whisper : remove comment

command : enable beam-search, add "no_timestamps", add "context", add p
grammars : add assistant + update comments
2025-06-24 17:15:19 +00:00 · 2023-09-10 13:27:06 +03:00 · 2023-09-10 12:22:57 +03:00 · 2023-09-09 20:24:58 +03:00 · 2023-09-09 20:05:57 +03:00 · 2023-09-06 17:05:05 +03:00
262 changed files with 15336 additions and 113990 deletions
--- a/.devops/cublas.Dockerfile
+++ b/.devops/cublas.Dockerfile
@ -1,28 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
-
-# Target the CUDA build image
-ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_CUDA_DEV_CONTAINER} as build
-
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
-
-RUN apt-get update && \
-    apt-get install -y build-essential git cmake
-
-WORKDIR /app
-
-COPY . .
-
-# Set nvcc architecture
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV WHISPER_CUBLAS=1
-
-RUN make
-
-ENTRYPOINT ["/app/main"]
--- a/.devops/main-cuda.Dockerfile
+++ b/.devops/main-cuda.Dockerfile
@ -1,38 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.3.1
-# Target the CUDA build image
-ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-# Target the CUDA runtime image
-ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
-WORKDIR /app
-
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
-# Set nvcc architecture
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV WHISPER_CUBLAS=1
-
-RUN apt-get update && \
-    apt-get install -y build-essential \
-    && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-# Ref: https://stackoverflow.com/a/53464012
-ENV CUDA_MAIN_VERSION=12.3
-ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
-
-COPY .. .
-RUN make
-
-FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY --from=build /app /app
-ENTRYPOINT [ "bash", "-c" ]
--- a/.devops/main.Dockerfile
+++ b/.devops/main.Dockerfile
@ -1,19 +0,0 @@
-FROM ubuntu:22.04 AS build
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y build-essential \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY .. .
-RUN make
-
-FROM ubuntu:22.04 AS runtime
-WORKDIR /app
-
-RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
-  && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
-
-COPY --from=build /app /app
-ENTRYPOINT [ "bash", "-c" ]
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@ -25,7 +25,6 @@ jobs:
          docker run --platform ${{ matrix.arch }} --rm \
            -v ${{ github.workspace }}:/workspace \
            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
            apt update
            apt install -y build-essential libsdl2-dev
            make
@ -87,10 +86,9 @@ jobs:
          docker run --platform ${{ matrix.arch }} --rm \
            -v ${{ github.workspace }}:/workspace \
            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
            apt update
            apt install -y build-essential cmake libsdl2-dev
-            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }}
            make
            ctest -L gh --output-on-failure'

@ -115,10 +113,9 @@ jobs:
          docker run --platform ${{ matrix.arch }} --rm \
            -v ${{ github.workspace }}:/workspace \
            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
            apt update
-            apt install -y clang build-essential cmake libsdl2-dev
-            cmake . -DWHISPER_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
+            apt install -y build-essential cmake libsdl2-dev
+            cmake . -DWHISPER_SUPPORT_SDL2=ON -DCMAKE_BUILD_TYPE=${{ matrix.build }} -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang
            make
            ctest -L gh --output-on-failure'

@ -143,113 +140,12 @@ jobs:
          docker run --platform ${{ matrix.arch }} --rm \
            -v ${{ github.workspace }}:/workspace \
            -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
-            set -e
            apt update
            apt install -y build-essential cmake
            cmake . -DCMAKE_BUILD_TYPE=Debug -DWHISPER_SANITIZE_${{ matrix.sanitizer }}=ON
            make
            ctest -L gh --output-on-failure'

-  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] 
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v3
-        
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-            
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          mkdir build
-          cd build
-          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] 
-        arch: [linux/amd64, linux/arm64, linux/arm/v7, linux/ppc64le]
-
-    continue-on-error: true
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v3
-        
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-            
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          mkdir build
-          cd build
-          cmake -DWHISPER_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
-          cmake --build . --config Release -j $(nproc)
-  
  windows:
    runs-on: windows-latest

@ -266,7 +162,7 @@ jobs:
            s2arc: x64
            jnaPath: win32-x86-64
          - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0

    steps:
      - name: Clone
@ -286,7 +182,7 @@ jobs:
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}

      - name: Build
        run: |
@ -321,16 +217,13 @@ jobs:
        sdl2: [ON]
        include:
          - arch: Win32
-            obzip: https://github.com/OpenMathLib/OpenBLAS/releases/download/v0.3.25/OpenBLAS-0.3.25-x86.zip
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-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
+            obzip: https://github.com/xianyi/OpenBLAS/releases/download/v0.3.21/OpenBLAS-0.3.21-x64.zip
            s2arc: x64
-            clblast: ON
-            clver: 1.6.1
          - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0

    steps:
      - name: Clone
@ -346,7 +239,7 @@ jobs:
          7z x blas.zip -oblas -y
          copy blas/include/cblas.h .
          copy blas/include/openblas_config.h .
-          echo "OPENBLAS_PATH=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV
+          echo "blasdir=$env:GITHUB_WORKSPACE/blas" >> $env:GITHUB_ENV

      - name: Fetch SDL2 and set SDL2_DIR
        if: matrix.sdl2 == 'ON'
@ -355,26 +248,13 @@ jobs:
          7z x sdl2.zip
          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV

-      - name: Install OpenCL
-        if: matrix.clblast == 'ON'
-        run: vcpkg.exe --triplet=${{ matrix.arch }}-windows install opencl
-
-      - name: Fetch CLBlast and set CLBlast_DIR
-        if: matrix.clblast == 'ON'
-        run: |
-          C:/msys64/usr/bin/wget.exe -qO clblast.zip https://github.com/CNugteren/CLBlast/releases/download/${{ matrix.clver }}/CLBlast-${{ matrix.clver }}-windows-x64.zip
-          7z x clblast.zip
-          7z x CLBlast-${{ matrix.clver }}-windows-x64.7z
-          echo "CLBlast_DIR=$env:GITHUB_WORKSPACE/CLBlast-${{ matrix.clver }}-windows-x64/lib/cmake/CLBlast" >> $env:GITHUB_ENV
-
      - name: Configure
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_OPENBLAS=${{ matrix.blas }}
-          -DCMAKE_LIBRARY_PATH="$env:OPENBLAS_PATH/lib"
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
-          -DWHISPER_CLBLAST=${{ matrix.clblast }}
+          -DWHISPER_SUPPORT_OPENBLAS=${{ matrix.blas }}
+          -DCMAKE_LIBRARY_PATH="$env:blasdir/lib"
+          -DWHISPER_SUPPORT_SDL2=${{ matrix.sdl2 }}

      - name: Build
        run: |
@ -383,21 +263,17 @@ jobs:

      - name: Copy libopenblas.dll
        if: matrix.blas == 'ON'
-        run: copy "$env:OPENBLAS_PATH/bin/libopenblas.dll" build/bin/${{ matrix.build }}
+        run: copy "$env:blasdir/bin/libopenblas.dll" build/bin/${{ matrix.build }}

      - name: Copy SDL2.dll
        if: matrix.sdl2 == 'ON'
        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}

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

  windows-cublas:
@ -409,12 +285,11 @@ jobs:
        arch: [x64]
        cublas: [ON]
        sdl2: [ON]
-        cuda-toolkit: [12.2.0, 11.8.0]
        include:
          - arch: x64
            s2arc: x64
          - sdl2: ON
-            s2ver: 2.28.5
+            s2ver: 2.26.0

    steps:
      - name: Clone
@ -425,9 +300,7 @@ jobs:

      - name: Install CUDA Toolkit
        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.11
-        with:
-          cuda: '${{ matrix.cuda-toolkit }}'
+        uses: Jimver/cuda-toolkit@v0.2.10

      - name: Fetch SDL2 and set SDL2_DIR
        if: matrix.sdl2 == 'ON'
@ -440,20 +313,12 @@ jobs:
        run: >
          cmake -S . -B ./build -A ${{ matrix.arch }}
          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DWHISPER_CUBLAS=${{ matrix.cublas }}
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+          -DWHISPER_CUBLAS=1

-      - name: Build ${{ matrix.cuda-toolkit }}
+      - name: Build
        run: |
          cd ./build
-          cmake --build . --config ${{ matrix.build }}
-
-      - name: Copy CUDA DLLs
-        run: >
-          Copy-Item -PassThru
-          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
-          -Include cudart64_*,cublas64_*,cublasLt64_*
-          -Destination build/bin/${{ matrix.build }}
+          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}

      - name: Copy SDL2.dll
        if: matrix.sdl2 == 'ON'
@ -463,7 +328,7 @@ jobs:
        if: matrix.sdl2 == 'ON'
        uses: actions/upload-artifact@v1
        with:
-          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
+          name: whisper-cublas-bin-${{ matrix.arch }}
          path: build/bin/${{ matrix.build }}

  emscripten:
@ -516,14 +381,6 @@ jobs:
    steps:
      - name: Clone
        uses: actions/checkout@v3
-        with:
-          path: whisper
-
-      - name: Clone
-        uses: actions/checkout@v3
-        with:
-          repository: ggerganov/ggml
-          path: ggml

      - name: Install Java
        uses: actions/setup-java@v3
@ -536,41 +393,9 @@ jobs:

      - name: Build
        run: |
-          cd whisper/examples/whisper.android
+          cd examples/whisper.android
          ./gradlew assembleRelease --no-daemon

-      - name: Build with external ggml
-        run: |
-          export PATH_TO_GGML=$PWD/ggml
-          cd whisper/examples/whisper.android
-          ./gradlew assembleRelease --no-daemon -PGGML_HOME=$PATH_TO_GGML
-
-  android_java:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v3
-
-      - name: set up JDK 11
-        uses: actions/setup-java@v3
-        with:
-          java-version: '11'
-          distribution: 'temurin'
-          cache: gradle
-
-      - name: Setup Android SDK
-        uses: android-actions/setup-android@v2
-        with:
-          api-level: 30
-          build-tools-version: 30.0.3
-
-      - name: Build
-        run: |
-          cd examples/whisper.android.java
-          chmod +x ./gradlew
-          ./gradlew assembleRelease
-
  java:
    needs: [ 'windows' ]
    runs-on: windows-latest
@ -603,15 +428,15 @@ jobs:

      - name: Publish package
        if: ${{ github.ref == 'refs/heads/master' }}
-        uses: gradle/gradle-build-action@v2.4.2
+        uses: gradle/gradle-build-action@v2
        with:
          arguments: publish
          build-root-directory: bindings/java
        env:
          MAVEN_USERNAME: ${{ secrets.JIRA_USER }}
          MAVEN_PASSWORD: ${{ secrets.JIRA_PASS }}
-          PGP_SECRET: ${{ secrets.GPG_PRIVATE_KEY }}
-          PGP_PASSPHRASE: ${{ secrets.GPG_PASSPHRASE }}
+          # MAVEN_USERNAME: ${{ secrets.OSSRH_USERNAME }}
+          # MAVEN_PASSWORD: ${{ secrets.OSSRH_TOKEN }}

  quantize:
    runs-on: ubuntu-latest
--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@ -1,57 +0,0 @@
-name: Publish Docker image
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-jobs:
-  push_to_registry:
-    name: Push Docker image to Docker Hub
-    if: github.event.pull_request.draft == false
-
-    runs-on: ubuntu-latest
-    env:
-      COMMIT_SHA: ${{ github.sha }}
-    strategy:
-      matrix:
-        config:
-          - { tag: "main", dockerfile: ".devops/main.Dockerfile", platform: "linux/amd64,linux/arm64" }
-          - { tag: "main-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platform: "linux/amd64" }
-
-    steps:
-      - name: Check out the repo
-        uses: actions/checkout@v3
-
-      - name: Set up QEMU
-        uses: docker/setup-qemu-action@v3
-
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-
-      - name: Log in to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          registry: ghcr.io
-          username: ${{ github.repository_owner }}
-          password: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Build and push Docker image (versioned)
-        if: github.event_name == 'push'
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          push: true
-          platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
-          file: ${{ matrix.config.dockerfile }}
-
-      - name: Build and push Docker image (tagged)
-        uses: docker/build-push-action@v4
-        with:
-          context: .
-          push: ${{ github.event_name == 'push' }}
-          platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository }}:${{ matrix.config.tag }}"
-          file: ${{ matrix.config.dockerfile }}
--- a/.gitignore
+++ b/.gitignore
@ -8,7 +8,6 @@
 .DS_Store

 build/
-build-coreml/
 build-em/
 build-debug/
 build-release/
@ -19,11 +18,6 @@ build-no-accel/
 build-sanitize-addr/
 build-sanitize-thread/

-# SPM
-.build/
-.swiftpm
-*.metallib
-
 /main
 /stream
 /command
@ -31,7 +25,6 @@ build-sanitize-thread/
 /talk-llama
 /bench
 /quantize
-/server
 /lsp

 arm_neon.h
@ -53,9 +46,3 @@ models/*.mlpackage
 bindings/java/.gradle/
 bindings/java/.idea/
 .idea/
-
-benchmark_results.csv
-cmake-build-debug/
-.cxx/
-.gradle/
-local.properties
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@ -1,7 +1,6 @@
-cmake_minimum_required (VERSION 3.5)
+cmake_minimum_required (VERSION 3.0)

-project(whisper.cpp VERSION 1.5.4)
-set(SOVERSION 1)
+project(whisper.cpp VERSION 1.4.2)

 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@ -36,12 +35,6 @@ endif()

 # options

-if (APPLE)
-    set(WHISPER_METAL_DEFAULT ON)
-else()
-    set(WHISPER_METAL_DEFAULT OFF)
-endif()
-
 option(BUILD_SHARED_LIBS              "whisper: build shared libs" ${BUILD_SHARED_LIBS_DEFAULT})

 option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
@ -65,20 +58,15 @@ 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)
+    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)
 endif()

 option(WHISPER_PERF "whisper: enable perf timings" OFF)
@ -109,13 +97,6 @@ endif()

 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
@ -126,61 +107,9 @@ if (APPLE)
            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)
+            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_USE_ACCELERATE)
        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})
+            message(WARNING "Accelerate framework not found")
        endif()
    endif()

@ -193,7 +122,7 @@ if (APPLE)

            set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
        else()
-            message(FATAL_ERROR "CoreML framework not found")
+            message(WARNING "CoreML framework not found")
        endif()

        if (WHISPER_COREML_ALLOW_FALLBACK)
@ -216,13 +145,13 @@ if (WHISPER_BLAS)
            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.")
+            message(WARNING "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_PREFER_PKGCONFIG 1)
        set(BLA_SIZEOF_INTEGER 8)
-        set(BLA_PREFER_PKGCONFIG 1)
        find_package(BLAS)

        if(BLAS_FOUND)
@ -233,7 +162,7 @@ if (WHISPER_BLAS)
            include_directories(${BLAS_INCLUDE_DIRS})
            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} ${BLAS_LIBRARIES})
        else()
-            message(FATAL_ERROR "BLAS library was not found")
+            message(WARNING "BLAS library was not found")
        endif()
    endif ()
 endif ()
@ -248,24 +177,18 @@ if (WHISPER_CUBLAS)

        enable_language(CUDA)

-        set(GGML_SOURCES_CUDA ggml-cuda.cu ggml-cuda.h)
+        set(GGML_CUDA_SOURCES 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()
+            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
        else()
            set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
        endif()

-        set(WHISPER_EXTRA_LIBS ${WHISPER_EXTRA_LIBS} CUDA::cuda_driver)
    else()
-        message(FATAL_ERROR "cuBLAS not found")
+        message(WARNING "cuBLAS not found")
    endif()
 endif()

@ -296,7 +219,7 @@ if (WHISPER_HIPBLAS)
        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")
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
    endif()
 endif()

@ -305,13 +228,13 @@ if (WHISPER_CLBLAST)
    if (CLBlast_FOUND)
        message(STATUS "CLBlast found")

-        set(GGML_SOURCES_OPENCL ggml-opencl.cpp ggml-opencl.h)
+        set(GGML_OPENCL_SOURCES 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")
+        message(WARNING "CLBlast not found")
    endif()
 endif()

@ -319,30 +242,6 @@ 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)
@ -374,8 +273,7 @@ if (WHISPER_ALL_WARNINGS)
 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} -Werror=vla")
    #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -fno-math-errno -ffinite-math-only -funsafe-math-optimizations")
 endif()

@ -404,8 +302,8 @@ else()
        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")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
        else()
            if(NOT WHISPER_NO_AVX)
                set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mavx")
@ -423,53 +321,6 @@ else()
    endif()
 endif()

-#
-# POSIX conformance
-#
-
-# 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 ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
 endif()
@ -528,27 +379,12 @@ set(TARGET whisper)
 add_library(${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}
+    ${GGML_CUDA_SOURCES}
+    ${GGML_OPENCL_SOURCES}
    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
@ -572,7 +408,6 @@ else()
 endif()

 if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
    target_link_libraries(${TARGET} PUBLIC
        ${CMAKE_DL_LIBS}
        )
@ -586,23 +421,11 @@ if (BUILD_SHARED_LIBS)
        WHISPER_BUILD
        GGML_BUILD
        )
-
-    if (WHISPER_METAL)
-        # TODO: I think this should make ggml-metal.m "see" the ggml-metal.metal file from the "bin" directory
-        #       but for some reason it does not work here like it does in llama.cpp
-        set_target_properties(${TARGET} PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-    endif()
 endif()

-if (GGML_SOURCES_CUDA)
+if (GGML_CUDA_SOURCES)
    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_ARCHITECTURES OFF)
    set_property(TARGET whisper PROPERTY CUDA_SELECT_NVCC_ARCH_FLAGS "Auto")
 endif()

@ -614,15 +437,12 @@ target_compile_definitions(${TARGET} PUBLIC
    ${WHISPER_EXTRA_FLAGS}
    )

-set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "ggml.h;whisper.h")
-
-include(GNUInstallDirs)
+set_target_properties(${TARGET} PROPERTIES PUBLIC_HEADER "whisper.h")

 install(TARGETS ${TARGET}
-    LIBRARY  DESTINATION lib
-    ARCHIVE  DESTINATION lib/static
-    RUNTIME  DESTINATION bin
-    RESOURCE DESTINATION bin
+    LIBRARY DESTINATION lib
+    ARCHIVE DESTINATION lib/static
+    RUNTIME DESTINATION bin
    PUBLIC_HEADER DESTINATION include
    )

--- a/173
+++ b/173
@ -1,4 +1,4 @@
-default: main bench quantize server
+default: main bench quantize

 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
@ -18,7 +18,7 @@ ifndef NVCC_VERSION
 	endif
 endif

-CCV  := $(shell $(CC) --version | head -n 1)
+CCV := $(shell $(CC) --version | head -n 1)
 CXXV := $(shell $(CXX) --version | head -n 1)

 # Mac OS + Arm can report x86_64
@ -42,61 +42,18 @@ CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
 CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
 LDFLAGS  =

-ifdef MACOSX_DEPLOYMENT_TARGET
-	CFLAGS   += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-	CXXFLAGS += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-	LDFLAGS  += -mmacosx-version-min=$(MACOSX_DEPLOYMENT_TARGET)
-endif
-
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-CFLAGS   += -D_XOPEN_SOURCE=600
-CXXFLAGS += -D_XOPEN_SOURCE=600
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-ifeq ($(UNAME_S),OpenBSD)
-	CFLAGS   += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-	CXXFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-endif
-
-# Data types, macros and functions related to controlling CPU affinity
-# are available on Linux through GNU extensions in libc
-ifeq ($(UNAME_S),Linux)
-	CFLAGS   += -D_GNU_SOURCE
-	CXXFLAGS += -D_GNU_SOURCE
+# ref: https://github.com/ggerganov/whisper.cpp/issues/37
+ifneq ($(wildcard /usr/include/musl/*),)
+	CFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
+	CXXFLAGS += -D_POSIX_SOURCE -D_GNU_SOURCE
 endif

 # 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
 ifeq ($(UNAME_S),Darwin)
 	CFLAGS   += -D_DARWIN_C_SOURCE
 	CXXFLAGS += -D_DARWIN_C_SOURCE
 endif
-ifeq ($(UNAME_S),DragonFly)
-	CFLAGS   += -D__BSD_VISIBLE
-	CXXFLAGS += -D__BSD_VISIBLE
-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
-ifeq ($(UNAME_S),FreeBSD)
-	CFLAGS   += -D__BSD_VISIBLE
-	CXXFLAGS += -D__BSD_VISIBLE
-endif
-ifeq ($(UNAME_S),NetBSD)
-	CFLAGS   += -D_NETBSD_SOURCE
-	CXXFLAGS += -D_NETBSD_SOURCE
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	CFLAGS   += -D_BSD_SOURCE
-	CXXFLAGS += -D_BSD_SOURCE
-endif

 # OS specific
 # TODO: support Windows
@ -105,25 +62,15 @@ ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$
 	CXXFLAGS += -pthread
 endif

-# detect Windows
-ifneq ($(findstring _NT,$(UNAME_S)),)
-	_WIN32 := 1
-endif
-
-# Windows Sockets 2 (Winsock) for network-capable apps
-ifeq ($(_WIN32),1)
-	LWINSOCK2 := -lws2_32
-endif
-
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	ifeq ($(UNAME_S),Darwin)
-		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
+		CPUINFO_CMD := sysctl machdep.cpu.features
 	else ifeq ($(UNAME_S),Linux)
 		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter MINGW32_NT% MINGW64_NT% MSYS_NT%,$(UNAME_S)))
+	else ifneq (,$(filter MINGW32_NT% MINGW64_NT%,$(UNAME_S)))
 		CPUINFO_CMD := cat /proc/cpuinfo
 	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
 		CPUINFO_CMD := grep Features /var/run/dmesg.boot
@ -185,8 +132,6 @@ ifndef WHISPER_NO_ACCELERATE
 	# Mac M1 - include Accelerate framework
 	ifeq ($(UNAME_S),Darwin)
 		CFLAGS  += -DGGML_USE_ACCELERATE
-		CFLAGS  += -DACCELERATE_NEW_LAPACK
-		CFLAGS  += -DACCELERATE_LAPACK_ILP64
 		LDFLAGS += -framework Accelerate
 	endif
 endif
@ -200,16 +145,6 @@ ifdef WHISPER_COREML_ALLOW_FALLBACK
 endif
 endif

-ifndef WHISPER_NO_METAL
-	ifeq ($(UNAME_S),Darwin)
-		WHISPER_METAL := 1
-
-		CFLAGS   += -DGGML_USE_METAL
-		CXXFLAGS += -DGGML_USE_METAL
-		LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	endif
-endif
-
 ifdef WHISPER_OPENBLAS
 	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	LDFLAGS += -lopenblas
@ -217,14 +152,14 @@ endif

 ifdef WHISPER_CUBLAS
 	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
-		CUDA_ARCH_FLAG ?= native
+		CUDA_ARCH_FLAG=native
 	else
-		CUDA_ARCH_FLAG ?= all
+		CUDA_ARCH_FLAG=all
 	endif

 	CFLAGS      += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 	CXXFLAGS    += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	LDFLAGS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	LDFLAGS     += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib
 	WHISPER_OBJ += ggml-cuda.o
 	NVCC        = nvcc
 	NVCCFLAGS   = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
@ -316,17 +251,6 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@

-ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-WHISPER_OBJ += ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o
-
 whisper.o: whisper.cpp whisper.h ggml.h ggml-cuda.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@

@ -342,39 +266,14 @@ whisper-encoder-impl.o: coreml/whisper-encoder-impl.m coreml/whisper-encoder-imp
 WHISPER_OBJ += whisper.o whisper-encoder.o whisper-encoder-impl.o
 endif

-ifdef WHISPER_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h
-	$(CC) $(CFLAGS) -c $< -o $@
+libwhisper.a: ggml.o $(WHISPER_OBJ)
+	$(AR) rcs libwhisper.a ggml.o $(WHISPER_OBJ)

-WHISPER_OBJ += ggml-metal.o
-
-ifdef WHISPER_METAL_EMBED_LIBRARY
-CFLAGS += -DGGML_METAL_EMBED_LIBRARY
-
-ggml-metal-embed.o: ggml-metal.metal
-	@echo "Embedding Metal library"
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-
-WHISPER_OBJ += ggml-metal-embed.o
-endif
-endif
-
-libwhisper.a: $(WHISPER_OBJ)
-	$(AR) rcs libwhisper.a $(WHISPER_OBJ)
-
-libwhisper.so: $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so $(WHISPER_OBJ) $(LDFLAGS)
+libwhisper.so: ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)

 clean:
-	rm -f *.o main stream command talk talk-llama bench quantize server lsp libwhisper.a libwhisper.so
+	rm -f *.o main stream command talk talk-llama bench quantize lsp libwhisper.a libwhisper.so

 #
 # Examples
@ -385,33 +284,30 @@ CC_SDL=`sdl2-config --cflags --libs`
 SRC_COMMON     = examples/common.cpp examples/common-ggml.cpp
 SRC_COMMON_SDL = examples/common-sdl.cpp

-main: examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o main $(LDFLAGS)
+main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o main $(LDFLAGS)
 	./main -h

-bench: examples/bench/bench.cpp $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp $(WHISPER_OBJ) -o bench $(LDFLAGS)
+bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)

-quantize: examples/quantize/quantize.cpp $(WHISPER_OBJ) $(SRC_COMMON)
-	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o quantize $(LDFLAGS)
+quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
+	$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)

-server: examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/server/server.cpp $(SRC_COMMON) $(WHISPER_OBJ) -o server $(LDFLAGS) $(LWINSOCK2)
+stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)

-stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)
+command: examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)

-command: examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/command/command.cpp examples/grammar-parser.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o command $(CC_SDL) $(LDFLAGS)
+lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)

-lsp: examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/lsp/lsp.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o lsp $(CC_SDL) $(LDFLAGS)
+talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)

-talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk/talk.cpp examples/talk/gpt-2.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk $(CC_SDL) $(LDFLAGS)
-
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ)
-	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/unicode.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
+	$(CXX) $(CXXFLAGS) examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o talk-llama $(CC_SDL) $(LDFLAGS)

 #
 # Audio samples
@ -456,10 +352,9 @@ samples:
 .PHONY: medium.en
 .PHONY: medium
 .PHONY: large-v1
-.PHONY: large-v2
-.PHONY: large-v3
+.PHONY: large

-tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large: main
 	bash ./models/download-ggml-model.sh $@
 	@echo ""
 	@echo "==============================================="
--- a/Package.swift
+++ b/Package.swift
@ -1,61 +0,0 @@
-// swift-tools-version:5.5
-
-import PackageDescription
-
-let package = Package(
-    name: "whisper",
-    platforms: [
-        .macOS(.v12),
-        .iOS(.v14),
-        .watchOS(.v4),
-        .tvOS(.v14)
-    ],
-    products: [
-        .library(name: "whisper", targets: ["whisper"]),
-    ],
-    targets: [
-        .target(
-            name: "whisper",
-            path: ".",
-            exclude: [
-               "bindings",
-               "cmake",
-               "coreml",
-               "examples",
-               "extra",
-               "models",
-               "samples",
-               "tests",
-               "CMakeLists.txt",
-               "ggml-cuda.cu",
-               "ggml-cuda.h",
-               "Makefile"
-            ],
-            sources: [
-                "ggml.c",
-                "whisper.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m"
-            ],
-            resources: [.process("ggml-metal.metal")],
-            publicHeadersPath: "spm-headers",
-            cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL")
-                // NOTE: NEW_LAPACK will required iOS version 16.4+
-                // We should consider add this in the future when we drop support for iOS 14
-                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-                // .define("ACCELERATE_NEW_LAPACK"),
-                // .define("ACCELERATE_LAPACK_ILP64")
-            ],
-            linkerSettings: [
-                .linkedFramework("Accelerate")
-            ]
-        )
-    ],
-    cxxLanguageStandard: .cxx11
-)
--- a/README.md
+++ b/README.md
@ -6,20 +6,22 @@
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)

-Stable: [v1.5.4](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.5.4) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Beta: [v1.4.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.4.2) / Stable: [v1.2.1](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.2.1) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)

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

 - Plain C/C++ implementation without dependencies
- Apple Silicon first-class citizen - optimized via ARM NEON, Accelerate framework, Metal and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
+- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate framework and [Core ML](https://github.com/ggerganov/whisper.cpp#core-ml-support)
 - AVX intrinsics support for x86 architectures
 - VSX intrinsics support for POWER architectures
 - Mixed F16 / F32 precision
 - [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
+- Low memory usage (Flash Attention)
 - Zero memory allocations at runtime
- Support for CPU-only inference
- [Efficient GPU support for NVIDIA](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
+- Runs on the CPU
+- [Partial GPU support for NVIDIA via cuBLAS](https://github.com/ggerganov/whisper.cpp#nvidia-gpu-support-via-cublas)
 - [Partial OpenCL GPU support via CLBlast](https://github.com/ggerganov/whisper.cpp#opencl-gpu-support-via-clblast)
+- [BLAS CPU support via OpenBLAS](https://github.com/ggerganov/whisper.cpp#blas-cpu-support-via-openblas)
 - [OpenVINO Support](https://github.com/ggerganov/whisper.cpp#openvino-support)
 - [C-style API](https://github.com/ggerganov/whisper.cpp/blob/master/whisper.h)

@ -33,10 +35,11 @@ Supported platforms:
 - [x] [WebAssembly](examples/whisper.wasm)
 - [x] Windows ([MSVC](https://github.com/ggerganov/whisper.cpp/blob/master/.github/workflows/build.yml#L117-L144) and [MinGW](https://github.com/ggerganov/whisper.cpp/issues/168)]
 - [x] [Raspberry Pi](https://github.com/ggerganov/whisper.cpp/discussions/166)
- [x] [docker](https://github.com/ggerganov/whisper.cpp/pkgs/container/whisper.cpp)

-The entire high-level implementation of the model is contained in [whisper.h](whisper.h) and [whisper.cpp](whisper.cpp).
-The rest of the code is part of the [`ggml`](https://github.com/ggerganov/ggml) machine learning library.
+The entire implementation of the model is contained in 2 source files:
+
+- Tensor operations: [ggml.h](ggml.h) / [ggml.c](ggml.c)
+- Transformer inference: [whisper.h](whisper.h) / [whisper.cpp](whisper.cpp)

 Having such a lightweight implementation of the model allows to easily integrate it in different platforms and applications.
 As an example, here is a video of running the model on an iPhone 13 device - fully offline, on-device: [whisper.objc](examples/whisper.objc)
@ -47,10 +50,6 @@ You can also easily make your own offline voice assistant application: [command]

 https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4

-On Apple Silicon, the inference runs fully on the GPU via Metal:
-
-https://github.com/ggerganov/whisper.cpp/assets/1991296/c82e8f86-60dc-49f2-b048-d2fdbd6b5225
-
 Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)

 ## Implementation details
@ -61,22 +60,22 @@ Or you can even run it straight in the browser: [talk.wasm](examples/talk.wasm)
 - Sample real-time audio transcription from the microphone is demonstrated in [stream.cpp](examples/stream)
 - Various other examples are available in the [examples](examples) folder

-The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.
+The tensor operators are optimized heavily for Apple silicon CPUs. Depending on the computation size, Arm Neon SIMD
+intrinsics or CBLAS Accelerate framework routines are used. The latter are especially effective for bigger sizes since
+the Accelerate framework utilizes the special-purpose AMX coprocessor available in modern Apple products.

 ## Quick start

-First clone the repository:
+First clone the repository.

-```bash
-git clone https://github.com/ggerganov/whisper.cpp.git
-```
-
-Then, download one of the Whisper [models](models/README.md) converted in [`ggml` format](#ggml-format). For example:
+Then, download one of the Whisper models converted in [ggml format](models). For example:

 ```bash
 bash ./models/download-ggml-model.sh base.en
 ```

+If you wish to convert the Whisper models to ggml format yourself, instructions are in [models/README.md](models/README.md).
+
 Now build the [main](examples/main) example and transcribe an audio file like this:

 ```bash
@ -91,7 +90,7 @@ make

 For a quick demo, simply run `make base.en`:

-```text
+```java
 $ make base.en

 cc  -I.              -O3 -std=c11   -pthread -DGGML_USE_ACCELERATE   -c ggml.c -o ggml.o
@ -110,39 +109,30 @@ options:
  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
  -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
  -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [5      ] beam size for beam search
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
  -tr,       --translate         [false  ] translate from source language to english
-  -di,       --diarize           [false  ] stereo audio diarization
  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
+  -di,       --diarize           [false  ] stereo audio diarization
  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
  -otxt,     --output-txt        [false  ] output result in a text file
  -ovtt,     --output-vtt        [false  ] output result in a vtt file
  -osrt,     --output-srt        [false  ] output result in a srt file
-  -olrc,     --output-lrc        [false  ] output result in a lrc file
  -owts,     --output-words      [false  ] output script for generating karaoke video
-  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
  -ocsv,     --output-csv        [false  ] output result in a CSV file
-  -oj,       --output-json       [false  ] output result in a JSON file
-  -ojf,      --output-json-full  [false  ] include more information in the JSON file
  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
  -ps,       --print-special     [false  ] print special tokens
  -pc,       --print-colors      [false  ] print colors
  -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -nt,       --no-timestamps     [true   ] do not print timestamps
  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
             --prompt PROMPT     [       ] initial prompt
  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
  -f FNAME,  --file FNAME        [       ] input WAV file path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of tokens
-  -ng,       --no-gpu            [false  ] disable GPU


 bash ./models/download-ggml-model.sh base.en
@ -207,7 +197,7 @@ For detailed usage instructions, run: `./main -h`
 Note that the [main](examples/main) example currently runs only with 16-bit WAV files, so make sure to convert your input before running the tool.
 For example, you can use `ffmpeg` like this:

-```bash
+```java
 ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le output.wav
 ```

@ -233,19 +223,18 @@ make small
 make medium.en
 make medium
 make large-v1
-make large-v2
-make large-v3
+make large
 ```

 ## Memory usage

-| Model  | Disk    | Mem     |
-| ------ | ------- | ------- |
-| tiny   | 75 MiB  | ~273 MB |
-| base   | 142 MiB | ~388 MB |
-| small  | 466 MiB | ~852 MB |
-| medium | 1.5 GiB | ~2.1 GB |
-| large  | 2.9 GiB | ~3.9 GB |
+| Model  | Disk   | Mem     | SHA                                        |
+| ---    | ---    | ---     | ---                                        |
+| tiny   |  75 MB | ~125 MB | `bd577a113a864445d4c299885e0cb97d4ba92b5f` |
+| base   | 142 MB | ~210 MB | `465707469ff3a37a2b9b8d8f89f2f99de7299dac` |
+| small  | 466 MB | ~600 MB | `55356645c2b361a969dfd0ef2c5a50d530afd8d5` |
+| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
+| large  | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |

 ## Quantization

@ -278,8 +267,7 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in

  - To ensure `coremltools` operates correctly, please confirm that [Xcode](https://developer.apple.com/xcode/) is installed and execute `xcode-select --install` to install the command-line tools.
  - Python 3.10 is recommended.
-  - MacOS Sonoma (version 14) or newer is recommended, as older versions of MacOS might experience issues with transcription hallucination.
-  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html) for this step:
+  - [OPTIONAL] It is recommended to utilize a Python version management system, such as [Miniconda](https://docs.conda.io/en/latest/miniconda.html)  for this step:
    - To create an environment, use: `conda create -n py310-whisper python=3.10 -y`
    - To activate the environment, use: `conda activate py310-whisper`

@ -305,8 +293,8 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in

 - Run the examples as usual. For example:

-  ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav

  ...

@ -334,23 +322,21 @@ This can result in significant speedup in encoder performance. Here are the inst
 - First, setup python virtual env. and install python dependencies. Python 3.10 is recommended.

  Windows:
-
-  ```powershell
+  ```
  cd models
  python -m venv openvino_conv_env
  openvino_conv_env\Scripts\activate
  python -m pip install --upgrade pip
-  pip install -r requirements-openvino.txt
+  pip install -r openvino-conversion-requirements.txt
  ```

  Linux and macOS:
-
-  ```bash
+  ```
  cd models
  python3 -m venv openvino_conv_env
  source openvino_conv_env/bin/activate
  python -m pip install --upgrade pip
-  pip install -r requirements-openvino.txt
+  pip install -r openvino-conversion-requirements.txt
  ```

 - Generate an OpenVINO encoder model. For example, to generate a `base.en` model, use:
@ -359,7 +345,7 @@ This can result in significant speedup in encoder performance. Here are the inst
  python convert-whisper-to-openvino.py --model base.en
  ```

-  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as `ggml` models, as that
+  This will produce ggml-base.en-encoder-openvino.xml/.bin IR model files. It's recommended to relocate these to the same folder as ggml models, as that
  is the default location that the OpenVINO extension will search at runtime.

 - Build `whisper.cpp` with OpenVINO support:
@ -369,28 +355,24 @@ This can result in significant speedup in encoder performance. Here are the inst
  After downloading & extracting package onto your development system, set up required environment by sourcing setupvars script. For example:

  Linux:
-
  ```bash
  source /path/to/l_openvino_toolkit_ubuntu22_2023.0.0.10926.b4452d56304_x86_64/setupvars.sh
  ```

  Windows (cmd):
-
-  ```powershell
+  ```
  C:\Path\To\w_openvino_toolkit_windows_2023.0.0.10926.b4452d56304_x86_64\setupvars.bat
  ```

  And then build the project using cmake:
-
  ```bash
  cmake -B build -DWHISPER_OPENVINO=1
  cmake --build build -j --config Release
  ```

 - Run the examples as usual. For example:
-
-  ```text
-  $ ./main -m models/ggml-base.en.bin -f samples/jfk.wav
+  ```bash
+  ./main -m models/ggml-base.en.bin -f samples/jfk.wav

  ...

@ -406,12 +388,12 @@ This can result in significant speedup in encoder performance. Here are the inst

  The first time run on an OpenVINO device is slow, since the OpenVINO framework will compile the IR (Intermediate Representation) model to a device-specific 'blob'. This device-specific blob will get
  cached for the next run.
-
+  
 For more information about the Core ML implementation please refer to PR [#1037](https://github.com/ggerganov/whisper.cpp/pull/1037).

-## NVIDIA GPU support
+## NVIDIA GPU support via cuBLAS

-With NVIDIA cards the processing of the models is done efficiently on the GPU via cuBLAS and custom CUDA kernels.
+With NVIDIA cards the Encoder processing can to a large extent be offloaded to the GPU through cuBLAS.
 First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-downloads

 Now build `whisper.cpp` with cuBLAS support:
@ -441,6 +423,7 @@ cmake -B build -DWHISPER_CLBLAST=ON
 cmake --build build -j --config Release
 ```

+
 Run all the examples as usual.

 ## BLAS CPU support via OpenBLAS
@ -455,38 +438,6 @@ make clean
 WHISPER_OPENBLAS=1 make -j
 ```

-## Docker
-
-### Prerequisites
-
- Docker must be installed and running on your system.
- Create a folder to store big models & intermediate files (ex. /whisper/models)
-
-### Images
-
-We have two Docker images available for this project:
-
-1. `ghcr.io/ggerganov/whisper.cpp:main`: This image includes the main executable file as well as `curl` and `ffmpeg`. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/whisper.cpp:main-cuda`: Same as `main` but compiled with CUDA support. (platforms: `linux/amd64`)
-
-### Usage
-
-```shell
-# download model and persist it in a local folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  whisper.cpp:main "./models/download-ggml-model.sh base /models"
-# transcribe an audio file
-docker run -it --rm \
-  -v path/to/models:/models \
-  -v path/to/audios:/audios \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f /audios/jfk.wav"
-# transcribe an audio file in samples folder
-docker run -it --rm \
-  -v path/to/models:/models \
-  whisper.cpp:main "./main -m /models/ggml-base.bin -f ./samples/jfk.wav"
-```
-
 ## Limitations

 - Inference only
@ -499,7 +450,7 @@ in about half a minute on a MacBook M1 Pro, using `medium.en` model:
 <details>
  <summary>Expand to see the result</summary>

-```text
+```java
 $ ./main -m models/ggml-medium.en.bin -f samples/gb1.wav -t 8

 whisper_init_from_file: loading model from 'models/ggml-medium.en.bin'
@ -571,7 +522,6 @@ whisper_print_timings:   encode time = 18665.10 ms /     9 runs ( 2073.90 ms per
 whisper_print_timings:   decode time = 13090.93 ms /   549 runs (   23.85 ms per run)
 whisper_print_timings:    total time = 32733.52 ms
 ```
-
 </details>

 ## Real-time audio input example
@ -580,7 +530,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The [stream](examples/stream) tool samples the audio every half a second and runs the transcription continuously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).

-```bash
+```java
 make stream
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
@ -592,7 +542,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Adding the `--print-colors` argument will print the transcribed text using an experimental color coding strategy
 to highlight words with high or low confidence:

-```bash
+```java
 ./main -m models/ggml-base.en.bin -f samples/gb0.wav --print-colors
 ```

@ -602,8 +552,8 @@ to highlight words with high or low confidence:

 For example, to limit the line length to a maximum of 16 characters, simply add `-ml 16`:

-```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16
+```java
+./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 16

 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@ -626,8 +576,8 @@ main: processing './samples/jfk.wav' (176000 samples, 11.0 sec), 4 threads, 1 pr

 The `--max-len` argument can be used to obtain word-level timestamps. Simply use `-ml 1`:

-```text
-$ ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1
+```java
+./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -ml 1

 whisper_model_load: loading model from './models/ggml-base.en.bin'
 ...
@ -697,7 +647,7 @@ This requires to have `ffmpeg` installed.

 Here are a few *"typical"* examples:

-```bash
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/jfk.wav -owts
 source ./samples/jfk.wav.wts
 ffplay ./samples/jfk.wav.mp4
@ -707,7 +657,7 @@ https://user-images.githubusercontent.com/1991296/199337465-dbee4b5e-9aeb-48a3-b

 ---

-```bash
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/mm0.wav -owts
 source ./samples/mm0.wav.wts
 ffplay ./samples/mm0.wav.mp4
@ -717,7 +667,7 @@ https://user-images.githubusercontent.com/1991296/199337504-cc8fd233-0cb7-4920-9

 ---

-```bash
+```java
 ./main -m ./models/ggml-base.en.bin -f ./samples/gb0.wav -owts
 source ./samples/gb0.wav.wts
 ffplay ./samples/gb0.wav.mp4
@ -731,7 +681,7 @@ https://user-images.githubusercontent.com/1991296/199337538-b7b0c7a3-2753-4a88-a

 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
+```java
 ./extra/bench-wts.sh samples/jfk.wav
 ffplay ./samples/jfk.wav.all.mp4
 ```
@ -748,19 +698,7 @@ took to execute it. The results are summarized in the following Github issue:

 [Benchmark results](https://github.com/ggerganov/whisper.cpp/issues/89)

-Additionally a script to run whisper.cpp with different models and audio files is provided [bench.py](bench.py).
-
-You can run it with the following command, by default it will run against any standard model in the models folder.
-
-```bash
-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.
-
-It outputs a csv file with the results of the benchmarking.
-
-## `ggml` format
+## ggml format

 The original models are converted to a custom binary format. This allows to pack everything needed into a single file:

@ -775,50 +713,49 @@ or manually from here:
 - https://huggingface.co/ggerganov/whisper.cpp
 - https://ggml.ggerganov.com

-For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or [models/README.md](models/README.md).
+For more details, see the conversion script [models/convert-pt-to-ggml.py](models/convert-pt-to-ggml.py) or the README
+in [models](models).

 ## [Bindings](https://github.com/ggerganov/whisper.cpp/discussions/categories/bindings)

- [x] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
- [x] JavaScript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
+- [X] Rust: [tazz4843/whisper-rs](https://github.com/tazz4843/whisper-rs) | [#310](https://github.com/ggerganov/whisper.cpp/discussions/310)
+- [X] Javascript: [bindings/javascript](bindings/javascript) | [#309](https://github.com/ggerganov/whisper.cpp/discussions/309)
  - React Native (iOS / Android): [whisper.rn](https://github.com/mybigday/whisper.rn)
- [x] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
- [x] Java:
+- [X] Go: [bindings/go](bindings/go) | [#312](https://github.com/ggerganov/whisper.cpp/discussions/312)
+- [X] Java:
  - [GiviMAD/whisper-jni](https://github.com/GiviMAD/whisper-jni)
- [x] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
- [x] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
+- [X] Ruby: [bindings/ruby](bindings/ruby) | [#507](https://github.com/ggerganov/whisper.cpp/discussions/507)
+- [X] Objective-C / Swift: [ggerganov/whisper.spm](https://github.com/ggerganov/whisper.spm) | [#313](https://github.com/ggerganov/whisper.cpp/discussions/313)
  - [exPHAT/SwiftWhisper](https://github.com/exPHAT/SwiftWhisper)
- [x] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
+- [X] .NET: | [#422](https://github.com/ggerganov/whisper.cpp/discussions/422)
  - [sandrohanea/whisper.net](https://github.com/sandrohanea/whisper.net)
  - [NickDarvey/whisper](https://github.com/NickDarvey/whisper)
- [x] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
+- [X] Python: | [#9](https://github.com/ggerganov/whisper.cpp/issues/9)
  - [stlukey/whispercpp.py](https://github.com/stlukey/whispercpp.py) (Cython)
  - [aarnphm/whispercpp](https://github.com/aarnphm/whispercpp) (Pybind11)
- [x] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
- [x] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)
+- [X] R: [bnosac/audio.whisper](https://github.com/bnosac/audio.whisper)
+- [X] Unity: [macoron/whisper.unity](https://github.com/Macoron/whisper.unity)

 ## Examples

 There are various examples of using the library for different projects in the [examples](examples) folder.
 Some of the examples are even ported to run in the browser using WebAssembly. Check them out!

-| Example                                             | Web                                   | Description                                                                                                                     |
-| --------------------------------------------------- | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------- |
-| [main](examples/main)                               | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper                                                                       |
-| [bench](examples/bench)                             | [bench.wasm](examples/bench.wasm)     | Benchmark the performance of Whisper on your machine                                                                            |
-| [stream](examples/stream)                           | [stream.wasm](examples/stream.wasm)   | Real-time transcription of raw microphone capture                                                                               |
-| [command](examples/command)                         | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic                                                         |
-| [wchess](examples/wchess)                           | [wchess.wasm](examples/wchess)        | Voice-controlled chess                                                                                                          |
-| [talk](examples/talk)                               | [talk.wasm](examples/talk.wasm)       | Talk with a GPT-2 bot                                                                                                           |
-| [talk-llama](examples/talk-llama)                   |                                       | Talk with a LLaMA bot                                                                                                           |
-| [whisper.objc](examples/whisper.objc)               |                                       | iOS mobile application using whisper.cpp                                                                                        |
-| [whisper.swiftui](examples/whisper.swiftui)         |                                       | SwiftUI iOS / macOS application using whisper.cpp                                                                               |
-| [whisper.android](examples/whisper.android)         |                                       | Android mobile application using whisper.cpp                                                                                    |
-| [whisper.nvim](examples/whisper.nvim)               |                                       | Speech-to-text plugin for Neovim                                                                                                |
-| [generate-karaoke.sh](examples/generate-karaoke.sh) |                                       | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture                           |
-| [livestream.sh](examples/livestream.sh)             |                                       | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185)                                           |
-| [yt-wsp.sh](examples/yt-wsp.sh)                     |                                       | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |
-| [server](examples/server)                           |                                       | HTTP transcription server with OAI-like API                                                                                     |
+| Example | Web | Description |
+| ---     | --- | ---         |
+| [main](examples/main) | [whisper.wasm](examples/whisper.wasm) | Tool for translating and transcribing audio using Whisper |
+| [bench](examples/bench) | [bench.wasm](examples/bench.wasm) | Benchmark the performance of Whisper on your machine |
+| [stream](examples/stream) | [stream.wasm](examples/stream.wasm) | Real-time transcription of raw microphone capture |
+| [command](examples/command) | [command.wasm](examples/command.wasm) | Basic voice assistant example for receiving voice commands from the mic |
+| [talk](examples/talk) | [talk.wasm](examples/talk.wasm) | Talk with a GPT-2 bot |
+| [talk-llama](examples/talk-llama) | | Talk with a LLaMA bot |
+| [whisper.objc](examples/whisper.objc) | | iOS mobile application using whisper.cpp |
+| [whisper.swiftui](examples/whisper.swiftui) | | SwiftUI iOS / macOS application using whisper.cpp |
+| [whisper.android](examples/whisper.android) | | Android mobile application using whisper.cpp |
+| [whisper.nvim](examples/whisper.nvim) | | Speech-to-text plugin for Neovim |
+| [generate-karaoke.sh](examples/generate-karaoke.sh) | | Helper script to easily [generate a karaoke video](https://youtu.be/uj7hVta4blM) of raw audio capture |
+| [livestream.sh](examples/livestream.sh) | | [Livestream audio transcription](https://github.com/ggerganov/whisper.cpp/issues/185) |
+| [yt-wsp.sh](examples/yt-wsp.sh) | | Download + transcribe and/or translate any VOD [(original)](https://gist.github.com/DaniruKun/96f763ec1a037cc92fe1a059b643b818) |

 ## [Discussions](https://github.com/ggerganov/whisper.cpp/discussions)

--- a/README_sycl.md
+++ b/README_sycl.md
@ -1,249 +0,0 @@
-# whisper.cpp for SYCL
-
-[Background](#background)
-
-[OS](#os)
-
-[Intel GPU](#intel-gpu)
-
-[Linux](#linux)
-
-[Environment Variable](#environment-variable)
-
-[Known Issue](#known-issue)
-
-[Todo](#todo)
-
-## Background
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators<72>such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
-
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
-
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
-
-To avoid  re-inventing the wheel, this code refers other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel<EFBFBD> DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
-
-The whisper.cpp for SYCL is used to support Intel GPUs.
-
-For Intel CPU, recommend to use whisper.cpp for X86 (Intel MKL build).
-
-## OS
-
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04|
-|Windows|Ongoing| |
-
-
-## Intel GPU
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
-
-
-## Linux
-
-### Setup Environment
-
-1. Install Intel GPU driver.
-
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
-
-Note: for iGPU, please install the client GPU driver.
-
-b. Add user to group: video, render.
-
-```
-sudo usermod -aG render username
-sudo usermod -aG video username
-```
-
-Note: re-login to enable it.
-
-c. Check
-
-```
-sudo apt install clinfo
-sudo clinfo -l
-```
-
-Output (example):
-
-```
-Platform #0: Intel(R) OpenCL Graphics
- `-- Device #0: Intel(R) Arc(TM) A770 Graphics
-
-
-Platform #0: Intel(R) OpenCL HD Graphics
- `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
-```
-
-2. Install Intel<65> oneAPI Base toolkit.
-
-
-a. Please follow the procedure in [Get the Intel<65> oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
-
-Recommend to install to default folder: **/opt/intel/oneapi**.
-
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
-
-b. Check
-
-```
-source /opt/intel/oneapi/setvars.sh
-
-sycl-ls
-```
-
-There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
-
-Output (example):
-```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
-
-```
-
-2. Build locally:
-
-```
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON 
-
-#for FP32
-cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v
-
-```
-
-or
-
-```
-./examples/sycl/build.sh
-```
-
-Note:
-
- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
-
-### Run
-
-1. Put model file to folder **models**
-
-2. Enable oneAPI running environment
-
-```
-source /opt/intel/oneapi/setvars.sh
-```
-
-3. List device ID
-
-Run without parameter:
-
-```
-./build/bin/ls-sycl-device
-
-or
-
-./build/bin/main
-```
-
-Check the ID in startup log, like:
-
-```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
-```
-
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
-
-4. Set device ID and execute whisper.cpp
-
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
-
-```
-GGML_SYCL_DEVICE=0 ./build/bin/main -m models/ggml-base.en.bin -f samples/jfk.wav
-```
-or run by script:
-
-```
-./examples/sycl/run_whisper.sh
-```
-
-
-
-5. Check the device ID in output
-
-Like:
-```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
-```
-
-
-## Environment Variable
-
-#### Build
-
-|Name|Value|Function|
-|-|-|-|
-|WHISPER_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, WHISPER_SYCL=ON is mandatory.|
-|WHISPER_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path.For FP32, do not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
-
-#### Running
-
-
-|Name|Value|Function|
-|-|-|-|
-|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
-|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-
-## Known Issue
-
- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
-
-  Miss to enable oneAPI running environment.
-
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
-
-
- Hang during startup
-
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
-
-  Solution: add **--no-mmap**.
-
-## Todo
-
- Support to build in Windows.
-
- Support multiple cards.
--- a/bindings/go/Makefile
+++ b/bindings/go/Makefile
@ -1,26 +1,9 @@
-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
-GGML_METAL_PATH_RESOURCES := $(abspath ../..)
 BUILD_DIR := build
 MODELS_DIR := models
 EXAMPLES_DIR := $(wildcard examples/*)
 INCLUDE_PATH := $(abspath ../..)
 LIBRARY_PATH := $(abspath ../..)

-ifeq ($(UNAME_S),Darwin)
-	EXT_LDFLAGS := -framework Foundation -framework Metal -framework MetalKit
-endif
-
 all: clean whisper examples

 whisper: mkdir
@ -28,13 +11,8 @@ whisper: mkdir
 	@${MAKE} -C ../.. libwhisper.a

 test: model-small whisper modtidy
-ifeq ($(UNAME_S),Darwin)
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v .
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go test -ldflags "-extldflags '$(EXT_LDFLAGS)'" -v ./pkg/whisper/...
-else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v .
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go test -v ./pkg/whisper/...
-endif

 examples: $(EXAMPLES_DIR)

@ -43,11 +21,7 @@ model-small: mkdir examples/go-model-download

 $(EXAMPLES_DIR): mkdir whisper modtidy
 	@echo Build example $(notdir $@)
-ifeq ($(UNAME_S),Darwin)
-	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} GGML_METAL_PATH_RESOURCES=${GGML_METAL_PATH_RESOURCES} go build ${BUILD_FLAGS} -ldflags "-extldflags '$(EXT_LDFLAGS)'" -o ${BUILD_DIR}/$(notdir $@) ./$@
-else
 	@C_INCLUDE_PATH=${INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} go build ${BUILD_FLAGS} -o ${BUILD_DIR}/$(notdir $@) ./$@
-endif

 mkdir:
 	@echo Mkdir ${BUILD_DIR}
--- a/bindings/go/examples/go-model-download/main.go
+++ b/bindings/go/examples/go-model-download/main.go
@ -24,7 +24,7 @@ const (

 var (
 	// The models which will be downloaded, if no model is specified as an argument
-	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large-v2", "ggml-large-v3"}
+	modelNames = []string{"ggml-tiny.en", "ggml-tiny", "ggml-base.en", "ggml-base", "ggml-small.en", "ggml-small", "ggml-medium.en", "ggml-medium", "ggml-large-v1", "ggml-large"}
 )

 var (
--- a/bindings/go/params.go
+++ b/bindings/go/params.go
@ -118,16 +118,6 @@ func (p *Params) SetMaxTokensPerSegment(n int) {
 	p.max_tokens = C.int(n)
 }

-// Set audio encoder context
-func (p *Params) SetAudioCtx(n int) {
-	p.audio_ctx = C.int(n)
-}
-
-// Set initial prompt
-func (p *Params) SetInitialPrompt(prompt string) {
-	p.initial_prompt = C.CString(prompt)
-}
-
 ///////////////////////////////////////////////////////////////////////////////
 // PRIVATE METHODS

@ -151,8 +141,6 @@ func (p *Params) String() string {
 	str += fmt.Sprintf(" n_max_text_ctx=%d", p.n_max_text_ctx)
 	str += fmt.Sprintf(" offset_ms=%d", p.offset_ms)
 	str += fmt.Sprintf(" duration_ms=%d", p.duration_ms)
-	str += fmt.Sprintf(" audio_ctx=%d", p.audio_ctx)
-	str += fmt.Sprintf(" initial_prompt=%s", C.GoString(p.initial_prompt))
 	if p.translate {
 		str += " translate"
 	}
--- a/bindings/go/pkg/whisper/context.go
+++ b/bindings/go/pkg/whisper/context.go
@ -82,7 +82,7 @@ func (context *context) SetSpeedup(v bool) {
 }

 func (context *context) SetSplitOnWord(v bool) {
-	context.params.SetSplitOnWord(v)
+        context.params.SetSplitOnWord(v)
 }

 // Set number of threads to use
@ -125,16 +125,6 @@ func (context *context) SetMaxTokensPerSegment(n uint) {
 	context.params.SetMaxTokensPerSegment(int(n))
 }

-// Set audio encoder context
-func (context *context) SetAudioCtx(n uint) {
-	context.params.SetAudioCtx(int(n))
-}
-
-// Set initial prompt
-func (context *context) SetInitialPrompt(prompt string) {
-	context.params.SetInitialPrompt(prompt)
-}
-
 // ResetTimings resets the mode timings. Should be called before processing
 func (context *context) ResetTimings() {
 	context.model.ctx.Whisper_reset_timings()
--- a/bindings/go/pkg/whisper/interface.go
+++ b/bindings/go/pkg/whisper/interface.go
@ -38,18 +38,16 @@ type Context interface {
 	IsMultilingual() bool     // Return true if the model is multilingual.
 	Language() string         // Get language

-	SetOffset(time.Duration)        // Set offset
-	SetDuration(time.Duration)      // Set duration
-	SetThreads(uint)                // Set number of threads to use
-	SetSpeedup(bool)                // Set speedup flag
-	SetSplitOnWord(bool)            // Set split on word flag
-	SetTokenThreshold(float32)      // Set timestamp token probability threshold
-	SetTokenSumThreshold(float32)   // Set timestamp token sum probability threshold
-	SetMaxSegmentLength(uint)       // Set max segment length in characters
-	SetTokenTimestamps(bool)        // Set token timestamps flag
-	SetMaxTokensPerSegment(uint)    // Set max tokens per segment (0 = no limit)
-	SetAudioCtx(uint)               // Set audio encoder context
-	SetInitialPrompt(prompt string) // Set initial prompt
+	SetOffset(time.Duration)      // Set offset
+	SetDuration(time.Duration)    // Set duration
+	SetThreads(uint)              // Set number of threads to use
+	SetSpeedup(bool)              // Set speedup flag
+	SetSplitOnWord(bool)          // Set split on word flag
+	SetTokenThreshold(float32)    // Set timestamp token probability threshold
+	SetTokenSumThreshold(float32) // Set timestamp token sum probability threshold
+	SetMaxSegmentLength(uint)     // Set max segment length in characters
+	SetTokenTimestamps(bool)      // Set token timestamps flag
+	SetMaxTokensPerSegment(uint)  // Set max tokens per segment (0 = no limit)

 	// Process mono audio data and return any errors.
 	// If defined, newly generated segments are passed to the
--- a/bindings/go/whisper.go
+++ b/bindings/go/whisper.go
@ -10,7 +10,7 @@ import (

 /*
 #cgo LDFLAGS: -lwhisper -lm -lstdc++
-#cgo darwin LDFLAGS: -framework Accelerate -framework Metal -framework Foundation -framework CoreGraphics
+#cgo darwin LDFLAGS: -framework Accelerate
 #include <whisper.h>
 #include <stdlib.h>

@ -83,6 +83,7 @@ const (
 	SampleRate = C.WHISPER_SAMPLE_RATE                 // Expected sample rate, samples per second
 	SampleBits = uint16(unsafe.Sizeof(C.float(0))) * 8 // Sample size in bits
 	NumFFT     = C.WHISPER_N_FFT
+	NumMEL     = C.WHISPER_N_MEL
 	HopLength  = C.WHISPER_HOP_LENGTH
 	ChunkSize  = C.WHISPER_CHUNK_SIZE
 )
@ -102,7 +103,7 @@ var (
 func Whisper_init(path string) *Context {
 	cPath := C.CString(path)
 	defer C.free(unsafe.Pointer(cPath))
-	if ctx := C.whisper_init_from_file_with_params(cPath, C.whisper_context_default_params()); ctx != nil {
+	if ctx := C.whisper_init_from_file(cPath); ctx != nil {
 		return (*Context)(ctx)
 	} else {
 		return nil
--- a/bindings/ios
+++ b/bindings/ios
--- a/bindings/java/build.gradle
+++ b/bindings/java/build.gradle
@ -2,14 +2,12 @@ plugins {
    id 'java'
    id 'java-library'
    id 'maven-publish'
-    id 'signing'
 }

 archivesBaseName = 'whispercpp'
 group = 'io.github.ggerganov'
 version = '1.4.0'

-
 sourceCompatibility = 1.8
 targetCompatibility = 1.8

@ -111,23 +109,4 @@ publishing {
            }
        }
    }
-
-    repositories {
-        maven {
-            def releasesRepoUrl = 'https://s01.oss.sonatype.org/service/local/staging/deploy/maven2/'
-            def snapshotsRepoUrl = 'https://s01.oss.sonatype.org/content/repositories/snapshots/'
-            url = version.endsWith('-SNAPSHOT') ? snapshotsRepoUrl : releasesRepoUrl
-            credentials {
-                username = System.getenv("MAVEN_USERNAME")
-                password = System.getenv("MAVEN_PASSWORD")
-            }
-        }
-    }
-}
-
-signing {
-    def signingKey = System.getenv("PGP_SECRET")
-    def signingPassword = System.getenv("PGP_PASSPHRASE")
-    useInMemoryPgpKeys(signingKey, signingPassword)
-    sign publishing.publications.mavenJava
 }
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperContext.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperContext.java
@ -4,7 +4,6 @@ import com.sun.jna.Structure;
 import com.sun.jna.ptr.PointerByReference;
 import io.github.ggerganov.whispercpp.ggml.GgmlType;
 import io.github.ggerganov.whispercpp.WhisperModel;
-import io.github.ggerganov.whispercpp.params.WhisperContextParams;

 import java.util.List;

@ -24,9 +23,8 @@ public class WhisperContext extends Structure {
    public PointerByReference vocab;
    public PointerByReference state;

-    /** populated by whisper_init_from_file_with_params() */
+    /** populated by whisper_init_from_file() */
    String path_model;
-    WhisperContextParams params;

 //    public static class ByReference extends WhisperContext implements Structure.ByReference {
 //    }
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCpp.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCpp.java
@ -2,16 +2,12 @@ package io.github.ggerganov.whispercpp;

 import com.sun.jna.Native;
 import com.sun.jna.Pointer;
-import io.github.ggerganov.whispercpp.bean.WhisperSegment;
-import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;

 import java.io.File;
 import java.io.FileNotFoundException;
 import java.io.IOException;
-import java.util.ArrayList;
-import java.util.List;

 /**
 * Before calling most methods, you must call `initContext(modelPath)` to initialise the `ctx` Pointer.
@ -19,9 +15,8 @@ import java.util.List;
 public class WhisperCpp implements AutoCloseable {
    private WhisperCppJnaLibrary lib = WhisperCppJnaLibrary.instance;
    private Pointer ctx = null;
-    private Pointer paramsPointer = null;
-    private Pointer greedyParamsPointer = null;
-    private Pointer beamParamsPointer = null;
+    private Pointer greedyPointer = null;
+    private Pointer beamPointer = null;

    public File modelDir() {
        String modelDirPath = System.getenv("XDG_CACHE_HOME");
@ -36,18 +31,6 @@ public class WhisperCpp implements AutoCloseable {
     * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
     */
    public void initContext(String modelPath) throws FileNotFoundException {
-        initContextImpl(modelPath, getContextDefaultParams());
-    }
-
-    /**
-     * @param modelPath - absolute path, or just the name (eg: "base", "base-en" or "base.en")
-     * @param params - params to use when initialising the context
-     */
-    public void initContext(String modelPath, WhisperContextParams params) throws FileNotFoundException {
-        initContextImpl(modelPath, params);
-    }
-
-    private void initContextImpl(String modelPath, WhisperContextParams params) throws FileNotFoundException {
        if (ctx != null) {
            lib.whisper_free(ctx);
        }
@ -60,26 +43,13 @@ public class WhisperCpp implements AutoCloseable {
            modelPath = new File(modelDir(), modelPath).getAbsolutePath();
        }

-        ctx = lib.whisper_init_from_file_with_params(modelPath, params);
+        ctx = lib.whisper_init_from_file(modelPath);

        if (ctx == null) {
            throw new FileNotFoundException(modelPath);
        }
    }

-    /**
-     * 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:
-     * - call `whisper_free_context_params()`
-     * - `Native.free(Pointer.nativeValue(pointer));`
-     */
-    public WhisperContextParams getContextDefaultParams() {
-        paramsPointer = lib.whisper_context_default_params_by_ref();
-        WhisperContextParams params = new WhisperContextParams(paramsPointer);
-        params.read();
-        return params;
-    }
-    
    /**
     * Provides default params which can be used with `whisper_full()` etc.
     * Because this function allocates memory for the params, the caller must call either:
@ -93,15 +63,15 @@ public class WhisperCpp implements AutoCloseable {

        // whisper_full_default_params_by_ref allocates memory which we need to delete, so only create max 1 pointer for each strategy.
        if (strategy == WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY) {
-            if (greedyParamsPointer == null) {
-                greedyParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            if (greedyPointer == null) {
+                greedyPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
            }
-            pointer = greedyParamsPointer;
+            pointer = greedyPointer;
        } else {
-            if (beamParamsPointer == null) {
-                beamParamsPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
+            if (beamPointer == null) {
+                beamPointer = lib.whisper_full_default_params_by_ref(strategy.ordinal());
            }
-            pointer = beamParamsPointer;
+            pointer = beamPointer;
        }

        WhisperFullParams params = new WhisperFullParams(pointer);
@ -123,17 +93,13 @@ public class WhisperCpp implements AutoCloseable {
    }

    private void freeParams() {
-        if (paramsPointer != null) {
-            Native.free(Pointer.nativeValue(paramsPointer));
-            paramsPointer = null;
+        if (greedyPointer != null) {
+            Native.free(Pointer.nativeValue(greedyPointer));
+            greedyPointer = null;
        }
-        if (greedyParamsPointer != null) {
-            Native.free(Pointer.nativeValue(greedyParamsPointer));
-            greedyParamsPointer = null;
-        }
-        if (beamParamsPointer != null) {
-            Native.free(Pointer.nativeValue(beamParamsPointer));
-            beamParamsPointer = null;
+        if (beamPointer != null) {
+            Native.free(Pointer.nativeValue(beamPointer));
+            beamPointer = null;
        }
    }

@ -163,28 +129,6 @@ public class WhisperCpp implements AutoCloseable {

        return str.toString().trim();
    }
-    public List<WhisperSegment> fullTranscribeWithTime(WhisperFullParams whisperParams, float[] audioData) throws IOException {
-        if (ctx == null) {
-            throw new IllegalStateException("Model not initialised");
-        }
-
-        if (lib.whisper_full(ctx, whisperParams, audioData, audioData.length) != 0) {
-            throw new IOException("Failed to process audio");
-        }
-
-        int nSegments = lib.whisper_full_n_segments(ctx);
-        List<WhisperSegment> segments= new ArrayList<>(nSegments);
-
-
-        for (int i = 0; i < nSegments; i++) {
-            long t0 = lib.whisper_full_get_segment_t0(ctx, i);
-            String text = lib.whisper_full_get_segment_text(ctx, i);
-            long t1 = lib.whisper_full_get_segment_t1(ctx, i);
-            segments.add(new WhisperSegment(t0,t1,text));
-        }
-
-        return segments;
-    }

 //    public int getTextSegmentCount(Pointer ctx) {
 //        return lib.whisper_full_n_segments(ctx);
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/WhisperCppJnaLibrary.java
@ -5,7 +5,6 @@ import com.sun.jna.Native;
 import com.sun.jna.Pointer;
 import io.github.ggerganov.whispercpp.model.WhisperModelLoader;
 import io.github.ggerganov.whispercpp.model.WhisperTokenData;
-import io.github.ggerganov.whispercpp.params.WhisperContextParams;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;

 public interface WhisperCppJnaLibrary extends Library {
@ -14,31 +13,12 @@ public interface WhisperCppJnaLibrary extends Library {
    String whisper_print_system_info();

    /**
-     * DEPRECATED. Allocate (almost) all memory needed for the model by loading from a file.
+     * Allocate (almost) all memory needed for the model by loading from a file.
     *
     * @param path_model Path to the model file
     * @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:
-     * - call `whisper_free_context_params()`
-     * - `Native.free(Pointer.nativeValue(pointer));`
-     */
-    Pointer whisper_context_default_params_by_ref();
-
-    void whisper_free_context_params(Pointer params);
-
-    /**
-     * Allocate (almost) all memory needed for the model by loading from a file.
-     *
-     * @param path_model Path to the model file
-     * @param params     Pointer to whisper_context_params
-     * @return Whisper context on success, null on failure
-     */
-    Pointer whisper_init_from_file_with_params(String path_model, WhisperContextParams params);

    /**
     * Allocate (almost) all memory needed for the model by loading from a buffer.
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/bean/WhisperSegment.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/bean/WhisperSegment.java
@ -1,47 +0,0 @@
-package io.github.ggerganov.whispercpp.bean;
-
-/**
- * Created by litonglinux@qq.com on 10/21/2023_7:48 AM
- */
-public class WhisperSegment {
-  private long start, end;
-  private String sentence;
-
-  public WhisperSegment() {
-  }
-
-  public WhisperSegment(long start, long end, String sentence) {
-    this.start = start;
-    this.end = end;
-    this.sentence = sentence;
-  }
-
-  public long getStart() {
-    return start;
-  }
-
-  public long getEnd() {
-    return end;
-  }
-
-  public String getSentence() {
-    return sentence;
-  }
-
-  public void setStart(long start) {
-    this.start = start;
-  }
-
-  public void setEnd(long end) {
-    this.end = end;
-  }
-
-  public void setSentence(String sentence) {
-    this.sentence = sentence;
-  }
-
-  @Override
-  public String toString() {
-    return "[" + start + " --> " + end + "]:" + sentence;
-  }
-}
--- a/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperContextParams.java
+++ b/bindings/java/src/main/java/io/github/ggerganov/whispercpp/params/WhisperContextParams.java
@ -1,31 +0,0 @@
-package io.github.ggerganov.whispercpp.params;
-
-import com.sun.jna.*;
-
-import java.util.Arrays;
-import java.util.List;
-
-/**
- * Parameters for the whisper_init_from_file_with_params() function.
- * If you change the order or add new parameters, make sure to update the default values in whisper.cpp:
- * whisper_context_default_params()
- */
-public class WhisperContextParams extends Structure {
-
-    public WhisperContextParams(Pointer p) {
-        super(p);
-    }
-
-    /** Use GPU for inference Number (default = true) */
-    public CBool use_gpu;
-
-    /** Use GPU for inference Number (default = true) */
-    public void useGpu(boolean enable) {
-        use_gpu = enable ? CBool.TRUE : CBool.FALSE;
-    }
-
-    @Override
-    protected List<String> getFieldOrder() {
-        return Arrays.asList("use_gpu");
-    }
-}
--- 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
@ -58,9 +58,6 @@ public class WhisperFullParams extends Structure {
        no_context = enable ? CBool.FALSE : CBool.TRUE;
    }

-    /** Generate timestamps or not? */
-    public CBool no_timestamps;
-
    /** Flag to force single segment output (useful for streaming). (default = false) */
    public CBool single_segment;

@ -307,16 +304,10 @@ public class WhisperFullParams extends Structure {
        logits_filter_callback = CallbackReference.getFunctionPointer(callback);
    }

-    /** Grammar stuff */
-    public Pointer grammar_rules;
-    public long n_grammar_rules;
-    public long i_start_rule;
-    public float grammar_penalty;
-
    @Override
    protected List<String> getFieldOrder() {
        return Arrays.asList("strategy", "n_threads", "n_max_text_ctx", "offset_ms", "duration_ms", "translate",
-                "no_context", "single_segment", "no_timestamps",
+                "no_context", "single_segment",
                "print_special", "print_progress", "print_realtime", "print_timestamps",  "token_timestamps",
                "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",
@ -325,7 +316,6 @@ public class WhisperFullParams extends Structure {
                "new_segment_callback", "new_segment_callback_user_data",
                "progress_callback", "progress_callback_user_data",
                "encoder_begin_callback", "encoder_begin_callback_user_data",
-                "logits_filter_callback", "logits_filter_callback_user_data",
-                "grammar_rules", "n_grammar_rules", "i_start_rule", "grammar_penalty");
+                "logits_filter_callback", "logits_filter_callback_user_data");
    }
 }
--- a/bindings/java/src/test/java/io/github/ggerganov/whispercpp/WhisperCppTest.java
+++ b/bindings/java/src/test/java/io/github/ggerganov/whispercpp/WhisperCppTest.java
@ -2,7 +2,6 @@ package io.github.ggerganov.whispercpp;

 import static org.junit.jupiter.api.Assertions.*;

-import io.github.ggerganov.whispercpp.bean.WhisperSegment;
 import io.github.ggerganov.whispercpp.params.CBool;
 import io.github.ggerganov.whispercpp.params.WhisperFullParams;
 import io.github.ggerganov.whispercpp.params.WhisperSamplingStrategy;
@ -12,7 +11,6 @@ import javax.sound.sampled.AudioInputStream;
 import javax.sound.sampled.AudioSystem;
 import java.io.File;
 import java.io.FileNotFoundException;
-import java.util.List;

 class WhisperCppTest {
    private static WhisperCpp whisper = new WhisperCpp();
@ -22,12 +20,11 @@ class WhisperCppTest {
    static void init() throws FileNotFoundException {
        // By default, models are loaded from ~/.cache/whisper/ and are usually named "ggml-${name}.bin"
        // or you can provide the absolute path to the model file.
-        //String modelName = "../../models/ggml-tiny.bin";
        String modelName = "../../models/ggml-tiny.en.bin";
        try {
            whisper.initContext(modelName);
-            //whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
-            //whisper.getJavaDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
+//            whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+//            whisper.getJavaDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
            modelInitialised = true;
        } catch (FileNotFoundException ex) {
            System.out.println("Model " + modelName + " not found");
@ -45,7 +42,7 @@ class WhisperCppTest {
        assertEquals(16384, params.n_max_text_ctx);
        assertFalse(params.translate);
        assertEquals(0.01f, params.thold_pt);
-        assertEquals(5, params.beam_search.beam_size);
+        assertEquals(2, params.beam_search.beam_size);
        assertEquals(-1.0f, params.beam_search.patience);
    }

@ -58,7 +55,7 @@ class WhisperCppTest {
        assertEquals(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY.ordinal(), params.strategy);
        assertNotEquals(0, params.n_threads);
        assertEquals(16384, params.n_max_text_ctx);
-        assertEquals(5, params.greedy.best_of);
+        assertEquals(2, params.greedy.best_of);
    }

    @Test
@ -75,11 +72,11 @@ class WhisperCppTest {
        byte[] b = new byte[audioInputStream.available()];
        float[] floats = new float[b.length / 2];

-        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
+//        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
        params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
        params.print_progress = CBool.FALSE;
-        //params.initial_prompt = "and so my fellow Americans um, like";
+//        params.initial_prompt = "and so my fellow Americans um, like";


        try {
@ -102,43 +99,4 @@ class WhisperCppTest {
            audioInputStream.close();
        }
    }
-
-    @Test
-    void testFullTranscribeWithTime() throws Exception {
-        if (!modelInitialised) {
-            System.out.println("Model not initialised, skipping test");
-            return;
-        }
-
-        // Given
-        File file = new File(System.getProperty("user.dir"), "../../samples/jfk.wav");
-        AudioInputStream audioInputStream = AudioSystem.getAudioInputStream(file);
-
-        byte[] b = new byte[audioInputStream.available()];
-        float[] floats = new float[b.length / 2];
-
-        //WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_GREEDY);
-        WhisperFullParams params = whisper.getFullDefaultParams(WhisperSamplingStrategy.WHISPER_SAMPLING_BEAM_SEARCH);
-        params.setProgressCallback((ctx, state, progress, user_data) -> System.out.println("progress: " + progress));
-        params.print_progress = CBool.FALSE;
-        //params.initial_prompt = "and so my fellow Americans um, like";
-
-        try {
-            audioInputStream.read(b);
-
-            for (int i = 0, j = 0; i < b.length; i += 2, j++) {
-                int intSample = (int) (b[i + 1]) << 8 | (int) (b[i]) & 0xFF;
-                floats[j] = intSample / 32767.0f;
-            }
-
-            List<WhisperSegment> segments = whisper.fullTranscribeWithTime(params, floats);
-            assertTrue(segments.size() > 0, "The size of segments should be greater than 0");
-            for (WhisperSegment segment : segments) {
-                System.out.println(segment);
-            }
-        } finally {
-            audioInputStream.close();
-        }
-    }
-
 }
--- a/bindings/javascript/README.md
+++ b/bindings/javascript/README.md
@ -41,7 +41,7 @@ make publish-npm

 ## Sample run

-```text
+```java
 $ node --experimental-wasm-threads --experimental-wasm-simd ../tests/test-whisper.js

 whisper_model_load: loading model from 'whisper.bin'
@ -63,7 +63,7 @@ whisper_model_load: ggml ctx size =  140.60 MB
 whisper_model_load: memory size   =   22.83 MB
 whisper_model_load: model size    =  140.54 MB

-system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 |
+system_info: n_threads = 8 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | NEON = 0 | F16C = 0 | FP16_VA = 0 | WASM_SIMD = 1 | BLAS = 0 | 

 operator(): processing 176000 samples, 11.0 sec, 8 threads, 1 processors, lang = en, task = transcribe ...

--- a/bindings/javascript/emscripten.cpp
+++ b/bindings/javascript/emscripten.cpp
@ -20,7 +20,7 @@ struct whisper_context * g_context;
 EMSCRIPTEN_BINDINGS(whisper) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        if (g_context == nullptr) {
-            g_context = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+            g_context = whisper_init_from_file(path_model.c_str());
            if (g_context != nullptr) {
                return true;
            } else {
--- a/bindings/javascript/package.json
+++ b/bindings/javascript/package.json
@ -1,6 +1,6 @@
 {
  "name": "whisper.cpp",
-  "version": "1.5.4",
+  "version": "1.4.2",
  "description": "Whisper speech recognition",
  "main": "whisper.js",
  "scripts": {
--- a/bindings/javascript/whisper.js
+++ b/bindings/javascript/whisper.js
--- a/bindings/ruby/ext/.gitignore
+++ b/bindings/ruby/ext/.gitignore
@ -1,8 +1,6 @@
 Makefile
 ggml.c
 ggml.h
-ggml-alloc.c
-ggml-alloc.h
 whisper.bundle
 whisper.cpp
 whisper.h
--- a/bindings/ruby/ext/extconf.rb
+++ b/bindings/ruby/ext/extconf.rb
@ -3,14 +3,6 @@ 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__),'..','..','..','ggml.h')} .")
 system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml.c')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-impl.h')} .")
-system("cp #{File.join(File.dirname(__FILE__),'..','..','..','ggml-alloc.h')} .")
-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-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
@ -1,87 +0,0 @@
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        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_t                backend;
-        ggml_backend_buffer_context_t context;
-
-        size_t size;
-    };
-
-    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);
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*get_name)(ggml_backend_t backend);
-
-        void (*free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);
-
-        // get buffer alignment
-        size_t (*get_alignment)(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);
-
-        // (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
-        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 (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-    };
-
-    struct ggml_backend {
-        struct ggml_backend_i iface;
-
-        ggml_backend_context_t context;
-    };
-
-#ifdef  __cplusplus
-}
-#endif
--- a/bindings/ruby/ext/ggml-backend.c
+++ b/bindings/ruby/ext/ggml-backend.c
@ -1,950 +0,0 @@
-#include "ggml-backend-impl.h"
-#include "ggml-alloc.h"
-#include "ggml-impl.h"
-
-#include <assert.h>
-#include <limits.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define UNUSED GGML_UNUSED
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-// backend buffer
-
-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) {
-    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
-
-    GGML_ASSERT(iface.get_base != NULL);
-
-    (*buffer) = (struct ggml_backend_buffer) {
-        /* .interface = */ iface,
-        /* .backend   = */ backend,
-        /* .context   = */ context,
-        /* .size      = */ size,
-    };
-
-    return buffer;
-}
-
-void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
-    if (buffer == NULL) {
-        return;
-    }
-
-    if (buffer->iface.free_buffer != NULL) {
-        buffer->iface.free_buffer(buffer);
-    }
-    free(buffer);
-}
-
-size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
-    return ggml_backend_get_alignment(buffer->backend);
-}
-
-size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-    return buffer->size;
-}
-
-void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
-    void * base = buffer->iface.get_base(buffer);
-
-    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
-
-    return base;
-}
-
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // get_alloc_size is optional, defaults to ggml_nbytes
-    if (buffer->iface.get_alloc_size) {
-        return buffer->iface.get_alloc_size(buffer, tensor);
-    }
-    return ggml_nbytes(tensor);
-}
-
-void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // init_tensor is optional
-    if (buffer->iface.init_tensor) {
-        buffer->iface.init_tensor(buffer, tensor);
-    }
-}
-
-void ggml_backend_buffer_free_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // free_tensor is optional
-    if (buffer->iface.free_tensor) {
-        buffer->iface.free_tensor(buffer, tensor);
-    }
-}
-
-// backend
-
-ggml_backend_t ggml_get_backend(const struct ggml_tensor * tensor) {
-    return tensor->buffer ? tensor->buffer->backend : NULL;
-}
-
-const char * ggml_backend_name(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return "NULL";
-    }
-    return backend->iface.get_name(backend);
-}
-
-void ggml_backend_free(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return;
-    }
-
-    backend->iface.free(backend);
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
-    return backend->iface.alloc_buffer(backend, size);
-}
-
-size_t ggml_backend_get_alignment(ggml_backend_t backend) {
-    return backend->iface.get_alignment(backend);
-}
-
-void ggml_backend_tensor_set_async(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.set_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_get_async(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_get_backend(tensor)->iface.get_tensor_async(ggml_get_backend(tensor), tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
-
-    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
-}
-
-void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_t backend = ggml_get_backend(tensor);
-
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(backend != NULL && "tensor backend not set");
-
-    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
-    backend->iface.synchronize(backend);
-}
-
-void ggml_backend_synchronize(ggml_backend_t backend) {
-    backend->iface.synchronize(backend);
-}
-
-ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_plan_create(backend, cgraph);
-}
-
-void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_free(backend, plan);
-}
-
-void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_compute(backend, plan);
-}
-
-bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_compute(backend, cgraph);
-}
-
-bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return backend->iface.supports_op(backend, op);
-}
-
-// backend copy
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
-    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
-    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
-    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
-
-    // fprintf(stderr, "cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
-
-    if (src == dst) {
-        return;
-    }
-
-    // TODO: allow backends to support copy to/from same backend
-
-    if (ggml_get_backend(dst)->iface.cpy_tensor_from != NULL) {
-        ggml_get_backend(dst)->iface.cpy_tensor_from(ggml_get_backend(dst)->context, src, dst);
-    } else if (ggml_get_backend(src)->iface.cpy_tensor_to != NULL) {
-        ggml_get_backend(src)->iface.cpy_tensor_to(ggml_get_backend(src)->context, src, dst);
-    } else {
-        // shouldn't be hit when copying from/to CPU
-        #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to are implemented for backends %s and %s, falling back to get/set\n", ggml_backend_name(src->buffer->backend), ggml_backend_name(dst->buffer->backend));
-        #endif
-        size_t nbytes = ggml_nbytes(src);
-        void * data = malloc(nbytes);
-        ggml_backend_tensor_get(src, data, 0, nbytes);
-        ggml_backend_tensor_set(dst, data, 0, nbytes);
-        free(data);
-    }
-}
-
-// backend CPU
-
-struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
-};
-
-static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
-    return "CPU";
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    free(cpu_ctx->work_data);
-    free(cpu_ctx);
-    free(backend);
-}
-
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
-
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-    UNUSED(buffer);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .free_buffer    = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL, // no initialization required
-    /* .free_tensor    = */ NULL, // no cleanup required
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .free_buffer    = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base       = */ ggml_backend_cpu_buffer_get_base,
-    /* .get_alloc_size = */ NULL, // defaults to ggml_nbytes
-    /* .init_tensor    = */ NULL,
-    /* .free_tensor    = */ NULL,
-};
-
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-
-static ggml_backend_buffer_t ggml_backend_cpu_alloc_buffer(ggml_backend_t backend, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
-
-    GGML_ASSERT(data != NULL && "failed to allocate buffer");
-
-    return ggml_backend_buffer_init(backend, cpu_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_cpu_get_alignment(ggml_backend_t backend) {
-    return TENSOR_ALIGNMENT;
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_set_tensor_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_get_tensor_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_synchronize(ggml_backend_t backend) {
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_from(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_cpy_tensor_to(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
-
-    UNUSED(backend);
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph;
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-    }
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    free(cpu_plan->cplan.work_data);
-    free(cpu_plan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
-        cpu_ctx->work_size = cplan.work_size;
-    }
-
-    cplan.work_data = cpu_ctx->work_data;
-
-    ggml_graph_compute(cgraph, &cplan);
-}
-
-static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return true;
-    UNUSED(backend);
-    UNUSED(op);
-}
-
-static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name            = */ ggml_backend_cpu_name,
-    /* .free                = */ ggml_backend_cpu_free,
-    /* .alloc_buffer        = */ ggml_backend_cpu_alloc_buffer,
-    /* .get_alignment       = */ ggml_backend_cpu_get_alignment,
-    /* .set_tensor_async    = */ ggml_backend_cpu_set_tensor_async,
-    /* .get_tensor_async    = */ ggml_backend_cpu_get_tensor_async,
-    /* .synchronize         = */ ggml_backend_cpu_synchronize,
-    /* .cpy_tensor_from     = */ ggml_backend_cpu_cpy_tensor_from,
-    /* .cpy_tensor_to       = */ ggml_backend_cpu_cpy_tensor_to,
-    /* .graph_plan_create   = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free     = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_compute  = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute       = */ ggml_backend_cpu_graph_compute,
-    /* .supports_op         = */ ggml_backend_cpu_supports_op,
-};
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
-
-    ctx->n_threads = GGML_DEFAULT_N_THREADS;
-    ctx->work_data = NULL;
-    ctx->work_size = 0;
-
-    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
-
-    *cpu_backend = (struct ggml_backend) {
-        /* .interface = */ cpu_backend_i,
-        /* .context   = */ ctx
-    };
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_cpu_name;
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size) {
-    return ggml_backend_buffer_init(backend_cpu, cpu_backend_buffer_i_from_ptr, ptr, size);
-}
-
-// scheduler
-
-#define GGML_MAX_BACKENDS 4
-#define GGML_MAX_SPLITS 256
-#define GGML_MAX_SPLIT_INPUTS 16
-
-struct ggml_backend_sched_split {
-    ggml_tallocr_t tallocr;
-    int i_start;
-    int i_end;
-    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
-    int n_inputs;
-    struct ggml_cgraph * graph;
-};
-
-struct ggml_backend_sched {
-    int n_backends;
-    ggml_backend_t backends[GGML_MAX_BACKENDS];
-    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
-
-    ggml_gallocr_t galloc;
-
-    struct ggml_hash_set    hash_set;
-    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
-    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
-
-    struct ggml_cgraph * graph;
-    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
-    int n_splits;
-
-    struct ggml_context * ctx;
-
-    // align context_buffer to GGML_MEM_ALIGN
-    #ifdef _MSC_VER
-    __declspec(align(GGML_MEM_ALIGN))
-    #else
-    __attribute__((aligned(GGML_MEM_ALIGN)))
-    #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + GGML_MAX_SPLITS*sizeof(struct ggml_cgraph)];
-};
-
-#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
-#define node_allocr(node) sched->node_talloc[hash_id(node)]
-
-static bool ggml_is_view_op(enum ggml_op op) {
-    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
-}
-
-// returns the priority of the backend, lower is better
-static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->backends[i] == backend) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->tallocs[i] == allocr) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-// returns the backend that should be used for the node based on the current locations
-char causes[GGML_DEFAULT_GRAPH_SIZE*4 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
-static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
-    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
-    // ie. kv cache updates
-    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
-    // dst
-    ggml_backend_t cur_backend = ggml_get_backend(node);
-    if (cur_backend != NULL) {
-        sprintf(causes[hash_id(node)], "1.dst");
-        return cur_backend;
-    }
-
-    // view_src
-    if (node->view_src != NULL && ggml_get_backend(node->view_src) != NULL) {
-        sprintf(causes[hash_id(node)], "1.vsrc");
-        return ggml_get_backend(node->view_src);
-    }
-
-    // src
-    int cur_prio = INT_MAX;
-    size_t cur_size = 0;
-
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        const struct ggml_tensor * src = node->src[i];
-        if (src == NULL) {
-            break;
-        }
-        ggml_backend_t src_backend = ggml_get_backend(src);
-        if (src_backend != NULL) {
-            int src_prio = sched_backend_prio(sched, src_backend);
-            size_t src_size = ggml_nbytes(src);
-            if (src_prio < cur_prio && src_size >= cur_size) {
-                cur_prio = src_prio;
-                cur_size = src_size;
-                cur_backend = src_backend;
-                sprintf(causes[hash_id(node)], "1.src%d", i);
-            }
-        }
-    }
-    return cur_backend;
-}
-
-static char * fmt_size(size_t size) {
-    static char buffer[128];
-    if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
-    } else {
-        sprintf(buffer, "%zuK", size/1024);
-    }
-    return buffer;
-}
-
-static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
-            ggml_backend_t split_backend = ggml_tallocr_get_buffer(sched->splits[cur_split].tallocr)->backend;
-            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend), sched->splits[cur_split].n_inputs);
-            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
-                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name, fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
-            }
-            fprintf(stderr, "\n");
-            cur_split++;
-        }
-        struct ggml_tensor * node = graph->nodes[i];
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        ggml_backend_t node_backend = node_allocr ? ggml_tallocr_get_buffer(node_allocr)->backend : NULL;
-        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name, fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", causes[hash_id(node)]);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            ggml_backend_t src_backend = src_allocr ? ggml_tallocr_get_buffer(src_allocr)->backend : NULL;
-            fprintf(stderr, " %20.20s (%4.4s) [%4.4s %8.8s]", src->name, fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", causes[hash_id(src)]);
-        }
-        fprintf(stderr, "\n");
-    }
-}
-
-// creates a copy of the tensor with the same memory layout
-static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
-    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        dup->nb[i] = tensor->nb[i];
-    }
-    return dup;
-}
-
-// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
-// TODO: merge passes
-static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    // reset state
-    size_t hash_size = sched->hash_set.size;
-    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
-    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
-    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
-    sched->n_splits = 0;
-
-    struct ggml_init_params params = {
-        /*.mem_size =   */ sizeof(sched->context_buffer),
-        /*.mem_buffer = */ sched->context_buffer,
-        /*.no_alloc =   */ true
-    };
-
-    if (sched->ctx != NULL) {
-        ggml_free(sched->ctx);
-    }
-
-    sched->ctx = ggml_init(params);
-
-    // pass 1: assign backends to ops with allocated inputs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        if (node_allocr(leaf) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
-        if (leaf_backend == NULL && leaf->view_src != NULL) {
-            leaf_backend = ggml_get_backend(leaf->view_src);
-        }
-        if (leaf_backend != NULL) {
-            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
-        }
-    }
-
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node_allocr(node) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
-        if (node_backend != NULL) {
-            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
-        }
-    }
-    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 2: assign backends to ops from current assignments
-    // TODO:
-    //  - reuse sched_backend_from_cur
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            int    cur_prio = INT_MAX;
-            size_t cur_size = 0;
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * src = node->src[j];
-                if (src == NULL) {
-                    break;
-                }
-                ggml_tallocr_t src_allocr = node_allocr(src);
-                if (src_allocr != NULL) {
-                    int    src_prio = sched_allocr_prio(sched, src_allocr);
-                    size_t src_size = ggml_nbytes(src);
-                    if (src_prio < cur_prio && src_size >= cur_size) {
-                        cur_prio = src_prio;
-                        cur_size = src_size;
-                        node_allocr = src_allocr;
-                        sprintf(causes[hash_id(node)], "2.src%d", j);
-                    }
-                }
-            }
-            if (node_allocr != NULL) {
-                node_allocr(node) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 3: assign backends to remaining src from dst (should only be leafs)
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr == NULL) {
-                node_allocr(src) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 4: split graph, find tensors that need to be copied
-    // TODO:
-    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
-    // find first backend
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node->view_src == NULL) {
-            sched->splits[0].tallocr = node_allocr(node);
-            break;
-        }
-    }
-    sched->splits[0].i_start = 0;
-    sched->splits[0].n_inputs = 0;
-    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
-    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
-    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-
-        ggml_tallocr_t node_allocr = node_allocr(node);
-
-        if (node_allocr != cur_allocr) {
-            sched->splits[cur_split].i_end = i;
-            cur_split++;
-            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
-            sched->splits[cur_split].tallocr = node_allocr;
-            sched->splits[cur_split].i_start = i;
-            sched->splits[cur_split].n_inputs = 0;
-            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
-            cur_allocr = node_allocr;
-            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-        }
-
-        // find inputs that are not on the same backend
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr) {
-                int n_inputs = sched->splits[cur_split].n_inputs++;
-                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
-                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
-
-                // create copies
-                size_t id = hash_id(src);
-                if (sched->node_copies[id][cur_backend_id] == NULL) {
-                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                    sched->node_copies[id][cur_backend_id] = tensor_copy;
-                    node_allocr(tensor_copy) = cur_allocr;
-                    ggml_backend_t backend = ggml_tallocr_get_buffer(cur_allocr)->backend;
-                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-                }
-                node->src[j] = sched->node_copies[id][cur_backend_id];
-            }
-        }
-    }
-    sched->splits[cur_split].i_end = graph->n_nodes;
-    sched->n_splits = cur_split + 1;
-
-    //fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); fflush(stdout);
-
-#if 1
-    // sanity check: all sources should have the same backend as the node
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
-        }
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
-                fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
-                    node->name, node_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(node_allocr)->backend) : "NULL",
-                    j, src->name, src_allocr ? ggml_backend_name(ggml_tallocr_get_buffer(src_allocr)->backend) : "NULL");
-            }
-        }
-    }
-#endif
-
-    // create copies of the graph for each split
-    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &sched->splits[i];
-        split->graph = ggml_graph_view(sched->ctx, graph, split->i_start, split->i_end);
-
-        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
-            input_cpy->src[0] = input;
-            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
-        }
-
-        for (int j = split->i_start; j < split->i_end; j++) {
-            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
-        }
-    }
-    sched->graph = graph_copy;
-}
-
-static void sched_alloc_splits(ggml_backend_sched_t sched) {
-    ggml_gallocr_alloc_graph_n(
-        sched->galloc,
-        sched->graph,
-        sched->hash_set,
-        sched->node_talloc);
-}
-
-static void sched_compute_splits(ggml_backend_sched_t sched) {
-    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
-    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
-
-    struct ggml_backend_sched_split * splits = sched->splits;
-
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
-        ggml_backend_t split_backend = ggml_tallocr_get_buffer(split->tallocr)->backend;
-        int split_backend_id = sched_backend_prio(sched, split_backend);
-
-        // copy the input tensors to the split backend
-        uint64_t copy_start_us = ggml_time_us();
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(split->inputs[j])][sched_backend_prio(sched, split_backend)];
-            if (split->inputs[j]->buffer == NULL) {
-                if (split->inputs[j]->view_src == NULL) {
-                    fprintf(stderr, "input %s has no buffer and no view_src\n", split->inputs[j]->name);
-                    exit(1);
-                }
-                struct ggml_tensor * view = split->inputs[j];
-                view->backend = view->view_src->backend;
-                view->buffer  = view->view_src->buffer;
-                view->data    = (char *)view->view_src->data + view->view_offs;
-                ggml_backend_buffer_init_tensor(ggml_backend_sched_get_buffer(sched, view->buffer->backend), view);
-            }
-            if (input_cpy->buffer == NULL) {
-                fprintf(stderr, "input_cpy %s has no buffer\n", input_cpy->name);
-                exit(1);
-            }
-            GGML_ASSERT(split->inputs[j]->buffer->backend != input_cpy->buffer->backend);
-            GGML_ASSERT(input_cpy->buffer->backend == split_backend);
-            ggml_backend_tensor_copy(split->inputs[j], input_cpy);
-        }
-        // ggml_backend_synchronize(split_backend);
-        int64_t copy_end_us = ggml_time_us();
-        copy_us[split_backend_id] += copy_end_us - copy_start_us;
-
-#if 0
-        char split_filename[GGML_MAX_NAME];
-        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
-        ggml_graph_dump_dot(split->graph, NULL, split_filename);
-#endif
-
-        uint64_t compute_start_us = ggml_time_us();
-        ggml_backend_graph_compute(split_backend, split->graph);
-        // ggml_backend_synchronize(split_backend);
-        uint64_t compute_end_us = ggml_time_us();
-        compute_us[split_backend_id] += compute_end_us - compute_start_us;
-    }
-
-#if 0
-    // per-backend timings
-    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (copy_us[i] > 0 || compute_us[i] > 0) {
-            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
-        }
-    }
-#endif
-}
-
-static void sched_reset(ggml_backend_sched_t sched) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_reset(sched->tallocs[i]);
-    }
-}
-
-ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
-    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
-
-    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
-    memset(sched, 0, sizeof(struct ggml_backend_sched));
-
-    fprintf(stderr, "ggml_backend_sched size: %lu KB\n", sizeof(struct ggml_backend_sched)/1024);
-
-    sched->n_backends = n_backends;
-    for (int i = 0; i < n_backends; i++) {
-        sched->backends[i] = backends[i];
-    }
-
-    sched->galloc = ggml_gallocr_new();
-
-    // init measure allocs for each backend
-    for (int i = 0; i < n_backends; i++) {
-        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
-    }
-
-    return sched;
-}
-
-void ggml_backend_sched_free(ggml_backend_sched_t sched) {
-    if (sched == NULL) {
-        return;
-    }
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_free(sched->tallocs[i]);
-    }
-    ggml_gallocr_free(sched->galloc);
-    free(sched->hash_set.keys);
-    free(sched->node_talloc);
-    free(sched->node_copies);
-    free(sched);
-}
-
-void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    // initialize hash tables
-    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
-    sched->hash_set.size = hash_size;
-    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
-    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
-    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
-
-    sched_split_graph(sched, measure_graph);
-    sched_alloc_splits(sched);
-
-    // allocate buffers and reset allocators
-    for (int i = 0; i < sched->n_backends; i++) {
-        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
-        ggml_tallocr_free(sched->tallocs[i]);
-        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
-    }
-
-    sched_reset(sched);
-}
-
-void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
-
-    sched_split_graph(sched, graph);
-    sched_alloc_splits(sched);
-    sched_compute_splits(sched);
-    sched_reset(sched);
-}
-
-ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return sched->tallocs[backend_index];
-}
-
-ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
-}
-
-void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
-    node_allocr(node) = sched->tallocs[backend_index];
-}
--- a/bindings/ruby/ext/ggml-backend.h
+++ b/bindings/ruby/ext/ggml-backend.h
@ -1,136 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    struct ggml_backend_buffer;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-
-    // 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
-    //
-
-    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_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 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 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);
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    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_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(ggml_backend_t backend_cpu, void * ptr, size_t size);
-
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        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
-
-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
-
-        // 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 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);
-        }
-
-        // 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);
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // Initialize a backend scheduler
-    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 void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    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 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
-}
-#endif
--- a/bindings/ruby/ext/ggml-impl.h
+++ b/bindings/ruby/ext/ggml-impl.h
@ -1,249 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#include <assert.h>
-#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// static_assert should be a #define, but if it's not,
-// fall back to the _Static_assert C11 keyword.
-// if C99 - static_assert is noop
-// ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#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__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#ifndef __SSE3__
-#define __SSE3__
-#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
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
-
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#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)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#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)
-
-#endif // __F16C__
-
-#endif // __ARM_NEON
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-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) || !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));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#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)
-
-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_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
-size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-#ifdef __cplusplus
-}
-#endif
--- 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
@ -1,224 +0,0 @@
-#pragma once
-
-#include "ggml-impl.h"
-
-// GGML internal header
-
-#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 * 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 * 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_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_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 * 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 * 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 * 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/ruby_whisper.cpp
+++ b/bindings/ruby/ext/ruby_whisper.cpp
@ -87,7 +87,7 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
  if (!rb_respond_to(whisper_model_file_path, rb_intern("to_s"))) {
    rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
  }
-  rw->context = whisper_init_from_file_with_params(StringValueCStr(whisper_model_file_path), whisper_context_default_params());
+  rw->context = whisper_init_from_file(StringValueCStr(whisper_model_file_path));
  if (rw->context == nullptr) {
    rb_raise(rb_eRuntimeError, "error: failed to initialize whisper context");
  }
--- a/coreml/whisper-encoder-impl.h
+++ b/coreml/whisper-encoder-impl.h
@ -123,7 +123,7 @@ API_AVAILABLE(macos(12.0), ios(15.0), watchos(8.0), tvos(15.0)) __attribute__((v

 /**
    Make a prediction using the convenience interface
-    @param logmel_data as 1 × n_mel × 3000 3-dimensional array of floats:
+    @param logmel_data as 1 × 80 × 3000 3-dimensional array of floats:
    @param error If an error occurs, upon return contains an NSError object that describes the problem. If you are not interested in possible errors, pass in NULL.
    @return the prediction as whisper_encoder_implOutput
 */
--- a/coreml/whisper-encoder.h
+++ b/coreml/whisper-encoder.h
@ -3,8 +3,6 @@
 // Code is derived from the work of Github user @wangchou
 // ref: https://github.com/wangchou/callCoreMLFromCpp

-#include <stdint.h>
-
 #if __cplusplus
 extern "C" {
 #endif
@ -16,8 +14,6 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx);

 void whisper_coreml_encode(
        const whisper_coreml_context * ctx,
-                             int64_t   n_ctx,
-                             int64_t   n_mel,
                               float * mel,
                               float * out);

--- a/coreml/whisper-encoder.mm
+++ b/coreml/whisper-encoder.mm
@ -22,13 +22,7 @@ struct whisper_coreml_context * whisper_coreml_init(const char * path_model) {

    NSURL * url_model = [NSURL fileURLWithPath: path_model_str];

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

    if (data == NULL) {
        return NULL;
@ -48,15 +42,13 @@ void whisper_coreml_free(struct whisper_coreml_context * ctx) {

 void whisper_coreml_encode(
        const whisper_coreml_context * ctx,
-                             int64_t   n_ctx,
-                             int64_t   n_mel,
                               float * mel,
                               float * out) {
    MLMultiArray * inMultiArray = [
        [MLMultiArray alloc] initWithDataPointer: mel
-                                           shape: @[@1, @(n_mel), @(n_ctx)]
+                                           shape: @[@1, @80, @3000]
                                        dataType: MLMultiArrayDataTypeFloat32
-                                         strides: @[@(n_ctx*n_mel), @(n_ctx), @1]
+                                         strides: @[@(240000), @(3000), @1]
                                     deallocator: nil
                                           error: nil
    ];
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@ -14,10 +14,6 @@ if (WHISPER_SDL2)
    message(STATUS "SDL2_LIBRARIES = ${SDL2_LIBRARIES}")
 endif()

-if (WHISPER_CLBLAST)
-    find_package(CLBlast REQUIRED)
-endif()
-
 # common

 set(TARGET common)
@ -54,9 +50,6 @@ if (WHISPER_SDL2)
    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()

-# add json lib
-add_library(json_cpp INTERFACE json.hpp)
-
 # examples

 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@ -72,16 +65,10 @@ elseif(CMAKE_JS_VERSION)
 else()
    add_subdirectory(main)
    add_subdirectory(stream)
-    add_subdirectory(server)
    add_subdirectory(command)
    add_subdirectory(bench)
    add_subdirectory(quantize)
    add_subdirectory(talk)
    add_subdirectory(talk-llama)
    add_subdirectory(lsp)
-    if (LLAMA_SYCL)
-        add_subdirectory(sycl)
-    endif()
 endif()
-
-add_subdirectory(wchess)
--- a/examples/addon.node/test/whisper.spec.js
+++ b/examples/addon.node/test/whisper.spec.js
@ -11,7 +11,6 @@ const whisperParamsMock = {
  language: "en",
  model: path.join(__dirname, "../../../models/ggml-base.en.bin"),
  fname_inp: path.join(__dirname, "../../../samples/jfk.wav"),
-  use_gpu: true,
 };

 describe("Run whisper.node", () => {
--- a/examples/addon.node/addon.cpp
+++ b/examples/addon.node/addon.cpp
@ -36,7 +36,6 @@ struct whisper_params {
    bool print_colors   = false;
    bool print_progress = false;
    bool no_timestamps  = false;
-    bool use_gpu        = true;

    std::string language = "en";
    std::string prompt;
@ -52,6 +51,27 @@ struct whisper_print_user_data {
    const std::vector<std::vector<float>> * pcmf32s;
 };

+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
 void whisper_print_segment_callback(struct whisper_context * ctx, 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;
@ -83,8 +103,8 @@ void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper
        if (params.diarize && pcmf32s.size() == 2) {
            const int64_t n_samples = pcmf32s[0].size();

-            const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
-            const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+            const int64_t is0 = timestamp_to_sample(t0, n_samples);
+            const int64_t is1 = timestamp_to_sample(t1, n_samples);

            double energy0 = 0.0f;
            double energy1 = 0.0f;
@ -133,9 +153,7 @@ int run(whisper_params &params, std::vector<std::vector<std::string>> &result) {

    // whisper init

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());

    if (ctx == nullptr) {
        fprintf(stderr, "error: failed to initialize whisper context\n");
@ -297,12 +315,10 @@ Napi::Value whisper(const Napi::CallbackInfo& info) {
  std::string language = whisper_params.Get("language").As<Napi::String>();
  std::string model = whisper_params.Get("model").As<Napi::String>();
  std::string input = whisper_params.Get("fname_inp").As<Napi::String>();
-  bool use_gpu = whisper_params.Get("use_gpu").As<Napi::Boolean>();

  params.language = language;
  params.model = model;
  params.fname_inp.emplace_back(input);
-  params.use_gpu = use_gpu;

  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
@ -11,7 +11,6 @@ const whisperParams = {
  language: "en",
  model: path.join(__dirname, "../../models/ggml-base.en.bin"),
  fname_inp: "../../samples/jfk.wav",
-  use_gpu: true,
 };

 const arguments = process.argv.slice(2);
--- a/examples/bench.wasm/emscripten.cpp
+++ b/examples/bench.wasm/emscripten.cpp
@ -23,9 +23,7 @@ void bench_main(size_t index) {

    fprintf(stderr, "%s: running benchmark with %d threads - please wait...\n", __func__, n_threads);

-    const int n_mels = whisper_model_n_mels(ctx);
-
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
        fprintf(stderr, "error: failed to set mel: %d\n", ret);
        return;
    }
@ -59,7 +57,7 @@ EMSCRIPTEN_BINDINGS(bench) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        for (size_t i = 0; i < g_contexts.size(); ++i) {
            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                if (g_contexts[i] != nullptr) {
                    if (g_worker.joinable()) {
                        g_worker.join();
--- a/examples/bench/bench.cpp
+++ b/examples/bench/bench.cpp
@ -1,18 +1,15 @@
 #include "whisper.h"

 #include <cstdio>
-#include <cstring>
 #include <string>
 #include <thread>

 // command-line parameters
 struct whisper_params {
    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t what = 0; // what to benchmark: 0 - whisper encoder, 1 - memcpy, 2 - ggml_mul_mat
+    int32_t what = 0; // what to benchmark: 0 - whisper ecoder, 1 - memcpy, 2 - ggml_mul_mat

    std::string model = "models/ggml-base.en.bin";
-
-    bool use_gpu = true;
 };

 void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
@ -25,10 +22,9 @@ 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 == "-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 {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -48,20 +44,16 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
    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, "                           %-7s  0 - whisper\n",                                 "");
+    fprintf(stderr, "                           %-7s  0 - whisper encoder\n",                         "");
    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
    fprintf(stderr, "\n");
 }

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

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());

    {
        fprintf(stderr, "\n");
@ -73,65 +65,16 @@ int whisper_bench_full(const whisper_params & params) {
        return 2;
    }

-    const int n_mels = whisper_model_n_mels(ctx);
-
-    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, WHISPER_N_MEL)) {
        fprintf(stderr, "error: failed to set mel: %d\n", ret);
        return 3;
    }
-    // heat encoder
+
    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        fprintf(stderr, "error: failed to encode model: %d\n", ret);
        return 4;
    }

-    whisper_token tokens[512];
-    memset(tokens, 0, sizeof(tokens));
-
-    // prompt heat
-    if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to decode: %d\n", ret);
-        return 4;
-    }
-
-    // text-generation heat
-    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to decode: %d\n", ret);
-        return 4;
-    }
-
-    whisper_reset_timings(ctx);
-
-    // actual run
-    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-        fprintf(stderr, "error: failed to encode: %d\n", ret);
-        return 4;
-    }
-
-    // text-generation
-    for (int i = 0; i < 256; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
-            return 4;
-        }
-    }
-
-    // batched decoding
-    for (int i = 0; i < 64; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
-            return 4;
-        }
-    }
-
-    // prompt processing
-    for (int i = 0; i < 16; i++) {
-        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
-            fprintf(stderr, "error: failed to decode: %d\n", ret);
-            return 4;
-        }
-    }
-
    whisper_print_timings(ctx);
    whisper_free(ctx);

@ -160,7 +103,7 @@ int main(int argc, char ** argv) {
    int ret = -1;

    switch (params.what) {
-        case 0: ret = whisper_bench_full(params);                break;
+        case 0: ret = whisper_bench_encoder(params);                break;
        case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
        case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
        default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;
--- a/examples/command.wasm/emscripten.cpp
+++ b/examples/command.wasm/emscripten.cpp
@ -243,7 +243,7 @@ EMSCRIPTEN_BINDINGS(command) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        for (size_t i = 0; i < g_contexts.size(); ++i) {
            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                if (g_contexts[i] != nullptr) {
                    g_running = true;
                    if (g_worker.joinable()) {
--- a/examples/command/command.cpp
+++ b/examples/command/command.cpp
@ -6,8 +6,8 @@
 // ref: https://github.com/ggerganov/whisper.cpp/issues/171
 //

-#include "common-sdl.h"
 #include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 #include "grammar-parser.h"

@ -22,6 +22,11 @@
 #include <vector>
 #include <map>

+bool file_exists(const std::string & fname) {
+    std::ifstream f(fname.c_str());
+    return f.good();
+}
+
 // command-line parameters
 struct whisper_params {
    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
@ -43,7 +48,6 @@ struct whisper_params {
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
-    bool use_gpu       = true;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -76,7 +80,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        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 == "-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]; }
@ -113,7 +116,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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, "  -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());
@ -171,14 +173,10 @@ std::string transcribe(
    auto grammar_rules = grammar_parsed.c_rules();

    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
-        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
-            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, 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(grammar_rule);
-            wparams.grammar_penalty = params.grammar_penalty;
-        }
+        wparams.grammar_rules   = grammar_rules.data();
+        wparams.n_grammar_rules = grammar_rules.size();
+        wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
+        wparams.grammar_penalty = params.grammar_penalty;
    }

    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
@ -688,10 +686,7 @@ int main(int argc, char ** argv) {

    // whisper init

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());

    // print some info about the processing
    {
@ -731,7 +726,7 @@ int main(int argc, char ** argv) {

    if (!params.grammar.empty()) {
        auto & grammar = params.grammar_parsed;
-        if (is_file_exist(params.grammar.c_str())) {
+        if (file_exists(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>());
--- a/examples/common-ggml.cpp
+++ b/examples/common-ggml.cpp
@ -9,11 +9,6 @@ static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
    {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
    {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
    {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
-    {"q2_k", GGML_FTYPE_MOSTLY_Q2_K},
-    {"q3_k", GGML_FTYPE_MOSTLY_Q3_K},
-    {"q4_k", GGML_FTYPE_MOSTLY_Q4_K},
-    {"q5_k", GGML_FTYPE_MOSTLY_Q5_K},
-    {"q6_k", GGML_FTYPE_MOSTLY_Q6_K},
 };

 void ggml_print_ftypes(FILE * fp) {
@ -53,23 +48,15 @@ bool ggml_common_quantize_0(
        case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
        case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
        case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
-        case GGML_FTYPE_MOSTLY_Q2_K: qtype = GGML_TYPE_Q2_K; break;
-        case GGML_FTYPE_MOSTLY_Q3_K: qtype = GGML_TYPE_Q3_K; break;
-        case GGML_FTYPE_MOSTLY_Q4_K: qtype = GGML_TYPE_Q4_K; break;
-        case GGML_FTYPE_MOSTLY_Q5_K: qtype = GGML_TYPE_Q5_K; break;
-        case GGML_FTYPE_MOSTLY_Q6_K: qtype = GGML_TYPE_Q6_K; break;
        case GGML_FTYPE_UNKNOWN:
        case GGML_FTYPE_ALL_F32:
        case GGML_FTYPE_MOSTLY_F16:
        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
-        case GGML_FTYPE_MOSTLY_IQ2_XXS:
-        case GGML_FTYPE_MOSTLY_IQ2_XS:
-        case GGML_FTYPE_MOSTLY_IQ2_S:
-        case GGML_FTYPE_MOSTLY_IQ3_XXS:
-        case GGML_FTYPE_MOSTLY_IQ3_S:
-        case GGML_FTYPE_MOSTLY_IQ1_S:
-        case GGML_FTYPE_MOSTLY_IQ4_NL:
-        case GGML_FTYPE_MOSTLY_IQ4_XS:
+        case GGML_FTYPE_MOSTLY_Q2_K:
+        case GGML_FTYPE_MOSTLY_Q3_K:
+        case GGML_FTYPE_MOSTLY_Q4_K:
+        case GGML_FTYPE_MOSTLY_Q5_K:
+        case GGML_FTYPE_MOSTLY_Q6_K:
                {
                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                    return false;
@ -90,6 +77,8 @@ bool ggml_common_quantize_0(
    std::vector<ggml_fp16_t> data_f16;
    std::vector<float>       data_f32;

+    std::vector<int64_t> hist_all(1 << 4, 0);
+
    while (true) {
        int32_t n_dims;
        int32_t length;
@ -174,19 +163,28 @@ bool ggml_common_quantize_0(
            work.resize(nelements); // for quantization

            size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
            switch ((ggml_type) ttype) {
                case GGML_TYPE_Q4_0:
-                case GGML_TYPE_Q4_1:
-                case GGML_TYPE_Q5_0:
-                case GGML_TYPE_Q5_1:
-                case GGML_TYPE_Q8_0:
-                case GGML_TYPE_Q2_K:
-                case GGML_TYPE_Q3_K:
-                case GGML_TYPE_Q4_K:
-                case GGML_TYPE_Q5_K:
-                case GGML_TYPE_Q6_K:
                    {
-                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], nullptr);
+                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_1:
+                    {
+                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_0:
+                    {
+                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_1:
+                    {
+                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q8_0:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
                    } break;
                case GGML_TYPE_F32:
                case GGML_TYPE_F16:
@ -194,15 +192,12 @@ bool ggml_common_quantize_0(
                case GGML_TYPE_I16:
                case GGML_TYPE_I32:
                case GGML_TYPE_Q8_1:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
                case GGML_TYPE_Q8_K:
-                case GGML_TYPE_IQ2_XXS:
-                case GGML_TYPE_IQ2_XS:
-                case GGML_TYPE_IQ2_S:
-                case GGML_TYPE_IQ3_XXS:
-                case GGML_TYPE_IQ3_S:
-                case GGML_TYPE_IQ1_S:
-                case GGML_TYPE_IQ4_NL:
-                case GGML_TYPE_IQ4_XS:
                case GGML_TYPE_COUNT:
                    {
                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
@ -213,7 +208,15 @@ bool ggml_common_quantize_0(
            fout.write(reinterpret_cast<char *>(work.data()), cur_size);
            total_size_new += cur_size;

-            printf("size = %8.2f MB -> %8.2f MB\n", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
        } else {
            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
            fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
@ -226,5 +229,18 @@ bool ggml_common_quantize_0(
    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));

+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
    return true;
 }
--- a/examples/common-sdl.cpp
+++ b/examples/common-sdl.cpp
@ -139,13 +139,10 @@ void audio_async::callback(uint8_t * stream, int len) {
        return;
    }

-    size_t n_samples = len / sizeof(float);
+    const size_t n_samples = len / sizeof(float);

-    if (n_samples > m_audio.size()) {
-        n_samples = m_audio.size();
-
-        stream += (len - (n_samples * sizeof(float)));
-    }
+    m_audio_new.resize(n_samples);
+    memcpy(m_audio_new.data(), stream, n_samples * sizeof(float));

    //fprintf(stderr, "%s: %zu samples, pos %zu, len %zu\n", __func__, n_samples, m_audio_pos, m_audio_len);

@ -156,7 +153,7 @@ void audio_async::callback(uint8_t * stream, int len) {
            const size_t n0 = m_audio.size() - m_audio_pos;

            memcpy(&m_audio[m_audio_pos], stream, n0 * sizeof(float));
-            memcpy(&m_audio[0], stream + n0 * sizeof(float), (n_samples - n0) * sizeof(float));
+            memcpy(&m_audio[0], &stream[n0], (n_samples - n0) * sizeof(float));

            m_audio_pos = (m_audio_pos + n_samples) % m_audio.size();
            m_audio_len = m_audio.size();
--- a/examples/common-sdl.h
+++ b/examples/common-sdl.h
@ -41,6 +41,7 @@ private:
    std::mutex       m_mutex;

    std::vector<float> m_audio;
+    std::vector<float> m_audio_new;
    size_t             m_audio_pos = 0;
    size_t             m_audio_len = 0;
 };
--- a/examples/common.cpp
+++ b/examples/common.cpp
@ -38,12 +38,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
            params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
        } else if (arg == "-t" || arg == "--threads") {
            params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
        } else if (arg == "-p" || arg == "--prompt") {
            params.prompt = get_next_arg(i, argc, argv, arg, params);
        } else if (arg == "-n" || arg == "--n_predict") {
            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
-        } else if (arg == "-np" || arg == "--n_parallel") {
-            params.n_parallel = std::stoi(get_next_arg(i, argc, argv, arg, params));
        } else if (arg == "--top_k") {
            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
        } else if (arg == "--top_p") {
@ -56,12 +56,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
            params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
        } else if (arg == "-b" || arg == "--batch_size") {
            params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
-        } else if (arg == "-c" || arg == "--context") {
-            params.n_ctx= std::stoi(get_next_arg(i, argc, argv, arg, params));
-        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
-            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
-        } else if (arg == "--ignore-eos") {
-            params.ignore_eos = true;
        } else if (arg == "-m" || arg == "--model") {
            params.model = get_next_arg(i, argc, argv, arg, params);
        } else if (arg == "-i" || arg == "--interactive") {
@ -103,6 +97,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    fprintf(stderr, "  -h, --help            show this help message and exit\n");
    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
    fprintf(stderr, "  -f FNAME, --file FNAME\n");
@ -116,9 +111,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
-    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
-    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
    fprintf(stderr, "  -m FNAME, --model FNAME\n");
    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
    fprintf(stderr, "\n");
@ -615,21 +607,6 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(

 }

-bool is_wav_buffer(const std::string buf) {
-    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
-    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
-    if (buf.size() < 12 || buf.substr(0, 4) != "RIFF" || buf.substr(8, 4) != "WAVE") {
-        return false;
-    }
-
-    uint32_t chunk_size = *reinterpret_cast<const uint32_t*>(buf.data() + 4);
-    if (chunk_size + 8 != buf.size()) {
-        return false;
-    }
-
-    return true;
-}
-
 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
@ -654,12 +631,6 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector

        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
    }
-    else if (is_wav_buffer(fname)) {
-        if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
-            fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
-            return false;
-        }
-    }
    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
        return false;
@ -821,7 +792,7 @@ bool sam_params_parse(int argc, char ** argv, sam_params & params) {
    return true;
 }

-void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
+void sam_print_usage(int argc, char ** argv, const sam_params & params) {
    fprintf(stderr, "usage: %s [options]\n", argv[0]);
    fprintf(stderr, "\n");
    fprintf(stderr, "options:\n");
@ -836,48 +807,3 @@ void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
    fprintf(stderr, "\n");
 }
-
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma) {
-    int64_t msec = t * 10;
-    int64_t hr = msec / (1000 * 60 * 60);
-    msec = msec - hr * (1000 * 60 * 60);
-    int64_t min = msec / (1000 * 60);
-    msec = msec - min * (1000 * 60);
-    int64_t sec = msec / 1000;
-    msec = msec - sec * 1000;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
-
-    return std::string(buf);
-}
-
-int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate) {
-    return std::max(0, std::min((int) n_samples - 1, (int) ((t*whisper_sample_rate)/100)));
-}
-
-bool is_file_exist(const char *fileName)
-{
-    std::ifstream infile(fileName);
-    return infile.good();
-}
-
-bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id)
-{
-    std::ofstream speak_file(path.c_str());
-    if (speak_file.fail()) {
-        fprintf(stderr, "%s: failed to open speak_file\n", __func__);
-        return false;
-    } else {
-        speak_file.write(text.c_str(), text.size());
-        speak_file.close();
-        int ret = system((command + " " + std::to_string(voice_id) + " " + path).c_str());
-        if (ret != 0) {
-            fprintf(stderr, "%s: failed to speak\n", __func__);
-            return false;
-        }
-    }
-    return true;
-}
--- a/examples/common.h
+++ b/examples/common.h
@ -7,8 +7,6 @@
 #include <vector>
 #include <random>
 #include <thread>
-#include <ctime>
-#include <fstream>

 #define COMMON_SAMPLE_RATE 16000

@ -17,15 +15,10 @@
 //

 struct gpt_params {
-    int32_t seed         = -1;   // RNG seed
-    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_predict    = 200;  // new tokens to predict
-    int32_t n_parallel   = 1;    // number of parallel streams
-    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
-
-    bool ignore_eos = false; // ignore EOS token when generating text
+    int32_t seed      = -1;  // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict = 200; // new tokens to predict
+    int32_t n_batch   = 8;   // batch size for prompt processing

    // sampling parameters
    int32_t top_k          = 40;
@ -40,6 +33,8 @@ struct gpt_params {

    bool    interactive      = false;
    int32_t interactive_port = -1;
+
+    int32_t n_gpu_layers     = 0;
 };

 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
@ -135,11 +130,7 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 // Audio utils
 //

-// Check if a buffer is a WAV audio file
-bool is_wav_buffer(const std::string buf);
-
 // Read WAV audio file and store the PCM data into pcmf32
-// fname can be a buffer of WAV data instead of a filename
 // The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
 // If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
 bool read_wav(
@ -148,104 +139,6 @@ bool read_wav(
        std::vector<std::vector<float>> & pcmf32s,
        bool stereo);

-// Write PCM data into WAV audio file
-class wav_writer {
-private:
-    std::ofstream file;
-    uint32_t dataSize = 0;
-    std::string wav_filename;
-
-    bool write_header(const uint32_t sample_rate,
-                      const uint16_t bits_per_sample,
-                      const uint16_t channels) {
-
-        file.write("RIFF", 4);
-        file.write("\0\0\0\0", 4);    // Placeholder for file size
-        file.write("WAVE", 4);
-        file.write("fmt ", 4);
-
-        const uint32_t sub_chunk_size = 16;
-        const uint16_t audio_format = 1;      // PCM format
-        const uint32_t byte_rate = sample_rate * channels * bits_per_sample / 8;
-        const uint16_t block_align = channels * bits_per_sample / 8;
-
-        file.write(reinterpret_cast<const char *>(&sub_chunk_size), 4);
-        file.write(reinterpret_cast<const char *>(&audio_format), 2);
-        file.write(reinterpret_cast<const char *>(&channels), 2);
-        file.write(reinterpret_cast<const char *>(&sample_rate), 4);
-        file.write(reinterpret_cast<const char *>(&byte_rate), 4);
-        file.write(reinterpret_cast<const char *>(&block_align), 2);
-        file.write(reinterpret_cast<const char *>(&bits_per_sample), 2);
-        file.write("data", 4);
-        file.write("\0\0\0\0", 4);    // Placeholder for data size
-
-        return true;
-    }
-
-    // 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 = data[i] * 32767;
-            file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
-            dataSize += sizeof(int16_t);
-        }
-        if (file.is_open()) {
-            file.seekp(4, std::ios::beg);
-            uint32_t fileSize = 36 + dataSize;
-            file.write(reinterpret_cast<char *>(&fileSize), 4);
-            file.seekp(40, std::ios::beg);
-            file.write(reinterpret_cast<char *>(&dataSize), 4);
-            file.seekp(0, std::ios::end);
-        }
-        return true;
-    }
-
-    bool open_wav(const std::string & filename) {
-        if (filename != wav_filename) {
-            if (file.is_open()) {
-                file.close();
-            }
-        }
-        if (!file.is_open()) {
-            file.open(filename, std::ios::binary);
-            wav_filename = filename;
-            dataSize = 0;
-        }
-        return file.is_open();
-    }
-
-public:
-    bool open(const std::string & filename,
-              const    uint32_t   sample_rate,
-              const    uint16_t   bits_per_sample,
-              const    uint16_t   channels) {
-
-        if (open_wav(filename)) {
-            write_header(sample_rate, bits_per_sample, channels);
-        } else {
-            return false;
-        }
-
-        return true;
-    }
-
-    bool close() {
-        file.close();
-        return true;
-    }
-
-    bool write(const float * data, size_t length) {
-        return write_audio(data, length);
-    }
-
-    ~wav_writer() {
-        if (file.is_open()) {
-            file.close();
-        }
-    }
-};
-
-
 // Apply a high-pass frequency filter to PCM audio
 // Suppresses frequencies below cutoff Hz
 void high_pass_filter(
@ -281,31 +174,3 @@ struct sam_params {
 bool sam_params_parse(int argc, char ** argv, sam_params & params);

 void sam_print_usage(int argc, char ** argv, const sam_params & params);
-
-//
-// Terminal utils
-//
-
-
-// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
-// Lowest is red, middle is yellow, highest is green.
-const std::vector<std::string> k_colors = {
-    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
-    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
-};
-
-//
-// Other utils
-//
-
-// convert timestamp to string, 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma = false);
-
-// given a timestamp get the sample
-int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate);
-
-// check if file exists using ifstream
-bool is_file_exist(const char *fileName);
-
-// write text to file, and call system("command voice_id file")
-bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id);
--- a/examples/helpers.js
+++ b/examples/helpers.js
@ -22,7 +22,6 @@ var printTextarea = (function() {
 async function clearCache() {
    if (confirm('Are you sure you want to clear the cache?\nAll the models will be downloaded again.')) {
        indexedDB.deleteDatabase(dbName);
-        location.reload();
    }
 }

--- a/examples/livestream.sh
+++ b/examples/livestream.sh
@ -48,7 +48,7 @@ if [ -n "$3" ]; then
 fi

 # Whisper models
-models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large-v2" "large-v3" )
+models=( "tiny.en" "tiny" "base.en" "base" "small.en" "small" "medium.en" "medium" "large-v1" "large" )

 # list available models
 function list_models {
--- a/examples/lsp/CMakeLists.txt
+++ b/examples/lsp/CMakeLists.txt
@ -5,5 +5,5 @@ if (WHISPER_SDL2)

    include(DefaultTargetOptions)

-    target_link_libraries(${TARGET} PRIVATE common json_cpp common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
 endif ()
--- a/examples/lsp/json.hpp
+++ b/examples/lsp/json.hpp
--- a/examples/lsp/lsp.cpp
+++ b/examples/lsp/lsp.cpp
@ -30,7 +30,6 @@ struct whisper_params {
    bool translate     = false;
    bool print_special = false;
    bool print_energy  = false;
-    bool use_gpu       = true;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -73,7 +72,6 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        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 == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
        else {
@ -104,7 +102,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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, "  -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");
@ -327,12 +324,12 @@ json register_commandset(struct whisper_context * ctx, json jparams, std::vector
    commandset_list.push_back(cs);
    return json{{"index",index}};
 }
-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,
    // but for now, it's unsupported
-    throw json {
+    throw json{
        {"code", -32601},
            {"message", "Seeking is not yet supported."}
    };
@ -415,7 +412,7 @@ void process_loop(struct whisper_context * ctx, audio_async &audio, const whispe
            jobqueue.pop_front();
            // send response
            std::string data = resp.dump(-1, ' ', false, json::error_handler_t::replace);
-            fprintf(stdout, "Content-Length: %d\r\n\r\n%s\n", (int)data.length()+1, data.c_str());
+            fprintf(stdout, "Content-Length: %d\r\n\r\n%s\n", data.length()+1, data.c_str());
            std::cout.flush();

        }
@ -435,9 +432,7 @@ int main(int argc, char ** argv) {
    }

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

    audio_async audio(30*1000);
--- a/examples/main/README.md
+++ b/examples/main/README.md
@ -17,37 +17,28 @@ options:
  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
  -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
  -bo N,     --best-of N         [5      ] number of best candidates to keep
-  -bs N,     --beam-size N       [5      ] beam size for beam search
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -su,       --speed-up          [false  ] speed up audio by x2 (reduced accuracy)
  -tr,       --translate         [false  ] translate from source language to english
  -di,       --diarize           [false  ] stereo audio diarization
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
  -otxt,     --output-txt        [false  ] output result in a text file
  -ovtt,     --output-vtt        [false  ] output result in a vtt file
  -osrt,     --output-srt        [false  ] output result in a srt file
-  -olrc,     --output-lrc        [false  ] output result in a lrc file
  -owts,     --output-words      [false  ] output script for generating karaoke video
-  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
  -ocsv,     --output-csv        [false  ] output result in a CSV file
  -oj,       --output-json       [false  ] output result in a JSON file
-  -ojf,      --output-json-full  [false  ] include more information in the JSON file
  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
  -ps,       --print-special     [false  ] print special tokens
  -pc,       --print-colors      [false  ] print colors
  -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -nt,       --no-timestamps     [true   ] do not print timestamps
  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
             --prompt PROMPT     [       ] initial prompt
  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
  -f FNAME,  --file FNAME        [       ] input WAV file path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  -ls,       --log-score         [false  ] log best decoder scores of tokens
-  -ng,       --no-gpu            [false  ] disable GPU
 ```
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@ -14,6 +14,34 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif

+// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
+// Lowest is red, middle is yellow, highest is green.
+const std::vector<std::string> k_colors = {
+    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
+    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
+};
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
+}
+
 // helper function to replace substrings
 void replace_all(std::string & s, const std::string & search, const std::string & replace) {
    for (size_t pos = 0; ; pos += replace.length()) {
@ -34,9 +62,8 @@ struct whisper_params {
    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 best_of      =  2;
+    int32_t beam_size    = -1;

    float word_thold    =  0.01f;
    float entropy_thold =  2.40f;
@ -56,15 +83,12 @@ struct whisper_params {
    bool output_wts      = false;
    bool output_csv      = false;
    bool output_jsn      = false;
-    bool output_jsn_full = false;
    bool output_lrc      = false;
-    bool no_prints       = false;
    bool print_special   = false;
    bool print_colors    = false;
    bool print_progress  = false;
    bool no_timestamps   = false;
    bool log_score       = false;
-    bool use_gpu         = true;

    std::string language  = "en";
    std::string prompt;
@ -82,14 +106,6 @@ struct whisper_params {

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

-char* whisper_param_turn_lowercase(char* in){
-    int string_len = strlen(in);
-    for(int i = 0; i < string_len; i++){
-        *(in+i) = tolower((unsigned char)*(in+i));
-    }
-    return in;
-}
-
 bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
    for (int i = 1; i < argc; i++) {
        std::string arg = argv[i];
@ -117,7 +133,6 @@ 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-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]); }
@ -136,21 +151,18 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
        else if (arg == "-ocsv" || arg == "--output-csv")      { params.output_csv      = true; }
        else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
-        else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
        else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
-        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = whisper_param_turn_lowercase(argv[++i]); }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = 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 == "-ls"   || arg == "--log-score")       { params.log_score = true; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
            whisper_print_usage(argc, argv, params);
@ -177,7 +189,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
-    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
    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);
@ -195,9 +206,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
    fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
    fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
-    fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
    fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
-    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
@ -209,7 +218,6 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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, "  -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, "\n");
 }

@ -224,8 +232,8 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s
    std::string speaker = "";
    const int64_t n_samples = pcmf32s[0].size();

-    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
-    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+    const int64_t is0 = timestamp_to_sample(t0, n_samples);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples);

    double energy0 = 0.0f;
    double energy1 = 0.0f;
@ -252,7 +260,7 @@ std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s

    return speaker;
 }
-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) {
@ -484,7 +492,7 @@ bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_
    return true;
 }

-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);

@ -503,12 +511,7 @@ bool output_score(struct whisper_context * ctx, const char * fname, const whispe
    return true;
 }

-bool output_json(
-             struct whisper_context * ctx,
-                         const char * fname,
-               const whisper_params & params,
-    std::vector<std::vector<float>>   pcmf32s,
-                               bool   full) {
+bool output_json(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
    std::ofstream fout(fname);
    int indent = 0;

@ -525,7 +528,7 @@ bool output_json(
    auto end_arr = [&](bool end) {
        indent--;
        doindent();
-        fout << (end ? "]\n" : "],\n");
+        fout << (end ? "]\n" : "},\n");
    };

    auto start_obj = [&](const char *name) {
@ -566,29 +569,12 @@ bool output_json(
        end_value(end);
    };

-    auto value_f = [&](const char *name, const float val, bool end) {
-        start_value(name);
-        fout << val;
-        end_value(end);
-    };
-
    auto value_b = [&](const char *name, const bool val, bool end) {
        start_value(name);
        fout << (val ? "true" : "false");
        end_value(end);
    };

-    auto times_o = [&](int64_t t0, int64_t t1, bool end) {
-        start_obj("timestamps");
-        value_s("from", to_timestamp(t0, true).c_str(), false);
-        value_s("to", to_timestamp(t1, true).c_str(), true);
-        end_obj(false);
-        start_obj("offsets");
-        value_i("from", t0 * 10, false);
-        value_i("to", t1 * 10, true);
-        end_obj(end);
-    };
-
    if (!fout.is_open()) {
        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
        return false;
@ -634,26 +620,15 @@ bool output_json(
                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

                start_obj(nullptr);
-                    times_o(t0, t1, false);
-                    value_s("text", text, !params.diarize && !params.tinydiarize && !full);
-
-                    if (full) {
-                        start_arr("tokens");
-                        const int n = whisper_full_n_tokens(ctx, i);
-                        for (int j = 0; j < n; ++j) {
-                            auto token = whisper_full_get_token_data(ctx, i, j);
-                            start_obj(nullptr);
-                                value_s("text", whisper_token_to_str(ctx, token.id), false);
-                                if(token.t0 > -1 && token.t1 > -1) {
-                                    // If we have per-token timestamps, write them out
-                                    times_o(token.t0, token.t1, false);
-                                }
-                                value_i("id", token.id, false);
-                                value_f("p", token.p, true);
-                            end_obj(j == (n - 1));
-                        }
-                        end_arr(!params.diarize && !params.tinydiarize);
-                    }
+                    start_obj("timestamps");
+                        value_s("from", to_timestamp(t0, true).c_str(), false);
+                        value_s("to", to_timestamp(t1, true).c_str(), true);
+                    end_obj(false);
+                    start_obj("offsets");
+                        value_i("from", t0 * 10, false);
+                        value_i("to", t1 * 10, true);
+                    end_obj(false);
+                    value_s("text", text, !params.diarize && !params.tinydiarize);

                    if (params.diarize && pcmf32s.size() == 2) {
                        value_s("speaker", estimate_diarization_speaker(pcmf32s, t0, t1, true).c_str(), true);
@ -838,9 +813,6 @@ bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_
    return true;
 }

-
-void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
-
 int main(int argc, char ** argv) {
    whisper_params params;

@ -849,19 +821,6 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    // remove non-existent files
-    for (auto it = params.fname_inp.begin(); it != params.fname_inp.end();) {
-        const auto fname_inp = it->c_str();
-
-        if (*it != "-" && !is_file_exist(fname_inp)) {
-            fprintf(stderr, "error: input file not found '%s'\n", fname_inp);
-            it = params.fname_inp.erase(it);
-            continue;
-        }
-
-        it++;
-    }
-
    if (params.fname_inp.empty()) {
        fprintf(stderr, "error: no input files specified\n");
        whisper_print_usage(argc, argv, params);
@ -880,16 +839,9 @@ int main(int argc, char ** argv) {
        exit(0);
    }

-    if (params.no_prints) {
-        whisper_log_set(cb_log_disable, NULL);
-    }
-
    // whisper init

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());

    if (ctx == nullptr) {
        fprintf(stderr, "error: failed to initialize whisper context\n");
@ -911,28 +863,29 @@ int main(int argc, char ** argv) {
            continue;
        }

-        if (!whisper_is_multilingual(ctx)) {
-            if (params.language != "en" || params.translate) {
-                params.language = "en";
-                params.translate = false;
-                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-            }
-        }
-        if (params.detect_language) {
-            params.language = "auto";
-        }
-
-        if (!params.no_prints) {
-            // print system information
+        // print system information
+        {
            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
+        // print some info about the processing
+        {
            fprintf(stderr, "\n");
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            if (params.detect_language) {
+                params.language = "auto";
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
                    __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
-                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
+                    params.n_threads, params.n_processors,
                    params.language.c_str(),
                    params.translate ? "translate" : "transcribe",
                    params.tinydiarize ? "tdrz = 1, " : "",
@ -959,11 +912,10 @@ int main(int argc, char ** argv) {
            wparams.offset_ms        = params.offset_t_ms;
            wparams.duration_ms      = params.duration_ms;

-            wparams.token_timestamps = params.output_wts || params.output_jsn_full || params.max_len > 0;
+            wparams.token_timestamps = params.output_wts || params.max_len > 0;
            wparams.thold_pt         = params.word_thold;
            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
            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;
@ -979,8 +931,6 @@ int main(int argc, char ** argv) {
            wparams.entropy_thold    = params.entropy_thold;
            wparams.logprob_thold    = params.logprob_thold;

-            wparams.no_timestamps    = params.no_timestamps;
-
            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };

            // this callback is called on each new segment
@ -994,9 +944,8 @@ int main(int argc, char ** argv) {
                wparams.progress_callback_user_data = &user_data;
            }

-            // examples for abort mechanism
-            // in examples below, we do not abort the processing, but we could if the flag is set to true
-
+            // example for abort mechanism
+            // in this example, we do not abort the processing, but we could if the flag is set to true
            // the callback is called before every encoder run - if it returns false, the processing is aborted
            {
                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
@ -1008,17 +957,6 @@ int main(int argc, char ** argv) {
                wparams.encoder_begin_callback_user_data = &is_aborted;
            }

-            // the callback is called before every computation - if it returns true, the computation is aborted
-            {
-                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
-
-                wparams.abort_callback = [](void * user_data) {
-                    bool is_aborted = *(bool*)user_data;
-                    return is_aborted;
-                };
-                wparams.abort_callback_user_data = &is_aborted;
-            }
-
            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
                fprintf(stderr, "%s: failed to process audio\n", argv[0]);
                return 10;
@ -1062,7 +1000,7 @@ int main(int argc, char ** argv) {
            // output to JSON file
            if (params.output_jsn) {
                const auto fname_jsn = fname_out + ".json";
-                output_json(ctx, fname_jsn.c_str(), params, pcmf32s, params.output_jsn_full);
+                output_json(ctx, fname_jsn.c_str(), params, pcmf32s);
            }

            // output to LRC file
--- a/examples/python/test_whisper_processor.py
+++ b/examples/python/test_whisper_processor.py
@ -1,7 +0,0 @@
-import whisper_processor
-
-try:
-    result = whisper_processor.process_audio("./audio/wake_word_detected16k.wav", "base.en")
-    print(result)
-except Exception as e:
-    print(f"Error: {e}")
--- a/examples/python/whisper_processor.py
+++ b/examples/python/whisper_processor.py
@ -1,54 +0,0 @@
-import subprocess
-import sys
-import os
-
-def process_audio(wav_file, model_name="base.en"):
-    """
-    Processes an audio file using a specified model and returns the processed string.
-
-    :param wav_file: Path to the WAV file
-    :param model_name: Name of the model to use
-    :return: Processed string output from the audio processing
-    :raises: Exception if an error occurs during processing
-    """
-
-    model = f"./models/ggml-{model_name}.bin"
-
-    # Check if the file exists
-    if not os.path.exists(model):
-        raise FileNotFoundError(f"Model file not found: {model} \n\nDownload a model with this command:\n\n> bash ./models/download-ggml-model.sh {model_name}\n\n")
-
-    if not os.path.exists(wav_file):
-        raise FileNotFoundError(f"WAV file not found: {wav_file}")
-
-    full_command = f"./main -m {model} -f {wav_file} -np -nt"
-
-    # Execute the command
-    process = subprocess.Popen(full_command, shell=True, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
-
-    # Get the output and error (if any)
-    output, error = process.communicate()
-
-    if error:
-        raise Exception(f"Error processing audio: {error.decode('utf-8')}")
-
-    # Process and return the output string
-    decoded_str = output.decode('utf-8').strip()
-    processed_str = decoded_str.replace('[BLANK_AUDIO]', '').strip()
-
-    return processed_str
-
-def main():
-    if len(sys.argv) >= 2:
-        wav_file = sys.argv[1]
-        model_name = sys.argv[2] if len(sys.argv) == 3 else "base.en"
-        try:
-            result = process_audio(wav_file, model_name)
-            print(result)
-        except Exception as e:
-            print(f"Error: {e}")
-    else:
-        print("Usage: python whisper_processor.py <wav_file> [<model_name>]")
-
-if __name__ == "__main__":
-    main()
--- a/examples/server/CMakeLists.txt
+++ b/examples/server/CMakeLists.txt
@ -1,10 +0,0 @@
-set(TARGET server)
-add_executable(${TARGET} server.cpp httplib.h)
-
-include(DefaultTargetOptions)
-
-target_link_libraries(${TARGET} PRIVATE common json_cpp whisper ${CMAKE_THREAD_LIBS_INIT})
-
-if (WIN32)
-    target_link_libraries(${TARGET} PRIVATE ws2_32)
-endif()
--- a/examples/server/README.md
+++ b/examples/server/README.md
@ -1,69 +0,0 @@
-# whisper.cpp http server
-
-Simple http server. WAV Files are passed to the inference model via http requests.
-
-https://github.com/ggerganov/whisper.cpp/assets/1991296/e983ee53-8741-4eb5-9048-afe5e4594b8f
-
-## Usage
-
-```
-./server -h
-
-usage: ./bin/server [options]
-
-options:
-  -h,        --help              [default] show this help message and exit
-  -t N,      --threads N         [4      ] number of threads to use during computation
-  -p N,      --processors N      [1      ] number of processors to use during computation
-  -ot N,     --offset-t N        [0      ] time offset in milliseconds
-  -on N,     --offset-n N        [0      ] segment index offset
-  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
-  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
-  -ml N,     --max-len N         [0      ] maximum segment length in characters
-  -sow,      --split-on-word     [false  ] split on word rather than on token
-  -bo N,     --best-of N         [2      ] number of best candidates to keep
-  -bs N,     --beam-size N       [-1     ] beam size for beam search
-  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
-  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
-  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
-  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
-  -tr,       --translate         [false  ] translate from source language to english
-  -di,       --diarize           [false  ] stereo audio diarization
-  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
-  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
-  -ps,       --print-special     [false  ] print special tokens
-  -pc,       --print-colors      [false  ] print colors
-  -pr,       --print-realtime    [false  ] print output in realtime
-  -pp,       --print-progress    [false  ] print progress
-  -nt,       --no-timestamps     [false  ] do not print timestamps
-  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
-  -dl,       --detect-language   [false  ] exit after automatically detecting language
-             --prompt PROMPT     [       ] initial prompt
-  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
-  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
-  --host HOST,                   [127.0.0.1] Hostname/ip-adress for the server
-  --port PORT,                   [8080   ] Port number for the server
-  --convert,                     [false  ] Convert audio to WAV, requires ffmpeg on the server
-```
-
-> [!WARNING]
-> **Do not run the server example with administrative privileges and ensure it's operated in a sandbox environment, especially since it involves risky operations like accepting user file uploads and using ffmpeg for format conversions. Always validate and sanitize inputs to guard against potential security threats.**
-
-## request examples
-
-**/inference**
-```
-curl 127.0.0.1:8080/inference \
-H "Content-Type: multipart/form-data" \
-F file="@<file-path>" \
-F temperature="0.0" \
-F temperature_inc="0.2" \
-F response_format="json"
-```
-
-**/load**
-```
-curl 127.0.0.1:8080/load \
-H "Content-Type: multipart/form-data" \
-F model="<path-to-model-file>"
-```
--- a/examples/server/httplib.h
+++ b/examples/server/httplib.h
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@ -1,988 +0,0 @@
-#include "common.h"
-
-#include "whisper.h"
-#include "httplib.h"
-#include "json.hpp"
-
-#include <cmath>
-#include <fstream>
-#include <cstdio>
-#include <string>
-#include <thread>
-#include <vector>
-#include <cstring>
-#include <sstream>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-using namespace httplib;
-using json = nlohmann::ordered_json;
-
-namespace {
-
-// output formats
-const std::string json_format   = "json";
-const std::string text_format   = "text";
-const std::string srt_format    = "srt";
-const std::string vjson_format  = "verbose_json";
-const std::string vtt_format    = "vtt";
-
-struct server_params
-{
-    std::string hostname = "127.0.0.1";
-    std::string public_path = "examples/server/public";
-    std::string request_path = "";
-
-    int32_t port          = 8080;
-    int32_t read_timeout  = 600;
-    int32_t write_timeout = 600;
-
-    bool ffmpeg_converter = false;
-};
-
-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       = 2;
-    int32_t beam_size     = -1;
-    int32_t audio_ctx     = 0;
-
-    float word_thold      =  0.01f;
-    float entropy_thold   =  2.40f;
-    float logprob_thold   = -1.00f;
-    float temperature     =  0.00f;
-    float temperature_inc =  0.20f;
-
-    bool speed_up        = false;
-    bool debug_mode      = false;
-    bool translate       = false;
-    bool detect_language = false;
-    bool diarize         = false;
-    bool tinydiarize     = false;
-    bool split_on_word   = false;
-    bool no_fallback     = false;
-    bool print_special   = false;
-    bool print_colors    = false;
-    bool print_realtime  = false;
-    bool print_progress  = false;
-    bool no_timestamps   = false;
-    bool use_gpu         = true;
-
-    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 response_format     = json_format;
-
-    // [TDRZ] speaker turn string
-    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
-
-    std::string openvino_encode_device = "CPU";
-};
-
-void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
-    fprintf(stderr, "\n");
-    fprintf(stderr, "usage: %s [options] \n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
-    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
-    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
-    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
-    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
-    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
-    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
-    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
-    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
-    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
-    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
-    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");
-    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
-    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
-    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
-    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
-    fprintf(stderr, "  -pr,       --print-realtime    [%-7s] print output in realtime\n",                       params.print_realtime ? "true" : "false");
-    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
-    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "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\n",                                 params.prompt.c_str());
-    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, "  --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());
-    fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
-    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
-    fprintf(stderr, "\n");
-}
-
-bool whisper_params_parse(int argc, char ** argv, whisper_params & params, server_params & sparams) {
-    for (int i = 1; i < argc; i++) {
-        std::string arg = argv[i];
-
-        if (arg == "-h" || arg == "--help") {
-            whisper_print_usage(argc, argv, params, sparams);
-            exit(0);
-        }
-        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
-        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
-        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
-        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
-        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
-        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
-        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
-        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-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; }
-        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
-        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
-        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
-        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
-        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
-        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
-        else if (arg == "-pr"   || arg == "--print-realtime")  { params.print_realtime  = true; }
-        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
-        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
-        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
-        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
-        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
-        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
-        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
-        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
-        // server params
-        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
-        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
-        else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
-        else if (                  arg == "--request-path")    { sparams.request_path = argv[++i]; }
-        else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
-        else {
-            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            whisper_print_usage(argc, argv, params, sparams);
-            exit(0);
-        }
-    }
-
-    return true;
-}
-
-struct whisper_print_user_data {
-    const whisper_params * params;
-
-    const std::vector<std::vector<float>> * pcmf32s;
-    int progress_prev;
-};
-
-void check_ffmpeg_availibility() {
-    int result = system("ffmpeg -version");
-
-    if (result == 0) {
-        std::cout << "ffmpeg is available." << std::endl;
-    } else {
-        // ffmpeg is not available
-        std::cout << "ffmpeg is not found. Please ensure that ffmpeg is installed ";
-        std::cout << "and that its executable is included in your system's PATH. ";
-        exit(0);
-    }
-}
-
-bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
-    std::ostringstream cmd_stream;
-    std::string converted_filename_temp = temp_filename + "_temp.wav";
-    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
-    std::string cmd = cmd_stream.str();
-
-    int status = std::system(cmd.c_str());
-    if (status != 0) {
-        error_resp = "{\"error\":\"FFmpeg conversion failed.\"}";
-        return false;
-    }
-
-    // Remove the original file
-    if (remove(temp_filename.c_str()) != 0) {
-        error_resp = "{\"error\":\"Failed to remove the original file.\"}";
-        return false;
-    }
-
-    // Rename the temporary file to match the original filename
-    if (rename(converted_filename_temp.c_str(), temp_filename.c_str()) != 0) {
-        error_resp = "{\"error\":\"Failed to rename the temporary file.\"}";
-        return false;
-    }
-    return true;
-}
-
-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();
-
-    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
-    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
-
-    double energy0 = 0.0f;
-    double energy1 = 0.0f;
-
-    for (int64_t j = is0; j < is1; j++) {
-        energy0 += fabs(pcmf32s[0][j]);
-        energy1 += fabs(pcmf32s[1][j]);
-    }
-
-    if (energy0 > 1.1*energy1) {
-        speaker = "0";
-    } else if (energy1 > 1.1*energy0) {
-        speaker = "1";
-    } else {
-        speaker = "?";
-    }
-
-    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
-
-    if (!id_only) {
-        speaker.insert(0, "(speaker ");
-        speaker.append(")");
-    }
-
-    return speaker;
-}
-
-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) {
-        *progress_prev += progress_step;
-        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
-    }
-}
-
-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;
-
-    const int n_segments = whisper_full_n_segments(ctx);
-
-    std::string speaker = "";
-
-    int64_t t0 = 0;
-    int64_t t1 = 0;
-
-    // print the last n_new segments
-    const int s0 = n_segments - n_new;
-
-    if (s0 == 0) {
-        printf("\n");
-    }
-
-    for (int i = s0; i < n_segments; i++) {
-        if (!params.no_timestamps || params.diarize) {
-            t0 = whisper_full_get_segment_t0(ctx, i);
-            t1 = whisper_full_get_segment_t1(ctx, i);
-        }
-
-        if (!params.no_timestamps) {
-            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
-        }
-
-        if (params.diarize && pcmf32s.size() == 2) {
-            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
-        }
-
-        if (params.print_colors) {
-            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-                if (params.print_special == false) {
-                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-                    if (id >= whisper_token_eot(ctx)) {
-                        continue;
-                    }
-                }
-
-                const char * text = whisper_full_get_token_text(ctx, i, j);
-                const float  p    = whisper_full_get_token_p   (ctx, i, j);
-
-                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
-
-                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
-            }
-        } else {
-            const char * text = whisper_full_get_segment_text(ctx, i);
-
-            printf("%s%s", speaker.c_str(), text);
-        }
-
-        if (params.tinydiarize) {
-            if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
-                printf("%s", params.tdrz_speaker_turn.c_str());
-            }
-        }
-
-        // with timestamps or speakers: each segment on new line
-        if (!params.no_timestamps || params.diarize) {
-            printf("\n");
-        }
-        fflush(stdout);
-    }
-}
-
-std::string output_str(struct whisper_context * ctx, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
-    std::stringstream result;
-    const int n_segments = whisper_full_n_segments(ctx);
-    for (int i = 0; i < n_segments; ++i) {
-        const char * text = whisper_full_get_segment_text(ctx, i);
-        std::string speaker = "";
-
-        if (params.diarize && pcmf32s.size() == 2)
-        {
-            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
-        }
-
-        result << speaker << text << "\n";
-    }
-    return result.str();
-}
-
-bool parse_str_to_bool(const std::string & s) {
-    if (s == "true" || s == "1" || s == "yes" || s == "y") {
-        return true;
-    }
-    return false;
-}
-
-void get_req_parameters(const Request & req, whisper_params & params)
-{
-    if (req.has_file("offset_t"))
-    {
-        params.offset_t_ms = std::stoi(req.get_file_value("offset_t").content);
-    }
-    if (req.has_file("offset_n"))
-    {
-        params.offset_n = std::stoi(req.get_file_value("offset_n").content);
-    }
-    if (req.has_file("duration"))
-    {
-        params.duration_ms = std::stoi(req.get_file_value("duration").content);
-    }
-    if (req.has_file("max_context"))
-    {
-        params.max_context = std::stoi(req.get_file_value("max_context").content);
-    }
-    if (req.has_file("max_len"))
-    {
-        params.max_len = std::stoi(req.get_file_value("max_len").content);
-    }
-    if (req.has_file("best_of"))
-    {
-        params.best_of = std::stoi(req.get_file_value("best_of").content);
-    }
-    if (req.has_file("beam_size"))
-    {
-        params.beam_size = std::stoi(req.get_file_value("beam_size").content);
-    }
-    if (req.has_file("audio_ctx"))
-    {
-        params.audio_ctx = std::stof(req.get_file_value("audio_ctx").content);
-    }
-    if (req.has_file("word_thold"))
-    {
-        params.word_thold = std::stof(req.get_file_value("word_thold").content);
-    }
-    if (req.has_file("entropy_thold"))
-    {
-        params.entropy_thold = std::stof(req.get_file_value("entropy_thold").content);
-    }
-    if (req.has_file("logprob_thold"))
-    {
-        params.logprob_thold = std::stof(req.get_file_value("logprob_thold").content);
-    }
-    if (req.has_file("debug_mode"))
-    {
-        params.debug_mode = parse_str_to_bool(req.get_file_value("debug_mode").content);
-    }
-    if (req.has_file("translate"))
-    {
-        params.translate = parse_str_to_bool(req.get_file_value("translate").content);
-    }
-    if (req.has_file("diarize"))
-    {
-        params.diarize = parse_str_to_bool(req.get_file_value("diarize").content);
-    }
-    if (req.has_file("tinydiarize"))
-    {
-        params.tinydiarize = parse_str_to_bool(req.get_file_value("tinydiarize").content);
-    }
-    if (req.has_file("split_on_word"))
-    {
-        params.split_on_word = parse_str_to_bool(req.get_file_value("split_on_word").content);
-    }
-    if (req.has_file("no_timestamps"))
-    {
-        params.no_timestamps = parse_str_to_bool(req.get_file_value("no_timestamps").content);
-    }
-    if (req.has_file("language"))
-    {
-        params.language = req.get_file_value("language").content;
-    }
-    if (req.has_file("detect_language"))
-    {
-        params.detect_language = parse_str_to_bool(req.get_file_value("detect_language").content);
-    }
-    if (req.has_file("prompt"))
-    {
-        params.prompt = req.get_file_value("prompt").content;
-    }
-    if (req.has_file("response_format"))
-    {
-        params.response_format = req.get_file_value("response_format").content;
-    }
-    if (req.has_file("temperature"))
-    {
-        params.temperature = std::stof(req.get_file_value("temperature").content);
-    }
-    if (req.has_file("temperature_inc"))
-    {
-        params.temperature_inc = std::stof(req.get_file_value("temperature_inc").content);
-    }
-}
-
-}  // namespace
-
-int main(int argc, char ** argv) {
-    whisper_params params;
-    server_params sparams;
-
-    std::mutex whisper_mutex;
-
-    if (whisper_params_parse(argc, argv, params, sparams) == false) {
-        whisper_print_usage(argc, argv, params, sparams);
-        return 1;
-    }
-
-    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
-        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
-        whisper_print_usage(argc, argv, params, sparams);
-        exit(0);
-    }
-
-    if (params.diarize && params.tinydiarize) {
-        fprintf(stderr, "error: cannot use both --diarize and --tinydiarize\n");
-        whisper_print_usage(argc, argv, params, sparams);
-        exit(0);
-    }
-
-    if (sparams.ffmpeg_converter) {
-        check_ffmpeg_availibility();
-    }
-    // whisper init
-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
-
-    if (ctx == nullptr) {
-        fprintf(stderr, "error: failed to initialize whisper context\n");
-        return 3;
-    }
-
-    // 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);
-
-    Server svr;
-    svr.set_default_headers({{"Server", "whisper.cpp"},
-                             {"Access-Control-Allow-Origin", "*"},
-                             {"Access-Control-Allow-Headers", "content-type, authorization"}});
-
-    std::string const default_content = R"(
-    <html>
-    <head>
-        <title>Whisper.cpp Server</title>
-        <meta charset="utf-8">
-        <meta name="viewport" content="width=device-width">
-        <style>
-        body {
-            font-family: sans-serif;
-        }
-        form {
-            display: flex;
-            flex-direction: column;
-            align-items: flex-start;
-        }
-        label {
-            margin-bottom: 0.5rem;
-        }
-        input, select {
-            margin-bottom: 1rem;
-        }
-        button {
-            margin-top: 1rem;
-        }
-        </style>
-    </head>
-    <body>
-        <h1>Whisper.cpp Server</h1>
-
-        <h2>/inference</h2>
-        <pre>
-    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/inference \
-    -H "Content-Type: multipart/form-data" \
-    -F file="@&lt;file-path&gt;" \
-    -F temperature="0.0" \
-    -F temperature_inc="0.2" \
-    -F response_format="json"
-        </pre>
-
-        <h2>/load</h2>
-        <pre>
-    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/load \
-    -H "Content-Type: multipart/form-data" \
-    -F model="&lt;path-to-model-file&gt;"
-        </pre>
-
-        <div>
-            <h2>Try it out</h2>
-            <form action="/inference" method="POST" enctype="multipart/form-data">
-                <label for="file">Choose an audio file:</label>
-                <input type="file" id="file" name="file" accept="audio/*" required><br>
-
-                <label for="temperature">Temperature:</label>
-                <input type="number" id="temperature" name="temperature" value="0.0" step="0.01" placeholder="e.g., 0.0"><br>
-
-                <label for="response_format">Response Format:</label>
-                <select id="response_format" name="response_format">
-                    <option value="verbose_json">Verbose JSON</option>
-                    <option value="json">JSON</option>
-                    <option value="text">Text</option>
-                    <option value="srt">SRT</option>
-                    <option value="vtt">VTT</option>
-                </select><br>
-
-                <button type="submit">Submit</button>
-            </form>
-        </div>
-    </body>
-    </html>
-    )";
-
-    // store default params so we can reset after each inference request
-    whisper_params default_params = params;
-
-    // this is only called if no index.html is found in the public --path
-    svr.Get(sparams.request_path + "/", [&default_content](const Request &, Response &res){
-        res.set_content(default_content, "text/html");
-        return false;
-    });
-
-    svr.Options(sparams.request_path + "/inference", [&](const Request &, Response &){
-    });
-
-    svr.Post(sparams.request_path + "/inference", [&](const Request &req, Response &res){
-        // acquire whisper model mutex lock
-        std::lock_guard<std::mutex> lock(whisper_mutex);
-
-        // first check user requested fields of the request
-        if (!req.has_file("file"))
-        {
-            fprintf(stderr, "error: no 'file' field in the request\n");
-            const std::string error_resp = "{\"error\":\"no 'file' field in the request\"}";
-            res.set_content(error_resp, "application/json");
-            return;
-        }
-        auto audio_file = req.get_file_value("file");
-
-        // check non-required fields
-        get_req_parameters(req, params);
-
-        std::string filename{audio_file.filename};
-        printf("Received request: %s\n", filename.c_str());
-
-        // audio arrays
-        std::vector<float> pcmf32;               // mono-channel F32 PCM
-        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
-
-        if (sparams.ffmpeg_converter) {
-            // if file is not wav, convert to wav
-            // write to temporary file
-            const std::string temp_filename = "whisper_server_temp_file.wav";
-            std::ofstream temp_file{temp_filename, std::ios::binary};
-            temp_file << audio_file.content;
-            temp_file.close();
-
-            std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
-            const bool is_converted = convert_to_wav(temp_filename, error_resp);
-            if (!is_converted) {
-                res.set_content(error_resp, "application/json");
-                return;
-            }
-
-            // read wav content into pcmf32
-            if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize))
-            {
-                fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
-                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
-                res.set_content(error_resp, "application/json");
-                std::remove(temp_filename.c_str());
-                return;
-            }
-            // remove temp file
-            std::remove(temp_filename.c_str());
-        } else {
-            if (!::read_wav(audio_file.content, pcmf32, pcmf32s, params.diarize))
-            {
-                fprintf(stderr, "error: failed to read WAV file\n");
-                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
-                res.set_content(error_resp, "application/json");
-                return;
-            }
-        }
-
-
-        printf("Successfully loaded %s\n", filename.c_str());
-
-        // print system information
-        {
-            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
-        {
-            fprintf(stderr, "\n");
-            if (!whisper_is_multilingual(ctx)) {
-                if (params.language != "en" || params.translate) {
-                    params.language = "en";
-                    params.translate = false;
-                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
-                }
-            }
-            if (params.detect_language) {
-                params.language = "auto";
-            }
-            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
-                    __func__, filename.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.tinydiarize ? "tdrz = 1, " : "",
-                    params.no_timestamps ? 0 : 1);
-
-            fprintf(stderr, "\n");
-        }
-
-        // run the inference
-        {
-            printf("Running whisper.cpp inference on %s\n", filename.c_str());
-            whisper_full_params wparams = whisper_full_default_params(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;
-            wparams.print_timestamps = !params.no_timestamps;
-            wparams.print_special    = params.print_special;
-            wparams.translate        = params.translate;
-            wparams.language         = params.language.c_str();
-            wparams.detect_language  = params.detect_language;
-            wparams.n_threads        = params.n_threads;
-            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
-            wparams.offset_ms        = params.offset_t_ms;
-            wparams.duration_ms      = params.duration_ms;
-
-            wparams.thold_pt         = params.word_thold;
-            wparams.max_len          = params.max_len == 0 ? 60 : params.max_len;
-            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.initial_prompt   = params.prompt.c_str();
-
-            wparams.greedy.best_of        = params.best_of;
-            wparams.beam_search.beam_size = params.beam_size;
-
-            wparams.temperature      = params.temperature;
-            wparams.temperature_inc  = params.temperature_inc;
-            wparams.entropy_thold    = params.entropy_thold;
-            wparams.logprob_thold    = params.logprob_thold;
-
-            wparams.no_timestamps    = params.no_timestamps;
-            wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
-
-            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
-
-            // this callback is called on each new segment
-            if (params.print_realtime) {
-                wparams.new_segment_callback           = whisper_print_segment_callback;
-                wparams.new_segment_callback_user_data = &user_data;
-            }
-
-            if (wparams.print_progress) {
-                wparams.progress_callback           = whisper_print_progress_callback;
-                wparams.progress_callback_user_data = &user_data;
-            }
-
-            // examples for abort mechanism
-            // in examples below, we do not abort the processing, but we could if the flag is set to true
-
-            // the callback is called before every encoder run - if it returns false, the processing is aborted
-            {
-                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
-
-                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
-                    bool is_aborted = *(bool*)user_data;
-                    return !is_aborted;
-                };
-                wparams.encoder_begin_callback_user_data = &is_aborted;
-            }
-
-            // the callback is called before every computation - if it returns true, the computation is aborted
-            {
-                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
-
-                wparams.abort_callback = [](void * user_data) {
-                    bool is_aborted = *(bool*)user_data;
-                    return is_aborted;
-                };
-                wparams.abort_callback_user_data = &is_aborted;
-            }
-
-            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
-                fprintf(stderr, "%s: failed to process audio\n", argv[0]);
-                const std::string error_resp = "{\"error\":\"failed to process audio\"}";
-                res.set_content(error_resp, "application/json");
-                return;
-            }
-        }
-
-        // return results to user
-        if (params.response_format == text_format)
-        {
-            std::string results = output_str(ctx, params, pcmf32s);
-            res.set_content(results.c_str(), "text/html");
-        }
-        else if (params.response_format == srt_format)
-        {
-            std::stringstream ss;
-            const int n_segments = whisper_full_n_segments(ctx);
-            for (int i = 0; i < n_segments; ++i) {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-                std::string speaker = "";
-
-                if (params.diarize && pcmf32s.size() == 2)
-                {
-                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
-                }
-
-                ss << i + 1 + params.offset_n << "\n";
-                ss << to_timestamp(t0, true) << " --> " << to_timestamp(t1, true) << "\n";
-                ss << speaker << text << "\n\n";
-            }
-            res.set_content(ss.str(), "application/x-subrip");
-        } else if (params.response_format == vtt_format) {
-            std::stringstream ss;
-
-            ss << "WEBVTT\n\n";
-
-            const int n_segments = whisper_full_n_segments(ctx);
-            for (int i = 0; i < n_segments; ++i) {
-                const char * text = whisper_full_get_segment_text(ctx, i);
-                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-                std::string speaker = "";
-
-                if (params.diarize && pcmf32s.size() == 2)
-                {
-                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1, true);
-                    speaker.insert(0, "<v Speaker");
-                    speaker.append(">");
-                }
-
-                ss << to_timestamp(t0) << " --> " << to_timestamp(t1) << "\n";
-                ss << speaker << text << "\n\n";
-            }
-            res.set_content(ss.str(), "text/vtt");
-        } else if (params.response_format == vjson_format) {
-            /* try to match openai/whisper's Python format */
-            std::string results = output_str(ctx, params, pcmf32s);
-            json jres = json{
-                {"task", params.translate ? "translate" : "transcribe"},
-                {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
-                {"duration", float(pcmf32.size())/WHISPER_SAMPLE_RATE},
-                {"text", results},
-                {"segments", json::array()}
-            };
-            const int n_segments = whisper_full_n_segments(ctx);
-            for (int i = 0; i < n_segments; ++i)
-            {
-                json segment = json{
-                    {"id", i},
-                    {"text", whisper_full_get_segment_text(ctx, i)},
-                };
-
-                if (!params.no_timestamps) {
-                    segment["start"] = whisper_full_get_segment_t0(ctx, i) * 0.01;
-                    segment["end"] = whisper_full_get_segment_t1(ctx, i) * 0.01;
-                }
-
-                float total_logprob = 0;
-                const int n_tokens = whisper_full_n_tokens(ctx, i);
-                for (int j = 0; j < n_tokens; ++j) {
-                    whisper_token_data token = whisper_full_get_token_data(ctx, i, j);
-                    if (token.id >= whisper_token_eot(ctx)) {
-                        continue;
-                    }
-
-                    segment["tokens"].push_back(token.id);
-                    json word = json{{"word", whisper_full_get_token_text(ctx, i, j)}};
-                    if (!params.no_timestamps) {
-                        word["start"] = token.t0 * 0.01;
-                        word["end"] = token.t1 * 0.01;
-                    }
-                    word["probability"] = token.p;
-                    total_logprob += token.plog;
-                    segment["words"].push_back(word);
-                }
-
-                segment["temperature"] = params.temperature;
-                segment["avg_logprob"] = total_logprob / n_tokens;
-
-                // TODO compression_ratio and no_speech_prob are not implemented yet
-                // segment["compression_ratio"] = 0;
-                // segment["no_speech_prob"] = 0;
-
-                jres["segments"].push_back(segment);
-            }
-            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
-                            "application/json");
-        }
-        // TODO add more output formats
-        else
-        {
-            std::string results = output_str(ctx, params, pcmf32s);
-            json jres = json{
-                {"text", results}
-            };
-            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
-                            "application/json");
-        }
-
-        // reset params to thier defaults
-        params = default_params;
-    });
-    svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
-        std::lock_guard<std::mutex> lock(whisper_mutex);
-        if (!req.has_file("model"))
-        {
-            fprintf(stderr, "error: no 'model' field in the request\n");
-            const std::string error_resp = "{\"error\":\"no 'model' field in the request\"}";
-            res.set_content(error_resp, "application/json");
-            return;
-        }
-        std::string model = req.get_file_value("model").content;
-        if (!is_file_exist(model.c_str()))
-        {
-            fprintf(stderr, "error: 'model': %s not found!\n", model.c_str());
-            const std::string error_resp = "{\"error\":\"model not found!\"}";
-            res.set_content(error_resp, "application/json");
-            return;
-        }
-
-        // clean up
-        whisper_free(ctx);
-
-        // whisper init
-        ctx = whisper_init_from_file_with_params(model.c_str(), cparams);
-
-        // TODO perhaps load prior model here instead of exit
-        if (ctx == nullptr) {
-            fprintf(stderr, "error: model init  failed, no model loaded must exit\n");
-            exit(1);
-        }
-
-        // 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);
-
-        const std::string success = "Load was successful!";
-        res.set_content(success, "application/text");
-
-        // check if the model is in the file system
-    });
-
-    svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) {
-        const char fmt[] = "500 Internal Server Error\n%s";
-        char buf[BUFSIZ];
-        try {
-            std::rethrow_exception(std::move(ep));
-        } catch (std::exception &e) {
-            snprintf(buf, sizeof(buf), fmt, e.what());
-        } catch (...) {
-            snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
-        }
-        res.set_content(buf, "text/plain");
-        res.status = 500;
-    });
-
-    svr.set_error_handler([](const Request &req, Response &res) {
-        if (res.status == 400) {
-            res.set_content("Invalid request", "text/plain");
-        } else if (res.status != 500) {
-            res.set_content("File Not Found (" + req.path + ")", "text/plain");
-            res.status = 404;
-        }
-    });
-
-    // set timeouts and change hostname and port
-    svr.set_read_timeout(sparams.read_timeout);
-    svr.set_write_timeout(sparams.write_timeout);
-
-    if (!svr.bind_to_port(sparams.hostname, sparams.port))
-    {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n",
-                sparams.hostname.c_str(), sparams.port);
-        return 1;
-    }
-
-    // Set the base directory for serving static files
-    svr.set_base_dir(sparams.public_path);
-
-    // to make it ctrl+clickable:
-    printf("\nwhisper server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
-
-    if (!svr.listen_after_bind())
-    {
-        return 1;
-    }
-
-    whisper_print_timings(ctx);
-    whisper_free(ctx);
-
-    return 0;
-}
--- a/examples/stream.wasm/emscripten.cpp
+++ b/examples/stream.wasm/emscripten.cpp
@ -103,11 +103,11 @@ void stream_main(size_t index) {

            {
                const int n_segments = whisper_full_n_segments(ctx);
-                if (n_segments > 0) {
-                    const char * text = whisper_full_get_segment_text(ctx, n_segments - 1);
+                for (int i = n_segments - 1; i < n_segments; ++i) {
+                    const char * text = whisper_full_get_segment_text(ctx, i);

-                    const int64_t t0 = whisper_full_get_segment_t0(ctx, n_segments - 1);
-                    const int64_t t1 = whisper_full_get_segment_t1(ctx, n_segments - 1);
+                    const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                    const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

                    printf("transcribed: %s\n", text);

@ -132,7 +132,7 @@ EMSCRIPTEN_BINDINGS(stream) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        for (size_t i = 0; i < g_contexts.size(); ++i) {
            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                if (g_contexts[i] != nullptr) {
                    g_running = true;
                    if (g_worker.joinable()) {
--- a/examples/stream/README.md
+++ b/examples/stream/README.md
@ -4,7 +4,7 @@ This is a naive example of performing real-time inference on audio from your mic
 The `stream` tool samples the audio every half a second and runs the transcription continously.
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).

-```bash
+```java
 ./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```

@ -14,7 +14,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a

 Setting the `--step` argument to `0` enables the sliding window mode:

-```bash
+```java
 ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```

@ -39,20 +39,6 @@ brew install sdl2
 make stream
 ```

-Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
-`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
-
-```bash
-whisper.cpp/examples/stream$ make stream
-g++     stream.cpp   -o stream
-stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
-    6 | #include "common/sdl.h"
-      |          ^~~~~~~~~~~~~~
-compilation terminated.
-make: *** [<builtin>: stream] Error 1
-```
-
 ## Web version

 This tool can also run in the browser: [examples/stream.wasm](/examples/stream.wasm)
--- a/examples/stream/stream.cpp
+++ b/examples/stream/stream.cpp
@ -2,8 +2,9 @@
 //
 // A very quick-n-dirty implementation serving mainly as a proof of concept.
 //
-#include "common-sdl.h"
+
 #include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"

 #include <cassert>
@ -13,6 +14,19 @@
 #include <vector>
 #include <fstream>

+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t) {
+    int64_t sec = t/100;
+    int64_t msec = t - sec*100;
+    int64_t min = sec/60;
+    sec = sec - min*60;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
+
+    return std::string(buf);
+}

 // command-line parameters
 struct whisper_params {
@ -34,8 +48,6 @@ struct whisper_params {
    bool no_context    = true;
    bool no_timestamps = false;
    bool tinydiarize   = false;
-    bool save_audio    = false; // save audio to wav file
-    bool use_gpu       = true;

    std::string language  = "en";
    std::string model     = "models/ggml-base.en.bin";
@ -52,26 +64,24 @@ 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 == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
-        else if (                  arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
-        else if (                  arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
-        else if (arg == "-c"    || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
-        else if (arg == "-mt"   || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
-        else if (arg == "-ac"   || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
-        else if (arg == "-vth"  || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
-        else if (arg == "-fth"  || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
-        else if (arg == "-su"   || arg == "--speed-up")      { params.speed_up      = true; }
-        else if (arg == "-tr"   || arg == "--translate")     { params.translate     = true; }
-        else if (arg == "-nf"   || arg == "--no-fallback")   { params.no_fallback   = true; }
-        else if (arg == "-ps"   || arg == "--print-special") { params.print_special = true; }
-        else if (arg == "-kc"   || arg == "--keep-context")  { params.no_context    = false; }
-        else if (arg == "-l"    || arg == "--language")      { params.language      = argv[++i]; }
-        else if (arg == "-m"    || arg == "--model")         { params.model         = argv[++i]; }
-        else if (arg == "-f"    || arg == "--file")          { params.fname_out     = argv[++i]; }
-        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 == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (                 arg == "--step")          { params.step_ms       = std::stoi(argv[++i]); }
+        else if (                 arg == "--length")        { params.length_ms     = std::stoi(argv[++i]); }
+        else if (                 arg == "--keep")          { params.keep_ms       = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-nf"  || arg == "--no-fallback")   { params.no_fallback   = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-kc"  || arg == "--keep-context")  { params.no_context    = false; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")  { params.tinydiarize   = true; }

        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -106,9 +116,7 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                                params.language.c_str());
    fprintf(stderr, "  -m FNAME, --model FNAME   [%-7s] model path\n",                                     params.model.c_str());
    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                          params.fname_out.c_str());
-    fprintf(stderr, "  -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, "  -tdrz,     --tinydiarize  [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
    fprintf(stderr, "\n");
 }

@ -146,16 +154,14 @@ int main(int argc, char ** argv) {
    audio.resume();

    // whisper init
+
    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1){
        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
        whisper_print_usage(argc, argv, params);
        exit(0);
    }

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+    struct whisper_context * ctx = whisper_init_from_file(params.model.c_str());

    std::vector<float> pcmf32    (n_samples_30s, 0.0f);
    std::vector<float> pcmf32_old;
@ -206,28 +212,14 @@ int main(int argc, char ** argv) {
        }
    }

-    wav_writer wavWriter;
-    // save wav file
-    if (params.save_audio) {
-        // Get current date/time for filename
-        time_t now = time(0);
-        char buffer[80];
-        strftime(buffer, sizeof(buffer), "%Y%m%d%H%M%S", localtime(&now));
-        std::string filename = std::string(buffer) + ".wav";
-
-        wavWriter.open(filename, WHISPER_SAMPLE_RATE, 16, 1);
-    }
-    printf("[Start speaking]\n");
+    printf("[Start speaking]");
    fflush(stdout);

-    auto t_last  = std::chrono::high_resolution_clock::now();
+          auto t_last  = std::chrono::high_resolution_clock::now();
    const auto t_start = t_last;

    // main audio loop
    while (is_running) {
-        if (params.save_audio) {
-            wavWriter.write(pcmf32_new.data(), pcmf32_new.size());
-        }
        // handle Ctrl + C
        is_running = sdl_poll_events();

@ -358,7 +350,7 @@ int main(int argc, char ** argv) {
                        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
                        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);

-                        std::string output = "[" + to_timestamp(t0, false) + " --> " + to_timestamp(t1, false) + "]  " + text;
+                        std::string output = "[" + to_timestamp(t0) + " --> " + to_timestamp(t1) + "]  " + text;

                        if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
                            output += " [SPEAKER_TURN]";
@ -379,7 +371,7 @@ int main(int argc, char ** argv) {
                    fout << std::endl;
                }

-                if (use_vad) {
+                if (use_vad){
                    printf("\n");
                    printf("### Transcription %d END\n", n_iter);
                }
--- a/examples/sycl/CMakeLists.txt
+++ b/examples/sycl/CMakeLists.txt
@ -1,9 +0,0 @@
-#  MIT license
-#  Copyright (C) 2024 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-set(TARGET ls-sycl-device)
-add_executable(${TARGET} ls-sycl-device.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)
--- a/examples/sycl/README.md
+++ b/examples/sycl/README.md
@ -1,47 +0,0 @@
-# llama.cpp/example/sycl
-
-This example program provide the tools for llama.cpp for SYCL on Intel GPU.
-
-## Tool
-
-|Tool Name| Function|Status|
-|-|-|-|
-|ls-sycl-device| List all SYCL devices with ID, compute capability, max work group size, ect.|Support|
-
-### ls-sycl-device
-
-List all SYCL devices with ID, compute capability, max work group size, ect.
-
-1. Build the llama.cpp for SYCL for all targets.
-
-2. Enable oneAPI running environment
-
-```
-source /opt/intel/oneapi/setvars.sh
-```
-
-3. Execute
-
-```
-./build/bin/ls-sycl-device
-```
-
-Check the ID in startup log, like:
-
-```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
-```
-
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
--- a/examples/sycl/build.sh
+++ b/examples/sycl/build.sh
@ -1,19 +0,0 @@
-#  MIT license
-#  Copyright (C) 2024 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DWHISPER_SYCL_F16=ON # faster for long-prompt inference
-
-#for FP32
-cmake .. -DWHISPER_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v
--- a/examples/sycl/ls-sycl-device.cpp
+++ b/examples/sycl/ls-sycl-device.cpp
@ -1,11 +0,0 @@
-/*MIT license
-  Copyright (C) 2024 Intel Corporation
-  SPDX-License-Identifier: MIT
-*/
-
-#include "ggml-sycl.h"
-
-int main(int argc, char ** argv) {
-    ggml_backend_sycl_print_sycl_devices();
-    return 0;
-}
--- a/examples/sycl/run-whisper.sh
+++ b/examples/sycl/run-whisper.sh
@ -1,17 +0,0 @@
-#!/bin/bash
-
-#  MIT license
-#  Copyright (C) 2024 Intel Corporation
-#  SPDX-License-Identifier: MIT
-
-INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
-source /opt/intel/oneapi/setvars.sh
-
-if [ $# -gt 0 ]; then
-    export GGML_SYCL_DEVICE=$1
-else
-    export GGML_SYCL_DEVICE=0
-fi
-echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
-#export GGML_SYCL_DEBUG=1
-./build/bin/main -m models/ggml-base.en.bin -f samples/jfk.wav
--- a/examples/talk-llama/.gitignore
+++ b/examples/talk-llama/.gitignore
@ -1,2 +1 @@
 audio.mp3
-to_speak.txt
--- a/examples/talk-llama/CMakeLists.txt
+++ b/examples/talk-llama/CMakeLists.txt
@ -1,18 +1,16 @@
 if (WHISPER_SDL2)
    # talk-llama
    set(TARGET talk-llama)
-    add_executable(${TARGET} talk-llama.cpp llama.cpp unicode.cpp)
-    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #add_executable(${TARGET} talk-llama.cpp llama.cpp)
+    #target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
+    #target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

-    if (WHISPER_CLBLAST)
-        set(CLBLAST_LIBNAME clblast)
-    endif ()
-    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CLBLAST_LIBNAME} ${CMAKE_THREAD_LIBS_INIT})
+    # TODO: this is temporary
+    #       need to export ggml symbols for MSVC, but too lazy ..
+    add_executable(${TARGET} talk-llama.cpp llama.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)

-    if(WIN32)
-        # It requires Windows 8.1 or later for PrefetchVirtualMemory
-        target_compile_definitions(${TARGET} PRIVATE -D_WIN32_WINNT=0x0602)
-    endif()
+    target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
+    target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})

    include(DefaultTargetOptions)
 endif ()
--- a/examples/talk-llama/README.md
+++ b/examples/talk-llama/README.md
@ -2,12 +2,6 @@

 Talk with an LLaMA AI in your terminal

-*Latest perf as of 2 Nov 2023 using Whisper Medium + LLaMA v2 13B Q8_0 on M2 Ultra:*
-
-https://github.com/ggerganov/whisper.cpp/assets/1991296/d97a3788-bf2a-4756-9a43-60c6b391649e
-
-*Previous demo running on CPUs*
-
 [Demo Talk](https://user-images.githubusercontent.com/1991296/228024237-848f998c-c334-46a6-bef8-3271590da83b.mp4)

 ## Building
@ -25,7 +19,7 @@ brew install sdl2
 make talk-llama

 # Run it
-./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
+./talk-llama -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
 ```

 - The `-mw` argument specifies the Whisper model that you would like to use. Recommended `base` or `small` for real-time experience
@ -42,7 +36,7 @@ This feature is especially helpful for maintaining context in long conversations
 Example usage:

 ```bash
-./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/llama-13b/ggml-model-q4_0.gguf -p "Georgi" -t 8
+./talk-llama --session ./my-session-file -mw ./models/ggml-small.en.bin -ml ../llama.cpp/models/13B/ggml-model-q4_0.bin -p "Georgi" -t 8
 ```

 ## TTS
--- a/examples/talk-llama/eleven-labs.py
+++ b/examples/talk-llama/eleven-labs.py
@ -1,80 +1,20 @@
 import sys
-import argparse
-import textwrap
+import importlib.util

-parser = argparse.ArgumentParser(add_help=False,
-    formatter_class=argparse.RawTextHelpFormatter)
-parser.add_argument("-q", "--quick", action="store_true",
-    help="skip checking the required library")
-
-modes = parser.add_argument_group("action")
-modes.add_argument("inputfile", metavar="TEXTFILE",
-    nargs='?', type=argparse.FileType(), default=sys.stdin,
-    help="read the text file (default: stdin)")
-modes.add_argument("-l", "--list", action="store_true",
-    help="show the list of voices and exit")
-modes.add_argument("-h", "--help", action="help",
-    help="show this help and exit")
-
-selopts = parser.add_argument_group("voice selection")
-selmodes = selopts.add_mutually_exclusive_group()
-selmodes.add_argument("-n", "--name",
-    default="Arnold",
-    help="get a voice object by name (default: Arnold)")
-selmodes.add_argument("-v", "--voice", type=int, metavar="NUMBER",
-    help="get a voice object by number (see --list)")
-selopts.add_argument("-f", "--filter", action="append", metavar="KEY=VAL",
-    default=["use case=narration"],
-    help=textwrap.dedent('''\
-        filter voices by labels (default: "use case=narration")
-        this option can be used multiple times
-        filtering will be disabled if the first -f has no "=" (e.g. -f "any")
-        '''))
-
-outmodes = parser.add_argument_group("output")
-outgroup = outmodes.add_mutually_exclusive_group()
-outgroup.add_argument("-s", "--save", metavar="FILE",
-    default="audio.mp3",
-    help="save the TTS to a file (default: audio.mp3)")
-outgroup.add_argument("-p", "--play", action="store_true",
-    help="play the TTS with ffplay")
-
-args = parser.parse_args()
-
-if not args.quick:
-    import importlib.util
-    if importlib.util.find_spec("elevenlabs") is None:
-        print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
-        sys.exit()
-
-from elevenlabs import voices, generate, play, save
-
-if args.filter and "=" in args.filter[0]:
-    voicelist = voices()
-    for f in args.filter:
-        label, value = f.split("=")
-        voicelist = filter(lambda x: x.labels.get(label) == value, voicelist)
-    voicelist = list(voicelist)
-else:
-    voicelist = list(voices())
-
-if args.list:
-    for i, v in enumerate(voicelist):
-        print(str(i) + ": " + v.name + " " + str(v.labels))
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
    sys.exit()

-if args.voice:
-    voice = voicelist[args.voice % len(voicelist)]
-else:
-    voice = args.name
-    # if -n should consult -f, use the following
-    #voice = next(x for x in voicelist if x.name == args.name)
+from elevenlabs import generate, play, save

+# Get a Voice object, by name or UUID
+voice = "Arnold" #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
 audio = generate(
-    text=str(args.inputfile.read()),
-    voice=voice
+  text=str(sys.argv[2:]),
+  voice=voice
 )
-if args.play:
-    play(audio)
-else:
-    save(audio, args.save) 
+
+# Save the TTS to a file
+save(audio, "audio.mp3") 
--- a/examples/talk-llama/llama-util.h
+++ b/examples/talk-llama/llama-util.h
@ -0,0 +1,474 @@
+// Internal header to be included only by llama.cpp.
+// Contains wrappers around OS interfaces.
+
+#ifndef LLAMA_UTIL_H
+#define LLAMA_UTIL_H
+
+#include <cstdio>
+#include <cstdint>
+#include <cerrno>
+#include <cstring>
+#include <cstdarg>
+#include <cstdlib>
+#include <climits>
+
+#include <string>
+#include <vector>
+#include <stdexcept>
+
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/mman.h>
+        #endif
+        #if defined(_POSIX_MEMLOCK_RANGE)
+            #include <sys/resource.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <io.h>
+    #include <stdio.h> // for _fseeki64
+#endif
+
+#define LLAMA_ASSERT(x) \
+    do { \
+        if (!(x)) { \
+            fprintf(stderr, "LLAMA_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
+            abort(); \
+        } \
+    } while (0)
+
+#ifdef __GNUC__
+#ifdef __MINGW32__
+__attribute__((format(gnu_printf, 1, 2)))
+#else
+__attribute__((format(printf, 1, 2)))
+#endif
+#endif
+static std::string format(const char * fmt, ...) {
+    va_list ap, ap2;
+    va_start(ap, fmt);
+    va_copy(ap2, ap);
+    int size = vsnprintf(NULL, 0, fmt, ap);
+    LLAMA_ASSERT(size >= 0 && size < INT_MAX);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    LLAMA_ASSERT(size2 == size);
+    va_end(ap2);
+    va_end(ap);
+    return std::string(buf.data(), size);
+}
+
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
+
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            throw std::runtime_error(format("failed to open %s: %s", fname, strerror(errno)));
+        }
+        seek(0, SEEK_END);
+        size = tell();
+        seek(0, SEEK_SET);
+    }
+
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        LLAMA_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
+    }
+
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        LLAMA_ASSERT(ret == 0); // same
+    }
+
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        std::size_t ret = std::fread(ptr, len, 1, fp);
+        if (ferror(fp)) {
+            throw std::runtime_error(format("read error: %s", strerror(errno)));
+        }
+        if (ret != 1) {
+            throw std::runtime_error(std::string("unexpectedly reached end of file"));
+        }
+    }
+
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, len, 1, fp);
+        if (ret != 1) {
+            throw std::runtime_error(format("write error: %s", strerror(errno)));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
+    }
+};
+
+#if defined(_WIN32)
+static std::string llama_format_win_err(DWORD err) {
+    LPSTR buf;
+    size_t size = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
+                                 NULL, err, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&buf, 0, NULL);
+    if (!size) {
+        return "FormatMessageA failed";
+    }
+    std::string ret(buf, size);
+    LocalFree(buf);
+    return ret;
+}
+#endif
+
+struct llama_mmap {
+    void * addr;
+    size_t size;
+
+    llama_mmap(const llama_mmap &) = delete;
+
+#ifdef _POSIX_MAPPED_FILES
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
+        size = file->size;
+        int fd = fileno(file->fp);
+        int flags = MAP_SHARED;
+#ifdef __linux__
+        flags |= MAP_POPULATE;
+#endif
+        addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
+        if (addr == MAP_FAILED) {
+            throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
+        }
+
+        if (prefetch > 0) {
+            // Advise the kernel to preload the mapped memory
+            if (posix_madvise(addr, std::min(file->size, prefetch), POSIX_MADV_WILLNEED)) {
+                fprintf(stderr, "warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
+                        strerror(errno));
+            }
+        }
+    }
+
+    ~llama_mmap() {
+        munmap(addr, size);
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    llama_mmap(struct llama_file * file, bool prefetch = true) {
+        size = file->size;
+
+        HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
+
+        HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+        DWORD error = GetLastError();
+
+        if (hMapping == NULL) {
+            throw std::runtime_error(format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+        error = GetLastError();
+        CloseHandle(hMapping);
+
+        if (addr == NULL) {
+            throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str()));
+        }
+
+        #if _WIN32_WINNT >= _WIN32_WINNT_WIN8
+        if (prefetch) {
+            // Advise the kernel to preload the mapped memory
+            WIN32_MEMORY_RANGE_ENTRY range;
+            range.VirtualAddress = addr;
+            range.NumberOfBytes = (SIZE_T)size;
+            if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
+                fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+            }
+        }
+        #else
+        #pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
+        #endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
+    }
+
+    ~llama_mmap() {
+        if (!UnmapViewOfFile(addr)) {
+            fprintf(stderr, "warning: UnmapViewOfFile failed: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    llama_mmap(struct llama_file *, bool prefetch = true) {
+        (void)prefetch;
+        throw std::runtime_error(std::string("mmap not supported"));
+    }
+#endif
+};
+
+// Represents some region of memory being locked using mlock or VirtualLock;
+// will automatically unlock on destruction.
+struct llama_mlock {
+    void * addr = NULL;
+    size_t size = 0;
+    bool failed_already = false;
+
+    llama_mlock() {}
+    llama_mlock(const llama_mlock &) = delete;
+
+    ~llama_mlock() {
+        if (size) {
+            raw_unlock(addr, size);
+        }
+    }
+
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
+    }
+
+    void grow_to(size_t target_size) {
+        LLAMA_ASSERT(addr);
+        if (failed_already) {
+            return;
+        }
+        size_t granularity = lock_granularity();
+        target_size = (target_size + granularity - 1) & ~(granularity - 1);
+        if (target_size > size) {
+            if (raw_lock((uint8_t *) addr + size, target_size - size)) {
+                size = target_size;
+            } else {
+                failed_already = true;
+            }
+        }
+    }
+
+#ifdef _POSIX_MEMLOCK_RANGE
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        return (size_t) sysconf(_SC_PAGESIZE);
+    }
+
+    #ifdef __APPLE__
+        #define MLOCK_SUGGESTION \
+            "Try increasing the sysctl values 'vm.user_wire_limit' and 'vm.global_user_wire_limit' and/or " \
+            "decreasing 'vm.global_no_user_wire_amount'.  Also try increasing RLIMIT_MLOCK (ulimit -l).\n"
+    #else
+        #define MLOCK_SUGGESTION \
+            "Try increasing RLIMIT_MLOCK ('ulimit -l' as root).\n"
+    #endif
+
+    bool raw_lock(const void * addr, size_t size) {
+        if (!mlock(addr, size)) {
+            return true;
+        } else {
+            char* errmsg = std::strerror(errno);
+            bool suggest = (errno == ENOMEM);
+
+            // Check if the resource limit is fine after all
+            struct rlimit lock_limit;
+            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
+                suggest = false;
+            if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
+                suggest = false;
+
+            fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
+                    size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
+            return false;
+        }
+    }
+
+    #undef MLOCK_SUGGESTION
+
+    void raw_unlock(void * addr, size_t size) {
+        if (munlock(addr, size)) {
+            fprintf(stderr, "warning: failed to munlock buffer: %s\n", std::strerror(errno));
+        }
+    }
+#elif defined(_WIN32)
+    static constexpr bool SUPPORTED = true;
+
+    size_t lock_granularity() {
+        SYSTEM_INFO si;
+        GetSystemInfo(&si);
+        return (size_t) si.dwPageSize;
+    }
+
+    bool raw_lock(void * ptr, size_t len) {
+        for (int tries = 1; ; tries++) {
+            if (VirtualLock(ptr, len)) {
+                return true;
+            }
+            if (tries == 2) {
+                fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
+                    len, size, llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+
+            // It failed but this was only the first try; increase the working
+            // set size and try again.
+            SIZE_T min_ws_size, max_ws_size;
+            if (!GetProcessWorkingSetSize(GetCurrentProcess(), &min_ws_size, &max_ws_size)) {
+                fprintf(stderr, "warning: GetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+            // Per MSDN: "The maximum number of pages that a process can lock
+            // is equal to the number of pages in its minimum working set minus
+            // a small overhead."
+            // Hopefully a megabyte is enough overhead:
+            size_t increment = len + 1048576;
+            // The minimum must be <= the maximum, so we need to increase both:
+            min_ws_size += increment;
+            max_ws_size += increment;
+            if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
+                fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
+                        llama_format_win_err(GetLastError()).c_str());
+                return false;
+            }
+        }
+    }
+
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
+            fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
+                    llama_format_win_err(GetLastError()).c_str());
+        }
+    }
+#else
+    static constexpr bool SUPPORTED = false;
+
+    size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t len) {
+        fprintf(stderr, "warning: mlock not supported on this system\n");
+        return false;
+    }
+
+    void raw_unlock(const void * addr, size_t len) {}
+#endif
+};
+
+// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    llama_buffer() = default;
+
+    void resize(size_t len) {
+        delete[] addr;
+        addr = new uint8_t[len];
+        size = len;
+    }
+
+    ~llama_buffer() {
+        delete[] addr;
+    }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
+};
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+struct llama_ctx_buffer {
+    uint8_t * addr = NULL;
+    bool is_cuda;
+    size_t size = 0;
+
+    llama_ctx_buffer() = default;
+
+    void resize(size_t size) {
+        free();
+
+        addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        if (addr) {
+            is_cuda = true;
+        }
+        else {
+            // fall back to pageable memory
+            addr = new uint8_t[size];
+            is_cuda = false;
+        }
+        this->size = size;
+    }
+
+    void free() {
+        if (addr) {
+            if (is_cuda) {
+                ggml_cuda_host_free(addr);
+            }
+            else {
+                delete[] addr;
+            }
+        }
+        addr = NULL;
+    }
+
+    ~llama_ctx_buffer() {
+        free();
+    }
+
+    // disable copy and move
+    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
+    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
+};
+#else
+typedef llama_buffer llama_ctx_buffer;
+#endif
+
+#endif
--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
--- a/examples/talk-llama/speak
+++ b/examples/talk-llama/speak
@ -1,40 +1,24 @@
 #!/bin/bash

 # Usage:
-#  speak <voice_id> <textfile>
+#  speak.sh <voice_id> <text-to-speak>

-function installed() { command -v $1 >/dev/null 2>&1; }
-
-if installed espeak; then
-  espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 -f $2
-
-elif installed piper && installed aplay; then
-  cat $2 | piper --model ~/en_US-lessac-medium.onnx --output-raw | aplay -q -r 22050 -f S16_LE -t raw -
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 "$2"

 # for Mac
-elif installed say; then
-  say -f $2
+say "$2"

 # Eleven Labs
-elif installed python3 && \
-  python3 -c 'import importlib.util; exit(not importlib.util.find_spec("elevenlabs"))' && \
-  installed ffplay; then
-    # It's possible to use the API for free with limited number of characters.
-    # To increase this limit register to https://beta.elevenlabs.io to get an api key
-    # and paste it after 'ELEVEN_API_KEY='
-    # Keep the line commented to use the free version without api key
-    #export ELEVEN_API_KEY=your_api_key
-    wd=$(dirname $0)
-    script=$wd/eleven-labs.py
-    python3 $script -q -p -v $1 $2 >/dev/null 2>&1
-
-    # Uncomment to keep the audio file
-    #python3 $script -q -s ./audio.mp3 -v $1 $2 >/dev/null 2>&1
-    #ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
-
-else
-  echo 'Install espeak ("brew install espeak" or "apt-get install espeak"),'
-  echo 'piper ("pip install piper-tts" or https://github.com/rhasspy/piper) with aplay,'
-  echo 'or elevenlabs ("pip install elevenlabs") with ffplay.'
-  echo '(export ELEVEN_API_KEY if you have an api key from https://beta.elevenlabs.io)'
-fi
+# To use it, install the elevenlabs module from pip (pip install elevenlabs)
+# It's possible to use the API for free with limited number of characters. To increase this limit register to https://beta.elevenlabs.io to get an api key and paste it after 'ELEVEN_API_KEY='
+#Keep the line commented to use the free version whitout api key
+#
+#export ELEVEN_API_KEY=your_api_key
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2" >/dev/null 2>&1
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
--- a/examples/talk-llama/speak.bat
+++ b/examples/talk-llama/speak.bat
@ -1 +1 @@
-@powershell -ExecutionPolicy Bypass -F examples\talk-llama\speak.ps1 %1 %2
+@powershell -ExecutionPolicy Bypass -F examples\talk\speak.ps1 %1 %2
--- a/examples/talk-llama/speak.ps1
+++ b/examples/talk-llama/speak.ps1
@ -1,14 +1,12 @@
 # Set-ExecutionPolicy -ExecutionPolicy Bypass -Scope CurrentUser
 param(
-  [Parameter(Mandatory=$true)][int]$voicenum,
-  [Parameter(Mandatory=$true)][string]$textfile
+  # voice options are David or Zira
+  [Parameter(Mandatory=$true)][string]$voice,
+  [Parameter(Mandatory=$true)][string]$text
 )

 Add-Type -AssemblyName System.Speech;
 $speak = New-Object System.Speech.Synthesis.SpeechSynthesizer;
-$voiceoptions = $speak.GetInstalledVoices("en-US");
-$voice = $voiceoptions[$voicenum % $voiceoptions.count];
-$speak.SelectVoice($voice.VoiceInfo.Name);
+$speak.SelectVoice("Microsoft $voice Desktop");
 $speak.Rate="0";
-$text = Get-Content -Path $textfile;
 $speak.Speak($text);
--- a/examples/talk-llama/talk-llama.cpp
+++ b/examples/talk-llama/talk-llama.cpp
@ -1,8 +1,8 @@
 // Talk with AI
 //

-#include "common-sdl.h"
 #include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 #include "llama.h"

@ -14,37 +14,15 @@
 #include <thread>
 #include <vector>
 #include <regex>
-#include <sstream>

 std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::string & text, bool add_bos) {
-    auto * model = llama_get_model(ctx);
+    // initialize to prompt numer of chars, since n_tokens <= n_prompt_chars
+    std::vector<llama_token> res(text.size() + (int)add_bos);
+    int n = llama_tokenize(ctx, text.c_str(), res.data(), res.size(), add_bos);
+    assert(n >= 0);
+    res.resize(n);

-    // upper limit for the number of tokens
-    int n_tokens = text.length() + add_bos;
-    std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, false);
-        GGML_ASSERT(check == -n_tokens);
-    } else {
-        result.resize(n_tokens);
-    }
-    return result;
-}
-
-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());
-    if (n_tokens < 0) {
-        result.resize(-n_tokens);
-        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);
-    }
-
-    return std::string(result.data(), result.size());
+    return res;
 }

 // command-line parameters
@ -54,28 +32,22 @@ struct whisper_params {
    int32_t capture_id = -1;
    int32_t max_tokens = 32;
    int32_t audio_ctx  = 0;
-    int32_t n_gpu_layers = 999;

-    float vad_thold  = 0.6f;
-    float freq_thold = 100.0f;
+    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 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;

    std::string person      = "Georgi";
-    std::string bot_name    = "LLaMA";
-    std::string wake_cmd    = "";
-    std::string heard_ok    = "";
    std::string language    = "en";
    std::string model_wsp   = "models/ggml-base.en.bin";
    std::string model_llama = "models/ggml-llama-7B.bin";
    std::string speak       = "./examples/talk-llama/speak";
-    std::string speak_file  = "./examples/talk-llama/to_speak.txt";
    std::string prompt      = "";
    std::string fname_out;
    std::string path_session = "";       // path to file for saving/loading model eval state
@ -91,31 +63,25 @@ 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 == "-vms" || arg == "--voice-ms")       { params.voice_ms       = std::stoi(argv[++i]); }
-        else if (arg == "-c"   || arg == "--capture")        { params.capture_id     = std::stoi(argv[++i]); }
-        else if (arg == "-mt"  || arg == "--max-tokens")     { params.max_tokens     = std::stoi(argv[++i]); }
-        else if (arg == "-ac"  || arg == "--audio-ctx")      { params.audio_ctx      = std::stoi(argv[++i]); }
-        else if (arg == "-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 == "-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]; }
-        else if (arg == "-w"   || arg == "--wake-command")   { params.wake_cmd       = argv[++i]; }
-        else if (arg == "-ho"  || arg == "--heard-ok")       { params.heard_ok       = argv[++i]; }
-        else if (arg == "-l"   || arg == "--language")       { params.language       = argv[++i]; }
-        else if (arg == "-mw"  || arg == "--model-whisper")  { params.model_wsp      = argv[++i]; }
-        else if (arg == "-ml"  || arg == "--model-llama")    { params.model_llama    = argv[++i]; }
-        else if (arg == "-s"   || arg == "--speak")          { params.speak          = argv[++i]; }
-        else if (arg == "-sf"  || arg == "--speak-file")     { params.speak_file     = argv[++i]; }
-        else if (arg == "--prompt-file")                     {
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-vms" || arg == "--voice-ms")      { params.voice_ms      = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-su"  || arg == "--speed-up")      { params.speed_up      = true; }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "--verbose-prompt")                 { params.verbose_prompt = true; }
+        else if (arg == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
+        else if (arg == "--session")                        { params.path_session  = argv[++i];}
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
+        else if (arg == "-ml"  || arg == "--model-llama")   { params.model_llama   = argv[++i]; }
+        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
+        else if (arg == "--prompt-file")                    {
            std::ifstream file(argv[++i]);
            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
            if (params.prompt.back() == '\n') {
@ -123,7 +89,6 @@ 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);
@ -139,33 +104,27 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    fprintf(stderr, "usage: %s [options]\n", argv[0]);
    fprintf(stderr, "\n");
    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h,       --help           [default] show this help message and exit\n");
-    fprintf(stderr, "  -t N,     --threads N      [%-7d] number of threads to use during computation\n", params.n_threads);
-    fprintf(stderr, "  -vms N,   --voice-ms N     [%-7d] voice duration in milliseconds\n",              params.voice_ms);
-    fprintf(stderr, "  -c ID,    --capture ID     [%-7d] capture device ID\n",                           params.capture_id);
-    fprintf(stderr, "  -mt N,    --max-tokens N   [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
-    fprintf(stderr, "  -ac N,    --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
-    fprintf(stderr, "  -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, "  -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());
-    fprintf(stderr, "  -ho TEXT, --heard-ok TEXT  [%-7s] said by TTS before generating reply\n",         params.heard_ok.c_str());
-    fprintf(stderr, "  -l LANG,  --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
-    fprintf(stderr, "  -mw FILE, --model-whisper  [%-7s] whisper model file\n",                          params.model_wsp.c_str());
-    fprintf(stderr, "  -ml FILE, --model-llama    [%-7s] llama model file\n",                            params.model_llama.c_str());
-    fprintf(stderr, "  -s FILE,  --speak TEXT     [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  -sf FILE, --speak-file     [%-7s] file to pass to TTS\n",                         params.speak_file.c_str());
-    fprintf(stderr, "  --prompt-file FNAME        [%-7s] file with custom prompt to start dialog\n",     "");
-    fprintf(stderr, "  --session FNAME                   file to cache model state in (may be large!) (default: none)\n");
-    fprintf(stderr, "  -f FNAME, --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -h,       --help          [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N     [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -vms N,   --voice-ms N    [%-7d] voice duration in milliseconds\n",              params.voice_ms);
+    fprintf(stderr, "  -c ID,    --capture ID    [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,    --max-tokens N  [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,    --audio-ctx N   [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,   --vad-thold N   [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,   --freq-thold N  [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -su,      --speed-up      [%-7s] speed up audio by x2 (reduced accuracy)\n",     params.speed_up ? "true" : "false");
+    fprintf(stderr, "  -tr,      --translate     [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,      --print-special [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,      --print-energy  [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
+    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
+    fprintf(stderr, "  -ml FILE, --model-llama   [%-7s] llama model file\n",                            params.model_llama.c_str());
+    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
+    fprintf(stderr, "  --prompt-file FNAME       [%-7s] file with custom prompt to start dialog\n",     "");
+    fprintf(stderr, "  --session FNAME       file to cache model state in (may be large!) (default: none)\n");
+    fprintf(stderr, "  --verbose-prompt          [%-7s] print prompt at start\n",                       params.verbose_prompt ? "true" : "false");
+    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
    fprintf(stderr, "\n");
 }

@ -237,18 +196,6 @@ std::string transcribe(
    return result;
 }

-std::vector<std::string> get_words(const std::string &txt) {
-    std::vector<std::string> words;
-
-    std::istringstream iss(txt);
-    std::string word;
-    while (iss >> word) {
-        words.push_back(word);
-    }
-
-    return words;
-}
-
 const std::string k_prompt_whisper = R"(A conversation with a person called {1}.)";

 const std::string k_prompt_llama = R"(Text transcript of a never ending dialog, where {0} interacts with an AI assistant named {1}.
@ -276,7 +223,7 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+    if (whisper_lang_id(params.language.c_str()) == -1) {
        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
        whisper_print_usage(argc, argv, params);
        exit(0);
@ -284,32 +231,20 @@ int main(int argc, char ** argv) {

    // whisper init

-    struct whisper_context_params cparams = whisper_context_default_params();
-    cparams.use_gpu = params.use_gpu;
-
-    struct whisper_context * ctx_wsp = whisper_init_from_file_with_params(params.model_wsp.c_str(), cparams);
+    struct whisper_context * ctx_wsp = whisper_init_from_file(params.model_wsp.c_str());

    // llama init

-    llama_backend_init();
+    llama_init_backend();

-    auto lmparams = llama_model_default_params();
-    if (!params.use_gpu) {
-        lmparams.n_gpu_layers = 0;
-    } else {
-        lmparams.n_gpu_layers = params.n_gpu_layers;
-    }
-
-    struct llama_model * model_llama = llama_load_model_from_file(params.model_llama.c_str(), lmparams);
-
-    llama_context_params lcparams = llama_context_default_params();
+    auto lparams = llama_context_default_params();

    // tune these to your liking
-    lcparams.n_ctx      = 2048;
-    lcparams.seed       = 1;
-    lcparams.n_threads  = params.n_threads;
+    lparams.n_ctx      = 2048;
+    lparams.seed       = 1;
+    lparams.f16_kv     = true;

-    struct llama_context * ctx_llama = llama_new_context_with_model(model_llama, lcparams);
+    struct llama_context * ctx_llama = llama_init_from_file(params.model_llama.c_str(), lparams);

    // print some info about the processing
    {
@ -332,6 +267,7 @@ int main(int argc, char ** argv) {
        fprintf(stderr, "\n");
    }

+
    // init audio

    audio_async audio(30*1000);
@ -342,17 +278,20 @@ int main(int argc, char ** argv) {

    audio.resume();

+    int n_iter = 0;
+
    bool is_running  = true;
    bool force_speak = false;

    float prob0 = 0.0f;

    const std::string chat_symb = ":";
+    const std::string bot_name  = "LLaMA";

    std::vector<float> pcmf32_cur;
    std::vector<float> pcmf32_prompt;

-    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", params.bot_name);
+    const std::string prompt_whisper = ::replace(k_prompt_whisper, "{1}", bot_name);

    // construct the initial prompt for LLaMA inference
    std::string prompt_llama = params.prompt.empty() ? k_prompt_llama : params.prompt;
@ -361,7 +300,7 @@ int main(int argc, char ** argv) {
    prompt_llama.insert(0, 1, ' ');

    prompt_llama = ::replace(prompt_llama, "{0}", params.person);
-    prompt_llama = ::replace(prompt_llama, "{1}", params.bot_name);
+    prompt_llama = ::replace(prompt_llama, "{1}", bot_name);

    {
        // get time string
@ -391,8 +330,6 @@ int main(int argc, char ** argv) {

    prompt_llama = ::replace(prompt_llama, "{4}", chat_symb);

-    llama_batch batch = llama_batch_init(llama_n_ctx(ctx_llama), 0, 1);
-
    // init session
    std::string path_session = params.path_session;
    std::vector<llama_token> session_tokens;
@ -406,7 +343,7 @@ int main(int argc, char ** argv) {
        if (fp != NULL) {
            std::fclose(fp);

-            session_tokens.resize(llama_n_ctx(ctx_llama));
+            session_tokens.resize(lparams.n_ctx);
            size_t n_token_count_out = 0;
            if (!llama_load_session_file(ctx_llama, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
@ -428,21 +365,8 @@ int main(int argc, char ** argv) {
    printf("\n");
    printf("%s : initializing - please wait ...\n", __func__);

-    // prepare batch
-    {
-        batch.n_tokens = embd_inp.size();
-
-        for (int i = 0; i < batch.n_tokens; i++) {
-            batch.token[i]     = embd_inp[i];
-            batch.pos[i]       = i;
-            batch.n_seq_id[i]  = 1;
-            batch.seq_id[i][0] = 0;
-            batch.logits[i]    = i == batch.n_tokens - 1;
-        }
-    }
-
-    if (llama_decode(ctx_llama, batch)) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    if (llama_eval(ctx_llama, embd_inp.data(), embd_inp.size(), 0, params.n_threads)) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
        return 1;
    }

@ -478,16 +402,6 @@ int main(int argc, char ** argv) {
    bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < (embd_inp.size() * 3 / 4);

    printf("%s : done! start speaking in the microphone\n", __func__);
-
-    // show wake command if enabled
-    const std::string wake_cmd = params.wake_cmd;
-    const int wake_cmd_length = get_words(wake_cmd).size();
-    const bool use_wake_cmd = wake_cmd_length > 0;
-
-    if (use_wake_cmd) {
-        printf("%s : the wake-up command is: '%s%s%s'\n", __func__, "\033[1m", wake_cmd.c_str(), "\033[0m");
-    }
-
    printf("\n");
    printf("%s%s", params.person.c_str(), chat_symb.c_str());
    fflush(stdout);
@ -533,38 +447,10 @@ int main(int argc, char ** argv) {

                audio.get(params.voice_ms, pcmf32_cur);

-                std::string all_heard;
-
-                if (!force_speak) {
-                    all_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
-                }
-
-                const auto words = get_words(all_heard);
-
-                std::string wake_cmd_heard;
                std::string text_heard;

-                for (int i = 0; i < (int) words.size(); ++i) {
-                    if (i < wake_cmd_length) {
-                        wake_cmd_heard += words[i] + " ";
-                    } else {
-                        text_heard += words[i] + " ";
-                    }
-                }
-
-                // check if audio starts with the wake-up command if enabled
-                if (use_wake_cmd) {
-                    const float sim = similarity(wake_cmd_heard, wake_cmd);
-
-                    if ((sim < 0.7f) || (text_heard.empty())) {
-                        audio.clear();
-                        continue;
-                    }
-                }
-
-                // optionally give audio feedback that the current text is being processed
-                if (!params.heard_ok.empty()) {
-                    speak_with_file(params.speak, params.heard_ok, params.speak_file, voice_id);
+                if (!force_speak) {
+                    text_heard = ::trim(::transcribe(ctx_wsp, params, pcmf32_cur, prompt_whisper, prob0, t_ms));
                }

                // remove text between brackets using regex
@ -601,7 +487,7 @@ int main(int argc, char ** argv) {
                force_speak = false;

                text_heard.insert(0, 1, ' ');
-                text_heard += "\n" + params.bot_name + chat_symb;
+                text_heard += "\n" + bot_name + chat_symb;
                fprintf(stdout, "%s%s%s", "\033[1m", text_heard.c_str(), "\033[0m");
                fflush(stdout);

@ -628,7 +514,7 @@ int main(int argc, char ** argv) {
                            //printf("\n---\n");
                            //printf("resetting: '");
                            //for (int i = 0; i < (int) embd.size(); i++) {
-                            //    printf("%s", llama_token_to_piece(ctx_llama, embd[i]));
+                            //    printf("%s", llama_token_to_str(ctx_llama, embd[i]));
                            //}
                            //printf("'\n");
                            //printf("\n---\n");
@ -662,21 +548,8 @@ int main(int argc, char ** argv) {
                            n_session_consumed = session_tokens.size();
                        }

-                        // prepare batch
-                        {
-                            batch.n_tokens = embd.size();
-
-                            for (int i = 0; i < batch.n_tokens; i++) {
-                                batch.token[i]     = embd[i];
-                                batch.pos[i]       = n_past + i;
-                                batch.n_seq_id[i]  = 1;
-                                batch.seq_id[i][0] = 0;
-                                batch.logits[i]    = i == batch.n_tokens - 1;
-                            }
-                        }
-
-                        if (llama_decode(ctx_llama, batch)) {
-                            fprintf(stderr, "%s : failed to decode\n", __func__);
+                        if (llama_eval(ctx_llama, embd.data(), embd.size(), n_past, params.n_threads)) {
+                            fprintf(stderr, "%s : failed to eval\n", __func__);
                            return 1;
                        }
                    }
@ -707,9 +580,9 @@ int main(int argc, char ** argv) {

                        {
                            auto logits = llama_get_logits(ctx_llama);
-                            auto n_vocab = llama_n_vocab(model_llama);
+                            auto n_vocab = llama_n_vocab(ctx_llama);

-                            logits[llama_token_eos(model_llama)] = 0;
+                            logits[llama_token_eos()] = 0;

                            std::vector<llama_token_data> candidates;
                            candidates.reserve(n_vocab);
@ -720,13 +593,13 @@ int main(int argc, char ** argv) {
                            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };

                            // apply repeat penalty
-                            const float nl_logit = logits[llama_token_nl(model_llama)];
+                            const float nl_logit = logits[llama_token_nl()];

-                            llama_sample_repetition_penalties(ctx_llama, &candidates_p,
+                            llama_sample_repetition_penalty(ctx_llama, &candidates_p,
                                    embd_inp.data() + std::max(0, n_past - repeat_last_n),
-                                    repeat_last_n, repeat_penalty, 0.0, 0.0f);
+                                    repeat_last_n, repeat_penalty);

-                            logits[llama_token_nl(model_llama)] = nl_logit;
+                            logits[llama_token_nl()] = nl_logit;

                            if (temp <= 0) {
                                // Greedy sampling
@ -735,28 +608,27 @@ int main(int argc, char ** argv) {
                                // Temperature sampling
                                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
                                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                                llama_sample_temp (ctx_llama, &candidates_p, temp);
+                                llama_sample_temperature(ctx_llama, &candidates_p, temp);
                                id = llama_sample_token(ctx_llama, &candidates_p);
                            }
                        }

-                        if (id != llama_token_eos(model_llama)) {
+                        if (id != llama_token_eos()) {
                            // add it to the context
                            embd.push_back(id);

-                            text_to_speak += llama_token_to_piece(ctx_llama, id);
+                            text_to_speak += llama_token_to_str(ctx_llama, id);

-                            printf("%s", llama_token_to_piece(ctx_llama, id).c_str());
-                            fflush(stdout);
+                            printf("%s", llama_token_to_str(ctx_llama, id));
                        }
                    }

                    {
                        std::string last_output;
                        for (int i = embd_inp.size() - 16; i < (int) embd_inp.size(); i++) {
-                            last_output += llama_token_to_piece(ctx_llama, embd_inp[i]);
+                            last_output += llama_token_to_str(ctx_llama, embd_inp[i]);
                        }
-                        last_output += llama_token_to_piece(ctx_llama, embd[0]);
+                        last_output += llama_token_to_str(ctx_llama, embd[0]);

                        for (std::string & antiprompt : antiprompts) {
                            if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
@ -776,9 +648,12 @@ int main(int argc, char ** argv) {
                    }
                }

-                speak_with_file(params.speak, text_to_speak, params.speak_file, voice_id);
+                text_to_speak = ::replace(text_to_speak, "\"", "");
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());

                audio.clear();
+
+                ++n_iter;
            }
        }
    }
--- a/examples/talk-llama/unicode.cpp
+++ b/examples/talk-llama/unicode.cpp
--- a/examples/talk-llama/unicode.h
+++ b/examples/talk-llama/unicode.h
@ -1,26 +0,0 @@
-#pragma once
-
-#include <cstdint>
-#include <string>
-#include <vector>
-
-#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);
-
-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);
-
--- a/examples/talk.wasm/emscripten.cpp
+++ b/examples/talk.wasm/emscripten.cpp
@ -29,6 +29,18 @@ std::string g_status_forced = "";

 std::vector<float> g_pcmf32;

+std::string to_timestamp(int64_t t) {
+    int64_t sec = t/100;
+    int64_t msec = t - sec*100;
+    int64_t min = sec/60;
+    sec = sec - min*60;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d.%03d", (int) min, (int) sec, (int) msec);
+
+    return std::string(buf);
+}
+
 void talk_set_status(const std::string & status) {
    std::lock_guard<std::mutex> lock(g_mutex);
    g_status = status;
@ -259,7 +271,7 @@ EMSCRIPTEN_BINDINGS(talk) {
    emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
        for (size_t i = 0; i < g_contexts.size(); ++i) {
            if (g_contexts[i] == nullptr) {
-                g_contexts[i] = whisper_init_from_file_with_params(path_model.c_str(), whisper_context_default_params());
+                g_contexts[i] = whisper_init_from_file(path_model.c_str());
                if (g_contexts[i] != nullptr) {
                    g_running = true;
                    if (g_worker.joinable()) {
--- a/examples/talk.wasm/gpt-2.cpp
+++ b/examples/talk.wasm/gpt-2.cpp
@ -155,33 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
        const int n_ctx   = hparams.n_ctx;
        const int n_vocab = hparams.n_vocab;

-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b

-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
-        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head

-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b

-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b

-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v

        ctx_size += (6 + 12*n_layer)*256; // object overhead

@ -524,7 +524,8 @@ bool gpt2_eval(
            struct ggml_tensor * KQ_scaled =
                ggml_scale(ctx0,
                        KQ,
-                        1.0f/sqrt(float(n_embd)/n_head));
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );

            // KQ_masked = mask_past(KQ_scaled)
            // [n_past + N, N, 12]
--- a/examples/talk/.gitignore
+++ b/examples/talk/.gitignore
@ -1,2 +1 @@
 audio.mp3
-to_speak.txt
--- a/examples/talk/README.md
+++ b/examples/talk/README.md
@ -31,7 +31,7 @@ To run this, you will need a ggml GPT-2 model: [instructions](https://github.com
 Alternatively, you can simply download the smallest ggml GPT-2 117M model (240 MB) like this:

 ```
-wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
+wget --quiet --show-progress -O models/ggml-gpt-2-117M.bin https://huggingface.co/ggerganov/ggml/raw/main/ggml-model-gpt-2-117M.bin
 ```

 ## TTS
--- a/examples/talk/eleven-labs.py
+++ b/examples/talk/eleven-labs.py
@ -1,80 +1,20 @@
 import sys
-import argparse
-import textwrap
+import importlib.util

-parser = argparse.ArgumentParser(add_help=False,
-    formatter_class=argparse.RawTextHelpFormatter)
-parser.add_argument("-q", "--quick", action="store_true",
-    help="skip checking the required library")
-
-modes = parser.add_argument_group("action")
-modes.add_argument("inputfile", metavar="TEXTFILE",
-    nargs='?', type=argparse.FileType(), default=sys.stdin,
-    help="read the text file (default: stdin)")
-modes.add_argument("-l", "--list", action="store_true",
-    help="show the list of voices and exit")
-modes.add_argument("-h", "--help", action="help",
-    help="show this help and exit")
-
-selopts = parser.add_argument_group("voice selection")
-selmodes = selopts.add_mutually_exclusive_group()
-selmodes.add_argument("-n", "--name",
-    default="Arnold",
-    help="get a voice object by name (default: Arnold)")
-selmodes.add_argument("-v", "--voice", type=int, metavar="NUMBER",
-    help="get a voice object by number (see --list)")
-selopts.add_argument("-f", "--filter", action="append", metavar="KEY=VAL",
-    default=["use case=narration"],
-    help=textwrap.dedent('''\
-        filter voices by labels (default: "use case=narration")
-        this option can be used multiple times
-        filtering will be disabled if the first -f has no "=" (e.g. -f "any")
-        '''))
-
-outmodes = parser.add_argument_group("output")
-outgroup = outmodes.add_mutually_exclusive_group()
-outgroup.add_argument("-s", "--save", metavar="FILE",
-    default="audio.mp3",
-    help="save the TTS to a file (default: audio.mp3)")
-outgroup.add_argument("-p", "--play", action="store_true",
-    help="play the TTS with ffplay")
-
-args = parser.parse_args()
-
-if not args.quick:
-    import importlib.util
-    if importlib.util.find_spec("elevenlabs") is None:
-        print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
-        sys.exit()
-
-from elevenlabs import voices, generate, play, save
-
-if args.filter and "=" in args.filter[0]:
-    voicelist = voices()
-    for f in args.filter:
-        label, value = f.split("=")
-        voicelist = filter(lambda x: x.labels.get(label) == value, voicelist)
-    voicelist = list(voicelist)
-else:
-    voicelist = list(voices())
-
-if args.list:
-    for i, v in enumerate(voicelist):
-        print(str(i) + ": " + v.name + " " + str(v.labels))
+if importlib.util.find_spec("elevenlabs") is None:
+    print("elevenlabs library is not installed, you can install it to your enviroment using 'pip install elevenlabs'")
    sys.exit()

-if args.voice:
-    voice = voicelist[args.voice % len(voicelist)]
-else:
-    voice = args.name
-    # if -n should consult -f, use the following
-    #voice = next(x for x in voicelist if x.name == args.name)
+from elevenlabs import generate, play, save

+# Get a Voice object, by name or UUID
+voice = "Arnold" #Possible Voices: Adam Antoni Arnold Bella Domi Elli Josh
+
+# Generate the TTS
 audio = generate(
-    text=str(args.inputfile.read()),
-    voice=voice
+  text=str(sys.argv[2:]),
+  voice=voice
 )
-if args.play:
-    play(audio)
-else:
-    save(audio, args.save) 
+
+# Save the TTS to a file
+save(audio, "audio.mp3") 
--- a/examples/talk/gpt-2.cpp
+++ b/examples/talk/gpt-2.cpp
@ -121,13 +121,13 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
            return false;
        }

-        char word[129];
-
+        std::string word;
        for (int i = 0; i < n_vocab; i++) {
            uint32_t len;
            fin.read((char *) &len, sizeof(len));
-            word[len] = '\0';
-            fin.read((char *) word, len);
+
+            word.resize(len);
+            fin.read((char *) word.data(), len);

            vocab.token_to_id[word] = i;
            vocab.id_to_token[i] = word;
@ -155,33 +155,33 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
        const int n_ctx   = hparams.n_ctx;
        const int n_vocab = hparams.n_vocab;

-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_g
-        ctx_size += ggml_row_size(GGML_TYPE_F32, n_embd); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b

-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // wte
-        ctx_size +=   n_ctx*ggml_row_size(GGML_TYPE_F32, n_embd); // wpe
-        ctx_size += n_vocab*ggml_row_size(wtype, n_embd);         // lm_head
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head

-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b

-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_g
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd)); // ln_2_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         3*n_embd*n_embd)); // c_attn_attn_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 3*n_embd));        // c_attn_attn_b
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         n_embd*n_embd)); // c_attn_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, n_embd));        // c_attn_proj_b
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_fc_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32, 4*n_embd));        // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b

-        ctx_size += n_layer*(ggml_row_size(wtype,         4*n_embd*n_embd)); // c_mlp_proj_w
-        ctx_size += n_layer*(ggml_row_size(GGML_TYPE_F32,   n_embd));        // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b

-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_k
-        ctx_size += n_ctx*n_layer*ggml_row_size(GGML_TYPE_F32, n_embd); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v

        ctx_size += (6 + 12*n_layer)*256; // object overhead

@ -191,9 +191,9 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
    // create the ggml context
    {
        struct ggml_init_params params = {
-            /*.mem_size   =*/ ctx_size,
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ false,
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+            .no_alloc   = false,
        };

        model.ctx = ggml_init(params);
@ -525,7 +525,8 @@ bool gpt2_eval(
            struct ggml_tensor * KQ_scaled =
                ggml_scale(ctx0,
                        KQ,
-                        1.0f/sqrt(float(n_embd)/n_head));
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );

            // KQ_masked = mask_past(KQ_scaled)
            // [n_past + N, N, 12]
--- a/examples/talk/speak
+++ b/examples/talk/speak
@ -1,40 +1,24 @@
 #!/bin/bash

 # Usage:
-#  speak <voice_id> <textfile>
+#  speak.sh <voice_id> <text-to-speak>

-function installed() { command -v $1 >/dev/null 2>&1; }
+# espeak
+# Mac OS: brew install espeak
+# Linux: apt-get install espeak
+#
+#espeak -v en-us+m$1 -s 175 -p 50 -a 200 -g 5 -k 5 "$2"

-if installed espeak; then
-  espeak -v en-us+m$1 -s 225 -p 50 -a 200 -g 5 -k 5 -f $2
-
-elif installed piper && installed aplay; then
-  cat $2 | piper --model ~/en_US-lessac-medium.onnx --output-raw | aplay -q -r 22050 -f S16_LE -t raw -
-
-# for Mac
-elif installed say; then
-  say -f $2
+# Mac OS "say" command
+say "$2"

 # Eleven Labs
-elif installed python3 && \
-  python3 -c 'import importlib.util; exit(not importlib.util.find_spec("elevenlabs"))' && \
-  installed ffplay; then
-    # It's possible to use the API for free with limited number of characters.
-    # To increase this limit register to https://beta.elevenlabs.io to get an api key
-    # and paste it after 'ELEVEN_API_KEY='
-    # Keep the line commented to use the free version without api key
-    #export ELEVEN_API_KEY=your_api_key
-    wd=$(dirname $0)
-    script=$wd/eleven-labs.py
-    python3 $script -q -p -v $1 $2 >/dev/null 2>&1
-
-    # Uncomment to keep the audio file
-    #python3 $script -q -s ./audio.mp3 -v $1 $2 >/dev/null 2>&1
-    #ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3 >/dev/null 2>&1
-
-else
-  echo 'Install espeak ("brew install espeak" or "apt-get install espeak"),'
-  echo 'piper ("pip install piper-tts" or https://github.com/rhasspy/piper) with aplay,'
-  echo 'or elevenlabs ("pip install elevenlabs") with ffplay.'
-  echo '(export ELEVEN_API_KEY if you have an api key from https://beta.elevenlabs.io)'
-fi
+# To use it, install the elevenlabs module from pip (pip install elevenlabs)
+# It's possible to use the API for free with limited number of characters. To increase this limit register to https://beta.elevenlabs.io to get an api key and paste it after 'ELEVEN_API_KEY='
+#Keep the line commented to use the free version without api key
+#
+#export ELEVEN_API_KEY=your_api_key
+#wd=$(dirname $0)
+#script=$wd/eleven-labs.py
+#python3 $script $1 "$2"
+#ffplay -autoexit -nodisp -loglevel quiet -hide_banner -i ./audio.mp3
--- a/examples/talk/speak.ps1
+++ b/examples/talk/speak.ps1
@ -1,14 +1,12 @@
 # Set-ExecutionPolicy -ExecutionPolicy Bypass -Scope CurrentUser
 param(
-  [Parameter(Mandatory=$true)][int]$voicenum,
-  [Parameter(Mandatory=$true)][string]$textfile
+  # voice options are David or Zira
+  [Parameter(Mandatory=$true)][string]$voice,
+  [Parameter(Mandatory=$true)][string]$text
 )

 Add-Type -AssemblyName System.Speech;
 $speak = New-Object System.Speech.Synthesis.SpeechSynthesizer;
-$voiceoptions = $speak.GetInstalledVoices("en-US");
-$voice = $voiceoptions[$voicenum % $voiceoptions.count];
-$speak.SelectVoice($voice.VoiceInfo.Name);
+$speak.SelectVoice("Microsoft $voice Desktop");
 $speak.Rate="0";
-$text = Get-Content -Path $textfile;
 $speak.Speak($text);
--- a/examples/talk/talk.cpp
+++ b/examples/talk/talk.cpp
@ -1,8 +1,8 @@
 // Talk with AI
 //

-#include "common-sdl.h"
 #include "common.h"
+#include "common-sdl.h"
 #include "whisper.h"
 #include "gpt-2.h"

@ -31,14 +31,12 @@ struct whisper_params {
    bool print_special = false;
    bool print_energy  = false;
    bool no_timestamps = true;
-    bool use_gpu       = true;

    std::string person    = "Santa";
    std::string language  = "en";
    std::string model_wsp = "models/ggml-base.en.bin";
    std::string model_gpt = "models/ggml-gpt-2-117M.bin";
    std::string speak     = "./examples/talk/speak";
-    std::string speak_file= "./examples/talk/to_speak.txt";
    std::string fname_out;
 };

@ -63,13 +61,11 @@ bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
        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 == "-p"   || arg == "--person")        { params.person        = argv[++i]; }
        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
        else if (arg == "-mw"  || arg == "--model-whisper") { params.model_wsp     = argv[++i]; }
        else if (arg == "-mg"  || arg == "--model-gpt")     { params.model_gpt     = argv[++i]; }
        else if (arg == "-s"   || arg == "--speak")         { params.speak         = argv[++i]; }
-        else if (arg == "-sf"  || arg == "--speak_file")    { params.speak_file    = argv[++i]; }
        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
        else {
            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
@ -98,13 +94,11 @@ void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & para
    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, "  -p NAME,  --person NAME   [%-7s] person name (for prompt selection)\n",          params.person.c_str());
    fprintf(stderr, "  -l LANG,  --language LANG [%-7s] spoken language\n",                             params.language.c_str());
    fprintf(stderr, "  -mw FILE, --model-whisper [%-7s] whisper model file\n",                          params.model_wsp.c_str());
    fprintf(stderr, "  -mg FILE, --model-gpt     [%-7s] gpt model file\n",                              params.model_gpt.c_str());
    fprintf(stderr, "  -s FILE,  --speak TEXT    [%-7s] command for TTS\n",                             params.speak.c_str());
-    fprintf(stderr, "  -sf FILE, --speak_file    [%-7s] file to pass to TTS\n",                         params.speak_file.c_str());
    fprintf(stderr, "  -f FNAME, --file FNAME    [%-7s] text output file name\n",                       params.fname_out.c_str());
    fprintf(stderr, "\n");
 }
@ -187,10 +181,8 @@ int main(int argc, char ** argv) {
    }

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

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

    // gpt init

@ -319,7 +311,7 @@ int main(int argc, char ** argv) {
                    std::string prompt = ::replace(::replace(k_prompt, "{0}", params.person), "{1}", prompt_base);

                    text_to_speak = gpt2_gen_text(ctx_gpt, prompt.c_str(), params.max_tokens);
-                    //text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
+                    text_to_speak = std::regex_replace(text_to_speak, std::regex("[^a-zA-Z0-9\\.,\\?!\\s\\:\\'\\-]"), "");
                    text_to_speak = text_to_speak.substr(0, text_to_speak.find_first_of('\n'));

                    // remove first 2 lines of base prompt
@ -357,7 +349,7 @@ int main(int argc, char ** argv) {
                gpt2_set_prompt(ctx_gpt, prompt_base.c_str());

                text_to_speak = ::replace(text_to_speak, params.person + ": ", "");
-                speak_with_file(params.speak, text_to_speak, params.speak_file, voice_id);
+                system((params.speak + " " + std::to_string(voice_id) + " \"" + text_to_speak + "\"").c_str());

                audio.clear();

--- a/examples/twitch.sh
+++ b/examples/twitch.sh
@ -21,7 +21,7 @@ help()
    echo "Usage: ./twitch.sh -s [step] -m [model] -t [threads] [url]"
    echo "options:"
    echo "-s       Step in seconds (default is $step)."
-    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large-v2' 'large-v3' (default is '$model')."
+    echo "-m       Choose model, options are: 'tiny.en' 'tiny' 'base.en' 'base' 'small.en' 'small' 'medium.en' 'medium' 'large-v1' 'large' (default is '$model')."
    echo "-t       Number of threads to use."
    echo "-h       Print this help page."
    echo
--- a/examples/wchess/CMakeLists.txt
+++ b/examples/wchess/CMakeLists.txt
@ -1,9 +0,0 @@
-set(CMAKE_CXX_STANDARD 11)
-
-add_subdirectory(libwchess)
-
-if (EMSCRIPTEN)
-    add_subdirectory(wchess.wasm)
-else()
-    add_subdirectory(wchess.cmd)
-endif()
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Georgi Gerganov	3c50be2217	whisper : remove comment	2023-09-10 13:27:06 +03:00
Georgi Gerganov	37de5dcf2b	command : enable beam-search, add "no_timestamps", add "context", add p	2023-09-10 12:22:57 +03:00
Georgi Gerganov	7a2abb311d	grammars : add assistant + update comments	2023-09-09 20:24:58 +03:00
Georgi Gerganov	54d168db67	command : grammar-related improvements - option to read grammar from file - add sample grammars for colors and chess moves - fine-tune the performance further	2023-09-09 20:05:57 +03:00
Georgi Gerganov	b8f34d1ed7	whisper : fine-tuning grammar functionality	2023-09-06 17:05:05 +03:00
Georgi Gerganov	97ebb48b99	command : fix exception when recognizing the command	2023-09-06 15:05:17 +03:00
Georgi Gerganov	b0306cd5cf	build : fix after master merge	2023-09-06 14:28:25 +03:00
Georgi Gerganov	afc84b35b0	Merge branch 'master' into HEAD	2023-09-06 13:33:34 +03:00
Evan Jones	31476ccc0e	whisper : add grammar-based sampling	2023-08-30 20:00:31 -04:00