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
 
@@ -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

.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 1.7.3)
+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: [v1.7.3](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.3) / [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/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": "1.7.3",
+  "version": "1.7.4",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {

bindings/ruby/.gitignore

@@ -1,5 +1,3 @@
 LICENSE
 pkg/
-lib/whisper.so
-lib/whisper.bundle
-lib/whisper.dll
+lib/whisper.*

bindings/ruby/README.md

@@ -60,10 +60,10 @@ You also can use shorthand for pre-converted models:
 whisper = Whisper::Context.new("base.en")
 ```
 
-You can see the list of prepared model names by `Whisper::Model.preconverted_models.keys`:
+You can see the list of prepared model names by `Whisper::Model.pre_converted_models.keys`:
 
 ```ruby
-puts Whisper::Model.preconverted_model_names
+puts Whisper::Model.pre_converted_models.keys
 # tiny
 # tiny.en
 # tiny-q5_1
@@ -87,8 +87,9 @@ whisper = Whisper::Context.new("path/to/your/model.bin")
 Or, you can download model files:
 
 ```ruby
-model_uri = Whisper::Model::URI.new("http://example.net/uri/of/your/model.bin")
-whisper = Whisper::Context.new(model_uri)
+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.
@@ -220,7 +221,18 @@ whisper.each_segment do |segment|
 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/ext/ruby_whisper.cpp

@@ -49,10 +49,14 @@ 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
@@ -187,6 +191,69 @@ 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);
@@ -217,6 +284,17 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   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);
   }
@@ -230,8 +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
@@ -353,80 +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");
@@ -631,6 +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;
@@ -719,6 +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;
@@ -823,6 +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...
  *
@@ -979,19 +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.
@@ -1547,6 +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);
 }
@@ -1719,7 +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";
@@ -1766,6 +1814,7 @@ void Init_whisper() {
   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");
@@ -1809,6 +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);
 
@@ -1832,8 +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);
@@ -1887,6 +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);
@@ -1903,6 +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.rb

@@ -1,2 +0,0 @@
-require "whisper.so"
-require "whisper/model/uri"

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

@@ -1,11 +1,11 @@
-require "whisper.so"
 require "uri"
 require "net/http"
 require "time"
 require "pathname"
 require "io/console/size"
 
-class Whisper::Model
+module Whisper
+  class Model
     class URI
       def initialize(uri)
         @uri = URI(uri)
@@ -79,20 +79,22 @@ class Whisper::Model
             downloaded += chunk.bytesize
             show_progress downloaded, size
           end
+          $stderr.puts
         end
         downloading_path.rename path
       end
 
       def show_progress(current, size)
-      return unless $stderr.tty?
-      return unless size
+        progress_rate_available = size && $stderr.tty?
 
         unless @prev
           @prev = Time.now
-        $stderr.puts "Downloading #{@uri}"
+          $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
@@ -102,7 +104,11 @@ class Whisper::Model
           line = "[#{arrow}] (#{format_bytesize(current)} / #{format_bytesize(size)})"
           padding = ' ' * ($stderr.winsize[1] - line.size)
           $stderr.print "\r#{line}#{padding}"
-      $stderr.puts if current >= size
+        else
+          return if now - @prev < 1
+
+          $stderr.print "."
+        end
         @prev = now
       end
 
@@ -115,8 +121,7 @@ class Whisper::Model
       end
     end
 
-  @pre_converted_models = {}
-  %w[
+    @pre_converted_models = %w[
       tiny
       tiny.en
       tiny-q5_1
@@ -147,11 +152,12 @@ class Whisper::Model
       large-v3-turbo
       large-v3-turbo-q5_0
       large-v3-turbo-q8_0
-  ].each do |name|
-    @pre_converted_models[name] = URI.new("https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-#{name}.bin")
-  end
+    ].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

@@ -4,4 +4,21 @@ require_relative "jfk_reader/jfk_reader"
 
 class TestBase < Test::Unit::TestCase
   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/test_model.rb

@@ -68,4 +68,42 @@ class TestModel < TestBase
     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

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("base.en")
-      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

@@ -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("base.en")
-        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

bindings/ruby/whispercpp.gemspec

@@ -3,8 +3,8 @@ 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']

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")
+    add_subdirectory(cli)
+    add_subdirectory(bench)
+    add_subdirectory(server)
+    add_subdirectory(quantize)
     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(bench)
-    set_target_properties(bench PROPERTIES FOLDER "examples")
-    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)
+
+    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/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
-./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
@@ -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
- ./build/bin/stream -m ./models/ggml-base.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
@@ -43,7 +43,7 @@ brew install sdl2
 cmake -B build -DWHISPER_SDL2=ON
 cmake --build build --config Release
 
-./build/bin/stream
+./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.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.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

@@ -75,7 +75,7 @@ endif()
 
 # general
 option(GGML_STATIC "ggml: static link libraries"                     OFF)
-option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
+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)
 
@@ -120,8 +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)
@@ -251,26 +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.h

@@ -1564,17 +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
@@ -1592,6 +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
@@ -1611,7 +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
@@ -1638,6 +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,

ggml/src/CMakeLists.txt

@@ -234,6 +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})
@@ -289,9 +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)
-    if (NOT MSVC)
-        # MSVC doesn't support AVX-VNNI or AMX
     ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 FMA AVX_VNNI)
+    if (NOT MSVC)
+        # 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 ()

ggml/src/ggml-backend-reg.cpp

@@ -66,6 +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>;
@@ -88,11 +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);
@@ -114,8 +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;
 }
@@ -202,11 +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;
         }
@@ -214,7 +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;
         }
@@ -222,7 +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;
         }
@@ -231,16 +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));
 
@@ -376,14 +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;
@@ -401,13 +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;
         }
@@ -423,57 +442,63 @@ 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 + "\\";
+    return base_path + L"\\";
+#else
+    return {};
 #endif
 }
 
-static std::string backend_filename_prefix() {
+static std::wstring backend_filename_prefix() {
 #ifdef _WIN32
-    return "ggml-";
+    return L"ggml-";
 #else
-    return "libggml-";
+    return L"libggml-";
 #endif
 }
 
-static std::string backend_filename_suffix() {
+static std::wstring backend_filename_suffix() {
 #ifdef _WIN32
-    return ".dll";
+    return L".dll";
 #else
-    return ".so";
+    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::string file_prefix = backend_filename_prefix() + name + "-";
-    std::vector<std::string> search_paths;
+    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("./");
+        search_paths.push_back(L"." + path_separator());
         search_paths.push_back(get_executable_path());
     } else {
-#if defined(_WIN32)
-        search_paths.push_back(std::string(user_search_path) + "\\");
-#else
-        search_paths.push_back(std::string(user_search_path) + "/");
-#endif
+        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) {
@@ -483,27 +508,27 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         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());
                             }
                         }
                     }
@@ -515,15 +540,15 @@ 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/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-cpu/CMakeLists.txt

@@ -74,112 +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)
+
+            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 (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-                if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                    # Android armeabi-v7a
-                    list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+                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()
-                    # Raspberry Pi 2
-                    list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+                        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|amd64)$"))
+
+        message(STATUS "x86 detected")
+
         if (MSVC)
             # instruction set detection for MSVC only
             if (GGML_NATIVE)
@@ -231,8 +215,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 list(APPEND ARCH_DEFINITIONS GGML_SSE42)
             endif()
             if (GGML_AVX_VNNI)
-                # MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2
-                #list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
+                list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
             endif()
         else ()
             if (GGML_NATIVE)
@@ -339,6 +322,11 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     target_compile_definitions(${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/ggml-cpu-aarch64.cpp

@@ -194,9 +194,12 @@ static inline __m256i sum_i16_pairs_int32x8(const __m256i x) {
 }
 
 static inline __m256i mul_sum_us8_pairs_int32x8(const __m256i ax, const __m256i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
     const __m256i zero = _mm256_setzero_si256();
     return _mm256_dpbusd_epi32(zero, ax, sy);
+#elif defined(__AVXVNNI__)
+    const __m256i zero = _mm256_setzero_si256();
+    return _mm256_dpbusd_avx_epi32(zero, ax, sy);
 #else
     // Perform multiplication and create 16-bit values
     const __m256i dot = _mm256_maddubs_epi16(ax, sy);
@@ -647,72 +650,52 @@ static void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, c
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
-    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx;
 
-        __asm__ __volatile__(
-            "movi v2.16b, #0x4\n"
-            "movi v1.16b, #0xf0\n"
-            "add %x[b_ptr], %x[b_ptr], #0x8\n"
-            "1:"  // Column loop
-            "add x23, %x[a_ptr], #0x2\n"
-            "movi v0.16b, #0x0\n"
-            "mov x22, %x[nb]\n"
-            "2:"  // Block loop
-            "ldr q31, [%x[b_ptr], #0x0]\n"
-            "ldr q30, [%x[b_ptr], #0x10]\n"
-            "mov x21, x23\n"
-            "movi v29.4s, #0x0\n"
-            "ldr q28, [%x[b_ptr], #0x20]\n"
-            "ldr q27, [%x[b_ptr], #0x30]\n"
-            "movi v26.4s, #0x0\n"
-            "sub x20, x23, #0x2\n"
-            "ld1r { v25.8h }, [x20]\n"
-            "ldr q24, [%x[b_ptr], #-0x8]\n"
-            "sub x22, x22, #0x1\n"
-            "add x23, x23, #0x22\n"
-            "ld1r { v23.2d }, [x21], #0x8\n"
-            "sshl v22.16b, v31.16b, v2.16b\n"
-            "sshl v16.16b, v30.16b, v2.16b\n"
-            "add %x[b_ptr], %x[b_ptr], #0x48\n"
-            "ld1r { v21.2d }, [x21], #0x8\n"
-            "sshl v20.16b, v28.16b, v2.16b\n"
-            "sshl v19.16b, v27.16b, v2.16b\n"
-            "ld1r { v18.2d }, [x21], #0x8\n"
-            "ld1r { v17.2d }, [x21], #0x8\n"
-            "and v31.16b, v31.16b, v1.16b\n"
-            "and v30.16b, v30.16b, v1.16b\n"
-            ".inst 0x4e9796dd  // sdot v29.4s, v22.16b, v23.16b\n"
-            ".inst 0x4e97961a  // sdot v26.4s, v16.16b, v23.16b\n"
-            "and v28.16b, v28.16b, v1.16b\n"
-            "and v27.16b, v27.16b, v1.16b\n"
-            "fcvtl v25.4s, v25.4h\n"
-            "fcvtl v16.4s, v24.4h\n"
-            ".inst 0x4e95969d  // sdot v29.4s, v20.16b, v21.16b\n"
-            ".inst 0x4e95967a  // sdot v26.4s, v19.16b, v21.16b\n"
-            "fmul v16.4s, v16.4s, v25.4s\n"
-            ".inst 0x4e9297fd  // sdot v29.4s, v31.16b, v18.16b\n"
-            ".inst 0x4e9297da  // sdot v26.4s, v30.16b, v18.16b\n"
-            ".inst 0x4e91979d  // sdot v29.4s, v28.16b, v17.16b\n"
-            ".inst 0x4e91977a  // sdot v26.4s, v27.16b, v17.16b\n"
-            "addp v29.4s, v29.4s, v26.4s\n"
-            "scvtf v29.4s, v29.4s, #0x4\n"
-            "fmla v0.4s, v29.4s, v16.4s\n"
-            "cbnz x22, 2b\n"
-            "sub %x[nc], %x[nc], #0x4\n"
-            "str q0, [%x[res_ptr], #0x0]\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "cbnz %x[nc], 1b\n"
-            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
-            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
-            : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
-        );
+        for (int c = 0; c < nc; c += ncols_interleaved) {
+            const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+            float32x4_t acc = vdupq_n_f32(0);
+            for (int b = 0; b < nb; b++) {
+                int8x16_t b0 = vld1q_s8((const int8_t *) b_ptr->qs);
+                int8x16_t b1 = vld1q_s8((const int8_t *) b_ptr->qs + 16);
+                int8x16_t b2 = vld1q_s8((const int8_t *) b_ptr->qs + 32);
+                int8x16_t b3 = vld1q_s8((const int8_t *) b_ptr->qs + 48);
+                float16x4_t bd = vld1_f16((const __fp16 *) b_ptr->d);
+
+                int8x16_t a0 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs);
+                int8x16_t a1 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 1);
+                int8x16_t a2 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 2);
+                int8x16_t a3 = (int8x16_t) vld1q_dup_s64((const int64_t *) a_ptr->qs + 3);
+                float16x4_t ad = vld1_dup_f16((const __fp16 *) &a_ptr->d);
+
+                int32x4_t ret0 = vdupq_n_s32(0);
+                int32x4_t ret1 = vdupq_n_s32(0);
+
+                ret0 = vdotq_s32(ret0, b0 << 4, a0);
+                ret1 = vdotq_s32(ret1, b1 << 4, a0);
+                ret0 = vdotq_s32(ret0, b2 << 4, a1);
+                ret1 = vdotq_s32(ret1, b3 << 4, a1);
+
+                ret0 = vdotq_s32(ret0, b0 & 0xf0U, a2);
+                ret1 = vdotq_s32(ret1, b1 & 0xf0U, a2);
+                ret0 = vdotq_s32(ret0, b2 & 0xf0U, a3);
+                ret1 = vdotq_s32(ret1, b3 & 0xf0U, a3);
+
+                int32x4_t ret = vpaddq_s32(ret0, ret1);
+
+                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                        vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+                a_ptr++;
+                b_ptr++;
+            }
+            vst1q_f32(s, acc);
+            s += ncols_interleaved;
+        }
         return;
     }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
     float sumf[4];
     int sumi;
 

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.c

@@ -986,7 +986,7 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 #define GGML_F16_STEP 32
 #define GGML_F16_EPR  4
 
-static inline __m128 __sse_f16x4_load(ggml_fp16_t *x) {
+static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
     float tmp[4];
 
     tmp[0] = GGML_FP16_TO_FP32(x[0]);
@@ -7419,14 +7419,14 @@ static void ggml_compute_forward_mul_mat(
     if (src1_cont) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(params,
+                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                      nb01/ggml_type_size(src0->type),
                                      (const char *)src1->data + i12*nb12 + i13*nb13,
                                      nb11/ggml_type_size(src1->type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
-                                     ith, nth,
                                      src0->type,
                                      src1->type,
                                      dst->type))
@@ -7471,14 +7471,14 @@ UseGgmlGemm1:;
 
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                if (!llamafile_sgemm(params,
+                                     ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
                                      nb01/ggml_type_size(src0->type),
                                      (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                      row_size/ggml_type_size(vec_dot_type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
-                                     ith, nth,
                                      src0->type,
                                      vec_dot_type,
                                      dst->type))

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

@@ -522,6 +522,12 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_sve()) {
             features.push_back({ "SVE", "1" });
         }
+        if (ggml_cpu_has_dotprod()) {
+            features.push_back({ "DOTPROD", "1" });
+        }
+        if (ggml_cpu_has_matmul_int8()) {
+            features.push_back({ "MATMUL_INT8", "1" });
+        }
         if (ggml_cpu_get_sve_cnt() > 0) {
             static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
             features.push_back({ "SVE_CNT", sve_cnt.c_str() });

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -53,6 +53,8 @@
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
+#include <atomic>
+
 #ifdef _MSC_VER
 #define NOINLINE __declspec(noinline)
 #else
@@ -134,6 +136,16 @@ inline __m512 madd(__m512 a, __m512 b, __m512 c) {
     return _mm512_fmadd_ps(a, b, c);
 }
 #endif
+#if defined(__AVX512BF16__)
+template <>
+inline __m512 madd(__m512bh a, __m512bh b, __m512 c) {
+    return _mm512_dpbf16_ps(c, a, b);
+}
+template <>
+inline __m256 madd(__m256bh a, __m256bh b, __m256 c) {
+    return _mm256_dpbf16_ps(c, a, b);
+}
+#endif
 #endif
 
 #if defined(__ARM_FEATURE_FMA)
@@ -204,6 +216,7 @@ template <> inline float32x4_t load(const float *p) {
     return vld1q_f32(p);
 }
 #if !defined(_MSC_VER)
+// FIXME: this should check for __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 template <> inline float16x8_t load(const ggml_fp16_t *p) {
     return vld1q_f16((const float16_t *)p);
 }
@@ -225,6 +238,13 @@ template <> inline __m256 load(const float *p) {
 }
 #endif // __AVX__
 
+#if defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m256 load(const ggml_bf16_t *p) {
+    return _mm256_castsi256_ps(
+        _mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)p)), 16));
+}
+#endif // __AVX2__
+
 #if defined(__F16C__)
 template <> inline __m256 load(const ggml_fp16_t *p) {
     return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
@@ -238,8 +258,27 @@ template <> inline __m512 load(const float *p) {
 template <> inline __m512 load(const ggml_fp16_t *p) {
     return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
 }
+template <> inline __m512 load(const ggml_bf16_t *p) {
+    return _mm512_castsi512_ps(
+        _mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)p)), 16));
+}
 #endif // __AVX512F__
 
+#if defined(__AVX512BF16__)
+template <> inline __m512bh load(const ggml_bf16_t *p) {
+    return (__m512bh)_mm512_loadu_ps((const float *)p);
+}
+template <> inline __m256bh load(const ggml_bf16_t *p) {
+    return (__m256bh)_mm256_loadu_ps((const float *)p);
+}
+template <> inline __m512bh load(const float *p) {
+    return _mm512_cvtne2ps_pbh(_mm512_loadu_ps(p + 16), _mm512_loadu_ps(p));
+}
+template <> inline __m256bh load(const float *p) {
+    return _mm512_cvtneps_pbh(_mm512_loadu_ps(p));
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // CONSTANTS
 
@@ -251,199 +290,170 @@ static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FLOATING POINT MATRIX MULTIPLICATION
 
+template <int M>
+static inline int64_t BLOCK_SIZE(size_t m) {
+    const int64_t NB_BLOC_M = (m + M - 1) / M;
+    return (m % NB_BLOC_M == 0) ? m / NB_BLOC_M : (m / NB_BLOC_M) + 1;
+}
+
+static constexpr inline int64_t BLOC_POS(int64_t ib, int64_t ibN, int64_t bloc_size) {
+    return ib < ibN ? ib * bloc_size : ibN * bloc_size + (ib - ibN) * (bloc_size - 1);
+}
+
 template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
 class tinyBLAS {
   public:
-    tinyBLAS(int64_t k,
+    tinyBLAS(const ggml_compute_params * params, int64_t k,
              const TA *A, int64_t lda,
              const TB *B, int64_t ldb,
-             TC *C, int64_t ldc,
-             int ith, int nth)
-        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+             TC *C, int64_t ldc)
+        : params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
     }
 
-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
+    bool matmul(int64_t m, int64_t n) {
+        if (k % KN != 0)
+            return false;
+        // compute RM for only need tile with size RM&RM-1
+#if VECTOR_REGISTERS == 32
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 4>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 2>(m, n, SIZE_N, 12);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<6>(n);
+            mnpack<4, 6, 1>(m, n, SIZE_N, 12);
+            return true;
+        }
+#else  // VECTOR_REGISTERS == 16
+        if (m % 16 == 0 && (m/16 >= params->nth)) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 4>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 8 == 0 ) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 2>(m, n, SIZE_N, 24);
+            return true;
+        }
+        if (m % 4 == 0) {
+            const int64_t SIZE_N = BLOCK_SIZE<3>(n);
+            mnpack<4, 3, 1>(m, n, SIZE_N, 24);
+            return true;
+        }
+#endif
+        return false;
     }
 
   private:
-    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t mc, nc, mp, np;
-        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
-#if VECTOR_REGISTERS == 32
-        case 0x55:
-            mc = 5;
-            nc = 5;
-            gemm<5, 5>(m0, m, n0, n);
-            break;
-        case 0x45:
-            mc = 4;
-            nc = 5;
-            gemm<4, 5>(m0, m, n0, n);
-            break;
-        case 0x54:
-            mc = 5;
-            nc = 4;
-            gemm<5, 4>(m0, m, n0, n);
-            break;
-        case 0x44:
-            mc = 4;
-            nc = 4;
-            gemm<4, 4>(m0, m, n0, n);
-            break;
-        case 0x53:
-            mc = 5;
-            nc = 3;
-            gemm<5, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-            mc = 3;
-            nc = 5;
-            gemm<3, 5>(m0, m, n0, n);
-            break;
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-#else
-        case 0x55:
-        case 0x54:
-        case 0x53:
-        case 0x45:
-        case 0x44:
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-#endif
-        case 0x34:
-            mc = 3;
-            nc = 4;
-            gemm<3, 4>(m0, m, n0, n);
-            break;
-        case 0x52:
-            mc = 5;
-            nc = 2;
-            gemm<5, 2>(m0, m, n0, n);
-            break;
-        case 0x33:
-            mc = 3;
-            nc = 3;
-            gemm<3, 3>(m0, m, n0, n);
-            break;
-        case 0x25:
-            mc = 2;
-            nc = 5;
-            gemm<2, 5>(m0, m, n0, n);
-            break;
-        case 0x42:
-            mc = 4;
-            nc = 2;
-            gemm<4, 2>(m0, m, n0, n);
-            break;
-        case 0x24:
-            mc = 2;
-            nc = 4;
-            gemm<2, 4>(m0, m, n0, n);
-            break;
-        case 0x32:
-            mc = 3;
-            nc = 2;
-            gemm<3, 2>(m0, m, n0, n);
-            break;
-        case 0x23:
-            mc = 2;
-            nc = 3;
-            gemm<2, 3>(m0, m, n0, n);
-            break;
-        case 0x51:
-            mc = 5;
-            nc = 1;
-            gemm<5, 1>(m0, m, n0, n);
-            break;
-        case 0x41:
-            mc = 4;
-            nc = 1;
-            gemm<4, 1>(m0, m, n0, n);
-            break;
-        case 0x22:
-            mc = 2;
-            nc = 2;
-            gemm<2, 2>(m0, m, n0, n);
-            break;
-        case 0x15:
-            mc = 1;
-            nc = 5;
-            gemm<1, 5>(m0, m, n0, n);
-            break;
-        case 0x14:
-            mc = 1;
-            nc = 4;
-            gemm<1, 4>(m0, m, n0, n);
-            break;
-        case 0x31:
-            mc = 3;
-            nc = 1;
-            gemm<3, 1>(m0, m, n0, n);
-            break;
-        case 0x13:
-            mc = 1;
-            nc = 3;
-            gemm<1, 3>(m0, m, n0, n);
-            break;
-        case 0x21:
-            mc = 2;
-            nc = 1;
-            gemm<2, 1>(m0, m, n0, n);
-            break;
-        case 0x12:
-            mc = 1;
-            nc = 2;
-            gemm<1, 2>(m0, m, n0, n);
-            break;
-        case 0x11:
-            mc = 1;
-            nc = 1;
-            gemm<1, 1>(m0, m, n0, n);
-            break;
-        default:
-            return;
+    template <int RM, int RN, int BM>
+    inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
+        if (SIZE_N == RN) {
+            return gemm<RM, RN, BM>(m, n, BN);
+        }
+        if constexpr (RN > 1) {
+            return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
+        } else {
+            GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
+            GGML_ASSERT(false); // we have miss something.
         }
-        mp = m0 + (m - m0) / mc * mc;
-        np = n0 + (n - n0) / nc * nc;
-        mnpack(mp, m, n0, np);
-        mnpack(m0, m, np, n);
     }
 
     template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t ytiles = (m - m0) / RM;
-        int64_t xtiles = (n - n0) / RN;
-        int64_t tiles = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles)
-            end = tiles;
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = m0 + job / xtiles * RM;
-            int64_t jj = n0 + job % xtiles * RN;
+    inline void gemm_bloc(int64_t ii, int64_t jj) {
         D Cv[RN][RM] = {};
-            for (int64_t l = 0; l < k; l += KN)
-                for (int64_t j = 0; j < RN; ++j)
-                    for (int64_t i = 0; i < RM; ++i)
-                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
-                                        load<V>(B + ldb * (jj + j) + l),
-                                        Cv[j][i]);
+        for (int64_t l = 0; l < k; l += KN) {
+            // help compiler for op order.
+            if constexpr (RM <= RN) {
+                V Av[RM];
+                for (int64_t i = 0; i < RM; ++i) {
+                    Av[i] = load<V>(A + lda * (ii + i) + l);
+                }
+                for (int64_t j = 0; j < RN; ++j) {
+                    V Bv = load<V>(B + ldb * (jj + j) + l);
+                    for (int64_t i = 0; i < RM; ++i) {
+                        Cv[j][i] = madd(Av[i], Bv, Cv[j][i]);
+                    }
+                }
+            } else {
+                V Bv[RN];
+                for (int64_t j = 0; j < RN; ++j) {
+                    Bv[j] = load<V>(B + ldb * (jj + j) + l);
+                }
+                for (int64_t i = 0; i < RM; ++i) {
+                    V Av = load<V>(A + lda * (ii + i) + l);
+                    for (int64_t j = 0; j < RN; ++j) {
+                        Cv[j][i] = madd(Av, Bv[j], Cv[j][i]);
+                    }
+                }
+            }
+        }
         for (int64_t j = 0; j < RN; ++j)
             for (int64_t i = 0; i < RM; ++i)
                 C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
     }
+
+    template <int RM, int RN, int BM>
+    NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
+        static std::atomic<int64_t> current_chunk;
+
+        GGML_ASSERT(m % (RM * BM) == 0);
+        const int64_t ytiles = m / (RM * BM);
+        const int64_t xtiles = (n + RN -1) / RN;
+        const int64_t jj_RN = (xtiles - (xtiles * RN - n));
+
+        // "round" bloc_size to "nearest" BN
+        const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
+        const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
+        const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
+        const int64_t nb_job = ytiles * NB_BN;
+
+        if (params->ith == 0) {
+            GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
+            // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+            std::atomic_store_explicit(&current_chunk, (int64_t)params->nth, std::memory_order_relaxed);
         }
 
+        ggml_barrier(params->threadpool);
+
+        int64_t job = params->ith;
+        while (job < nb_job) {
+            const int64_t ii = (job % ytiles) * RM * BM;
+            const int64_t jb =  job / ytiles;
+            const int64_t jr0 = BLOC_POS(jb  , jj_BN, SIZE_BN);
+            const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
+
+            const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
+            const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
+            const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
+
+            for (int64_t bi = 0; bi < BM * RM; bi += RM) {
+                int64_t jj = jj0;
+                for (; jj < jj1; jj += RN) {
+                    gemm_bloc<RM, RN>(ii + bi, jj);
+                }
+                if constexpr (RN > 1) {
+                    for (; jj < jj2; jj += RN - 1) {
+                        gemm_bloc<RM, RN-1>(ii + bi, jj);
+                    }
+                }
+                GGML_ASSERT(jj == jj2);
+            }
+
+            // next step.
+            job = std::atomic_fetch_add_explicit(&current_chunk, (int64_t)1, std::memory_order_relaxed);
+        }
+
+        ggml_barrier(params->threadpool);
+        return;
+    }
+
+    const ggml_compute_params * params;
     const TA *const A;
     const TB *const B;
     TC *const C;
@@ -451,8 +461,6 @@ class tinyBLAS {
     const int64_t lda;
     const int64_t ldb;
     const int64_t ldc;
-    const int ith;
-    const int nth;
 };
 
 //////////////////////////////////////////////////////////////////////////////////////////
@@ -992,8 +1000,10 @@ class tinyBLAS_Q0_AVX {
 
     inline __m256 updot(__m256i u, __m256i s) {
         __m256i res;
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
         res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
+#elif defined(__AVXVNNI__)
+        res = _mm256_dpbusd_avx_epi32(_mm256_setzero_si256(), u, s);
 #else
         res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
 #endif
@@ -1656,8 +1666,9 @@ class tinyBLAS_PPC {
  * @param Ctype is GGML data type of `C`
  * @return true if this function was able to service the matmul request
  */
-bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
-                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
+bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
+                     const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int Atype, int Btype, int Ctype) {
 
     assert(m >= 0);
     assert(n >= 0);
@@ -1665,8 +1676,8 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
     assert(lda >= k);
     assert(ldb >= k);
     assert(ldc >= m);
-    assert(nth > 0);
-    assert(ith < nth);
+    assert(params->nth > 0);
+    assert(params->ith < params->nth);
 
     // only enable sgemm for prompt processing
     if (n < 2)
@@ -1681,37 +1692,25 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
         if (Btype != GGML_TYPE_F32)
             return false;
 #if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        tinyBLAS<16, __m512, __m512, float, float, float> tb{
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{ params,
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__AVX__) || defined(__AVX2__)
-        if (k % 8)
-            return false;
-        tinyBLAS<8, __m256, __m256, float, float, float> tb{
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{ params,
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__ARM_NEON)
         if (n < 4)
             return false;
-        if (k % 4)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
             k, (const float *)A, lda,
             (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__MMA__)
         if (k % 8)
             return false;
@@ -1719,7 +1718,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const float *)A, lda,
             (const float *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1727,60 +1726,71 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
 #endif
     }
 
+    case GGML_TYPE_BF16: {
+#if defined(__AVX512BF16__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#elif defined(__AVX512F__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#elif defined(__AVX2__)
+        if (Btype == GGML_TYPE_BF16) {
+            tinyBLAS<8, __m256, __m256, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
+                (const ggml_bf16_t *)A, lda,
+                (const ggml_bf16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
+#endif
+        return false;
+    }
     case GGML_TYPE_F16: {
 #if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
+                (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
         if (n < 8)
             return false;
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F16)
-            return false;
-        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
                 k, (const ggml_fp16_t *)A, lda,
                 (const ggml_fp16_t *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #elif defined(__ARM_NEON) && !defined(_MSC_VER)
-        if (k % 4)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
+        if (Btype == GGML_TYPE_F32) {
+            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{ params,
                 k, (const ggml_fp16_t *)A, lda,
                 (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #endif
+        return false;
     }
 
     case GGML_TYPE_Q8_0: {
@@ -1791,7 +1801,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_q8_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@@ -1799,7 +1809,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_q8_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1815,7 +1825,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_q4_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #elif defined(__ARM_FEATURE_DOTPROD)
@@ -1823,7 +1833,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_q4_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1839,7 +1849,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_q5_0 *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1855,7 +1865,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
             k, (const block_iq4_nl *)A, lda,
             (const block_q8_0 *)B, ldb,
             (float *)C, ldc,
-            ith, nth};
+            params->ith, params->nth};
         tb.matmul(m, n);
         return true;
 #else
@@ -1867,6 +1877,7 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
         return false;
     }
 
+    (void)params;
     (void)m;
     (void)n;
     (void)k;
@@ -1876,8 +1887,6 @@ bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda
     (void)ldb;
     (void)C;
     (void)ldc;
-    (void)ith;
-    (void)nth;
     (void)Atype;
     (void)Btype;
     (void)Ctype;

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -5,8 +5,8 @@
 extern "C" {
 #endif
 
-bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
-                     const void *, int64_t, void *, int64_t, int, int,
+bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t, int64_t, int64_t,
+                     const void *, int64_t, const void *, int64_t, void *, int64_t,
                      int, int, int);
 
 #ifdef __cplusplus

ggml/src/ggml-impl.h

@@ -551,6 +551,22 @@ static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
 #define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
 #define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
+// expose GGUF internals for test code
+
+GGML_API size_t gguf_type_size(enum gguf_type type);
+
+GGML_API struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_params params);
+
+struct gguf_buf {
+    void * data;
+    size_t size;
+    size_t offset;
+};
+GGML_API struct gguf_buf gguf_buf_init(size_t size);
+GGML_API void gguf_buf_free(struct gguf_buf buf);
+
+GGML_API void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf * buf, bool only_meta);
+
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-metal/CMakeLists.txt

@@ -103,3 +103,19 @@ else()
         DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
         )
 endif() # GGML_METAL_EMBED_LIBRARY
+
+if (NOT 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})
+
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+endif()

ggml/src/ggml-metal/ggml-metal.m

@@ -2067,8 +2067,8 @@ static void ggml_metal_encode_node(
                 GGML_ASSERT(ne12 % ne02 == 0);
                 GGML_ASSERT(ne13 % ne03 == 0);
 
-                const uint r2 = ne12/ne02;
-                const uint r3 = ne13/ne03;
+                const uint32_t r2 = ne12/ne02;
+                const uint32_t r3 = ne13/ne03;
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel

ggml/src/ggml-sycl/common.cpp

@@ -11,6 +11,8 @@
 //
 
 #include "common.hpp"
+
+#include "ggml-backend-impl.h"
 #include "ggml-impl.h"
 
 int get_current_device_id() {
@@ -65,9 +67,9 @@ void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
                                  const ggml_sycl_op_flatten_t op) try {
 
     const bool use_src1 = src1 != nullptr;
-
-    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
-    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    if(use_src1)
+      GGML_ASSERT(strcmp(src1->buffer->buft->iface.get_name(src1->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
+    GGML_ASSERT(strcmp(dst->buffer->buft->iface.get_name(dst->buffer->buft), GGML_SYCL_NAME "_Split") != 0);
 
     // dd = data device
     float * src0_ddf = (float *) src0->data;

ggml/src/ggml-sycl/common.hpp

@@ -26,7 +26,11 @@
 
 #define GGML_COMMON_DECL_SYCL
 #define GGML_COMMON_IMPL_SYCL
+/* suppress warning spam */
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wnested-anon-types"
 #include "ggml-common.h"
+#pragma clang diagnostic pop
 
 void* ggml_sycl_host_malloc(size_t size);
 void ggml_sycl_host_free(void* ptr);

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -288,10 +288,8 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                      ggml_tensor *tensor) try {
     ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
 
-    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+    if (tensor->view_src != NULL) {
         assert(tensor->view_src->buffer->buft == buffer->buft);
-        tensor->backend = tensor->view_src->backend;
-        tensor->extra = tensor->view_src->extra;
         return;
     }
 
@@ -539,7 +537,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
     auto dev_count = ggml_backend_sycl_get_device_count();
 
     if (device>=dev_count or device<0) {
-        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+        GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
             device, dev_count-1);
         GGML_ASSERT(device<dev_count);
     }
@@ -567,7 +565,7 @@ ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_conte
 
     int device = ctx->device;
     if (device>=ggml_sycl_info().device_count or device<0) {
-        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+        GGML_LOG_ERROR("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
             device, ggml_sycl_info().device_count-1);
         GGML_ASSERT(device<ggml_sycl_info().device_count);
     }
@@ -746,7 +744,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
             size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
         }
 
-        // FIXME: do not crash if cudaMalloc fails
+        // FIXME: do not crash if SYCL Buffer alloc fails
         // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
         ggml_sycl_set_device(i);
         const queue_ptr stream = ctx->streams[i];
@@ -788,7 +786,6 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                 CHECK_TRY_ERROR(extra->events[i][is] = new sycl::event()));
         }
     }
-    tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT;
     tensor->extra = extra;
 }
 catch (sycl::exception const &exc) {
@@ -2349,12 +2346,22 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
 
     dpct::memcpy_direction kind;
     char * src_ptr;
-    if (src->backend == GGML_BACKEND_TYPE_CPU) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
         kind = dpct::host_to_device;
+        //GGML_SYCL_DEBUG("%s: Host buffer type src tensor\n", __func__);
         src_ptr = (char *) src->data;
         // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d  GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
-    } else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
-        GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
+    } else if (ggml_backend_buffer_is_sycl(src->buffer)) {
+        // If buffer is a SYCL buffer
+        //GGML_SYCL_DEBUG("%s: SYCL buffer type src tensor\n", __func__);
+        kind    = dpct::device_to_device;
+        src_ptr = (char *) src->data;
+    } else if (ggml_backend_buffer_is_sycl_split(src->buffer)) {
+        /*
+        If buffer is a SYCL split buffer
+        */
+        //GGML_SYCL_DEBUG("%s: Split buffer type src tensor\n", __func__);
+        GGML_ASSERT(i1_low == 0 && i1_high == src->ne[1]);
         kind = dpct::device_to_device;
         ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
         int id;
@@ -2857,8 +2864,8 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
     const int nb2 = dst->nb[2];
     const int nb3 = dst->nb[3];
 
-    GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
-    GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(dst->buffer));
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src1->buffer));
     GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
 
     GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
@@ -2878,7 +2885,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
     int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
 
-    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
     GGML_ASSERT(!(split && ne02 > 1));
     GGML_ASSERT(!(split && ne03 > 1));
     GGML_ASSERT(!(split && ne02 < ne12));
@@ -3198,7 +3205,7 @@ static void ggml_sycl_mul_mat_vec_p021(ggml_backend_sycl_context & ctx, const gg
                                        const ggml_tensor *src1,
                                        ggml_tensor *dst) try {
     GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
     GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
@@ -3231,7 +3238,7 @@ static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
     GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -3293,7 +3300,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
                                              ggml_tensor *dst) try {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -4638,10 +4645,9 @@ static ggml_backend_dev_t ggml_backend_sycl_reg_get_device(ggml_backend_reg_t re
 static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, const char *name) {
     GGML_UNUSED(reg);
 
-    // TODO: update to the current function signature
-    //if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
-    //    return (void *)ggml_backend_sycl_split_buffer_type;
-    //}
+    if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
+        return (void *)ggml_backend_sycl_split_buffer_type;
+    }
 
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -145,6 +145,8 @@ class vk_perf_logger;
 #endif
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
 
+static constexpr uint32_t mul_mat_vec_max_cols = 8;
+
 struct vk_device_struct {
     std::mutex mutex;
 
@@ -202,8 +204,8 @@ struct vk_device_struct {
     vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
-    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT][mul_mat_vec_max_cols];
     vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
@@ -411,7 +413,7 @@ struct vk_op_unary_push_constants {
     uint32_t ne;
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2;
     uint32_t ne0_012mp; uint32_t ne0_012L;
     uint32_t ne0_01mp;  uint32_t ne0_01L;
@@ -459,7 +461,7 @@ struct vk_op_binary_push_constants {
     uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
     uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
-    uint32_t d_offset;
+    uint32_t misalign_offsets;
     float param1; float param2; int32_t param3;
 };
 
@@ -546,7 +548,7 @@ struct vk_staging_memcpy {
 };
 
 struct vk_op_upscale_push_constants {
-    uint32_t ne; uint32_t d_offset;
+    uint32_t ne; uint32_t a_offset; uint32_t d_offset;
     uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
     float sf0; float sf1; float sf2; float sf3;
@@ -1404,10 +1406,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
         // spec constants and tile sizes for non-quant matmul/matmul_id
         l_warptile = { 256, 128, 256, 64 };
         m_warptile = { 256, 128, 128, 64 };
-        s_warptile = { 128,  32,  16, 64 };
+        s_warptile = { 128,  64,  64, 64 };
         l_wg_denoms = {128, 256, 1 };
         m_wg_denoms = {128, 128, 1 };
-        s_wg_denoms = { 32,  16, 1 };
+        s_wg_denoms = { 64,  64, 1 };
 
         // spec constants and tile sizes for quant matmul (non-Qi_K)
         l_warptile_mmq = { 256, 128, 256, 64 };
@@ -1855,53 +1857,60 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     // mul mat vec
 
-    // AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
-    uint32_t rm = 1;
-    if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
-        rm = 2;
+    // the number of rows computed per shader depends on GPU model and quant
+    uint32_t rm_stdq = 1;
+    uint32_t rm_kq = 2;
+    if (device->vendor_id == VK_VENDOR_ID_AMD) {
+        if (device->subgroup_min_size == 64 && device->subgroup_max_size == 64) { // GCN
+            rm_stdq = 2;
+            rm_kq = 4;
+        }
+    } else if (device->vendor_id == VK_VENDOR_ID_INTEL)
+        rm_stdq = 2;
 
-    // computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    for (uint32_t i = 0; i < mul_mat_vec_max_cols; ++i) {
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f32_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f32_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f32_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f32_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
 
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ][i], "mul_mat_vec_f32_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ][i], "mul_mat_vec_f16_f16_f32_"+std::to_string(i+1),  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2, i+1}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0][i], "mul_mat_vec_q4_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1][i], "mul_mat_vec_q4_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0][i], "mul_mat_vec_q5_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1][i], "mul_mat_vec_q5_1_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0][i], "mul_mat_vec_q8_0_f16_f32_"+std::to_string(i+1), mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K][i], "mul_mat_vec_q2_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K][i], "mul_mat_vec_q3_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K][i], "mul_mat_vec_q4_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K][i], "mul_mat_vec_q5_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K][i], "mul_mat_vec_q6_k_f16_f32_"+std::to_string(i+1), mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq, i+1}, 1, true);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL][i], "mul_mat_vec_iq4_nl_f16_f32_"+std::to_string(i+1), mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq, i+1}, 1, true);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {device->subgroup_size, 2*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq, 1, 1}, {device->subgroup_size, 1*rm_stdq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {rm_kq, 1, 1}, {subgroup_size_16, rm_kq}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq, 1, 1}, {subgroup_size_16, 2*rm_stdq}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -2012,11 +2021,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     if (device->float_controls_rte_fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {512, 1, 1}, { device->subgroup_size }, 1, true);
     }
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
@@ -2887,9 +2896,10 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 
-static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type, uint32_t num_cols) {
     VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
     GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+    GGML_ASSERT(num_cols >= 1 && num_cols <= mul_mat_vec_max_cols);
 
     switch (a_type) {
         case GGML_TYPE_F32:
@@ -2910,7 +2920,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
             return nullptr;
     }
 
-    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type][num_cols-1] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type][num_cols-1];
 }
 
 static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
@@ -3205,8 +3215,8 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
         GGML_ABORT("fatal error");
     }
     // Check if src is pinned memory
-    vk_buffer buf;
-    size_t buf_offset;
+    vk_buffer buf = nullptr;
+    size_t buf_offset = 0;
     ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset);
 
     const uint64_t ne0 = tensor->ne[0];
@@ -3269,7 +3279,7 @@ static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_cont
     VkBufferCopy buf_copy{ 0, offset, copy_size };
 
     ggml_vk_sync_buffers(subctx);
-    vkCmdCopyBuffer(subctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
+    vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     for (uint64_t i3 = 0; i3 < ne3; i3++) {
         for (uint64_t i2 = 0; i2 < ne2; i2++) {
@@ -3302,7 +3312,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
     }
     // Check if src is pinned memory
     vk_buffer buf = nullptr;
-    size_t buf_offset;
+    size_t buf_offset = 0;
     ggml_vk_host_get(dst->device, src, buf, buf_offset);
 
     if (buf != nullptr) {
@@ -3344,7 +3354,7 @@ static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, siz
         copy_size};
 
     ggml_vk_sync_buffers(subctx);
-    vkCmdCopyBuffer(subctx->s->buffer, staging_buffer->buffer, dst->buffer, 1, &buf_copy);
+    vkCmdCopyBuffer(subctx->s->buffer, (VkBuffer)staging_buffer->buffer, (VkBuffer)dst->buffer, 1, &buf_copy);
 
     if (width == spitch) {
         deferred_memcpy((uint8_t *)staging_buffer->ptr, src, width * height, &subctx->in_memcpys);
@@ -3400,7 +3410,7 @@ static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size
 
     // Check if dst is pinned memory
     vk_buffer buf = nullptr;
-    size_t buf_offset;
+    size_t buf_offset = 0;
     ggml_vk_host_get(src->device, dst, buf, buf_offset);
 
     std::vector<vk::BufferCopy> slices(1);
@@ -3480,7 +3490,7 @@ static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t ds
 
     VkBufferCopy bc{ src_offset, dst_offset, size };
 
-    vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
+    vkCmdCopyBuffer(ctx->s->buffer, (VkBuffer)src->buffer, (VkBuffer)dst->buffer, 1, &bc);
 }
 
 static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
@@ -3732,9 +3742,9 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
 
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
 
     bool src0_uma = false;
@@ -3920,8 +3930,6 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t ne12 = src1->ne[2];
     const uint64_t ne13 = src1->ne[3];
 
-    GGML_ASSERT(ne11 == 1);
-
     const uint64_t ne20 = dst->ne[0];
     const uint64_t ne21 = dst->ne[1];
     const uint64_t ne22 = dst->ne[2];
@@ -3930,13 +3938,18 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t r2 = ne12 / ne02;
     const uint64_t r3 = ne13 / ne03;
 
+    // batch_n indicates that we need to compute a few vector results, and this assumes
+    // ne12 and ne13 are 1. It overloads the batch_strides to hold the row strides.
+    GGML_ASSERT(ne11 == 1 || ne12 * ne13 == 1);
+    bool batch_n = ne11 > 1;
+
     ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
 
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
 
     bool src0_uma = false;
@@ -3980,7 +3993,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     } else {
         to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
     }
-    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type);
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type, ne11);
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
     GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
     GGML_ASSERT(dmmv != nullptr);
@@ -4052,8 +4065,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
     }
 
-    uint32_t stride_batch_x = ne00*ne01;
-    uint32_t stride_batch_y = ne10*ne11;
+    // For batch_n, the A matrix is the same for each batch, and B/D use the row stride as the batch stride
+    uint32_t stride_batch_x = batch_n ? 0 : ne00*ne01;
+    uint32_t stride_batch_y = batch_n ? ne10 : (ne10*ne11);
+    uint32_t stride_batch_d = batch_n ? ne20 : (ne20*ne21);
 
     if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
         stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
@@ -4076,7 +4091,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
     // compute
     const vk_mat_vec_push_constants pc = {
         (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
-        stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
+        stride_batch_x, stride_batch_y, stride_batch_d,
         (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
     };
     ggml_vk_sync_buffers(subctx);
@@ -4112,7 +4127,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
     ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
 
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
 
     bool src1_uma = false;
@@ -4256,7 +4271,10 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1 &&
                !ggml_is_permuted(src0) && !ggml_is_permuted(src1)) {
         ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
-    } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+    // mul_mat_vec supports batching ne12*ne13 when ne11==1, or treating ne11 as the batch size (up to four)
+    // when ne12 and ne13 are one.
+    } else if ((dst->ne[1] == 1 || (dst->ne[1] <= mul_mat_vec_max_cols && src1->ne[2] * src1->ne[3] == 1)) &&
+               (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
         ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
     } else {
         ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
@@ -4300,11 +4318,11 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
     ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
 
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
-    vk_buffer d_ids;
+    vk_buffer d_ids = nullptr;
     size_t ids_buf_offset = 0;
 
     bool src0_uma = false;
@@ -4505,11 +4523,11 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
     ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
 
-    vk_buffer d_Qx;
+    vk_buffer d_Qx = nullptr;
     size_t qx_buf_offset = 0;
-    vk_buffer d_Qy;
+    vk_buffer d_Qy = nullptr;
     size_t qy_buf_offset = 0;
-    vk_buffer d_ids;
+    vk_buffer d_ids = nullptr;
     size_t ids_buf_offset = 0;
 
     bool src0_uma = false;
@@ -4768,8 +4786,8 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
 
     ggml_vk_sync_buffers(subctx);
 
-    vk_buffer d_Q, d_K, d_V, d_D, d_M;
-    uint64_t q_buf_offset, k_buf_offset, v_buf_offset, d_buf_offset, m_buf_offset;
+    vk_buffer d_Q = nullptr, d_K = nullptr, d_V = nullptr, d_D = nullptr, d_M = nullptr;
+    size_t q_buf_offset = 0, k_buf_offset = 0, v_buf_offset = 0, d_buf_offset = 0, m_buf_offset = 0;
 
     bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false;
 
@@ -5071,6 +5089,57 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
     }
 }
 
+static uint32_t get_misalign_bytes(ggml_backend_vk_context * ctx, const ggml_tensor * t)
+{
+    return ((vk_tensor_offset(t) + t->view_offs) & (ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1));;
+}
+
+template <typename T> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, T &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    GGML_UNUSED(p);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+    GGML_UNUSED(dst);
+    static_assert(!std::is_const<T>::value, "unexpected type");
+    GGML_ASSERT(!src0 || get_misalign_bytes(ctx, src0) == 0);
+    GGML_ASSERT(!src1 || get_misalign_bytes(ctx, src1) == 0);
+    GGML_ASSERT(!src2 || get_misalign_bytes(ctx, src2) == 0);
+    GGML_ASSERT(!dst  || get_misalign_bytes(ctx, dst) == 0);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_unary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.misalign_offsets = (a_offset << 16) | d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_binary_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t b_offset = get_misalign_bytes(ctx, src1) / ggml_type_size(src1->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    GGML_ASSERT(dst->op != GGML_OP_GET_ROWS || (a_offset == 0 && b_offset == 0 && d_offset == 0));
+
+    p.misalign_offsets = (a_offset << 16) | (b_offset << 8) | d_offset;
+
+    GGML_UNUSED(src2);
+}
+
+template <> void init_pushconst_tensor_offsets(ggml_backend_vk_context * ctx, vk_op_upscale_push_constants &p, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
+    const uint32_t a_offset = get_misalign_bytes(ctx, src0) / ggml_type_size(src0->type);
+    const uint32_t d_offset = get_misalign_bytes(ctx, dst) / ggml_type_size(dst->type);
+
+    p.a_offset = a_offset;
+    p.d_offset = d_offset;
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src2);
+}
+
 template<typename PC>
 static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
@@ -5174,8 +5243,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     }
 
     GGML_ASSERT(d_D != nullptr);
-    uint64_t d_buf_offset = ((vk_tensor_offset(dst) + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
-    GGML_ASSERT(d_buf_offset == vk_tensor_offset(dst) || op == GGML_OP_CPY);  // NOLINT
+    uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
     if(!src0_uma) {
         d_X = src0_buf_ctx->dev_buffer;
         x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
@@ -5191,6 +5259,12 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
         GGML_ASSERT(d_Z != nullptr);
     }
+    // Compute misalignment offset for descriptors and store it in in push constants, then align the descriptor offsets.
+    init_pushconst_tensor_offsets(ctx, pc, src0, src1, src2, dst);
+    x_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    y_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    z_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
+    d_buf_offset &= ~(ctx->device->properties.limits.minStorageBufferOffsetAlignment - 1);
 
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
@@ -5378,7 +5452,6 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
     int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
     int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
@@ -5390,7 +5463,7 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f, offset,
     }, dryrun);
 }
@@ -5474,8 +5547,8 @@ static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subc
 
     ggml_vk_sync_buffers(subctx);
 
-    vk_buffer d_D, d_K, d_V, d_R, d_TF, d_TD, d_State;
-    uint64_t k_offset, v_offset, r_offset, tf_offset, td_offset, state_offset, dst_offset;
+    vk_buffer d_D = nullptr, d_K = nullptr, d_V = nullptr, d_R = nullptr, d_TF = nullptr, d_TD = nullptr, d_State = nullptr;
+    size_t k_offset = 0, v_offset = 0, r_offset = 0, tf_offset = 0, td_offset = 0, state_offset = 0, dst_offset = 0;
     bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false;
 
     if (ctx->device->uma) {
@@ -5594,7 +5667,7 @@ static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, c
     const float sf3 = (float)dst->ne[3] / src0->ne[3];
 
     ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
-        (uint32_t)ggml_nelements(dst), 0,
+        (uint32_t)ggml_nelements(dst), 0, 0,
         (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
         sf0, sf1, sf2, sf3,
@@ -5704,13 +5777,12 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
 static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
-    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
     ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
-        d_offset,
+        0,
         0.0f, 0.0f,
         0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);

ggml/src/ggml-vulkan/vulkan-shaders/acc.comp

@@ -21,9 +21,9 @@ void main() {
     get_indices(idx, i00, i01, i02, i03);
 
     if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
     } else {
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]));
     }
 }
 

ggml/src/ggml-vulkan/vulkan-shaders/add.comp

@@ -22,7 +22,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
 }

ggml/src/ggml-vulkan/vulkan-shaders/concat.comp

@@ -30,12 +30,12 @@ void main() {
     const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+    data_d[get_doffset() + dst_idx] = D_TYPE(is_src0 ? data_a[get_aoffset() + src0_idx] : data_b[get_boffset() + src1_idx]);
 #else
     if (is_src0) {
-        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+        data_d[get_doffset() + dst_idx] = data_a[get_aoffset() + src0_idx];
     } else {
-        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+        data_d[get_doffset() + dst_idx] = data_b[get_boffset() + src1_idx];
     }
 #endif
 }

ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp

@@ -19,9 +19,9 @@ void main() {
     if (idx + (num_iter-1)*num_threads < p.ne) {
         [[unroll]] for (uint i = 0; i < num_iter; ++i) {
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }
@@ -32,9 +32,9 @@ void main() {
             }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-            data_d[p.d_offset + idx] = D_TYPE(data_a[idx]);
+            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
 #else
-            data_d[p.d_offset + idx] = data_a[idx];
+            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
 #endif
             idx += num_threads;
         }

ggml/src/ggml-vulkan/vulkan-shaders/copy.comp

@@ -13,8 +13,8 @@ void main() {
     }
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
 #else
-    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
+    data_d[get_doffset() + dst_idx(idx)] = data_a[get_aoffset() + src0_idx(idx)];
 #endif
 }

ggml/src/ggml-vulkan/vulkan-shaders/cos.comp

@@ -12,6 +12,6 @@ void main() {
         return;
     }
 
-    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(cos(val));
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[get_aoffset() + src0_idx(idx)]);
+    data_d[get_doffset() + dst_idx(idx)] = D_TYPE(cos(val));
 }

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -10,9 +10,10 @@ float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2
     const float16_t d = bl.block.d;
     const uint idx = coordInBlock[1];
     const uint shift = (idx & 0x10) >> 2;
-    uint32_t qs = unpack8(uint32_t(bl.block.qs[(idx & 0xE) >> 1]))[idx & 1];
+    uint32_t qs = uint32_t(bl.block.qs[(idx & 0xE) >> 1]);
     qs >>= shift;
-    qs &= 0xF;
+    qs &= 0x0F0F;
+    qs = unpack8(qs)[idx & 1];
     float16_t ret = (float16_t(qs) - float16_t(8)) * d;
     return ret;
 }
@@ -152,15 +153,17 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4
    block_q4_K block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K_packed16 {
+   block_q4_K_packed16 block;
+};
+
 float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
+    decodeBufQ4_K_packed16 bl16 = decodeBufQ4_K_packed16(bl);
     const uint idx = coordInBlock[1];
-    const uint iqs = idx;
 
-    const uint n = iqs / 64;                   // 0,1,2,3
-    const uint b = (iqs % 64) / 32;            // 0,1
+    const uint b = (idx & 0x20) >> 5;            // 0,1
     const uint is = (idx & 0xE0) >> 5;         // 0..7
-    const uint qsi = n * 32 + (iqs % 32);      // 0..127
 
     const f16vec2 loadd = bl.block.d;
 
@@ -184,9 +187,11 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2
     const float16_t d = loadd.x * float16_t(sc);
     const float16_t m = loadd.y * float16_t(mbyte);
 
-    uint32_t dmask = 0xF << (b * 4);
+    uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
+    qs = (qs >> (b * 4)) & 0x0F0F;
+    qs = unpack8(qs)[idx & 1];
 
-    float16_t ret = d * float16_t((bl.block.qs[qsi    ] & dmask) >> (b * 4)) - m;
+    float16_t ret = d * float16_t(qs) - m;
 
     return ret;
 }
@@ -195,18 +200,19 @@ layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5
    block_q5_K block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K_packed16 {
+   block_q5_K_packed16 block;
+};
+
 float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
+    decodeBufQ5_K_packed16 bl16 = decodeBufQ5_K_packed16(bl);
     const uint idx = coordInBlock[1];
-    const uint iqs = idx;
 
-    const uint n = iqs / 64;                   // 0,1,2,3
-    const uint b = (iqs % 64) / 32;            // 0,1
+    const uint b = (idx & 0x20) >> 5;          // 0,1
     const uint is = (idx & 0xE0) >> 5;         // 0..7
-    const uint qsi = n * 32 + (iqs % 32);      // 0..127
-    const uint qhi = (iqs % 32);               // 0..31
 
-    const uint8_t hm = uint8_t(1 << (iqs / 32));
+    const uint32_t hm = 0x0101 << is;
 
     const f16vec2 loadd = bl.block.d;
 
@@ -230,9 +236,15 @@ float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2
     const float16_t d = loadd.x * float16_t(sc);
     const float16_t m = loadd.y * float16_t(mbyte);
 
-    uint32_t dmask = 0xF << (b * 4);
+    uint qh = uint32_t(bl16.block.qh[(idx & 0x1E) >> 1]);
+    qh = qh & hm;
+    qh = unpack8(qh)[idx & 1];
 
-    float16_t ret = d * (float16_t((bl.block.qs[qsi    ] & dmask) >> (b * 4)) + float16_t((bl.block.qh[qhi    ] & hm) != 0 ? 16 : 0)) - m;
+    uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
+    qs = (qs >> (b * 4)) & 0x0F0F;
+    qs = unpack8(qs)[idx & 1];
+
+    float16_t ret = d * (float16_t(qs) + (qh != 0 ? float16_t(16) : float16_t(0))) - m;
 
     return ret;
 }
@@ -241,22 +253,30 @@ layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_
    block_q6_K block;
 };
 
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ6_K_packed16 {
+   block_q6_K_packed16 block;
+};
+
 float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
 {
+    decodeBufQ6_K_packed16 bl16 = decodeBufQ6_K_packed16(bl);
     const uint idx = coordInBlock[1];
-    const uint iqs = idx;
 
-    const uint n = iqs / 128;                   // 0,1
-    const uint b = (iqs % 128) / 64;            // 0,1
-    const uint is_b = (iqs % 32) / 16;          // 0,1
-    const uint qhshift = ((iqs % 128) / 32) * 2;// 0,2,4,6
-    const uint is = 8 * n + qhshift + is_b;     // 0..15
-    const uint qsi = n * 64 + (iqs % 64);       // 0..127
-    const uint qhi = n * 32 + (iqs % 32);       // 0..63
+    const uint b = (idx & 0x40) >> 6;           // 0,1
+    const uint qhshift = (idx & 0x60) >> 4;    // 0,2,4,6
+    const uint is = (idx & 0xF0) >> 4;          // 0..15
 
     const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]);
 
-    float16_t ret = dscale * float16_t(int8_t(((bl.block.ql[qsi    ] >> (b * 4)) & 0xF) | (((bl.block.qh[qhi    ] >> qhshift) & 3) << 4)) - 32);
+    uint ql = uint32_t(bl16.block.ql[((idx & 0x80) >> 2) + ((idx & 0x3E) >> 1)]);
+    ql = (ql >> (b * 4)) & 0x0F0F;
+
+    uint qh = uint32_t(bl16.block.qh[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
+    qh = ((qh >> qhshift) & 0x0303) << 4;
+
+    int q = unpack8(ql | qh)[idx & 1];
+
+    float16_t ret = dscale * float16_t(q - 32);
 
     return ret;
 }

ggml/src/ggml-vulkan/vulkan-shaders/div.comp

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp

@@ -7,7 +7,7 @@ layout (push_constant) uniform parameter
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2; int param3;
 } p;
 
@@ -22,6 +22,10 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
+uint get_doffset() { return p.misalign_offsets & 0xFF; }
+
 // mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
 uint fastmod(uint a, uint b) {
     if ((b & (b-1)) == 0) {

ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp

@@ -6,7 +6,7 @@ layout (push_constant) uniform parameter
     uint ne;
     uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
-    uint d_offset;
+    uint misalign_offsets;
     float param1; float param2;
 
     uint ne0_012mp; uint ne0_012L;
@@ -24,6 +24,9 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+uint get_aoffset() { return p.misalign_offsets >> 16; }
+uint get_doffset() { return p.misalign_offsets & 0xFFFF; }
+
 // see init_fastdiv_values in ggml-vulkan.cpp
 uint fastdiv(uint n, uint mp, uint L) {
     uint msbs, lsbs;

ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp

@@ -15,10 +15,10 @@ void main() {
         return;
     }
 
-    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+    const uint i01 = data_b[get_boffset() + i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
 
-    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
-    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+    const uint a_offset = get_aoffset() + i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = get_doffset() + i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
 
 #ifndef OPTIMIZATION_ERROR_WORKAROUND
     data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);

ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp

@@ -2,6 +2,7 @@
 
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_spirv_intrinsics: enable
+#extension GL_EXT_control_flow_attributes : require
 
 #if RTE16
 spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
@@ -23,40 +24,64 @@ layout (push_constant) uniform parameter
 
 #include "types.comp"
 
-#define BLOCK_SIZE 256
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
 
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+const uint NUM_ITER = 512 / BLOCK_SIZE;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
 
 void main() {
-    const uint i = gl_GlobalInvocationID.x;
-    if (i >= p.pelements) {
-        return;
+    const uint gidx = gl_GlobalInvocationID.x;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    A_TYPE values[NUM_ITER];
+    uint offset_dst[NUM_ITER];
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+        values[idx] = A_TYPE(0);
     }
 
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
         const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
         const uint kx = i / ksize;
         const uint kd = kx * ksize;
         const uint ky = (i - kd) / p.OW;
         const uint ix = i % p.OW;
 
-    const uint oh = gl_GlobalInvocationID.y;
-    const uint batch = gl_GlobalInvocationID.z / p.IC;
-    const uint ic = gl_GlobalInvocationID.z % p.IC;
-
         const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
         const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
 
-    const uint offset_dst =
+        offset_dst[idx] =
             ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
             (ic * (p.KW * p.KH) + ky * p.KW + kx);
 
-    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
-        data_d[offset_dst] = D_TYPE(0.0f);
-    } else {
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        if (iih < p.IH && iiw < p.IW) {
             const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
-        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+            values[idx] = data_a[offset_src + iih * p.IW + iiw];
         }
     }
+
+    [[unroll]] for (uint idx = 0; idx < NUM_ITER; ++idx) {
+
+        const uint i = gidx * NUM_ITER + idx;
+
+        if (i >= p.pelements) {
+            continue;
+        }
+
+        data_d[offset_dst[idx]] = D_TYPE(values[idx]);
+    }
+
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul.comp

@@ -20,7 +20,7 @@ void main() {
         uint i00, i01, i02, i03;
         get_indices(idx, i00, i01, i02, i03);
 
-        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+        data_d[get_doffset() + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[get_boffset() + src1_idx(i00, i01, i02, i03)]));
 
         idx += num_threads;
     }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp

@@ -9,9 +9,6 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-layout (constant_id = 1) const uint NUM_ROWS = 1;
-
 #if !defined(DATA_A_F32) && !defined(DATA_A_F16)
 #define K_PER_ITER 8
 #else
@@ -21,23 +18,22 @@ layout (constant_id = 1) const uint NUM_ROWS = 1;
 
 uint a_offset, b_offset, d_offset, y_offset;
 
-shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
-
-void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
+void iter(inout FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
 {
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
         const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
         const uint iybs = col - col%QUANT_K; // y block start index
 
 #if K_PER_ITER == 8
 #if QUANT_R == 2
-    const B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
-    const B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
+        const B_TYPE_VEC4 bv02 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4];
+        const B_TYPE_VEC4 bv13 = data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs + y_offset) / 4];
         const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
         const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
 #else
-    const vec4 bv0 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4]);
-    const vec4 bv1 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4 + 1]);
+        const vec4 bv0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4]);
+        const vec4 bv1 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + iybs + iqs) / 4 + 1]);
 #endif
 #else
         // Check if the second of the pair of elements is OOB, and don't fetch B or
@@ -48,9 +44,9 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
         const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
 
         FLOAT_TYPE b0 = 0, b1 = 0;
-    b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
+        b0 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs]);
         if (!OOB) {
-        b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+            b1 = FLOAT_TYPE(data_b[j*p.batch_stride_b + b_offset + iybs + iqs + y_offset]);
         }
 #endif
         uint ibi = first_row*p.ncols;
@@ -75,18 +71,19 @@ void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_
             if (dm.y == 0)
                 rowtmp *= dm.x;
 
-        temp[n] += rowtmp;
+            temp[j][n] += rowtmp;
 #else
             const vec2 v = dequantize(ib, iqs, a_offset);
 
             // matrix multiplication
-        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+            temp[j][n] = fma(FLOAT_TYPE(v.x), b0, temp[j][n]);
             if (!OOB) {
-            temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+                temp[j][n] = fma(FLOAT_TYPE(v.y), b1, temp[j][n]);
             }
 #endif
         }
     }
+}
 
 void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
@@ -96,10 +93,12 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
-    FLOAT_TYPE temp[NUM_ROWS];
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-    for (uint i = 0; i < NUM_ROWS; ++i) {
-        temp[i] = FLOAT_TYPE(0);
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
     }
 
     uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
@@ -131,24 +130,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
         i++;
     }
 
-    // sum up partial sums and write back result
-    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        tmpsh[n][tid] = temp[n];
-    }
-    barrier();
-    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-                tmpsh[n][tid] += tmpsh[n][tid + s];
-            }
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
-        }
-    }
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
 }
 
 void main() {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp

@@ -83,3 +83,36 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
             batch_idx * p.batch_stride_d;
 #endif
 }
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
+layout (constant_id = 2) const uint NUM_COLS = 1;
+
+shared FLOAT_TYPE tmpsh[NUM_COLS][NUM_ROWS][BLOCK_SIZE];
+
+void reduce_result(const in FLOAT_TYPE temp[NUM_COLS][NUM_ROWS], const in uint32_t d_offset, const in uint32_t first_row, const in uint32_t num_rows, const in uint32_t tid) {
+    // sum up partial sums and write back result
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            tmpsh[j][n][tid] = temp[j][n];
+        }
+    }
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                    tmpsh[j][n][tid] += tmpsh[j][n][tid + s];
+                }
+            }
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                data_d[j*p.batch_stride_d + d_offset + first_row + n] = D_TYPE(tmpsh[j][n][0]);
+            }
+        }
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -5,22 +5,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -38,24 +27,23 @@ void main() {
     const uint s_offset = 8*v_im;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
             f16vec2 d = data_a[ib0 + i].d;
             const FLOAT_TYPE dall = d.x;
             const FLOAT_TYPE dmin = d.y;
 
-        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
-        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
-        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
-        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
-        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
-        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
-        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
-        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
-
             uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
             uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
 
@@ -74,6 +62,16 @@ void main() {
             uvec2 qs0 =  uvec2(unpack8(qs0_u16));
             uvec2 qs16 = uvec2(unpack8(qs16_u16));
 
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
+                B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
+                B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
+                B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
+                B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
+                B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
+                B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
+                B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
+
                 FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
                 FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
                 [[unroll]] for (int l = 0; l < 2; ++l) {
@@ -94,20 +92,24 @@ void main() {
                            fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
                            fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
                 }
-        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
+                temp[j][n] = fma(dall, sum1, fma(-dmin, sum2, temp[j][n]));
+            }
+        }
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -5,22 +5,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -39,24 +28,23 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 128*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     const uint s_shift = 4 * v_im;
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y_idx = i * QUANT_K + y_offset;
 
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
             const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
 
-        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
-        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
-        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
-        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
-        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
-        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
-        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
-        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
-
             uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
             uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
             uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
@@ -70,6 +58,17 @@ void main() {
             u8vec2 s8 = unpack8(s8_16);
             u8vec2 s10 = unpack8(s10_16);
 
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+
+                B_TYPE_VEC2 b0 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 0];
+                B_TYPE_VEC2 b16 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 8];
+                B_TYPE_VEC2 b32 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 16];
+                B_TYPE_VEC2 b48 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 24];
+                B_TYPE_VEC2 b64 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 32];
+                B_TYPE_VEC2 b80 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 40];
+                B_TYPE_VEC2 b96 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 48];
+                B_TYPE_VEC2 b112 = data_b_v2[(j*p.batch_stride_b + b_offset + y_idx) / 2 + 56];
+
                 FLOAT_TYPE sum = FLOAT_TYPE(0.0);
                 [[unroll]] for (int l = 0; l < 2; ++l) {
                     sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
@@ -81,20 +80,24 @@ void main() {
                           fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
                           fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
                 }
-        temp = fma(d, sum, temp);
+                temp[j][n] = fma(d, sum, temp[j][n]);
+            }
+        }
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -6,22 +6,11 @@
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    if (row >= p.stride_d) {
-        return;
-    }
-
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     uint a_offset, b_offset, d_offset;
     get_offsets(a_offset, b_offset, d_offset);
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
 
     // 16 threads are used to process each block
     const uint it_size = gl_WorkGroupSize.x/16;
@@ -42,12 +31,20 @@ void main() {
     const uint q_offset = 32*v_im + l0;
     const uint y_offset = 64*v_im + l0;
 
-    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+    FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
+
+    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
+            temp[j][i] = FLOAT_TYPE(0);
+        }
+    }
 
     [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
         const uint y1_idx = i * QUANT_K + y_offset;
         const uint y2_idx = y1_idx + 128;
 
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
             f16vec2 d = data_a[ib0 + i].d;
             const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
             const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
@@ -98,10 +95,11 @@ void main() {
             const uint32_t q4_14 = qs64_hi4.z;
             const uint32_t q4_15 = qs64_hi4.w;
 
-        B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
-        B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
-        B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
-        B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
+            [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
+                B_TYPE_VEC4 by10 =  data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4];
+                B_TYPE_VEC4 by132 = data_b_v4[(j*p.batch_stride_b + b_offset + y1_idx) / 4 + 8];
+                B_TYPE_VEC4 by20 =  data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4];
+                B_TYPE_VEC4 by232 = data_b_v4[(j*p.batch_stride_b + b_offset + y2_idx) / 4 + 8];
 
                 const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
                 const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
@@ -112,20 +110,24 @@ void main() {
                     fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
                     fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
                     fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
-        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
+                temp[j][n] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp[j][n]));
+            }
+        }
     }
 
-    tmp[gl_LocalInvocationID.x] = temp;
+    reduce_result(temp, d_offset, first_row, num_rows, tid);
+}
 
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
         }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
+        compute_outputs(first_row, p.stride_d - first_row);
     }
 }