alloc : fix allocation data of pre-allocated leafs

whisper : allocate encoder results in dedicated buffer
2025-06-24 17:15:19 +00:00 · 2024-03-16 15:47:14 +01:00 · 2024-03-16 16:15:12 +02:00
439 changed files with 53783 additions and 240194 deletions
--- a/.devops/cublas.Dockerfile
+++ b/.devops/cublas.Dockerfile
@ -21,7 +21,7 @@ COPY . .
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV GGML_CUDA=1
+ENV WHISPER_CUBLAS=1

 RUN make

--- a/.devops/main-cuda.Dockerfile
+++ b/.devops/main-cuda.Dockerfile
@ -14,7 +14,7 @@ ARG CUDA_DOCKER_ARCH=all
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
-ENV GGML_CUDA=1
+ENV WHISPER_CUBLAS=1

 RUN apt-get update && \
    apt-get install -y build-essential \
@ -28,8 +28,6 @@ COPY .. .
 RUN make

 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
-ENV CUDA_MAIN_VERSION=12.3
-ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 WORKDIR /app

 RUN apt-get update && \
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -15,10 +15,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -36,7 +36,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Dependencies
        run: |
@ -53,10 +53,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Build
-        uses: cross-platform-actions/action@v0.24.0
+        uses: cross-platform-actions/action@v0.15.0
        with:
          operating_system: freebsd
          version: '13.2'
@ -77,10 +77,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -101,17 +101,14 @@ jobs:
      fail-fast: false
      matrix:
        build: [Debug, Release]
-        #arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
-        # TODO: arm/v7 disabled due to clang bug
-        #       https://github.com/ggerganov/whisper.cpp/actions/runs/9657764109/job/26637633042?pr=2256#step:4:1990
-        arch: [linux/amd64, linux/arm64, linux/ppc64le]
+        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -136,10 +133,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
+        uses: docker/setup-qemu-action@v2

      - name: Build ${{ matrix.arch }}
        run: |
@ -155,21 +152,21 @@ jobs:

  ubuntu-22-cmake-sycl:
    runs-on: ubuntu-22.04
-
+    
    strategy:
      fail-fast: false
      matrix:
        dwhisper_sycl: [ON]
        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
+        dcmake_cxx_compiler: [icpx] 
        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]

    continue-on-error: true

    steps:
      - name: Clone
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+        
      - name: add oneAPI to apt
        shell: bash
        run: |
@ -192,34 +189,34 @@ jobs:

      - name: Clone
        id: checkout
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+            
      - name: Build
        id: cmake_build
        run: |
          source /opt/intel/oneapi/setvars.sh
          mkdir build
          cd build
-          cmake -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
          cmake --build . --config Release -j $(nproc)

  ubuntu-22-cmake-sycl-fp16:
    runs-on: ubuntu-22.04
-
+    
    strategy:
      fail-fast: false
      matrix:
        dwhisper_sycl: [ON]
        dcmake_c_compiler: [icx]
-        dcmake_cxx_compiler: [icpx]
+        dcmake_cxx_compiler: [icpx] 
        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]

    continue-on-error: true

    steps:
      - name: Clone
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+        
      - name: add oneAPI to apt
        shell: bash
        run: |
@ -242,75 +239,17 @@ jobs:

      - name: Clone
        id: checkout
-        uses: actions/checkout@v4
-
+        uses: actions/checkout@v3
+            
      - name: Build
        id: cmake_build
        run: |
          source /opt/intel/oneapi/setvars.sh
          mkdir build
          cd build
-          cmake -DGGML_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
+          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
          cmake --build . --config Release -j $(nproc)
-
-  windows-msys2:
-    runs-on: windows-latest
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
-          - { sys: CLANG64, env: clang-x86_64, build: Release }
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Setup ${{ matrix.sys }}
-        uses: msys2/setup-msys2@v2
-        with:
-          update: true
-          msystem: ${{matrix.sys}}
-          install: >-
-            base-devel
-            mingw-w64-${{matrix.env}}-toolchain
-            mingw-w64-${{matrix.env}}-cmake
-            mingw-w64-${{matrix.env}}-SDL2
-            mingw-w64-${{matrix.env}}-openblas
-
-      - name: Build using make
-        shell: msys2 {0}
-        run: |
-            make -j $(nproc)
-
-      - name: Clean after building using make
-        shell: msys2 {0}
-        run: |
-            make clean
-
-      - name: Build using make w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            make GGML_OPENBLAS=1 -j $(nproc)
-
-      - name: Build using CMake
-        shell: msys2 {0}
-        run: |
-            cmake -B build
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
-      - name: Clean after building using CMake
-        shell: msys2 {0}
-        run: |
-            rm -rf build
-
-      - name: Build using CMake w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            cmake -B build -DGGML_OPENBLAS=ON
-            cmake --build build --config ${{ matrix.build }} -j $(nproc)
-
+  
  windows:
    runs-on: windows-latest

@ -331,10 +270,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Fetch SDL2 and set SDL2_DIR
        if: matrix.sdl2 == 'ON'
@ -359,14 +298,14 @@ jobs:
        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}

      - name: Upload dll
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
        with:
          name: ${{ matrix.jnaPath }}_whisper.dll
          path: build/bin/${{ matrix.build }}/whisper.dll

      - name: Upload binaries
        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
          name: whisper-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}
@ -384,18 +323,21 @@ jobs:
          - arch: Win32
            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
            s2arc: x86
+            clblast: OFF
          - arch: x64
            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x64.zip
            s2arc: x64
+            clblast: ON
+            clver: 1.6.1
          - sdl2: ON
            s2ver: 2.28.5

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Fetch OpenBLAS
        if: matrix.blas == 'ON'
@ -413,13 +355,26 @@ jobs:
          7z x sdl2.zip
          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV

+      - name: Install OpenCL
+        if: matrix.clblast == 'ON'
+        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
+
+      - name: Fetch CLBlast and set CLBlast_DIR
+        if: matrix.clblast == 'ON'
+        run: |
+          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
+          7z x clblast.zip
+          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
+          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
+
      - name: Configure
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_OPENBLAS=${{ matrix.blas }}
+          -DWHISPER_OPENBLAS=${{ matrix.blas }}
          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CLBLAST=${{ matrix.clblast }}

      - name: Build
        run: |
@ -434,15 +389,19 @@ jobs:
        if: matrix.sdl2 == 'ON'
        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}

+      - name: Copy clblast.dll
+        if: matrix.clblast == 'ON'
+        run: copy "$env:CLBlast_DIR/../../clblast.dll" build/bin/${{ matrix.build }}
+
      - name: Upload binaries
        if: matrix.blas == 'ON' && matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
-          name: whisper-blas-bin-${{ matrix.arch }}
+          name: whisper-blas${{ matrix.clblast == 'ON' && '-clblast' || ''}}-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}

  windows-cublas:
-    runs-on: windows-2019
+    runs-on: windows-latest

    strategy:
      matrix:
@ -459,14 +418,14 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
+        uses: microsoft/setup-msbuild@v1

      - name: Install CUDA Toolkit
        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
+        uses: Jimver/cuda-toolkit@v0.2.11
        with:
          cuda: '${{ matrix.cuda-toolkit }}'

@ -481,7 +440,7 @@ jobs:
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_CUDA=${{ matrix.cublas }}
+          -DWHISPER_CUBLAS=${{ matrix.cublas }}
          -DWHISPER_SDL2=${{ matrix.sdl2 }}

      - name: Build ${{ matrix.cuda-toolkit }}
@ -502,7 +461,7 @@ jobs:

      - name: Upload binaries
        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v1
        with:
          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}
@ -516,10 +475,10 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Setup emsdk
-        uses: mymindstorm/setup-emsdk@v14
+        uses: mymindstorm/setup-emsdk@v12

      - name: Verify
        run: emcc -v
@ -538,7 +497,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Configure
        run: |
@ -556,24 +515,24 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
        with:
          path: whisper

      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
        with:
          repository: ggerganov/ggml
          path: ggml

      - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v3
        with:
          distribution: zulu
-          java-version: 21
+          java-version: 17

      - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
+        uses: android-actions/setup-android@v2

      - name: Build
        run: |
@ -591,19 +550,20 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: set up JDK 11
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v3
        with:
          java-version: '11'
          distribution: 'temurin'
          cache: gradle

      - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
+        uses: android-actions/setup-android@v2
        with:
-          cmdline-tools-version: 9.0
+          api-level: 30
+          build-tools-version: 30.0.3

      - name: Build
        run: |
@ -615,16 +575,15 @@ jobs:
    needs: [ 'windows' ]
    runs-on: windows-latest
    steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v3

      - name: Install Java
-        uses: actions/setup-java@v4
+        uses: actions/setup-java@v1
        with:
-          distribution: zulu
-          java-version: 20
+          java-version: 17

      - name: Download Windows lib
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v3
        with:
          name: win32-x86-64_whisper.dll
          path: bindings/java/build/generated/resources/main/win32-x86-64
@ -637,7 +596,7 @@ jobs:
          ./gradlew build

      - name: Upload jar
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v3
        with:
          name: whispercpp.jar
          path: bindings/java/build/libs/whispercpp-*.jar
@ -659,7 +618,7 @@ jobs:

    steps:
      - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3

      - name: Test quantize
        run: |
--- a/.github/workflows/examples.yml
+++ b/.github/workflows/examples.yml
@ -37,7 +37,7 @@ jobs:
        run: npm install

      - name: Compile addon.node
-        run: npx cmake-js compile -T addon.node -B Release
+        run: npx cmake-js compile -T whisper-addon -B Release

      - name: Download test model
        run: |
--- a/.gitignore
+++ b/.gitignore
@ -6,11 +6,18 @@
 .vs/
 .vscode/
 .DS_Store
-.vimspector.json
-/CMakeSettings.json

 build/
-build-*/
+build-coreml/
+build-em/
+build-debug/
+build-release/
+build-rwdi/
+build-static/
+build-cublas/
+build-no-accel/
+build-sanitize-addr/
+build-sanitize-thread/

 # SPM
 .build/
@ -51,4 +58,4 @@ benchmark_results.csv
 cmake-build-debug/
 .cxx/
 .gradle/
-local.properties
+local.properties
--- a/.gitmodules
+++ b/.gitmodules
@ -0,0 +1,3 @@
+[submodule "bindings/ios"]
+	path = bindings/ios
+	url = https://github.com/ggerganov/whisper.spm
--- a/301
+++ b/301
@ -1,301 +0,0 @@
-# date: Tue Apr  9 20:27:03 EEST 2024
-# this file is auto-generated by scripts/gen-authors.sh
-
-0/0 <zero@imaskeleton.me>
-0cc4m <picard12@live.de>
-0xsourcecode <134374803+0xsourcecode@users.noreply.github.com>
-AT <manyoso@users.noreply.github.com>
-Aarni Koskela <akx@iki.fi>
-Aaron Pham <29749331+aarnphm@users.noreply.github.com>
-Aaron Taylor <aaron@exphat.com>
-Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
-Abitofevrything <54505189+abitofevrything@users.noreply.github.com>
-AfryMask <AfryMask@163.com>
-Ahmad Bilal <ahmad.bilal@empglabs.com>
-AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
-Akash Mahajan <akash7190@gmail.com>
-Akash Mahajan <akashmjn@stanford.edu>
-Al Hoang <3811822-hoanga@users.noreply.gitlab.com>
-Alan <unknown>
-Aleksander Andrzejewski <18704749+aleksanderandrzejewski@users.noreply.github.com>
-Alex Azarov <alex@azarov.by>
-Alex Bacart <13940752+alex-bacart@users.noreply.github.com>
-Alex Evgrashin <aevgrashin@yandex.ru>
-Alexandr Graschenkov <alexandr.graschenkov91@gmail.com>
-Alexandru Mariuti <alex@mariuti.com>
-Alexey Kharlamov <alexey@kharlamov.biz>
-Alfredo Montesinos <alfredo.montesinos@g.austincc.edu>
-Ali Alameh <ali.alameh@isae.edu.lb>
-Ananta Bastola <anantarajbastola@gmail.com>
-Andreu Huguet <andreuhuguet@gmail.com>
-Andrew Huynh <a5thuynh@gmail.com>
-Andrew S <andrews54757@gmail.com>
-Andy Maloney <asmaloney@gmail.com>
-Anton Kostin <masguit42@users.noreply.github.com>
-Artyom Mezin <psycho.fading@gmail.com>
-Asad Memon <asad.lionpk@gmail.com>
-Ashraful Islam <ashraful.meche@gmail.com>
-AsukaMinato <asukaminato@nyan.eu.org>
-AustinMroz <austinmroz@utexas.edu>
-Avik Sengupta <avik@sengupta.net>
-Bader-eddine Ouaich <49657842+baderouaich@users.noreply.github.com>
-Baffin Lee <baffinlee@gmail.com>
-Ben Nortier <bjnortier@gmail.com>
-Benjamin Heiniger <benjamin.heiniger@bluewin.ch>
-Bo-Yi Wu <appleboy.tw@gmail.com>
-Boris Bliznioukov <blib@mail.com>
-Borislav Stanimirov <b.stanimirov@abv.bg>
-Brad Murray <59848399+bradmurray-dt@users.noreply.github.com>
-Brian Murray <brian@bmurray.ca>
-CRD716 <crd716@gmail.com>
-Canis Lupus <Canis-UK@users.noreply.github.com>
-Carolinabanana <140120812+Carolinabanana@users.noreply.github.com>
-ChangSeok Oh <shivamidow@users.noreply.github.com>
-Chaoqun <27287694+OpenWaygate@users.noreply.github.com>
-Chia-Hsiang Cheng <88014292+garychia@users.noreply.github.com>
-Chidi Williams <williamschidi1@gmail.com>
-Christian <12550267+iceychris@users.noreply.github.com>
-Clifford Heath <clifford.heath@gmail.com>
-Colin <github@whoisc.cc>
-DGdev91 <DGdev91@users.noreply.github.com>
-Damian Czaja <trojan295@protonmail.com>
-Daniel Bevenius <daniel.bevenius@gmail.com>
-David <dnhkng@gmail.com>
-David Thorpe <djt@mutablelogic.com>
-Davidson Francis <davidsondfgl@gmail.com>
-Dener Stassun <denerstassun@gmail.com>
-Didzis Gosko <didzis@users.noreply.github.com>
-Digipom <admin@digipom.com>
-Dimo <dimo@ieee.org>
-Dody Suria Wijaya <dodysw@gmail.com>
-Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
-Duncan McConnell <ddmcconnell4@gmail.com>
-Egor Egorov <me@egorfine.com>
-Elkana Bardugo <ttv200@gmail.com>
-Emmanuel Schmidbauer <eschmidbauer@gmail.com>
-Engininja2 <139037756+Engininja2@users.noreply.github.com>
-Eric Swanson <eswanson@alloscomp.com>
-Eric Tendian <erictendian@gmail.com>
-Erik Scholz <Green-Sky@users.noreply.github.com>
-Evan Jones <evan.q.jones@gmail.com>
-Evan Martin <evan.martin@gmail.com>
-Eve <139727413+netrunnereve@users.noreply.github.com>
-Evgeny Kuznetsov <evgeny@kuznetsov.md>
-F1L1P <78918286+F1L1Pv2@users.noreply.github.com>
-Fangjun Kuang <csukuangfj@gmail.com>
-Felix <stenbackfelix@gmail.com>
-Finn Voorhees <finnvoorhees@gmail.com>
-FlippFuzz <41221030+FlippFuzz@users.noreply.github.com>
-Gang Chen <goncha@gmail.com>
-Gavin Cai <gavin1818@hotmail.com>
-George Hindle <george@georgehindle.com>
-Georgi Gerganov <ggerganov@gmail.com>
-GitAritron <103900385+GitAritron@users.noreply.github.com>
-GiviMAD <GiviMAD@users.noreply.github.com>
-Gleicon Moraes <gleicon@gmail.com>
-Gregor Jasny <gjasny@googlemail.com>
-Guillaume Wenzek <gwenzek@users.noreply.github.com>
-HY. Kelvin Lee <34256578+hykelvinlee42@users.noreply.github.com>
-Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
-Hang <bebound@gmail.com>
-Herman Semenov <GermanAizek@yandex.ru>
-Hrishikesh Barman <geekodour@users.noreply.github.com>
-Ian Bicking <ian@ianbicking.org>
-Ian Bull <irbull@eclipsesource.com>
-Ikko Ashimine <eltociear@gmail.com>
-InconsolableCellist <23345188+InconsolableCellist@users.noreply.github.com>
-Ismatulla Mansurov <47342870+sapoepsilon@users.noreply.github.com>
-Ivan Gorin <ivangorin21@gmail.com>
-JJ <103335846+computerscienceiscool@users.noreply.github.com>
-Jack Mousseau <jmousseau@users.noreply.github.com>
-JacobLinCool <jacoblincool@gmail.com>
-Jakub Ráček <blizzcz@gmail.com>
-Jared Van Bortel <jared@nomic.ai>
-Jay Binks <jaybinks@gmail.com>
-Jhen-Jie Hong <developer@jhen.me>
-Jhen-Jie Hong <iainst0409@gmail.com>
-JidongZhang-THU <1119708529@qq.com>
-Jo Liss <joliss42@gmail.com>
-Johan <jr.raffin@gmail.com>
-Johannes Gäßler <johannesg@5d6.de>
-John Balis <phobossystems@gmail.com>
-Jonathan Soo <jcsoo@agora.com>
-Jonno <1160532+razodactyl@users.noreply.github.com>
-Joonas Pihlajamaa <joonas.pihlajamaa@iki.fi>
-Jose <34888496+Jerry-Master@users.noreply.github.com>
-Josh Bleecher Snyder <josharian@gmail.com>
-Judd <foldl@users.noreply.github.com>
-Jumper775 <78500318+jumpers775@users.noreply.github.com>
-Justine Tunney <jtunney@gmail.com>
-KP Kaiser <kirk@zothcorp.com>
-Kamilake <exjang0@gmail.com>
-Kartik Saranathan <278928+Kartiku@users.noreply.github.com>
-Kasumi <90275229+kasumi-1@users.noreply.github.com>
-Kawrakow <48489457+ikawrakow@users.noreply.github.com>
-Kevin Brothaler <admin@digipom.com>
-Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
-Kreijstal <rainb@tfwno.gf>
-Kylin <56434533+KyL0N@users.noreply.github.com>
-LBlue <153975653+lbluep@users.noreply.github.com>
-Larry Battle <larry.battle.tech@gmail.com>
-Laytan Laats <laytanlaats@hotmail.com>
-Leo Moll <leo.moll@yeasoft.com>
-Lexevolution <31176843+Lexevolution@users.noreply.github.com>
-LittleLoli <26589867+WhichWho@users.noreply.github.com>
-Lucas Zanek <57494138+LucasZNK@users.noreply.github.com>
-Luis Herrera <herrera-luis@users.noreply.github.com>
-Lukas Rist <glaslos@gmail.com>
-M. A. Ali <73258591+MightyStud@users.noreply.github.com>
-M. Eren Akbiyik <erenakbiyik@gmail.com>
-Maciek <maciek.mab122@gmail.com>
-Marcin Mielniczuk <marmistrz.dev@zoho.eu>
-Martin Warnaar <martinwarnaar@gmail.com>
-Matheus de Sousa <23645013+keyehzy@users.noreply.github.com>
-Mathijs de Bruin <mathijs@mathijsfietst.nl>
-Matija Pevec <mightymatth@users.noreply.github.com>
-Maximiliano Levi <8160966+maxilevi@users.noreply.github.com>
-Meng, Hengyu <hengyu.meng@intel.com>
-Michael Podvitskiy <podvitskiymichael@gmail.com>
-Michael Rienstra <mrienstra@gmail.com>
-Mikhail Grigorev <sleuthhound@gmail.com>
-Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
-Mohit Agarwal <mohit@sdf.org>
-Murilo Santana <mvrilo@gmail.com>
-Neil Chudleigh <nchudleigh@users.noreply.github.com>
-Neo Zhang Jianyu <jianyu.zhang@intel.com>
-Neuman Vong <neuman.vong@gmail.com>
-Nicholas Albion <nalbion@yahoo.com>
-Niels Mayer <Niels.Mayer@gmail.com>
-Okabintaro <103938900+Okabintaro@users.noreply.github.com>
-Oleg Sidorov <me@whitebox.io>
-Oleg Sidorov <oleg@sidorov.nl>
-Ondrej Kokes <ondrej.kokes@gmail.com>
-Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
-Paul Tsochantaris <ptsochantaris@icloud.com>
-Philipp Zabel <philipp.zabel@gmail.com>
-Philippe Normand <phil@base-art.net>
-Przemysław Pawełczyk <przemoc@gmail.com>
-Qianhe Chen <54462604+chenqianhe@users.noreply.github.com>
-Radosław Gryta <radek.gryta@gmail.com>
-Reinforce-II <fate@eastal.com>
-Reinis Muiznieks <muiznieks.reinis@gmail.com>
-RelatedTitle <r3latedtitle@gmail.com>
-RhinoDevel <RhinoDevel@users.noreply.github.com>
-Rich Jones <miserlou@gmail.com>
-Robin <robin.xw@hotmail.com>
-Roddur Dasgupta <roddurd@gmail.com>
-Roland Rabien <figbug@gmail.com>
-Rotem Dan <rotemdan@gmail.com>
-Ryan Hitchman <hitchmanr@gmail.com>
-Ryan Metcalfe <107415876+RyanMetcalfeInt8@users.noreply.github.com>
-RyanChang <ftes90015@gmail.com>
-Sam <49637763+Onlyartist9@users.noreply.github.com>
-Sam Pullara <spullara@gmail.com>
-Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
-Sergio López <slp@sinrega.org>
-Siddharth Ramakrishnan <srr2141@columbia.edu>
-Simon Moisselin <simon.moisstoll@gmail.com>
-Sindre Sorhus <sindresorhus@gmail.com>
-Slava Primenko <primenko.s@gmail.com>
-Syahmi Azhar <prsyahmi@gmail.com>
-Syed Jafri <syedjafri97@gmail.com>
-Sơn Phan Trung <phantrungson17@gmail.com>
-Taisei Mima <bhbstar.me@gmail.com>
-Takeshi Inoue <inoue.takeshi@gmail.com>
-Tamotsu Takahashi <ttakah+github@gmail.com>
-Taras Glek <taras@thegp.com>
-Tauseef Mohiuddin <35351464+tauseefmohammed2@users.noreply.github.com>
-Thijs Raymakers <thijs@raymakers.nl>
-Thomas Fitzsimmons <fitzsim@fitzsim.org>
-Tiago Fassoni <tiagofassoni@users.noreply.github.com>
-Tienshiao Ma <tienshiao@tienshiao.org>
-Timothy Cronin <40186632+4imothy@users.noreply.github.com>
-Tobrun <tobrun.van.nuland@gmail.com>
-Todd <taf2@users.noreply.github.com>
-Tong Li <31761981+litongjava@users.noreply.github.com>
-Topping1 <78745143+Topping1@users.noreply.github.com>
-Travis Cline <travis.cline@gmail.com>
-UEXTM.com <84163508+uextm@users.noreply.github.com>
-Vadim Peretokin <vperetokin@hey.com>
-Valentin Gosu <1454649+valenting@users.noreply.github.com>
-Vulcan <93451215+trholding@users.noreply.github.com>
-WhiteOlivierus <36532695+WhiteOlivierus@users.noreply.github.com>
-Xiang (Kevin) Li <kevinli020508@gmail.com>
-Xiao-Yong Jin <jinxiaoyong@gmail.com>
-XiaotaoChen <chenxiaotao1234@gmail.com>
-Yajing Tang <phillis@google.com>
-Yang Shen <aplshenyang@gmail.com>
-Yunès <jean.baptiste.yunes@free.fr>
-ZaBlazzingZephyrus <119159668+blazingzephyr@users.noreply.github.com>
-Zigfrid Zvezdin <ziggerZZ@gmail.com>
-Zollner <24618122+Zolliner@users.noreply.github.com>
-ai-at-home <149282006+ai-at-home@users.noreply.github.com>
-alonfaraj <alonfaraj@gmail.com>
-andypayne <apayne@gmail.com>
-ardfork <134447697+ardfork@users.noreply.github.com>
-automaticcat <daogiatuank54@gmail.com>
-be-next <jerome.ramette@gmail.com>
-bert hubert <bert@hubertnet.nl>
-bmwl <brian.marshall@tolko.com>
-bobqianic <129547291+bobqianic@users.noreply.github.com>
-bocytko <bocytko+github@gmail.com>
-boolemancer <48014766+boolemancer@users.noreply.github.com>
-boolemancer <boolemancer@gmail.com>
-bradmit <151883577+bradmit@users.noreply.github.com>
-brunofaustino <b.fa.amorim@gmail.com>
-bssrdf <merlintiger@hotmail.com>
-byte-6174 <88070277+byte-6174@users.noreply.github.com>
-cdosoftei <ciprian.dosoftei@gmail.com>
-clach04 <Chris.Clark@actian.com>
-compilade <113953597+compilade@users.noreply.github.com>
-conradg <conradjgodfrey@gmail.com>
-ddpasa <112642920+ddpasa@users.noreply.github.com>
-denersc <denerstassun@gmail.com>
-dscripka <dscripka@users.noreply.github.com>
-duthils <duthils@duthils.net>
-ecneladis <ecneladis@users.noreply.github.com>
-faker <nspyia2002@gmail.com>
-fitzsim <fitzsim@fitzsim.org>
-fraxy-v <65565042+fraxy-v@users.noreply.github.com>
-genevera (she/her) <genevera@users.noreply.github.com>
-geniusnut <geniusnut@gmail.com>
-greeshmay <greeshmay@gmail.com>
-hydai <z54981220@gmail.com>
-iamthad <thadeus.j.fleming@gmail.com>
-james wolf <contractorwolf@hotmail.com>
-joecryptotoo <80373433+joecryptotoo@users.noreply.github.com>
-jorismertz <35079666+jorismertz@users.noreply.github.com>
-junkfood <69683722+JunkFood02@users.noreply.github.com>
-jwijffels <jwijffels@bnosac.be>
-kamranjon <kamranjon@gmail.com>
-katsu560 <katsu560oo-@docomo.ne.jp>
-kennethge <57784063+kenneth-ge@users.noreply.github.com>
-keyehzy <msamuel@aluno.puc-rio.br>
-leejet <leejet714@gmail.com>
-litong <31761981+litongjava@users.noreply.github.com>
-lnyan <lkwq007@gmail.com>
-m.bell <m.bell@techsmith.com>
-mkiol <mkiol@users.noreply.github.com>
-novag <7754358+novag@users.noreply.github.com>
-pajowu <pajowu@pajowu.de>
-polarmoon <90010972+polarmoon@users.noreply.github.com>
-rlapray <lapray.romain@gmail.com>
-sandrohanea <40202887+sandrohanea@users.noreply.github.com>
-semiformal-net <84111142+semiformal-net@users.noreply.github.com>
-shibukazu <61775791+shibukazu@users.noreply.github.com>
-shikokuchuo <53399081+shikokuchuo@users.noreply.github.com>
-slaren <slarengh@gmail.com>
-slashlib <slashlib@users.noreply.github.com>
-snadampal <87143774+snadampal@users.noreply.github.com>
-st-gr <38470677+st-gr@users.noreply.github.com>
-texmex76 <40733439+texmex76@users.noreply.github.com>
-thefinaldegree <thefinaldegree@gmail.com>
-trixirt <trix@redhat.com>
-ulatekh <ulatekh@yahoo.com>
-undef <undefdev@gmail.com>
-venkr <venkateshrameshkumar+1@gmail.com>
-vicalloy <zbirder@gmail.com>
-xdrudis <xavierdrudis@yahoo.es>
-zhouwg <6889919+zhouwg@users.noreply.github.com>
-布客飞龙 <562826179@qq.com>
-Артём Земляк <azemlyak@smart-consulting.ru>
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -1,31 +1,22 @@
-cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
-project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.6.2)
-include(CheckIncludeFileCXX)
+cmake_minimum_required (VERSION 3.5)

+project(whisper.cpp VERSION 1.5.4)
 set(SOVERSION 1)

-#set(CMAKE_WARN_DEPRECATED YES)
-set(CMAKE_WARN_UNUSED_CLI YES)
-
-set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
-
-if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
-    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
-    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
-endif()
-
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")

 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)

-if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
    set(WHISPER_STANDALONE ON)
-
-    include(git-vars)
+    include(GitVars)
+    include(BuildTypes)

    # configure project version
+    if (EXISTS "${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl")
+        configure_file(${CMAKE_SOURCE_DIR}/bindings/ios/Makefile-tmpl ${CMAKE_SOURCE_DIR}/bindings/ios/Makefile @ONLY)
+    endif()
    configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
 else()
    set(WHISPER_STANDALONE OFF)
@ -35,11 +26,6 @@ if (EMSCRIPTEN)
    set(BUILD_SHARED_LIBS_DEFAULT OFF)

    option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
-
-    # TODO: without these, we get the following error:
-    #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
-    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
 else()
    if (MINGW)
        set(BUILD_SHARED_LIBS_DEFAULT OFF)
@ -48,145 +34,611 @@ else()
    endif()
 endif()

-option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+# options

-#
-# option list
-#
+if (APPLE)
+    set(WHISPER_METAL_DEFAULT ON)
+else()
+    set(WHISPER_METAL_DEFAULT OFF)
+endif()

-# general
-option(WHISPER_CCACHE "whisper: use ccache if available" ON)
+option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})

-# debug
 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
 option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)

-# build
-option(WHISPER_FATAL_WARNINGS "whisper: enable -Werror flag" OFF)
+option(WHISPER_SANITIZE_THREAD        "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS       "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED     "whisper: enable undefined sanitizer" OFF)

-# sanitizers
-option(WHISPER_SANITIZE_THREAD    "whisper: enable thread sanitizer"    OFF)
-option(WHISPER_SANITIZE_ADDRESS   "whisper: enable address sanitizer"   OFF)
-option(WHISPER_SANITIZE_UNDEFINED "whisper: enable undefined sanitizer" OFF)
+option(WHISPER_BUILD_TESTS            "whisper: build tests"    ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES         "whisper: build examples" ${WHISPER_STANDALONE})

-# extra artifacts
-option(WHISPER_BUILD_TESTS    "whisper: build tests"          ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_EXAMPLES "whisper: build examples"       ${WHISPER_STANDALONE})
-option(WHISPER_BUILD_SERVER   "whisper: build server example" ${WHISPER_STANDALONE})
+option(WHISPER_SDL2                   "whisper: support for libSDL2" OFF)

-# 3rd party libs
-option(WHISPER_CURL "whisper: use libcurl to download model from an URL" OFF)
-option(WHISPER_SDL2 "whisper: support for libSDL2" OFF)
+option(WHISPER_NO_AVX                 "whisper: disable AVX"  OFF)
+option(WHISPER_NO_AVX2                "whisper: disable AVX2" OFF)
+option(WHISPER_NO_FMA                 "whisper: disable FMA"  OFF)
+option(WHISPER_NO_F16C                "whisper: disable F16c" OFF)

-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    option(WHISPER_FFMPEG "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
+option(WHISPER_OPENVINO               "whisper: support for OpenVINO" OFF)
+
+if (APPLE)
+    option(WHISPER_NO_ACCELERATE         "whisper: disable Accelerate framework" OFF)
+    option(WHISPER_METAL                 "whisper: use Metal"                    ${WHISPER_METAL_DEFAULT})
+    option(WHISPER_METAL_NDEBUG          "whisper: disable Metal debugging"      OFF)
+    option(WHISPER_COREML                "whisper: enable Core ML framework"     OFF)
+    option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback"    OFF)
+    option(WHISPER_METAL_EMBED_LIBRARY   "whisper: embed Metal library"          OFF)
+else()
+    option(WHISPER_BLAS                  "whisper: use BLAS libraries"                        OFF)
+    option(WHISPER_BLAS_VENDOR           "whisper: BLAS library vendor"                       Generic)
+    option(WHISPER_OPENBLAS              "whisper: prefer OpenBLAS"                           OFF)
+    option(WHISPER_CUBLAS                "whisper: support for cuBLAS"                        OFF)
+    option(WHISPER_HIPBLAS               "whisper: support for hipBLAS"                       OFF)
+    option(WHISPER_CLBLAST               "whisper: use CLBlast"                               OFF)
+    option(WHISPER_SYCL                  "whisper: use SYCL"                                  OFF)
+    option(WHISPER_SYCL_F16              "whisper: use 16 bit floats for sycl calculations"   OFF)
 endif()

-option(WHISPER_COREML                "whisper: enable Core ML framework"  OFF)
-option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
-option(WHISPER_OPENVINO              "whisper: support for OpenVINO"      OFF)
+option(WHISPER_PERF "whisper: enable perf timings" OFF)

-# Required for relocatable CMake package
-include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+# sanitizers

-# override ggml options
-set(GGML_CCACHE             ${WHISPER_CCACHE})
-set(GGML_SANITIZE_THREAD    ${WHISPER_SANITIZE_THREAD})
-set(GGML_SANITIZE_ADDRESS   ${WHISPER_SANITIZE_ADDRESS})
-set(GGML_SANITIZE_UNDEFINED ${WHISPER_SANITIZE_UNDEFINED})
-set(GGML_ALL_WARNINGS       ${WHISPER_ALL_WARNINGS})
-set(GGML_FATAL_WARNINGS     ${WHISPER_FATAL_WARNINGS})
-
-# transition helpers
-function (whisper_option_depr TYPE OLD NEW)
-    if (${OLD})
-        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
-        set(${NEW} ON)
+if (NOT MSVC)
+    if (WHISPER_SANITIZE_THREAD)
+        set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
    endif()
-endfunction()

-whisper_option_depr(FATAL_ERROR WHISPER_CUBLAS              GGML_CUDA)
-whisper_option_depr(WARNING     WHISPER_CUDA                GGML_CUDA)
-whisper_option_depr(WARNING     WHISPER_KOMPUTE             GGML_KOMPUTE)
-whisper_option_depr(WARNING     WHISPER_METAL               GGML_METAL)
-whisper_option_depr(WARNING     WHISPER_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
-whisper_option_depr(WARNING     WHISPER_NATIVE              GGML_NATIVE)
-whisper_option_depr(WARNING     WHISPER_OPENMP              GGML_OPENMP)
-whisper_option_depr(WARNING     WHISPER_RPC                 GGML_RPC)
-whisper_option_depr(WARNING     WHISPER_SYCL                GGML_SYCL)
-whisper_option_depr(WARNING     WHISPER_SYCL_F16            GGML_SYCL_F16)
+    if (WHISPER_SANITIZE_ADDRESS)
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=address -fno-omit-frame-pointer")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
+    endif()
+
+    if (WHISPER_SANITIZE_UNDEFINED)
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=undefined")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
+    endif()
+endif()
+
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
+
+# dependencies
+
+find_package(Threads REQUIRED)
+
+#compile flag sycl
+if (WHISPER_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
+
+# on APPLE
+if (APPLE)
+    # include Accelerate framework
+    if (NOT WHISPER_NO_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
+
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            set(WHISPER_EXTRA_LIBS  ${WHISPER_EXTRA_LIBS}  ${ACCELERATE_FRAMEWORK})
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64)
+        else()
+            message(FATAL_ERROR "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (WHISPER_METAL)
+        find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+        find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+        find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
+
+        if (METAL_FRAMEWORK)
+            message(STATUS "Metal framework found")
+
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS}
+                ${FOUNDATION_LIBRARY}
+                ${METAL_FRAMEWORK}
+                ${METALKIT_FRAMEWORK}
+                )
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_METAL)
+
+            if (WHISPER_METAL_NDEBUG)
+                set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_NDEBUG)
+            endif()
+        else()
+            message(FATAL_ERROR "Metal framework not found")
+        endif()
+
+        set(GGML_SOURCES_METAL ggml-metal.m ggml-metal.h)
+
+        # copy ggml-metal.metal to bin directory
+        configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
+
+        if (WHISPER_METAL_EMBED_LIBRARY)
+            enable_language(ASM)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_METAL_EMBED_LIBRARY)
+
+            set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
+
+            file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+            set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
+
+            add_custom_command(
+                OUTPUT ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
+                COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
+                DEPENDS ${METALLIB_SOURCE}
+                COMMENT "Generate assembly for embedded Metal library"
+            )
+
+            set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
+        endif()
+    endif()
+
+    if (WHISPER_COREML)
+        find_library(FOUNDATION_FRAMEWORK Foundation)
+        find_library(COREML_FRAMEWORK CoreML)
+
+        if (COREML_FRAMEWORK)
+            message(STATUS "CoreML framework found")
+
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+        else()
+            message(FATAL_ERROR "CoreML framework not found")
+        endif()
+
+        if (WHISPER_COREML_ALLOW_FALLBACK)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
+        endif()
+    endif()
+endif()
+
+if (WHISPER_OPENBLAS)
+    set(WHISPER_BLAS_VENDOR "OpenBLAS")
+    set(WHISPER_BLAS ON)
+endif()
+
+if (WHISPER_BLAS)
+    if (WIN32)
+        if(DEFINED ENV{OPENBLAS_PATH})
+            set(BLAS_LIBRARIES $ENV{OPENBLAS_PATH}/lib/libopenblas.dll.a)
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories($ENV{OPENBLAS_PATH}/include)
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else ()
+            message(FATAL_ERROR "BLAS library was not found. Environment variable OPENBLAS_PATH not defined.")
+        endif ()
+    else ()
+        set(BLA_STATIC 1)
+        set(BLA_VENDOR ${WHISPER_BLAS_VENDOR})
+        set(BLA_SIZEOF_INTEGER 8)
+        set(BLA_PREFER_PKGCONFIG 1)
+        find_package(BLAS)
+
+        if(BLAS_FOUND)
+            message(STATUS "BLAS compatible library found")
+            message(STATUS "Libraries ${BLAS_LIBRARIES}")
+            find_path(BLAS_INCLUDE_DIRS cblas.h /usr/include/openblas /usr/local/include/openblas $ENV{BLAS_HOME}/include)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_OPENBLAS)
+            include_directories(${BLAS_INCLUDE_DIRS})
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
+        else()
+            message(FATAL_ERROR "BLAS library was not found")
+        endif()
+    endif ()
+endif ()
+
+if (WHISPER_CUBLAS)
+    cmake_minimum_required(VERSION 3.17)
+
+    find_package(CUDAToolkit)
+
+    if (CUDAToolkit_FOUND)
+        message(STATUS "cuBLAS found")
+
+        enable_language(CUDA)
+
+        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+
+        add_compile_definitions(GGML_USE_CUBLAS)
+
+        if (WHISPER_STATIC)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
+        else()
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
+        endif()
+
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
+    else()
+        message(FATAL_ERROR "cuBLAS not found")
+    endif()
+endif()
+
+
+if (WHISPER_HIPBLAS)
+    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    endif()
+    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+    endif()
+
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
+
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        set_property(TARGET ggml-rocm PROPERTY POSITION_INDEPENDENT_CODE ON)
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+
+        if (WHISPER_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(FATAL_ERROR "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    endif()
+endif()
+
+if (WHISPER_CLBLAST)
+    find_package(CLBlast)
+    if (CLBlast_FOUND)
+        message(STATUS "CLBlast found")
+
+        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
+
+        add_compile_definitions(GGML_USE_CLBLAST)
+
+        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} clblast)
+    else()
+        message(FATAL_ERROR "CLBlast not found")
+    endif()
+endif()
+
+if( WHISPER_OPENVINO )
+    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+endif()
+
+if (WHISPER_SYCL)
+    if ( NOT DEFINED ENV{ONEAPI_ROOT})
+        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
+    endif()
+    #todo: AOT
+
+    find_package(IntelSYCL REQUIRED)
+    if (WHISPER_SYCL_F16)
+        add_compile_definitions(GGML_SYCL_F16)
+    endif()
+    add_compile_definitions(GGML_USE_SYCL)
+
+    add_compile_options(-I./) #include DPCT
+    add_compile_options(-I/${SYCL_INCLUDE_DIR})
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
+
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
+    set(GGML_SOURCES_SYCL ggml-sycl.cpp)
+
+    set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+endif()
+# compiler flags
+
+if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
+endif ()
+
+if (WHISPER_ALL_WARNINGS)
+    if (NOT MSVC)
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} \
+            -Wall                           \
+            -Wextra                         \
+            -Wpedantic                      \
+            -Wshadow                        \
+            -Wcast-qual                     \
+            -Wstrict-prototypes             \
+            -Wpointer-arith                 \
+            -Wno-unused-function            \
+        ")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} \
+            -Wall                           \
+            -Wextra                         \
+            -Wpedantic                      \
+            -Wcast-qual                     \
+        ")
+    else()
+        # todo : msvc
+    endif()
+endif()
+
+if (NOT MSVC)
+    # TODO: temporary disabled until we figure out ggml-metal.m
+    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -Werror=vla")
+    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
+endif()
+
+message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+
+if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+    message(STATUS "ARM detected")
+elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+    message(STATUS "PowerPC detected")
+else()
+    message(STATUS "x86 detected")
+    if (MSVC)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /utf-8")
+        set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /utf-8")
+        set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /utf-8")
+        if(NOT WHISPER_NO_AVX2)
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX2")
+            set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX2")
+            set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX2")
+        else()
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /arch:AVX")
+                set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} /arch:AVX")
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} /arch:AVX")
+            endif()
+        endif()
+    else()
+        if (EMSCRIPTEN)
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
+        else()
+            if(NOT WHISPER_NO_AVX)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
+            endif()
+            if(NOT WHISPER_NO_AVX2)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx2")
+            endif()
+            if(NOT WHISPER_NO_FMA)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mfma")
+            endif()
+            if(NOT WHISPER_NO_F16C)
+                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mf16c")
+            endif()
+        endif()
+    endif()
+endif()

 #
-# build the library
+# POSIX conformance
 #

-add_subdirectory(ggml)
-add_subdirectory(src)
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
+if (WHISPER_PERF)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
+endif()

 #
-# install
+# whisper.coreml - Core ML support
 #

+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        )
+endif()
+
+if (WHISPER_OPENVINO)
+    set(TARGET whisper.openvino)
+
+    add_library(${TARGET} OBJECT
+        openvino/whisper-openvino-encoder.h
+        openvino/whisper-openvino-encoder.cpp
+        )
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
+
+    target_link_libraries(${TARGET} PRIVATE openvino::runtime)
+endif()
+
+#
+# whisper - this is the main library of the project
+#
+
+set(TARGET whisper)
+
+add_library(${TARGET}
+    ggml.h
+    ggml.c
+    ggml-alloc.h
+    ggml-alloc.c
+    ggml-backend.h
+    ggml-backend.c
+    ggml-quants.h
+    ggml-quants.c
+    ${GGML_SOURCES_METAL}
+    ${GGML_SOURCES_CUDA}
+    ${GGML_SOURCES_OPENCL}
+    ${GGML_SOURCES_SYCL}
+    ${GGML_HEADERS_SYCL}
+    whisper.h
+    whisper.cpp
+    )
+
+# Set the version numbers
+set_target_properties(whisper PROPERTIES
+    VERSION ${PROJECT_VERSION}
+    SOVERSION ${SOVERSION}
+)
+
+include(DefaultTargetOptions)
+
+target_include_directories(${TARGET} PUBLIC
+    .
+    )
+
+if (WHISPER_COREML)
+    target_link_libraries(${TARGET} PRIVATE whisper.coreml)
+endif()
+
+if (WHISPER_OPENVINO)
+    target_link_libraries(${TARGET} PRIVATE whisper.openvino)
+endif()
+
+if (MSVC)
+    target_link_libraries(${TARGET} PRIVATE ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
+
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -D_CRT_SECURE_NO_WARNINGS)
+else()
+    target_link_libraries(${TARGET} PRIVATE m ${WHISPER_EXTRA_LIBS} ${CMAKE_THREAD_LIBS_INIT})
+endif()
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_link_libraries(${TARGET} PUBLIC
+        ${CMAKE_DL_LIBS}
+        )
+
+    target_compile_definitions(${TARGET} PUBLIC
+        WHISPER_SHARED
+        GGML_SHARED
+        )
+
+    target_compile_definitions(${TARGET} PRIVATE
+        WHISPER_BUILD
+        GGML_BUILD
+        )
+
+    if (WHISPER_METAL)
+        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
+        #       but for some reason it does not work here like it does in llama.cpp
+        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    endif()
+endif()
+
+if (GGML_SOURCES_CUDA)
+    message(STATUS "GGML CUDA sources found, configuring CUDA architecture")
+    # Only configure gmml CUDA architectures is not globally set
+    if (NOT DEFINED GGML_CUDA_ARCHITECTURES)
+        # Not overriden by user, so set defaults
+        set(GGML_CUDA_ARCHITECTURES 52 61 70)
+    endif()
+    message(STATUS "GGML Configuring CUDA architectures ${GGML_CUDA_ARCHITECTURES}")
+    set_property(TARGET whisper PROPERTY CUDA_ARCHITECTURES ${GGML_CUDA_ARCHITECTURES})
+    set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
+endif()
+
+if (EMSCRIPTEN)
+    set_target_properties(${TARGET} PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()
+
+target_compile_definitions(${TARGET} PUBLIC
+    ${WHISPER_EXTRA_FLAGS}
+    )
+
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
+
 include(GNUInstallDirs)
-include(CMakePackageConfigHelpers)

-set(WHISPER_BUILD_NUMBER        ${BUILD_NUMBER})
-set(WHISPER_BUILD_COMMIT        ${BUILD_COMMIT})
-set(WHISPER_INSTALL_VERSION     ${CMAKE_PROJECT_VERSION})
+install(TARGETS ${TARGET}
+    LIBRARY  DESTINATION lib
+    ARCHIVE  DESTINATION lib/static
+    RUNTIME  DESTINATION bin
+    RESOURCE DESTINATION bin
+    PUBLIC_HEADER DESTINATION include
+    )

-set(WHISPER_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
-set(WHISPER_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
-set(WHISPER_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
+#
+# bindings
+#

-get_directory_property(WHISPER_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
-
-set_target_properties(whisper PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/whisper.h)
-install(TARGETS whisper LIBRARY PUBLIC_HEADER)
-
-configure_package_config_file(
-        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/whisper-config.cmake.in
-        ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
-    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper
-    PATH_VARS
-    WHISPER_INCLUDE_INSTALL_DIR
-    WHISPER_LIB_INSTALL_DIR
-    WHISPER_BIN_INSTALL_DIR )
-
-write_basic_package_version_file(
-    ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
-    VERSION ${WHISPER_INSTALL_VERSION}
-    COMPATIBILITY SameMajorVersion)
-
-install(FILES ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
-              ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
-        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper)
-
-install(
-    FILES convert-hf-to-gguf.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
-
-configure_file(cmake/whisper.pc.in
-        "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
-        @ONLY)
-
-install(FILES "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
-        DESTINATION lib/pkgconfig)
+add_subdirectory(bindings)

 #
 # programs, examples and tests
 #

 if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
-    #include(CTest)
-    #add_subdirectory(tests)
+    enable_testing()
+    add_subdirectory(tests)
 endif ()

 if (WHISPER_BUILD_EXAMPLES)
--- a/2
+++ b/2
@ -1,6 +1,6 @@
 MIT License

-Copyright (c) 2023-2024 The ggml authors
+Copyright (c) 2023 Georgi Gerganov

 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
--- a/1222
+++ b/1222
--- a/Package.swift
+++ b/Package.swift
@ -27,15 +27,17 @@ let package = Package(
               "samples",
               "tests",
               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
               "Makefile"
            ],
            sources: [
-                "ggml/src/ggml.c",
-                "src/whisper.cpp",
-                "ggml/src/ggml-alloc.c",
-                "ggml/src/ggml-backend.c",
-                "ggml/src/ggml-quants.c",
-                "ggml/src/ggml-metal.m"
+                "ggml.c",
+                "whisper.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c",
+                "ggml-metal.m"
            ],
            resources: [.process("ggml-metal.metal")],
            publicHeadersPath: "spm-headers",
--- a/README.md
+++ b/README.md
@ -4,10 +4,9 @@

 [![Actions Status](https://github.com/ggerganov/whisper.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/whisper.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![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: [v1.6.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.6.0) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)

 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:

@ -20,6 +19,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
 - Zero memory allocations at runtime
 - Support for CPU-only inference
 - [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
 - [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)

@ -414,13 +414,35 @@ For more information about the Core ML implementation please refer to PR [#1037]
 With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads

-Now build `whisper.cpp` with CUDA support:
+Now build `whisper.cpp` with cuBLAS support:

 ```
 make clean
-GGML_CUDA=1 make -j
+WHISPER_CUBLAS=1 make -j
 ```

+## OpenCL GPU support via CLBlast
+
+For cards and integrated GPUs that support OpenCL, the Encoder processing can be largely offloaded to the GPU through CLBlast. This is especially useful for users with AMD APUs or low end devices for up to ~2x speedup.
+
+First, make sure you have installed `CLBlast` for your OS or Distribution: https://github.com/CNugteren/CLBlast
+
+Now build `whisper.cpp` with CLBlast support:
+
+```
+Makefile:
+cd whisper.cpp
+make clean
+WHISPER_CLBLAST=1 make -j
+
+CMake:
+cd whisper.cpp
+cmake -B build -DWHISPER_CLBLAST=ON
+cmake --build build -j --config Release
+```
+
+Run all the examples as usual.
+
 ## BLAS CPU support via OpenBLAS

 Encoder processing can be accelerated on the CPU via OpenBLAS.
@ -430,22 +452,7 @@ Now build `whisper.cpp` with OpenBLAS support:

 ```
 make clean
-GGML_OPENBLAS=1 make -j
-```
-
-## BLAS CPU support via Intel MKL
-
-Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
-First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
-
-Now build `whisper.cpp` with Intel MKL BLAS support:
-
-```
-source /opt/intel/oneapi/setvars.sh
-mkdir build
-cd build
-cmake -DWHISPER_MKL=ON ..
-WHISPER_MKL=1 make -j
+WHISPER_OPENBLAS=1 make -j
 ```

 ## Docker
@ -480,16 +487,6 @@ docker run -it --rm \
  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:
-
-```
-conan install --requires="whisper-cpp/[*]" --build=missing
-```
-
-For detailed instructions on how to use Conan, please refer to the [Conan documentation](https://docs.conan.io/2/).
-
 ## Limitations

 - Inference only
@ -698,7 +695,7 @@ The [main](examples/main) example provides support for output of karaoke-style m
 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:
+Here are a few *"typical"* examples:

 ```bash
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
@ -732,10 +729,10 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a

 ## Video comparison of different models

-Use the [scripts/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/scripts/bench-wts.sh) script to generate a video in the following format:
+Use the [extra/bench-wts.sh](https://github.com/ggerganov/whisper.cpp/blob/master/extra/bench-wts.sh) script to generate a video in the following format:

 ```bash
-./scripts/bench-wts.sh samples/jfk.wav
+./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```

@ -756,7 +753,7 @@ Additionally a script to run whisper.cpp with different models and audio files i
 You can run it with the following command, by default it will run against any standard model in the models folder.

 ```bash
-python3 scripts/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
+python3 extra/bench.py -f samples/jfk.wav -t 2,4,8 -p 1,2
 ```

 It is written in python with the intention of being easy to modify and extend for your benchmarking use case.
@ -796,7 +793,6 @@ For more details, see the conversion script [models/convert-pt-to-ggml.py](model
  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
 - [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
-  - [AIWintermuteAI/whispercpp](https://github.com/AIWintermuteAI/whispercpp) (Updated fork of aarnphm/whispercpp)
  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
 - [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
 - [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
--- a/bindings/go/examples/go-whisper/flags.go
+++ b/bindings/go/examples/go-whisper/flags.go
@ -68,6 +68,10 @@ func (flags *Flags) GetOut() string {
 	return strings.ToLower(flags.Lookup("out").Value.String())
 }

+func (flags *Flags) IsSpeedup() bool {
+	return flags.Lookup("speedup").Value.String() == "true"
+}
+
 func (flags *Flags) IsTokens() bool {
 	return flags.Lookup("tokens").Value.String() == "true"
 }
@ -107,6 +111,10 @@ func (flags *Flags) SetParams(context whisper.Context) error {
 		fmt.Fprintf(flags.Output(), "Setting duration to %v\n", duration)
 		context.SetDuration(duration)
 	}
+	if flags.IsSpeedup() {
+		fmt.Fprintf(flags.Output(), "Setting speedup to true\n")
+		context.SetSpeedup(true)
+	}
 	if threads := flags.GetThreads(); threads != 0 {
 		fmt.Fprintf(flags.Output(), "Setting threads to %d\n", threads)
 		context.SetThreads(threads)
@ -138,6 +146,7 @@ func registerFlags(flag *Flags) {
 	flag.Duration("offset", 0, "Time offset")
 	flag.Duration("duration", 0, "Duration of audio to process")
 	flag.Uint("threads", 0, "Number of threads to use")
+	flag.Bool("speedup", false, "Enable speedup")
 	flag.Uint("max-len", 0, "Maximum segment length in characters")
 	flag.Uint("max-tokens", 0, "Maximum tokens per segment")
 	flag.Float64("word-thold", 0, "Maximum segment score")
--- a/bindings/go/params.go
+++ b/bindings/go/params.go
@ -47,6 +47,10 @@ func (p *Params) SetPrintTimestamps(v bool) {
 	p.print_timestamps = toBool(v)
 }

+func (p *Params) SetSpeedup(v bool) {
+	p.speed_up = toBool(v)
+}
+
 // Set language id
 func (p *Params) SetLanguage(lang int) error {
 	if lang == -1 {
@ -173,6 +177,9 @@ func (p *Params) String() string {
 	if p.token_timestamps {
 		str += " token_timestamps"
 	}
+	if p.speed_up {
+		str += " speed_up"
+	}

 	return str + ">"
 }
--- a/bindings/go/pkg/whisper/context.go
+++ b/bindings/go/pkg/whisper/context.go
@ -76,6 +76,11 @@ func (context *context) SetTranslate(v bool) {
 	context.params.SetTranslate(v)
 }

+// Set speedup flag
+func (context *context) SetSpeedup(v bool) {
+	context.params.SetSpeedup(v)
+}
+
 func (context *context) SetSplitOnWord(v bool) {
 	context.params.SetSplitOnWord(v)
 }
--- a/bindings/go/pkg/whisper/interface.go
+++ b/bindings/go/pkg/whisper/interface.go
@ -41,6 +41,7 @@ type Context interface {
 	SetOffset(time.Duration)        // Set offset
 	SetDuration(time.Duration)      // Set duration
 	SetThreads(uint)                // Set number of threads to use
+	SetSpeedup(bool)                // Set speedup flag
 	SetSplitOnWord(bool)            // Set split on word flag
 	SetTokenThreshold(float32)      // Set timestamp token probability threshold
 	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
--- a/bindings/ios
+++ b/bindings/ios
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
@ -20,7 +20,7 @@ public interface WhisperCppJnaLibrary extends Library {
     * @return Whisper context on success, null on failure
     */
    Pointer whisper_init_from_file(String path_model);
-
+    
    /**
     * Provides default params which can be used with `whisper_init_from_file_with_params()` etc.
     * Because this function allocates memory for the params, the caller must call either:
@ -304,6 +304,14 @@ public interface WhisperCppJnaLibrary extends Library {
    /** Language id associated with the provided state */
    int whisper_full_lang_id_from_state(Pointer state);

+    /**
+     * Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
+     * The resulting spectrogram is stored inside the default state of the provided whisper context.
+     * @return 0 on success
+     */
+    int whisper_pcm_to_mel_phase_vocoder(Pointer ctx, final float[] samples, int n_samples, int n_threads);
+
+    int whisper_pcm_to_mel_phase_vocoder_with_state(Pointer ctx, Pointer state, final float[] samples, int n_samples, int n_threads);

    /** Get the start time of the specified segment. */
    long whisper_full_get_segment_t0(Pointer ctx, int i_segment);
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperFullParams.java
@ -129,6 +129,14 @@ public class WhisperFullParams extends Structure {
    /** Maximum tokens per segment (0, default = no limit) */
    public int max_tokens;

+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public CBool speed_up;
+
+    /** Flag to speed up the audio by 2x using Phase Vocoder. (default = false) */
+    public void speedUp(boolean enable) {
+        speed_up = enable ? CBool.TRUE : CBool.FALSE;
+    }
+
    /** Overwrite the audio context size (0 = use default). */
    public int audio_ctx;

@ -140,9 +148,6 @@ public class WhisperFullParams extends Structure {
        tdrz_enable = enable ? CBool.TRUE : CBool.FALSE;
    }

-    /** Regular expression matching tokens to suppress. */
-    public String suppress_regex;
-
    /** Tokens to provide to the whisper decoder as an initial prompt.
     * These are prepended to any existing text context from a previous call. */
    public String initial_prompt;
@ -313,8 +318,8 @@ public class WhisperFullParams extends Structure {
        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
                "no_context", "single_segment", "no_timestamps",
                "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",
+                "thold_pt", "thold_ptsum", "max_len", "split_on_word", "max_tokens", "speed_up", "audio_ctx",
+                "tdrz_enable", "initial_prompt", "prompt_tokens", "prompt_n_tokens", "language", "detect_language",
                "suppress_blank", "suppress_non_speech_tokens", "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",
--- a/bindings/javascript/package.json
+++ b/bindings/javascript/package.json
@ -1,6 +1,6 @@
 {
  "name": "whisper.cpp",
-  "version": "1.6.2",
+  "version": "1.5.4",
  "description": "Whisper speech recognition",
  "main": "whisper.js",
  "scripts": {
--- a/bindings/ruby/Rakefile
+++ b/bindings/ruby/Rakefile
@ -1,12 +0,0 @@
-require 'rake/clean'
-  require 'rubygems/package'
-
-desc 'Build gem'
-task :package do
-  spec_source = File.read File.join(File.dirname(__FILE__),'whispercpp.gemspec')
-  spec = nil
-  # see: http://gist.github.com/16215
-  Thread.new { spec = eval("#{spec_source}") }.join
-  spec.validate
-  Gem::Package.build(spec)
-end
--- a/bindings/ruby/ext/extconf.rb
+++ b/bindings/ruby/ext/extconf.rb
@ -1,7 +1,6 @@
 require 'mkmf'
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.cpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','whisper-mel.hpp')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
@ -10,7 +9,6 @@ system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.c')} ."
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend-impl.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-backend.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-common.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-quants.c')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','examples','dr_wav.h')} .")
--- a/bindings/ruby/ext/ggml-backend-impl.h
+++ b/bindings/ruby/ext/ggml-backend-impl.h
@ -12,63 +12,31 @@ extern "C" {
    // Backend buffer
    //

-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        const char *          (*GGML_CALL get_name)        (ggml_backend_buffer_type_t buft);
-        ggml_backend_buffer_t (*GGML_CALL alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*GGML_CALL get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*GGML_CALL get_max_size)    (ggml_backend_buffer_type_t buft); // allocation max size
-        size_t                (*GGML_CALL get_alloc_size)  (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*GGML_CALL supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*GGML_CALL is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
    typedef void * ggml_backend_buffer_context_t;

    struct ggml_backend_buffer_i {
-        const char * (*GGML_CALL get_name)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL free_buffer)(ggml_backend_buffer_t buffer);
-        void *       (*GGML_CALL get_base)   (ggml_backend_buffer_t buffer);
-        void         (*GGML_CALL init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void         (*GGML_CALL set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void         (*GGML_CALL get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool         (*GGML_CALL cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst); // dst is in the buffer, src may be in any buffer
-        void         (*GGML_CALL clear)      (ggml_backend_buffer_t buffer, uint8_t value);
-        void         (*GGML_CALL reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
+        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
+        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
+        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
+        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
+        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
    };

    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
+        struct ggml_backend_buffer_i iface;
+
+        ggml_backend_t                backend;
        ggml_backend_buffer_context_t context;
+
        size_t size;
-        enum ggml_backend_buffer_usage usage;
    };

-    GGML_CALL ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+            struct ggml_backend                  * backend,
            struct ggml_backend_buffer_i           iface,
                   ggml_backend_buffer_context_t   context,
                   size_t                          size);

-    // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // buffer that contains a collection of buffers
-    GGML_CALL ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    GGML_CALL bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    GGML_CALL void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-
    //
    // Backend
    //
@ -76,66 +44,44 @@ extern "C" {
    typedef void * ggml_backend_context_t;

    struct ggml_backend_i {
-        const char * (*GGML_CALL get_name)(ggml_backend_t backend);
+        const char * (*get_name)(ggml_backend_t backend);

-        void (*GGML_CALL free)(ggml_backend_t backend);
+        void (*free)(ggml_backend_t backend);

        // buffer allocation
-        ggml_backend_buffer_type_t (*GGML_CALL get_default_buffer_type)(ggml_backend_t backend);
+        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);

-        // (optional) asynchronous tensor data access
-        void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        // get buffer alignment
+        size_t (*get_alignment)(ggml_backend_t backend);

-        // (optional) complete all pending operations
-        void (*GGML_CALL synchronize)(ggml_backend_t backend);
+        // tensor data access
+        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void (*synchronize)     (ggml_backend_t backend);

-        // compute graph with a plan (not used currently)
-        ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
-        void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // (optional) copy tensor between different backends, allow for single-copy tranfers
+        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);

        // compute graph with a plan
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph without a plan
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);

        // check if the backend supports an operation
-        bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
-        // these should be expensive operations with large batch sizes that may benefit from running on this backend
-        // even if the weight has to be copied from the CPU temporarily
-        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-
-        // (optional) event synchronization
-        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
-        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
-        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
-        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
+        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
    };

    struct ggml_backend {
-        ggml_guid_t guid;
-
        struct ggml_backend_i iface;
+
        ggml_backend_context_t context;
    };

-    struct ggml_backend_event {
-        ggml_backend_t backend;
-        void * context;
-    };
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*GGML_CALL ggml_backend_init_fn)(const char * params, void * user_data);
-
-    GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
 #ifdef  __cplusplus
 }
 #endif
--- a/bindings/ruby/ext/ggml-backend.c
+++ b/bindings/ruby/ext/ggml-backend.c
--- a/bindings/ruby/ext/ggml-backend.h
+++ b/bindings/ruby/ext/ggml-backend.h
@ -7,123 +7,69 @@
 extern "C" {
 #endif

-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend_event * ggml_backend_event_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
    //
    // Backend buffer
    //

-    // buffer type
-    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
-    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
-    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
+    struct ggml_backend_buffer;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;

-    // buffer
-    enum ggml_backend_buffer_usage {
-        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
-        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
-    };
-
-    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
-    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
-    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
-    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+    // backend buffer functions
+    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_free_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);

    //
    // Backend
    //

-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+    struct ggml_backend;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+
+    GGML_API ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor);
+
    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
    GGML_API void         ggml_backend_free(ggml_backend_t backend);

-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);

-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API size_t ggml_backend_get_alignment(ggml_backend_t backend);

-    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_set_async(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);

-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);

-    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
-    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);

-    // asynchronous copy
-    // the copy is performed after all the currently queued operations in backend_src
-    // backend_dst will wait for the copy to complete before performing other operations
-    // automatic fallback to sync copy if async is not supported
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-    // events
-    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
-    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
-    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event
-
    //
    // CPU backend
    //

    GGML_API ggml_backend_t ggml_backend_cpu_init(void);

-    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);

    // Create a backend buffer from an existing pointer
-    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);

-    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);

    //
    // Backend scheduler
@ -137,96 +83,53 @@ extern "C" {
    /*
      Example usage:

-        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
-        // preferrably to run on the same backend as the buffer
-        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
+        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph

-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+        // initialize buffers from a measure graph
+        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed

-        // initialize buffers from a max size graph (optional)
-        reserve_graph = build_graph(sched, max_batch_size);
+        // in build_graph:
+        build_graph(...) {
+            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
+            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
+            ggml_allocr_alloc(alloc_cpu, tensor);

-        // manually assign nodes to a backend (optional, should not be needed in most cases)
-        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
+        }

-        ggml_backend_sched_reserve(sched, reserve_graph);
+        // allocate backend buffers from measure graph
+        ggml_backend_sched_init_measure(sched, measure_graph);
+
+        // the scheduler is now ready to compute graphs

        // compute
        graph = build_graph(sched);
        ggml_backend_sched_graph_compute(sched, graph);
-
-        // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
-    }
    */

    struct ggml_backend_sched;
    typedef struct ggml_backend_sched * ggml_backend_sched_t;

-    // when ask == true, the scheduler wants to know if the user wants to observe this node
-    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
-    //
-    // when ask == false, the scheduler is passing the node tensor to the user for observation
-    // if the user returns false, the scheduler will cancel the graph compute
-    //
-    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
-
    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
+
+    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);

    // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);

-    // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
-    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);

-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
-
-    // Allocate and compute graph on the backend scheduler
-    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
-    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
-
-    // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
-
-    // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);

+    // Allocate a graph on the backend scheduler
+    GGML_API void ggml_backend_sched_graph_compute(
+            ggml_backend_sched_t sched,
+            struct ggml_cgraph * graph);

 #ifdef  __cplusplus
 }
--- a/bindings/ruby/ext/ggml-cuda.h
+++ b/bindings/ruby/ext/ggml-cuda.h
@ -1,43 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API GGML_CALL bool ggml_backend_is_cuda(ggml_backend_t backend);
-
-// device buffer
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-GGML_API GGML_CALL int  ggml_backend_cuda_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-impl.h
+++ b/bindings/ruby/ext/ggml-impl.h
@ -5,7 +5,6 @@
 // GGML internal header

 #include <assert.h>
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stddef.h>
 #include <stdbool.h>
 #include <string.h> // memcpy
@ -19,7 +18,6 @@ extern "C" {
 // fall back to the _Static_assert C11 keyword.
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef __cplusplus
 #ifndef static_assert
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
 #define static_assert(cond, msg) _Static_assert(cond, msg)
@ -27,7 +25,6 @@ extern "C" {
 #define static_assert(cond, msg) struct global_scope_noop_trick
 #endif
 #endif
-#endif

 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
@ -37,18 +34,17 @@ extern "C" {
 #ifndef __F16C__
 #define __F16C__
 #endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __SSE3__
 #define __SSE3__
 #endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
 #endif

+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
@ -60,30 +56,14 @@ extern "C" {
 //
 #include <arm_neon.h>

-typedef __fp16 ggml_fp16_internal_t;
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)

-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)

 #else

-typedef uint16_t ggml_fp16_internal_t;
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@ -237,7 +217,8 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
    uint16_t s;
    memcpy(&s, &f, sizeof(uint16_t));
@ -245,23 +226,19 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }

 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #endif

 #define GGML_HASHTABLE_FULL ((size_t)-1)
 #define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)

-struct ggml_hash_set ggml_hash_set_new(size_t size);
-
 bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);

 // returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
 size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);

-// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+// returns GGML_HAHSHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);

 // return index, asserts if table is full
--- a/bindings/ruby/ext/ggml-quants.c
+++ b/bindings/ruby/ext/ggml-quants.c
--- a/bindings/ruby/ext/ggml-quants.h
+++ b/bindings/ruby/ext/ggml-quants.h
@ -1,133 +1,224 @@
 #pragma once

-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
+#include "ggml-impl.h"

 // GGML internal header

-#ifdef __cplusplus
-extern "C" {
+#include <stdint.h>
+#include <stddef.h>
+
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    ggml_fp16_t m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    int8_t  qs[QK8_0];     // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    float d;               // delta
+    float s;               // d * sum(qs[i])
+    int8_t  qs[QK8_1];     // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// Super-block size
+#ifdef GGML_QKK_64
+#define QK_K 64
+#define K_SCALE_SIZE 4
+#else
+#define QK_K 256
+#define K_SCALE_SIZE 12
 #endif

+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.5625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    ggml_fp16_t d;           // super-block scale for quantized scales
+    ggml_fp16_t dmin;        // super-block scale for quantized mins
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[2];
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
+#else
+typedef struct {
+    uint8_t hmask[QK_K/8];     // quants - high bit
+    uint8_t qs[QK_K/4];        // quants - low 2 bits
+    uint8_t scales[12];        // scales, quantized with 6 bits
+    ggml_fp16_t d;             // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+#endif
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d[2];          // super-block scales/mins
+    uint8_t scales[2];         // 4-bit block scales/mins
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;             // super-block scale for quantized scales
+    ggml_fp16_t dmin;          // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];        // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+#endif
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+#ifdef GGML_QKK_64
+typedef struct {
+    ggml_fp16_t d;               // super-block scale
+    int8_t  scales[QK_K/16];     // 8-bit block scales
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
+#else
+typedef struct {
+    ggml_fp16_t d;               // super-block scale for quantized scales
+    ggml_fp16_t dmin;            // super-block scale for quantized mins
+    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];          // quants, high bit
+    uint8_t qs[QK_K/2];          // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+#endif
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_fp16_t d;           // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+
 // Quantization
-void quantize_row_q4_0_reference(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_reference(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_reference(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_reference(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_reference(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_reference(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
+void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
+void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
+void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
+void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
+void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);

-void quantize_row_q2_K_reference(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_reference(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
+void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
+void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
+void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
+void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);

-void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);

-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);

 // Dequantization
-void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
+//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);

-void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);

 // Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
-
-#ifdef __cplusplus
-}
-#endif
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);

+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
--- a/bindings/ruby/ext/ggml-sycl.h
+++ b/bindings/ruby/ext/ggml-sycl.h
@ -1,49 +0,0 @@
-//
-//  MIT license
-//  Copyright (C) 2024 Intel Corporation
-//  SPDX-License-Identifier: MIT
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_SYCL_MAX_DEVICES       48
-#define GGML_SYCL_NAME "SYCL"
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
-
-// devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
-
-// split tensor buffer that splits matrices by rows across multiple devices
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
-
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
-
-GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
-GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
-// TODO: these are temporary
-//       ref: https://github.com/ggerganov/llama.cpp/pull/6022#issuecomment-1992615670
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_id(int device_index);
-GGML_API GGML_CALL void ggml_backend_sycl_set_single_device_mode(int main_gpu_id);
-GGML_API GGML_CALL void ggml_backend_sycl_set_mul_device_mode();
-
-// SYCL doesn't support registering host memory, keep here for reference
-// GGML_API GGML_CALL bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API GGML_CALL void ggml_backend_sycl_unregister_host_buffer(void * buffer);
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-vulkan.h
+++ b/bindings/ruby/ext/ggml-vulkan.h
@ -1,29 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_VK_NAME "Vulkan"
-#define GGML_VK_MAX_DEVICES 16
-
-GGML_API void ggml_vk_instance_init(void);
-
-// backend API
-GGML_API GGML_CALL ggml_backend_t ggml_backend_vk_init(size_t dev_num);
-
-GGML_API GGML_CALL bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API GGML_CALL int  ggml_backend_vk_get_device_count(void);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API GGML_CALL void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
-
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
-// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ruby_whisper.cpp
+++ b/bindings/ruby/ext/ruby_whisper.cpp
@ -311,6 +311,12 @@ static VALUE ruby_whisper_params_get_split_on_word(VALUE self) {
 static VALUE ruby_whisper_params_set_split_on_word(VALUE self, VALUE value) {
  BOOL_PARAMS_SETTER(self, split_on_word, value)
 }
+static VALUE ruby_whisper_params_get_speed_up(VALUE self) {
+  BOOL_PARAMS_GETTER(self, speed_up)
+}
+static VALUE ruby_whisper_params_set_speed_up(VALUE self, VALUE value) {
+  BOOL_PARAMS_SETTER(self, speed_up, value)
+}
 static VALUE ruby_whisper_params_get_diarize(VALUE self) {
  ruby_whisper_params *rwp;
  Data_Get_Struct(self, ruby_whisper_params, rwp);
@ -402,6 +408,8 @@ void Init_whisper() {
  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);
  rb_define_method(cParams, "split_on_word=", ruby_whisper_params_set_split_on_word, 1);
+  rb_define_method(cParams, "speed_up", ruby_whisper_params_get_speed_up, 0);
+  rb_define_method(cParams, "speed_up=", ruby_whisper_params_set_speed_up, 1);
  rb_define_method(cParams, "diarize", ruby_whisper_params_get_diarize, 0);
  rb_define_method(cParams, "diarize=", ruby_whisper_params_set_diarize, 1);

--- a/bindings/ruby/tests/test_whisper.rb
+++ b/bindings/ruby/tests/test_whisper.rb
@ -117,6 +117,13 @@ class TestWhisper < Test::Unit::TestCase
    assert !@params.split_on_word
  end

+  def test_speed_up
+    @params.speed_up = true
+    assert @params.speed_up
+    @params.speed_up = false
+    assert !@params.speed_up
+  end
+
  def test_whisper
    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
    params  = Whisper::Params.new
--- a/bindings/ruby/whispercpp.gemspec
+++ b/bindings/ruby/whispercpp.gemspec
@ -1,28 +0,0 @@
-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.description = %q{High-performance inference of OpenAI's Whisper automatic speech recognition (ASR) model via Ruby}
-  s.email   = 'todd.fisher@gmail.com'
-  s.extra_rdoc_files = ['LICENSE', 'README.md']
-  
-  s.files = ["LICENSE", "README.md", "Rakefile", "ext/extconf.rb", "ext/ggml.c", "ext/ruby_whisper.cpp", "ext/whisper.cpp", "ext/dr_wav.h", "ext/ggml.h", "ext/ruby_whisper.h", "ext/whisper.h"]
-
-  #### Load-time details
-  s.require_paths = ['lib','ext']
-  s.summary = %q{Ruby whisper.cpp bindings}
-  s.test_files = ["tests/test_whisper.rb"]
-  
-  s.extensions << 'ext/extconf.rb'
-  
-
-  #### Documentation and testing.
-  s.homepage = 'https://github.com/ggerganov/whisper.cpp'
-  s.rdoc_options = ['--main', '../../README.md']
-
-  
-    s.platform = Gem::Platform::RUBY
-  
-  s.licenses = ['MIT']
-end
--- a/cmake/BuildTypes.cmake
+++ b/cmake/BuildTypes.cmake
@ -0,0 +1,54 @@
+# Add new build types
+
+# ReleaseGG - Release with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the c++ compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELEASEGG
+    CMAKE_C_FLAGS_RELEASEGG
+    CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    CMAKE_SHARED_LINKER_FLAGS_RELEASEGG )
+
+# RelWithDebInfoGG - RelWithDebInfo with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the c++ compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with debug symbols and enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG )
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo" "ReleaseGG" "RelWithDebInfoGG")
+endif()
--- a/cmake/FindFFmpeg.cmake
+++ b/cmake/FindFFmpeg.cmake
@ -1,163 +0,0 @@
-# From
-# https://github.com/snikulov/cmake-modules/blob/master/FindFFmpeg.cmake
-#
-# vim: ts=2 sw=2
-# - Try to find the required ffmpeg components(default: AVFORMAT, AVUTIL, AVCODEC)
-#
-# Once done this will define
-#  FFMPEG_FOUND         - System has the all required components.
-#  FFMPEG_INCLUDE_DIRS  - Include directory necessary for using the required components headers.
-#  FFMPEG_LIBRARIES     - Link these to use the required ffmpeg components.
-#  FFMPEG_DEFINITIONS   - Compiler switches required for using the required ffmpeg components.
-#
-# For each of the components it will additionally set.
-#   - AVCODEC
-#   - AVDEVICE
-#   - AVFORMAT
-#   - AVFILTER
-#   - AVUTIL
-#   - POSTPROC
-#   - SWSCALE
-# the following variables will be defined
-#  <component>_FOUND        - System has <component>
-#  <component>_INCLUDE_DIRS - Include directory necessary for using the <component> headers
-#  <component>_LIBRARIES    - Link these to use <component>
-#  <component>_DEFINITIONS  - Compiler switches required for using <component>
-#  <component>_VERSION      - The components version
-#
-# Copyright (c) 2006, Matthias Kretz, <kretz@kde.org>
-# Copyright (c) 2008, Alexander Neundorf, <neundorf@kde.org>
-# Copyright (c) 2011, Michael Jansen, <kde@michael-jansen.biz>
-#
-# Redistribution and use is allowed according to the terms of the BSD license.
-# For details see the accompanying COPYING-CMAKE-SCRIPTS file.
-
-include(FindPackageHandleStandardArgs)
-
-# The default components were taken from a survey over other FindFFMPEG.cmake files
-if (NOT FFmpeg_FIND_COMPONENTS)
-  set(FFmpeg_FIND_COMPONENTS AVFORMAT AVCODEC AVUTIL SWRESAMPLE) 
-endif()
-
-#
-### Macro: set_component_found
-#
-# Marks the given component as found if both *_LIBRARIES AND *_INCLUDE_DIRS is present.
-#
-macro(set_component_found _component )
-  if (${_component}_LIBRARIES AND ${_component}_INCLUDE_DIRS)
-    message(DEBUG "  - ${_component} found.")
-    set(${_component}_FOUND TRUE)
-  else ()
-  message(DEBUG "  - ${_component} not found.")
-  endif ()
-endmacro()
-
-#
-### Macro: find_component
-#
-# Checks for the given component by invoking pkgconfig and then looking up the libraries and
-# include directories.
-#
-macro(find_component _component _pkgconfig _library _header)
-
-  if (NOT WIN32)
-     # use pkg-config to get the directories and then use these values
-     # in the FIND_PATH() and FIND_LIBRARY() calls
-     find_package(PkgConfig)
-     if (PKG_CONFIG_FOUND)
-       pkg_check_modules(PC_${_component} ${_pkgconfig})
-       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDEDIR}")
-       message(STATUS "Pkgconfig found: ${PC_${_component}_INCLUDE_DIRS}")
-       message(STATUS "${PC_${_component}_CFLAGS}")
-     endif ()
-  endif (NOT WIN32)
-
-
-  find_path(${_component}_INCLUDE_DIRS ${_header}
-    HINTS
-      ${PC_${_component}_INCLUDEDIR}
-      ${PC_${_component}_INCLUDE_DIRS}
-    PATH_SUFFIXES
-      ffmpeg
-  )
-
-  # CMake's default is to search first for shared libraries and then for static libraries.
-  # Todo later: add option to prefer static libs over dynamic:
-  find_library(${_component}_LIBRARIES NAMES ${_library} lib${_library}.a  
-      HINTS
-      ${PC_${_component}_LIBDIR}
-      ${PC_${_component}_LIBRARY_DIRS}
-  )
-
-  set(${_component}_DEFINITIONS  ${PC_${_component}_CFLAGS_OTHER} CACHE STRING "The ${_component} CFLAGS.")
-  set(${_component}_VERSION      ${PC_${_component}_VERSION}      CACHE STRING "The ${_component} version number.")
-
-  set_component_found(${_component})
-
-  mark_as_advanced(
-    ${_component}_INCLUDE_DIRS
-    ${_component}_LIBRARIES
-    ${_component}_DEFINITIONS
-    ${_component}_VERSION)
-
-endmacro()
-
-
-# Check for cached results. If there are skip the costly part.
-if (NOT FFMPEG_LIBRARIES)
-
-  # Check for all possible component.
-  find_component(AVCODEC    libavcodec    avcodec  libavcodec/avcodec.h)
-  find_component(AVFORMAT   libavformat   avformat libavformat/avformat.h)
-  find_component(AVDEVICE   libavdevice   avdevice libavdevice/avdevice.h)
-  #find_component(AVRESAMPLE libavresample avresample libavresample/avresample.h) # old name for swresample
-  find_component(AVUTIL     libavutil     avutil   libavutil/avutil.h)
-  find_component(AVFILTER   libavfilter   avfilter libavfilter/avfilter.h)
-  find_component(SWSCALE    libswscale    swscale  libswscale/swscale.h)
-  find_component(POSTPROC   libpostproc   postproc libpostproc/postprocess.h)
-  find_component(SWRESAMPLE libswresample swresample libswresample/swresample.h)
-
-  # Check if the required components were found and add their stuff to the FFMPEG_* vars.
-  foreach (_component ${FFmpeg_FIND_COMPONENTS})
-    if (${_component}_FOUND)
-      # message(STATUS "Required component ${_component} present.")
-      set(FFMPEG_LIBRARIES   ${FFMPEG_LIBRARIES}   ${${_component}_LIBRARIES})
-      set(FFMPEG_DEFINITIONS ${FFMPEG_DEFINITIONS} ${${_component}_DEFINITIONS})
-      list(APPEND FFMPEG_INCLUDE_DIRS ${${_component}_INCLUDE_DIRS})
-    else ()
-      # message(STATUS "Required component ${_component} missing.")
-    endif ()
-  endforeach ()
-
-  # Build the include path with duplicates removed.
-  if (FFMPEG_INCLUDE_DIRS)
-    list(REMOVE_DUPLICATES FFMPEG_INCLUDE_DIRS)
-  endif ()
-
-  # cache the vars.
-  set(FFMPEG_INCLUDE_DIRS ${FFMPEG_INCLUDE_DIRS} CACHE STRING "The FFmpeg include directories." FORCE)
-  set(FFMPEG_LIBRARIES    ${FFMPEG_LIBRARIES}    CACHE STRING "The FFmpeg libraries." FORCE)
-  set(FFMPEG_DEFINITIONS  ${FFMPEG_DEFINITIONS}  CACHE STRING "The FFmpeg cflags." FORCE)
-
-  mark_as_advanced(FFMPEG_INCLUDE_DIRS
-                   FFMPEG_LIBRARIES
-                   FFMPEG_DEFINITIONS)
-
-endif ()
-
-# Now set the noncached _FOUND vars for the components.
-# whisper.cpp does not need SWSCALE
-foreach (_component AVCODEC AVDEVICE AVFORMAT AVRESAMPLE AVUTIL POSTPROCESS)
-  set_component_found(${_component})
-endforeach ()
-
-# Compile the list of required vars
-set(_FFmpeg_REQUIRED_VARS FFMPEG_LIBRARIES FFMPEG_INCLUDE_DIRS)
-foreach (_component ${FFmpeg_FIND_COMPONENTS})
-  list(APPEND _FFmpeg_REQUIRED_VARS ${_component}_LIBRARIES ${_component}_INCLUDE_DIRS)
-endforeach ()
-
-# Give a nice error message if some of the required vars are missing.
-find_package_handle_standard_args(FFmpeg DEFAULT_MSG ${_FFmpeg_REQUIRED_VARS})
-
--- a/cmake/git-vars.cmake
+++ b/cmake/git-vars.cmake
--- a/cmake/build-info.cmake
+++ b/cmake/build-info.cmake
@ -1,58 +0,0 @@
-set(BUILD_NUMBER 0)
-set(BUILD_COMMIT "unknown")
-set(BUILD_COMPILER "unknown")
-set(BUILD_TARGET "unknown")
-
-# Look for git
-find_package(Git)
-if(NOT Git_FOUND)
-    find_program(GIT_EXECUTABLE NAMES git git.exe)
-    if(GIT_EXECUTABLE)
-        set(Git_FOUND TRUE)
-        message(STATUS "Found Git: ${GIT_EXECUTABLE}")
-    else()
-        message(WARNING "Git not found. Build info will not be accurate.")
-    endif()
-endif()
-
-# Get the commit count and hash
-if(Git_FOUND)
-    execute_process(
-        COMMAND ${GIT_EXECUTABLE} rev-parse --short HEAD
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE HEAD
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE RES
-    )
-    if (RES EQUAL 0)
-        set(BUILD_COMMIT ${HEAD})
-    endif()
-    execute_process(
-        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        OUTPUT_VARIABLE COUNT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-        RESULT_VARIABLE RES
-    )
-    if (RES EQUAL 0)
-        set(BUILD_NUMBER ${COUNT})
-    endif()
-endif()
-
-if(MSVC)
-    set(BUILD_COMPILER "${CMAKE_C_COMPILER_ID} ${CMAKE_C_COMPILER_VERSION}")
-    set(BUILD_TARGET ${CMAKE_VS_PLATFORM_NAME})
-else()
-    execute_process(
-        COMMAND sh -c "$@ --version | head -1" _ ${CMAKE_C_COMPILER}
-        OUTPUT_VARIABLE OUT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    set(BUILD_COMPILER ${OUT})
-    execute_process(
-        COMMAND ${CMAKE_C_COMPILER} -dumpmachine
-        OUTPUT_VARIABLE OUT
-        OUTPUT_STRIP_TRAILING_WHITESPACE
-    )
-    set(BUILD_TARGET ${OUT})
-endif()
--- a/cmake/whisper-config.cmake.in
+++ b/cmake/whisper-config.cmake.in
@ -1,65 +0,0 @@
-set(LLAMA_VERSION      @LLAMA_INSTALL_VERSION@)
-set(LLAMA_BUILD_COMMIT @LLAMA_BUILD_COMMIT@)
-set(LLAMA_BUILD_NUMBER @LLAMA_BUILD_NUMBER@)
-set(LLAMA_SHARED_LIB   @BUILD_SHARED_LIBS@)
-
-set(GGML_BLAS       @GGML_BLAS@)
-set(GGML_CUDA       @GGML_CUDA@)
-set(GGML_METAL      @GGML_METAL@)
-set(GGML_HIPBLAS    @GGML_HIPBLAS@)
-set(GGML_ACCELERATE @GGML_ACCELERATE@)
-
-@PACKAGE_INIT@
-
-set_and_check(LLAMA_INCLUDE_DIR "@PACKAGE_LLAMA_INCLUDE_INSTALL_DIR@")
-set_and_check(LLAMA_LIB_DIR     "@PACKAGE_LLAMA_LIB_INSTALL_DIR@")
-set_and_check(LLAMA_BIN_DIR     "@PACKAGE_LLAMA_BIN_INSTALL_DIR@")
-
-# Ensure transient dependencies satisfied
-
-find_package(Threads REQUIRED)
-
-if (APPLE AND GGML_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate REQUIRED)
-endif()
-
-if (GGML_BLAS)
-    find_package(BLAS REQUIRED)
-endif()
-
-if (GGML_CUDA)
-    find_package(CUDAToolkit REQUIRED)
-endif()
-
-if (GGML_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK Metal REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
-endif()
-
-if (GGML_HIPBLAS)
-    find_package(hip REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
-endif()
-
-find_library(llama_LIBRARY llama
-    REQUIRED
-    HINTS ${LLAMA_LIB_DIR})
-
-set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
-set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
-
-add_library(llama UNKNOWN IMPORTED)
-
-set_target_properties(llama
-    PROPERTIES
-        INTERFACE_INCLUDE_DIRECTORIES "${LLAMA_INCLUDE_DIR}"
-        INTERFACE_LINK_LIBRARIES "${_llama_link_deps}"
-        INTERFACE_COMPILE_DEFINITIONS "${_llama_transient_defines}"
-        IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
-        IMPORTED_LOCATION "${llama_LIBRARY}"
-        INTERFACE_COMPILE_FEATURES cxx_std_11
-        POSITION_INDEPENDENT_CODE ON )
-
-check_required_components(Llama)
--- a/cmake/whisper.pc.in
+++ b/cmake/whisper.pc.in
@ -1,10 +0,0 @@
-prefix=@CMAKE_INSTALL_PREFIX@
-exec_prefix=${prefix}
-libdir=${exec_prefix}/lib
-includedir=${prefix}/include
-
-Name: whisper
-Description: Port of OpenAI's Whisper model in C/C++
-Version: @PROJECT_VERSION@
-Libs: -L${libdir} -lwhisper
-Cflags: -I${includedir}
--- a/src/coreml/whisper-decoder-impl.h
+++ b/src/coreml/whisper-decoder-impl.h
--- a/src/coreml/whisper-decoder-impl.m
+++ b/src/coreml/whisper-decoder-impl.m
--- a/src/coreml/whisper-encoder-impl.h
+++ b/src/coreml/whisper-encoder-impl.h
--- a/src/coreml/whisper-encoder-impl.m
+++ b/src/coreml/whisper-encoder-impl.m
--- a/src/coreml/whisper-encoder.h
+++ b/src/coreml/whisper-encoder.h
--- a/src/coreml/whisper-encoder.mm
+++ b/src/coreml/whisper-encoder.mm
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@ -11,7 +11,7 @@ if (WHISPER_SDL2)
    string(STRIP "${SDL2_LIBRARIES}" SDL2_LIBRARIES)

    message(STATUS "SDL2_INCLUDE_DIRS = ${SDL2_INCLUDE_DIRS}")
-    message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
+    message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()

 if (WHISPER_CLBLAST)
@ -22,51 +22,19 @@ endif()

 set(TARGET common)

-unset(COMMON_EXTRA_LIBS)
-
-if (WHISPER_FFMPEG)
-    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
-    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
-    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
-    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
-    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
-    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
-    find_package(FFmpeg REQUIRED)
-
-    if (NOT ${FFMPEG_FOUND})
-        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
-    endif()
-
-    message(STATUS "Found ffmpeg libs:       ${FFMPEG_LIBRARIES}")
-    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
-    message(STATUS "ffmpeg definitions:      ${FFMPEG_DEFINITIONS}")
-    message(STATUS "Found avformat ${AVFORMAT_VERSION}")
-
-    include_directories(${FFMPEG_INCLUDE_DIRS})
-    add_compile_definitions(WHISPER_FFMPEG)
-
-    list(APPEND COMMON_EXTRA_LIBS ${FFMPEG_LIBRARIES})
-
-    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
-endif()
-
-
 add_library(${TARGET} STATIC
    common.h
    common.cpp
    common-ggml.h
    common-ggml.cpp
-    grammar-parser.h
    grammar-parser.cpp
-    ${COMMON_SOURCES_FFMPEG}
    )

 include(DefaultTargetOptions)

-target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
+target_link_libraries(${TARGET} PRIVATE whisper)

 set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-set_target_properties(${TARGET} PROPERTIES FOLDER "libs")

 if (WHISPER_SDL2)
    # common-sdl
@ -80,16 +48,14 @@ if (WHISPER_SDL2)

    include(DefaultTargetOptions)

-    target_include_directories(${TARGET} PUBLIC  ${SDL2_INCLUDE_DIRS})
-    target_link_libraries     (${TARGET} PRIVATE ${SDL2_LIBRARIES})
+    target_include_directories(${TARGET} PUBLIC ${SDL2_INCLUDE_DIRS})
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES})

    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
 endif()

 # add json lib
-add_library(json_cpp INTERFACE)
-target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
+add_library(json_cpp INTERFACE json.hpp)

 # examples

@ -97,50 +63,25 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})

 if (EMSCRIPTEN)
    add_subdirectory(whisper.wasm)
-    set_target_properties(libmain PROPERTIES FOLDER "libs")
    add_subdirectory(stream.wasm)
-    set_target_properties(libstream PROPERTIES FOLDER "libs")
    add_subdirectory(command.wasm)
-    set_target_properties(libcommand PROPERTIES FOLDER "libs")
    add_subdirectory(talk.wasm)
-    set_target_properties(libtalk PROPERTIES FOLDER "libs")
    add_subdirectory(bench.wasm)
-    set_target_properties(libbench PROPERTIES FOLDER "libs")
 elseif(CMAKE_JS_VERSION)
    add_subdirectory(addon.node)
-    set_target_properties(addon.node PROPERTIES FOLDER "examples")
 else()
    add_subdirectory(main)
-    set_target_properties(main PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    add_subdirectory(stream)
-    set_target_properties(stream PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
    add_subdirectory(server)
-    set_target_properties(server PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    add_subdirectory(command)
-    set_target_properties(command PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
    add_subdirectory(bench)
-    set_target_properties(bench PROPERTIES FOLDER "examples")
    add_subdirectory(quantize)
-    set_target_properties(quantize PROPERTIES FOLDER "examples")
-if (WHISPER_SDL2)
    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)
+    if (LLAMA_SYCL)
        add_subdirectory(sycl)
-        set_target_properties(sycl PROPERTIES FOLDER "examples")
    endif()
-endif (WHISPER_SDL2)
 endif()

-if (WHISPER_SDL2)
-    add_subdirectory(wchess)
-    set_target_properties(wchess PROPERTIES FOLDER "examples")
-endif (WHISPER_SDL2)
+add_subdirectory(wchess)
--- a/examples/addon.node/CMakeLists.txt
+++ b/examples/addon.node/CMakeLists.txt
@ -1,4 +1,4 @@
-set(TARGET addon.node)
+set(TARGET whisper-addon)

 # Base settings
 #==================================================================
--- a/examples/addon.node/README.md
+++ b/examples/addon.node/README.md
@ -14,14 +14,14 @@ npm install
 Make sure it is in the project root directory and compiled with make-js.

 ```shell
-npx cmake-js compile -T addon.node -B Release
+npx cmake-js compile -T whisper-addon -B Release
 ```

 For Electron addon and cmake-js options, you can see [cmake-js](https://github.com/cmake-js/cmake-js) and make very few configuration changes.

 > Such as appointing special cmake path:
 > ```shell
-> npx cmake-js compile -c 'xxx/cmake' -T addon.node -B Release
+> npx cmake-js compile -c 'xxx/cmake' -T whisper-addon -B Release
 > ```

 ## Run
--- a/examples/addon.node/test/whisper.spec.js
+++ b/examples/addon.node/test/whisper.spec.js
@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
  __dirname,
-  "../../../build/Release/addon.node"
+  "../../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");

@ -12,12 +12,6 @@ const whisperParamsMock = {
  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
  use_gpu: true,
-  flash_attn: false,
-  no_prints: true,
-  comma_in_time: false,
-  translate: true,
-  no_timestamps: false,
-  audio_ctx: 0,
 };

 describe("Run whisper.node", () => {
--- a/examples/addon.node/addon.cpp
+++ b/examples/addon.node/addon.cpp
@ -19,12 +19,12 @@ struct whisper_params {
    int32_t max_len      = 0;
    int32_t best_of      = 5;
    int32_t beam_size    = -1;
-    int32_t audio_ctx    = 0;

    float word_thold    = 0.01f;
    float entropy_thold = 2.4f;
    float logprob_thold = -1.0f;

+    bool speed_up       = false;
    bool translate      = false;
    bool diarize        = false;
    bool output_txt     = false;
@ -36,10 +36,7 @@ struct whisper_params {
    bool print_colors   = false;
    bool print_progress = false;
    bool no_timestamps  = false;
-    bool no_prints      = false;
    bool use_gpu        = true;
-    bool flash_attn     = false;
-    bool comma_in_time  = true;

    std::string language = "en";
    std::string prompt;
@ -47,8 +44,6 @@ struct whisper_params {

    std::vector<std::string> fname_inp = {};
    std::vector<std::string> fname_out = {};
-
-    std::vector<float> pcmf32 = {}; // mono-channel F32 PCM
 };

 struct whisper_print_user_data {
@ -125,15 +120,9 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
    }
 }

-void cb_log_disable(enum ggml_log_level, const char *, void *) {}
-
 int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
-    if (params.no_prints) {
-        whisper_log_set(cb_log_disable, NULL);
-    }
-
-    if (params.fname_inp.empty() && params.pcmf32.empty()) {
-        fprintf(stderr, "error: no input files or audio buffer specified\n");
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
        return 2;
    }

@ -146,7 +135,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {

    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 = whisper_init_from_file_with_params(params.model.c_str(), cparams);

    if (ctx == nullptr) {
@ -154,14 +142,6 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        return 3;
    }

-    // if params.pcmf32 is provided, set params.fname_inp to "buffer"
-    // this is simpler than further modifications in the code
-    if (!params.pcmf32.empty()) {
-        fprintf(stderr, "info: using audio buffer as input\n");
-        params.fname_inp.clear();
-        params.fname_inp.emplace_back("buffer");
-    }
-
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
        const auto fname_out = f < (int)params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -169,25 +149,20 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        std::vector<float> pcmf32; // mono-channel F32 PCM
        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM

-        // read the input audio file if params.pcmf32 is not provided
-        if (params.pcmf32.empty()) {
-            if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-                fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
-                continue;
-            }
-        } else {
-            pcmf32 = params.pcmf32;
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
        }

        // print system information
-        if (!params.no_prints) {
+        {
            fprintf(stderr, "\n");
            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
        }

        // print some info about the processing
-        if (!params.no_prints) {
+        {
            fprintf(stderr, "\n");
            if (!whisper_is_multilingual(ctx)) {
                if (params.language != "en" || params.translate) {
@ -196,13 +171,12 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
                }
            }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d, audio_ctx = %d ...\n",
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, timestamps = %d ...\n",
                    __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
                    params.n_threads, params.n_processors,
                    params.language.c_str(),
                    params.translate ? "translate" : "transcribe",
-                    params.no_timestamps ? 0 : 1,
-                    params.audio_ctx);
+                    params.no_timestamps ? 0 : 1);

            fprintf(stderr, "\n");
        }
@ -229,15 +203,14 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
            wparams.entropy_thold    = params.entropy_thold;
            wparams.logprob_thold    = params.logprob_thold;
            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
-            wparams.audio_ctx        = params.audio_ctx;
+
+            wparams.speed_up         = params.speed_up;

            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;

            wparams.initial_prompt   = params.prompt.c_str();

-            wparams.no_timestamps    = params.no_timestamps;
-
            whisper_print_user_data user_data = { &params, &pcmf32s };

            // this callback is called on each new segment
@ -273,8 +246,8 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {
        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

-        result[i].emplace_back(to_timestamp(t0, params.comma_in_time));
-        result[i].emplace_back(to_timestamp(t1, params.comma_in_time));
+        result[i].emplace_back(to_timestamp(t0, true));
+        result[i].emplace_back(to_timestamp(t1, true));
        result[i].emplace_back(text);
    }

@ -325,33 +298,11 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
  std::string model = whisper_params.Get("model").As<Napi::String>();
  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();
-  bool flash_attn = whisper_params.Get("flash_attn").As<Napi::Boolean>();
-  bool no_prints = whisper_params.Get("no_prints").As<Napi::Boolean>();
-  bool no_timestamps = whisper_params.Get("no_timestamps").As<Napi::Boolean>();
-  int32_t audio_ctx = whisper_params.Get("audio_ctx").As<Napi::Number>();
-  bool comma_in_time = whisper_params.Get("comma_in_time").As<Napi::Boolean>();
-
-  Napi::Value pcmf32Value = whisper_params.Get("pcmf32");
-  std::vector<float> pcmf32_vec;
-  if (pcmf32Value.IsTypedArray()) {
-    Napi::Float32Array pcmf32 = pcmf32Value.As<Napi::Float32Array>();
-    size_t length = pcmf32.ElementLength();
-    pcmf32_vec.reserve(length);
-    for (size_t i = 0; i < length; i++) {
-      pcmf32_vec.push_back(pcmf32[i]);
-    }
-  }

  params.language = language;
  params.model = model;
  params.fname_inp.emplace_back(input);
  params.use_gpu = use_gpu;
-  params.flash_attn = flash_attn;
-  params.no_prints = no_prints;
-  params.no_timestamps = no_timestamps;
-  params.audio_ctx = audio_ctx;
-  params.pcmf32 = pcmf32_vec;
-  params.comma_in_time = comma_in_time;

  Napi::Function callback = info[1].As<Napi::Function>();
  Worker* worker = new Worker(callback, params);
--- a/examples/addon.node/index.js
+++ b/examples/addon.node/index.js
@ -1,7 +1,7 @@
 const path = require("path");
 const { whisper } = require(path.join(
  __dirname,
-  "../../build/Release/addon.node"
+  "../../build/Release/whisper-addon"
 ));
 const { promisify } = require("util");

@ -10,27 +10,15 @@ const whisperAsync = promisify(whisper);
 const whisperParams = {
  language: "en",
  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
-  fname_inp: path.join(__dirname, "../../samples/jfk.wav"),
+  fname_inp: "../../samples/jfk.wav",
  use_gpu: true,
-  flash_attn: false,
-  no_prints: true,
-  comma_in_time: false,
-  translate: true,
-  no_timestamps: false,
-  audio_ctx: 0,
 };

 const arguments = process.argv.slice(2);
 const params = Object.fromEntries(
  arguments.reduce((pre, item) => {
    if (item.startsWith("--")) {
-      const [key, value] = item.slice(2).split("=");
-      if (key === "audio_ctx") {
-        whisperParams[key] = parseInt(value);
-      } else {
-        whisperParams[key] = value;
-      }
-      return pre;
+      return [...pre, item.slice(2).split("=")];
    }
    return pre;
  }, [])
@ -45,6 +33,5 @@ for (const key in params) {
 console.log("whisperParams =", whisperParams);

 whisperAsync(whisperParams).then((result) => {
-  console.log();
-  console.log(result);
+  console.log(`Result from whisper: ${result}`);
 });
--- a/examples/addon.node/package.json
+++ b/examples/addon.node/package.json
@ -1,5 +1,5 @@
 {
-  "name": "addon.node",
+  "name": "whisper-addon",
  "version": "0.0.0",
  "description": "",
  "main": "index.js",
--- a/examples/bench/bench.cpp
+++ b/examples/bench/bench.cpp
@ -12,13 +12,12 @@ struct whisper_params {

    std::string model = "models/ggml-base.en.bin";

-    bool use_gpu    = true;
-    bool flash_attn = false;
+    bool use_gpu = true;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -26,11 +25,10 @@ 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 == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
-        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(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 == "-t"  || arg == "--threads") { params.n_threads = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")   { params.model     = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")    { params.what      = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")  { params.use_gpu   = false; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -51,20 +49,17 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
    fprintf(stderr, "\n");
 }

-static int whisper_bench_full(const whisper_params & params) {
+int whisper_bench_full(const whisper_params & params) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

--- a/examples/command/README.md
+++ b/examples/command/README.md
@ -37,13 +37,9 @@ https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9
 The `command` 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:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/command/command.cpp
+++ b/examples/command/command.cpp
@ -38,12 +38,12 @@ struct whisper_params {

    grammar_parser::parse_state grammar_parsed;

+    bool speed_up      = false;
    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 language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -52,14 +52,11 @@ struct whisper_params {
    std::string prompt;
    std::string context;
    std::string grammar;
-
-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -75,11 +72,11 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
@ -88,7 +85,6 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
-        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -113,11 +109,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -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, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@ -126,11 +122,10 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
-    fprintf(stderr, "  --suppress-regex REGEX       [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
    fprintf(stderr, "\n");
 }

-static std::string transcribe(
+std::string transcribe(
                 whisper_context * ctx,
            const whisper_params & params,
        const std::vector<float> & pcmf32,
@ -162,6 +157,7 @@ static std::string transcribe(
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx = params.audio_ctx;
+    wparams.speed_up  = params.speed_up;

    wparams.temperature     = 0.4f;
    wparams.temperature_inc = 1.0f;
@ -171,8 +167,6 @@ static std::string transcribe(

    wparams.initial_prompt = params.context.data();

-    wparams.suppress_regex = params.suppress_regex.c_str();
-
    const auto & grammar_parsed = params.grammar_parsed;
    auto grammar_rules = grammar_parsed.c_rules();

@ -216,7 +210,7 @@ static std::string transcribe(
    return result;
 }

-static std::vector<std::string> read_allowed_commands(const std::string & fname) {
+std::vector<std::string> read_allowed_commands(const std::string & fname) {
    std::vector<std::string> allowed_commands;

    std::ifstream ifs(fname);
@ -238,7 +232,7 @@ static std::vector<std::string> read_allowed_commands(const std::string & fname)
    return allowed_commands;
 }

-static std::vector<std::string> get_words(const std::string &txt) {
+std::vector<std::string> get_words(const std::string &txt) {
    std::vector<std::string> words;

    std::istringstream iss(txt);
@ -252,7 +246,7 @@ static std::vector<std::string> get_words(const std::string &txt) {

 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
    fprintf(stderr, "\n");
    fprintf(stderr, "%s: guided mode\n", __func__);

@ -367,6 +361,7 @@ static int process_command_list(struct whisper_context * ctx, audio_async &audio
            wparams.n_threads        = params.n_threads;

            wparams.audio_ctx        = params.audio_ctx;
+            wparams.speed_up         = params.speed_up;

            wparams.prompt_tokens    = k_tokens.data();
            wparams.prompt_n_tokens  = k_tokens.size();
@ -463,7 +458,7 @@ static int process_command_list(struct whisper_context * ctx, audio_async &audio

 // always-prompt mode
 // transcribe the voice into text after valid prompt
-static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
    bool is_running = true;
    bool ask_prompt = true;

@ -543,7 +538,7 @@ static int always_prompt_transcription(struct whisper_context * ctx, audio_async

 // general-purpose mode
 // freely transcribe the voice into text
-static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
    bool is_running  = true;
    bool have_prompt = false;
    bool ask_prompt  = true;
@ -694,9 +689,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

--- a/examples/common-ggml.cpp
+++ b/examples/common-ggml.cpp
@ -70,8 +70,6 @@ bool ggml_common_quantize_0(
        case GGML_FTYPE_MOSTLY_IQ1_S:
        case GGML_FTYPE_MOSTLY_IQ4_NL:
        case GGML_FTYPE_MOSTLY_IQ4_XS:
-        case GGML_FTYPE_MOSTLY_IQ1_M:
-        case GGML_FTYPE_MOSTLY_BF16:
                {
                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                    return false;
@ -195,8 +193,6 @@ bool ggml_common_quantize_0(
                case GGML_TYPE_I8:
                case GGML_TYPE_I16:
                case GGML_TYPE_I32:
-                case GGML_TYPE_I64:
-                case GGML_TYPE_F64:
                case GGML_TYPE_Q8_1:
                case GGML_TYPE_Q8_K:
                case GGML_TYPE_IQ2_XXS:
@ -207,8 +203,6 @@ bool ggml_common_quantize_0(
                case GGML_TYPE_IQ1_S:
                case GGML_TYPE_IQ4_NL:
                case GGML_TYPE_IQ4_XS:
-                case GGML_TYPE_IQ1_M:
-                case GGML_TYPE_BF16:
                case GGML_TYPE_COUNT:
                    {
                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
--- a/examples/common-sdl.cpp
+++ b/examples/common-sdl.cpp
@ -219,7 +219,7 @@ bool sdl_poll_events() {
            case SDL_QUIT:
                {
                    return false;
-                }
+                } break;
            default:
                break;
        }
--- a/examples/common.cpp
+++ b/examples/common.cpp
@ -19,18 +19,8 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

-#ifdef _WIN32
-#include <fcntl.h>
-#include <io.h>
-#endif
-
-#ifdef WHISPER_FFMPEG
-// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
-extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
-#endif
-
 // Function to check if the next argument exists
-static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
    if (i + 1 < argc && argv[i + 1][0] != '-') {
        return argv[++i];
    } else {
@ -346,7 +336,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
    return tokens;
 }

-static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
    std::vector<gpt_vocab::id> output;
    std::stringstream ss(input);
    std::string token;
@ -358,7 +348,7 @@ static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& in
    return output;
 }

-static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
    if (fpath_test.empty()){
        fprintf(stderr, "%s : No test file found.\n", __func__);
        return std::map<std::string, std::vector<gpt_vocab::id>>();
@ -642,14 +632,10 @@ bool is_wav_buffer(const std::string buf) {

 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
    drwav wav;
-    std::vector<uint8_t> wav_data; // used for pipe input from stdin or ffmpeg decoding output
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin

    if (fname == "-") {
        {
-            #ifdef _WIN32
-            _setmode(_fileno(stdin), _O_BINARY);
-            #endif
-
            uint8_t buf[1024];
            while (true)
            {
@ -675,42 +661,27 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
        }
    }
    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
-#if defined(WHISPER_FFMPEG)
-        if (ffmpeg_decode_audio(fname, wav_data) != 0) {
-            fprintf(stderr, "error: failed to ffmpeg decode '%s' \n", fname.c_str());
-            return false;
-        }
-        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-            fprintf(stderr, "error: failed to read wav data as wav \n");
-            return false;
-        }
-#else
        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
        return false;
-#endif
    }

    if (wav.channels != 1 && wav.channels != 2) {
        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

    if (stereo && wav.channels != 2) {
        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
-        drwav_uninit(&wav);
        return false;
    }

    if (wav.bitsPerSample != 16) {
        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
-        drwav_uninit(&wav);
        return false;
    }

--- a/examples/common.h
+++ b/examples/common.h
@ -21,7 +21,7 @@ struct gpt_params {
    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
    int32_t n_predict    = 200;  // new tokens to predict
    int32_t n_parallel   = 1;    // number of parallel streams
-    int32_t n_batch      = 32;   // batch size for prompt processing
+    int32_t n_batch      = 8;    // batch size for prompt processing
    int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
    int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU

@ -185,7 +185,7 @@ private:
    // It is assumed that PCM data is normalized to a range from -1 to 1
    bool write_audio(const float * data, size_t length) {
        for (size_t i = 0; i < length; ++i) {
-            const int16_t intSample = int16_t(data[i] * 32767);
+            const int16_t intSample = data[i] * 32767;
            file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
            dataSize += sizeof(int16_t);
        }
--- a/examples/ffmpeg-transcode.cpp
+++ b/examples/ffmpeg-transcode.cpp
@ -1,350 +0,0 @@
-/* SPDX-License-Identifier: GPL-2.0 */
-
-/*
- * transcode.c - convert audio file to WAVE
- *
- * Copyright (C) 2019		Andrew Clayton <andrew@digital-domain.net>
- * Copyright (C) 2024       William Tambellini <william.tambellini@gmail.com>
- */
-
-// Just for conveninent C++ API
-#include <vector>
-#include <string>
-
-// C
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-#include <stdbool.h>
-#include <stdint.h>
-#include <sys/types.h>
-#include <sys/stat.h>
-#include <fcntl.h>
-#include <unistd.h>
-#include <sys/mman.h>
-
-extern "C" {
-#include <libavutil/opt.h>
-#include <libavcodec/avcodec.h>
-#include <libavformat/avformat.h>
-#include <libswresample/swresample.h>
-}
-
-typedef uint64_t u64;
-typedef int64_t  s64;
-typedef uint32_t u32;
-typedef int32_t  s32;
-typedef uint16_t u16;
-typedef int16_t  s16;
-typedef uint8_t   u8;
-typedef int8_t    s8;
-
-#define WAVE_SAMPLE_RATE	16000
-#define AVIO_CTX_BUF_SZ		 4096
-
-static const char* ffmpegLog = getenv("FFMPEG_LOG");
-// Todo: add __FILE__ __LINE__
-#define LOG(...) \
-  do { if (ffmpegLog) fprintf(stderr, __VA_ARGS__); } while(0) // C99
-
-/*
- * WAVE file header based on definition from
- * https://gist.github.com/Jon-Schneider/8b7c53d27a7a13346a643dac9c19d34f
- *
- * We must ensure this structure doesn't have any holes or
- * padding so we can just map it straight to the WAVE data.
- */
-struct wave_hdr {
-	/* RIFF Header: "RIFF" */
-	char riff_header[4];
-	/* size of audio data + sizeof(struct wave_hdr) - 8 */
-	int wav_size;
-	/* "WAVE" */
-	char wav_header[4];
-
-	/* Format Header */
-	/* "fmt " (includes trailing space) */
-	char fmt_header[4];
-	/* Should be 16 for PCM */
-	int fmt_chunk_size;
-	/* Should be 1 for PCM. 3 for IEEE Float */
-	s16 audio_format;
-	s16 num_channels;
-	int sample_rate;
-	/*
-	 * Number of bytes per second
-	 * sample_rate * num_channels * bit_depth/8
-	 */
-	int byte_rate;
-	/* num_channels * bytes per sample */
-	s16 sample_alignment;
-	/* bits per sample */
-	s16 bit_depth;
-
-	/* Data Header */
-	/* "data" */
-	char data_header[4];
-	/*
-	 * size of audio
-	 * number of samples * num_channels * bit_depth/8
-	 */
-	int data_bytes;
-} __attribute__((__packed__));
-
-struct audio_buffer {
-	u8 *ptr;
-	int size; /* size left in the buffer */
-};
-
-static void set_wave_hdr(wave_hdr& wh, size_t size) {
-    memcpy(&wh.riff_header, "RIFF", 4);
-    wh.wav_size = size + sizeof(struct wave_hdr) - 8;
-    memcpy(&wh.wav_header, "WAVE", 4);
-    memcpy(&wh.fmt_header, "fmt ", 4);
-    wh.fmt_chunk_size = 16;
-    wh.audio_format = 1;
-    wh.num_channels = 1;
-    wh.sample_rate = WAVE_SAMPLE_RATE;
-    wh.sample_alignment = 2;
-    wh.bit_depth = 16;
-    wh.byte_rate = wh.sample_rate * wh.sample_alignment;
-    memcpy(&wh.data_header, "data", 4);
-    wh.data_bytes = size;
-}
-
-static void write_wave_hdr(int fd, size_t size) {
-	struct wave_hdr wh;
-    set_wave_hdr(wh, size);
-	write(fd, &wh, sizeof(struct wave_hdr));
-}
-
-static int map_file(int fd, u8 **ptr, size_t *size)
-{
-	struct stat sb;
-
-	fstat(fd, &sb);
-	*size = sb.st_size;
-
-    *ptr = (u8*)mmap(NULL, *size, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0);
-	if (*ptr == MAP_FAILED) {
-		perror("mmap");
-		return -1;
-	}
-
-	return 0;
-}
-
-static int read_packet(void *opaque, u8 *buf, int buf_size)
-{
-    struct audio_buffer *audio_buf = (audio_buffer*)opaque;
-
-	buf_size = FFMIN(buf_size, audio_buf->size);
-
-	/* copy internal buffer data to buf */
-	memcpy(buf, audio_buf->ptr, buf_size);
-	audio_buf->ptr += buf_size;
-	audio_buf->size -= buf_size;
-
-	return buf_size;
-}
-
-static void convert_frame(struct SwrContext *swr, AVCodecContext *codec,
-			  AVFrame *frame, s16 **data, int *size, bool flush)
-{
-	int nr_samples;
-	s64 delay;
-	u8 *buffer;
-
-	delay = swr_get_delay(swr, codec->sample_rate);
-	nr_samples = av_rescale_rnd(delay + frame->nb_samples,
-				    WAVE_SAMPLE_RATE, codec->sample_rate,
-				    AV_ROUND_UP);
-	av_samples_alloc(&buffer, NULL, 1, nr_samples, AV_SAMPLE_FMT_S16, 0);
-
-	/*
-	 * !flush is used to check if we are flushing any remaining
-	 * conversion buffers...
-	 */
-	nr_samples = swr_convert(swr, &buffer, nr_samples,
-				 !flush ? (const u8 **)frame->data : NULL,
-				 !flush ? frame->nb_samples : 0);
-
-    *data = (s16*)realloc(*data, (*size + nr_samples) * sizeof(s16));
-	memcpy(*data + *size, buffer, nr_samples * sizeof(s16));
-	*size += nr_samples;
-	av_freep(&buffer);
-}
-
-static bool is_audio_stream(const AVStream *stream)
-{
-	if (stream->codecpar->codec_type == AVMEDIA_TYPE_AUDIO)
-		return true;
-
-	return false;
-}
-
-// Return non zero on error, 0 on success
-// audio_buffer: input memory
-// data: decoded output audio data (wav file)
-// size: size of output data
-static int decode_audio(struct audio_buffer *audio_buf, s16 **data, int *size)
-{
-    LOG("decode_audio: input size: %d\n", audio_buf->size);
-	AVFormatContext *fmt_ctx;
-	AVIOContext *avio_ctx;
-	AVStream *stream;
-	AVCodecContext *codec;
-	AVPacket packet;
-	AVFrame *frame;
-	struct SwrContext *swr;
-	u8 *avio_ctx_buffer;
-	unsigned int i;
-	int stream_index = -1;
-	int err;
-    const size_t errbuffsize = 1024;
-    char errbuff[errbuffsize];
-
-    av_register_all(); // from avformat. Still a must-have call for ffmpeg v3! (can be skipped for later versions)
-
-    fmt_ctx = avformat_alloc_context();
-    avio_ctx_buffer = (u8*)av_malloc(AVIO_CTX_BUF_SZ);
-    LOG("Creating an avio context: AVIO_CTX_BUF_SZ=%d\n", AVIO_CTX_BUF_SZ);
-    avio_ctx = avio_alloc_context(avio_ctx_buffer, AVIO_CTX_BUF_SZ, 0, audio_buf, &read_packet, NULL, NULL);
-	fmt_ctx->pb = avio_ctx;
-
-    // open the input stream and read header
-	err = avformat_open_input(&fmt_ctx, NULL, NULL, NULL);
-	if (err) {
-        LOG("Could not read audio buffer: %d: %s\n", err, av_make_error_string(errbuff, errbuffsize, err));
-        return err;
-	}
-
-	err = avformat_find_stream_info(fmt_ctx, NULL);
-	if (err < 0) {
-        LOG("Could not retrieve stream info from audio buffer: %d\n", err);
-        return err;
-	}
-
-	for (i = 0; i < fmt_ctx->nb_streams; i++) {
-		if (is_audio_stream(fmt_ctx->streams[i])) {
-			stream_index = i;
-			break;
-		}
-	}
-
-	if (stream_index == -1) {
-        LOG("Could not retrieve audio stream from buffer\n");
-		return -1;
-	}
-
-	stream = fmt_ctx->streams[stream_index];
-	codec = avcodec_alloc_context3(
-			avcodec_find_decoder(stream->codecpar->codec_id));
-	avcodec_parameters_to_context(codec, stream->codecpar);
-	err = avcodec_open2(codec, avcodec_find_decoder(codec->codec_id),
-							NULL);
-	if (err) {
-        LOG("Failed to open decoder for stream #%d in audio buffer\n", stream_index);
-        return err;
-	}
-
-	/* prepare resampler */
-	swr = swr_alloc();
-
-	av_opt_set_int(swr, "in_channel_count", codec->channels, 0);
-	av_opt_set_int(swr, "out_channel_count", 1, 0);
-	av_opt_set_int(swr, "in_channel_layout", codec->channel_layout, 0);
-	av_opt_set_int(swr, "out_channel_layout", AV_CH_LAYOUT_MONO, 0);
-	av_opt_set_int(swr, "in_sample_rate", codec->sample_rate, 0);
-	av_opt_set_int(swr, "out_sample_rate", WAVE_SAMPLE_RATE, 0);
-	av_opt_set_sample_fmt(swr, "in_sample_fmt", codec->sample_fmt, 0);
-	av_opt_set_sample_fmt(swr, "out_sample_fmt", AV_SAMPLE_FMT_S16, 0);
-
-	swr_init(swr);
-	if (!swr_is_initialized(swr)) {
-        LOG("Resampler has not been properly initialized\n");
-		return -1;
-	}
-
-	av_init_packet(&packet);
-	frame = av_frame_alloc();
-	if (!frame) {
-        LOG("Error allocating the frame\n");
-		return -1;
-	}
-
-	/* iterate through frames */
-	*data = NULL;
-	*size = 0;
-	while (av_read_frame(fmt_ctx, &packet) >= 0) {
-		avcodec_send_packet(codec, &packet);
-
-		err = avcodec_receive_frame(codec, frame);
-		if (err == AVERROR(EAGAIN))
-			continue;
-
-		convert_frame(swr, codec, frame, data, size, false);
-	}
-	/* Flush any remaining conversion buffers... */
-	convert_frame(swr, codec, frame, data, size, true);
-
-	av_frame_free(&frame);
-	swr_free(&swr);
-    //avio_context_free(); // todo?
-	avcodec_close(codec);
-	avformat_close_input(&fmt_ctx);
-	avformat_free_context(fmt_ctx);
-
-	if (avio_ctx) {
-		av_freep(&avio_ctx->buffer);
-		av_freep(&avio_ctx);
-	}
-
-	return 0;
-}
-
-// in mem decoding/conversion/resampling:
-// ifname: input file path
-// owav_data: in mem wav file. Can be forwarded as it to whisper/drwav
-// return 0 on success
-int ffmpeg_decode_audio(const std::string &ifname, std::vector<uint8_t>& owav_data) {
-    LOG("ffmpeg_decode_audio: %s\n", ifname.c_str());
-    int ifd = open(ifname.c_str(), O_RDONLY);
-    if (ifd == -1) {
-        fprintf(stderr, "Couldn't open input file %s\n", ifname.c_str());
-        return -1;
-    }
-    u8 *ibuf = NULL;
-    size_t ibuf_size;
-    int err = map_file(ifd, &ibuf, &ibuf_size);
-    if (err) {
-        LOG("Couldn't map input file %s\n", ifname.c_str());
-        return err;
-    }
-    LOG("Mapped input file: %x size: %d\n", ibuf, ibuf_size);
-    struct audio_buffer inaudio_buf;
-    inaudio_buf.ptr = ibuf;
-    inaudio_buf.size = ibuf_size;
-
-    s16 *odata=NULL;
-    int osize=0;
-
-    err = decode_audio(&inaudio_buf, &odata, &osize);
-    LOG("decode_audio returned %d \n", err);
-    if (err != 0) {
-        LOG("decode_audio failed\n");
-        return err;
-    }
-    LOG("decode_audio output size: %d\n", osize);
-
-    wave_hdr wh;
-    const size_t outdatasize = osize * sizeof(s16);
-    set_wave_hdr(wh, outdatasize);
-    owav_data.resize(sizeof(wave_hdr) + outdatasize);
-    // header:
-    memcpy(owav_data.data(), &wh, sizeof(wave_hdr));
-    // the data:
-    memcpy(owav_data.data() + sizeof(wave_hdr), odata, osize* sizeof(s16));
-
-    return 0;
-}
--- a/examples/grammar-parser.cpp
+++ b/examples/grammar-parser.cpp
@ -9,7 +9,7 @@
 namespace grammar_parser {
    // NOTE: assumes valid utf8 (but checks for overrun)
    // copied from whisper.cpp
-    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+    std::pair<uint32_t, const char *> decode_utf8(const char * src) {
        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
        uint8_t  first_byte = static_cast<uint8_t>(*src);
        uint8_t  highbits   = first_byte >> 4;
@ -24,19 +24,19 @@ namespace grammar_parser {
        return std::make_pair(value, pos);
    }

-    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+    uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
        return result.first->second;
    }

-    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+    uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
        return next_id;
    }

-    static void add_rule(
+    void add_rule(
            parse_state & state,
            uint32_t      rule_id,
            const std::vector<whisper_grammar_element> & rule) {
@ -46,11 +46,11 @@ namespace grammar_parser {
        state.rules[rule_id] = rule;
    }

-    static bool is_word_char(char c) {
+    bool is_word_char(char c) {
        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
    }

-    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+    std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
        const char * pos   = src;
        const char * end   = src + size;
        uint32_t     value = 0;
@ -73,7 +73,7 @@ namespace grammar_parser {
        return std::make_pair(value, pos);
    }

-    static const char * parse_space(const char * src, bool newline_ok) {
+    const char * parse_space(const char * src, bool newline_ok) {
        const char * pos = src;
        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
@ -88,7 +88,7 @@ namespace grammar_parser {
        return pos;
    }

-    static const char * parse_name(const char * src) {
+    const char * parse_name(const char * src) {
        const char * pos = src;
        while (is_word_char(*pos)) {
            pos++;
@ -99,7 +99,7 @@ namespace grammar_parser {
        return pos;
    }

-    static std::pair<uint32_t, const char *> parse_char(const char * src) {
+    std::pair<uint32_t, const char *> parse_char(const char * src) {
        if (*src == '\\') {
            switch (src[1]) {
                case 'x': return parse_hex(src + 2, 2);
@ -122,14 +122,14 @@ namespace grammar_parser {
        throw std::runtime_error("unexpected end of input");
    }

-    static const char * parse_alternates(
+    const char * parse_alternates(
            parse_state       & state,
            const char        * src,
            const std::string & rule_name,
            uint32_t            rule_id,
            bool                is_nested);

-    static const char * parse_sequence(
+    const char * parse_sequence(
            parse_state                        & state,
            const char                         * src,
            const std::string                  & rule_name,
@ -190,7 +190,7 @@ namespace grammar_parser {
                pos = parse_space(pos + 1, is_nested);
            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
                }

                // apply transformation to previous symbol (last_sym_start to end) according to
@ -229,7 +229,7 @@ namespace grammar_parser {
        return pos;
    }

-    static const char * parse_alternates(
+    const char * parse_alternates(
            parse_state       & state,
            const char        * src,
            const std::string & rule_name,
@ -247,7 +247,7 @@ namespace grammar_parser {
        return pos;
    }

-    static const char * parse_rule(parse_state & state, const char * src) {
+    const char * parse_rule(parse_state & state, const char * src) {
        const char * name_end = parse_name(src);
        const char * pos      = parse_space(name_end, false);
        size_t       name_len = name_end - src;
@ -285,7 +285,7 @@ namespace grammar_parser {
        }
    }

-    static void print_grammar_char(FILE * file, uint32_t c) {
+    void print_grammar_char(FILE * file, uint32_t c) {
        if (0x20 <= c && c <= 0x7f) {
            fprintf(file, "%c", static_cast<char>(c));
        } else {
@ -294,7 +294,7 @@ namespace grammar_parser {
        }
    }

-    static bool is_char_element(whisper_grammar_element elem) {
+    bool is_char_element(whisper_grammar_element elem) {
        switch (elem.type) {
            case WHISPER_GRETYPE_CHAR:           return true;
            case WHISPER_GRETYPE_CHAR_NOT:       return true;
@ -304,7 +304,7 @@ namespace grammar_parser {
        }
    }

-    static void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+    void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
        for (auto elem : rule) {
            switch (elem.type) {
                case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
@ -334,7 +334,7 @@ namespace grammar_parser {
        fprintf(file, "\n");
    }

-    static void print_rule(
+    void print_rule(
            FILE     * file,
            uint32_t   rule_id,
            const std::vector<whisper_grammar_element> & rule,
@ -413,7 +413,7 @@ namespace grammar_parser {
        }
    }

-    std::vector<const whisper_grammar_element *> parse_state::c_rules() const {
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const{
        std::vector<const whisper_grammar_element *> ret;
        for (const auto & rule : rules) {
            ret.push_back(rule.data());
--- a/examples/helpers.js
+++ b/examples/helpers.js
@ -34,6 +34,9 @@ async function fetchRemote(url, cbProgress, cbPrint) {
        url,
        {
            method: 'GET',
+            headers: {
+                'Content-Type': 'application/octet-stream',
+            },
        }
    );

--- a/examples/lsp/lsp.cpp
+++ b/examples/lsp/lsp.cpp
@ -26,11 +26,11 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -53,7 +53,7 @@ struct commandset {

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -69,11 +69,11 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else {
@ -100,16 +100,16 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -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, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "\n");
 }
-static uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
+uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_params & params, uint64_t maxlength_ms, std::vector<float> & pcmf32) {
    using namespace std::chrono;
    uint64_t time_now = time_point_cast<milliseconds>(system_clock::now()).time_since_epoch().count();
    uint64_t start_time = time_now;
@ -153,7 +153,7 @@ static uint64_t wait_for_vad(audio_async & audio, json jparams, const whisper_pa
    return time_now;
 }

-static json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
+json unguided_transcription(struct whisper_context * ctx, audio_async &audio, json jparams, const whisper_params &params) {
    std::vector<whisper_token> prompt_tokens;
    std::vector<float> pcmf32;
    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 10000U, pcmf32);
@ -181,6 +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.speed_up         = params.speed_up;
    wparams.suppress_non_speech_tokens = true;
    // run the transformer and a single decoding pass
    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
@ -199,7 +200,7 @@ static json unguided_transcription(struct whisper_context * ctx, audio_async &au

 // command-list mode
 // guide the transcription to match the most likely command from a provided list
-static json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
+json guided_transcription(struct whisper_context * ctx, audio_async &audio, const whisper_params &params, json jparams, std::vector<struct commandset> commandset_list) {
    struct commandset cs = commandset_list[jparams.value("commandset_index", commandset_list.size()-1)];
    std::vector<float> pcmf32;
    uint64_t unprocessed_audio_timestamp = wait_for_vad(audio, jparams, params, 2000U, pcmf32);
@ -219,6 +220,7 @@ static json guided_transcription(struct whisper_context * ctx, audio_async &audi
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    // TODO: Do some time testing. Does an overly long prompt slow down processing?
    // Set up command sets/precompute prompts
@ -285,7 +287,7 @@ static json guided_transcription(struct whisper_context * ctx, audio_async &audi
    }
 }

-static json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
+json register_commandset(struct whisper_context * ctx, json jparams, std::vector<struct commandset> &commandset_list) {
    // TODO: check for token collision
    struct commandset cs;

@ -325,8 +327,7 @@ static json register_commandset(struct whisper_context * ctx, json jparams, std:
    commandset_list.push_back(cs);
    return json{{"index",index}};
 }
-
-static json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
+json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json /*params*/) {
    // whisper_state has the pertinent offsets, but there also seem to be a large
    // number of scratch buffers that would prevent rewinding context in a manner similar to llama
    // I'll give this a another pass once everything else is implemented,
@ -336,8 +337,7 @@ static json seek(struct whisper_context * /*ctx*/, audio_async & /*audio*/, json
            {"message", "Seeking is not yet supported."}
    };
 }
-
-static json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
+json parse_job(const json &body, struct whisper_context * ctx, audio_async &audio, const whisper_params &params, std::vector<struct commandset> &commandset_list) {
    // See: https://www.jsonrpc.org/specification
    json id = body.at("id");
    try {
@ -377,7 +377,7 @@ static json parse_job(const json &body, struct whisper_context * ctx, audio_asyn
    }
 }

-static void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+void process_loop(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
    std::deque<json> jobqueue;
    std::vector<struct commandset> commandset_list;
    while (true) {
@ -436,10 +436,7 @@ int main(int argc, char ** argv) {

    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
+    cparams.use_gpu = params.use_gpu;
    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
    // init audio

--- a/examples/main/CMakeLists.txt
+++ b/examples/main/CMakeLists.txt
@ -3,4 +3,4 @@ add_executable(${TARGET} main.cpp)

 include(DefaultTargetOptions)

-target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@ -1,12 +1,10 @@
 #include "common.h"

 #include "whisper.h"
-#include "grammar-parser.h"

 #include <cmath>
 #include <fstream>
 #include <cstdio>
-#include <regex>
 #include <string>
 #include <thread>
 #include <vector>
@ -17,7 +15,7 @@
 #endif

 // helper function to replace substrings
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+void replace_all(std::string & s, const std::string & search, const std::string & replace) {
    for (size_t pos = 0; ; pos += replace.length()) {
        pos = s.find(search, pos);
        if (pos == std::string::npos) break;
@ -28,25 +26,23 @@ static void replace_all(std::string & s, const std::string & search, const std::

 // command-line parameters
 struct whisper_params {
-    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors  = 1;
-    int32_t offset_t_ms   = 0;
-    int32_t offset_n      = 0;
-    int32_t duration_ms   = 0;
-    int32_t progress_step = 5;
-    int32_t max_context   = -1;
-    int32_t max_len       = 0;
-    int32_t best_of       = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
-    int32_t beam_size     = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
-    int32_t audio_ctx     = 0;
+    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors =  1;
+    int32_t offset_t_ms  =  0;
+    int32_t offset_n     =  0;
+    int32_t duration_ms  =  0;
+    int32_t progress_step =  5;
+    int32_t max_context  = -1;
+    int32_t max_len      =  0;
+    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t audio_ctx   = 0;

-    float word_thold      =  0.01f;
-    float entropy_thold   =  2.40f;
-    float logprob_thold   = -1.00f;
-    float grammar_penalty = 100.0f;
-    float temperature     = 0.0f;
-    float temperature_inc = 0.2f;
+    float word_thold    =  0.01f;
+    float entropy_thold =  2.40f;
+    float logprob_thold = -1.00f;

+    bool speed_up        = false;
    bool debug_mode      = false;
    bool translate       = false;
    bool detect_language = false;
@ -69,42 +65,32 @@ struct whisper_params {
    bool no_timestamps   = false;
    bool log_score       = false;
    bool use_gpu         = true;
-    bool flash_attn      = false;

    std::string language  = "en";
    std::string prompt;
    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
    std::string model     = "models/ggml-base.en.bin";
-    std::string grammar;
-    std::string grammar_rule;

    // [TDRZ] speaker turn string
    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line

-    // A regular expression that matches tokens to suppress
-    std::string suppress_regex;
-
    std::string openvino_encode_device = "CPU";

-    std::string dtw = "";
-
    std::vector<std::string> fname_inp = {};
    std::vector<std::string> fname_out = {};
-
-    grammar_parser::parse_state grammar_parsed;
 };

-static void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static char * whisper_param_turn_lowercase(char * in){
+char* whisper_param_turn_lowercase(char* in){
    int string_len = strlen(in);
-    for (int i = 0; i < string_len; i++){
+    for(int i = 0; i < string_len; i++){
        *(in+i) = tolower((unsigned char)*(in+i));
    }
    return in;
 }

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -131,12 +117,11 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-ac"   || arg == "--audio-context")   { 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]); }
+        // else if (arg == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
        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; }
@ -164,14 +149,8 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-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 {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -182,7 +161,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
    return true;
 }

-static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
    fprintf(stderr, "\n");
    fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
    fprintf(stderr, "\n");
@ -202,8 +181,7 @@ 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, "  -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, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@ -226,18 +204,12 @@ static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params
    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
-    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt (max n_text_ctx/2 tokens)\n",       params.prompt.c_str());
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
    fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
-    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
    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, "  --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());
-    fprintf(stderr, "  --grammar-penalty N            [%-7.1f] scales down logits of nongrammar tokens\n",      params.grammar_penalty);
    fprintf(stderr, "\n");
 }

@ -248,7 +220,7 @@ struct whisper_print_user_data {
    int progress_prev;
 };

-static std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
    std::string speaker = "";
    const int64_t n_samples = pcmf32s[0].size();

@ -280,8 +252,7 @@ static std::string estimate_diarization_speaker(std::vector<std::vector<float>>

    return speaker;
 }
-
-static void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
    if (progress >= *progress_prev + progress_step) {
@ -290,7 +261,7 @@ static void whisper_print_progress_callback(struct whisper_context * /*ctx*/, st
    }
 }

-static void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
    const auto & params  = *((whisper_print_user_data *) user_data)->params;
    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;

@ -359,7 +330,7 @@ static void whisper_print_segment_callback(struct whisper_context * ctx, struct
    }
 }

-static bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@ -386,7 +357,7 @@ static bool output_txt(struct whisper_context * ctx, const char * fname, const w
    return true;
 }

-static bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@ -418,7 +389,7 @@ static bool output_vtt(struct whisper_context * ctx, const char * fname, const w
    return true;
 }

-static bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@ -447,7 +418,7 @@ static bool output_srt(struct whisper_context * ctx, const char * fname, const w
    return true;
 }

-static char * escape_double_quotes_and_backslashes(const char * str) {
+char *escape_double_quotes_and_backslashes(const char *str) {
    if (str == NULL) {
        return NULL;
    }
@ -460,7 +431,7 @@ static char * escape_double_quotes_and_backslashes(const char * str) {
        }
    }

-    char * escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
    if (escaped == NULL) {
        return NULL;
    }
@ -478,39 +449,7 @@ static char * escape_double_quotes_and_backslashes(const char * str) {
    return escaped;
 }

-// double quote should be escaped by another double quote. (rfc4180)
-static char * escape_double_quotes_in_csv(const char * str) {
-    if (str == NULL) {
-        return NULL;
-    }
-
-    size_t escaped_length = strlen(str) + 1;
-
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped_length++;
-        }
-    }
-
-    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
-    if (escaped == NULL) {
-        return NULL;
-    }
-
-    size_t pos = 0;
-    for (size_t i = 0; str[i] != '\0'; i++) {
-        if (str[i] == '"') {
-            escaped[pos++] = '"';
-        }
-        escaped[pos++] = str[i];
-    }
-
-    // no need to set zero due to calloc() being used prior
-
-    return escaped;
-}
-
-static bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@ -531,7 +470,7 @@ static bool output_csv(struct whisper_context * ctx, const char * fname, const w
        const char * text = whisper_full_get_segment_text(ctx, i);
        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-        char * text_escaped = escape_double_quotes_in_csv(text);
+        char * text_escaped = escape_double_quotes_and_backslashes(text);

        //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
        fout << 10 * t0 << "," << 10 * t1 << ",";
@ -545,7 +484,7 @@ static bool output_csv(struct whisper_context * ctx, const char * fname, const w
    return true;
 }

-static bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
    std::ofstream fout(fname);
    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);

@ -564,7 +503,7 @@ static bool output_score(struct whisper_context * ctx, const char * fname, const
    return true;
 }

-static bool output_json(
+bool output_json(
             struct whisper_context * ctx,
                         const char * fname,
               const whisper_params & params,
@ -710,8 +649,7 @@ static bool output_json(
                                    times_o(token.t0, token.t1, false);
                                }
                                value_i("id", token.id, false);
-                                value_f("p", token.p, false);
-                                value_f("t_dtw", token.t_dtw, true);
+                                value_f("p", token.p, true);
                            end_obj(j == (n - 1));
                        }
                        end_arr(!params.diarize && !params.tinydiarize);
@ -735,7 +673,7 @@ static bool output_json(
 // karaoke video generation
 // outputs a bash script that uses ffmpeg to generate a video with the subtitles
 // TODO: font parameter adjustments
-static bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
+bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);

    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
@ -860,7 +798,7 @@ static bool output_wts(struct whisper_context * ctx, const char * fname, const c
    return true;
 }

-static bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
@ -901,40 +839,11 @@ static bool output_lrc(struct whisper_context * ctx, const char * fname, const w
 }


-static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+void cb_log_disable(enum ggml_log_level , const char * , void * ) { }

 int main(int argc, char ** argv) {
    whisper_params params;

-    // If the only argument starts with "@", read arguments line-by-line
-    // from the given file.
-    std::vector<std::string> vec_args;
-    if (argc == 2 && argv != nullptr && argv[1] != nullptr && argv[1][0] == '@') {
-        // Save the name of the executable.
-        vec_args.push_back(argv[0]);
-
-        // Open the response file.
-        char const * rspfile = argv[1] + sizeof(char);
-        std::ifstream fin(rspfile);
-        if (fin.is_open() == false) {
-            fprintf(stderr, "error: response file '%s' not found\n", rspfile);
-            return 1;
-        }
-
-        // Read the entire response file.
-        std::string line;
-        while (std::getline(fin, line)) {
-            vec_args.push_back(line);
-        }
-
-        // Use the contents of the response file as the command-line arguments.
-        argc = static_cast<int>(vec_args.size());
-        argv = static_cast<char **>(alloca(argc * sizeof (char *)));
-        for (int i = 0; i < argc; ++i) {
-            argv[i] = const_cast<char *>(vec_args[i].c_str());
-        }
-    }
-
    if (whisper_params_parse(argc, argv, params) == false) {
        whisper_print_usage(argc, argv, params);
        return 1;
@ -978,31 +887,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
-    if (!params.dtw.empty()) {
-        cparams.dtw_token_timestamps = true;
-        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
-
-        if (params.dtw == "tiny")      cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
-        if (params.dtw == "tiny.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
-        if (params.dtw == "base")      cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
-        if (params.dtw == "base.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
-        if (params.dtw == "small")     cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
-        if (params.dtw == "small.en")  cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
-        if (params.dtw == "medium")    cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
-        if (params.dtw == "medium.en") cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
-        if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
-        if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
-        if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
-
-        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
-            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
-            return 3;
-        }
-    }
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -1014,29 +899,6 @@ int main(int argc, char ** argv) {
    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);

-    if (!params.grammar.empty()) {
-        auto & grammar = params.grammar_parsed;
-        if (is_file_exist(params.grammar.c_str())) {
-            // read grammar from file
-            std::ifstream ifs(params.grammar.c_str());
-            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
-            grammar = grammar_parser::parse(txt.c_str());
-        } else {
-            // read grammar from string
-            grammar = grammar_parser::parse(params.grammar.c_str());
-        }
-
-        // will be empty (default) if there are parse errors
-        if (grammar.rules.empty()) {
-            fprintf(stderr, "error: failed to parse grammar \"%s\"\n", params.grammar.c_str());
-            return 4;
-        } else {
-            fprintf(stderr, "%s: grammar:\n", __func__);
-            grammar_parser::print_grammar(stderr, grammar);
-            fprintf(stderr, "\n");
-        }
-    }
-
    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
        const auto fname_inp = params.fname_inp[f];
 		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
@ -1083,8 +945,7 @@ int main(int argc, char ** argv) {
        {
            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);

-            const bool use_grammar = (!params.grammar_parsed.rules.empty() && !params.grammar_rule.empty());
-            wparams.strategy = (params.beam_size > 1 || use_grammar) ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;

            wparams.print_realtime   = false;
            wparams.print_progress   = params.print_progress;
@ -1104,20 +965,17 @@ int main(int argc, char ** argv) {
            wparams.split_on_word    = params.split_on_word;
            wparams.audio_ctx        = params.audio_ctx;

+            wparams.speed_up         = params.speed_up;
            wparams.debug_mode       = params.debug_mode;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]

-            wparams.suppress_regex   = params.suppress_regex.empty() ? nullptr : params.suppress_regex.c_str();
-
            wparams.initial_prompt   = params.prompt.c_str();

            wparams.greedy.best_of        = params.best_of;
            wparams.beam_search.beam_size = params.beam_size;

-            wparams.temperature_inc  = params.no_fallback ? 0.0f : params.temperature_inc;
-            wparams.temperature      = params.temperature;
-
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
            wparams.entropy_thold    = params.entropy_thold;
            wparams.logprob_thold    = params.logprob_thold;

@ -1125,20 +983,6 @@ int main(int argc, char ** argv) {

            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };

-            const auto & grammar_parsed = params.grammar_parsed;
-            auto grammar_rules = grammar_parsed.c_rules();
-
-            if (use_grammar) {
-                if (grammar_parsed.symbol_ids.find(params.grammar_rule) == grammar_parsed.symbol_ids.end()) {
-                    fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, params.grammar_rule.c_str());
-                } else {
-                    wparams.grammar_rules = grammar_rules.data();
-                    wparams.n_grammar_rules = grammar_rules.size();
-                    wparams.i_start_rule = grammar_parsed.symbol_ids.at(params.grammar_rule);
-                    wparams.grammar_penalty = params.grammar_penalty;
-                }
-            }
-
            // this callback is called on each new segment
            if (!wparams.print_realtime) {
                wparams.new_segment_callback           = whisper_print_segment_callback;
@ -1235,9 +1079,7 @@ int main(int argc, char ** argv) {
        }
    }

-    if (!params.no_prints) {
-        whisper_print_timings(ctx);
-    }
+    whisper_print_timings(ctx);
    whisper_free(ctx);

    return 0;
--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@ -36,7 +36,7 @@ struct whisper_filters {
 };

 // quantize a model
-static bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
    gpt_vocab vocab;

    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@ -61,6 +61,7 @@ struct whisper_params {
    float temperature     =  0.00f;
    float temperature_inc =  0.20f;

+    bool speed_up        = false;
    bool debug_mode      = false;
    bool translate       = false;
    bool detect_language = false;
@ -74,7 +75,6 @@ struct whisper_params {
    bool print_progress  = false;
    bool no_timestamps   = false;
    bool use_gpu         = true;
-    bool flash_attn      = false;

    std::string language        = "en";
    std::string prompt          = "";
@ -87,8 +87,6 @@ struct whisper_params {
    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line

    std::string openvino_encode_device = "CPU";
-
-    std::string dtw = "";
 };

 void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
@ -111,6 +109,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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, "  -su,       --speed-up          [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
@ -127,7 +126,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
    // server params
-    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n", params.dtw.c_str());
    fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
    fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
    fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
@ -153,10 +151,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
        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 == "-ac"   || arg == "--audio-context")   { 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 == "-su"   || arg == "--speed-up")        { params.speed_up        = true; }
        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; }
@ -174,9 +173,7 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params, serve
        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
        else if (arg == "-m"    || arg == "--model")           { params.model           = 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 == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
        // server params
        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
@ -501,53 +498,7 @@ int main(int argc, char ** argv) {
    }
    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
-
-    if (!params.dtw.empty()) {
-        cparams.dtw_token_timestamps = true;
-        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
-
-        if (params.dtw == "tiny") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
-        }
-        if (params.dtw == "tiny.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
-        }
-        if (params.dtw == "base") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
-        }
-        if (params.dtw == "base.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
-        }
-        if (params.dtw == "small") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
-        }
-        if (params.dtw == "small.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
-        }
-        if (params.dtw == "medium") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
-        }
-        if (params.dtw == "medium.en") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
-        }
-        if (params.dtw == "large.v1") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
-        }
-        if (params.dtw == "large.v2") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
-        }
-        if (params.dtw == "large.v3") {
-            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
-        }
-
-        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
-            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
-            return 3;
-        }
-    }
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -765,6 +716,7 @@ int main(int argc, char ** argv) {
            wparams.split_on_word    = params.split_on_word;
            wparams.audio_ctx        = params.audio_ctx;

+            wparams.speed_up         = params.speed_up;
            wparams.debug_mode       = params.debug_mode;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
@ -832,7 +784,7 @@ int main(int argc, char ** argv) {
        if (params.response_format == text_format)
        {
            std::string results = output_str(ctx, params, pcmf32s);
-            res.set_content(results.c_str(), "text/html; charset=utf-8");
+            res.set_content(results.c_str(), "text/html");
        }
        else if (params.response_format == srt_format)
        {
@ -913,7 +865,6 @@ int main(int argc, char ** argv) {
                    if (!params.no_timestamps) {
                        word["start"] = token.t0 * 0.01;
                        word["end"] = token.t1 * 0.01;
-                        word["t_dtw"] = token.t_dtw;
                    }
                    word["probability"] = token.p;
                    total_logprob += token.plog;
@ -943,7 +894,7 @@ int main(int argc, char ** argv) {
                            "application/json");
        }

-        // reset params to their defaults
+        // reset params to thier defaults
        params = default_params;
    });
    svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
--- a/examples/stream/README.md
+++ b/examples/stream/README.md
@ -30,13 +30,9 @@ a transcription block that is suitable for parsing.
 The `stream` 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:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/stream/stream.cpp
+++ b/examples/stream/stream.cpp
@ -27,6 +27,7 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool no_fallback   = false;
    bool print_special = false;
@ -35,7 +36,6 @@ struct whisper_params {
    bool tinydiarize   = false;
    bool save_audio    = false; // save audio to wav file
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -44,7 +44,7 @@ struct whisper_params {

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -61,6 +61,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
        else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
@ -71,7 +72,6 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        else if (arg == "-tdrz" || arg == "--tinydiarize")   { params.tinydiarize   = true; }
        else if (arg == "-sa"   || arg == "--save-audio")    { params.save_audio    = true; }
        else if (arg == "-ng"   || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"   || arg == "--flash-attn")    { params.flash_attn    = true; }

        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -98,6 +98,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",           params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                   params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",        params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
    fprintf(stderr, "  -nf,      --no-fallback   [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
@ -108,7 +109,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -tdrz,    --tinydiarize   [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
    fprintf(stderr, "  -sa,      --save-audio    [%-7s] save the recorded audio to a file\n",              params.save_audio ? "true" : "false");
    fprintf(stderr, "  -ng,      --no-gpu        [%-7s] disable GPU inference\n",                          params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,      --flash-attn    [%-7s] flash attention during inference\n",               params.flash_attn ? "true" : "false");
    fprintf(stderr, "\n");
 }

@ -153,9 +153,7 @@ int main(int argc, char ** argv) {
    }

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);

@ -311,6 +309,7 @@ int main(int argc, char ** argv) {
            wparams.n_threads        = params.n_threads;

            wparams.audio_ctx        = params.audio_ctx;
+            wparams.speed_up         = params.speed_up;

            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]

--- a/examples/talk-llama/CMakeLists.txt
+++ b/examples/talk-llama/CMakeLists.txt
@ -1,7 +1,7 @@
 if (WHISPER_SDL2)
    # talk-llama
    set(TARGET talk-llama)
-    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp unicode-data.cpp)
+    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp)
    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})

    if (WHISPER_CLBLAST)
--- a/examples/talk-llama/README.md
+++ b/examples/talk-llama/README.md
@ -15,13 +15,9 @@ https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-
 The `talk-llama` 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:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@ -37,13 +37,9 @@

 #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
-#define LLAMA_FILE_MAGIC_GGSQ 0x67677371u // 'ggsq'

 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 6
-
-#define LLAMA_STATE_SEQ_MAGIC   LLAMA_FILE_MAGIC_GGSQ
-#define LLAMA_STATE_SEQ_VERSION 1
+#define LLAMA_SESSION_VERSION 4

 #ifdef __cplusplus
 extern "C" {
@ -64,29 +60,9 @@ extern "C" {

    enum llama_vocab_type {
        LLAMA_VOCAB_TYPE_NONE = 0, // For models without vocab
-        LLAMA_VOCAB_TYPE_SPM  = 1, // LLaMA tokenizer based on byte-level BPE with byte fallback
-        LLAMA_VOCAB_TYPE_BPE  = 2, // GPT-2 tokenizer based on byte-level BPE
-        LLAMA_VOCAB_TYPE_WPM  = 3, // BERT tokenizer based on WordPiece
-    };
-
-    // pre-tokenization types
-    enum llama_vocab_pre_type {
-        LLAMA_VOCAB_PRE_TYPE_DEFAULT        = 0,
-        LLAMA_VOCAB_PRE_TYPE_LLAMA3         = 1,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_LLM   = 2,
-        LLAMA_VOCAB_PRE_TYPE_DEEPSEEK_CODER = 3,
-        LLAMA_VOCAB_PRE_TYPE_FALCON         = 4,
-        LLAMA_VOCAB_PRE_TYPE_MPT            = 5,
-        LLAMA_VOCAB_PRE_TYPE_STARCODER      = 6,
-        LLAMA_VOCAB_PRE_TYPE_GPT2           = 7,
-        LLAMA_VOCAB_PRE_TYPE_REFACT         = 8,
-        LLAMA_VOCAB_PRE_TYPE_COMMAND_R      = 9,
-        LLAMA_VOCAB_PRE_TYPE_STABLELM2      = 10,
-        LLAMA_VOCAB_PRE_TYPE_QWEN2          = 11,
-        LLAMA_VOCAB_PRE_TYPE_OLMO           = 12,
-        LLAMA_VOCAB_PRE_TYPE_DBRX           = 13,
-        LLAMA_VOCAB_PRE_TYPE_SMAUG          = 14,
-        LLAMA_VOCAB_PRE_TYPE_PORO           = 15,
+        LLAMA_VOCAB_TYPE_SPM  = 1, // SentencePiece
+        LLAMA_VOCAB_TYPE_BPE  = 2, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM  = 3, // WordPiece
    };

    // note: these values should be synchronized with ggml_rope
@ -98,7 +74,7 @@ extern "C" {
        LLAMA_ROPE_TYPE_GLM  =  4,
    };

-    enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
+    enum llama_token_type {
        LLAMA_TOKEN_TYPE_UNDEFINED    = 0,
        LLAMA_TOKEN_TYPE_NORMAL       = 1,
        LLAMA_TOKEN_TYPE_UNKNOWN      = 2,
@ -108,20 +84,6 @@ extern "C" {
        LLAMA_TOKEN_TYPE_BYTE         = 6,
    };

-    enum llama_token_attr {
-        LLAMA_TOKEN_ATTR_UNDEFINED    = 0,
-        LLAMA_TOKEN_ATTR_UNKNOWN      = 1 << 0,
-        LLAMA_TOKEN_ATTR_UNUSED       = 1 << 1,
-        LLAMA_TOKEN_ATTR_NORMAL       = 1 << 2,
-        LLAMA_TOKEN_ATTR_CONTROL      = 1 << 3,  // SPECIAL?
-        LLAMA_TOKEN_ATTR_USER_DEFINED = 1 << 4,
-        LLAMA_TOKEN_ATTR_BYTE         = 1 << 5,
-        LLAMA_TOKEN_ATTR_NORMALIZED   = 1 << 6,
-        LLAMA_TOKEN_ATTR_LSTRIP       = 1 << 7,
-        LLAMA_TOKEN_ATTR_RSTRIP       = 1 << 8,
-        LLAMA_TOKEN_ATTR_SINGLE_WORD  = 1 << 9,
-    };
-
    // model file types
    enum llama_ftype {
        LLAMA_FTYPE_ALL_F32              = 0,
@ -155,8 +117,6 @@ extern "C" {
        LLAMA_FTYPE_MOSTLY_IQ2_S         = 28, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_IQ2_M         = 29, // except 1d tensors
        LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors

        LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
    };
@ -174,7 +134,6 @@ extern "C" {
        LLAMA_POOLING_TYPE_NONE = 0,
        LLAMA_POOLING_TYPE_MEAN = 1,
        LLAMA_POOLING_TYPE_CLS  = 2,
-        LLAMA_POOLING_TYPE_LAST = 3,
    };

    enum llama_split_mode {
@ -195,7 +154,7 @@ extern "C" {
        bool sorted;
    } llama_token_data_array;

-    typedef bool (*llama_progress_callback)(float progress, void * user_data);
+    typedef bool (*llama_progress_callback)(float progress, void *ctx);

    // Input data for llama_decode
    // A llama_batch object can contain input about one or many sequences
@ -231,19 +190,15 @@ extern "C" {
        LLAMA_KV_OVERRIDE_TYPE_INT,
        LLAMA_KV_OVERRIDE_TYPE_FLOAT,
        LLAMA_KV_OVERRIDE_TYPE_BOOL,
-        LLAMA_KV_OVERRIDE_TYPE_STR,
    };

    struct llama_model_kv_override {
-        enum llama_model_kv_override_type tag;
-
        char key[128];
-
+        enum llama_model_kv_override_type tag;
        union {
-            int64_t val_i64;
-            double  val_f64;
-            bool    val_bool;
-            char    val_str[128];
+            int64_t int_value;
+            double float_value;
+            bool bool_value;
        };
    };

@ -260,9 +215,6 @@ extern "C" {
        // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
        const float * tensor_split;

-        // comma separated list of RPC servers to use for offloading
-        const char * rpc_servers;
-
        // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
        // If the provided progress_callback returns true, model loading continues.
        // If it returns false, model loading is immediately aborted.
@ -275,14 +227,11 @@ extern "C" {
        const struct llama_model_kv_override * kv_overrides;

        // Keep the booleans together to avoid misalignment during copy-by-value.
-        bool vocab_only;    // only load the vocabulary, no weights
-        bool use_mmap;      // use mmap if possible
-        bool use_mlock;     // force system to keep model in RAM
-        bool check_tensors; // validate model tensor data
+        bool vocab_only; // only load the vocabulary, no weights
+        bool use_mmap;   // use mmap if possible
+        bool use_mlock;  // force system to keep model in RAM
    };

-    // NOTE: changing the default values of parameters marked as [EXPERIMENTAL] may cause crashes or incorrect results in certain configurations
-    //       https://github.com/ggerganov/llama.cpp/pull/7544
    struct llama_context_params {
        uint32_t seed;              // RNG seed, -1 for random
        uint32_t n_ctx;             // text context, 0 = from model
@ -294,6 +243,7 @@ extern "C" {

        enum llama_rope_scaling_type rope_scaling_type; // RoPE scaling type, from `enum llama_rope_scaling_type`
        enum llama_pooling_type      pooling_type;      // whether to pool (sum) embedding results by sequence id
+                                                        // (ignored if no pooling layer)

        // ref: https://github.com/ggerganov/llama.cpp/pull/2054
        float    rope_freq_base;   // RoPE base frequency, 0 = from model
@ -308,14 +258,13 @@ extern "C" {
        ggml_backend_sched_eval_callback cb_eval;
        void * cb_eval_user_data;

-        enum ggml_type type_k; // data type for K cache [EXPERIMENTAL]
-        enum ggml_type type_v; // data type for V cache [EXPERIMENTAL]
+        enum ggml_type type_k; // data type for K cache
+        enum ggml_type type_v; // data type for V cache

        // Keep the booleans together to avoid misalignment during copy-by-value.
        bool logits_all;  // the llama_decode() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
        bool embeddings;  // if true, extract embeddings (together with logits)
        bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
-        bool flash_attn;  // whether to use flash attention [EXPERIMENTAL]

        // Abort callback
        // if it returns true, execution of llama_decode() will be aborted
@ -326,17 +275,13 @@ extern "C" {

    // model quantization parameters
    typedef struct llama_model_quantize_params {
-        int32_t nthread;                     // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
-        enum llama_ftype ftype;              // quantize to this llama_ftype
-        enum ggml_type output_tensor_type;   // output tensor type
-        enum ggml_type token_embedding_type; // itoken embeddings tensor type
-        bool allow_requantize;               // allow quantizing non-f32/f16 tensors
-        bool quantize_output_tensor;         // quantize output.weight
-        bool only_copy;                      // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
-        bool pure;                           // quantize all tensors to the default type
-        bool keep_split;                     // quantize to the same number of shards
-        void * imatrix;                      // pointer to importance matrix data
-        void * kv_overrides;                 // pointer to vector containing overrides
+        int32_t nthread;             // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency()
+        enum llama_ftype ftype;      // quantize to this llama_ftype
+        bool allow_requantize;       // allow quantizing non-f32/f16 tensors
+        bool quantize_output_tensor; // quantize output.weight
+        bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
+        bool pure;                   // quantize all tensors to the default type
+        void * imatrix;              // pointer to importance matrix data
    } llama_model_quantize_params;

    // grammar types
@ -366,9 +311,6 @@ extern "C" {
        // modifies a preceding LLAMA_GRETYPE_CHAR or
        // LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
        LLAMA_GRETYPE_CHAR_ALT       = 6,
-
-        // any character (.)
-        LLAMA_GRETYPE_CHAR_ANY       = 7,
    };

    typedef struct llama_grammar_element {
@ -440,15 +382,12 @@ extern "C" {
    LLAMA_API uint32_t llama_n_ubatch   (const struct llama_context * ctx);
    LLAMA_API uint32_t llama_n_seq_max  (const struct llama_context * ctx);

-    LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx);
-
-    LLAMA_API enum llama_vocab_type   llama_vocab_type  (const struct llama_model * model);
-    LLAMA_API enum llama_rope_type    llama_rope_type   (const struct llama_model * model);
+    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_model * model);
+    LLAMA_API enum llama_rope_type  llama_rope_type (const struct llama_model * model);

    LLAMA_API int32_t llama_n_vocab    (const struct llama_model * model);
    LLAMA_API int32_t llama_n_ctx_train(const struct llama_model * model);
    LLAMA_API int32_t llama_n_embd     (const struct llama_model * model);
-    LLAMA_API int32_t llama_n_layer    (const struct llama_model * model);

    // Get the model's RoPE frequency scaling factor
    LLAMA_API float llama_rope_freq_scale_train(const struct llama_model * model);
@ -496,24 +435,10 @@ extern "C" {
    // Returns 0 on success
    LLAMA_API int32_t llama_model_apply_lora_from_file(
            const struct llama_model * model,
-                          const char * path_lora,
-                               float   scale,
-                          const char * path_base_model,
-                             int32_t   n_threads);
-
-    // Apply a loaded control vector to a llama_context, or if data is NULL, clear
-    // the currently loaded vector.
-    // n_embd should be the size of a single layer's control, and data should point
-    // to an n_embd x n_layers buffer starting from layer 1.
-    // il_start and il_end are the layer range the vector should apply to (both inclusive)
-    // See llama_control_vector_load in common to load a control vector.
-    LLAMA_API int32_t llama_control_vector_apply(
-            struct llama_context * lctx,
-                     const float * data,
-                          size_t   len,
-                         int32_t   n_embd,
-                         int32_t   il_start,
-                         int32_t   il_end);
+                      const char * path_lora,
+                           float   scale,
+                      const char * path_base_model,
+                         int32_t   n_threads);

    //
    // KV cache
@ -574,12 +499,11 @@ extern "C" {
    // Returns the number of used KV cells (i.e. have at least one sequence assigned to them)
    LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx);

-    // Clear the KV cache - both cell info is erased and KV data is zeroed
+    // Clear the KV cache
    LLAMA_API void llama_kv_cache_clear(
            struct llama_context * ctx);

    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
-    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
    // seq_id < 0 : match any sequence
    // p0 < 0     : [0,  p1]
    // p1 < 0     : [p0, inf)
@ -651,92 +575,34 @@ extern "C" {

    // Returns the maximum size in bytes of the state (rng, logits, embedding
    // and kv_cache) - will often be smaller after compacting tokens
-    LLAMA_API size_t llama_state_get_size(const struct llama_context * ctx);
-    LLAMA_API DEPRECATED(size_t llama_get_state_size(const struct llama_context * ctx),
-        "use llama_state_get_size instead");
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);

    // Copies the state to the specified destination address.
    // Destination needs to have allocated enough memory.
    // Returns the number of bytes copied
-    LLAMA_API size_t llama_state_get_data(
+    LLAMA_API size_t llama_copy_state_data(
            struct llama_context * ctx,
                         uint8_t * dst);
-    LLAMA_API DEPRECATED(size_t llama_copy_state_data(
-            struct llama_context * ctx,
-                         uint8_t * dst),
-        "use llama_state_get_data instead");

    // Set the state reading from the specified address
    // Returns the number of bytes read
-    LLAMA_API size_t llama_state_set_data(
+    LLAMA_API size_t llama_set_state_data(
            struct llama_context * ctx,
                   const uint8_t * src);
-    LLAMA_API DEPRECATED(size_t llama_set_state_data(
-            struct llama_context * ctx,
-                   const uint8_t * src),
-        "use llama_state_set_data instead");

    // Save/load session file
-    LLAMA_API bool llama_state_load_file(
+    LLAMA_API bool llama_load_session_file(
            struct llama_context * ctx,
                      const char * path_session,
                     llama_token * tokens_out,
                          size_t   n_token_capacity,
                          size_t * n_token_count_out);
-    LLAMA_API DEPRECATED(bool llama_load_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out),
-        "use llama_state_load_file instead");

-    LLAMA_API bool llama_state_save_file(
+    LLAMA_API bool llama_save_session_file(
            struct llama_context * ctx,
                      const char * path_session,
               const llama_token * tokens,
                          size_t   n_token_count);
-    LLAMA_API DEPRECATED(bool llama_save_session_file(
-            struct llama_context * ctx,
-                      const char * path_session,
-               const llama_token * tokens,
-                          size_t   n_token_count),
-        "use llama_state_save_file instead");
-
-    // Get the exact size needed to copy the KV cache of a single sequence
-    LLAMA_API size_t llama_state_seq_get_size(
-            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
-
-    // Copy the KV cache of a single sequence into the specified buffer
-    LLAMA_API size_t llama_state_seq_get_data(
-            struct llama_context * ctx,
-                         uint8_t * dst,
-                    llama_seq_id   seq_id);
-
-    // Copy the sequence data (originally copied with `llama_state_seq_get_data`) into the specified sequence
-    // Returns:
-    //  - Positive: Ok
-    //  - Zero: Failed to load
-    LLAMA_API size_t llama_state_seq_set_data(
-            struct llama_context * ctx,
-                   const uint8_t * src,
-                    llama_seq_id   dest_seq_id);
-
-    LLAMA_API size_t llama_state_seq_save_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   seq_id,
-               const llama_token * tokens,
-                          size_t   n_token_count);
-
-    LLAMA_API size_t llama_state_seq_load_file(
-            struct llama_context * ctx,
-                      const char * filepath,
-                    llama_seq_id   dest_seq_id,
-                     llama_token * tokens_out,
-                          size_t   n_token_capacity,
-                          size_t * n_token_count_out);

    //
    // Decoding
@ -780,16 +646,6 @@ extern "C" {
    // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
    LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);

-    // Get the number of threads used for generation of a single token.
-    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
-
-    // Get the number of threads used for prompt and batch processing (multiple token).
-    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
-
-    // Set whether the model is in embeddings mode or not
-    // If true, embeddings will be returned but logits will not
-    LLAMA_API void llama_set_embeddings(struct llama_context * ctx, bool embeddings);
-
    // Set whether to use causal attention or not
    // If set to true, the model will only attend to the past tokens
    LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);
@ -803,31 +659,23 @@ extern "C" {
    LLAMA_API void llama_synchronize(struct llama_context * ctx);

    // Token logits obtained from the last call to llama_decode()
-    // The logits for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // Rows: number of tokens for which llama_batch.logits[i] != 0
+    // The logits for the last token are stored in the last row
+    // Logits for which llama_batch.logits[i] == 0 are undefined
+    // Rows: n_tokens provided with llama_batch
    // Cols: n_vocab
    LLAMA_API float * llama_get_logits(struct llama_context * ctx);

-    // Logits for the ith token. For positive indices, Equivalent to:
-    // llama_get_logits(ctx) + ctx->output_ids[i]*n_vocab
-    // Negative indicies can be used to access logits in reverse order, -1 is the last logit.
-    // returns NULL for invalid ids.
+    // Logits for the ith token. Equivalent to:
+    // llama_get_logits(ctx) + i*n_vocab
    LLAMA_API float * llama_get_logits_ith(struct llama_context * ctx, int32_t i);

-    // Get all output token embeddings.
-    // when pooling_type == LLAMA_POOLING_TYPE_NONE or when using a generative model,
-    // the embeddings for which llama_batch.logits[i] != 0 are stored contiguously
-    // in the order they have appeared in the batch.
-    // shape: [n_outputs*n_embd]
-    // Otherwise, returns NULL.
+    // Get all output token embeddings
+    // shape: [n_tokens*n_embd] (1-dimensional)
    LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);

-    // Get the embeddings for the ith token. For positive indices, Equivalent to:
-    // llama_get_embeddings(ctx) + ctx->output_ids[i]*n_embd
-    // Negative indicies can be used to access embeddings in reverse order, -1 is the last embedding.
+    // Get the embeddings for the ith token
+    // llama_get_embeddings(ctx) + i*n_embd
    // shape: [n_embd] (1-dimensional)
-    // returns NULL for invalid ids.
    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);

    // Get the embeddings for a sequence id
@ -843,19 +691,11 @@ extern "C" {

    LLAMA_API float llama_token_get_score(const struct llama_model * model, llama_token token);

-    LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_model * model, llama_token token);
-
-    // Check if the token is supposed to end generation (end-of-generation, eg. EOS, EOT, etc.)
-    LLAMA_API bool llama_token_is_eog(const struct llama_model * model, llama_token token);
-
-    // Identify if Token Id is a control token or a render-able token
-    LLAMA_API bool llama_token_is_control(const struct llama_model * model, llama_token token);
+    LLAMA_API enum llama_token_type llama_token_get_type(const struct llama_model * model, llama_token token);

    // Special tokens
    LLAMA_API llama_token llama_token_bos(const struct llama_model * model); // beginning-of-sentence
    LLAMA_API llama_token llama_token_eos(const struct llama_model * model); // end-of-sentence
-    LLAMA_API llama_token llama_token_cls(const struct llama_model * model); // classification
-    LLAMA_API llama_token llama_token_sep(const struct llama_model * model); // sentence separator
    LLAMA_API llama_token llama_token_nl (const struct llama_model * model); // next-line

    // Returns -1 if unknown, 1 for true or 0 for false.
@ -864,7 +704,7 @@ extern "C" {
    // Returns -1 if unknown, 1 for true or 0 for false.
    LLAMA_API int32_t         llama_add_eos_token(const struct llama_model * model);

-    // Codellama infill tokens
+    // codellama infill tokens
    LLAMA_API llama_token llama_token_prefix(const struct llama_model * model); // Beginning of infill prefix
    LLAMA_API llama_token llama_token_middle(const struct llama_model * model); // Beginning of infill middle
    LLAMA_API llama_token llama_token_suffix(const struct llama_model * model); // Beginning of infill suffix
@ -878,28 +718,26 @@ extern "C" {
    /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
    /// @return Returns the number of tokens on success, no more than n_tokens_max
    /// @return Returns a negative number on failure - the number of tokens that would have been returned
-    /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
-    ///                      as plaintext. Does not insert a leading space.
+    /// @param special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated as plaintext.
+    ///                Does not insert a leading space.
    LLAMA_API int32_t llama_tokenize(
        const struct llama_model * model,
                      const char * text,
                         int32_t   text_len,
                     llama_token * tokens,
                         int32_t   n_tokens_max,
-                            bool   add_special,
-                            bool   parse_special);
+                            bool   add_bos,
+                            bool   special);

    // Token Id -> Piece.
    // Uses the vocabulary in the provided context.
    // Does not write null terminator to the buffer.
    // User code is responsible to remove the leading whitespace of the first non-BOS token when decoding multiple tokens.
-    // @param special If true, special tokens are rendered in the output.
    LLAMA_API int32_t llama_token_to_piece(
              const struct llama_model * model,
                           llama_token   token,
                                  char * buf,
-                               int32_t   length,
-                                  bool   special);
+                               int32_t   length);

    /// Apply chat template. Inspired by hf apply_chat_template() on python.
    /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model"
@ -1052,7 +890,7 @@ extern "C" {
            struct llama_context * ctx,
          llama_token_data_array * candidates);

-    /// @details Randomly selects a token from the candidates based on their probabilities using the RNG of ctx.
+    /// @details Randomly selects a token from the candidates based on their probabilities.
    LLAMA_API llama_token llama_sample_token(
            struct llama_context * ctx,
          llama_token_data_array * candidates);
@ -1064,18 +902,48 @@ extern "C" {
                     llama_token   token);

    //
-    // Model split
+    // Beam search
    //

-    /// @details Build a split GGUF final path for this chunk.
-    ///          llama_split_path(split_path, sizeof(split_path), "/models/ggml-model-q4_0", 2, 4) => split_path = "/models/ggml-model-q4_0-00002-of-00004.gguf"
-    //  Returns the split_path length.
-    LLAMA_API int llama_split_path(char * split_path, size_t maxlen, const char * path_prefix, int split_no, int split_count);
+    struct llama_beam_view {
+        const llama_token * tokens;

-    /// @details Extract the path prefix from the split_path if and only if the split_no and split_count match.
-    ///          llama_split_prefix(split_prefix, 64, "/models/ggml-model-q4_0-00002-of-00004.gguf", 2, 4) => split_prefix = "/models/ggml-model-q4_0"
-    //  Returns the split_prefix length.
-    LLAMA_API int llama_split_prefix(char * split_prefix, size_t maxlen, const char * split_path, int split_no, int split_count);
+        size_t n_tokens;
+        float  p;        // Cumulative beam probability (renormalized relative to all beams)
+        bool   eob;      // Callback should set this to true when a beam is at end-of-beam.
+    };
+
+    // Passed to beam_search_callback function.
+    // Whenever 0 < common_prefix_length, this number of tokens should be copied from any of the beams
+    // (e.g. beams[0]) as they will be removed (shifted) from all beams in all subsequent callbacks.
+    // These pointers are valid only during the synchronous callback, so should not be saved.
+    struct llama_beams_state {
+        struct llama_beam_view * beam_views;
+
+        size_t n_beams;               // Number of elements in beam_views[].
+        size_t common_prefix_length;  // Current max length of prefix tokens shared by all beams.
+        bool   last_call;             // True iff this is the last callback invocation.
+    };
+
+    // Type of pointer to the beam_search_callback function.
+    // void* callback_data is any custom data passed to llama_beam_search, that is subsequently
+    // passed back to beam_search_callback. This avoids having to use global variables in the callback.
+    typedef void (*llama_beam_search_callback_fn_t)(void * callback_data, struct llama_beams_state);
+
+    /// @details Deterministically returns entire sentence constructed by a beam search.
+    /// @param ctx Pointer to the llama_context.
+    /// @param callback Invoked for each iteration of the beam_search loop, passing in beams_state.
+    /// @param callback_data A pointer that is simply passed back to callback.
+    /// @param n_beams Number of beams to use.
+    /// @param n_past Number of tokens already evaluated.
+    /// @param n_predict Maximum number of tokens to predict. EOS may occur earlier.
+    LLAMA_API void llama_beam_search(
+                   struct llama_context * ctx,
+        llama_beam_search_callback_fn_t   callback,
+                                   void * callback_data,
+                                 size_t   n_beams,
+                                int32_t   n_past,
+                                int32_t   n_predict);

    // Performance information
    LLAMA_API struct llama_timings llama_get_timings(struct llama_context * ctx);
@ -1099,49 +967,15 @@ extern "C" {
 // Internal API to be implemented by llama.cpp and used by tests/benchmarks only
 #ifdef LLAMA_API_INTERNAL

-#include <random>
-#include <string>
 #include <vector>
+#include <string>

 struct ggml_tensor;

-struct llama_partial_utf8 {
-    uint32_t value;    // bit value so far (unshifted)
-    int      n_remain; // num bytes remaining; -1 indicates invalid sequence
-};
-
-struct llama_grammar {
-    const std::vector<std::vector<llama_grammar_element>>   rules;
-    std::vector<std::vector<const llama_grammar_element *>> stacks;
-
-    // buffer for partially generated UTF-8 sequence from accepted tokens
-    llama_partial_utf8                                      partial_utf8;
-};
-
-struct llama_grammar_candidate {
-    size_t               index;
-    const uint32_t     * code_points;
-    llama_partial_utf8   partial_utf8;
-};
-
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
    struct llama_context * ctx
 );

-void llama_grammar_accept(
-        const std::vector<std::vector<llama_grammar_element>>         & rules,
-        const std::vector<std::vector<const llama_grammar_element *>> & stacks,
-        const uint32_t                                                  chr,
-        std::vector<std::vector<const llama_grammar_element *>>       & new_stacks);
-
-std::pair<std::vector<uint32_t>, llama_partial_utf8> decode_utf8(
-        const std::string & src,
-        llama_partial_utf8   partial_start);
-
-// Randomly selects a token from the candidates based on their probabilities using given std::mt19937.
-// This is a temporary workaround in order to fix race conditions when sampling with multiple sequences.
-llama_token llama_sample_token_with_rng(struct llama_context * ctx, llama_token_data_array * candidates, std::mt19937 & rng);
-
 #endif // LLAMA_API_INTERNAL

 #endif // LLAMA_H
--- a/examples/talk-llama/talk-llama.cpp
+++ b/examples/talk-llama/talk-llama.cpp
@ -16,7 +16,7 @@
 #include <regex>
 #include <sstream>

-static std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
+std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
    auto * model = llama_get_model(ctx);

    // upper limit for the number of tokens
@ -33,12 +33,12 @@ static std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const
    return result;
 }

-static std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
    std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
    if (n_tokens < 0) {
        result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), false);
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
        GGML_ASSERT(check == -n_tokens);
    } else {
        result.resize(n_tokens);
@ -59,13 +59,13 @@ struct whisper_params {
    float vad_thold  = 0.6f;
    float freq_thold = 100.0f;

+    bool speed_up       = false;
    bool translate      = false;
    bool print_special  = false;
    bool print_energy   = false;
    bool no_timestamps  = true;
    bool verbose_prompt = false;
    bool use_gpu        = true;
-    bool flash_attn     = false;

    std::string person      = "Georgi";
    std::string bot_name    = "LLaMA";
@ -83,7 +83,7 @@ struct whisper_params {

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -99,12 +99,12 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        else if (arg == "-ngl" || arg == "--n-gpu-layers")   { params.n_gpu_layers   = 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 == "-su"  || arg == "--speed-up")       { params.speed_up       = true; }
        else if (arg == "-tr"  || arg == "--translate")      { params.translate      = true; }
        else if (arg == "-ps"  || arg == "--print-special")  { params.print_special  = true; }
        else if (arg == "-pe"  || arg == "--print-energy")   { params.print_energy   = true; }
        else if (arg == "-vp"  || arg == "--verbose-prompt") { params.verbose_prompt = 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 == "-bn"   || arg == "--bot-name")      { params.bot_name       = argv[++i]; }
        else if (arg == "--session")                         { params.path_session   = argv[++i]; }
@ -123,6 +123,7 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
            }
        }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -147,12 +148,12 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ngl N,   --n-gpu-layers N [%-7d] number of layers to store in VRAM\n",           params.n_gpu_layers);
    fprintf(stderr, "  -vth N,   --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,      --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -vp,      --verbose-prompt [%-7s] print prompt at start\n",                       params.verbose_prompt ? "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, "  -bn NAME, --bot-name NAME  [%-7s] bot name (to display)\n",                       params.bot_name.c_str());
    fprintf(stderr, "  -w TEXT,  --wake-command T [%-7s] wake-up command to listen for\n",               params.wake_cmd.c_str());
@ -168,7 +169,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "\n");
 }

-static std::string transcribe(
+std::string transcribe(
        whisper_context * ctx,
        const whisper_params & params,
        const std::vector<float> & pcmf32,
@ -202,6 +203,7 @@ static std::string transcribe(
    wparams.prompt_n_tokens  = prompt_tokens.empty() ? 0       : prompt_tokens.size();

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
        return "";
@ -235,7 +237,7 @@ static std::string transcribe(
    return result;
 }

-static std::vector<std::string> get_words(const std::string &txt) {
+std::vector<std::string> get_words(const std::string &txt) {
    std::vector<std::string> words;

    std::istringstream iss(txt);
@ -283,15 +285,9 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
-    if (!ctx_wsp) {
-        fprintf(stderr, "No whisper.cpp model specified. Please provide using -mw <modelfile>\n");
-        return 1;
-    }

    // llama init

@ -305,10 +301,6 @@ int main(int argc, char ** argv) {
    }

    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
-    if (!model_llama) {
-        fprintf(stderr, "No llama.cpp model specified. Please provide using -ml <modelfile>\n");
-        return 1;
-    }

    llama_context_params lcparams = llama_context_default_params();

@ -316,7 +308,6 @@ int main(int argc, char ** argv) {
    lcparams.n_ctx      = 2048;
    lcparams.seed       = 1;
    lcparams.n_threads  = params.n_threads;
-    lcparams.flash_attn = params.flash_attn;

    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);

--- a/examples/talk-llama/unicode-data.cpp
+++ b/examples/talk-llama/unicode-data.cpp
--- a/examples/talk-llama/unicode-data.h
+++ b/examples/talk-llama/unicode-data.h
@ -1,20 +0,0 @@
-#pragma once
-
-#include <cstdint>
-#include <vector>
-#include <unordered_map>
-#include <unordered_set>
-
-struct range_nfd {
-    uint32_t first;
-    uint32_t last;
-    uint32_t nfd;
-};
-
-static const uint32_t MAX_CODEPOINTS = 0x110000;
-
-extern const std::vector<std::pair<uint32_t, uint16_t>> unicode_ranges_flags;
-extern const std::unordered_set<uint32_t> unicode_set_whitespace;
-extern const std::unordered_map<uint32_t, uint32_t> unicode_map_lowercase;
-extern const std::unordered_map<uint32_t, uint32_t> unicode_map_uppercase;
-extern const std::vector<range_nfd> unicode_ranges_nfd;
--- a/examples/talk-llama/unicode.cpp
+++ b/examples/talk-llama/unicode.cpp
--- a/examples/talk-llama/unicode.h
+++ b/examples/talk-llama/unicode.h
@ -4,60 +4,23 @@
 #include <string>
 #include <vector>

-struct codepoint_flags {
-    enum {
-        UNDEFINED       = 0x0001,
-        NUMBER          = 0x0002,  // regex: \p{N}
-        LETTER          = 0x0004,  // regex: \p{L}
-        SEPARATOR       = 0x0008,  // regex: \p{Z}
-        ACCENT_MARK     = 0x0010,  // regex: \p{M}
-        PUNCTUATION     = 0x0020,  // regex: \p{P}
-        SYMBOL          = 0x0040,  // regex: \p{S}
-        CONTROL         = 0x0080,  // regex: \p{C}
-        MASK_CATEGORIES = 0x00FF,
-    };
-
-    // codepoint type
-    uint16_t is_undefined   : 1;
-    uint16_t is_number      : 1;  // regex: \p{N}
-    uint16_t is_letter      : 1;  // regex: \p{L}
-    uint16_t is_separator   : 1;  // regex: \p{Z}
-    uint16_t is_accent_mark : 1;  // regex: \p{M}
-    uint16_t is_punctuation : 1;  // regex: \p{P}
-    uint16_t is_symbol      : 1;  // regex: \p{S}
-    uint16_t is_control     : 1;  // regex: \p{C}
-    // helper flags
-    uint16_t is_whitespace  : 1;  // regex: \s
-    uint16_t is_lowercase   : 1;
-    uint16_t is_uppercase   : 1;
-    uint16_t is_nfd         : 1;
-
-    // decode from uint16
-    inline codepoint_flags(const uint16_t flags=0) {
-        *reinterpret_cast<uint16_t*>(this) = flags;
-    }
-
-    inline uint16_t as_uint() const {
-        return *reinterpret_cast<const uint16_t*>(this);
-    }
-
-    inline uint16_t category_flag() const {
-        return this->as_uint() & MASK_CATEGORIES;
-    }
-};
-
+#define CODEPOINT_TYPE_UNIDENTIFIED 0
+#define CODEPOINT_TYPE_DIGIT        1
+#define CODEPOINT_TYPE_LETTER       2
+#define CODEPOINT_TYPE_WHITESPACE   3
+#define CODEPOINT_TYPE_ACCENT_MARK  4
+#define CODEPOINT_TYPE_PUNCTUATION  5
+#define CODEPOINT_TYPE_SYMBOL       6
+#define CODEPOINT_TYPE_CONTROL      7

 std::string unicode_cpt_to_utf8(uint32_t cp);
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);

 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);

-codepoint_flags unicode_cpt_flags(const uint32_t cp);
-codepoint_flags unicode_cpt_flags(const std::string & utf8);
+int unicode_cpt_type(uint32_t cp);
+int unicode_cpt_type(const std::string & utf8);

 std::string unicode_byte_to_utf8(uint8_t byte);
 uint8_t unicode_utf8_to_byte(const std::string & utf8);

-uint32_t unicode_tolower(uint32_t cp);
-
-std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);
--- a/examples/talk/README.md
+++ b/examples/talk/README.md
@ -11,13 +11,9 @@ Web version: [examples/talk.wasm](/examples/talk.wasm)
 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:
+# Install SDL2 on Linux
 sudo apt-get install libsdl2-dev

-# On Fedora Linux:
-sudo dnf install SDL2 SDL2-devel
-
 # Install SDL2 on Mac OS
 brew install sdl2

--- a/examples/talk/gpt-2.cpp
+++ b/examples/talk/gpt-2.cpp
@ -72,7 +72,7 @@ struct gpt2_model {
 };

 // load the model's weights from a file
-static bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
+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);
@ -380,7 +380,7 @@ static bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_v
 //   - embd_w:    the predicted logits for the next token
 //
 // TODO: sync latest version from ggml repo
-static bool gpt2_eval(
+bool gpt2_eval(
        const gpt2_model & model,
        const int n_threads,
        const int n_past,
--- a/examples/talk/talk.cpp
+++ b/examples/talk/talk.cpp
@ -26,12 +26,12 @@ struct whisper_params {
    float vad_thold    = 0.6f;
    float freq_thold   = 100.0f;

+    bool speed_up      = false;
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
    bool use_gpu       = true;
-    bool flash_attn    = false;

    std::string person    = "Santa";
    std::string language  = "en";
@ -44,7 +44,7 @@ struct whisper_params {

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);

-static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];

@ -59,11 +59,11 @@ static bool whisper_params_parse(int argc, char ** argv, whisper_params & params
        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 == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-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]; }
@ -94,11 +94,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -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());
@ -109,7 +109,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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) {
+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;
@ -129,6 +129,7 @@ static std::string transcribe(whisper_context * ctx, const whisper_params & para
    wparams.n_threads        = params.n_threads;

    wparams.audio_ctx        = params.audio_ctx;
+    wparams.speed_up         = params.speed_up;

    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
        return "";
@ -187,9 +188,7 @@ int main(int argc, char ** argv) {

    // whisper init
    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);

--- a/examples/wchess/CMakeLists.txt
+++ b/examples/wchess/CMakeLists.txt
@ -1,10 +1,9 @@
+set(CMAKE_CXX_STANDARD 11)
+
 add_subdirectory(libwchess)
-set_target_properties(wchess-core PROPERTIES FOLDER "libs")

 if (EMSCRIPTEN)
    add_subdirectory(wchess.wasm)
-    set_target_properties(wchess.wasm PROPERTIES FOLDER "libs")
 else()
    add_subdirectory(wchess.cmd)
-    set_target_properties(wchess PROPERTIES FOLDER "libs")
 endif()
--- a/examples/wchess/wchess.cmd/wchess.cmd.cpp
+++ b/examples/wchess/wchess.cmd/wchess.cmd.cpp
@ -26,12 +26,12 @@ struct whisper_params {

    float grammar_penalty = 100.0f;

+    bool speed_up      = false;
    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 language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -56,11 +56,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,        --speed-up       [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
-    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention during decoding\n",             params.flash_attn ? "true" : "false");
    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
@ -87,11 +87,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        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 == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
-        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
@ -183,9 +183,7 @@ int main(int argc, char ** argv) {
    // whisper init

    struct whisper_context_params cparams = whisper_context_default_params();
-
-    cparams.use_gpu    = params.use_gpu;
-    cparams.flash_attn = params.flash_attn;
+    cparams.use_gpu = params.use_gpu;

    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
    if (!ctx) {
--- a/examples/whisper.android.java/app/src/main/jni/whisper/CMakeLists.txt
+++ b/examples/whisper.android.java/app/src/main/jni/whisper/CMakeLists.txt
@ -5,14 +5,15 @@ project(whisper.cpp)
 set(CMAKE_CXX_STANDARD 11)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../../)

-set(SOURCE_FILES
-    ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-    ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
-    ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.c
-    ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
-    ${WHISPER_LIB_DIR}/src/whisper.cpp
-    ${CMAKE_SOURCE_DIR}/jni.c
-    )
+set(
+        SOURCE_FILES
+        ${WHISPER_LIB_DIR}/ggml.c
+        ${WHISPER_LIB_DIR}/ggml-alloc.c
+        ${WHISPER_LIB_DIR}/ggml-backend.c
+        ${WHISPER_LIB_DIR}/ggml-quants.c
+        ${WHISPER_LIB_DIR}/whisper.cpp
+        ${CMAKE_SOURCE_DIR}/jni.c
+)

 find_library(LOG_LIB log)

@ -40,6 +41,7 @@ function(build_library target_name)
        #target_link_options(${target_name} PRIVATE -Wl,--gc-sections)
        #target_link_options(${target_name} PRIVATE -Wl,--exclude-libs,ALL)
        #target_link_options(${target_name} PRIVATE -flto)
+
    endif ()
 endfunction()

@ -52,7 +54,3 @@ elseif (${ANDROID_ABI} STREQUAL "armeabi-v7a")
 endif ()

 include_directories(${WHISPER_LIB_DIR})
-include_directories(${WHISPER_LIB_DIR}/src)
-include_directories(${WHISPER_LIB_DIR}/include)
-include_directories(${WHISPER_LIB_DIR}/ggml/include)
-include_directories(${WHISPER_LIB_DIR}/ggml/src)
--- a/examples/whisper.android/README.md
+++ b/examples/whisper.android/README.md
@ -12,3 +12,47 @@ To use:
 (PS: Do not move this android project folder individually to other folders, because this android project folder depends on the files of the whole project.)

 <img width="300" alt="image" src="https://user-images.githubusercontent.com/1670775/221613663-a17bf770-27ef-45ab-9a46-a5f99ba65d2a.jpg">
+
+## CLBlast
+
+> [!NOTE]
+> - OpenCL does not have the same level of support as CUDA or Metal.
+> - Turning on CLBlast may degrade OpenCL performance if your device isn't already tuned. See [tuning.md](https://github.com/CNugteren/CLBlast/blob/162783a414969464ce3aa5adf5c2554afa5ee93e/doc/tuning.md#already-tuned-for-devices) for a list of devices that are already tuned and what to do if yours is missing.
+
+Build CLBlast.
+
+```
+# In path/to/CLBlast (we assume OpenCL-Headers relative location)
+$ANDROID_SDK_PATH/cmake/3.22.1/bin/cmake .. \
+    -DCMAKE_SYSTEM_NAME=Android \
+    -DCMAKE_SYSTEM_VERSION=33 \
+    -DCMAKE_ANDROID_ARCH_ABI=arm64-v8a \
+    -DCMAKE_ANDROID_NDK=$ANDROID_NDK_PATH \
+    -DCMAKE_ANDROID_STL_TYPE=c++_static \
+    -DOPENCL_ROOT=$(readlink -f ../../OpenCL-Headers) \
+    -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=BOTH \
+    -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+# Build libclblast.so
+make -j4
+```
+
+Pull `libGLES_mali.so` to `libOpenCL.so`.
+
+```bash
+# In path/to/whisper.android
+mkdir lib/src/main/jniLibs/arm64-v8a
+adb pull /system/vendor/lib64/egl/libGLES_mali.so lib/src/main/jniLibs/arm64-v8a/libOpenCL.so
+```
+
+In gradle.properties, set `GGML_HOME` to the location of GGML, as well as
+required options for turning on CLBlast.
+
+```
+GGML_HOME=/path/to/ggml
+GGML_CLBLAST=ON
+CLBLAST_HOME=/path/to/CLBlast
+OPENCL_LIB=/path/to/libOpenCL.so
+OPENCL_ROOT=/path/to/OpenCL-Headers
+```
+
--- a/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
+++ b/examples/whisper.android/app/src/main/java/com/whispercppdemo/ui/main/MainScreenViewModel.kt
@ -145,7 +145,7 @@ class MainScreenViewModel(private val application: Application) : ViewModel() {
            val start = System.currentTimeMillis()
            val text = whisperContext?.transcribeData(data)
            val elapsed = System.currentTimeMillis() - start
-            printMessage("Done ($elapsed ms): \n$text\n")
+            printMessage("Done ($elapsed ms): $text\n")
        } catch (e: Exception) {
            Log.w(LOG_TAG, e)
            printMessage("${e.localizedMessage}\n")
--- a/examples/whisper.android/lib/src/main/java/com/whispercpp/whisper/LibWhisper.kt
+++ b/examples/whisper.android/lib/src/main/java/com/whispercpp/whisper/LibWhisper.kt
@ -16,7 +16,7 @@ class WhisperContext private constructor(private var ptr: Long) {
        Executors.newSingleThreadExecutor().asCoroutineDispatcher()
    )

-    suspend fun transcribeData(data: FloatArray, printTimestamp: Boolean = true): String = withContext(scope.coroutineContext) {
+    suspend fun transcribeData(data: FloatArray): String = withContext(scope.coroutineContext) {
        require(ptr != 0L)
        val numThreads = WhisperCpuConfig.preferredThreadCount
        Log.d(LOG_TAG, "Selecting $numThreads threads")
@ -24,13 +24,7 @@ class WhisperContext private constructor(private var ptr: Long) {
        val textCount = WhisperLib.getTextSegmentCount(ptr)
        return@withContext buildString {
            for (i in 0 until textCount) {
-                if (printTimestamp) {
-                    val textTimestamp = "[${toTimestamp(WhisperLib.getTextSegmentT0(ptr, i))} --> ${toTimestamp(WhisperLib.getTextSegmentT1(ptr, i))}]"
-                    val textSegment = WhisperLib.getTextSegment(ptr, i)
-                    append("$textTimestamp: $textSegment\n")
-                } else {
-                    append(WhisperLib.getTextSegment(ptr, i))
-                }
+                append(WhisperLib.getTextSegment(ptr, i))
            }
        }
    }
@ -137,29 +131,12 @@ private class WhisperLib {
        external fun fullTranscribe(contextPtr: Long, numThreads: Int, audioData: FloatArray)
        external fun getTextSegmentCount(contextPtr: Long): Int
        external fun getTextSegment(contextPtr: Long, index: Int): String
-        external fun getTextSegmentT0(contextPtr: Long, index: Int): Long
-        external fun getTextSegmentT1(contextPtr: Long, index: Int): Long
        external fun getSystemInfo(): String
        external fun benchMemcpy(nthread: Int): String
        external fun benchGgmlMulMat(nthread: Int): String
    }
 }

-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-private fun toTimestamp(t: Long, comma: Boolean = false): String {
-    var msec = t * 10
-    val hr = msec / (1000 * 60 * 60)
-    msec -= hr * (1000 * 60 * 60)
-    val min = msec / (1000 * 60)
-    msec -= min * (1000 * 60)
-    val sec = msec / 1000
-    msec -= sec * 1000
-
-    val delimiter = if (comma) "," else "."
-    return String.format("%02d:%02d:%02d%s%03d", hr, min, sec, delimiter, msec)
-}
-
 private fun isArmEabiV7a(): Boolean {
    return Build.SUPPORTED_ABIS[0].equals("armeabi-v7a")
 }
--- a/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
+++ b/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
@ -10,7 +10,7 @@ option(GGML_HOME       "whisper: Path to external GGML source" OFF)

 set(
    SOURCE_FILES
-    ${WHISPER_LIB_DIR}/src/whisper.cpp
+    ${WHISPER_LIB_DIR}/whisper.cpp
    ${CMAKE_SOURCE_DIR}/jni.c
    )

@ -18,10 +18,10 @@ if (NOT GGML_HOME)
    set(
        SOURCE_FILES
        ${SOURCE_FILES}
-        ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
+        ${WHISPER_LIB_DIR}/ggml.c
+        ${WHISPER_LIB_DIR}/ggml-alloc.c
+        ${WHISPER_LIB_DIR}/ggml-backend.c
+        ${WHISPER_LIB_DIR}/ggml-quants.c
        )
 endif()

@ -75,7 +75,3 @@ endif ()
 build_library("whisper") # Default target

 include_directories(${WHISPER_LIB_DIR})
-include_directories(${WHISPER_LIB_DIR}/src)
-include_directories(${WHISPER_LIB_DIR}/include)
-include_directories(${WHISPER_LIB_DIR}/ggml/include)
-include_directories(${WHISPER_LIB_DIR}/ggml/src)
--- a/examples/whisper.android/lib/src/main/jni/whisper/jni.c
+++ b/examples/whisper.android/lib/src/main/jni/whisper/jni.c
@ -212,22 +212,6 @@ Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegment(
    return string;
 }

-JNIEXPORT jlong JNICALL
-Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegmentT0(
-        JNIEnv *env, jobject thiz, jlong context_ptr, jint index) {
-    UNUSED(thiz);
-    struct whisper_context *context = (struct whisper_context *) context_ptr;
-    return whisper_full_get_segment_t0(context, index);
-}
-
-JNIEXPORT jlong JNICALL
-Java_com_whispercpp_whisper_WhisperLib_00024Companion_getTextSegmentT1(
-        JNIEnv *env, jobject thiz, jlong context_ptr, jint index) {
-    UNUSED(thiz);
-    struct whisper_context *context = (struct whisper_context *) context_ptr;
-    return whisper_full_get_segment_t1(context, index);
-}
-
 JNIEXPORT jstring JNICALL
 Java_com_whispercpp_whisper_WhisperLib_00024Companion_getSystemInfo(
        JNIEnv *env, jobject thiz
--- a/examples/whisper.nvim/whisper.nvim
+++ b/examples/whisper.nvim/whisper.nvim
@ -45,6 +45,6 @@ if [ ! -f ./models/ggml-${model}.bin ] ; then
 fi

 # fine-tune the parameters according to your machine specs
-./stream -t 8 -m models/ggml-${model}.bin --step 350 --length 10000 -f /tmp/whisper.nvim 2> /dev/null
+./stream -t 8 -m models/ggml-base.en.bin --step 350 --length 10000 -f /tmp/whisper.nvim 2> /dev/null

 exit 0
--- a/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
+++ b/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
@ -9,6 +9,7 @@
 /* Begin PBXBuildFile section */
 		1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 184447182AB211A2007D6BFE /* ggml-alloc.c */; };
 		1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 1844471B2AB21655007D6BFE /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-framework Foundation -framework Metal -framework MetalKit -fno-objc-arc"; }; };
+		184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7A29052BDF00BD2A04 /* AppDelegate.m */; };
 		18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C7D29052BDF00BD2A04 /* SceneDelegate.m */; };
 		18627C8129052BDF00BD2A04 /* ViewController.m in Sources */ = {isa = PBXBuildFile; fileRef = 18627C8029052BDF00BD2A04 /* ViewController.m */; };
@ -19,8 +20,6 @@
 		18627C9429052C4900BD2A04 /* whisper.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9329052C4900BD2A04 /* whisper.cpp */; settings = {COMPILER_FLAGS = "-DWHISPER_USE_COREML -DWHISPER_COREML_ALLOW_FALLBACK -DGGML_USE_METAL"; }; };
 		18627C9629052C5800BD2A04 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 18627C9529052C5800BD2A04 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL"; }; };
 		18627C9B29052CFF00BD2A04 /* ggml-base.en.bin in Resources */ = {isa = PBXBuildFile; fileRef = 18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */; };
-		18A276062C2A98A5001C8D37 /* ggml-metal.metal in Copy Files */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
-		18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 1844471D2AB2195F007D6BFE /* ggml-metal.metal */; };
 		18ABE15A2AF556340044A204 /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.c */; };
 		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
@ -30,24 +29,23 @@
 /* End PBXBuildFile section */

 /* Begin PBXCopyFilesBuildPhase section */
-		184447202AB21B25007D6BFE /* Copy Files */ = {
+		184447202AB21B25007D6BFE /* CopyFiles */ = {
 			isa = PBXCopyFilesBuildPhase;
 			buildActionMask = 2147483647;
 			dstPath = "";
 			dstSubfolderSpec = 7;
 			files = (
-				18A276062C2A98A5001C8D37 /* ggml-metal.metal in Copy Files */,
+				184447212AB21B43007D6BFE /* ggml-metal.metal in CopyFiles */,
 			);
-			name = "Copy Files";
 			runOnlyForDeploymentPostprocessing = 0;
 		};
 /* End PBXCopyFilesBuildPhase section */

 /* Begin PBXFileReference section */
-		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml/src/ggml-alloc.c"; sourceTree = "<group>"; };
-		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml/include/ggml-alloc.h"; sourceTree = "<group>"; };
-		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal.m"; sourceTree = "<group>"; };
-		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal.metal"; sourceTree = "<group>"; };
+		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml-alloc.c"; sourceTree = "<group>"; };
+		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml-alloc.h"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@ -60,19 +58,17 @@
 		18627C8829052BE000BD2A04 /* Base */ = {isa = PBXFileReference; lastKnownFileType = file.storyboard; name = Base; path = Base.lproj/LaunchScreen.storyboard; sourceTree = "<group>"; };
 		18627C8A29052BE000BD2A04 /* Info.plist */ = {isa = PBXFileReference; lastKnownFileType = text.plist.xml; path = Info.plist; sourceTree = "<group>"; };
 		18627C8B29052BE000BD2A04 /* main.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = main.m; sourceTree = "<group>"; };
-		18627C9229052C2B00BD2A04 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../include/whisper.h; sourceTree = "<group>"; };
-		18627C9329052C4900BD2A04 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../src/whisper.cpp; sourceTree = "<group>"; };
-		18627C9529052C5800BD2A04 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml/src/ggml.c; sourceTree = "<group>"; };
-		18627C9729052C6600BD2A04 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml/include/ggml.h; sourceTree = "<group>"; };
+		18627C9229052C2B00BD2A04 /* whisper.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = whisper.h; path = ../../../whisper.h; sourceTree = "<group>"; };
+		18627C9329052C4900BD2A04 /* whisper.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = whisper.cpp; path = ../../../whisper.cpp; sourceTree = "<group>"; };
+		18627C9529052C5800BD2A04 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../../ggml.c; sourceTree = "<group>"; };
+		18627C9729052C6600BD2A04 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../../ggml.h; sourceTree = "<group>"; };
 		18627C9A29052CFF00BD2A04 /* ggml-base.en.bin */ = {isa = PBXFileReference; lastKnownFileType = archive.macbinary; name = "ggml-base.en.bin"; path = "../../../models/ggml-base.en.bin"; sourceTree = "<group>"; };
-		18A275FE2C2A94DE001C8D37 /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../../ggml/include/ggml-metal.h"; sourceTree = "<group>"; };
-		18A275FF2C2A9563001C8D37 /* ggml-common.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-common.h"; path = "../../../ggml/src/ggml-common.h"; sourceTree = "<group>"; };
-		18ABE1542AF556340044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../../ggml/src/ggml-quants.h"; sourceTree = "<group>"; };
-		18ABE1552AF556340044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../../ggml/include/ggml-backend.h"; sourceTree = "<group>"; };
-		18ABE1562AF556340044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend-impl.h"; path = "../../../ggml/src/ggml-backend-impl.h"; sourceTree = "<group>"; };
-		18ABE1572AF556340044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../../ggml/src/ggml-backend.c"; sourceTree = "<group>"; };
-		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
-		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
+		18ABE1542AF556340044A204 /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../../ggml-quants.h"; sourceTree = "<group>"; };
+		18ABE1552AF556340044A204 /* ggml-backend.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../../ggml-backend.h"; sourceTree = "<group>"; };
+		18ABE1562AF556340044A204 /* ggml-backend-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-backend-impl.h"; path = "../../../ggml-backend-impl.h"; sourceTree = "<group>"; };
+		18ABE1572AF556340044A204 /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../../ggml-backend.c"; sourceTree = "<group>"; };
+		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml-impl.h"; sourceTree = "<group>"; };
+		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml-quants.c"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@ -112,8 +108,6 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
-				18A275FF2C2A9563001C8D37 /* ggml-common.h */,
-				18A275FE2C2A94DE001C8D37 /* ggml-metal.h */,
 				18ABE1562AF556340044A204 /* ggml-backend-impl.h */,
 				18ABE1572AF556340044A204 /* ggml-backend.c */,
 				18ABE1552AF556340044A204 /* ggml-backend.h */,
@ -157,7 +151,7 @@
 				7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */,
 			);
 			name = coreml;
-			path = ../../../src/coreml;
+			path = ../../../coreml;
 			sourceTree = "<group>";
 		};
 /* End PBXGroup section */
@ -170,7 +164,7 @@
 				18627C7229052BDF00BD2A04 /* Sources */,
 				18627C7329052BDF00BD2A04 /* Frameworks */,
 				18627C7429052BDF00BD2A04 /* Resources */,
-				184447202AB21B25007D6BFE /* Copy Files */,
+				184447202AB21B25007D6BFE /* CopyFiles */,
 			);
 			buildRules = (
 			);
@ -188,7 +182,7 @@
 			isa = PBXProject;
 			attributes = {
 				BuildIndependentTargetsInParallel = 1;
-				LastUpgradeCheck = 1540;
+				LastUpgradeCheck = 1400;
 				TargetAttributes = {
 					18627C7529052BDF00BD2A04 = {
 						CreatedOnToolsVersion = 14.0.1;
@ -218,7 +212,6 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				18A2760B2C2A9B43001C8D37 /* ggml-metal.metal in Resources */,
 				18627C8929052BE000BD2A04 /* LaunchScreen.storyboard in Resources */,
 				7FE3424F2A0C418A0015A058 /* ggml-base.en-encoder.mlmodelc in Resources */,
 				18627C8629052BE000BD2A04 /* Assets.xcassets in Resources */,
@ -308,7 +301,6 @@
 				DEBUG_INFORMATION_FORMAT = dwarf;
 				ENABLE_STRICT_OBJC_MSGSEND = YES;
 				ENABLE_TESTABILITY = YES;
-				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GCC_C_LANGUAGE_STANDARD = gnu11;
 				GCC_DYNAMIC_NO_PIC = NO;
 				GCC_NO_COMMON_BLOCKS = YES;
@ -367,7 +359,6 @@
 				DEBUG_INFORMATION_FORMAT = "dwarf-with-dsym";
 				ENABLE_NS_ASSERTIONS = NO;
 				ENABLE_STRICT_OBJC_MSGSEND = YES;
-				ENABLE_USER_SCRIPT_SANDBOXING = YES;
 				GCC_C_LANGUAGE_STANDARD = gnu11;
 				GCC_NO_COMMON_BLOCKS = YES;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = YES;
@ -409,7 +400,6 @@
 					"@executable_path/Frameworks",
 				);
 				MARKETING_VERSION = 1.0;
-				MTL_HEADER_SEARCH_PATHS = "";
 				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SWIFT_EMIT_LOC_STRINGS = YES;
@ -438,7 +428,6 @@
 					"@executable_path/Frameworks",
 				);
 				MARKETING_VERSION = 1.0;
-				MTL_HEADER_SEARCH_PATHS = "";
 				PRODUCT_BUNDLE_IDENTIFIER = "com.ggerganov.whisper-objc";
 				PRODUCT_NAME = "$(TARGET_NAME)";
 				SWIFT_EMIT_LOC_STRINGS = YES;
--- a/examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift
+++ b/examples/whisper.swiftui/whisper.swiftui.demo/Models/WhisperState.swift
@ -15,7 +15,7 @@ class WhisperState: NSObject, ObservableObject, AVAudioRecorderDelegate {
    private var audioPlayer: AVAudioPlayer?
    
    private var modelUrl: URL? {
-        Bundle.main.url(forResource: "ggml-base.en", withExtension: "bin", subdirectory: "models")
+        Bundle.main.url(forResource: "ggml-tiny.en", withExtension: "bin", subdirectory: "models")
    }
    
    private var sampleUrl: URL? {
--- a/scripts/bench-all.sh
+++ b/scripts/bench-all.sh
@ -2,7 +2,7 @@

 # Helper script to run the bench tool on all models and print the results in share-able format

-printf "Usage: ./bench.sh [n_threads] [encoder-only] [flash-attn]\n"
+printf "Usage: ./bench.sh [n_threads] [encoder-only]\n"

 if [ -z "$1" ]; then
    n_threads=4
@ -11,19 +11,12 @@ else
 fi

 encoder_only=0
-if [ -z "$2" ] || [ "$2" -eq 0 ]; then
+if [ -z "$2" ]; then
    encoder_only=0
 else
    encoder_only=$2
 fi

-fattn=""
-if [ -z "$3" ] || [ "$3" -eq 0 ]; then
-    fattn=""
-else
-    fattn="-fa"
-fi
-
 models=(                                                                                                    \
      "tiny"     "tiny-q4_0"     "tiny-q4_1"     "tiny-q5_0"     "tiny-q5_1"     "tiny-q8_0"                \
      "base"     "base-q4_0"     "base-q4_1"     "base-q5_0"     "base-q5_1"     "base-q8_0"                \
@ -51,19 +44,13 @@ if [ "$encoder_only" -eq 0 ]; then
    printf "\n"
 fi

-if [ "$fattn" == "-fa" ]; then
-    fattn_i=1
-else
-    fattn_i=0
-fi
-
-printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "FA" "Enc." "Dec." "Bch5" "PP" "Commit"
-printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"
+printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "CPU" "OS" "Config" "Model" "Th" "Enc." "Dec." "Bch5" "PP" "Commit"
+printf "| %6s | %6s | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "---" "---" "---" "---" "---" "---" "---" "---" "---" "---"

 for model in "${models[@]}"; do
    # actual run
    # store stderr output in a variable in order to parse it later
-    output=$(./bench -m ./models/ggml-$model.bin -t $n_threads $fattn 2>&1)
+    output=$(./bench -m ./models/ggml-$model.bin -t $n_threads 2>&1)
    ret=$?

    # parse the output:
@ -108,6 +95,6 @@ for model in "${models[@]}"; do
    commit=$(git rev-parse --short HEAD)

    if [ $ret -eq 0 ]; then
-        printf "| <todo> | <todo> | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$fattn_i" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
+        printf "| <todo> | <todo> | %16s | %13s | %3s | %7s | %7s | %7s | %7s | %7s |\n" "$config" "$model" "$n_threads" "$encode_time" "$decode_time" "$batchd_time" "$prompt_time" "$commit"
    fi
 done
--- a/scripts/bench-wts.sh
+++ b/scripts/bench-wts.sh
--- a/scripts/bench.py
+++ b/scripts/bench.py
--- a/scripts/convert-all.sh
+++ b/scripts/convert-all.sh
--- a/scripts/deploy-wasm.sh
+++ b/scripts/deploy-wasm.sh
@ -4,7 +4,7 @@
 # Run from the build directory:
 #
 # cd build-em
-# ../scripts/deploy-wasm.sh
+# ../extra/deploy-wasm.sh
 #

 # check if emcmake is available
--- a/scripts/quantize-all.sh
+++ b/scripts/quantize-all.sh
--- a/scripts/sha-all.sh
+++ b/scripts/sha-all.sh
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
slaren	66df44b0b7	alloc : fix allocation data of pre-allocated leafs	2024-03-16 15:47:14 +01:00
Georgi Gerganov	c6c94de43a	whisper : allocate encoder results in dedicated buffer	2024-03-16 16:15:12 +02:00