WIP speculative

Signed-off-by: Gianluca Boiano <morf3089@gmail.com>

.bruno/LocalAI Test Requests/model gallery/model delete.bru

@@ -0,0 +1,11 @@
+meta {
+  name: model delete
+  type: http
+  seq: 7
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/galleries
+  body: none
+  auth: none
+}

.bruno/LocalAI Test Requests/transcription/transcribe.bru

@@ -0,0 +1,16 @@
+meta {
+  name: transcribe
+  type: http
+  seq: 1
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/v1/audio/transcriptions
+  body: multipartForm
+  auth: none
+}
+
+body:multipart-form {
+  file: @file(transcription/gb1.ogg)
+  model: whisper-1
+}

.bruno/LocalAI Test Requests/tts/musicgen.bru

@@ -16,7 +16,7 @@ headers {
 
 body:json {
   {
-      "backend": "transformers-musicgen",
+      "backend": "transformers",
       "model": "facebook/musicgen-small",
       "input": "80s Synths playing Jazz"
   }

.devcontainer/docker-compose-devcontainer.yml

@@ -7,7 +7,7 @@ services:
       args:
       - FFMPEG=true
       - IMAGE_TYPE=extras
-      - GO_TAGS=stablediffusion p2p tts
+      - GO_TAGS=p2p tts
     env_file:
       - ../.env
     ports:

.env

@@ -38,12 +38,12 @@
 ## Uncomment and set to true to enable rebuilding from source
 # REBUILD=true
 
-## Enable go tags, available: stablediffusion, tts
-## stablediffusion: image generation with stablediffusion
+## Enable go tags, available: p2p, tts
+## p2p: enable distributed inferencing
 ## tts: enables text-to-speech with go-piper 
 ## (requires REBUILD=true)
 #
-# GO_TAGS=stablediffusion
+# GO_TAGS=p2p
 
 ## Path where to store generated images
 # LOCALAI_IMAGE_PATH=/tmp/generated/images
@@ -82,6 +82,15 @@
 # Enable to allow p2p mode
 # LOCALAI_P2P=true
 
+# Enable to use federated mode
+# LOCALAI_FEDERATED=true
+
+# Enable to start federation server
+# FEDERATED_SERVER=true
+
+# Define to use federation token
+# TOKEN=""
+
 ### Watchdog settings
 ###
 # Enables watchdog to kill backends that are inactive for too much time

.gitattributes

@@ -1 +1,2 @@
 *.sh text eol=lf
+backend/cpp/llama/*.hpp linguist-vendored

.github/dependabot.yml

@@ -81,14 +81,6 @@ updates:
     directory: "/backend/python/transformers"
     schedule:
       interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/transformers-musicgen"
-    schedule:
-      interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/vall-e-x"
-    schedule:
-      interval: "weekly"
   - package-ecosystem: "pip"
     directory: "/backend/python/vllm"
     schedule:

.github/labeler.yml

@@ -1,6 +1,15 @@
 enhancements:
  - head-branch: ['^feature', 'feature']
 
+dependencies:
+- any:
+  - changed-files:
+    - any-glob-to-any-file: 'Makefile'
+  - changed-files:
+    - any-glob-to-any-file: '*.mod'
+  - changed-files:
+    - any-glob-to-any-file: '*.sum'
+
 kind/documentation:
 - any:
   - changed-files:

.github/workflows/bump_deps.yaml

@@ -12,23 +12,14 @@ jobs:
           - repository: "ggerganov/llama.cpp"
             variable: "CPPLLAMA_VERSION"
             branch: "master"
-          - repository: "go-skynet/go-ggml-transformers.cpp"
-            variable: "GOGGMLTRANSFORMERS_VERSION"
-            branch: "master"
-          - repository: "donomii/go-rwkv.cpp"
-            variable: "RWKV_VERSION"
-            branch: "main"
           - repository: "ggerganov/whisper.cpp"
             variable: "WHISPER_CPP_VERSION"
             branch: "master"
-          - repository: "go-skynet/go-bert.cpp"
-            variable: "BERT_VERSION"
-            branch: "master"
-          - repository: "go-skynet/bloomz.cpp"
-            variable: "BLOOMZ_VERSION"
+          - repository: "PABannier/bark.cpp"
+            variable: "BARKCPP_VERSION"
             branch: "main"
-          - repository: "mudler/go-ggllm.cpp"
-            variable: "GOGGLLM_VERSION"
+          - repository: "leejet/stable-diffusion.cpp"
+            variable: "STABLEDIFFUSION_GGML_VERSION"
             branch: "master"
           - repository: "mudler/go-stable-diffusion"
             variable: "STABLEDIFFUSION_VERSION"

.github/workflows/checksum_checker.yaml

@@ -23,7 +23,7 @@ jobs:
           sudo pip install --upgrade pip
           pip install huggingface_hub
       - name: 'Setup yq'
-        uses: dcarbone/install-yq-action@v1.1.1
+        uses: dcarbone/install-yq-action@v1.3.1
         with:
           version: 'v4.44.2'
           download-compressed: true

.github/workflows/deploy-explorer.yaml

@@ -33,7 +33,7 @@ jobs:
         run: |
           CGO_ENABLED=0 make build-api
       - name: rm
-        uses: appleboy/ssh-action@v1.1.0
+        uses: appleboy/ssh-action@v1.2.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}
@@ -53,7 +53,7 @@ jobs:
             rm: true
             target: ./local-ai
       - name: restarting
-        uses: appleboy/ssh-action@v1.1.0
+        uses: appleboy/ssh-action@v1.2.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}

.github/workflows/generate_intel_image.yaml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - base-image: intel/oneapi-basekit:2024.2.0-devel-ubuntu22.04
+          - base-image: intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04
             runs-on: 'ubuntu-latest'
             platforms: 'linux/amd64'
     runs-on: ${{matrix.runs-on}}

.github/workflows/image.yml

@@ -280,6 +280,7 @@ jobs:
       makeflags: ${{ matrix.makeflags }}
       latest-image: ${{ matrix.latest-image }}
       latest-image-aio: ${{ matrix.latest-image-aio }}
+      skip-drivers: ${{ matrix.skip-drivers }}
     secrets:
       dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
       dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
@@ -301,6 +302,7 @@ jobs:
             latest-image: 'latest-cpu'
             latest-image-aio: 'latest-aio-cpu'
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
@@ -312,6 +314,7 @@ jobs:
             base-image: "ubuntu:22.04"
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
@@ -323,6 +326,7 @@ jobs:
             base-image: "ubuntu:22.04"
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
@@ -334,6 +338,7 @@ jobs:
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
@@ -344,6 +349,7 @@ jobs:
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
           - build-type: 'vulkan'
             platforms: 'linux/amd64'
@@ -354,4 +360,45 @@ jobs:
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
+  gh-runner:
+    uses: ./.github/workflows/image_build.yml
+    with:
+      tag-latest: ${{ matrix.tag-latest }}
+      tag-suffix: ${{ matrix.tag-suffix }}
+      ffmpeg: ${{ matrix.ffmpeg }}
+      image-type: ${{ matrix.image-type }}
+      build-type: ${{ matrix.build-type }}
+      cuda-major-version: ${{ matrix.cuda-major-version }}
+      cuda-minor-version: ${{ matrix.cuda-minor-version }}
+      platforms: ${{ matrix.platforms }}
+      runs-on: ${{ matrix.runs-on }}
+      aio: ${{ matrix.aio }}
+      base-image: ${{ matrix.base-image }}
+      grpc-base-image: ${{ matrix.grpc-base-image }}
+      makeflags: ${{ matrix.makeflags }}
+      latest-image: ${{ matrix.latest-image }}
+      latest-image-aio: ${{ matrix.latest-image-aio }}
+      skip-drivers: ${{ matrix.skip-drivers }}
+    secrets:
+      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+    strategy:
+      matrix:
+        include:
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/arm64'
+            tag-latest: 'false'
+            tag-suffix: '-nvidia-l4t-arm64-core'
+            latest-image: 'latest-nvidia-l4t-arm64-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+            runs-on: 'ubuntu-24.04-arm'
+            makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'true'

.github/workflows/image_build.yml

@@ -49,6 +49,10 @@ on:
         description: 'FFMPEG'
         default: ''
         type: string
+      skip-drivers:
+        description: 'Skip drivers by default'
+        default: 'false'
+        type: string
       image-type:
         description: 'Image type'
         default: ''
@@ -234,6 +238,7 @@ jobs:
             GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
             GRPC_VERSION=v1.65.0
             MAKEFLAGS=${{ inputs.makeflags }}
+            SKIP_DRIVERS=${{ inputs.skip-drivers }}
           context: .
           file: ./Dockerfile
           cache-from: type=gha
@@ -262,6 +267,7 @@ jobs:
             GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
             GRPC_VERSION=v1.65.0
             MAKEFLAGS=${{ inputs.makeflags }}
+            SKIP_DRIVERS=${{ inputs.skip-drivers }}
           context: .
           file: ./Dockerfile
           cache-from: type=gha

.github/workflows/release.yaml

@@ -237,40 +237,7 @@ jobs:
           detached: true
           connect-timeout-seconds: 180
           limit-access-to-actor: true
-  build-stablediffusion:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - uses: actions/setup-go@v5
-        with:
-          go-version: '1.21.x'
-          cache: false
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install -y --no-install-recommends libopencv-dev protobuf-compiler ccache upx-ucl
-          go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
-          go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
-      - name: Build stablediffusion
-        run: |
-          export PATH=$PATH:$GOPATH/bin
-          make backend-assets/grpc/stablediffusion
-          mkdir -p release && cp backend-assets/grpc/stablediffusion release
-        env:
-          GO_TAGS: stablediffusion
-      - uses: actions/upload-artifact@v4
-        with:
-          name: stablediffusion
-          path: release/
-      - name: Release
-        uses: softprops/action-gh-release@v2
-        if: startsWith(github.ref, 'refs/tags/')
-        with:
-          files: |
-            release/*
+
 
   build-macOS-x86_64:
     runs-on: macos-13

.github/workflows/secscan.yaml

@@ -18,7 +18,7 @@ jobs:
         if: ${{ github.actor != 'dependabot[bot]' }}
       - name: Run Gosec Security Scanner
         if: ${{ github.actor != 'dependabot[bot]' }}
-        uses: securego/gosec@v2.21.4
+        uses: securego/gosec@v2.22.0
         with:
           # we let the report trigger content trigger a failure using the GitHub Security features.
           args: '-no-fail -fmt sarif -out results.sarif ./...'

.github/workflows/test-extra.yml

@@ -35,30 +35,6 @@ jobs:
         run: |
            make --jobs=5 --output-sync=target -C backend/python/transformers
            make --jobs=5 --output-sync=target -C backend/python/transformers test
-
-  tests-sentencetransformers:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential ffmpeg
-          # Install UV
-          curl -LsSf https://astral.sh/uv/install.sh | sh
-          sudo apt-get install -y ca-certificates cmake curl patch python3-pip
-          sudo apt-get install -y libopencv-dev
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
-
-      - name: Test sentencetransformers
-        run: |
-           make --jobs=5 --output-sync=target -C backend/python/sentencetransformers
-           make --jobs=5 --output-sync=target -C backend/python/sentencetransformers test
-
-
   tests-rerankers:
     runs-on: ubuntu-latest
     steps:
@@ -102,71 +78,27 @@ jobs:
           make --jobs=5 --output-sync=target -C backend/python/diffusers
           make --jobs=5 --output-sync=target -C backend/python/diffusers test
 
-  tests-parler-tts:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential ffmpeg
-          # Install UV
-          curl -LsSf https://astral.sh/uv/install.sh | sh
-          sudo apt-get install -y ca-certificates cmake curl patch python3-pip
-          sudo apt-get install -y libopencv-dev
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
+  # tests-transformers-musicgen:
+  #   runs-on: ubuntu-latest
+  #   steps:
+  #     - name: Clone
+  #       uses: actions/checkout@v4
+  #       with:
+  #         submodules: true
+  #     - name: Dependencies
+  #       run: |
+  #         sudo apt-get update
+  #         sudo apt-get install build-essential ffmpeg
+  #         # Install UV
+  #         curl -LsSf https://astral.sh/uv/install.sh | sh
+  #         sudo apt-get install -y ca-certificates cmake curl patch python3-pip
+  #         sudo apt-get install -y libopencv-dev
+  #         pip install --user --no-cache-dir grpcio-tools==1.64.1
 
-      - name: Test parler-tts
-        run: |
-           make --jobs=5 --output-sync=target -C backend/python/parler-tts
-           make --jobs=5 --output-sync=target -C backend/python/parler-tts test
-
-  tests-openvoice:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential ffmpeg
-          # Install UV
-          curl -LsSf https://astral.sh/uv/install.sh | sh
-          sudo apt-get install -y ca-certificates cmake curl patch python3-pip
-          sudo apt-get install -y libopencv-dev
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
-
-      - name: Test openvoice
-        run: |
-           make --jobs=5 --output-sync=target -C backend/python/openvoice
-           make --jobs=5 --output-sync=target -C backend/python/openvoice test
-
-  tests-transformers-musicgen:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential ffmpeg
-          # Install UV
-          curl -LsSf https://astral.sh/uv/install.sh | sh
-          sudo apt-get install -y ca-certificates cmake curl patch python3-pip
-          sudo apt-get install -y libopencv-dev
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
-
-      - name: Test transformers-musicgen
-        run: |
-           make --jobs=5 --output-sync=target -C backend/python/transformers-musicgen
-           make --jobs=5 --output-sync=target -C backend/python/transformers-musicgen test
+  #     - name: Test transformers-musicgen
+  #       run: |
+  #          make --jobs=5 --output-sync=target -C backend/python/transformers-musicgen
+  #          make --jobs=5 --output-sync=target -C backend/python/transformers-musicgen test
 
   # tests-bark:
   #   runs-on: ubuntu-latest
@@ -253,26 +185,6 @@ jobs:
   #       run: |
   #          make --jobs=5 --output-sync=target -C backend/python/vllm
   #          make --jobs=5 --output-sync=target -C backend/python/vllm test
-  tests-vallex:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-        with:
-          submodules: true
-      - name: Dependencies
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential ffmpeg
-          # Install UV
-          curl -LsSf https://astral.sh/uv/install.sh | sh
-          sudo apt-get install -y ca-certificates cmake curl patch python3-pip
-          sudo apt-get install -y libopencv-dev
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
-      - name: Test vall-e-x
-        run: |
-           make --jobs=5 --output-sync=target -C backend/python/vall-e-x
-           make --jobs=5 --output-sync=target -C backend/python/vall-e-x test
 
   tests-coqui:
     runs-on: ubuntu-latest

.github/workflows/test.yml

@@ -100,15 +100,12 @@ jobs:
           # The python3-grpc-tools package in 22.04 is too old
           pip install --user grpcio-tools
 
-          sudo rm -rfv /usr/bin/conda || true
-          PATH=$PATH:/opt/conda/bin make -C backend/python/sentencetransformers
+          make -C backend/python/transformers
 
           # Pre-build piper before we start tests in order to have shared libraries in place
           make sources/go-piper && \
           GO_TAGS="tts" make -C sources/go-piper piper.o && \
-          sudo cp -rfv sources/go-piper/piper-phonemize/pi/lib/. /usr/lib/ && \
-          # Pre-build stable diffusion before we install a newer version of abseil (not compatible with stablediffusion-ncn)
-          PATH="$PATH:/root/go/bin" GO_TAGS="stablediffusion tts" GRPC_BACKENDS=backend-assets/grpc/stablediffusion make build
+          sudo cp -rfv sources/go-piper/piper-phonemize/pi/lib/. /usr/lib/
         env:
           CUDA_VERSION: 12-4
       - name: Cache grpc
@@ -130,7 +127,7 @@ jobs:
           cd grpc && cd cmake/build && sudo make --jobs 5 install
       - name: Test
         run: |
-          PATH="$PATH:/root/go/bin" GO_TAGS="stablediffusion tts" make --jobs 5 --output-sync=target test
+          PATH="$PATH:/root/go/bin" GO_TAGS="tts" make --jobs 5 --output-sync=target test
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
         uses: mxschmitt/action-tmate@v3.19
@@ -224,7 +221,7 @@ jobs:
       - name: Dependencies
         run: |
           brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
+          pip install --user --no-cache-dir grpcio-tools
       - name: Test
         run: |
           export C_INCLUDE_PATH=/usr/local/include

.gitignore

@@ -2,6 +2,7 @@
 /sources/
 __pycache__/
 *.a
+*.o
 get-sources
 prepare-sources
 /backend/cpp/llama/grpc-server
@@ -12,7 +13,6 @@ prepare-sources
 
 go-ggml-transformers
 go-gpt2
-go-rwkv
 whisper.cpp
 /bloomz
 go-bert

.vscode/launch.json

@@ -26,7 +26,7 @@
                 "LOCALAI_P2P": "true",
                 "LOCALAI_FEDERATED": "true"
             },
-            "buildFlags": ["-tags", "stablediffusion p2p tts", "-v"],
+            "buildFlags": ["-tags", "p2p tts", "-v"],
             "envFile": "${workspaceFolder}/.env",
             "cwd": "${workspaceRoot}"
         }

Dockerfile

@@ -15,8 +15,7 @@ ARG TARGETARCH
 ARG TARGETVARIANT
 
 ENV DEBIAN_FRONTEND=noninteractive
-ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,huggingface-embeddings:/build/backend/python/sentencetransformers/run.sh,transformers:/build/backend/python/transformers/run.sh,sentencetransformers:/build/backend/python/sentencetransformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,autogptq:/build/backend/python/autogptq/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,openvoice:/build/backend/python/openvoice/run.sh,vall-e-x:/build/backend/python/vall-e-x/run.sh,vllm:/build/backend/python/vllm/run.sh,mamba:/build/backend/python/mamba/run.sh,exllama2:/build/backend/python/exllama2/run.sh,transformers-musicgen:/build/backend/python/transformers-musicgen/run.sh,parler-tts:/build/backend/python/parler-tts/run.sh"
-
+ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,transformers:/build/backend/python/transformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,autogptq:/build/backend/python/autogptq/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,faster-whisper:/build/backend/python/faster-whisper/run.sh,kokoro:/build/backend/python/kokoro/run.sh,vllm:/build/backend/python/vllm/run.sh,exllama2:/build/backend/python/exllama2/run.sh"
 
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
@@ -69,14 +68,10 @@ ENV PATH=/opt/rocm/bin:${PATH}
 # OpenBLAS requirements and stable diffusion
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
-        libopenblas-dev \
-        libopencv-dev && \
+        libopenblas-dev && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
-# Set up OpenCV
-RUN ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
-
 WORKDIR /build
 
 ###################################
@@ -85,7 +80,8 @@ WORKDIR /build
 # The requirements-extras target is for any builds with IMAGE_TYPE=extras. It should not be placed in this target unless every IMAGE_TYPE=extras build will use it
 FROM requirements-core AS requirements-extras
 
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+# Install uv as a system package
+RUN curl -LsSf https://astral.sh/uv/install.sh | UV_INSTALL_DIR=/usr/bin sh
 ENV PATH="/root/.cargo/bin:${PATH}"
 
 RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
@@ -114,12 +110,13 @@ FROM requirements-${IMAGE_TYPE} AS requirements-drivers
 ARG BUILD_TYPE
 ARG CUDA_MAJOR_VERSION=12
 ARG CUDA_MINOR_VERSION=0
+ARG SKIP_DRIVERS=false
 
 ENV BUILD_TYPE=${BUILD_TYPE}
 
 # Vulkan requirements
 RUN <<EOT bash
-    if [ "${BUILD_TYPE}" = "vulkan" ]; then
+    if [ "${BUILD_TYPE}" = "vulkan" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         apt-get update && \
         apt-get install -y  --no-install-recommends \
             software-properties-common pciutils wget gpg-agent && \
@@ -135,7 +132,7 @@ EOT
 
 # CuBLAS requirements
 RUN <<EOT bash
-    if [ "${BUILD_TYPE}" = "cublas" ]; then
+    if [ "${BUILD_TYPE}" = "cublas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         apt-get update && \
         apt-get install -y  --no-install-recommends \
             software-properties-common pciutils
@@ -161,7 +158,7 @@ RUN <<EOT bash
 EOT
 
 # If we are building with clblas support, we need the libraries for the builds
-RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
+RUN if [ "${BUILD_TYPE}" = "clblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
         apt-get update && \
         apt-get install -y --no-install-recommends \
             libclblast-dev && \
@@ -169,7 +166,7 @@ RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
         rm -rf /var/lib/apt/lists/* \
     ; fi
 
-RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
+RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
         apt-get update && \
         apt-get install -y --no-install-recommends \
             hipblas-dev \
@@ -249,7 +246,7 @@ RUN git clone --recurse-submodules --jobs 4 -b ${GRPC_VERSION} --depth 1 --shall
 
 FROM requirements-drivers AS builder-base
 
-ARG GO_TAGS="stablediffusion tts p2p"
+ARG GO_TAGS="tts p2p"
 ARG GRPC_BACKENDS
 ARG MAKEFLAGS
 ARG LD_FLAGS="-s -w"
@@ -283,35 +280,12 @@ RUN <<EOT bash
     fi
 EOT
 
-
-###################################
-###################################
-
-# This first portion of builder holds the layers specifically used to build backend-assets/grpc/stablediffusion
-# In most cases, builder is the image you should be using - however, this can save build time if one just needs to copy backend-assets/grpc/stablediffusion and nothing else.
-FROM builder-base AS builder-sd
-
-# stablediffusion does not tolerate a newer version of abseil, copy only over enough elements to build it
-COPY Makefile .
-COPY go.mod .
-COPY go.sum .
-COPY backend/backend.proto ./backend/backend.proto
-COPY backend/go/image/stablediffusion ./backend/go/image/stablediffusion
-COPY pkg/grpc ./pkg/grpc
-COPY pkg/stablediffusion ./pkg/stablediffusion
-RUN git init
-RUN make sources/go-stable-diffusion
-RUN touch prepare-sources
-
-# Actually build the backend
-RUN GRPC_BACKENDS=backend-assets/grpc/stablediffusion make backend-assets/grpc/stablediffusion
-
 ###################################
 ###################################
 
 # The builder target compiles LocalAI. This target is not the target that will be uploaded to the registry.
 # Adjustments to the build process should likely be made here.
-FROM builder-sd AS builder
+FROM builder-base AS builder
 
 # Install the pre-built GRPC
 COPY --from=grpc /opt/grpc /usr/local
@@ -329,7 +303,7 @@ RUN make prepare
 ## We only leave the most CPU-optimized variant and the fallback for the cublas/hipblas build
 ## (both will use CUDA or hipblas for the actual computation)
 RUN if [ "${BUILD_TYPE}" = "cublas" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
-        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
+        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx512 backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
     else \
         make build; \
     fi
@@ -351,8 +325,6 @@ ARG FFMPEG
 
 COPY --from=grpc /opt/grpc /usr/local
 
-COPY --from=builder-sd /build/backend-assets/grpc/stablediffusion /build/backend-assets/grpc/stablediffusion
-
 COPY .devcontainer-scripts /.devcontainer-scripts
 
 # Add FFmpeg
@@ -425,36 +397,28 @@ COPY --from=builder /build/local-ai ./
 # Copy shared libraries for piper
 COPY --from=builder /build/sources/go-piper/piper-phonemize/pi/lib/* /usr/lib/
 
-# do not let stablediffusion rebuild (requires an older version of absl)
-COPY --from=builder-sd /build/backend-assets/grpc/stablediffusion ./backend-assets/grpc/stablediffusion
-
 # Change the shell to bash so we can use [[ tests below
 SHELL ["/bin/bash", "-c"]
 # We try to strike a balance between individual layer size (as that affects total push time) and total image size
 # Splitting the backends into more groups with fewer items results in a larger image, but a smaller size for the largest layer
 # Splitting the backends into fewer groups with more items results in a smaller image, but a larger size for the largest layer
 
+RUN if [[ ( "${IMAGE_TYPE}" == "extras ")]]; then \
+        apt-get -qq -y install espeak-ng \
+    ; fi
+
 RUN if [[ ( "${EXTRA_BACKENDS}" =~ "coqui" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/coqui \
     ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "parler-tts" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/parler-tts \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "faster-whisper" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/faster-whisper \
     ; fi && \
     if [[ ( "${EXTRA_BACKENDS}" =~ "diffusers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/diffusers \
-    ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "transformers-musicgen" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/transformers-musicgen \
     ; fi
 
-RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vall-e-x" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/vall-e-x \
-    ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "openvoice" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/openvoice \
-    ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "sentencetransformers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/sentencetransformers \
+RUN if [[ ( "${EXTRA_BACKENDS}" =~ "kokoro" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/kokoro \
     ; fi && \
     if [[ ( "${EXTRA_BACKENDS}" =~ "exllama2" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/exllama2 \
@@ -474,9 +438,6 @@ RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vllm" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE
     ; fi && \
     if [[ ( "${EXTRA_BACKENDS}" =~ "rerankers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/rerankers \
-    ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "mamba" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/mamba \
     ; fi
 
 # Make sure the models directory exists

Makefile

@@ -8,31 +8,27 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=0a1c750c80147687df267114c81956757cc14382
-
-# go-rwkv version
-RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
-RWKV_VERSION?=661e7ae26d442f5cfebd2a0881b44e8c55949ec6
+CPPLLAMA_VERSION?=6152129d05870cb38162c422c6ba80434e021e9f
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
-WHISPER_CPP_VERSION?=0fbaac9c891055796456df7b9122a70c220f9ca1
-
-# bert.cpp version
-BERT_REPO?=https://github.com/go-skynet/go-bert.cpp
-BERT_VERSION?=710044b124545415f555e4260d16b146c725a6e4
+WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
 
 # go-piper version
 PIPER_REPO?=https://github.com/mudler/go-piper
-PIPER_VERSION?=9d0100873a7dbb0824dfea40e8cec70a1b110759
+PIPER_VERSION?=e10ca041a885d4a8f3871d52924b47792d5e5aa0
 
-# stablediffusion version
-STABLEDIFFUSION_REPO?=https://github.com/mudler/go-stable-diffusion
-STABLEDIFFUSION_VERSION?=4a3cd6aeae6f66ee57eae9a0075f8c58c3a6a38f
+# bark.cpp
+BARKCPP_REPO?=https://github.com/PABannier/bark.cpp.git
+BARKCPP_VERSION?=v1.0.0
 
-# tinydream version
-TINYDREAM_REPO?=https://github.com/M0Rf30/go-tiny-dream
-TINYDREAM_VERSION?=c04fa463ace9d9a6464313aa5f9cd0f953b6c057
+# stablediffusion.cpp (ggml)
+STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
+STABLEDIFFUSION_GGML_VERSION?=5eb15ef4d022bef4a391de4f5f6556e81fbb5024
+
+ONNX_VERSION?=1.20.0
+ONNX_ARCH?=x64
+ONNX_OS?=linux
 
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
@@ -45,6 +41,7 @@ CGO_LDFLAGS_WHISPER+=-lggml
 CUDA_LIBPATH?=/usr/local/cuda/lib64/
 GO_TAGS?=
 BUILD_ID?=
+NATIVE?=false
 
 TEST_DIR=/tmp/test
 
@@ -83,7 +80,25 @@ ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
 
+# IF native is false, we add -DGGML_NATIVE=OFF to CMAKE_ARGS
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# Detect if we are running on arm64
+ifneq (,$(findstring aarch64,$(shell uname -m)))
+	ONNX_ARCH=aarch64
+endif
+
 ifeq ($(OS),Darwin)
+	ONNX_OS=osx
+	ifneq (,$(findstring aarch64,$(shell uname -m)))
+		ONNX_ARCH=arm64
+	else ifneq (,$(findstring arm64,$(shell uname -m)))
+		ONNX_ARCH=arm64
+	else
+		ONNX_ARCH=x86_64
+	endif
 
 	ifeq ($(OSX_SIGNING_IDENTITY),)
 		OSX_SIGNING_IDENTITY := $(shell security find-identity -v -p codesigning | grep '"' | head -n 1 | sed -E 's/.*"(.*)"/\1/')
@@ -138,10 +153,10 @@ ifeq ($(BUILD_TYPE),hipblas)
 	export CC=$(ROCM_HOME)/llvm/bin/clang
 	# llama-ggml has no hipblas support, so override it here.
 	export STABLE_BUILD_TYPE=
-	export GGML_HIPBLAS=1
+	export GGML_HIP=1
 	GPU_TARGETS ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
 	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
-	CMAKE_ARGS+=-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
+	CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
 	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link -L${ROCM_HOME}/lib/llvm/lib
 endif
 
@@ -160,16 +175,6 @@ ifeq ($(STATIC),true)
 	LD_FLAGS+=-linkmode external -extldflags -static
 endif
 
-ifeq ($(findstring stablediffusion,$(GO_TAGS)),stablediffusion)
-#	OPTIONAL_TARGETS+=go-stable-diffusion/libstablediffusion.a
-	OPTIONAL_GRPC+=backend-assets/grpc/stablediffusion
-endif
-
-ifeq ($(findstring tinydream,$(GO_TAGS)),tinydream)
-#	OPTIONAL_TARGETS+=go-tiny-dream/libtinydream.a
-	OPTIONAL_GRPC+=backend-assets/grpc/tinydream
-endif
-
 ifeq ($(findstring tts,$(GO_TAGS)),tts)
 #	OPTIONAL_TARGETS+=go-piper/libpiper_binding.a
 #	OPTIONAL_TARGETS+=backend-assets/espeak-ng-data
@@ -179,16 +184,24 @@ ifeq ($(findstring tts,$(GO_TAGS)),tts)
 endif
 
 ALL_GRPC_BACKENDS=backend-assets/grpc/huggingface
-ALL_GRPC_BACKENDS+=backend-assets/grpc/bert-embeddings
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx2
+ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx512
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-fallback
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-ggml
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-grpc
 ALL_GRPC_BACKENDS+=backend-assets/util/llama-cpp-rpc-server
-ALL_GRPC_BACKENDS+=backend-assets/grpc/rwkv
 ALL_GRPC_BACKENDS+=backend-assets/grpc/whisper
+
+ifeq ($(ONNX_OS),linux)
+ifeq ($(ONNX_ARCH),x64)
+	ALL_GRPC_BACKENDS+=backend-assets/grpc/bark-cpp
+	ALL_GRPC_BACKENDS+=backend-assets/grpc/stablediffusion-ggml
+endif
+endif
+
 ALL_GRPC_BACKENDS+=backend-assets/grpc/local-store
+ALL_GRPC_BACKENDS+=backend-assets/grpc/silero-vad
 ALL_GRPC_BACKENDS+=$(OPTIONAL_GRPC)
 # Use filter-out to remove the specified backends
 ALL_GRPC_BACKENDS := $(filter-out $(SKIP_GRPC_BACKEND),$(ALL_GRPC_BACKENDS))
@@ -209,19 +222,6 @@ endif
 
 all: help
 
-## BERT embeddings
-sources/go-bert.cpp:
-	mkdir -p sources/go-bert.cpp
-	cd sources/go-bert.cpp && \
-	git init && \
-	git remote add origin $(BERT_REPO) && \
-	git fetch origin && \
-	git checkout $(BERT_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-sources/go-bert.cpp/libgobert.a: sources/go-bert.cpp
-	$(MAKE) -C sources/go-bert.cpp libgobert.a
-
 ## go-llama.cpp
 sources/go-llama.cpp:
 	mkdir -p sources/go-llama.cpp
@@ -235,6 +235,23 @@ sources/go-llama.cpp:
 sources/go-llama.cpp/libbinding.a: sources/go-llama.cpp
 	$(MAKE) -C sources/go-llama.cpp BUILD_TYPE=$(STABLE_BUILD_TYPE) libbinding.a
 
+## bark.cpp
+sources/bark.cpp:
+	git clone --recursive $(BARKCPP_REPO) sources/bark.cpp && \
+	cd sources/bark.cpp && \
+	git checkout $(BARKCPP_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/bark.cpp/build/libbark.a: sources/bark.cpp
+	cd sources/bark.cpp && \
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) .. && \
+	cmake --build . --config Release
+
+backend/go/bark/libbark.a: sources/bark.cpp/build/libbark.a
+	$(MAKE) -C backend/go/bark libbark.a
+
 ## go-piper
 sources/go-piper:
 	mkdir -p sources/go-piper
@@ -248,45 +265,37 @@ sources/go-piper:
 sources/go-piper/libpiper_binding.a: sources/go-piper
 	$(MAKE) -C sources/go-piper libpiper_binding.a example/main piper.o
 
-
-## RWKV
-sources/go-rwkv.cpp:
-	mkdir -p sources/go-rwkv.cpp
-	cd sources/go-rwkv.cpp && \
-	git init && \
-	git remote add origin $(RWKV_REPO) && \
-	git fetch origin && \
-	git checkout $(RWKV_VERSION) && \
+## stablediffusion (ggml)
+sources/stablediffusion-ggml.cpp:
+	git clone --recursive $(STABLEDIFFUSION_GGML_REPO) sources/stablediffusion-ggml.cpp && \
+	cd sources/stablediffusion-ggml.cpp && \
+	git checkout $(STABLEDIFFUSION_GGML_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-sources/go-rwkv.cpp/librwkv.a: sources/go-rwkv.cpp
-	cd sources/go-rwkv.cpp && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
+backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp
+	$(MAKE) -C backend/go/image/stablediffusion-ggml build/libstable-diffusion.a
+	$(MAKE) -C backend/go/image/stablediffusion-ggml libsd.a
 
-## stable diffusion
-sources/go-stable-diffusion:
-	mkdir -p sources/go-stable-diffusion
-	cd sources/go-stable-diffusion && \
-	git init && \
-	git remote add origin $(STABLEDIFFUSION_REPO) && \
-	git fetch origin && \
-	git checkout $(STABLEDIFFUSION_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
+backend-assets/grpc/stablediffusion-ggml: backend/go/image/stablediffusion-ggml/libsd.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ LIBRARY_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion-ggml ./backend/go/image/stablediffusion-ggml/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/stablediffusion-ggml
+endif
 
-sources/go-stable-diffusion/libstablediffusion.a: sources/go-stable-diffusion
-	CPATH="$(CPATH):/usr/include/opencv4" $(MAKE) -C sources/go-stable-diffusion libstablediffusion.a
+sources/onnxruntime:
+	mkdir -p sources/onnxruntime
+	curl -L https://github.com/microsoft/onnxruntime/releases/download/v$(ONNX_VERSION)/onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz -o sources/onnxruntime/onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz
+	cd sources/onnxruntime && tar -xvf onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz && rm onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz
+	cd sources/onnxruntime && mv onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION)/* ./
 
-## tiny-dream
-sources/go-tiny-dream:
-	mkdir -p sources/go-tiny-dream
-	cd sources/go-tiny-dream && \
-	git init && \
-	git remote add origin $(TINYDREAM_REPO) && \
-	git fetch origin && \
-	git checkout $(TINYDREAM_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-sources/go-tiny-dream/libtinydream.a: sources/go-tiny-dream
-	$(MAKE) -C sources/go-tiny-dream libtinydream.a
+backend-assets/lib/libonnxruntime.so.1: backend-assets/lib sources/onnxruntime
+	cp -rfv sources/onnxruntime/lib/* backend-assets/lib/
+ifeq ($(OS),Darwin)
+	mv backend-assets/lib/libonnxruntime.$(ONNX_VERSION).dylib backend-assets/lib/libonnxruntime.dylib
+else
+	mv backend-assets/lib/libonnxruntime.so.$(ONNX_VERSION) backend-assets/lib/libonnxruntime.so.1
+endif
 
 ## whisper
 sources/whisper.cpp:
@@ -301,26 +310,18 @@ sources/whisper.cpp:
 sources/whisper.cpp/libwhisper.a: sources/whisper.cpp
 	cd sources/whisper.cpp && $(MAKE) libwhisper.a libggml.a
 
-get-sources: sources/go-llama.cpp sources/go-piper sources/go-rwkv.cpp sources/whisper.cpp sources/go-bert.cpp sources/go-stable-diffusion sources/go-tiny-dream backend/cpp/llama/llama.cpp
+get-sources: sources/go-llama.cpp sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
 
 replace:
-	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(CURDIR)/sources/go-rwkv.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(CURDIR)/sources/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp/bindings/go=$(CURDIR)/sources/whisper.cpp/bindings/go
-	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(CURDIR)/sources/go-bert.cpp
-	$(GOCMD) mod edit -replace github.com/M0Rf30/go-tiny-dream=$(CURDIR)/sources/go-tiny-dream
 	$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(CURDIR)/sources/go-piper
-	$(GOCMD) mod edit -replace github.com/mudler/go-stable-diffusion=$(CURDIR)/sources/go-stable-diffusion
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(CURDIR)/sources/go-llama.cpp
 
 dropreplace:
-	$(GOCMD) mod edit -dropreplace github.com/donomii/go-rwkv.cpp
 	$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp
 	$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp/bindings/go
-	$(GOCMD) mod edit -dropreplace github.com/go-skynet/go-bert.cpp
-	$(GOCMD) mod edit -dropreplace github.com/M0Rf30/go-tiny-dream
 	$(GOCMD) mod edit -dropreplace github.com/mudler/go-piper
-	$(GOCMD) mod edit -dropreplace github.com/mudler/go-stable-diffusion
 	$(GOCMD) mod edit -dropreplace github.com/go-skynet/go-llama.cpp
 
 prepare-sources: get-sources replace
@@ -330,12 +331,8 @@ prepare-sources: get-sources replace
 rebuild: ## Rebuilds the project
 	$(GOCMD) clean -cache
 	$(MAKE) -C sources/go-llama.cpp clean
-	$(MAKE) -C sources/go-rwkv.cpp clean
 	$(MAKE) -C sources/whisper.cpp clean
-	$(MAKE) -C sources/go-stable-diffusion clean
-	$(MAKE) -C sources/go-bert.cpp clean
 	$(MAKE) -C sources/go-piper clean
-	$(MAKE) -C sources/go-tiny-dream clean
 	$(MAKE) build
 
 prepare: prepare-sources $(OPTIONAL_TARGETS)
@@ -348,7 +345,9 @@ clean: ## Remove build related file
 	rm -rf release/
 	rm -rf backend-assets/*
 	$(MAKE) -C backend/cpp/grpc clean
+	$(MAKE) -C backend/go/bark clean
 	$(MAKE) -C backend/cpp/llama clean
+	$(MAKE) -C backend/go/image/stablediffusion-ggml clean
 	rm -rf backend/cpp/llama-* || true
 	$(MAKE) dropreplace
 	$(MAKE) protogen-clean
@@ -439,8 +438,6 @@ test-models/testmodel.ggml:
 	wget -q https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
 	wget -q https://huggingface.co/mudler/all-MiniLM-L6-v2/resolve/main/ggml-model-q4_0.bin -O test-models/bert
 	wget -q https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
-	wget -q https://huggingface.co/mudler/rwkv-4-raven-1.5B-ggml/resolve/main/RWKV-4-Raven-1B5-v11-Eng99%2525-Other1%2525-20230425-ctx4096_Q4_0.bin -O test-models/rwkv
-	wget -q https://raw.githubusercontent.com/saharNooby/rwkv.cpp/5eb8f09c146ea8124633ab041d9ea0b1f1db4459/rwkv/20B_tokenizer.json -O test-models/rwkv.tokenizer.json
 	cp tests/models_fixtures/* test-models
 
 prepare-test: grpcs
@@ -449,9 +446,9 @@ prepare-test: grpcs
 
 test: prepare test-models/testmodel.ggml grpcs
 	@echo 'Running tests'
-	export GO_TAGS="tts stablediffusion debug"
+	export GO_TAGS="tts debug"
 	$(MAKE) prepare-test
-	HUGGINGFACE_GRPC=$(abspath ./)/backend/python/sentencetransformers/run.sh TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	HUGGINGFACE_GRPC=$(abspath ./)/backend/python/transformers/run.sh TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
 	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama && !llama-gguf"  --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
 	$(MAKE) test-llama
 	$(MAKE) test-llama-gguf
@@ -537,10 +534,10 @@ protogen-go-clean:
 	$(RM) bin/*
 
 .PHONY: protogen-python
-protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen mamba-protogen rerankers-protogen sentencetransformers-protogen transformers-protogen parler-tts-protogen transformers-musicgen-protogen vall-e-x-protogen vllm-protogen openvoice-protogen
+protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
 
 .PHONY: protogen-python-clean
-protogen-python-clean: autogptq-protogen-clean bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean mamba-protogen-clean sentencetransformers-protogen-clean rerankers-protogen-clean transformers-protogen-clean transformers-musicgen-protogen-clean parler-tts-protogen-clean vall-e-x-protogen-clean vllm-protogen-clean openvoice-protogen-clean
+protogen-python-clean: autogptq-protogen-clean bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean rerankers-protogen-clean transformers-protogen-clean kokoro-protogen-clean vllm-protogen-clean faster-whisper-protogen-clean
 
 .PHONY: autogptq-protogen
 autogptq-protogen:
@@ -574,6 +571,14 @@ diffusers-protogen:
 diffusers-protogen-clean:
 	$(MAKE) -C backend/python/diffusers protogen-clean
 
+.PHONY: faster-whisper-protogen
+faster-whisper-protogen:
+	$(MAKE) -C backend/python/faster-whisper protogen
+
+.PHONY: faster-whisper-protogen-clean
+faster-whisper-protogen-clean:
+	$(MAKE) -C backend/python/faster-whisper protogen-clean
+
 .PHONY: exllama2-protogen
 exllama2-protogen:
 	$(MAKE) -C backend/python/exllama2 protogen
@@ -582,14 +587,6 @@ exllama2-protogen:
 exllama2-protogen-clean:
 	$(MAKE) -C backend/python/exllama2 protogen-clean
 
-.PHONY: mamba-protogen
-mamba-protogen:
-	$(MAKE) -C backend/python/mamba protogen
-
-.PHONY: mamba-protogen-clean
-mamba-protogen-clean:
-	$(MAKE) -C backend/python/mamba protogen-clean
-
 .PHONY: rerankers-protogen
 rerankers-protogen:
 	$(MAKE) -C backend/python/rerankers protogen
@@ -598,14 +595,6 @@ rerankers-protogen:
 rerankers-protogen-clean:
 	$(MAKE) -C backend/python/rerankers protogen-clean
 
-.PHONY: sentencetransformers-protogen
-sentencetransformers-protogen:
-	$(MAKE) -C backend/python/sentencetransformers protogen
-
-.PHONY: sentencetransformers-protogen-clean
-sentencetransformers-protogen-clean:
-	$(MAKE) -C backend/python/sentencetransformers protogen-clean
-
 .PHONY: transformers-protogen
 transformers-protogen:
 	$(MAKE) -C backend/python/transformers protogen
@@ -614,37 +603,13 @@ transformers-protogen:
 transformers-protogen-clean:
 	$(MAKE) -C backend/python/transformers protogen-clean
 
-.PHONY: parler-tts-protogen
-parler-tts-protogen:
-	$(MAKE) -C backend/python/parler-tts protogen
+.PHONY: kokoro-protogen
+kokoro-protogen:
+	$(MAKE) -C backend/python/kokoro protogen
 
-.PHONY: parler-tts-protogen-clean
-parler-tts-protogen-clean:
-	$(MAKE) -C backend/python/parler-tts protogen-clean
-
-.PHONY: transformers-musicgen-protogen
-transformers-musicgen-protogen:
-	$(MAKE) -C backend/python/transformers-musicgen protogen
-
-.PHONY: transformers-musicgen-protogen-clean
-transformers-musicgen-protogen-clean:
-	$(MAKE) -C backend/python/transformers-musicgen protogen-clean
-
-.PHONY: vall-e-x-protogen
-vall-e-x-protogen:
-	$(MAKE) -C backend/python/vall-e-x protogen
-
-.PHONY: vall-e-x-protogen-clean
-vall-e-x-protogen-clean:
-	$(MAKE) -C backend/python/vall-e-x protogen-clean
-
-.PHONY: openvoice-protogen
-openvoice-protogen:
-	$(MAKE) -C backend/python/openvoice protogen
-
-.PHONY: openvoice-protogen-clean
-openvoice-protogen-clean:
-	$(MAKE) -C backend/python/openvoice protogen-clean
+.PHONY: kokoro-protogen-clean
+kokoro-protogen-clean:
+	$(MAKE) -C backend/python/kokoro protogen-clean
 
 .PHONY: vllm-protogen
 vllm-protogen:
@@ -661,15 +626,11 @@ prepare-extra-conda-environments: protogen-python
 	$(MAKE) -C backend/python/bark
 	$(MAKE) -C backend/python/coqui
 	$(MAKE) -C backend/python/diffusers
+	$(MAKE) -C backend/python/faster-whisper
 	$(MAKE) -C backend/python/vllm
-	$(MAKE) -C backend/python/mamba
-	$(MAKE) -C backend/python/sentencetransformers
 	$(MAKE) -C backend/python/rerankers
 	$(MAKE) -C backend/python/transformers
-	$(MAKE) -C backend/python/transformers-musicgen
-	$(MAKE) -C backend/python/parler-tts
-	$(MAKE) -C backend/python/vall-e-x
-	$(MAKE) -C backend/python/openvoice
+	$(MAKE) -C backend/python/kokoro
 	$(MAKE) -C backend/python/exllama2
 
 prepare-test-extra: protogen-python
@@ -693,13 +654,6 @@ backend-assets/espeak-ng-data: sources/go-piper sources/go-piper/libpiper_bindin
 backend-assets/grpc: protogen-go replace
 	mkdir -p backend-assets/grpc
 
-backend-assets/grpc/bert-embeddings: sources/go-bert.cpp sources/go-bert.cpp/libgobert.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-bert.cpp LIBRARY_PATH=$(CURDIR)/sources/go-bert.cpp \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bert-embeddings ./backend/go/llm/bert/
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/bert-embeddings
-endif
-
 backend-assets/grpc/huggingface: backend-assets/grpc
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/huggingface ./backend/go/llm/langchain/
 ifneq ($(UPX),)
@@ -746,6 +700,13 @@ backend-assets/grpc/llama-cpp-avx2: backend-assets/grpc backend/cpp/llama/llama.
 	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on" $(MAKE) VARIANT="llama-avx2" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-avx2/grpc-server backend-assets/grpc/llama-cpp-avx2
 
+backend-assets/grpc/llama-cpp-avx512: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-avx512
+	$(MAKE) -C backend/cpp/llama-avx512 purge
+	$(info ${GREEN}I llama-cpp build info:avx512${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=on -DGGML_FMA=on -DGGML_F16C=on" $(MAKE) VARIANT="llama-avx512" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-avx512/grpc-server backend-assets/grpc/llama-cpp-avx512
+
 backend-assets/grpc/llama-cpp-avx: backend-assets/grpc backend/cpp/llama/llama.cpp
 	cp -rf backend/cpp/llama backend/cpp/llama-avx
 	$(MAKE) -C backend/cpp/llama-avx purge
@@ -759,10 +720,6 @@ backend-assets/grpc/llama-cpp-fallback: backend-assets/grpc backend/cpp/llama/ll
 	$(info ${GREEN}I llama-cpp build info:fallback${RESET})
 	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" $(MAKE) VARIANT="llama-fallback" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-fallback/grpc-server backend-assets/grpc/llama-cpp-fallback
-# TODO: every binary should have its own folder instead, so can have different metal implementations
-ifeq ($(BUILD_TYPE),metal)
-	cp backend/cpp/llama-fallback/llama.cpp/build/bin/default.metallib backend-assets/grpc/
-endif
 
 backend-assets/grpc/llama-cpp-cuda: backend-assets/grpc backend/cpp/llama/llama.cpp
 	cp -rf backend/cpp/llama backend/cpp/llama-cuda
@@ -775,7 +732,7 @@ backend-assets/grpc/llama-cpp-hipblas: backend-assets/grpc backend/cpp/llama/lla
 	cp -rf backend/cpp/llama backend/cpp/llama-hipblas
 	$(MAKE) -C backend/cpp/llama-hipblas purge
 	$(info ${GREEN}I llama-cpp build info:hipblas${RESET})
-	BUILD_TYPE="hipblas" $(MAKE) VARIANT="llama-hipblas" build-llama-cpp-grpc-server
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" BUILD_TYPE="hipblas" $(MAKE) VARIANT="llama-hipblas" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-hipblas/grpc-server backend-assets/grpc/llama-cpp-hipblas
 
 backend-assets/grpc/llama-cpp-sycl_f16: backend-assets/grpc backend/cpp/llama/llama.cpp
@@ -810,6 +767,13 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/llama-ggml
 endif
 
+backend-assets/grpc/bark-cpp: backend/go/bark/libbark.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/bark/ LIBRARY_PATH=$(CURDIR)/backend/go/bark/ \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bark-cpp ./backend/go/bark/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/bark-cpp
+endif
+
 backend-assets/grpc/piper: sources/go-piper sources/go-piper/libpiper_binding.a backend-assets/grpc backend-assets/espeak-ng-data
 	CGO_CXXFLAGS="$(PIPER_CGO_CXXFLAGS)" CGO_LDFLAGS="$(PIPER_CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/sources/go-piper \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/piper ./backend/go/tts/
@@ -817,25 +781,11 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/piper
 endif
 
-backend-assets/grpc/rwkv: sources/go-rwkv.cpp sources/go-rwkv.cpp/librwkv.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-rwkv.cpp LIBRARY_PATH=$(CURDIR)/sources/go-rwkv.cpp \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/rwkv ./backend/go/llm/rwkv
+backend-assets/grpc/silero-vad: backend-assets/grpc backend-assets/lib/libonnxruntime.so.1
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" CPATH="$(CPATH):$(CURDIR)/sources/onnxruntime/include/" LIBRARY_PATH=$(CURDIR)/backend-assets/lib \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/silero-vad ./backend/go/vad/silero
 ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/rwkv
-endif
-
-backend-assets/grpc/stablediffusion: sources/go-stable-diffusion sources/go-stable-diffusion/libstablediffusion.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" CPATH="$(CPATH):$(CURDIR)/sources/go-stable-diffusion/:/usr/include/opencv4" LIBRARY_PATH=$(CURDIR)/sources/go-stable-diffusion/ \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion ./backend/go/image/stablediffusion
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/stablediffusion
-endif
-
-backend-assets/grpc/tinydream: sources/go-tiny-dream sources/go-tiny-dream/libtinydream.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/go-tiny-dream \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/tinydream ./backend/go/image/tinydream
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/tinydream
+	$(UPX) backend-assets/grpc/silero-vad
 endif
 
 backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/libwhisper.a backend-assets/grpc
@@ -891,7 +841,7 @@ docker-aio-all:
 
 docker-image-intel:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-devel-ubuntu22.04 \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
@@ -899,7 +849,7 @@ docker-image-intel:
 
 docker-image-intel-xpu:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-devel-ubuntu22.04 \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \

README.md

@@ -38,9 +38,13 @@
 </a>
 </p>
 
+<p align="center">
+<a href="https://trendshift.io/repositories/5539" target="_blank"><img src="https://trendshift.io/api/badge/repositories/5539" alt="mudler%2FLocalAI | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</p>
+
 > :bulb: Get help - [❓FAQ](https://localai.io/faq/) [💭Discussions](https://github.com/go-skynet/LocalAI/discussions) [:speech_balloon: Discord](https://discord.gg/uJAeKSAGDy) [:book: Documentation website](https://localai.io/)
 >
-> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/go-skynet/LocalAI/tree/master/examples/) 
+> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) 
 
 [![tests](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml)[![Build and Release](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml)[![build container images](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml)[![Bump dependencies](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml)[![Artifact Hub](https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/localai)](https://artifacthub.io/packages/search?repo=localai)
 
@@ -56,14 +60,17 @@ curl https://localai.io/install.sh | sh
 
 Or run with docker:
 ```bash
+# CPU only image:
+docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-cpu
+
+# Nvidia GPU:
+docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-gpu-nvidia-cuda-12
+
+# CPU and GPU image (bigger size):
+docker run -ti --name local-ai -p 8080:8080 localai/localai:latest
+
+# AIO images (it will pre-download a set of models ready for use, see https://localai.io/basics/container/)
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-cpu
-# Alternative images:
-# - if you have an Nvidia GPU:
-# docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-aio-gpu-nvidia-cuda-12
-# - without preconfigured models
-# docker run -ti --name local-ai -p 8080:8080 localai/localai:latest
-# - without preconfigured models for Nvidia GPUs
-# docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-gpu-nvidia-cuda-12 
 ```
 
 To load models:
@@ -85,15 +92,15 @@ local-ai run oci://localai/phi-2:latest
 
 ## 📰 Latest project news
 
+- Jan 2025: LocalAI model release: https://huggingface.co/mudler/LocalAI-functioncall-phi-4-v0.3, SANA support in diffusers: https://github.com/mudler/LocalAI/pull/4603
+- Dec 2024: stablediffusion.cpp backend (ggml) added ( https://github.com/mudler/LocalAI/pull/4289 )
+- Nov 2024: Bark.cpp backend added ( https://github.com/mudler/LocalAI/pull/4287 )
+- Nov 2024: Voice activity detection models (**VAD**) added to the API: https://github.com/mudler/LocalAI/pull/4204
+- Oct 2024: examples moved to [LocalAI-examples](https://github.com/mudler/LocalAI-examples)
 - Aug 2024:  🆕 FLUX-1, [P2P Explorer](https://explorer.localai.io)
-- July 2024: 🔥🔥 🆕 P2P Dashboard, LocalAI Federated mode and AI Swarms: https://github.com/mudler/LocalAI/pull/2723
-- June 2024: 🆕 You can browse now the model gallery without LocalAI! Check out https://models.localai.io
-- June 2024: Support for models from OCI registries: https://github.com/mudler/LocalAI/pull/2628
+- July 2024: 🔥🔥 🆕 P2P Dashboard, LocalAI Federated mode and AI Swarms: https://github.com/mudler/LocalAI/pull/2723. P2P Global community pools: https://github.com/mudler/LocalAI/issues/3113
 - May 2024: 🔥🔥 Decentralized P2P llama.cpp:  https://github.com/mudler/LocalAI/pull/2343 (peer2peer llama.cpp!) 👉 Docs  https://localai.io/features/distribute/
-- May 2024: 🔥🔥 Openvoice: https://github.com/mudler/LocalAI/pull/2334
-- May 2024: 🆕 Function calls without grammars and mixed mode: https://github.com/mudler/LocalAI/pull/2328
 - May 2024: 🔥🔥 Distributed inferencing: https://github.com/mudler/LocalAI/pull/2324
-- May 2024: Chat, TTS, and Image generation in the WebUI: https://github.com/mudler/LocalAI/pull/2222
 - April 2024: Reranker API: https://github.com/mudler/LocalAI/pull/2121
 
 Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap)
@@ -102,12 +109,10 @@ Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3A
 
 - Multimodal with vLLM and Video understanding: https://github.com/mudler/LocalAI/pull/3729
 - Realtime API https://github.com/mudler/LocalAI/issues/3714
-- 🔥🔥 Distributed, P2P Global community pools: https://github.com/mudler/LocalAI/issues/3113
 - WebUI improvements: https://github.com/mudler/LocalAI/issues/2156
 - Backends v2: https://github.com/mudler/LocalAI/issues/1126
 - Improving UX v2: https://github.com/mudler/LocalAI/issues/1373
 - Assistant API: https://github.com/mudler/LocalAI/issues/1273
-- Moderation endpoint: https://github.com/mudler/LocalAI/issues/999
 - Vulkan: https://github.com/mudler/LocalAI/issues/1647
 - Anthropic API: https://github.com/mudler/LocalAI/issues/1808
 
@@ -115,10 +120,10 @@ If you want to help and contribute, issues up for grabs: https://github.com/mudl
 
 ## 🚀 [Features](https://localai.io/features/)
 
-- 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `gpt4all.cpp`, ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
+- 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `transformers`, `vllm` ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
 - 🗣 [Text to Audio](https://localai.io/features/text-to-audio/)
 - 🔈 [Audio to Text](https://localai.io/features/audio-to-text/) (Audio transcription with `whisper.cpp`)
-- 🎨 [Image generation with stable diffusion](https://localai.io/features/image-generation)
+- 🎨 [Image generation](https://localai.io/features/image-generation)
 - 🔥 [OpenAI-alike tools API](https://localai.io/features/openai-functions/) 
 - 🧠 [Embeddings generation for vector databases](https://localai.io/features/embeddings/)
 - ✍️ [Constrained grammars](https://localai.io/features/constrained_grammars/)
@@ -126,6 +131,7 @@ If you want to help and contribute, issues up for grabs: https://github.com/mudl
 - 🥽 [Vision API](https://localai.io/features/gpt-vision/)
 - 📈 [Reranker API](https://localai.io/features/reranker/)
 - 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
+- 🔊 Voice activity detection (Silero-VAD support)
 - 🌍 Integrated WebUI!
 
 ## 💻 Usage
@@ -148,6 +154,7 @@ Model galleries
 Other:
 - Helm chart https://github.com/go-skynet/helm-charts
 - VSCode extension https://github.com/badgooooor/localai-vscode-plugin
+- Langchain: https://python.langchain.com/docs/integrations/providers/localai/
 - Terminal utility https://github.com/djcopley/ShellOracle
 - Local Smart assistant https://github.com/mudler/LocalAGI
 - Home Assistant https://github.com/sammcj/homeassistant-localai / https://github.com/drndos/hass-openai-custom-conversation / https://github.com/valentinfrlch/ha-gpt4vision
@@ -155,6 +162,9 @@ Other:
 - Slack bot https://github.com/mudler/LocalAGI/tree/main/examples/slack
 - Shell-Pilot(Interact with LLM using LocalAI models via pure shell scripts on your Linux or MacOS system) https://github.com/reid41/shell-pilot
 - Telegram bot https://github.com/mudler/LocalAI/tree/master/examples/telegram-bot
+- Another Telegram Bot https://github.com/JackBekket/Hellper
+- Auto-documentation https://github.com/JackBekket/Reflexia
+- Github bot which answer on issues, with code and documentation as context https://github.com/JackBekket/GitHelper
 - Github Actions: https://github.com/marketplace/actions/start-localai
 - Examples: https://github.com/mudler/LocalAI/tree/master/examples/
   
@@ -229,7 +239,6 @@ LocalAI couldn't have been built without the help of great software already avai
 - https://github.com/antimatter15/alpaca.cpp
 - https://github.com/EdVince/Stable-Diffusion-NCNN
 - https://github.com/ggerganov/whisper.cpp
-- https://github.com/saharNooby/rwkv.cpp
 - https://github.com/rhasspy/piper
 
 ## 🤗 Contributors

aio/cpu/embeddings.yaml

@@ -1,7 +1,7 @@
 name: text-embedding-ada-002
-backend: bert-embeddings
+embeddings: true
 parameters:
-  model: huggingface://mudler/all-MiniLM-L6-v2/ggml-model-q4_0.bin
+  model: huggingface://hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF/llama-3.2-1b-instruct-q4_k_m.gguf
 
 usage: |
     You can test this model with curl like this:

aio/cpu/image-gen.yaml

@@ -1,56 +1,17 @@
 name: stablediffusion
-backend: stablediffusion
+backend: stablediffusion-ggml
+cfg_scale: 4.5
+
+options:
+- sampler:euler
 parameters:
-  model: stablediffusion_assets
-
-license: "BSD-3"
-urls:
-- https://github.com/EdVince/Stable-Diffusion-NCNN
-- https://github.com/EdVince/Stable-Diffusion-NCNN/blob/main/LICENSE
-
-description: |
-     Stable Diffusion in NCNN with c++, supported txt2img and img2img
+  model: stable-diffusion-v1-5-pruned-emaonly-Q4_0.gguf
+step: 25
 
 download_files:
-- filename: "stablediffusion_assets/AutoencoderKL-256-256-fp16-opt.param"
-  sha256: "18ca4b66685e21406bcf64c484b3b680b4949900415536d599cc876579c85c82"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-256-256-fp16-opt.param"
-- filename: "stablediffusion_assets/AutoencoderKL-512-512-fp16-opt.param"
-  sha256: "cf45f63aacf3dbbab0f59ed92a6f2c14d9a1801314631cd3abe91e3c85639a20"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-512-512-fp16-opt.param"
-- filename: "stablediffusion_assets/AutoencoderKL-base-fp16.param"
-  sha256: "0254a056dce61b0c27dc9ec1b78b53bcf55315c540f55f051eb841aa992701ba"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-base-fp16.param"
-- filename: "stablediffusion_assets/AutoencoderKL-encoder-512-512-fp16.bin"
-  sha256: "ddcb79a9951b9f91e05e087739ed69da2c1c4ae30ba4168cce350b49d617c9fa"
-  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/AutoencoderKL-encoder-512-512-fp16.bin"
-- filename: "stablediffusion_assets/AutoencoderKL-fp16.bin"
-  sha256: "f02e71f80e70252734724bbfaed5c4ddd3a8ed7e61bb2175ff5f53099f0e35dd"
-  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/AutoencoderKL-fp16.bin"
-- filename: "stablediffusion_assets/FrozenCLIPEmbedder-fp16.bin"
-  sha256: "1c9a12f4e1dd1b295a388045f7f28a2352a4d70c3dc96a542189a3dd7051fdd6"
-  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/FrozenCLIPEmbedder-fp16.bin"
-- filename: "stablediffusion_assets/FrozenCLIPEmbedder-fp16.param"
-  sha256: "471afbe678dd1fd3fe764ef9c6eccaccb0a7d7e601f27b462aa926b20eb368c9"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/FrozenCLIPEmbedder-fp16.param"
-- filename: "stablediffusion_assets/log_sigmas.bin"
-  sha256: "a2089f8aa4c61f9c200feaec541ab3f5c94233b28deb6d5e8bcd974fa79b68ac"
-  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/raw/main/x86/linux/assets/log_sigmas.bin"
-- filename: "stablediffusion_assets/UNetModel-256-256-MHA-fp16-opt.param"
-  sha256: "a58c380229f09491776df837b7aa7adffc0a87821dc4708b34535da2e36e3da1"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-256-256-MHA-fp16-opt.param"
-- filename: "stablediffusion_assets/UNetModel-512-512-MHA-fp16-opt.param"
-  sha256: "f12034067062827bd7f43d1d21888d1f03905401acf6c6eea22be23c259636fa"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-512-512-MHA-fp16-opt.param"
-- filename: "stablediffusion_assets/UNetModel-base-MHA-fp16.param"
-  sha256: "696f6975de49f4325b53ce32aff81861a6d6c07cd9ce3f0aae2cc405350af38d"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-base-MHA-fp16.param"
-- filename: "stablediffusion_assets/UNetModel-MHA-fp16.bin"
-  sha256: "d618918d011bfc1f644c0f2a33bf84931bd53b28a98492b0a8ed6f3a818852c3"
-  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/UNetModel-MHA-fp16.bin"
-- filename: "stablediffusion_assets/vocab.txt"
-  sha256: "e30e57b6f1e47616982ef898d8922be24e535b4fa3d0110477b3a6f02ebbae7d"
-  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/vocab.txt"
+- filename: "stable-diffusion-v1-5-pruned-emaonly-Q4_0.gguf"
+  sha256: "b8944e9fe0b69b36ae1b5bb0185b3a7b8ef14347fe0fa9af6c64c4829022261f"
+  uri: "huggingface://second-state/stable-diffusion-v1-5-GGUF/stable-diffusion-v1-5-pruned-emaonly-Q4_0.gguf"
 
 usage: |
         curl http://localhost:8080/v1/images/generations \

backend/backend.proto

@@ -28,6 +28,8 @@ service Backend {
   rpc Rerank(RerankRequest) returns (RerankResult) {}
 
   rpc GetMetrics(MetricsRequest) returns (MetricsResponse);
+
+  rpc VAD(VADRequest) returns (VADResponse) {}
 }
 
 // Define the empty request
@@ -157,6 +159,8 @@ message Reply {
   bytes message = 1;
   int32 tokens = 2;
   int32 prompt_tokens = 3;
+  double timing_prompt_processing = 4;
+  double timing_token_generation = 5;
 }
 
 message ModelOptions {
@@ -233,6 +237,16 @@ message ModelOptions {
 
   bool FlashAttention = 56;
   bool NoKVOffload = 57;
+
+  string ModelPath = 59;
+
+  repeated string LoraAdapters = 60;
+  repeated float LoraScales = 61;
+
+  repeated string Options = 62;
+
+  string CacheTypeKey = 63;
+  string CacheTypeValue = 64;
 }
 
 message Result {
@@ -288,6 +302,19 @@ message TTSRequest {
   optional string language = 5;
 }
 
+message VADRequest {
+  repeated float audio = 1;
+}
+
+message VADSegment {
+  float start = 1;
+  float end = 2;
+}
+
+message VADResponse {
+  repeated VADSegment segments = 1;
+}
+
 message SoundGenerationRequest {
   string text = 1;
   string model = 2;
@@ -323,4 +350,4 @@ message StatusResponse {
 message Message {
   string role = 1;
   string content = 2;
-}
+}

backend/cpp/llama/Makefile

@@ -22,7 +22,7 @@ else ifeq ($(BUILD_TYPE),clblas)
 	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
 # If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
 else ifeq ($(BUILD_TYPE),hipblas)
-	CMAKE_ARGS+=-DGGML_HIPBLAS=ON
+	CMAKE_ARGS+=-DGGML_HIP=ON
 # If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
 # But if it's OSX without metal, disable it here
 else ifeq ($(OS),Darwin)
@@ -30,9 +30,7 @@ else ifeq ($(OS),Darwin)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
 		CMAKE_ARGS+=-DGGML_METAL=ON
-# Until this is tested properly, we disable embedded metal file
-# as we already embed it as part of the LocalAI assets
-		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=OFF
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
 		TARGET+=--target ggml-metal
 	endif
 endif

backend/cpp/llama/grpc-server.cpp

@@ -22,6 +22,7 @@
 #include "backend.grpc.pb.h"
 #include "utils.hpp"
 #include "sampling.h"
+#include "speculative.h"
 // include std::regex
 #include <cstddef>
 #include <thread>
@@ -134,6 +135,32 @@ static std::string tokens_to_output_formatted_string(const llama_context *ctx, c
     return out;
 }
 
+// Adds an RPC server
+// https://github.com/ggerganov/llama.cpp/compare/4dbc8b9cb71876e005724f4e8f73a3544646bcf5..3edfa7d3753c29e44b964c0ff424d2ea8d5fdee6
+static void add_rpc_devices(std::string servers) {
+    auto rpc_servers = string_split<std::string>(servers, ',');
+    if (rpc_servers.empty()) {
+        throw std::invalid_argument("no RPC servers specified");
+    }
+    ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
+    if (!rpc_reg) {
+        throw std::invalid_argument("failed to find RPC backend");
+    }
+    typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
+    ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
+    if (!ggml_backend_rpc_add_device_fn) {
+        throw std::invalid_argument("failed to find RPC device add function");
+    }
+    for (const auto & server : rpc_servers) {
+        ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
+        if (dev) {
+            ggml_backend_device_register(dev);
+        } else {
+            throw std::invalid_argument("failed to register RPC device");
+        }
+    }
+}
+
 // convert a vector of completion_token_output to json
 static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
 {
@@ -159,12 +186,45 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector<com
     return out;
 }
 
+struct llama_slot_params {
+    uint32_t seed      = -1; // RNG seed
+    bool stream        = true;
+    bool cache_prompt  = true; // remember the prompt to avoid reprocessing all prompt
+    bool return_tokens = false;
+
+    int32_t n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
+    int32_t n_predict = -1; // new tokens to predict
+    int32_t n_indent  =  0; // mininum line indentation for the generated text in number of whitespace characters
+
+    int64_t t_max_prompt_ms  = -1; // TODO: implement
+    int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
+
+    std::vector<common_adapter_lora_info> lora;
+
+    std::vector<std::string> antiprompt;
+    std::vector<std::string> response_fields;
+    bool timings_per_token = false;
+    bool post_sampling_probs = false;
+    bool ignore_eos = false;
+
+    json input_prefix;
+    json input_suffix;
+
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
+};
+
+
 struct llama_client_slot
 {
     int id;
     int task_id = -1;
 
-    struct slot_params params;
+    struct llama_slot_params params;
+    common_speculative * spec = nullptr;
+    llama_batch batch_spec = {};
+
 
     slot_state state = IDLE;
     slot_command command = NONE;
@@ -203,7 +263,7 @@ struct llama_client_slot
     std::string stopping_word;
 
     // sampling
-    struct common_sampler_params sparams;
+    struct common_params_sampling sparams;
     common_sampler *ctx_sampling = nullptr;
 
     int32_t ga_i = 0;   // group-attention state
@@ -257,6 +317,7 @@ struct llama_client_slot
         images.clear();
     }
 
+
     bool has_budget(common_params &global_params) {
         if (params.n_predict == -1 && global_params.n_predict == -1)
         {
@@ -428,6 +489,11 @@ struct llama_server_context
 {
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
+    common_init_result llama_init_dft;
+    llama_context * ctx_dft = nullptr;
+    llama_model * model_dft = nullptr;
+    llama_context_params cparams_dft;
+    const llama_vocab * vocab = nullptr;
 
     clip_ctx *clp_ctx = nullptr;
 
@@ -439,6 +505,7 @@ struct llama_server_context
     bool clean_kv_cache     = true;
     bool all_slots_are_idle = false;
     bool add_bos_token      = true;
+    bool has_eos_token      = true;
 
     int32_t n_ctx;  // total context for all clients / slots
 
@@ -474,6 +541,7 @@ struct llama_server_context
         }
     }
 
+
     bool load_model(const common_params &params_)
     {
         params = params_;
@@ -492,8 +560,8 @@ struct llama_server_context
         }
 
         common_init_result common_init = common_init_from_params(params);
-        model = common_init.model;
-        ctx = common_init.context;
+        model = common_init.model.release();
+        ctx = common_init.context.release();
         if (model == nullptr)
         {
             LOG_ERR("unable to load model: %s", params.model.c_str());
@@ -502,7 +570,7 @@ struct llama_server_context
 
         if (multimodal) {
             const int n_embd_clip = clip_n_mmproj_embd(clp_ctx);
-            const int n_embd_llm  = llama_n_embd(model);
+            const int n_embd_llm  = llama_model_n_embd(model);
             if (n_embd_clip != n_embd_llm) {
                 LOG("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm);
                 llama_free(ctx);
@@ -511,23 +579,54 @@ struct llama_server_context
             }
         }
 
+        vocab = llama_model_get_vocab(model);
         n_ctx = llama_n_ctx(ctx);
 
-        add_bos_token = llama_add_bos_token(model);
+        add_bos_token = llama_vocab_get_add_bos(vocab);
+        has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+
+        if (!params.speculative.model.empty()) {
+            LOG("loading draft model '%s'\n", params.speculative.model.c_str());
+
+            auto params_dft = params;
+
+            params_dft.devices      = params.speculative.devices;
+            params_dft.model        = params.speculative.model;
+            params_dft.n_ctx        = params.speculative.n_ctx == 0 ? params.n_ctx / params.n_parallel : params.speculative.n_ctx;
+            params_dft.n_gpu_layers = params.speculative.n_gpu_layers;
+            params_dft.n_parallel   = 1;
+
+            llama_init_dft = common_init_from_params(params_dft);
+
+            model_dft = llama_init_dft.model.get();
+
+            if (model_dft == nullptr) {
+                LOG("failed to load draft model, '%s'\n", params.speculative.model.c_str());
+                return false;
+            }
+
+            if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
+                LOG("the draft model '%s' is not compatible with the target model '%s'\n", params.speculative.model.c_str(), params.model.c_str());
+
+                return false;
+            }
+
+            const int n_ctx_dft = llama_n_ctx(llama_init_dft.context.get());
+
+            cparams_dft = common_context_params_to_llama(params_dft);
+            cparams_dft.n_batch = n_ctx_dft;
+
+            // force F16 KV cache for the draft model for extra performance
+            cparams_dft.type_k = GGML_TYPE_F16;
+            cparams_dft.type_v = GGML_TYPE_F16;
+
+            // the context is not needed - we will create one for each slot
+            llama_init_dft.context.reset();
+        }
 
         return true;
     }
 
-    void validate_model_chat_template(server_params & sparams) {
-        llama_chat_message chat[] = {{"user", "test"}};
-        std::vector<char> buf(1);
-        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
-        if (res < 0) {
-            LOG_ERR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", __func__);
-            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
-        }
-    }
-
     llama_client_slot* get_active_slot() {
         for (llama_client_slot& slot : slots) {
             // Check if the slot is currently processing
@@ -553,6 +652,22 @@ struct llama_server_context
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
 
+            if (model_dft) {
+                slot.batch_spec = llama_batch_init(params.speculative.n_max + 1, 0, 1);
+
+                ctx_dft = llama_init_from_model(model_dft, cparams_dft);
+                if (ctx_dft == nullptr) {
+                    LOG("%s", "failed to create draft context\n");
+                    return;
+                }
+
+                slot.spec = common_speculative_init(ctx_dft);
+                if (slot.spec == nullptr) {
+                    LOG("%s", "failed to create speculator\n");
+                    return;
+                }
+            }
+
             LOG_INFO("new slot", {
                 {"slot_id",    slot.id},
                 {"n_ctx_slot", slot.n_ctx}
@@ -661,16 +776,17 @@ struct llama_server_context
     }
 
     bool launch_slot_with_data(llama_client_slot* &slot, json data) {
-        slot_params default_params;
-        common_sampler_params default_sparams;
- 
+        llama_slot_params default_params;
+        common_params_sampling default_sparams;
+        
+        default_sparams.speculative = params_base.speculative;
+
         slot->params.stream             = json_value(data, "stream",            false);
         slot->params.cache_prompt       = json_value(data, "cache_prompt",      false);
         slot->params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
         slot->sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot->sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot->sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
-        slot->sparams.tfs_z             = json_value(data, "tfs_z",             default_sparams.tfs_z);
         slot->sparams.typ_p             = json_value(data, "typical_p",         default_sparams.typ_p);
         slot->sparams.temp              = json_value(data, "temperature",       default_sparams.temp);
         slot->sparams.dynatemp_range    = json_value(data, "dynatemp_range",    default_sparams.dynatemp_range);
@@ -682,13 +798,21 @@ struct llama_server_context
         slot->sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
         slot->sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
         slot->sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-        slot->sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
         slot->params.n_keep             = json_value(data, "n_keep",            slot->params.n_keep);
         slot->sparams.seed               = json_value(data, "seed",              default_sparams.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
 
+
+        slot->sparams.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min);
+        slot->sparams.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max);
+        slot->sparams.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
+
+        slot->sparams.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min);
+        slot->sparams.speculative.n_min = std::max(params.speculative.n_min, 2);
+        slot->sparams.speculative.n_max = std::max(params.speculative.n_max, 0);
+
         if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
             // Might be better to reject the request with a 400 ?
             LOG_WARNING("Max tokens to predict exceeds server configuration", {
@@ -727,8 +851,8 @@ struct llama_server_context
             slot->prompt = "";
         }
 
-        if (json_value(data, "ignore_eos", false)) {
-                slot->sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
+        if (json_value(data, "ignore_eos", false) && has_eos_token) {
+                slot->sparams.logit_bias.push_back({llama_vocab_eos(vocab), -INFINITY});
         }
         /*
         slot->sparams.penalty_prompt_tokens.clear();
@@ -767,13 +891,13 @@ struct llama_server_context
             }
         }
       */
-
         slot->sparams.logit_bias.clear();
 
         const auto &logit_bias = data.find("logit_bias");
         if (logit_bias != data.end() && logit_bias->is_array())
         {
-            const int n_vocab = llama_n_vocab(model);
+            const llama_vocab * vocab = llama_model_get_vocab(model);
+            const int n_vocab = llama_vocab_n_tokens(vocab);
             for (const auto &el : *logit_bias)
             {
                 if (el.is_array() && el.size() == 2)
@@ -802,7 +926,7 @@ struct llama_server_context
                     }
                     else if (el[0].is_string())
                     {
-                        auto toks = common_tokenize(model, el[0].get<std::string>(), false);
+                        auto toks = common_tokenize(vocab, el[0].get<std::string>(), false);
                         for (auto tok : toks)
                         {
                             slot->sparams.logit_bias.push_back({tok, bias});
@@ -1132,7 +1256,7 @@ struct llama_server_context
             slot.has_next_token = false;
         }
 
-        if (result.tok == llama_token_eos(model))
+        if (result.tok == llama_vocab_eos(vocab) || llama_vocab_is_eog(vocab, result.tok))
         {
             slot.stopped_eos = true;
             slot.has_next_token = false;
@@ -1206,7 +1330,6 @@ struct llama_server_context
             {"top_k",             slot.sparams.top_k},
             {"top_p",             slot.sparams.top_p},
             {"min_p",             slot.sparams.min_p},
-            {"tfs_z",             slot.sparams.tfs_z},
             {"typical_p",         slot.sparams.typ_p},
             {"repeat_last_n",     slot.sparams.penalty_last_n},
             {"repeat_penalty",    slot.sparams.penalty_repeat},
@@ -1215,13 +1338,12 @@ struct llama_server_context
             {"mirostat",          slot.sparams.mirostat},
             {"mirostat_tau",      slot.sparams.mirostat_tau},
             {"mirostat_eta",      slot.sparams.mirostat_eta},
-            {"penalize_nl",       slot.sparams.penalize_nl},
             {"stop",              slot.params.antiprompt},
             {"n_predict",         slot.params.n_predict},
             {"n_keep",            params.n_keep},
             {"ignore_eos",        slot.sparams.ignore_eos},
             {"stream",            slot.params.stream},
-      //      {"logit_bias",        slot.sparams.logit_bias},
+             //      {"logit_bias",        slot.sparams.logit_bias},
             {"n_probs",           slot.sparams.n_probs},
             {"min_keep",          slot.sparams.min_keep},
             {"grammar",           slot.sparams.grammar},
@@ -1329,7 +1451,7 @@ struct llama_server_context
         res.error = false;
         res.stop = true;
 
-        const int n_embd = llama_n_embd(model);
+        const int n_embd = llama_model_n_embd(model);
         if (!params.embedding)
         {
             LOG_WARNING("embedding disabled", {
@@ -1428,7 +1550,7 @@ struct llama_server_context
                     n_eval = n_batch;
                 }
 
-                const int n_embd = llama_n_embd(model);
+                const int n_embd = llama_model_n_embd(model);
                 float * embd = img.image_embedding + i * n_embd;
                 llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, slot.n_past, 0);
                 if (llama_decode(ctx, llava_batch.batch))
@@ -1709,11 +1831,11 @@ struct llama_server_context
                             suffix_tokens.erase(suffix_tokens.begin());
                         }
 
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
-                        prefix_tokens.insert(prefix_tokens.end(),   llama_token_suffix(model));
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_vocab_fim_pre(vocab));
+                        prefix_tokens.insert(prefix_tokens.begin(), llama_vocab_bos(vocab)); // always add BOS
+                        prefix_tokens.insert(prefix_tokens.end(),   llama_vocab_fim_suf(vocab));
                         prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
-                        prefix_tokens.push_back(llama_token_middle(model));
+                        prefix_tokens.push_back(llama_vocab_fim_mid(vocab));
                         prompt_tokens = prefix_tokens;
                     }
                     else
@@ -2008,6 +2130,97 @@ struct llama_server_context
             }
         }
 
+        // do speculative decoding
+        for (auto & slot : slots) {
+            if (!slot.is_processing() || !(ctx_dft && params.speculative.n_max > 0)) {
+                continue;
+            }
+
+            if (slot.state != PROCESSING) {
+                continue;
+            }
+
+            // determine the max draft that fits the current slot state
+            int n_draft_max = slot.params.speculative.n_max;
+
+            // note: n_past is not yet increased for the `id` token sampled above
+            //       also, need to leave space for 1 extra token to allow context shifts
+            n_draft_max = std::min(n_draft_max, slot.n_ctx - slot.n_past - 2);
+
+            if (slot.n_remaining > 0) {
+                n_draft_max = std::min(n_draft_max, slot.n_remaining - 1);
+            }
+
+            LOG("max possible draft: %d\n", n_draft_max);
+
+            if (n_draft_max < slot.params.speculative.n_min) {
+                LOG("the max possible draft is too small: %d < %d - skipping speculative decoding\n", n_draft_max, slot.params.speculative.n_min);
+
+                continue;
+            }
+
+            llama_token id = slot.sampled;
+
+            struct common_speculative_params params_spec;
+            params_spec.n_draft   = n_draft_max;
+            params_spec.n_reuse   = llama_n_ctx(ctx_dft) - slot.params.speculative.n_max;
+            params_spec.p_min     = slot.params.speculative.p_min;
+
+            llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+
+            // ignore small drafts
+            if (slot.params.speculative.n_min > (int) draft.size()) {
+                LOG("ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
+
+                continue;
+            }
+
+            // construct the speculation batch
+            common_batch_clear(slot.batch_spec);
+            common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
+
+            for (size_t i = 0; i < draft.size(); ++i) {
+                common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true);
+            }
+
+            LOG("decoding speculative batch, size = %d\n", slot.batch_spec.n_tokens);
+
+            llama_decode(ctx, slot.batch_spec);
+
+            // the accepted tokens from the speculation
+            const auto ids = common_sampler_sample_and_accept_n(slot.ctx_sampling, ctx, draft);
+
+            slot.n_past    += ids.size();
+            slot.n_decoded += ids.size();
+
+            slot.cache_tokens.push_back(id);
+            slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+
+            llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+
+            for (size_t i = 0; i < ids.size(); ++i) {
+                completion_token_output result;
+
+                result.tok          = ids[i];
+                result.text_to_send = common_token_to_piece(ctx, result.tok, params.special);
+                //result.prob         = 1.0f; // set later
+
+                // TODO: set result.probs
+
+                if (!process_token(result, slot)) {
+                    // release slot because of stop condition
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    metrics.on_prediction(slot);
+                    break;
+                }
+            }
+
+        LOG("accepted %d/%d draft tokens, new n_past = %d\n", (int) ids.size() - 1, (int) draft.size(), slot.n_past);
+        }
+    
+
         LOG_VERBOSE("slots updated", {});
         return true;
     }
@@ -2105,7 +2318,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     //     slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
     //     slot->sparams.top_k           = json_value(data, "top_k",             default_sparams.top_k);
     //     slot->sparams.top_p           = json_value(data, "top_p",             default_sparams.top_p);
-    //     slot->sparams.tfs_z           = json_value(data, "tfs_z",             default_sparams.tfs_z);
     //     slot->sparams.typical_p       = json_value(data, "typical_p",         default_sparams.typical_p);
     //     slot->sparams.temp            = json_value(data, "temperature",       default_sparams.temp);
     //     slot->sparams.penalty_last_n  = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
@@ -2115,7 +2327,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     //     slot->sparams.mirostat        = json_value(data, "mirostat",          default_sparams.mirostat);
     //     slot->sparams.mirostat_tau    = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
     //     slot->sparams.mirostat_eta    = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-    //     slot->sparams.penalize_nl     = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
     //     slot->params.n_keep           = json_value(data, "n_keep",            slot->params.n_keep);
     //     slot->params.seed             = json_value(data, "seed",              default_params.seed);
     //     slot->sparams.grammar         = json_value(data, "grammar",           default_sparams.grammar);
@@ -2129,7 +2340,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     data["n_predict"] = predict->tokens() == 0 ? -1 : predict->tokens();
     data["top_k"] = predict->topk();
     data["top_p"] = predict->topp();
-    data["tfs_z"] = predict->tailfreesamplingz();
     data["typical_p"] = predict->typicalp();
     data["temperature"] = predict->temperature();
     data["repeat_last_n"] = predict->repeat();
@@ -2139,7 +2349,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     data["mirostat"] = predict->mirostat();
     data["mirostat_tau"] = predict->mirostattau();
     data["mirostat_eta"] = predict->mirostateta();
-    data["penalize_nl"] = predict->penalizenl();
     data["n_keep"] = predict->nkeep();
     data["seed"] = predict->seed();
     data["grammar"] = predict->grammar();
@@ -2176,7 +2385,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     llama.params.n_predict = predict->tokens() == 0 ? -1 : predict->tokens();
 //     llama.params.sparams.top_k = predict->topk();
 //     llama.params.sparams.top_p = predict->topp();
-//     llama.params.sparams.tfs_z = predict->tailfreesamplingz();
 //     llama.params.sparams.typical_p = predict->typicalp();
 //     llama.params.sparams.penalty_last_n = predict->repeat();
 //     llama.params.sparams.temp = predict->temperature();
@@ -2186,7 +2394,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     llama.params.sparams.mirostat = predict->mirostat();
 //     llama.params.sparams.mirostat_tau = predict->mirostattau();
 //     llama.params.sparams.mirostat_eta = predict->mirostateta();
-//     llama.params.sparams.penalize_nl = predict->penalizenl();
 //     llama.params.n_keep = predict->nkeep();
 //     llama.params.seed = predict->seed();
 //     llama.params.sparams.grammar = predict->grammar();
@@ -2233,6 +2440,35 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     }
 // }
 
+const std::vector<ggml_type> kv_cache_types = {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_BF16,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_IQ4_NL,
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+};
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    for (const auto & type : kv_cache_types) {
+        if (ggml_type_name(type) == s) {
+            return type;
+        }
+    }
+    throw std::runtime_error("Unsupported cache type: " + s);
+}
+
+static std::string get_all_kv_cache_types() {
+    std::ostringstream msg;
+    for (const auto & type : kv_cache_types) {
+        msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 static void params_parse(const backend::ModelOptions* request,
                                 common_params & params) {
    
@@ -2246,11 +2482,41 @@ static void params_parse(const backend::ModelOptions* request,
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
+    if (!request->cachetypekey().empty()) {
+        params.cache_type_k = kv_cache_type_from_str(request->cachetypekey());
+    }
+    if (!request->cachetypevalue().empty()) {
+        params.cache_type_v = kv_cache_type_from_str(request->cachetypevalue());
+    }
     params.n_ctx = request->contextsize();
     //params.memory_f16 = request->f16memory();
     params.cpuparams.n_threads = request->threads();
     params.n_gpu_layers = request->ngpulayers();
     params.n_batch = request->nbatch();
+    params.speculative.model = request->draftmodel();
+
+    // If options is not NULL, parse options
+    for (int i = 0; request->options()[i] != NULL; i++) {
+        char *optname = strtok(request->options()[i], ":");
+        char *optval = strtok(NULL, ":");
+        if (optval == NULL) {
+            optval = "true";
+        }
+
+        if (!strcmp(optname, "speculative.n_gpu_layers")) {
+            params.speculative.n_gpu_layers = std::stoi(optval);
+        }
+        if (!strcmp(optname, "speculative.n_ctx")) {
+            params.speculative.n_ctx = std::stoi(optval);
+        }
+    }
+
+    if params.speculative.n_gpu_layers == 0 {
+        params.speculative.n_gpu_layers = params.n_gpu_layers;
+    }
+    if params.speculative.n_ctx == 0 {
+        params.speculative.n_ctx = params.n_ctx;
+    }
     // Set params.n_parallel by environment variable (LLAMA_PARALLEL), defaults to 1
     //params.n_parallel = 1;
     const char *env_parallel = std::getenv("LLAMACPP_PARALLEL");
@@ -2263,7 +2529,7 @@ static void params_parse(const backend::ModelOptions* request,
 
     const char *llama_grpc_servers = std::getenv("LLAMACPP_GRPC_SERVERS");
     if (llama_grpc_servers != NULL) {
-        params.rpc_servers = std::string(llama_grpc_servers);
+        add_rpc_devices(std::string(llama_grpc_servers));
     }
     
     // TODO: Add yarn
@@ -2304,6 +2570,7 @@ static void params_parse(const backend::ModelOptions* request,
     params.use_mmap = request->mmap();
     params.flash_attn = request->flashattention();
     params.no_kv_offload = request->nokvoffload();
+    params.ctx_shift = false; // We control context-shifting in any case (and we disable it as it could just lead to infinite loops)
 
     params.embedding = request->embeddings();
 
@@ -2388,6 +2655,13 @@ public:
                 int32_t tokens_evaluated = result.result_json.value("tokens_evaluated", 0);
                 reply.set_prompt_tokens(tokens_evaluated);
 
+                if (result.result_json.contains("timings")) {
+                    double timing_prompt_processing = result.result_json.at("timings").value("prompt_ms", 0.0);
+                    reply.set_timing_prompt_processing(timing_prompt_processing);
+                    double timing_token_generation = result.result_json.at("timings").value("predicted_ms", 0.0);
+                    reply.set_timing_token_generation(timing_token_generation);
+                }
+                
                 // Log Request Correlation Id
                 LOG_VERBOSE("correlation:", {
                     { "id", data["correlation_id"] }
@@ -2428,6 +2702,13 @@ public:
             reply->set_prompt_tokens(tokens_evaluated);
             reply->set_tokens(tokens_predicted);
             reply->set_message(completion_text);
+
+            if (result.result_json.contains("timings")) {
+                double timing_prompt_processing = result.result_json.at("timings").value("prompt_ms", 0.0);
+                reply->set_timing_prompt_processing(timing_prompt_processing);
+                double timing_token_generation = result.result_json.at("timings").value("predicted_ms", 0.0);
+                reply->set_timing_token_generation(timing_token_generation);
+            }
         }
         else
         {

backend/cpp/llama/patches/01-llava.patch

@@ -1,13 +1,13 @@
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 342042ff..224db9b5 100644
+index 3cd0d2fa..6c5e811a 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -2419,7 +2419,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
-             struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
-             int* patches_data = (int*)malloc(ggml_nbytes(patches));
-             for (int i = 0; i < num_patches; i++) {
--                patches_data[i] = i + 1;
-+                patches_data[i] = i;
-             }
-             ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-             free(patches_data);
+@@ -2608,7 +2608,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
+                 struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+                 int* patches_data = (int*)malloc(ggml_nbytes(patches));
+                 for (int i = 0; i < num_patches; i++) {
+-                    patches_data[i] = i + 1;
++                    patches_data[i] = i;
+                 }
+                 ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+                 free(patches_data);

backend/go/bark/Makefile

@@ -0,0 +1,25 @@
+INCLUDE_PATH := $(abspath ./)
+LIBRARY_PATH := $(abspath ./)
+
+AR?=ar
+
+BUILD_TYPE?=
+# keep standard at C11 and C++11
+CXXFLAGS = -I. -I$(INCLUDE_PATH)/../../../sources/bark.cpp/examples -I$(INCLUDE_PATH)/../../../sources/bark.cpp/spm-headers -I$(INCLUDE_PATH)/../../../sources/bark.cpp -O3 -DNDEBUG -std=c++17 -fPIC
+LDFLAGS  = -L$(LIBRARY_PATH) -L$(LIBRARY_PATH)/../../../sources/bark.cpp/build/examples -lbark -lstdc++ -lm
+
+# warnings
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+
+gobark.o:
+	$(CXX) $(CXXFLAGS) gobark.cpp -o gobark.o -c $(LDFLAGS)
+
+libbark.a: gobark.o
+	cp $(INCLUDE_PATH)/../../../sources/bark.cpp/build/libbark.a ./
+	$(AR) rcs libbark.a gobark.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml.c.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml-alloc.c.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml-backend.c.o
+
+clean:
+	rm -f gobark.o libbark.a

backend/go/bark/gobark.cpp

@@ -0,0 +1,85 @@
+#include <iostream>
+#include <tuple>
+
+#include "bark.h"
+#include "gobark.h"
+#include "common.h"
+#include "ggml.h"
+
+struct bark_context *c;
+
+void bark_print_progress_callback(struct bark_context *bctx, enum bark_encoding_step step, int progress, void *user_data) {
+    if (step == bark_encoding_step::SEMANTIC) {
+        printf("\rGenerating semantic tokens... %d%%", progress);
+    } else if (step == bark_encoding_step::COARSE) {
+        printf("\rGenerating coarse tokens... %d%%", progress);
+    } else if (step == bark_encoding_step::FINE) {
+        printf("\rGenerating fine tokens... %d%%", progress);
+    }
+    fflush(stdout);
+}
+
+int load_model(char *model) {
+    // initialize bark context
+    struct bark_context_params ctx_params = bark_context_default_params();
+    bark_params params;
+
+    params.model_path = model;
+
+   // ctx_params.verbosity = verbosity;
+    ctx_params.progress_callback = bark_print_progress_callback;
+    ctx_params.progress_callback_user_data = nullptr;
+
+    struct bark_context *bctx = bark_load_model(params.model_path.c_str(), ctx_params, params.seed);
+    if (!bctx) {
+        fprintf(stderr, "%s: Could not load model\n", __func__);
+        return 1;
+    }
+
+    c = bctx;
+
+    return 0;
+}
+
+int tts(char *text,int  threads, char *dst ) {
+
+    ggml_time_init();
+    const int64_t t_main_start_us = ggml_time_us();
+
+    // generate audio
+    if (!bark_generate_audio(c, text, threads)) {
+        fprintf(stderr, "%s: An error occured. If the problem persists, feel free to open an issue to report it.\n", __func__);
+        return 1;
+    }
+
+    const float *audio_data = bark_get_audio_data(c);
+    if (audio_data == NULL) {
+        fprintf(stderr, "%s: Could not get audio data\n", __func__);
+        return 1;
+    }
+
+    const int audio_arr_size = bark_get_audio_data_size(c);
+
+    std::vector<float> audio_arr(audio_data, audio_data + audio_arr_size);
+
+    write_wav_on_disk(audio_arr, dst);
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+        const int64_t t_load_us = bark_get_load_time(c);
+        const int64_t t_eval_us = bark_get_eval_time(c);
+
+        printf("\n\n");
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:     eval time = %8.2f ms\n", __func__, t_eval_us / 1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+    
+    return 0;
+}
+
+int unload() {
+    bark_free(c);
+}
+

backend/go/bark/gobark.go

@@ -0,0 +1,52 @@
+package main
+
+// #cgo CXXFLAGS: -I${SRCDIR}/../../../sources/bark.cpp/ -I${SRCDIR}/../../../sources/bark.cpp/encodec.cpp -I${SRCDIR}/../../../sources/bark.cpp/examples -I${SRCDIR}/../../../sources/bark.cpp/spm-headers
+// #cgo LDFLAGS: -L${SRCDIR}/ -L${SRCDIR}/../../../sources/bark.cpp/build/examples -L${SRCDIR}/../../../sources/bark.cpp/build/encodec.cpp/ -lbark -lencodec -lcommon
+// #include <gobark.h>
+// #include <stdlib.h>
+import "C"
+
+import (
+	"fmt"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+type Bark struct {
+	base.SingleThread
+	threads int
+}
+
+func (sd *Bark) Load(opts *pb.ModelOptions) error {
+
+	sd.threads = int(opts.Threads)
+
+	modelFile := C.CString(opts.ModelFile)
+	defer C.free(unsafe.Pointer(modelFile))
+
+	ret := C.load_model(modelFile)
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}
+
+func (sd *Bark) TTS(opts *pb.TTSRequest) error {
+	t := C.CString(opts.Text)
+	defer C.free(unsafe.Pointer(t))
+
+	dst := C.CString(opts.Dst)
+	defer C.free(unsafe.Pointer(dst))
+
+	threads := C.int(sd.threads)
+
+	ret := C.tts(t, threads, dst)
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}

backend/go/bark/gobark.h

@@ -0,0 +1,8 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+int load_model(char *model);
+int tts(char *text,int  threads, char *dst );
+#ifdef __cplusplus
+}
+#endif

backend/go/bark/main.go

@@ -1,7 +1,6 @@
 package main
 
 // Note: this is started internally by LocalAI and a server is allocated for each model
-
 import (
 	"flag"
 
@@ -15,7 +14,7 @@ var (
 func main() {
 	flag.Parse()
 
-	if err := grpc.StartServer(*addr, &Image{}); err != nil {
+	if err := grpc.StartServer(*addr, &Bark{}); err != nil {
 		panic(err)
 	}
 }

backend/go/image/stablediffusion-ggml/Makefile

@@ -0,0 +1,96 @@
+INCLUDE_PATH := $(abspath ./)
+LIBRARY_PATH := $(abspath ./)
+
+AR?=ar
+CMAKE_ARGS?=
+BUILD_TYPE?=
+ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
+# keep standard at C11 and C++11
+CXXFLAGS = -I. -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/thirdparty -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/ggml/include -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp -O3 -DNDEBUG -std=c++17 -fPIC
+
+# Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
+CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
+
+# If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DGGML_CUDA=ON
+# If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+# to CMAKE_ARGS automatically
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+# If build type is clblas (openCL) we set -DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
+else ifeq ($(BUILD_TYPE),clblas)
+	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
+# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
+else ifeq ($(BUILD_TYPE),hipblas)
+	CMAKE_ARGS+=-DGGML_HIP=ON
+# If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
+# But if it's OSX without metal, disable it here
+else ifeq ($(OS),Darwin)
+	ifneq ($(BUILD_TYPE),metal)
+		CMAKE_ARGS+=-DGGML_METAL=OFF
+	else
+		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		TARGET+=--target ggml-metal
+	endif
+endif
+
+# ifeq ($(BUILD_TYPE),sycl_f16)
+# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DSD_SYCL=ON -DGGML_SYCL_F16=ON
+# endif
+
+# ifeq ($(BUILD_TYPE),sycl_f32)
+# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON
+# endif
+
+# warnings
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+
+# Find all .a archives in ARCHIVE_DIR
+# (ggml can have different backends cpu, cuda, etc., each backend generates a .a archive)
+GGML_ARCHIVE_DIR := build/ggml/src/
+ALL_ARCHIVES := $(shell find $(GGML_ARCHIVE_DIR) -type f -name '*.a')
+
+# Name of the single merged library
+COMBINED_LIB := libggmlall.a
+
+# Rule to merge all the .a files into one
+$(COMBINED_LIB): $(ALL_ARCHIVES)
+	@echo "Merging all .a into $(COMBINED_LIB)"
+	rm -f $@
+	mkdir -p merge-tmp
+	for a in $(ALL_ARCHIVES); do \
+		( cd merge-tmp && ar x ../$$a ); \
+	done
+	( cd merge-tmp && ar rcs ../$@ *.o )
+	# Ensure we have a proper index
+	ranlib $@
+	# Clean up
+	rm -rf merge-tmp
+
+build/libstable-diffusion.a:
+	@echo "Building SD with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
+ifneq (,$(findstring sycl,$(BUILD_TYPE)))
+	+bash -c "source $(ONEAPI_VARS); \
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) ../../../../../sources/stablediffusion-ggml.cpp && \
+	cmake --build . --config Release"
+else
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) ../../../../../sources/stablediffusion-ggml.cpp && \
+	cmake --build . --config Release
+endif
+	$(MAKE) $(COMBINED_LIB)
+
+gosd.o:
+	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c
+
+libsd.a: gosd.o
+	cp $(INCLUDE_PATH)/build/libstable-diffusion.a ./libsd.a
+	$(AR) rcs libsd.a gosd.o
+
+clean:
+	rm -rf gosd.o libsd.a build $(COMBINED_LIB)

backend/go/image/stablediffusion-ggml/gosd.cpp

@@ -0,0 +1,228 @@
+#include <stdio.h>
+#include <string.h>
+#include <time.h>
+#include <iostream>
+#include <random>
+#include <string>
+#include <vector>
+#include "gosd.h"
+
+// #include "preprocessing.hpp"
+#include "flux.hpp"
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#define STB_IMAGE_STATIC
+#include "stb_image.h"
+
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#define STB_IMAGE_WRITE_STATIC
+#include "stb_image_write.h"
+
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#define STB_IMAGE_RESIZE_STATIC
+#include "stb_image_resize.h"
+
+// Names of the sampler method, same order as enum sample_method in stable-diffusion.h
+const char* sample_method_str[] = {
+    "euler_a",
+    "euler",
+    "heun",
+    "dpm2",
+    "dpm++2s_a",
+    "dpm++2m",
+    "dpm++2mv2",
+    "ipndm",
+    "ipndm_v",
+    "lcm",
+};
+
+// Names of the sigma schedule overrides, same order as sample_schedule in stable-diffusion.h
+const char* schedule_str[] = {
+    "default",
+    "discrete",
+    "karras",
+    "exponential",
+    "ays",
+    "gits",
+};
+
+sd_ctx_t* sd_c;
+
+sample_method_t sample_method;
+
+int load_model(char *model, char* options[], int threads, int diff) {
+    fprintf (stderr, "Loading model!\n");
+
+    char *stableDiffusionModel = "";
+    if (diff == 1 ) {
+        stableDiffusionModel = model;
+        model = "";
+    }
+
+    // decode options. Options are in form optname:optvale, or if booleans only optname.
+    char *clip_l_path  = "";
+    char *clip_g_path  = "";
+    char *t5xxl_path  = "";
+    char *vae_path  = "";
+    char *scheduler = "";
+    char *sampler = "";
+
+    // If options is not NULL, parse options
+    for (int i = 0; options[i] != NULL; i++) {
+        char *optname = strtok(options[i], ":");
+        char *optval = strtok(NULL, ":");
+        if (optval == NULL) {
+            optval = "true";
+        }
+
+        if (!strcmp(optname, "clip_l_path")) {
+            clip_l_path = optval;
+        }
+        if (!strcmp(optname, "clip_g_path")) {
+            clip_g_path = optval;
+        }
+        if (!strcmp(optname, "t5xxl_path")) {
+            t5xxl_path = optval;
+        }
+        if (!strcmp(optname, "vae_path")) {
+            vae_path = optval;
+        }
+        if (!strcmp(optname, "scheduler")) {
+            scheduler = optval;
+        }
+        if (!strcmp(optname, "sampler")) {
+            sampler = optval;
+        }
+    }
+
+    int sample_method_found = -1;
+    for (int m = 0; m < N_SAMPLE_METHODS; m++) {
+        if (!strcmp(sampler, sample_method_str[m])) {
+            sample_method_found = m;
+        }
+    }
+    if (sample_method_found == -1) {
+        fprintf(stderr, "Invalid sample method, default to EULER_A!\n");
+        sample_method_found = EULER_A;
+    }
+    sample_method = (sample_method_t)sample_method_found;
+
+    int schedule_found            = -1;
+    for (int d = 0; d < N_SCHEDULES; d++) {
+        if (!strcmp(scheduler, schedule_str[d])) {
+            schedule_found = d;
+                fprintf (stderr, "Found scheduler: %s\n", scheduler);
+
+        }
+    }
+
+    if (schedule_found == -1) {
+        fprintf (stderr, "Invalid scheduler! using DEFAULT\n");
+        schedule_found = DEFAULT;
+    }
+
+    schedule_t schedule = (schedule_t)schedule_found;
+    
+    fprintf (stderr, "Creating context\n");
+    sd_ctx_t* sd_ctx = new_sd_ctx(model,
+                                  clip_l_path,
+                                  clip_g_path,
+                                  t5xxl_path,
+                                  stableDiffusionModel,
+                                  vae_path,
+                                  "",
+                                  "",
+                                  "",
+                                  "",
+                                  "",
+                                  false,
+                                  false,
+                                  false,
+                                  threads,
+                                  SD_TYPE_COUNT,
+                                  STD_DEFAULT_RNG,
+                                  schedule,
+                                  false,
+                                  false,
+                                  false,
+                                  false);
+
+    if (sd_ctx == NULL) {
+        fprintf (stderr, "failed loading model (generic error)\n");
+        return 1;
+    }
+    fprintf (stderr, "Created context: OK\n");
+
+    sd_c = sd_ctx;
+
+    return 0;
+}
+
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed , char *dst, float cfg_scale) {
+
+    sd_image_t* results;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+
+    fprintf (stderr, "Generating image\n");
+
+    results = txt2img(sd_c,
+                            text,
+                            negativeText,
+                            -1, //clip_skip
+                            cfg_scale, // sfg_scale
+                            3.5f,
+                            width,
+                            height,
+                            sample_method, 
+                            steps,
+                            seed,
+                            1,
+                            NULL,
+                            0.9f,
+                            20.f,
+                            false,
+                            "",
+                            skip_layers.data(),
+                            skip_layers.size(),
+                            0,
+                            0.01,
+                            0.2);
+
+    if (results == NULL) {
+        fprintf (stderr, "NO results\n");
+        return 1;
+    }
+
+    if (results[0].data == NULL) {
+        fprintf (stderr, "Results with no data\n");
+        return 1;
+    }
+
+    fprintf (stderr, "Writing PNG\n");
+
+    fprintf (stderr, "DST: %s\n", dst);
+    fprintf (stderr, "Width: %d\n", results[0].width);
+    fprintf (stderr, "Height: %d\n", results[0].height);
+    fprintf (stderr, "Channel: %d\n", results[0].channel);
+    fprintf (stderr, "Data: %p\n", results[0].data);
+
+    stbi_write_png(dst, results[0].width, results[0].height, results[0].channel,
+                       results[0].data, 0, NULL);
+    fprintf (stderr, "Saved resulting image to '%s'\n", dst);
+
+    // TODO: free results. Why does it crash?
+
+    free(results[0].data);
+    results[0].data = NULL;
+    free(results);
+    fprintf (stderr, "gen_image is done", dst);
+
+    return 0;
+}
+
+int unload() {
+    free_sd_ctx(sd_c);
+}
+

backend/go/image/stablediffusion-ggml/gosd.go

@@ -0,0 +1,96 @@
+package main
+
+// #cgo CXXFLAGS: -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/thirdparty -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/ggml/include
+// #cgo LDFLAGS: -L${SRCDIR}/ -lsd -lstdc++ -lm -lggmlall -lgomp
+// #include <gosd.h>
+// #include <stdlib.h>
+import "C"
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/LocalAI/pkg/utils"
+)
+
+type SDGGML struct {
+	base.SingleThread
+	threads      int
+	sampleMethod string
+	cfgScale     float32
+}
+
+func (sd *SDGGML) Load(opts *pb.ModelOptions) error {
+
+	sd.threads = int(opts.Threads)
+
+	modelFile := C.CString(opts.ModelFile)
+	defer C.free(unsafe.Pointer(modelFile))
+
+	var options **C.char
+	// prepare the options array to pass to C
+
+	size := C.size_t(unsafe.Sizeof((*C.char)(nil)))
+	length := C.size_t(len(opts.Options))
+	options = (**C.char)(C.malloc(length * size))
+	view := (*[1 << 30]*C.char)(unsafe.Pointer(options))[0:len(opts.Options):len(opts.Options)]
+
+	var diffusionModel int
+
+	var oo []string
+	for _, op := range opts.Options {
+		if op == "diffusion_model" {
+			diffusionModel = 1
+			continue
+		}
+
+		// If it's an option path, we resolve absolute path from the model path
+		if strings.Contains(op, ":") && strings.Contains(op, "path") {
+			data := strings.Split(op, ":")
+			data[1] = filepath.Join(opts.ModelPath, data[1])
+			if err := utils.VerifyPath(data[1], opts.ModelPath); err == nil {
+				oo = append(oo, strings.Join(data, ":"))
+			}
+		} else {
+			oo = append(oo, op)
+		}
+	}
+
+	fmt.Fprintf(os.Stderr, "Options: %+v\n", oo)
+
+	for i, x := range oo {
+		view[i] = C.CString(x)
+	}
+
+	sd.cfgScale = opts.CFGScale
+
+	ret := C.load_model(modelFile, options, C.int(opts.Threads), C.int(diffusionModel))
+	if ret != 0 {
+		return fmt.Errorf("could not load model")
+	}
+
+	return nil
+}
+
+func (sd *SDGGML) GenerateImage(opts *pb.GenerateImageRequest) error {
+	t := C.CString(opts.PositivePrompt)
+	defer C.free(unsafe.Pointer(t))
+
+	dst := C.CString(opts.Dst)
+	defer C.free(unsafe.Pointer(dst))
+
+	negative := C.CString(opts.NegativePrompt)
+	defer C.free(unsafe.Pointer(negative))
+
+	ret := C.gen_image(t, negative, C.int(opts.Width), C.int(opts.Height), C.int(opts.Step), C.int(opts.Seed), dst, C.float(sd.cfgScale))
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}

backend/go/image/stablediffusion-ggml/gosd.h

@@ -0,0 +1,8 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+int load_model(char *model, char* options[], int threads, int diffusionModel);
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed, char *dst, float cfg_scale);
+#ifdef __cplusplus
+}
+#endif

backend/go/image/stablediffusion-ggml/main.go

@@ -1,7 +1,6 @@
 package main
 
 // Note: this is started internally by LocalAI and a server is allocated for each model
-
 import (
 	"flag"
 
@@ -15,7 +14,7 @@ var (
 func main() {
 	flag.Parse()
 
-	if err := grpc.StartServer(*addr, &Image{}); err != nil {
+	if err := grpc.StartServer(*addr, &SDGGML{}); err != nil {
 		panic(err)
 	}
 }

backend/go/image/stablediffusion/stablediffusion.go

@@ -1,33 +0,0 @@
-package main
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-	"github.com/mudler/LocalAI/pkg/stablediffusion"
-)
-
-type Image struct {
-	base.SingleThread
-	stablediffusion *stablediffusion.StableDiffusion
-}
-
-func (image *Image) Load(opts *pb.ModelOptions) error {
-	var err error
-	// Note: the Model here is a path to a directory containing the model files
-	image.stablediffusion, err = stablediffusion.New(opts.ModelFile)
-	return err
-}
-
-func (image *Image) GenerateImage(opts *pb.GenerateImageRequest) error {
-	return image.stablediffusion.GenerateImage(
-		int(opts.Height),
-		int(opts.Width),
-		int(opts.Mode),
-		int(opts.Step),
-		int(opts.Seed),
-		opts.PositivePrompt,
-		opts.NegativePrompt,
-		opts.Dst)
-}

backend/go/image/tinydream/tinydream.go

@@ -1,32 +0,0 @@
-package main
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-	"github.com/mudler/LocalAI/pkg/tinydream"
-)
-
-type Image struct {
-	base.SingleThread
-	tinydream *tinydream.TinyDream
-}
-
-func (image *Image) Load(opts *pb.ModelOptions) error {
-	var err error
-	// Note: the Model here is a path to a directory containing the model files
-	image.tinydream, err = tinydream.New(opts.ModelFile)
-	return err
-}
-
-func (image *Image) GenerateImage(opts *pb.GenerateImageRequest) error {
-	return image.tinydream.GenerateImage(
-		int(opts.Height),
-		int(opts.Width),
-		int(opts.Step),
-		int(opts.Seed),
-		opts.PositivePrompt,
-		opts.NegativePrompt,
-		opts.Dst)
-}

backend/go/llm/bert/bert.go

@@ -1,34 +0,0 @@
-package main
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	bert "github.com/go-skynet/go-bert.cpp"
-
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-type Embeddings struct {
-	base.SingleThread
-	bert *bert.Bert
-}
-
-func (llm *Embeddings) Load(opts *pb.ModelOptions) error {
-	model, err := bert.New(opts.ModelFile)
-	llm.bert = model
-	return err
-}
-
-func (llm *Embeddings) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
-
-	if len(opts.EmbeddingTokens) > 0 {
-		tokens := []int{}
-		for _, t := range opts.EmbeddingTokens {
-			tokens = append(tokens, int(t))
-		}
-		return llm.bert.TokenEmbeddings(tokens, bert.SetThreads(int(opts.Threads)))
-	}
-
-	return llm.bert.Embeddings(opts.Embeddings, bert.SetThreads(int(opts.Threads)))
-}

backend/go/llm/bert/main.go

@@ -1,21 +0,0 @@
-package main
-
-// Note: this is started internally by LocalAI and a server is allocated for each model
-
-import (
-	"flag"
-
-	grpc "github.com/mudler/LocalAI/pkg/grpc"
-)
-
-var (
-	addr = flag.String("addr", "localhost:50051", "the address to connect to")
-)
-
-func main() {
-	flag.Parse()
-
-	if err := grpc.StartServer(*addr, &Embeddings{}); err != nil {
-		panic(err)
-	}
-}

backend/go/llm/rwkv/rwkv.go

@@ -1,95 +0,0 @@
-package main
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	"fmt"
-	"path/filepath"
-
-	"github.com/donomii/go-rwkv.cpp"
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-const tokenizerSuffix = ".tokenizer.json"
-
-type LLM struct {
-	base.SingleThread
-
-	rwkv *rwkv.RwkvState
-}
-
-func (llm *LLM) Load(opts *pb.ModelOptions) error {
-	tokenizerFile := opts.Tokenizer
-	if tokenizerFile == "" {
-		modelFile := filepath.Base(opts.ModelFile)
-		tokenizerFile = modelFile + tokenizerSuffix
-	}
-	modelPath := filepath.Dir(opts.ModelFile)
-	tokenizerPath := filepath.Join(modelPath, tokenizerFile)
-
-	model := rwkv.LoadFiles(opts.ModelFile, tokenizerPath, uint32(opts.GetThreads()))
-
-	if model == nil {
-		return fmt.Errorf("rwkv could not load model")
-	}
-	llm.rwkv = model
-	return nil
-}
-
-func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
-	stopWord := "\n"
-	if len(opts.StopPrompts) > 0 {
-		stopWord = opts.StopPrompts[0]
-	}
-
-	if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
-		return "", err
-	}
-
-	response := llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), nil)
-
-	return response, nil
-}
-
-func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
-	go func() {
-
-		stopWord := "\n"
-		if len(opts.StopPrompts) > 0 {
-			stopWord = opts.StopPrompts[0]
-		}
-
-		if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
-			fmt.Println("Error processing input: ", err)
-			return
-		}
-
-		llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), func(s string) bool {
-			results <- s
-			return true
-		})
-		close(results)
-	}()
-
-	return nil
-}
-
-func (llm *LLM) TokenizeString(opts *pb.PredictOptions) (pb.TokenizationResponse, error) {
-	tokens, err := llm.rwkv.Tokenizer.Encode(opts.Prompt)
-	if err != nil {
-		return pb.TokenizationResponse{}, err
-	}
-
-	l := len(tokens)
-	i32Tokens := make([]int32, l)
-
-	for i, t := range tokens {
-		i32Tokens[i] = int32(t.ID)
-	}
-
-	return pb.TokenizationResponse{
-		Length: int32(l),
-		Tokens: i32Tokens,
-	}, nil
-}

backend/go/stores/store.go

@@ -311,12 +311,16 @@ func (s *Store) StoresGet(opts *pb.StoresGetOptions) (pb.StoresGetResult, error)
 }
 
 func isNormalized(k []float32) bool {
-	var sum float32
+	var sum float64
+
 	for _, v := range k {
-		sum += v
+		v64 := float64(v)
+		sum += v64*v64
 	}
 
-	return sum == 1.0
+	s := math.Sqrt(sum)
+
+	return s >= 0.99 && s <= 1.01
 }
 
 // TODO: This we could replace with handwritten SIMD code
@@ -328,7 +332,7 @@ func normalizedCosineSimilarity(k1, k2 []float32) float32 {
 		dot += k1[i] * k2[i]
 	}
 
-	assert(dot >= -1 && dot <= 1, fmt.Sprintf("dot = %f", dot))
+	assert(dot >= -1.01 && dot <= 1.01, fmt.Sprintf("dot = %f", dot))
 
 	// 2.0 * (1.0 - dot) would be the Euclidean distance
 	return dot
@@ -418,7 +422,7 @@ func cosineSimilarity(k1, k2 []float32, mag1 float64) float32 {
 
 	sim := float32(dot / (mag1 * math.Sqrt(mag2)))
 
-	assert(sim >= -1 && sim <= 1, fmt.Sprintf("sim = %f", sim))
+	assert(sim >= -1.01 && sim <= 1.01, fmt.Sprintf("sim = %f", sim))
 
 	return sim
 }

backend/go/vad/silero/main.go

@@ -15,7 +15,7 @@ var (
 func main() {
 	flag.Parse()
 
-	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+	if err := grpc.StartServer(*addr, &VAD{}); err != nil {
 		panic(err)
 	}
 }

backend/go/vad/silero/vad.go

@@ -0,0 +1,54 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/streamer45/silero-vad-go/speech"
+)
+
+type VAD struct {
+	base.SingleThread
+	detector *speech.Detector
+}
+
+func (vad *VAD) Load(opts *pb.ModelOptions) error {
+	v, err := speech.NewDetector(speech.DetectorConfig{
+		ModelPath:  opts.ModelFile,
+		SampleRate: 16000,
+		//WindowSize:           1024,
+		Threshold:            0.5,
+		MinSilenceDurationMs: 0,
+		SpeechPadMs:          0,
+	})
+	if err != nil {
+		return fmt.Errorf("create silero detector: %w", err)
+	}
+
+	vad.detector = v
+	return err
+}
+
+func (vad *VAD) VAD(req *pb.VADRequest) (pb.VADResponse, error) {
+	audio := req.Audio
+
+	segments, err := vad.detector.Detect(audio)
+	if err != nil {
+		return pb.VADResponse{}, fmt.Errorf("detect: %w", err)
+	}
+
+	vadSegments := []*pb.VADSegment{}
+	for _, s := range segments {
+		vadSegments = append(vadSegments, &pb.VADSegment{
+			Start: float32(s.SpeechStartAt),
+			End:   float32(s.SpeechEndAt),
+		})
+	}
+
+	return pb.VADResponse{
+		Segments: vadSegments,
+	}, nil
+}

backend/python/autogptq/requirements-intel.txt

@@ -1,5 +1,6 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools

backend/python/autogptq/requirements.txt

@@ -1,6 +1,6 @@
 accelerate
 auto-gptq==0.7.1
-grpcio==1.67.0
+grpcio==1.69.0
 protobuf
 certifi
 transformers

backend/python/bark/requirements-intel.txt

@@ -1,8 +1,9 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchaudio
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 transformers
 accelerate

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.67.0
+grpcio==1.69.0
 protobuf
 certifi

backend/python/common/libbackend.sh

@@ -17,6 +17,9 @@
 # LIMIT_TARGETS="cublas12"
 # source $(dirname $0)/../common/libbackend.sh
 #
+
+PYTHON_VERSION="3.10"
+
 function init() {
     # Name of the backend (directory name)
     BACKEND_NAME=${PWD##*/}
@@ -88,7 +91,7 @@ function getBuildProfile() {
 # always result in an activated virtual environment
 function ensureVenv() {
     if [ ! -d "${EDIR}/venv" ]; then
-        uv venv ${EDIR}/venv
+        uv venv --python ${PYTHON_VERSION} ${EDIR}/venv
         echo "virtualenv created"
     fi
 

backend/python/common/template/Makefile

@@ -1,8 +1,9 @@
 .DEFAULT_GOAL := install
 
 .PHONY: install
-install: protogen
+install:
 	bash install.sh
+	$(MAKE) protogen
 
 .PHONY: protogen
 protogen: backend_pb2_grpc.py backend_pb2.py
@@ -12,7 +13,7 @@ protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
 backend_pb2_grpc.py backend_pb2.py:
-	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+	bash protogen.sh
 
 .PHONY: clean
 clean: protogen-clean

backend/python/common/template/protogen.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto

backend/python/common/template/requirements-intel.txt

@@ -1,4 +1,5 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]

backend/python/common/template/requirements.txt

@@ -1,2 +1,3 @@
-grpcio==1.67.0
-protobuf
+grpcio==1.69.0
+protobuf
+grpcio-tools

backend/python/coqui/requirements-intel.txt

@@ -1,9 +1,10 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchaudio
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.69.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/backend.py

@@ -17,7 +17,7 @@ import backend_pb2_grpc
 
 import grpc
 
-from diffusers import StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
+from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
     EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
 from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
@@ -247,11 +247,16 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                         use_safetensors=True,
                         variant=variant)
             elif request.PipelineType == "FluxPipeline":
+                if fromSingleFile:
+                    self.pipe = FluxPipeline.from_single_file(modelFile,
+                                                              torch_dtype=torchType,
+                                                              use_safetensors=True)
+                else:
                     self.pipe = FluxPipeline.from_pretrained(
                         request.Model,
                         torch_dtype=torch.bfloat16)
-                    if request.LowVRAM:
-                        self.pipe.enable_model_cpu_offload()
+                if request.LowVRAM:
+                    self.pipe.enable_model_cpu_offload()
             elif request.PipelineType == "FluxTransformer2DModel":
                     dtype = torch.bfloat16
                     # specify from environment or default to "ChuckMcSneed/FLUX.1-dev"
@@ -270,6 +275,13 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
                     if request.LowVRAM:
                         self.pipe.enable_model_cpu_offload()
+            elif request.PipelineType == "SanaPipeline":
+                self.pipe = SanaPipeline.from_pretrained(
+                    request.Model,
+                    variant="bf16",
+                    torch_dtype=torch.bfloat16)
+                self.pipe.vae.to(torch.bfloat16)
+                self.pipe.text_encoder.to(torch.bfloat16)
 
             if CLIPSKIP and request.CLIPSkip != 0:
                 self.clip_skip = request.CLIPSkip
@@ -296,22 +308,34 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 self.pipe.controlnet = self.controlnet
             else:
                 self.controlnet = None
-            # Assume directory from request.ModelFile.
-            # Only if request.LoraAdapter it's not an absolute path
-            if request.LoraAdapter and request.ModelFile != "" and not os.path.isabs(request.LoraAdapter) and request.LoraAdapter:
-                # get base path of modelFile
-                modelFileBase = os.path.dirname(request.ModelFile)
+
+            if request.LoraAdapter and not os.path.isabs(request.LoraAdapter):
                 # modify LoraAdapter to be relative to modelFileBase
-                request.LoraAdapter = os.path.join(modelFileBase, request.LoraAdapter)
+                request.LoraAdapter = os.path.join(request.ModelPath, request.LoraAdapter)
+
             device = "cpu" if not request.CUDA else "cuda"
             self.device = device
             if request.LoraAdapter:
                 # Check if its a local file and not a directory ( we load lora differently for a safetensor file )
                 if os.path.exists(request.LoraAdapter) and not os.path.isdir(request.LoraAdapter):
-                    # self.load_lora_weights(request.LoraAdapter, 1, device, torchType)
                     self.pipe.load_lora_weights(request.LoraAdapter)
                 else:
                     self.pipe.unet.load_attn_procs(request.LoraAdapter)
+            if len(request.LoraAdapters) > 0:
+                i = 0
+                adapters_name = []
+                adapters_weights = []
+                for adapter in request.LoraAdapters:
+                    if not os.path.isabs(adapter):
+                        adapter = os.path.join(request.ModelPath, adapter)
+                    self.pipe.load_lora_weights(adapter, adapter_name=f"adapter_{i}")
+                    adapters_name.append(f"adapter_{i}")
+                    i += 1
+
+                for adapters_weight in request.LoraScales:
+                    adapters_weights.append(adapters_weight)
+
+                self.pipe.set_adapters(adapters_name, adapter_weights=adapters_weights)
 
             if request.CUDA:
                 self.pipe.to('cuda')
@@ -392,8 +416,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # create a dictionary of values for the parameters
         options = {
             "negative_prompt": request.negative_prompt,
-            "width": request.width,
-            "height": request.height,
             "num_inference_steps": steps,
         }
 
@@ -411,13 +433,13 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         keys = options.keys()
 
         if request.EnableParameters != "":
-            keys = request.EnableParameters.split(",")
+            keys = [key.strip() for key in request.EnableParameters.split(",")]
 
         if request.EnableParameters == "none":
             keys = []
 
         # create a dictionary of parameters by using the keys from EnableParameters and the values from defaults
-        kwargs = {key: options[key] for key in keys}
+        kwargs = {key: options.get(key) for key in keys if key in options}
 
         # Set seed
         if request.seed > 0:
@@ -428,6 +450,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         if self.PipelineType == "FluxPipeline":
             kwargs["max_sequence_length"] = 256
 
+        if request.width:
+            kwargs["width"] = request.width
+
+        if request.height:
+            kwargs["height"] = request.height
+
         if self.PipelineType == "FluxTransformer2DModel":
             kwargs["output_type"] = "pil"
             kwargs["generator"] = torch.Generator("cpu").manual_seed(0)
@@ -447,6 +475,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             export_to_video(video_frames, request.dst)
             return backend_pb2.Result(message="Media generated successfully", success=True)
 
+        print(f"Generating image with {kwargs=}", file=sys.stderr)
         image = {}
         if COMPEL:
             conditioning, pooled = self.compel.build_conditioning_tensor(prompt)

backend/python/diffusers/requirements-intel.txt

@@ -1,9 +1,10 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchvision
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchvision==0.18.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 diffusers
 opencv-python
 transformers

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.67.0
+grpcio==1.69.0
 pillow
 protobuf
 certifi

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.69.0
 protobuf
 certifi
 wheel

backend/python/faster-whisper/Makefile

@@ -1,8 +1,9 @@
 .DEFAULT_GOAL := install
 
 .PHONY: install
-install: protogen
+install:
 	bash install.sh
+	$(MAKE) protogen
 
 .PHONY: protogen
 protogen: backend_pb2_grpc.py backend_pb2.py
@@ -12,14 +13,8 @@ protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
 backend_pb2_grpc.py backend_pb2.py:
-	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+	bash protogen.sh
 
 .PHONY: clean
 clean: protogen-clean
-	rm -rf venv __pycache__
-
-.PHONY: test
-test: protogen
-	@echo "Testing openvoice..."
-	bash test.sh
-	@echo "openvoice tested."
+	rm -rf venv __pycache__

backend/python/faster-whisper/backend.py

@@ -1,85 +1,65 @@
 #!/usr/bin/env python3
 """
-Extra gRPC server for HuggingFace SentenceTransformer models.
+This is an extra gRPC server of LocalAI for Bark TTS
 """
 from concurrent import futures
-
+import time
 import argparse
 import signal
 import sys
 import os
-
-import time
 import backend_pb2
 import backend_pb2_grpc
 
+from faster_whisper import WhisperModel
+
 import grpc
 
-from sentence_transformers import SentenceTransformer
 
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 
 # If MAX_WORKERS are specified in the environment use it, otherwise default to 1
 MAX_WORKERS = int(os.environ.get('PYTHON_GRPC_MAX_WORKERS', '1'))
+COQUI_LANGUAGE = os.environ.get('COQUI_LANGUAGE', None)
 
 # Implement the BackendServicer class with the service methods
 class BackendServicer(backend_pb2_grpc.BackendServicer):
     """
-    A gRPC servicer for the backend service.
-
-    This class implements the gRPC methods for the backend service, including Health, LoadModel, and Embedding.
+    BackendServicer is the class that implements the gRPC service
     """
     def Health(self, request, context):
-        """
-        A gRPC method that returns the health status of the backend service.
-
-        Args:
-            request: A HealthRequest object that contains the request parameters.
-            context: A grpc.ServicerContext object that provides information about the RPC.
-
-        Returns:
-            A Reply object that contains the health status of the backend service.
-        """
         return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
-
     def LoadModel(self, request, context):
-        """
-        A gRPC method that loads a model into memory.
+        device = "cpu"
+        # Get device
+        # device = "cuda" if request.CUDA else "cpu"
+        if request.CUDA:
+            device = "cuda"
 
-        Args:
-            request: A LoadModelRequest object that contains the request parameters.
-            context: A grpc.ServicerContext object that provides information about the RPC.
-
-        Returns:
-            A Result object that contains the result of the LoadModel operation.
-        """
-        model_name = request.Model
         try:
-            self.model = SentenceTransformer(model_name, trust_remote_code=request.TrustRemoteCode)
+            print("Preparing models, please wait", file=sys.stderr)
+            self.model = WhisperModel(request.Model, device=device, compute_type="float16")
         except Exception as err:
             return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
-
         # Implement your logic here for the LoadModel service
         # Replace this with your desired response
         return backend_pb2.Result(message="Model loaded successfully", success=True)
 
-    def Embedding(self, request, context):
-        """
-        A gRPC method that calculates embeddings for a given sentence.
-
-        Args:
-            request: An EmbeddingRequest object that contains the request parameters.
-            context: A grpc.ServicerContext object that provides information about the RPC.
-
-        Returns:
-            An EmbeddingResult object that contains the calculated embeddings.
-        """
-        # Implement your logic here for the Embedding service
-        # Replace this with your desired response
-        print("Calculated embeddings for: " + request.Embeddings, file=sys.stderr)
-        sentence_embeddings = self.model.encode(request.Embeddings)
-        return backend_pb2.EmbeddingResult(embeddings=sentence_embeddings)
+    def AudioTranscription(self, request, context):
+        resultSegments = []
+        text = ""
+        try:
+            segments, info = self.model.transcribe(request.dst, beam_size=5, condition_on_previous_text=False)
+            id = 0
+            for segment in segments:
+                print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
+                resultSegments.append(backend_pb2.TranscriptSegment(id=id, start=segment.start, end=segment.end, text=segment.text))
+                text += segment.text
+                id += 1            
+        except Exception as err:
+            print(f"Unexpected {err=}, {type(err)=}", file=sys.stderr)
 
+        return backend_pb2.TranscriptResult(segments=resultSegments, text=text)
 
 def serve(address):
     server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))

backend/python/faster-whisper/protogen.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto

backend/python/faster-whisper/requirements-cpu.txt

@@ -0,0 +1,8 @@
+faster-whisper
+opencv-python
+accelerate
+compel
+peft
+sentencepiece
+torch==2.4.1
+optimum-quanto

backend/python/faster-whisper/requirements-cublas11.txt

@@ -1,5 +1,9 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.4.1+cu118
+faster-whisper
+opencv-python
 accelerate
-sentence-transformers==3.2.0
-transformers
+compel
+peft
+sentencepiece
+optimum-quanto

backend/python/faster-whisper/requirements-cublas12.txt

@@ -0,0 +1,8 @@
+torch==2.4.1
+faster-whisper
+opencv-python
+accelerate
+compel
+peft
+sentencepiece
+optimum-quanto

backend/python/faster-whisper/requirements-hipblas.txt

@@ -1,4 +1,3 @@
 --extra-index-url https://download.pytorch.org/whl/rocm6.0
-transformers
-accelerate
-torch==2.4.1+rocm6.0
+torch
+faster-whisper

backend/python/faster-whisper/requirements-intel.txt

@@ -0,0 +1,6 @@
+--extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
+optimum[openvino]
+faster-whisper

backend/python/faster-whisper/requirements.txt

@@ -0,0 +1,3 @@
+grpcio==1.69.0
+protobuf
+grpcio-tools

backend/python/kokoro/Makefile

@@ -0,0 +1,20 @@
+.DEFAULT_GOAL := install
+
+.PHONY: install
+install:
+	bash install.sh
+	$(MAKE) protogen
+
+.PHONY: protogen
+protogen: backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+backend_pb2_grpc.py backend_pb2.py:
+	bash protogen.sh
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__

backend/python/kokoro/backend.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-Extra gRPC server for MusicgenForConditionalGeneration models.
+Extra gRPC server for Kokoro models.
 """
 from concurrent import futures
 
@@ -8,20 +8,17 @@ import argparse
 import signal
 import sys
 import os
-
 import time
 import backend_pb2
 import backend_pb2_grpc
-
+import soundfile as sf
 import grpc
 
-from scipy.io.wavfile import write as write_wav
-
-from parler_tts import ParlerTTSForConditionalGeneration
-from transformers import AutoTokenizer
-import soundfile as sf  
+from models import build_model
+from kokoro import generate
 import torch
 
+SAMPLE_RATE = 22050
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 
 # If MAX_WORKERS are specified in the environment use it, otherwise default to 1
@@ -59,10 +56,31 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             A Result object that contains the result of the LoadModel operation.
         """
         model_name = request.Model
-        device = "cuda:0" if torch.cuda.is_available() else "cpu"
         try:
-            self.model = ParlerTTSForConditionalGeneration.from_pretrained(model_name).to(device)
-            self.tokenizer = AutoTokenizer.from_pretrained(model_name)
+            device = "cuda:0" if torch.cuda.is_available() else "cpu"
+            self.MODEL = build_model(request.ModelFile, device)
+            options = request.Options
+            # Find the voice from the options, options are a list of strings in this form optname:optvalue:
+            VOICE_NAME = None
+            for opt in options:
+                if opt.startswith("voice:"):
+                    VOICE_NAME = opt.split(":")[1]
+                    break
+            if VOICE_NAME is None:
+                return backend_pb2.Result(success=False, message=f"No voice specified in options")
+            MODELPATH = request.ModelPath
+            # If voice name contains a plus, split it and load the two models and combine them
+            if "+" in VOICE_NAME:
+                voice1, voice2 = VOICE_NAME.split("+")
+                voice1 = torch.load(f'{MODELPATH}/{voice1}.pt', weights_only=True).to(device)
+                voice2 = torch.load(f'{MODELPATH}/{voice2}.pt', weights_only=True).to(device)
+                self.VOICEPACK = torch.mean(torch.stack([voice1, voice2]), dim=0)
+            else:
+                self.VOICEPACK = torch.load(f'{MODELPATH}/{VOICE_NAME}.pt', weights_only=True).to(device)
+
+            self.VOICE_NAME = VOICE_NAME
+
+            print(f'Loaded voice: {VOICE_NAME}')
         except Exception as err:
             return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
 
@@ -70,38 +88,26 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
     def TTS(self, request, context):
         model_name = request.model
-        voice = request.voice
-        if voice == "":
-            voice = "A female speaker with a slightly low-pitched voice delivers her words quite expressively, in a very confined sounding environment with clear audio quality. She speaks very fast."
         if model_name == "":
             return backend_pb2.Result(success=False, message="request.model is required")
         try:
-            device = "cuda:0" if torch.cuda.is_available() else "cpu"
-            input_ids = self.tokenizer(voice, return_tensors="pt").input_ids.to(device)
-            prompt_input_ids = self.tokenizer(request.text, return_tensors="pt").input_ids.to(device)
-           
-            generation = self.model.generate(input_ids=input_ids, prompt_input_ids=prompt_input_ids)
-            audio_arr = generation.cpu().numpy().squeeze()
-            print("[parler-tts] TTS generated!", file=sys.stderr)
-            sf.write(request.dst, audio_arr, self.model.config.sampling_rate)
-            print("[parler-tts] TTS saved to", request.dst, file=sys.stderr)
-            print("[parler-tts] TTS for", file=sys.stderr)
-            print(request, file=sys.stderr)
+            audio, out_ps = generate(self.MODEL, request.text, self.VOICEPACK, lang=self.VOICE_NAME)
+            print(out_ps)
+            sf.write(request.dst, audio, SAMPLE_RATE)
         except Exception as err:
             return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
         return backend_pb2.Result(success=True)
 
-
 def serve(address):
     server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()
-    print("[parler-tts] Server started. Listening on: " + address, file=sys.stderr)
+    print("[Kokoro] Server started. Listening on: " + address, file=sys.stderr)
 
     # Define the signal handler function
     def signal_handler(sig, frame):
-        print("[parler-tts] Received termination signal. Shutting down...")
+        print("[Kokoro] Received termination signal. Shutting down...")
         server.stop(0)
         sys.exit(0)
 
@@ -121,5 +127,5 @@ if __name__ == "__main__":
         "--addr", default="localhost:50051", help="The address to bind the server to."
     )
     args = parser.parse_args()
-    print(f"[parler-tts] startup: {args}", file=sys.stderr)
+    print(f"[Kokoro] startup: {args}", file=sys.stderr)
     serve(args.addr)

backend/python/kokoro/istftnet.py

@@ -0,0 +1,524 @@
+# https://huggingface.co/hexgrad/Kokoro-82M/blob/main/istftnet.py
+# https://github.com/yl4579/StyleTTS2/blob/main/Modules/istftnet.py
+from scipy.signal import get_window
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn.utils import weight_norm, remove_weight_norm
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+# https://github.com/yl4579/StyleTTS2/blob/main/Modules/utils.py
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size*dilation - dilation)/2)
+
+LRELU_SLOPE = 0.1
+
+class AdaIN1d(nn.Module):
+    def __init__(self, style_dim, num_features):
+        super().__init__()
+        self.norm = nn.InstanceNorm1d(num_features, affine=False)
+        self.fc = nn.Linear(style_dim, num_features*2)
+
+    def forward(self, x, s):
+        h = self.fc(s)
+        h = h.view(h.size(0), h.size(1), 1)
+        gamma, beta = torch.chunk(h, chunks=2, dim=1)
+        return (1 + gamma) * self.norm(x) + beta
+
+class AdaINResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5), style_dim=64):
+        super(AdaINResBlock1, self).__init__()
+        self.convs1 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
+                               padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.ModuleList([
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+        
+        self.adain1 = nn.ModuleList([
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+        ])
+        
+        self.adain2 = nn.ModuleList([
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+            AdaIN1d(style_dim, channels),
+        ])
+        
+        self.alpha1 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs1))])
+        self.alpha2 = nn.ParameterList([nn.Parameter(torch.ones(1, channels, 1)) for i in range(len(self.convs2))])
+
+
+    def forward(self, x, s):
+        for c1, c2, n1, n2, a1, a2 in zip(self.convs1, self.convs2, self.adain1, self.adain2, self.alpha1, self.alpha2):
+            xt = n1(x, s)
+            xt = xt + (1 / a1) * (torch.sin(a1 * xt) ** 2)  # Snake1D
+            xt = c1(xt)
+            xt = n2(xt, s)
+            xt = xt + (1 / a2) * (torch.sin(a2 * xt) ** 2)  # Snake1D
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+            
+class TorchSTFT(torch.nn.Module):
+    def __init__(self, filter_length=800, hop_length=200, win_length=800, window='hann'):
+        super().__init__()
+        self.filter_length = filter_length
+        self.hop_length = hop_length
+        self.win_length = win_length
+        self.window = torch.from_numpy(get_window(window, win_length, fftbins=True).astype(np.float32))
+
+    def transform(self, input_data):
+        forward_transform = torch.stft(
+            input_data,
+            self.filter_length, self.hop_length, self.win_length, window=self.window.to(input_data.device),
+            return_complex=True)
+
+        return torch.abs(forward_transform), torch.angle(forward_transform)
+
+    def inverse(self, magnitude, phase):
+        inverse_transform = torch.istft(
+            magnitude * torch.exp(phase * 1j),
+            self.filter_length, self.hop_length, self.win_length, window=self.window.to(magnitude.device))
+
+        return inverse_transform.unsqueeze(-2)  # unsqueeze to stay consistent with conv_transpose1d implementation
+
+    def forward(self, input_data):
+        self.magnitude, self.phase = self.transform(input_data)
+        reconstruction = self.inverse(self.magnitude, self.phase)
+        return reconstruction
+    
+class SineGen(torch.nn.Module):
+    """ Definition of sine generator
+    SineGen(samp_rate, harmonic_num = 0,
+            sine_amp = 0.1, noise_std = 0.003,
+            voiced_threshold = 0,
+            flag_for_pulse=False)
+    samp_rate: sampling rate in Hz
+    harmonic_num: number of harmonic overtones (default 0)
+    sine_amp: amplitude of sine-wavefrom (default 0.1)
+    noise_std: std of Gaussian noise (default 0.003)
+    voiced_thoreshold: F0 threshold for U/V classification (default 0)
+    flag_for_pulse: this SinGen is used inside PulseGen (default False)
+    Note: when flag_for_pulse is True, the first time step of a voiced
+        segment is always sin(np.pi) or cos(0)
+    """
+
+    def __init__(self, samp_rate, upsample_scale, harmonic_num=0,
+                 sine_amp=0.1, noise_std=0.003,
+                 voiced_threshold=0,
+                 flag_for_pulse=False):
+        super(SineGen, self).__init__()
+        self.sine_amp = sine_amp
+        self.noise_std = noise_std
+        self.harmonic_num = harmonic_num
+        self.dim = self.harmonic_num + 1
+        self.sampling_rate = samp_rate
+        self.voiced_threshold = voiced_threshold
+        self.flag_for_pulse = flag_for_pulse
+        self.upsample_scale = upsample_scale
+
+    def _f02uv(self, f0):
+        # generate uv signal
+        uv = (f0 > self.voiced_threshold).type(torch.float32)
+        return uv
+
+    def _f02sine(self, f0_values):
+        """ f0_values: (batchsize, length, dim)
+            where dim indicates fundamental tone and overtones
+        """
+        # convert to F0 in rad. The interger part n can be ignored
+        # because 2 * np.pi * n doesn't affect phase
+        rad_values = (f0_values / self.sampling_rate) % 1
+
+        # initial phase noise (no noise for fundamental component)
+        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \
+                              device=f0_values.device)
+        rand_ini[:, 0] = 0
+        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+
+        # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
+        if not self.flag_for_pulse:
+#             # for normal case
+
+#             # To prevent torch.cumsum numerical overflow,
+#             # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1.
+#             # Buffer tmp_over_one_idx indicates the time step to add -1.
+#             # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi
+#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
+#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
+#             cumsum_shift = torch.zeros_like(rad_values)
+#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+
+#             phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
+            rad_values = torch.nn.functional.interpolate(rad_values.transpose(1, 2), 
+                                                         scale_factor=1/self.upsample_scale, 
+                                                         mode="linear").transpose(1, 2)
+    
+#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
+#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
+#             cumsum_shift = torch.zeros_like(rad_values)
+#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+    
+            phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
+            phase = torch.nn.functional.interpolate(phase.transpose(1, 2) * self.upsample_scale, 
+                                                    scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
+            sines = torch.sin(phase)
+            
+        else:
+            # If necessary, make sure that the first time step of every
+            # voiced segments is sin(pi) or cos(0)
+            # This is used for pulse-train generation
+
+            # identify the last time step in unvoiced segments
+            uv = self._f02uv(f0_values)
+            uv_1 = torch.roll(uv, shifts=-1, dims=1)
+            uv_1[:, -1, :] = 1
+            u_loc = (uv < 1) * (uv_1 > 0)
+
+            # get the instantanouse phase
+            tmp_cumsum = torch.cumsum(rad_values, dim=1)
+            # different batch needs to be processed differently
+            for idx in range(f0_values.shape[0]):
+                temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
+                temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
+                # stores the accumulation of i.phase within
+                # each voiced segments
+                tmp_cumsum[idx, :, :] = 0
+                tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
+
+            # rad_values - tmp_cumsum: remove the accumulation of i.phase
+            # within the previous voiced segment.
+            i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
+
+            # get the sines
+            sines = torch.cos(i_phase * 2 * np.pi)
+        return sines
+
+    def forward(self, f0):
+        """ sine_tensor, uv = forward(f0)
+        input F0: tensor(batchsize=1, length, dim=1)
+                  f0 for unvoiced steps should be 0
+        output sine_tensor: tensor(batchsize=1, length, dim)
+        output uv: tensor(batchsize=1, length, 1)
+        """
+        f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim,
+                             device=f0.device)
+        # fundamental component
+        fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
+
+        # generate sine waveforms
+        sine_waves = self._f02sine(fn) * self.sine_amp
+
+        # generate uv signal
+        # uv = torch.ones(f0.shape)
+        # uv = uv * (f0 > self.voiced_threshold)
+        uv = self._f02uv(f0)
+
+        # noise: for unvoiced should be similar to sine_amp
+        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
+        # .       for voiced regions is self.noise_std
+        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+        noise = noise_amp * torch.randn_like(sine_waves)
+
+        # first: set the unvoiced part to 0 by uv
+        # then: additive noise
+        sine_waves = sine_waves * uv + noise
+        return sine_waves, uv, noise
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+    """ SourceModule for hn-nsf
+    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0)
+    sampling_rate: sampling_rate in Hz
+    harmonic_num: number of harmonic above F0 (default: 0)
+    sine_amp: amplitude of sine source signal (default: 0.1)
+    add_noise_std: std of additive Gaussian noise (default: 0.003)
+        note that amplitude of noise in unvoiced is decided
+        by sine_amp
+    voiced_threshold: threhold to set U/V given F0 (default: 0)
+    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+    F0_sampled (batchsize, length, 1)
+    Sine_source (batchsize, length, 1)
+    noise_source (batchsize, length 1)
+    uv (batchsize, length, 1)
+    """
+
+    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
+                 add_noise_std=0.003, voiced_threshod=0):
+        super(SourceModuleHnNSF, self).__init__()
+
+        self.sine_amp = sine_amp
+        self.noise_std = add_noise_std
+
+        # to produce sine waveforms
+        self.l_sin_gen = SineGen(sampling_rate, upsample_scale, harmonic_num,
+                                 sine_amp, add_noise_std, voiced_threshod)
+
+        # to merge source harmonics into a single excitation
+        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+        self.l_tanh = torch.nn.Tanh()
+
+    def forward(self, x):
+        """
+        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+        F0_sampled (batchsize, length, 1)
+        Sine_source (batchsize, length, 1)
+        noise_source (batchsize, length 1)
+        """
+        # source for harmonic branch
+        with torch.no_grad():
+            sine_wavs, uv, _ = self.l_sin_gen(x)
+        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
+
+        # source for noise branch, in the same shape as uv
+        noise = torch.randn_like(uv) * self.sine_amp / 3
+        return sine_merge, noise, uv
+def padDiff(x):
+    return F.pad(F.pad(x, (0,0,-1,1), 'constant', 0) - x, (0,0,0,-1), 'constant', 0)
+
+    
+class Generator(torch.nn.Module):
+    def __init__(self, style_dim, resblock_kernel_sizes, upsample_rates, upsample_initial_channel, resblock_dilation_sizes, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size):
+        super(Generator, self).__init__()
+
+        self.num_kernels = len(resblock_kernel_sizes)
+        self.num_upsamples = len(upsample_rates)
+        resblock = AdaINResBlock1
+
+        self.m_source = SourceModuleHnNSF(
+                    sampling_rate=24000,
+                    upsample_scale=np.prod(upsample_rates) * gen_istft_hop_size,
+                    harmonic_num=8, voiced_threshod=10)
+        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * gen_istft_hop_size)
+        self.noise_convs = nn.ModuleList()
+        self.noise_res = nn.ModuleList()
+        
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            self.ups.append(weight_norm(
+                ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
+                                k, u, padding=(k-u)//2)))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel//(2**(i+1))
+            for j, (k, d) in enumerate(zip(resblock_kernel_sizes,resblock_dilation_sizes)):
+                self.resblocks.append(resblock(ch, k, d, style_dim))
+                
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            
+            if i + 1 < len(upsample_rates):  #
+                stride_f0 = np.prod(upsample_rates[i + 1:])
+                self.noise_convs.append(Conv1d(
+                    gen_istft_n_fft + 2, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
+                self.noise_res.append(resblock(c_cur, 7, [1,3,5], style_dim))
+            else:
+                self.noise_convs.append(Conv1d(gen_istft_n_fft + 2, c_cur, kernel_size=1))
+                self.noise_res.append(resblock(c_cur, 11, [1,3,5], style_dim))
+                
+                
+        self.post_n_fft = gen_istft_n_fft
+        self.conv_post = weight_norm(Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
+        self.ups.apply(init_weights)
+        self.conv_post.apply(init_weights)
+        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
+        self.stft = TorchSTFT(filter_length=gen_istft_n_fft, hop_length=gen_istft_hop_size, win_length=gen_istft_n_fft)
+        
+        
+    def forward(self, x, s, f0):
+        with torch.no_grad():
+            f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
+
+            har_source, noi_source, uv = self.m_source(f0)
+            har_source = har_source.transpose(1, 2).squeeze(1)
+            har_spec, har_phase = self.stft.transform(har_source)
+            har = torch.cat([har_spec, har_phase], dim=1)
+        
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x_source = self.noise_convs[i](har)
+            x_source = self.noise_res[i](x_source, s)
+
+            x = self.ups[i](x)
+            if i == self.num_upsamples - 1:
+                x = self.reflection_pad(x)
+
+            x = x + x_source
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i*self.num_kernels+j](x, s)
+                else:
+                    xs += self.resblocks[i*self.num_kernels+j](x, s)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.conv_post(x)
+        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
+        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
+        return self.stft.inverse(spec, phase)
+    
+    def fw_phase(self, x, s):
+        for i in range(self.num_upsamples):
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            x = self.ups[i](x)
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i*self.num_kernels+j](x, s)
+                else:
+                    xs += self.resblocks[i*self.num_kernels+j](x, s)
+            x = xs / self.num_kernels
+        x = F.leaky_relu(x)
+        x = self.reflection_pad(x)
+        x = self.conv_post(x)
+        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
+        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
+        return spec, phase
+
+    def remove_weight_norm(self):
+        print('Removing weight norm...')
+        for l in self.ups:
+            remove_weight_norm(l)
+        for l in self.resblocks:
+            l.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+        
+class AdainResBlk1d(nn.Module):
+    def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2),
+                 upsample='none', dropout_p=0.0):
+        super().__init__()
+        self.actv = actv
+        self.upsample_type = upsample
+        self.upsample = UpSample1d(upsample)
+        self.learned_sc = dim_in != dim_out
+        self._build_weights(dim_in, dim_out, style_dim)
+        self.dropout = nn.Dropout(dropout_p)
+        
+        if upsample == 'none':
+            self.pool = nn.Identity()
+        else:
+            self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
+        
+        
+    def _build_weights(self, dim_in, dim_out, style_dim):
+        self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
+        self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
+        self.norm1 = AdaIN1d(style_dim, dim_in)
+        self.norm2 = AdaIN1d(style_dim, dim_out)
+        if self.learned_sc:
+            self.conv1x1 = weight_norm(nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
+
+    def _shortcut(self, x):
+        x = self.upsample(x)
+        if self.learned_sc:
+            x = self.conv1x1(x)
+        return x
+
+    def _residual(self, x, s):
+        x = self.norm1(x, s)
+        x = self.actv(x)
+        x = self.pool(x)
+        x = self.conv1(self.dropout(x))
+        x = self.norm2(x, s)
+        x = self.actv(x)
+        x = self.conv2(self.dropout(x))
+        return x
+
+    def forward(self, x, s):
+        out = self._residual(x, s)
+        out = (out + self._shortcut(x)) / np.sqrt(2)
+        return out
+    
+class UpSample1d(nn.Module):
+    def __init__(self, layer_type):
+        super().__init__()
+        self.layer_type = layer_type
+
+    def forward(self, x):
+        if self.layer_type == 'none':
+            return x
+        else:
+            return F.interpolate(x, scale_factor=2, mode='nearest')
+
+class Decoder(nn.Module):
+    def __init__(self, dim_in=512, F0_channel=512, style_dim=64, dim_out=80, 
+                resblock_kernel_sizes = [3,7,11],
+                upsample_rates = [10, 6],
+                upsample_initial_channel=512,
+                resblock_dilation_sizes=[[1,3,5], [1,3,5], [1,3,5]],
+                upsample_kernel_sizes=[20, 12], 
+                gen_istft_n_fft=20, gen_istft_hop_size=5):
+        super().__init__()
+        
+        self.decode = nn.ModuleList()
+        
+        self.encode = AdainResBlk1d(dim_in + 2, 1024, style_dim)
+        
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
+        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 512, style_dim, upsample=True))
+
+        self.F0_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+        
+        self.N_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
+        
+        self.asr_res = nn.Sequential(
+            weight_norm(nn.Conv1d(512, 64, kernel_size=1)),
+        )
+        
+        
+        self.generator = Generator(style_dim, resblock_kernel_sizes, upsample_rates, 
+                                   upsample_initial_channel, resblock_dilation_sizes, 
+                                   upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size)
+        
+    def forward(self, asr, F0_curve, N, s):
+        F0 = self.F0_conv(F0_curve.unsqueeze(1))
+        N = self.N_conv(N.unsqueeze(1))
+        
+        x = torch.cat([asr, F0, N], axis=1)
+        x = self.encode(x, s)
+        
+        asr_res = self.asr_res(asr)
+        
+        res = True
+        for block in self.decode:
+            if res:
+                x = torch.cat([x, asr_res, F0, N], axis=1)
+            x = block(x, s)
+            if block.upsample_type != "none":
+                res = False
+                
+        x = self.generator(x, s, F0_curve)
+        return x

backend/python/kokoro/kokoro.py

@@ -0,0 +1,166 @@
+# https://huggingface.co/hexgrad/Kokoro-82M/blob/main/kokoro.py
+import phonemizer
+import re
+import torch
+import numpy as np
+
+def split_num(num):
+    num = num.group()
+    if '.' in num:
+        return num
+    elif ':' in num:
+        h, m = [int(n) for n in num.split(':')]
+        if m == 0:
+            return f"{h} o'clock"
+        elif m < 10:
+            return f'{h} oh {m}'
+        return f'{h} {m}'
+    year = int(num[:4])
+    if year < 1100 or year % 1000 < 10:
+        return num
+    left, right = num[:2], int(num[2:4])
+    s = 's' if num.endswith('s') else ''
+    if 100 <= year % 1000 <= 999:
+        if right == 0:
+            return f'{left} hundred{s}'
+        elif right < 10:
+            return f'{left} oh {right}{s}'
+    return f'{left} {right}{s}'
+
+def flip_money(m):
+    m = m.group()
+    bill = 'dollar' if m[0] == '$' else 'pound'
+    if m[-1].isalpha():
+        return f'{m[1:]} {bill}s'
+    elif '.' not in m:
+        s = '' if m[1:] == '1' else 's'
+        return f'{m[1:]} {bill}{s}'
+    b, c = m[1:].split('.')
+    s = '' if b == '1' else 's'
+    c = int(c.ljust(2, '0'))
+    coins = f"cent{'' if c == 1 else 's'}" if m[0] == '$' else ('penny' if c == 1 else 'pence')
+    return f'{b} {bill}{s} and {c} {coins}'
+
+def point_num(num):
+    a, b = num.group().split('.')
+    return ' point '.join([a, ' '.join(b)])
+
+def normalize_text(text):
+    text = text.replace(chr(8216), "'").replace(chr(8217), "'")
+    text = text.replace('«', chr(8220)).replace('»', chr(8221))
+    text = text.replace(chr(8220), '"').replace(chr(8221), '"')
+    text = text.replace('(', '«').replace(')', '»')
+    for a, b in zip('、。！，：；？', ',.!,:;?'):
+        text = text.replace(a, b+' ')
+    text = re.sub(r'[^\S \n]', ' ', text)
+    text = re.sub(r'  +', ' ', text)
+    text = re.sub(r'(?<=\n) +(?=\n)', '', text)
+    text = re.sub(r'\bD[Rr]\.(?= [A-Z])', 'Doctor', text)
+    text = re.sub(r'\b(?:Mr\.|MR\.(?= [A-Z]))', 'Mister', text)
+    text = re.sub(r'\b(?:Ms\.|MS\.(?= [A-Z]))', 'Miss', text)
+    text = re.sub(r'\b(?:Mrs\.|MRS\.(?= [A-Z]))', 'Mrs', text)
+    text = re.sub(r'\betc\.(?! [A-Z])', 'etc', text)
+    text = re.sub(r'(?i)\b(y)eah?\b', r"\1e'a", text)
+    text = re.sub(r'\d*\.\d+|\b\d{4}s?\b|(?<!:)\b(?:[1-9]|1[0-2]):[0-5]\d\b(?!:)', split_num, text)
+    text = re.sub(r'(?<=\d),(?=\d)', '', text)
+    text = re.sub(r'(?i)[$£]\d+(?:\.\d+)?(?: hundred| thousand| (?:[bm]|tr)illion)*\b|[$£]\d+\.\d\d?\b', flip_money, text)
+    text = re.sub(r'\d*\.\d+', point_num, text)
+    text = re.sub(r'(?<=\d)-(?=\d)', ' to ', text)
+    text = re.sub(r'(?<=\d)S', ' S', text)
+    text = re.sub(r"(?<=[BCDFGHJ-NP-TV-Z])'?s\b", "'S", text)
+    text = re.sub(r"(?<=X')S\b", 's', text)
+    text = re.sub(r'(?:[A-Za-z]\.){2,} [a-z]', lambda m: m.group().replace('.', '-'), text)
+    text = re.sub(r'(?i)(?<=[A-Z])\.(?=[A-Z])', '-', text)
+    return text.strip()
+
+def get_vocab():
+    _pad = "$"
+    _punctuation = ';:,.!?¡¿—…"«»“” '
+    _letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
+    _letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
+    symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa)
+    dicts = {}
+    for i in range(len((symbols))):
+        dicts[symbols[i]] = i
+    return dicts
+
+VOCAB = get_vocab()
+def tokenize(ps):
+    return [i for i in map(VOCAB.get, ps) if i is not None]
+
+phonemizers = dict(
+    a=phonemizer.backend.EspeakBackend(language='en-us', preserve_punctuation=True, with_stress=True),
+    b=phonemizer.backend.EspeakBackend(language='en-gb', preserve_punctuation=True, with_stress=True),
+)
+def phonemize(text, lang, norm=True):
+    if norm:
+        text = normalize_text(text)
+    ps = phonemizers[lang].phonemize([text])
+    ps = ps[0] if ps else ''
+    # https://en.wiktionary.org/wiki/kokoro#English
+    ps = ps.replace('kəkˈoːɹoʊ', 'kˈoʊkəɹoʊ').replace('kəkˈɔːɹəʊ', 'kˈəʊkəɹəʊ')
+    ps = ps.replace('ʲ', 'j').replace('r', 'ɹ').replace('x', 'k').replace('ɬ', 'l')
+    ps = re.sub(r'(?<=[a-zɹː])(?=hˈʌndɹɪd)', ' ', ps)
+    ps = re.sub(r' z(?=[;:,.!?¡¿—…"«»“” ]|$)', 'z', ps)
+    if lang == 'a':
+        ps = re.sub(r'(?<=nˈaɪn)ti(?!ː)', 'di', ps)
+    ps = ''.join(filter(lambda p: p in VOCAB, ps))
+    return ps.strip()
+
+def length_to_mask(lengths):
+    mask = torch.arange(lengths.max()).unsqueeze(0).expand(lengths.shape[0], -1).type_as(lengths)
+    mask = torch.gt(mask+1, lengths.unsqueeze(1))
+    return mask
+
+@torch.no_grad()
+def forward(model, tokens, ref_s, speed):
+    device = ref_s.device
+    tokens = torch.LongTensor([[0, *tokens, 0]]).to(device)
+    input_lengths = torch.LongTensor([tokens.shape[-1]]).to(device)
+    text_mask = length_to_mask(input_lengths).to(device)
+    bert_dur = model.bert(tokens, attention_mask=(~text_mask).int())
+    d_en = model.bert_encoder(bert_dur).transpose(-1, -2)
+    s = ref_s[:, 128:]
+    d = model.predictor.text_encoder(d_en, s, input_lengths, text_mask)
+    x, _ = model.predictor.lstm(d)
+    duration = model.predictor.duration_proj(x)
+    duration = torch.sigmoid(duration).sum(axis=-1) / speed
+    pred_dur = torch.round(duration).clamp(min=1).long()
+    pred_aln_trg = torch.zeros(input_lengths, pred_dur.sum().item())
+    c_frame = 0
+    for i in range(pred_aln_trg.size(0)):
+        pred_aln_trg[i, c_frame:c_frame + pred_dur[0,i].item()] = 1
+        c_frame += pred_dur[0,i].item()
+    en = d.transpose(-1, -2) @ pred_aln_trg.unsqueeze(0).to(device)
+    F0_pred, N_pred = model.predictor.F0Ntrain(en, s)
+    t_en = model.text_encoder(tokens, input_lengths, text_mask)
+    asr = t_en @ pred_aln_trg.unsqueeze(0).to(device)
+    return model.decoder(asr, F0_pred, N_pred, ref_s[:, :128]).squeeze().cpu().numpy()
+
+def generate(model, text, voicepack, lang='a', speed=1, ps=None):
+    ps = ps or phonemize(text, lang)
+    tokens = tokenize(ps)
+    if not tokens:
+        return None
+    elif len(tokens) > 510:
+        tokens = tokens[:510]
+        print('Truncated to 510 tokens')
+    ref_s = voicepack[len(tokens)]
+    out = forward(model, tokens, ref_s, speed)
+    ps = ''.join(next(k for k, v in VOCAB.items() if i == v) for i in tokens)
+    return out, ps
+
+def generate_full(model, text, voicepack, lang='a', speed=1, ps=None):
+    ps = ps or phonemize(text, lang)
+    tokens = tokenize(ps)
+    if not tokens:
+        return None
+    outs = []
+    loop_count = len(tokens)//510 + (1 if len(tokens) % 510 != 0 else 0)
+    for i in range(loop_count):
+        ref_s = voicepack[len(tokens[i*510:(i+1)*510])]
+        out = forward(model, tokens[i*510:(i+1)*510], ref_s, speed)
+        outs.append(out)
+    outs = np.concatenate(outs)
+    ps = ''.join(next(k for k, v in VOCAB.items() if i == v) for i in tokens)
+    return outs, ps