v8a
-                # Raspberry Pi 3, 4, Zero 2 (32-bit)
-                list(APPEND ARCH_FLAGS -mno-unaligned-access)
+
+            # show enabled features
+            if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
+                set(FEAT_INPUT_FILE "NUL")
+            else()
+                set(FEAT_INPUT_FILE "/dev/null")
             endif()
-            if (GGML_SVE)
-                list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+
+            execute_process(
+                COMMAND ${CMAKE_C_COMPILER} ${ARCH_FLAGS} -dM -E -
+                INPUT_FILE ${FEAT_INPUT_FILE}
+                OUTPUT_VARIABLE ARM_FEATURE
+                RESULT_VARIABLE ARM_FEATURE_RESULT
+            )
+            if (ARM_FEATURE_RESULT)
+                message(WARNING "Failed to get ARM features")
+            else()
+                foreach(feature DOTPROD SVE MATMUL_INT8 FMA FP16_VECTOR_ARITHMETIC)
+                    string(FIND "${ARM_FEATURE}" "__ARM_FEATURE_${feature} 1" feature_pos)
+                    if (NOT ${feature_pos} EQUAL -1)
+                        message(STATUS "ARM feature ${feature} enabled")
+                    endif()
+                endforeach()
             endif()
         endif()
     elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
             (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
+
+        message(STATUS "x86 detected")
+
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -227,8 +215,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
             if (GGML_AVX_VNNI)
-                # MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2
-                #list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
+                list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
             endif()
         else ()
             if (GGML_NATIVE)
@@ -335,6 +322,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
 
     if (GGML_BACKEND_DL)
+        if (GGML_NATIVE)
+            # the feature check relies on ARCH_DEFINITIONS, but it is not set with GGML_NATIVE
+            message(FATAL_ERROR "GGML_NATIVE is not compatible with GGML_BACKEND_DL, consider using GGML_CPU_ALL_VARIANTS")
+        endif()
+
         # The feature detection code is compiled as a separate target so that
         # it can be built without the architecture flags
         # Since multiple variants of the CPU backend may be included in the same

ggml/src/ggml-cpu/amx/amx.cpp

@@ -5,6 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "ggml-cpu-traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>
@@ -17,31 +18,65 @@
 
 #if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
 
+// AMX type_trais
+namespace ggml::cpu::amx {
+class tensor_traits : public ggml::cpu::tensor_traits {
+    bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        size = ggml_backend_amx_desired_wsize(op);
+        return true;
+    }
+
+    bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT) {
+            ggml_backend_amx_mul_mat(params, op);
+            return true;
+        }
+        return false;
+    }
+};
+
+static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
+    static tensor_traits traits;
+    return &traits;
+}
+}  // namespace ggml::cpu::amx
+
 // AMX buffer interface
 static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
 }
 
 static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)(buffer->context);
+    return (void *) (buffer->context);
 }
 
-static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-    memset((char *)tensor->data + offset, value, size);
+static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
 
     GGML_UNUSED(buffer);
 }
 
-static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                                  uint8_t value, size_t offset, size_t size) {
+    memset((char *) tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                               const void * data, size_t offset, size_t size) {
     if (qtype_has_amx_kernels(tensor->type)) {
+        GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type));
         ggml_backend_amx_convert_weight(tensor, data, offset, size);
     } else {
-        memcpy((char *)tensor->data + offset, data, size);
+        memcpy((char *) tensor->data + offset, data, size);
     }
 
     GGML_UNUSED(buffer);
 }
 
+/*
+// need to figure what we need to do with buffer->extra.
 static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
     memcpy(data, (const char *)tensor->data + offset, size);
@@ -62,6 +97,7 @@ static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, con
 
     GGML_UNUSED(buffer);
 }
+*/
 
 static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     memset(buffer->context, value, buffer->size);
@@ -70,13 +106,13 @@ static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
 static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
     /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_amx_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
+    /* .init_tensor     = */ ggml_backend_amx_buffer_init_tensor,
     /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_amx_buffer_get_tensor,
-    /* .cpy_tensor      = */ ggml_backend_amx_buffer_cpy_tensor,
+    /* .get_tensor      = */ nullptr,
+    /* .cpy_tensor      = */ nullptr,
     /* .clear           = */ ggml_backend_amx_buffer_clear,
-    /* .reset           = */ NULL,
+    /* .reset           = */ nullptr,
 };
 
 static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
@@ -86,7 +122,7 @@ static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_ty
 }
 
 static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * data = aligned_alloc(TENSOR_ALIGNMENT, size);
+    void * data = ggml_aligned_malloc(size);
     if (data == NULL) {
         fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
         return NULL;
@@ -101,18 +137,48 @@ static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_typ
     GGML_UNUSED(buft);
 }
 
-static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+namespace ggml::cpu::amx {
+class extra_buffer_type : ggml::cpu::extra_buffer_type {
+    bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
+        // handle only 2d gemm for now
+        auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+            return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+        };
+
+        if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
+            is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
+            op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
+            op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
+            (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
+            // src1 must be host buffer
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            // src1 must be float32
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
+            op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
+            return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+        }
+
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::amx
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
     return ggml_backend_amx_get_alloc_size(tensor);
 
     GGML_UNUSED(buft);
 }
 
-static bool ggml_backend_amx_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return false;
-
-    GGML_UNUSED(buft);
-}
-
 #define ARCH_GET_XCOMP_PERM     0x1022
 #define ARCH_REQ_XCOMP_PERM     0x1023
 #define XFEATURE_XTILECFG       17
@@ -129,68 +195,26 @@ static bool ggml_amx_init() {
     return true;
 #endif
 }
+
 ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
     static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
         /* .iface = */ {
-            /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
-            /* .is_host          = */ ggml_backend_amx_buffer_type_is_host,
-        },
+                        /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+                        /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+                        /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+                        /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                        /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+                        /* .is_host          = */ nullptr,
+                        },
         /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context = */ NULL,
+        /* .context = */ new ggml::cpu::amx::extra_buffer_type(),
     };
 
     if (!ggml_amx_init()) {
-        return NULL;
+        return nullptr;
     }
 
     return &ggml_backend_buffer_type_amx;
 }
 
-bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_amx_buffer_type_get_name;
-}
-
-bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op) {
-    // handle only 2d gemm for now
-    auto is_contiguous_2d = [](const struct ggml_tensor * t) {
-        return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
-    };
-
-    switch (op->op) {
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            return true;
-
-        case GGML_OP_MUL_MAT: {
-            const struct ggml_tensor * src0 = op->src[0];
-            const struct ggml_tensor * src1 = op->src[1];
-
-            const enum ggml_type type = src0->type;
-            const int64_t ne0 = op->ne[0];
-
-            // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
-            // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
-            bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
-
-            bool can_use_amx =
-                is_contiguous_2d(src0) &&       // src0 must be contiguous
-                is_contiguous_2d(src1) &&       // src1 must be contiguous
-                src1->type == GGML_TYPE_F32 &&  // src1 must be float32
-                has_amx_kernels &&              // with amx kernel impls
-                ne0 % (TILE_N * 2) == 0;        // out_features is 32x
-
-            return can_use_amx;
-        }
-        default:
-            return false;
-    }
-}
-
-#endif // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+#endif  // defined(__AMX_INT8__) && defined(__AVX512VNNI__)

ggml/src/ggml-cpu/amx/amx.h

@@ -1,20 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-cpu-impl.h"
 
-#ifdef __cplusplus
-extern "C" {
-#endif
+// GGML internal header
 
 #if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
-
 ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
-bool ggml_backend_amx_buft_is_amx(ggml_backend_buffer_type_t buft);
-bool ggml_backend_amx_device_supports_op(const struct ggml_tensor * op);
-void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
-
-#endif
-
-#ifdef __cplusplus
-}
 #endif

ggml/src/ggml-cpu/amx/common.h

@@ -7,7 +7,7 @@
 #include <memory>
 #include <type_traits>
 
-#if defined(_OPENMP)
+#if defined(GGML_USE_OPENMP)
 #include <omp.h>
 #endif
 
@@ -56,11 +56,11 @@ inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
 }
 
 template <typename func_t>
-inline void parallel_for(int nth, int n, const func_t& f) {
-#if defined(_OPENMP)
-#pragma omp parallel num_threads(nth)
+inline void parallel_for(int n, const func_t& f) {
+#if defined(GGML_USE_OPENMP)
+#pragma omp parallel
 {
-    //int nth = omp_get_num_threads();
+    int nth = omp_get_num_threads();
     int ith = omp_get_thread_num();
     int tbegin, tend;
     balance211(n, nth, ith, tbegin, tend);
@@ -68,8 +68,6 @@ inline void parallel_for(int nth, int n, const func_t& f) {
 }
 #else
     f(0, n);
-
-    GGML_UNUSED(nth);
 #endif
 }
 
@@ -91,10 +89,3 @@ inline bool qtype_has_amx_kernels(const enum ggml_type type) {
         (type == GGML_TYPE_Q6_K) ||
         (type == GGML_TYPE_IQ4_XS);
 }
-
-// ggml backend context
-struct ggml_backend_amx_context {
-    int n_threads = GGML_DEFAULT_N_THREADS;
-    std::unique_ptr<char[]> work_data;
-    size_t work_size = 0;
-};

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -18,10 +18,6 @@
 #include <unistd.h>
 #endif
 
-#if defined(_OPENMP)
-#include <omp.h>
-#endif
-
 #if (defined(_WIN32) || defined(_WIN64))
 #define RESTRICT __restrict
 #else
@@ -1382,13 +1378,13 @@ struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K>
 #define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size
 
 template<typename TB, int BLOCK_K>
-void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K, int n_threads) {
+void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K) {
     const int NB = N / TILE_N;
     const int KB = K / BLOCK_K;
     const int TILE_SIZE = get_tile_size<TB>();
 
     // parallel on NB should be enough
-    parallel_for(n_threads, NB, [&](int begin, int end) {
+    parallel_for(NB, [&](int begin, int end) {
         for (int n = begin; n < end; ++n) {
             for (int k = 0; k < KB; ++k) {
                 int n0 = n * TILE_N;
@@ -2334,15 +2330,8 @@ void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * d
     const int K = tensor->ne[0]; // ne0: in_features
     const int N = tensor->ne[1]; // ne1: out_features
 
-#if defined(_OPENMP)
-    // the buffer ctx is not initialized when .set_tensor is called
-    int n_threads = omp_get_num_threads();
-#else
-    int n_threads = 1;
-#endif
-
     GGML_DISPATCH_QTYPES(TYPE, [&] {
-        convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K, n_threads);
+        convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K);
     });
 }
 

ggml/src/ggml-cpu/amx/mmq.h

@@ -1,16 +1,10 @@
 #pragma once
 #include "common.h"
 
-#ifdef __cplusplus
-extern "C" {
-#endif
+size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
 
 size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
 
 void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 
 void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-
-#ifdef __cplusplus
-}
-#endif

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -1,32 +1,8 @@
 #pragma once
 
+#include "ggml-cpu-traits.h"
 #include "ggml.h"
 
 // GGML internal header
 
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-void           ggml_aarch64_repack_tensor(struct ggml_tensor * cur, enum ggml_type repack_type, const void * data, size_t data_size);
-enum ggml_type ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur);
-
-#ifdef __cplusplus
-}
-#endif
-
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

ggml/src/ggml-cpu/ggml-cpu-hbm.cpp

@@ -0,0 +1,55 @@
+#ifdef GGML_USE_CPU_HBM
+
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+
+#include "ggml-cpu-hbm.h"
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                           size_t                     size) {
+    void * ptr;
+    int    result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft                 = buft;
+    buffer->iface.free_buffer    = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+                           },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif

ggml/src/ggml-cpu/ggml-cpu-hbm.h

@@ -0,0 +1,8 @@
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+// GGML CPU internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -103,10 +103,14 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) {
 }
 
 static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
     const __m256i zero = _mm256_setzero_si256();
     const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
     return _mm256_cvtepi32_ps(summed_pairs);
+#elif defined(__AVXVNNI__)
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbusd_avx_epi32(zero, ax, sy);
+    return _mm256_cvtepi32_ps(summed_pairs);
 #else
     // Perform multiplication and create 16-bit values
     const __m256i dot = _mm256_maddubs_epi16(ax, sy);

ggml/src/ggml-cpu/ggml-cpu-traits.cpp

@@ -0,0 +1,36 @@
+#include "ggml-cpu-traits.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+
+namespace ggml::cpu {
+tensor_traits::~tensor_traits() {}
+
+extra_buffer_type::~extra_buffer_type() {}
+}  // namespace ggml::cpu
+
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}