chore: ⬆️ Update ggerganov/llama.cpp to 26a8406ba9198eb6fdd8329fa717555b4f77f05f (#4358 )

chore: ⬆️ Update ggerganov/llama.cpp to `26a8406ba9198eb6fdd8329fa717555b4f77f05f` (#4353) ⬆️ Update ggerganov/llama.cpp Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com> Co-authored-by: LocalAI [bot] <139863280+localai-bot@users.noreply.github.com> Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>
Revert "feat: include tokens usage for streamed output (#4282 )"
2026-02-03 11:13:31 -05:00 · 2024-12-10 15:51:45 +01:00 · 2024-12-08 16:31:48 +01:00
247 changed files with 3795 additions and 7098 deletions
--- a/Requests/tts/musicgen.bru
+++ b/Requests/tts/musicgen.bru
@@ -16,7 +16,7 @@ headers {

 body:json {
  {
-      "backend": "transformers",
+      "backend": "transformers-musicgen",
      "model": "facebook/musicgen-small",
      "input": "80s Synths playing Jazz"
  }
--- a/.devcontainer/docker-compose-devcontainer.yml
+++ b/.devcontainer/docker-compose-devcontainer.yml
@@ -7,7 +7,7 @@ services:
      args:
      - FFMPEG=true
      - IMAGE_TYPE=extras
-      - GO_TAGS=p2p tts
+      - GO_TAGS=stablediffusion p2p tts
    env_file:
      - ../.env
    ports:
--- a/.env
+++ b/.env
@@ -38,12 +38,12 @@
 ## Uncomment and set to true to enable rebuilding from source
 # REBUILD=true

-## Enable go tags, available: p2p, tts
-## p2p: enable distributed inferencing
+## Enable go tags, available: stablediffusion, tts
+## stablediffusion: image generation with stablediffusion
 ## tts: enables text-to-speech with go-piper 
 ## (requires REBUILD=true)
 #
-# GO_TAGS=p2p
+# GO_TAGS=stablediffusion

 ## Path where to store generated images
 # LOCALAI_IMAGE_PATH=/tmp/generated/images
@@ -82,15 +82,6 @@
 # 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
--- a/.github/dependabot.yml
+++ b/.github/dependabot.yml
@@ -81,6 +81,14 @@ 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:
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -5,10 +5,6 @@ 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:
--- a/.github/workflows/image.yml
+++ b/.github/workflows/image.yml
@@ -280,7 +280,6 @@ 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 }}
@@ -302,7 +301,6 @@ 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"
@@ -314,7 +312,6 @@ 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"
@@ -326,7 +323,6 @@ 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"
@@ -338,7 +334,6 @@ 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"
@@ -349,7 +344,6 @@ 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'
@@ -360,45 +354,4 @@ 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: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
-    strategy:
-      matrix:
-        include:
-          - build-type: 'cublas'
-            cuda-major-version: "12"
-            cuda-minor-version: "0"
-            platforms: 'linux/arm64'
-            tag-latest: 'false'
-            tag-suffix: '-nvidia-l4t-arm64-core'
-            latest-image: 'latest-nvidia-l4t-arm64-core'
-            ffmpeg: 'true'
-            image-type: 'core'
-            base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
-            runs-on: 'ubuntu-24.04-arm'
-            makeflags: "--jobs=4 --output-sync=target"
-            skip-drivers: 'true'
--- a/.github/workflows/image_build.yml
+++ b/.github/workflows/image_build.yml
@@ -49,10 +49,6 @@ on:
        description: 'FFMPEG'
        default: ''
        type: string
-      skip-drivers:
-        description: 'Skip drivers by default'
-        default: 'false'
-        type: string
      image-type:
        description: 'Image type'
        default: ''
@@ -238,7 +234,6 @@ jobs:
            GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
            GRPC_VERSION=v1.65.0
            MAKEFLAGS=${{ inputs.makeflags }}
-            SKIP_DRIVERS=${{ inputs.skip-drivers }}
          context: .
          file: ./Dockerfile
          cache-from: type=gha
@@ -267,7 +262,6 @@ jobs:
            GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
            GRPC_VERSION=v1.65.0
            MAKEFLAGS=${{ inputs.makeflags }}
-            SKIP_DRIVERS=${{ inputs.skip-drivers }}
          context: .
          file: ./Dockerfile
          cache-from: type=gha
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@@ -237,7 +237,40 @@ jobs:
          detached: true
          connect-timeout-seconds: 180
          limit-access-to-actor: true
-
+  build-stablediffusion:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with:
+          submodules: true
+      - uses: actions/setup-go@v5
+        with:
+          go-version: '1.21.x'
+          cache: false
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends libopencv-dev protobuf-compiler ccache upx-ucl
+          go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+          go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
+      - name: Build stablediffusion
+        run: |
+          export PATH=$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
--- a/.github/workflows/secscan.yaml
+++ b/.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@v2.22.0
+        uses: securego/gosec@v2.21.4
        with:
          # we let the report trigger content trigger a failure using the GitHub Security features.
          args: '-no-fail -fmt sarif -out results.sarif ./...'
--- a/.github/workflows/test-extra.yml
+++ b/.github/workflows/test-extra.yml
@@ -35,6 +35,30 @@ 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:
@@ -78,27 +102,78 @@ jobs:
          make --jobs=5 --output-sync=target -C backend/python/diffusers
          make --jobs=5 --output-sync=target -C backend/python/diffusers 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
+  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

-  #     - 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 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
+      - name: Setup tmate session if tests fail
+        if: ${{ failure() }}
+        uses: mxschmitt/action-tmate@v3.19
+        with:
+          detached: true
+          connect-timeout-seconds: 180
+          limit-access-to-actor: true
+
+  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

  # tests-bark:
  #   runs-on: ubuntu-latest
@@ -185,6 +260,26 @@ 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
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -100,12 +100,15 @@ jobs:
          # The python3-grpc-tools package in 22.04 is too old
          pip install --user grpcio-tools

-          make -C backend/python/transformers
+          sudo rm -rfv /usr/bin/conda || true
+          PATH=$PATH:/opt/conda/bin make -C backend/python/sentencetransformers

          # 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/
+          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="$PATH:/root/go/bin" GO_TAGS="stablediffusion tts" GRPC_BACKENDS=backend-assets/grpc/stablediffusion make build
        env:
          CUDA_VERSION: 12-4
      - name: Cache grpc
@@ -127,7 +130,7 @@ jobs:
          cd grpc && cd cmake/build && sudo make --jobs 5 install
      - name: Test
        run: |
-          PATH="$PATH:/root/go/bin" GO_TAGS="tts" make --jobs 5 --output-sync=target test
+          PATH="$PATH:/root/go/bin" GO_TAGS="stablediffusion tts" make --jobs 5 --output-sync=target test
      - name: Setup tmate session if tests fail
        if: ${{ failure() }}
        uses: mxschmitt/action-tmate@v3.19
--- a/.vscode/launch.json
+++ b/.vscode/launch.json
@@ -26,7 +26,7 @@
                "LOCALAI_P2P": "true",
                "LOCALAI_FEDERATED": "true"
            },
-            "buildFlags": ["-tags", "p2p tts", "-v"],
+            "buildFlags": ["-tags", "stablediffusion p2p tts", "-v"],
            "envFile": "${workspaceFolder}/.env",
            "cwd": "${workspaceRoot}"
        }
--- a/76
+++ b/76
@@ -15,7 +15,8 @@ ARG TARGETARCH
 ARG TARGETVARIANT

 ENV DEBIAN_FRONTEND=noninteractive
-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"
+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"
+

 RUN apt-get update && \
    apt-get install -y --no-install-recommends \
@@ -68,10 +69,14 @@ ENV PATH=/opt/rocm/bin:${PATH}
 # OpenBLAS requirements and stable diffusion
 RUN apt-get update && \
    apt-get install -y --no-install-recommends \
-        libopenblas-dev && \
+        libopenblas-dev \
+        libopencv-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

 ###################################
@@ -110,13 +115,12 @@ 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 [ "${BUILD_TYPE}" = "vulkan" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
+    if [ "${BUILD_TYPE}" = "vulkan" ]; then
        apt-get update && \
        apt-get install -y  --no-install-recommends \
            software-properties-common pciutils wget gpg-agent && \
@@ -132,7 +136,7 @@ EOT

 # CuBLAS requirements
 RUN <<EOT bash
-    if [ "${BUILD_TYPE}" = "cublas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
+    if [ "${BUILD_TYPE}" = "cublas" ]; then
        apt-get update && \
        apt-get install -y  --no-install-recommends \
            software-properties-common pciutils
@@ -158,7 +162,7 @@ RUN <<EOT bash
 EOT

 # If we are building with clblas support, we need the libraries for the builds
-RUN if [ "${BUILD_TYPE}" = "clblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
+RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
        apt-get update && \
        apt-get install -y --no-install-recommends \
            libclblast-dev && \
@@ -166,7 +170,7 @@ RUN if [ "${BUILD_TYPE}" = "clblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
        rm -rf /var/lib/apt/lists/* \
    ; fi

-RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
+RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
        apt-get update && \
        apt-get install -y --no-install-recommends \
            hipblas-dev \
@@ -246,7 +250,7 @@ RUN git clone --recurse-submodules --jobs 4 -b ${GRPC_VERSION} --depth 1 --shall

 FROM requirements-drivers AS builder-base

-ARG GO_TAGS="tts p2p"
+ARG GO_TAGS="stablediffusion tts p2p"
 ARG GRPC_BACKENDS
 ARG MAKEFLAGS
 ARG LD_FLAGS="-s -w"
@@ -280,12 +284,35 @@ 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-base AS builder
+FROM builder-sd AS builder

 # Install the pre-built GRPC
 COPY --from=grpc /opt/grpc /usr/local
@@ -303,7 +330,7 @@ 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-avx512 backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
+        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
    else \
        make build; \
    fi
@@ -325,6 +352,8 @@ 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
@@ -397,28 +426,36 @@ 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 [[ ( "${EXTRA_BACKENDS}" =~ "faster-whisper" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/faster-whisper \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "parler-tts" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/parler-tts \
    ; 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 [[ ( "${EXTRA_BACKENDS}" =~ "kokoro" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/kokoro \
+RUN if [[ ( "${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 \
    ; fi && \
    if [[ ( "${EXTRA_BACKENDS}" =~ "exllama2" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
        make -C backend/python/exllama2 \
@@ -438,6 +475,9 @@ RUN if [[ ( "${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
--- a/164
+++ b/164
@@ -8,7 +8,7 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=6152129d05870cb38162c422c6ba80434e021e9f
+CPPLLAMA_VERSION?=26a8406ba9198eb6fdd8329fa717555b4f77f05f

 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
@@ -18,13 +18,21 @@ WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
 PIPER_REPO?=https://github.com/mudler/go-piper
 PIPER_VERSION?=e10ca041a885d4a8f3871d52924b47792d5e5aa0

+# stablediffusion version
+STABLEDIFFUSION_REPO?=https://github.com/mudler/go-stable-diffusion
+STABLEDIFFUSION_VERSION?=4a3cd6aeae6f66ee57eae9a0075f8c58c3a6a38f
+
+# tinydream version
+TINYDREAM_REPO?=https://github.com/M0Rf30/go-tiny-dream
+TINYDREAM_VERSION?=c04fa463ace9d9a6464313aa5f9cd0f953b6c057
+
 # bark.cpp
 BARKCPP_REPO?=https://github.com/PABannier/bark.cpp.git
 BARKCPP_VERSION?=v1.0.0

 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=5eb15ef4d022bef4a391de4f5f6556e81fbb5024
+STABLEDIFFUSION_GGML_VERSION?=4570715727f35e5a07a76796d823824c8f42206c

 ONNX_VERSION?=1.20.0
 ONNX_ARCH?=x64
@@ -175,6 +183,16 @@ 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
@@ -186,7 +204,6 @@ endif
 ALL_GRPC_BACKENDS=backend-assets/grpc/huggingface
 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
@@ -265,6 +282,19 @@ sources/go-piper:
 sources/go-piper/libpiper_binding.a: sources/go-piper
 	$(MAKE) -C sources/go-piper libpiper_binding.a example/main piper.o

+## stable diffusion (onnx)
+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 $(STABLEDIFFUSION_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-stable-diffusion/libstablediffusion.a: sources/go-stable-diffusion
+	CPATH="$(CPATH):/usr/include/opencv4" $(MAKE) -C sources/go-stable-diffusion libstablediffusion.a
+
 ## stablediffusion (ggml)
 sources/stablediffusion-ggml.cpp:
 	git clone --recursive $(STABLEDIFFUSION_GGML_REPO) sources/stablediffusion-ggml.cpp && \
@@ -272,8 +302,14 @@ sources/stablediffusion-ggml.cpp:
 	git checkout $(STABLEDIFFUSION_GGML_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch

-backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp
-	$(MAKE) -C backend/go/image/stablediffusion-ggml build/libstable-diffusion.a
+sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a: sources/stablediffusion-ggml.cpp
+	cd sources/stablediffusion-ggml.cpp && \
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) .. && \
+	cmake --build . --config Release
+
+backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a
 	$(MAKE) -C backend/go/image/stablediffusion-ggml libsd.a

 backend-assets/grpc/stablediffusion-ggml: backend/go/image/stablediffusion-ggml/libsd.a backend-assets/grpc
@@ -297,6 +333,19 @@ else
 	mv backend-assets/lib/libonnxruntime.so.$(ONNX_VERSION) backend-assets/lib/libonnxruntime.so.1
 endif

+## 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
+
 ## whisper
 sources/whisper.cpp:
 	mkdir -p sources/whisper.cpp
@@ -310,18 +359,22 @@ 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/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
+get-sources: sources/go-llama.cpp sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp sources/go-stable-diffusion sources/go-tiny-dream backend/cpp/llama/llama.cpp

 replace:
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(CURDIR)/sources/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp/bindings/go=$(CURDIR)/sources/whisper.cpp/bindings/go
+	$(GOCMD) mod edit -replace github.com/M0Rf30/go-tiny-dream=$(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/ggerganov/whisper.cpp
 	$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp/bindings/go
+	$(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
@@ -332,7 +385,9 @@ rebuild: ## Rebuilds the project
 	$(GOCMD) clean -cache
 	$(MAKE) -C sources/go-llama.cpp clean
 	$(MAKE) -C sources/whisper.cpp clean
+	$(MAKE) -C sources/go-stable-diffusion clean
 	$(MAKE) -C sources/go-piper clean
+	$(MAKE) -C sources/go-tiny-dream clean
 	$(MAKE) build

 prepare: prepare-sources $(OPTIONAL_TARGETS)
@@ -446,9 +501,9 @@ prepare-test: grpcs

 test: prepare test-models/testmodel.ggml grpcs
 	@echo 'Running tests'
-	export GO_TAGS="tts debug"
+	export GO_TAGS="tts stablediffusion debug"
 	$(MAKE) prepare-test
-	HUGGINGFACE_GRPC=$(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 \
+	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 \
 	$(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
@@ -534,10 +589,10 @@ protogen-go-clean:
 	$(RM) bin/*

 .PHONY: protogen-python
-protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
+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

 .PHONY: protogen-python-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
+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

 .PHONY: autogptq-protogen
 autogptq-protogen:
@@ -571,14 +626,6 @@ 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
@@ -587,6 +634,14 @@ 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
@@ -595,6 +650,14 @@ 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
@@ -603,13 +666,37 @@ transformers-protogen:
 transformers-protogen-clean:
 	$(MAKE) -C backend/python/transformers protogen-clean

-.PHONY: kokoro-protogen
-kokoro-protogen:
-	$(MAKE) -C backend/python/kokoro protogen
+.PHONY: parler-tts-protogen
+parler-tts-protogen:
+	$(MAKE) -C backend/python/parler-tts protogen

-.PHONY: kokoro-protogen-clean
-kokoro-protogen-clean:
-	$(MAKE) -C backend/python/kokoro protogen-clean
+.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: vllm-protogen
 vllm-protogen:
@@ -626,11 +713,15 @@ 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/kokoro
+	$(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/exllama2

 prepare-test-extra: protogen-python
@@ -700,13 +791,6 @@ 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
 	$(MAKE) -C backend/cpp/llama-avx purge
@@ -781,6 +865,13 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/piper
 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),)
+	$(UPX) backend-assets/grpc/stablediffusion
+endif
+
 backend-assets/grpc/silero-vad: backend-assets/grpc backend-assets/lib/libonnxruntime.so.1
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" CPATH="$(CPATH):$(CURDIR)/sources/onnxruntime/include/" LIBRARY_PATH=$(CURDIR)/backend-assets/lib \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/silero-vad ./backend/go/vad/silero
@@ -788,6 +879,13 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/silero-vad
 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),)
+	$(UPX) backend-assets/grpc/tinydream
+endif
+
 backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/libwhisper.a backend-assets/grpc
 	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="$(CURDIR)/sources/whisper.cpp/include:$(CURDIR)/sources/whisper.cpp/ggml/include" LIBRARY_PATH=$(CURDIR)/sources/whisper.cpp \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./backend/go/transcribe/whisper
--- a/README.md
+++ b/README.md
@@ -39,7 +39,7 @@
 </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>
+<a href="https://trendshift.io/repositories/1484" target="_blank"><img src="https://trendshift.io/api/badge/repositories/1484" alt="go-skynet%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/)
@@ -92,15 +92,19 @@ 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. P2P Global community pools: https://github.com/mudler/LocalAI/issues/3113
+- 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
 - 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)
@@ -109,10 +113,12 @@ 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

@@ -120,10 +126,10 @@ 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`, `transformers`, `vllm` ... [: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`, `gpt4all.cpp`, ... [: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](https://localai.io/features/image-generation)
+- 🎨 [Image generation with stable diffusion](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/)
@@ -131,7 +137,6 @@ 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
@@ -154,7 +159,6 @@ 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
--- a/aio/cpu/image-gen.yaml
+++ b/aio/cpu/image-gen.yaml
@@ -1,17 +1,56 @@
 name: stablediffusion
-backend: stablediffusion-ggml
-cfg_scale: 4.5
-
-options:
- sampler:euler
+backend: stablediffusion
 parameters:
-  model: stable-diffusion-v1-5-pruned-emaonly-Q4_0.gguf
-step: 25
+  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

 download_files:
- 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"
+- 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"

 usage: |
        curl http://localhost:8080/v1/images/generations \
--- a/backend/backend.proto
+++ b/backend/backend.proto
@@ -159,8 +159,6 @@ message Reply {
  bytes message = 1;
  int32 tokens = 2;
  int32 prompt_tokens = 3;
-  double timing_prompt_processing = 4;
-  double timing_token_generation = 5;
 }

 message ModelOptions {
@@ -244,9 +242,6 @@ message ModelOptions {
  repeated float LoraScales = 61;

  repeated string Options = 62;
-
-  string CacheTypeKey = 63;
-  string CacheTypeValue = 64;
 }

 message Result {
@@ -350,4 +345,4 @@ message StatusResponse {
 message Message {
  string role = 1;
  string content = 2;
-}
+}
--- a/backend/cpp/llama/grpc-server.cpp
+++ b/backend/cpp/llama/grpc-server.cpp
@@ -22,7 +22,6 @@
 #include "backend.grpc.pb.h"
 #include "utils.hpp"
 #include "sampling.h"
-#include "speculative.h"
 // include std::regex
 #include <cstddef>
 #include <thread>
@@ -135,32 +134,6 @@ 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)
 {
@@ -186,45 +159,12 @@ 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 llama_slot_params params;
-    common_speculative * spec = nullptr;
-    llama_batch batch_spec = {};
-
+    struct slot_params params;

    slot_state state = IDLE;
    slot_command command = NONE;
@@ -317,7 +257,6 @@ struct llama_client_slot
        images.clear();
    }

-
    bool has_budget(common_params &global_params) {
        if (params.n_predict == -1 && global_params.n_predict == -1)
        {
@@ -489,11 +428,6 @@ 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;

@@ -505,7 +439,6 @@ 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

@@ -541,7 +474,6 @@ struct llama_server_context
        }
    }

-
    bool load_model(const common_params &params_)
    {
        params = params_;
@@ -560,8 +492,8 @@ struct llama_server_context
        }

        common_init_result common_init = common_init_from_params(params);
-        model = common_init.model.release();
-        ctx = common_init.context.release();
+        model = common_init.model;
+        ctx = common_init.context;
        if (model == nullptr)
        {
            LOG_ERR("unable to load model: %s", params.model.c_str());
@@ -570,7 +502,7 @@ struct llama_server_context

        if (multimodal) {
            const int n_embd_clip = clip_n_mmproj_embd(clp_ctx);
-            const int n_embd_llm  = llama_model_n_embd(model);
+            const int n_embd_llm  = llama_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);
@@ -579,54 +511,23 @@ struct llama_server_context
            }
        }

-        vocab = llama_model_get_vocab(model);
        n_ctx = llama_n_ctx(ctx);

-        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();
-        }
+        add_bos_token = llama_add_bos_token(model);

        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
@@ -652,22 +553,6 @@ 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}
@@ -776,11 +661,9 @@ struct llama_server_context
    }

    bool launch_slot_with_data(llama_client_slot* &slot, json data) {
-        llama_slot_params default_params;
+        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);
@@ -798,21 +681,13 @@ 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", {
@@ -851,8 +726,8 @@ struct llama_server_context
            slot->prompt = "";
        }

-        if (json_value(data, "ignore_eos", false) && has_eos_token) {
-                slot->sparams.logit_bias.push_back({llama_vocab_eos(vocab), -INFINITY});
+        if (json_value(data, "ignore_eos", false)) {
+                slot->sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
        }
        /*
        slot->sparams.penalty_prompt_tokens.clear();
@@ -891,13 +766,13 @@ 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 llama_vocab * vocab = llama_model_get_vocab(model);
-            const int n_vocab = llama_vocab_n_tokens(vocab);
+            const int n_vocab = llama_n_vocab(model);
            for (const auto &el : *logit_bias)
            {
                if (el.is_array() && el.size() == 2)
@@ -926,7 +801,7 @@ struct llama_server_context
                    }
                    else if (el[0].is_string())
                    {
-                        auto toks = common_tokenize(vocab, el[0].get<std::string>(), false);
+                        auto toks = common_tokenize(model, el[0].get<std::string>(), false);
                        for (auto tok : toks)
                        {
                            slot->sparams.logit_bias.push_back({tok, bias});
@@ -1256,7 +1131,7 @@ struct llama_server_context
            slot.has_next_token = false;
        }

-        if (result.tok == llama_vocab_eos(vocab) || llama_vocab_is_eog(vocab, result.tok))
+        if (result.tok == llama_token_eos(model))
        {
            slot.stopped_eos = true;
            slot.has_next_token = false;
@@ -1338,12 +1213,13 @@ 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},
@@ -1451,7 +1327,7 @@ struct llama_server_context
        res.error = false;
        res.stop = true;

-        const int n_embd = llama_model_n_embd(model);
+        const int n_embd = llama_n_embd(model);
        if (!params.embedding)
        {
            LOG_WARNING("embedding disabled", {
@@ -1550,7 +1426,7 @@ struct llama_server_context
                    n_eval = n_batch;
                }

-                const int n_embd = llama_model_n_embd(model);
+                const int n_embd = llama_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))
@@ -1831,11 +1707,11 @@ struct llama_server_context
                            suffix_tokens.erase(suffix_tokens.begin());
                        }

-                        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.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.end(),   suffix_tokens.begin(), suffix_tokens.end());
-                        prefix_tokens.push_back(llama_vocab_fim_mid(vocab));
+                        prefix_tokens.push_back(llama_token_middle(model));
                        prompt_tokens = prefix_tokens;
                    }
                    else
@@ -2130,97 +2006,6 @@ 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;
    }
@@ -2327,6 +2112,7 @@ 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);
@@ -2349,6 +2135,7 @@ 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();
@@ -2394,6 +2181,7 @@ 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();
@@ -2440,35 +2228,6 @@ 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) {
   
@@ -2482,41 +2241,11 @@ 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");
@@ -2529,7 +2258,7 @@ static void params_parse(const backend::ModelOptions* request,

    const char *llama_grpc_servers = std::getenv("LLAMACPP_GRPC_SERVERS");
    if (llama_grpc_servers != NULL) {
-        add_rpc_devices(std::string(llama_grpc_servers));
+        params.rpc_servers = std::string(llama_grpc_servers);
    }
    
    // TODO: Add yarn
@@ -2655,13 +2384,6 @@ 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"] }
@@ -2702,13 +2424,6 @@ 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
        {
--- a/backend/cpp/llama/patches/01-llava.patch
+++ b/backend/cpp/llama/patches/01-llava.patch
@@ -1,13 +1,13 @@
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 3cd0d2fa..6c5e811a 100644
+index 342042ff..224db9b5 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -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);
+@@ -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);
--- a/backend/go/image/stablediffusion-ggml/Makefile
+++ b/backend/go/image/stablediffusion-ggml/Makefile
@@ -2,95 +2,20 @@ 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
+	cp $(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a ./libsd.a
 	$(AR) rcs libsd.a gosd.o

 clean:
-	rm -rf gosd.o libsd.a build $(COMBINED_LIB)
+	rm -f gosd.o libsd.a
--- a/backend/go/image/stablediffusion-ggml/gosd.go
+++ b/backend/go/image/stablediffusion-ggml/gosd.go
@@ -1,7 +1,7 @@
 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
+// #cgo LDFLAGS: -L${SRCDIR}/ -L${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/build/ggml/src/ggml-cpu -L${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/build/ggml/src -lsd -lstdc++ -lm -lggml -lggml-base -lggml-cpu -lgomp
 // #include <gosd.h>
 // #include <stdlib.h>
 import "C"
--- a/backend/go/image/stablediffusion/main.go
+++ b/backend/go/image/stablediffusion/main.go
@@ -0,0 +1,21 @@
+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, &Image{}); err != nil {
+		panic(err)
+	}
+}
--- a/backend/go/image/stablediffusion/stablediffusion.go
+++ b/backend/go/image/stablediffusion/stablediffusion.go
@@ -0,0 +1,33 @@
+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)
+}
--- a/backend/go/image/tinydream/main.go
+++ b/backend/go/image/tinydream/main.go
@@ -0,0 +1,21 @@
+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, &Image{}); err != nil {
+		panic(err)
+	}
+}
--- a/backend/go/image/tinydream/tinydream.go
+++ b/backend/go/image/tinydream/tinydream.go
@@ -0,0 +1,32 @@
+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)
+}
--- a/backend/go/stores/store.go
+++ b/backend/go/stores/store.go
@@ -311,16 +311,12 @@ func (s *Store) StoresGet(opts *pb.StoresGetOptions) (pb.StoresGetResult, error)
 }

 func isNormalized(k []float32) bool {
-	var sum float64
-
+	var sum float32
 	for _, v := range k {
-		v64 := float64(v)
-		sum += v64*v64
+		sum += v
 	}

-	s := math.Sqrt(sum)
-
-	return s >= 0.99 && s <= 1.01
+	return sum == 1.0
 }

 // TODO: This we could replace with handwritten SIMD code
@@ -332,7 +328,7 @@ func normalizedCosineSimilarity(k1, k2 []float32) float32 {
 		dot += k1[i] * k2[i]
 	}

-	assert(dot >= -1.01 && dot <= 1.01, fmt.Sprintf("dot = %f", dot))
+	assert(dot >= -1 && dot <= 1, fmt.Sprintf("dot = %f", dot))

 	// 2.0 * (1.0 - dot) would be the Euclidean distance
 	return dot
@@ -422,7 +418,7 @@ func cosineSimilarity(k1, k2 []float32, mag1 float64) float32 {

 	sim := float32(dot / (mag1 * math.Sqrt(mag2)))

-	assert(sim >= -1.01 && sim <= 1.01, fmt.Sprintf("sim = %f", sim))
+	assert(sim >= -1 && sim <= 1, fmt.Sprintf("sim = %f", sim))

 	return sim
 }
--- a/backend/python/autogptq/requirements-intel.txt
+++ b/backend/python/autogptq/requirements-intel.txt
@@ -1,6 +1,5 @@
 --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
+intel-extension-for-pytorch
+torch
 optimum[openvino]
-setuptools
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
--- a/backend/python/autogptq/requirements.txt
+++ b/backend/python/autogptq/requirements.txt
@@ -1,6 +1,6 @@
 accelerate
 auto-gptq==0.7.1
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 certifi
 transformers
--- a/backend/python/bark/requirements-intel.txt
+++ b/backend/python/bark/requirements-intel.txt
@@ -1,9 +1,8 @@
 --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
-torchaudio==2.3.1+cxx11.abi
-oneccl_bind_pt==2.3.100+xpu
+intel-extension-for-pytorch
+torch
+torchaudio
 optimum[openvino]
-setuptools
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
 transformers
 accelerate
--- a/backend/python/bark/requirements.txt
+++ b/backend/python/bark/requirements.txt
@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 certifi
--- a/backend/python/common/libbackend.sh
+++ b/backend/python/common/libbackend.sh
@@ -17,9 +17,6 @@
 # LIMIT_TARGETS="cublas12"
 # source $(dirname $0)/../common/libbackend.sh
 #
-
-PYTHON_VERSION="3.10"
-
 function init() {
    # Name of the backend (directory name)
    BACKEND_NAME=${PWD##*/}
@@ -91,7 +88,7 @@ function getBuildProfile() {
 # always result in an activated virtual environment
 function ensureVenv() {
    if [ ! -d "${EDIR}/venv" ]; then
-        uv venv --python ${PYTHON_VERSION} ${EDIR}/venv
+        uv venv ${EDIR}/venv
        echo "virtualenv created"
    fi

--- a/backend/python/common/template/requirements-intel.txt
+++ b/backend/python/common/template/requirements-intel.txt
@@ -1,5 +1,4 @@
 --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
+intel-extension-for-pytorch
+torch
 optimum[openvino]
--- a/backend/python/common/template/requirements.txt
+++ b/backend/python/common/template/requirements.txt
@@ -1,3 +1,3 @@
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 grpcio-tools
--- a/backend/python/coqui/requirements-intel.txt
+++ b/backend/python/coqui/requirements-intel.txt
@@ -1,10 +1,9 @@
 --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
-torchaudio==2.3.1+cxx11.abi
-oneccl_bind_pt==2.3.100+xpu
+intel-extension-for-pytorch
+torch
+torchaudio
 optimum[openvino]
-setuptools
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
 transformers
 accelerate
 coqui-tts
--- a/backend/python/coqui/requirements.txt
+++ b/backend/python/coqui/requirements.txt
@@ -1,4 +1,4 @@
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 certifi
 packaging==24.1
--- a/backend/python/diffusers/backend.py
+++ b/backend/python/diffusers/backend.py
@@ -17,7 +17,7 @@ import backend_pb2_grpc

 import grpc

-from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
+from diffusers import StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
 from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
@@ -275,13 +275,6 @@ 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
--- a/backend/python/diffusers/requirements-intel.txt
+++ b/backend/python/diffusers/requirements-intel.txt
@@ -1,10 +1,9 @@
 --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
-torchvision==0.18.1+cxx11.abi
-oneccl_bind_pt==2.3.100+xpu
+intel-extension-for-pytorch
+torch
+torchvision
 optimum[openvino]
-setuptools
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
 diffusers
 opencv-python
 transformers
--- a/backend/python/diffusers/requirements.txt
+++ b/backend/python/diffusers/requirements.txt
@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.69.0
+grpcio==1.68.1
 pillow
 protobuf
 certifi
--- a/backend/python/exllama2/requirements.txt
+++ b/backend/python/exllama2/requirements.txt
@@ -1,4 +1,4 @@
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 certifi
 wheel
--- a/backend/python/faster-whisper/requirements-cpu.txt
+++ b/backend/python/faster-whisper/requirements-cpu.txt
@@ -1,8 +0,0 @@
-faster-whisper
-opencv-python
-accelerate
-compel
-peft
-sentencepiece
-torch==2.4.1
-optimum-quanto
--- a/backend/python/faster-whisper/requirements-cublas12.txt
+++ b/backend/python/faster-whisper/requirements-cublas12.txt
@@ -1,8 +0,0 @@
-torch==2.4.1
-faster-whisper
-opencv-python
-accelerate
-compel
-peft
-sentencepiece
-optimum-quanto
--- a/backend/python/faster-whisper/requirements-intel.txt
+++ b/backend/python/faster-whisper/requirements-intel.txt
@@ -1,6 +0,0 @@
--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
--- a/backend/python/faster-whisper/requirements.txt
+++ b/backend/python/faster-whisper/requirements.txt
@@ -1,3 +0,0 @@
-grpcio==1.69.0
-protobuf
-grpcio-tools
--- a/backend/python/kokoro/Makefile
+++ b/backend/python/kokoro/Makefile
@@ -1,20 +0,0 @@
-.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__
--- a/backend/python/kokoro/istftnet.py
+++ b/backend/python/kokoro/istftnet.py
@@ -1,524 +0,0 @@
-# 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
--- a/backend/python/kokoro/kokoro.py
+++ b/backend/python/kokoro/kokoro.py
@@ -1,166 +0,0 @@
-# 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
--- a/backend/python/kokoro/models.py
+++ b/backend/python/kokoro/models.py
@@ -1,373 +0,0 @@
-# https://github.com/yl4579/StyleTTS2/blob/main/models.py
-# https://huggingface.co/hexgrad/Kokoro-82M/blob/main/models.py
-from istftnet import AdaIN1d, Decoder
-from munch import Munch
-from pathlib import Path
-from plbert import load_plbert
-from torch.nn.utils import weight_norm, spectral_norm
-import json
-import numpy as np
-import os
-import os.path as osp
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-
-class LinearNorm(torch.nn.Module):
-    def __init__(self, in_dim, out_dim, bias=True, w_init_gain='linear'):
-        super(LinearNorm, self).__init__()
-        self.linear_layer = torch.nn.Linear(in_dim, out_dim, bias=bias)
-
-        torch.nn.init.xavier_uniform_(
-            self.linear_layer.weight,
-            gain=torch.nn.init.calculate_gain(w_init_gain))
-
-    def forward(self, x):
-        return self.linear_layer(x)
-
-class LayerNorm(nn.Module):
-    def __init__(self, channels, eps=1e-5):
-        super().__init__()
-        self.channels = channels
-        self.eps = eps
-
-        self.gamma = nn.Parameter(torch.ones(channels))
-        self.beta = nn.Parameter(torch.zeros(channels))
-
-    def forward(self, x):
-        x = x.transpose(1, -1)
-        x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
-        return x.transpose(1, -1)
-    
-class TextEncoder(nn.Module):
-    def __init__(self, channels, kernel_size, depth, n_symbols, actv=nn.LeakyReLU(0.2)):
-        super().__init__()
-        self.embedding = nn.Embedding(n_symbols, channels)
-
-        padding = (kernel_size - 1) // 2
-        self.cnn = nn.ModuleList()
-        for _ in range(depth):
-            self.cnn.append(nn.Sequential(
-                weight_norm(nn.Conv1d(channels, channels, kernel_size=kernel_size, padding=padding)),
-                LayerNorm(channels),
-                actv,
-                nn.Dropout(0.2),
-            ))
-        # self.cnn = nn.Sequential(*self.cnn)
-
-        self.lstm = nn.LSTM(channels, channels//2, 1, batch_first=True, bidirectional=True)
-
-    def forward(self, x, input_lengths, m):
-        x = self.embedding(x)  # [B, T, emb]
-        x = x.transpose(1, 2)  # [B, emb, T]
-        m = m.to(input_lengths.device).unsqueeze(1)
-        x.masked_fill_(m, 0.0)
-        
-        for c in self.cnn:
-            x = c(x)
-            x.masked_fill_(m, 0.0)
-            
-        x = x.transpose(1, 2)  # [B, T, chn]
-
-        input_lengths = input_lengths.cpu().numpy()
-        x = nn.utils.rnn.pack_padded_sequence(
-            x, input_lengths, batch_first=True, enforce_sorted=False)
-
-        self.lstm.flatten_parameters()
-        x, _ = self.lstm(x)
-        x, _ = nn.utils.rnn.pad_packed_sequence(
-            x, batch_first=True)
-                
-        x = x.transpose(-1, -2)
-        x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]])
-
-        x_pad[:, :, :x.shape[-1]] = x
-        x = x_pad.to(x.device)
-        
-        x.masked_fill_(m, 0.0)
-        
-        return x
-
-    def inference(self, x):
-        x = self.embedding(x)
-        x = x.transpose(1, 2)
-        x = self.cnn(x)
-        x = x.transpose(1, 2)
-        self.lstm.flatten_parameters()
-        x, _ = self.lstm(x)
-        return x
-    
-    def length_to_mask(self, 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
-
-
-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 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 AdaLayerNorm(nn.Module):
-    def __init__(self, style_dim, channels, eps=1e-5):
-        super().__init__()
-        self.channels = channels
-        self.eps = eps
-
-        self.fc = nn.Linear(style_dim, channels*2)
-
-    def forward(self, x, s):
-        x = x.transpose(-1, -2)
-        x = x.transpose(1, -1)
-                
-        h = self.fc(s)
-        h = h.view(h.size(0), h.size(1), 1)
-        gamma, beta = torch.chunk(h, chunks=2, dim=1)
-        gamma, beta = gamma.transpose(1, -1), beta.transpose(1, -1)
-        
-        
-        x = F.layer_norm(x, (self.channels,), eps=self.eps)
-        x = (1 + gamma) * x + beta
-        return x.transpose(1, -1).transpose(-1, -2)
-
-class ProsodyPredictor(nn.Module):
-
-    def __init__(self, style_dim, d_hid, nlayers, max_dur=50, dropout=0.1):
-        super().__init__() 
-        
-        self.text_encoder = DurationEncoder(sty_dim=style_dim, 
-                                            d_model=d_hid,
-                                            nlayers=nlayers, 
-                                            dropout=dropout)
-
-        self.lstm = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
-        self.duration_proj = LinearNorm(d_hid, max_dur)
-        
-        self.shared = nn.LSTM(d_hid + style_dim, d_hid // 2, 1, batch_first=True, bidirectional=True)
-        self.F0 = nn.ModuleList()
-        self.F0.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
-        self.F0.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
-        self.F0.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
-
-        self.N = nn.ModuleList()
-        self.N.append(AdainResBlk1d(d_hid, d_hid, style_dim, dropout_p=dropout))
-        self.N.append(AdainResBlk1d(d_hid, d_hid // 2, style_dim, upsample=True, dropout_p=dropout))
-        self.N.append(AdainResBlk1d(d_hid // 2, d_hid // 2, style_dim, dropout_p=dropout))
-        
-        self.F0_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
-        self.N_proj = nn.Conv1d(d_hid // 2, 1, 1, 1, 0)
-
-
-    def forward(self, texts, style, text_lengths, alignment, m):
-        d = self.text_encoder(texts, style, text_lengths, m)
-        
-        batch_size = d.shape[0]
-        text_size = d.shape[1]
-        
-        # predict duration
-        input_lengths = text_lengths.cpu().numpy()
-        x = nn.utils.rnn.pack_padded_sequence(
-            d, input_lengths, batch_first=True, enforce_sorted=False)
-        
-        m = m.to(text_lengths.device).unsqueeze(1)
-        
-        self.lstm.flatten_parameters()
-        x, _ = self.lstm(x)
-        x, _ = nn.utils.rnn.pad_packed_sequence(
-            x, batch_first=True)
-        
-        x_pad = torch.zeros([x.shape[0], m.shape[-1], x.shape[-1]])
-
-        x_pad[:, :x.shape[1], :] = x
-        x = x_pad.to(x.device)
-                
-        duration = self.duration_proj(nn.functional.dropout(x, 0.5, training=self.training))
-        
-        en = (d.transpose(-1, -2) @ alignment)
-
-        return duration.squeeze(-1), en
-    
-    def F0Ntrain(self, x, s):
-        x, _ = self.shared(x.transpose(-1, -2))
-        
-        F0 = x.transpose(-1, -2)
-        for block in self.F0:
-            F0 = block(F0, s)
-        F0 = self.F0_proj(F0)
-
-        N = x.transpose(-1, -2)
-        for block in self.N:
-            N = block(N, s)
-        N = self.N_proj(N)
-        
-        return F0.squeeze(1), N.squeeze(1)
-    
-    def length_to_mask(self, 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
-
-class DurationEncoder(nn.Module):
-
-    def __init__(self, sty_dim, d_model, nlayers, dropout=0.1):
-        super().__init__()
-        self.lstms = nn.ModuleList()
-        for _ in range(nlayers):
-            self.lstms.append(nn.LSTM(d_model + sty_dim, 
-                                 d_model // 2, 
-                                 num_layers=1, 
-                                 batch_first=True, 
-                                 bidirectional=True, 
-                                 dropout=dropout))
-            self.lstms.append(AdaLayerNorm(sty_dim, d_model))
-        
-        
-        self.dropout = dropout
-        self.d_model = d_model
-        self.sty_dim = sty_dim
-
-    def forward(self, x, style, text_lengths, m):
-        masks = m.to(text_lengths.device)
-        
-        x = x.permute(2, 0, 1)
-        s = style.expand(x.shape[0], x.shape[1], -1)
-        x = torch.cat([x, s], axis=-1)
-        x.masked_fill_(masks.unsqueeze(-1).transpose(0, 1), 0.0)
-                
-        x = x.transpose(0, 1)
-        input_lengths = text_lengths.cpu().numpy()
-        x = x.transpose(-1, -2)
-        
-        for block in self.lstms:
-            if isinstance(block, AdaLayerNorm):
-                x = block(x.transpose(-1, -2), style).transpose(-1, -2)
-                x = torch.cat([x, s.permute(1, -1, 0)], axis=1)
-                x.masked_fill_(masks.unsqueeze(-1).transpose(-1, -2), 0.0)
-            else:
-                x = x.transpose(-1, -2)
-                x = nn.utils.rnn.pack_padded_sequence(
-                    x, input_lengths, batch_first=True, enforce_sorted=False)
-                block.flatten_parameters()
-                x, _ = block(x)
-                x, _ = nn.utils.rnn.pad_packed_sequence(
-                    x, batch_first=True)
-                x = F.dropout(x, p=self.dropout, training=self.training)
-                x = x.transpose(-1, -2)
-                
-                x_pad = torch.zeros([x.shape[0], x.shape[1], m.shape[-1]])
-
-                x_pad[:, :, :x.shape[-1]] = x
-                x = x_pad.to(x.device)
-        
-        return x.transpose(-1, -2)
-    
-    def inference(self, x, style):
-        x = self.embedding(x.transpose(-1, -2)) * np.sqrt(self.d_model)
-        style = style.expand(x.shape[0], x.shape[1], -1)
-        x = torch.cat([x, style], axis=-1)
-        src = self.pos_encoder(x)
-        output = self.transformer_encoder(src).transpose(0, 1)
-        return output
-    
-    def length_to_mask(self, 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
-
-# https://github.com/yl4579/StyleTTS2/blob/main/utils.py
-def recursive_munch(d):
-    if isinstance(d, dict):
-        return Munch((k, recursive_munch(v)) for k, v in d.items())
-    elif isinstance(d, list):
-        return [recursive_munch(v) for v in d]
-    else:
-        return d
-
-def build_model(path, device):
-    config = Path(__file__).parent / 'config.json'
-    assert config.exists(), f'Config path incorrect: config.json not found at {config}'
-    with open(config, 'r') as r:
-        args = recursive_munch(json.load(r))
-    assert args.decoder.type == 'istftnet', f'Unknown decoder type: {args.decoder.type}'
-    decoder = Decoder(dim_in=args.hidden_dim, style_dim=args.style_dim, dim_out=args.n_mels,
-            resblock_kernel_sizes = args.decoder.resblock_kernel_sizes,
-            upsample_rates = args.decoder.upsample_rates,
-            upsample_initial_channel=args.decoder.upsample_initial_channel,
-            resblock_dilation_sizes=args.decoder.resblock_dilation_sizes,
-            upsample_kernel_sizes=args.decoder.upsample_kernel_sizes,
-            gen_istft_n_fft=args.decoder.gen_istft_n_fft, gen_istft_hop_size=args.decoder.gen_istft_hop_size)
-    text_encoder = TextEncoder(channels=args.hidden_dim, kernel_size=5, depth=args.n_layer, n_symbols=args.n_token)
-    predictor = ProsodyPredictor(style_dim=args.style_dim, d_hid=args.hidden_dim, nlayers=args.n_layer, max_dur=args.max_dur, dropout=args.dropout)
-    bert = load_plbert()
-    bert_encoder = nn.Linear(bert.config.hidden_size, args.hidden_dim)
-    for parent in [bert, bert_encoder, predictor, decoder, text_encoder]:
-        for child in parent.children():
-            if isinstance(child, nn.RNNBase):
-                child.flatten_parameters()
-    model = Munch(
-        bert=bert.to(device).eval(),
-        bert_encoder=bert_encoder.to(device).eval(),
-        predictor=predictor.to(device).eval(),
-        decoder=decoder.to(device).eval(),
-        text_encoder=text_encoder.to(device).eval(),
-    )
-    for key, state_dict in torch.load(path, map_location='cpu', weights_only=True)['net'].items():
-        assert key in model, key
-        try:
-            model[key].load_state_dict(state_dict)
-        except:
-            state_dict = {k[7:]: v for k, v in state_dict.items()}
-            model[key].load_state_dict(state_dict, strict=False)
-    return model
--- a/backend/python/kokoro/plbert.py
+++ b/backend/python/kokoro/plbert.py
@@ -1,16 +0,0 @@
-# https://huggingface.co/hexgrad/Kokoro-82M/blob/main/plbert.py
-# https://github.com/yl4579/StyleTTS2/blob/main/Utils/PLBERT/util.py
-from transformers import AlbertConfig, AlbertModel
-
-class CustomAlbert(AlbertModel):
-    def forward(self, *args, **kwargs):
-        # Call the original forward method
-        outputs = super().forward(*args, **kwargs)
-        # Only return the last_hidden_state
-        return outputs.last_hidden_state
-
-def load_plbert():
-    plbert_config = {'vocab_size': 178, 'hidden_size': 768, 'num_attention_heads': 12, 'intermediate_size': 2048, 'max_position_embeddings': 512, 'num_hidden_layers': 12, 'dropout': 0.1}
-    albert_base_configuration = AlbertConfig(**plbert_config)
-    bert = CustomAlbert(albert_base_configuration)
-    return bert
--- a/backend/python/kokoro/protogen.sh
+++ b/backend/python/kokoro/protogen.sh
@@ -1,6 +0,0 @@
-#!/bin/bash
-set -e
-
-source $(dirname $0)/../common/libbackend.sh
-
-python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
--- a/backend/python/kokoro/requirements-intel.txt
+++ b/backend/python/kokoro/requirements-intel.txt
@@ -1,5 +0,0 @@
--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
-transformers
--- a/backend/python/kokoro/requirements.txt
+++ b/backend/python/kokoro/requirements.txt
@@ -1,7 +0,0 @@
-grpcio==1.69.0
-protobuf
-phonemizer
-scipy
-munch
-setuptools
-soundfile
--- a/backend/python/mamba/Makefile
+++ b/backend/python/mamba/Makefile
@@ -0,0 +1,29 @@
+.PHONY: mamba
+mamba: protogen
+	bash install.sh 
+
+.PHONY: run
+run: protogen
+	@echo "Running mamba..."
+	bash run.sh
+	@echo "mamba run."
+
+.PHONY: test
+test: protogen
+	@echo "Testing mamba..."
+	bash test.sh
+	@echo "mamba tested."
+
+.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:
+	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+
+.PHONY: clean
+clean: protogen-clean
+	$(RM) -r venv __pycache__
--- a/backend/python/mamba/README.md
+++ b/backend/python/mamba/README.md
@@ -0,0 +1,5 @@
+# Creating a separate environment for the mamba project
+
+```
+make mamba
+```
--- a/backend/python/mamba/backend.py
+++ b/backend/python/mamba/backend.py
@@ -0,0 +1,179 @@
+#!/usr/bin/env python3
+from concurrent import futures
+import time
+import argparse
+import signal
+import sys
+import os
+
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel
+
+_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'))
+MAMBA_CHAT= os.environ.get('MAMBA_CHAT', '1') == '1'
+
+# Implement the BackendServicer class with the service methods
+class BackendServicer(backend_pb2_grpc.BackendServicer):
+    """
+    A gRPC servicer that implements the Backend service defined in backend.proto.
+    """
+    def generate(self,prompt, max_new_tokens):
+        """
+        Generates text based on the given prompt and maximum number of new tokens.
+
+        Args:
+            prompt (str): The prompt to generate text from.
+            max_new_tokens (int): The maximum number of new tokens to generate.
+
+        Returns:
+            str: The generated text.
+        """
+        self.generator.end_beam_search()
+
+        # Tokenizing the input
+        ids = self.generator.tokenizer.encode(prompt)
+
+        self.generator.gen_begin_reuse(ids)
+        initial_len = self.generator.sequence[0].shape[0]
+        has_leading_space = False
+        decoded_text = ''
+        for i in range(max_new_tokens):
+            token = self.generator.gen_single_token()
+            if i == 0 and self.generator.tokenizer.tokenizer.IdToPiece(int(token)).startswith('▁'):
+                has_leading_space = True
+
+            decoded_text = self.generator.tokenizer.decode(self.generator.sequence[0][initial_len:])
+            if has_leading_space:
+                decoded_text = ' ' + decoded_text
+
+            if token.item() == self.generator.tokenizer.eos_token_id:
+                break
+        return decoded_text
+
+    def Health(self, request, context):
+        """
+        Returns a health check message.
+
+        Args:
+            request: The health check request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Reply: The health check reply.
+        """
+        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
+
+    def LoadModel(self, request, context):
+        """
+        Loads a language model.
+
+        Args:
+            request: The load model request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Result: The load model result.
+        """
+        try:
+            tokenizerModel = request.Tokenizer
+            if tokenizerModel == "":
+                tokenizerModel = request.Model
+
+            tokenizer = AutoTokenizer.from_pretrained(tokenizerModel)
+            if MAMBA_CHAT:
+                tokenizer.eos_token = "<|endoftext|>"
+                tokenizer.pad_token = tokenizer.eos_token
+            self.tokenizer = tokenizer
+            self.model = MambaLMHeadModel.from_pretrained(request.Model, device="cuda", dtype=torch.float16)
+        except Exception as err:
+            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
+        return backend_pb2.Result(message="Model loaded successfully", success=True)
+
+    def Predict(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters.
+
+        Args:
+            request: The predict request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Result: The predict result.
+        """
+        if request.TopP == 0:
+            request.TopP = 0.9
+
+        max_tokens = request.Tokens
+
+        if request.Tokens == 0:
+            max_tokens = 2000
+
+        # encoded_input = self.tokenizer(request.Prompt)
+        tokens = self.tokenizer(request.Prompt, return_tensors="pt")
+        input_ids = tokens.input_ids.to(device="cuda")
+        out = self.model.generate(input_ids=input_ids, max_length=max_tokens, temperature=request.Temperature,
+                                     top_p=request.TopP, eos_token_id=self.tokenizer.eos_token_id)
+
+        decoded = self.tokenizer.batch_decode(out)
+       
+        generated_text = decoded[0]
+
+        # Remove prompt from response if present
+        if request.Prompt in generated_text:
+            generated_text = generated_text.replace(request.Prompt, "")
+
+        return backend_pb2.Reply(message=bytes(generated_text, encoding='utf-8'))
+
+    def PredictStream(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters, and streams the results.
+
+        Args:
+            request: The predict stream request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Result: The predict stream result.
+        """
+        yield self.Predict(request, context)
+
+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("Server started. Listening on: " + address, file=sys.stderr)
+
+    # Define the signal handler function
+    def signal_handler(sig, frame):
+        print("Received termination signal. Shutting down...")
+        server.stop(0)
+        sys.exit(0)
+
+    # Set the signal handlers for SIGINT and SIGTERM
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    try:
+        while True:
+            time.sleep(_ONE_DAY_IN_SECONDS)
+    except KeyboardInterrupt:
+        server.stop(0)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the gRPC server.")
+    parser.add_argument(
+        "--addr", default="localhost:50051", help="The address to bind the server to."
+    )
+    args = parser.parse_args()
+
+    serve(args.addr)
--- a/backend/python/mamba/install.sh
+++ b/backend/python/mamba/install.sh
@@ -0,0 +1,9 @@
+#!/bin/bash
+set -e
+
+LIMIT_TARGETS="cublas"
+EXTRA_PIP_INSTALL_FLAGS="--no-build-isolation"
+
+source $(dirname $0)/../common/libbackend.sh
+
+installRequirements
--- a/backend/python/mamba/requirements-after.txt
+++ b/backend/python/mamba/requirements-after.txt
@@ -0,0 +1,2 @@
+causal-conv1d==1.4.0
+mamba-ssm==2.2.2
--- a/backend/python/kokoro/requirements-cpu.txt
+++ b/backend/python/kokoro/requirements-cpu.txt
--- a/backend/python/kokoro/requirements-cublas11.txt
+++ b/backend/python/kokoro/requirements-cublas11.txt
--- a/backend/python/kokoro/requirements-cublas12.txt
+++ b/backend/python/kokoro/requirements-cublas12.txt
--- a/backend/python/mamba/requirements-install.txt
+++ b/backend/python/mamba/requirements-install.txt
@@ -0,0 +1,6 @@
+# mabma does not specify it's build dependencies per PEP517, so we need to disable build isolation
+# this also means that we need to install the basic build dependencies into the venv ourselves
+# https://github.com/Dao-AILab/causal-conv1d/issues/24
+packaging
+setuptools
+wheel
--- a/backend/python/mamba/requirements.txt
+++ b/backend/python/mamba/requirements.txt
@@ -0,0 +1,3 @@
+grpcio==1.68.1
+protobuf
+certifi
--- a/backend/python/mamba/run.sh
+++ b/backend/python/mamba/run.sh
@@ -0,0 +1,6 @@
+#!/bin/bash
+LIMIT_TARGETS="cublas"
+
+source $(dirname $0)/../common/libbackend.sh
+
+startBackend $@
--- a/backend/python/mamba/test.py
+++ b/backend/python/mamba/test.py
@@ -0,0 +1,76 @@
+import unittest
+import subprocess
+import time
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+
+import unittest
+import subprocess
+import time
+import grpc
+import backend_pb2_grpc
+import backend_pb2
+
+class TestBackendServicer(unittest.TestCase):
+    """
+    TestBackendServicer is the class that tests the gRPC service.
+
+    This class contains methods to test the startup and shutdown of the gRPC service.
+    """
+    def setUp(self):
+        self.service = subprocess.Popen(["python", "backend.py", "--addr", "localhost:50051"])
+        time.sleep(10)
+
+    def tearDown(self) -> None:
+        self.service.terminate()
+        self.service.wait()
+
+    def test_server_startup(self):
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.Health(backend_pb2.HealthMessage())
+                self.assertEqual(response.message, b'OK')
+        except Exception as err:
+            print(err)
+            self.fail("Server failed to start")
+        finally:
+            self.tearDown()
+    def test_load_model(self):
+        """
+        This method tests if the model is loaded successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="facebook/opt-125m"))
+                self.assertTrue(response.success)
+                self.assertEqual(response.message, "Model loaded successfully")
+        except Exception as err:
+            print(err)
+            self.fail("LoadModel service failed")
+        finally:
+            self.tearDown()
+
+    def test_text(self):
+        """
+        This method tests if the embeddings are generated successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="facebook/opt-125m"))
+                self.assertTrue(response.success)
+                req = backend_pb2.PredictOptions(Prompt="The capital of France is")
+                resp = stub.Predict(req)
+                self.assertIsNotNone(resp.message)
+        except Exception as err:
+            print(err)
+            self.fail("text service failed")
+        finally:
+            self.tearDown()
--- a/backend/python/faster-whisper/test.sh
+++ b/backend/python/faster-whisper/test.sh
--- a/backend/python/faster-whisper/Makefile
+++ b/backend/python/faster-whisper/Makefile
@@ -1,9 +1,8 @@
 .DEFAULT_GOAL := install

 .PHONY: install
-install:
+install: protogen
 	bash install.sh
-	$(MAKE) protogen

 .PHONY: protogen
 protogen: backend_pb2_grpc.py backend_pb2.py
@@ -13,8 +12,14 @@ protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py

 backend_pb2_grpc.py backend_pb2.py:
-	bash protogen.sh
+	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto

 .PHONY: clean
 clean: protogen-clean
-	rm -rf venv __pycache__
+	rm -rf venv __pycache__
+
+.PHONY: test
+test: protogen
+	@echo "Testing openvoice..."
+	bash test.sh
+	@echo "openvoice tested."
--- a/backend/python/openvoice/backend.py
+++ b/backend/python/openvoice/backend.py
@@ -0,0 +1,158 @@
+#!/usr/bin/env python3
+"""
+Extra gRPC server for OpenVoice models.
+"""
+from concurrent import futures
+
+import argparse
+import signal
+import sys
+import os
+import torch
+from openvoice import se_extractor
+from openvoice.api import ToneColorConverter
+from melo.api import TTS
+
+import time
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+
+
+_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'))
+
+# 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.
+    """
+    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.
+
+        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.clonedVoice = False
+            # Assume directory from request.ModelFile.
+            # Only if request.LoraAdapter it's not an absolute path
+            if request.AudioPath and request.ModelFile != "" and not os.path.isabs(request.AudioPath):
+                # get base path of modelFile
+                modelFileBase = os.path.dirname(request.ModelFile)
+                request.AudioPath = os.path.join(modelFileBase, request.AudioPath)
+            if request.AudioPath != "":
+                self.clonedVoice = True
+
+            self.modelpath = request.ModelFile
+            self.speaker = request.Type
+            self.ClonedVoicePath = request.AudioPath
+            
+            ckpt_converter = request.Model+'/converter'
+            device = "cuda:0" if torch.cuda.is_available() else "cpu"
+            self.device = device
+            self.tone_color_converter = None
+            if self.clonedVoice:
+                self.tone_color_converter = ToneColorConverter(f'{ckpt_converter}/config.json', device=device)
+                self.tone_color_converter.load_ckpt(f'{ckpt_converter}/checkpoint.pth')
+       
+        except Exception as err:
+            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
+
+        return backend_pb2.Result(message="Model loaded successfully", success=True)
+
+    def TTS(self, request, context):
+        model_name = request.model
+        if model_name == "":
+            return backend_pb2.Result(success=False, message="request.model is required")
+        try:
+            # Speed is adjustable
+            speed = 1.0
+            voice = "EN"
+            if request.voice:
+                voice = request.voice
+            model = TTS(language=voice, device=self.device)
+            speaker_ids = model.hps.data.spk2id
+            speaker_key = self.speaker
+            modelpath = self.modelpath
+            for s in speaker_ids.keys():
+                print(f"Speaker: {s} - ID: {speaker_ids[s]}")
+            speaker_id = speaker_ids[speaker_key]
+            speaker_key = speaker_key.lower().replace('_', '-')
+            source_se = torch.load(f'{modelpath}/base_speakers/ses/{speaker_key}.pth', map_location=self.device)
+            model.tts_to_file(request.text, speaker_id, request.dst, speed=speed)
+            if self.clonedVoice:
+                reference_speaker = self.ClonedVoicePath
+                target_se, audio_name = se_extractor.get_se(reference_speaker, self.tone_color_converter, vad=False)
+                # Run the tone color converter
+                encode_message = "@MyShell"
+                self.tone_color_converter.convert(
+                    audio_src_path=request.dst, 
+                    src_se=source_se, 
+                    tgt_se=target_se, 
+                    output_path=request.dst,
+                    message=encode_message)
+           
+            print("[OpenVoice] TTS generated!", file=sys.stderr)
+            print("[OpenVoice] TTS saved to", request.dst, file=sys.stderr)
+            print(request, file=sys.stderr)
+        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("[OpenVoice] Server started. Listening on: " + address, file=sys.stderr)
+
+    # Define the signal handler function
+    def signal_handler(sig, frame):
+        print("[OpenVoice] Received termination signal. Shutting down...")
+        server.stop(0)
+        sys.exit(0)
+
+    # Set the signal handlers for SIGINT and SIGTERM
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    try:
+        while True:
+            time.sleep(_ONE_DAY_IN_SECONDS)
+    except KeyboardInterrupt:
+        server.stop(0)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the gRPC server.")
+    parser.add_argument(
+        "--addr", default="localhost:50051", help="The address to bind the server to."
+    )
+    args = parser.parse_args()
+    print(f"[OpenVoice] startup: {args}", file=sys.stderr)
+    serve(args.addr)
--- a/backend/python/openvoice/install.sh
+++ b/backend/python/openvoice/install.sh
@@ -0,0 +1,16 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+# This is here because the Intel pip index is broken and returns 200 status codes for every package name, it just doesn't return any package links.
+# This makes uv think that the package exists in the Intel pip index, and by default it stops looking at other pip indexes once it finds a match.
+# We need uv to continue falling through to the pypi default index to find optimum[openvino] in the pypi index
+# the --upgrade actually allows us to *downgrade* torch to the version provided in the Intel pip index
+if [ "x${BUILD_PROFILE}" == "xintel" ]; then
+    EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
+fi
+
+installRequirements
+
+python -m unidic download
--- a/backend/python/openvoice/requirements-cpu.txt
+++ b/backend/python/openvoice/requirements-cpu.txt
@@ -0,0 +1,3 @@
+torch==2.4.1
+git+https://github.com/myshell-ai/MeloTTS.git
+git+https://github.com/myshell-ai/OpenVoice.git
--- a/backend/python/openvoice/requirements-cublas11.txt
+++ b/backend/python/openvoice/requirements-cublas11.txt
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+torch==2.4.1+cu118
+git+https://github.com/myshell-ai/MeloTTS.git
+git+https://github.com/myshell-ai/OpenVoice.git
--- a/backend/python/openvoice/requirements-cublas12.txt
+++ b/backend/python/openvoice/requirements-cublas12.txt
@@ -0,0 +1,3 @@
+torch==2.4.1
+git+https://github.com/myshell-ai/MeloTTS.git
+git+https://github.com/myshell-ai/OpenVoice.git
--- a/backend/python/openvoice/requirements-hipblas.txt
+++ b/backend/python/openvoice/requirements-hipblas.txt
@@ -0,0 +1,4 @@
+--extra-index-url https://download.pytorch.org/whl/rocm6.0
+torch==2.4.1+rocm6.0
+git+https://github.com/myshell-ai/MeloTTS.git
+git+https://github.com/myshell-ai/OpenVoice.git
--- a/backend/python/openvoice/requirements-intel.txt
+++ b/backend/python/openvoice/requirements-intel.txt
@@ -0,0 +1,23 @@
+--extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+intel-extension-for-pytorch
+torch
+optimum[openvino]
+grpcio==1.68.1
+protobuf
+librosa==0.9.1
+faster-whisper==0.9.0
+pydub==0.25.1
+wavmark==0.0.3
+numpy==1.22.0
+eng_to_ipa==0.0.2
+inflect==7.0.0
+unidecode==1.3.7
+whisper-timestamped==1.14.2
+openai
+python-dotenv
+pypinyin==0.50.0
+cn2an==0.5.22
+jieba==0.42.1
+langid==1.1.6
+git+https://github.com/myshell-ai/MeloTTS.git
+git+https://github.com/myshell-ai/OpenVoice.git
--- a/backend/python/openvoice/requirements.txt
+++ b/backend/python/openvoice/requirements.txt
@@ -0,0 +1,20 @@
+grpcio==1.68.1
+protobuf
+librosa
+faster-whisper
+pydub==0.25.1
+wavmark==0.0.3
+numpy==1.22.0
+eng_to_ipa==0.0.2
+inflect
+unidecode
+whisper-timestamped
+openai
+python-dotenv
+pypinyin
+cn2an==0.5.22
+networkx==2.8.8
+jieba==0.42.1
+gradio==3.48.0
+langid==1.1.6
+llvmlite==0.43.0
--- a/backend/python/faster-whisper/run.sh
+++ b/backend/python/faster-whisper/run.sh
--- a/backend/python/openvoice/test.py
+++ b/backend/python/openvoice/test.py
@@ -0,0 +1,82 @@
+"""
+A test script to test the gRPC service
+"""
+import unittest
+import subprocess
+import time
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+
+
+class TestBackendServicer(unittest.TestCase):
+    """
+    TestBackendServicer is the class that tests the gRPC service
+    """
+    def setUp(self):
+        """
+        This method sets up the gRPC service by starting the server
+        """
+        self.service = subprocess.Popen(["python3", "backend.py", "--addr", "localhost:50051"])
+        time.sleep(30)
+
+    def tearDown(self) -> None:
+        """
+        This method tears down the gRPC service by terminating the server
+        """
+        self.service.terminate()
+        self.service.wait()
+
+    def test_server_startup(self):
+        """
+        This method tests if the server starts up successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.Health(backend_pb2.HealthMessage())
+                self.assertEqual(response.message, b'OK')
+        except Exception as err:
+            print(err)
+            self.fail("Server failed to start")
+        finally:
+            self.tearDown()
+
+    def test_load_model(self):
+        """
+        This method tests if the model is loaded successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="checkpoints_v2", 
+                                                                    Type="en-us"))
+                self.assertTrue(response.success)
+                self.assertEqual(response.message, "Model loaded successfully")
+        except Exception as err:
+            print(err)
+            self.fail("LoadModel service failed")
+        finally:
+            self.tearDown()
+
+    def test_tts(self):
+        """
+        This method tests if the embeddings are generated successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="dingzhen"))
+                self.assertTrue(response.success)
+                tts_request = backend_pb2.TTSRequest(text="80s TV news production music hit for tonight's biggest story", voice="EN")
+                tts_response = stub.TTS(tts_request)
+                self.assertIsNotNone(tts_response)
+        except Exception as err:
+            print(err)
+            self.fail("TTS service failed")
+        finally:
+            self.tearDown()
--- a/backend/python/openvoice/test.sh
+++ b/backend/python/openvoice/test.sh
@@ -0,0 +1,12 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+# Download checkpoints if not present
+if [ ! -d "checkpoints_v2" ]; then
+    wget https://myshell-public-repo-host.s3.amazonaws.com/openvoice/checkpoints_v2_0417.zip -O checkpoints_v2.zip
+    unzip checkpoints_v2.zip
+fi
+
+runUnittests
--- a/backend/python/parler-tts/Makefile
+++ b/backend/python/parler-tts/Makefile
@@ -0,0 +1,44 @@
+export CONDA_ENV_PATH = "parler.yml"
+SKIP_CONDA?=0
+ifeq ($(BUILD_TYPE), cublas)
+export CONDA_ENV_PATH = "parler-nvidia.yml"
+endif
+
+# Intel GPU are supposed to have dependencies installed in the main python
+# environment, so we skip conda installation for SYCL builds.
+# https://github.com/intel/intel-extension-for-pytorch/issues/538
+ifneq (,$(findstring sycl,$(BUILD_TYPE)))
+export SKIP_CONDA=1
+endif
+
+.PHONY: parler-tts
+parler-tts:
+	@echo "Installing $(CONDA_ENV_PATH)..."
+	bash install.sh $(CONDA_ENV_PATH)
+	$(MAKE) protogen
+
+.PHONY: run
+run: protogen
+	@echo "Running transformers..."
+	bash run.sh
+	@echo "transformers run."
+
+.PHONY: test
+test: protogen
+	@echo "Testing transformers..."
+	bash test.sh
+	@echo "transformers tested."
+
+.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) -r venv __pycache__
--- a/backend/python/parler-tts/backend.py
+++ b/backend/python/parler-tts/backend.py
@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 """
-Extra gRPC server for Kokoro models.
+Extra gRPC server for MusicgenForConditionalGeneration models.
 """
 from concurrent import futures

@@ -8,17 +8,20 @@ import argparse
 import signal
 import sys
 import os
+
 import time
 import backend_pb2
 import backend_pb2_grpc
-import soundfile as sf
+
 import grpc

-from models import build_model
-from kokoro import generate
+from scipy.io.wavfile import write as write_wav
+
+from parler_tts import ParlerTTSForConditionalGeneration
+from transformers import AutoTokenizer
+import soundfile as sf  
 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
@@ -56,31 +59,10 @@ 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:
-            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}')
+            self.model = ParlerTTSForConditionalGeneration.from_pretrained(model_name).to(device)
+            self.tokenizer = AutoTokenizer.from_pretrained(model_name)
        except Exception as err:
            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")

@@ -88,26 +70,38 @@ 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:
-            audio, out_ps = generate(self.MODEL, request.text, self.VOICEPACK, lang=self.VOICE_NAME)
-            print(out_ps)
-            sf.write(request.dst, audio, SAMPLE_RATE)
+            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)
        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("[Kokoro] Server started. Listening on: " + address, file=sys.stderr)
+    print("[parler-tts] Server started. Listening on: " + address, file=sys.stderr)

    # Define the signal handler function
    def signal_handler(sig, frame):
-        print("[Kokoro] Received termination signal. Shutting down...")
+        print("[parler-tts] Received termination signal. Shutting down...")
        server.stop(0)
        sys.exit(0)

@@ -127,5 +121,5 @@ if __name__ == "__main__":
        "--addr", default="localhost:50051", help="The address to bind the server to."
    )
    args = parser.parse_args()
-    print(f"[Kokoro] startup: {args}", file=sys.stderr)
+    print(f"[parler-tts] startup: {args}", file=sys.stderr)
    serve(args.addr)
--- a/backend/python/parler-tts/install.sh
+++ b/backend/python/parler-tts/install.sh
@@ -0,0 +1,28 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+# This is here because the Intel pip index is broken and returns 200 status codes for every package name, it just doesn't return any package links.
+# This makes uv think that the package exists in the Intel pip index, and by default it stops looking at other pip indexes once it finds a match.
+# We need uv to continue falling through to the pypi default index to find optimum[openvino] in the pypi index
+# the --upgrade actually allows us to *downgrade* torch to the version provided in the Intel pip index
+if [ "x${BUILD_PROFILE}" == "xintel" ]; then
+    EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
+fi
+
+
+installRequirements
+
+
+# https://github.com/descriptinc/audiotools/issues/101
+# incompatible protobuf versions.
+PYDIR=python3.10
+pyenv="${MY_DIR}/venv/lib/${PYDIR}/site-packages/google/protobuf/internal/"
+
+if [ ! -d ${pyenv} ]; then
+    echo "(parler-tts/install.sh): Error: ${pyenv} does not exist"
+    exit 1
+fi
+
+curl -L https://raw.githubusercontent.com/protocolbuffers/protobuf/main/python/google/protobuf/internal/builder.py -o ${pyenv}/builder.py
--- a/backend/python/faster-whisper/protogen.sh
+++ b/backend/python/faster-whisper/protogen.sh
--- a/backend/python/parler-tts/requirements-after.txt
+++ b/backend/python/parler-tts/requirements-after.txt
@@ -0,0 +1,4 @@
+git+https://github.com/huggingface/parler-tts.git@8e465f1b5fcd223478e07175cb40494d19ffbe17
+llvmlite==0.43.0
+numba==0.60.0
+grpcio-tools==1.42.0
--- a/backend/python/parler-tts/requirements-cpu.txt
+++ b/backend/python/parler-tts/requirements-cpu.txt
@@ -0,0 +1,3 @@
+transformers
+accelerate
+torch==2.4.1
--- a/backend/python/parler-tts/requirements-cublas11.txt
+++ b/backend/python/parler-tts/requirements-cublas11.txt
@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+torch==2.4.1+cu118
+torchaudio==2.4.1+cu118
+transformers
+accelerate
--- a/backend/python/parler-tts/requirements-cublas12.txt
+++ b/backend/python/parler-tts/requirements-cublas12.txt
@@ -0,0 +1,4 @@
+torch==2.4.1
+torchaudio==2.4.1
+transformers
+accelerate
--- a/backend/python/parler-tts/requirements-hipblas.txt
+++ b/backend/python/parler-tts/requirements-hipblas.txt
@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/rocm6.0
+torch==2.3.0+rocm6.0
+torchaudio==2.3.0+rocm6.0
+transformers
+accelerate
--- a/backend/python/parler-tts/requirements-intel.txt
+++ b/backend/python/parler-tts/requirements-intel.txt
@@ -0,0 +1,8 @@
+--extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+intel-extension-for-pytorch
+torch
+torchaudio
+optimum[openvino]
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+transformers
+accelerate
--- a/backend/python/parler-tts/requirements.txt
+++ b/backend/python/parler-tts/requirements.txt
@@ -0,0 +1,3 @@
+grpcio==1.68.1
+certifi
+llvmlite==0.43.0
--- a/backend/python/parler-tts/run.sh
+++ b/backend/python/parler-tts/run.sh
--- a/backend/python/parler-tts/test.py
+++ b/backend/python/parler-tts/test.py
@@ -0,0 +1,81 @@
+"""
+A test script to test the gRPC service
+"""
+import unittest
+import subprocess
+import time
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+
+
+class TestBackendServicer(unittest.TestCase):
+    """
+    TestBackendServicer is the class that tests the gRPC service
+    """
+    def setUp(self):
+        """
+        This method sets up the gRPC service by starting the server
+        """
+        self.service = subprocess.Popen(["python3", "backend.py", "--addr", "localhost:50051"])
+        time.sleep(10)
+
+    def tearDown(self) -> None:
+        """
+        This method tears down the gRPC service by terminating the server
+        """
+        self.service.terminate()
+        self.service.wait()
+
+    def test_server_startup(self):
+        """
+        This method tests if the server starts up successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.Health(backend_pb2.HealthMessage())
+                self.assertEqual(response.message, b'OK')
+        except Exception as err:
+            print(err)
+            self.fail("Server failed to start")
+        finally:
+            self.tearDown()
+
+    def test_load_model(self):
+        """
+        This method tests if the model is loaded successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="parler-tts/parler_tts_mini_v0.1"))
+                self.assertTrue(response.success)
+                self.assertEqual(response.message, "Model loaded successfully")
+        except Exception as err:
+            print(err)
+            self.fail("LoadModel service failed")
+        finally:
+            self.tearDown()
+
+    def test_tts(self):
+        """
+        This method tests if the embeddings are generated successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(Model="parler-tts/parler_tts_mini_v0.1"))
+                self.assertTrue(response.success)
+                tts_request = backend_pb2.TTSRequest(text="Hey, how are you doing today?")
+                tts_response = stub.TTS(tts_request)
+                self.assertIsNotNone(tts_response)
+        except Exception as err:
+            print(err)
+            self.fail("TTS service failed")
+        finally:
+            self.tearDown()
--- a/backend/python/parler-tts/test.sh
+++ b/backend/python/parler-tts/test.sh
--- a/backend/python/rerankers/requirements-intel.txt
+++ b/backend/python/rerankers/requirements-intel.txt
@@ -1,9 +1,8 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch==2.3.110+xpu
+intel-extension-for-pytorch
 transformers
 accelerate
-torch==2.3.1+cxx11.abi
-oneccl_bind_pt==2.3.100+xpu
+torch
 rerankers[transformers]
 optimum[openvino]
-setuptools
+setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
--- a/backend/python/rerankers/requirements.txt
+++ b/backend/python/rerankers/requirements.txt
@@ -1,3 +1,3 @@
-grpcio==1.69.0
+grpcio==1.68.1
 protobuf
 certifi
--- a/backend/python/sentencetransformers/Makefile
+++ b/backend/python/sentencetransformers/Makefile
@@ -0,0 +1,31 @@
+.PHONY: sentencetransformers
+sentencetransformers: protogen
+	bash ./install.sh
+
+
+.PHONY: run
+run: protogen
+	@echo "Running sentencetransformers..."
+	bash run.sh
+	@echo "sentencetransformers run."
+
+# It is not working well by using command line. It only6 works with IDE like VSCode.
+.PHONY: test
+test: protogen
+	@echo "Testing sentencetransformers..."
+	bash test.sh
+	@echo "sentencetransformers tested."
+
+.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:
+	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__
--- a/backend/python/sentencetransformers/README.md
+++ b/backend/python/sentencetransformers/README.md
@@ -0,0 +1,5 @@
+# Creating a separate environment for the sentencetransformers project
+
+```
+make sentencetransformers
+```
--- a/backend/python/sentencetransformers/backend.py
+++ b/backend/python/sentencetransformers/backend.py
@@ -1,65 +1,85 @@
 #!/usr/bin/env python3
 """
-This is an extra gRPC server of LocalAI for Bark TTS
+Extra gRPC server for HuggingFace SentenceTransformer models.
 """
 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):
    """
-    BackendServicer is the class that implements the gRPC service
+    A gRPC servicer for the backend service.
+
+    This class implements the gRPC methods for the backend service, including Health, LoadModel, and Embedding.
    """
    def Health(self, request, context):
-        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
-    def LoadModel(self, request, context):
-        device = "cpu"
-        # Get device
-        # device = "cuda" if request.CUDA else "cpu"
-        if request.CUDA:
-            device = "cuda"
+        """
+        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.
+
+        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:
-            print("Preparing models, please wait", file=sys.stderr)
-            self.model = WhisperModel(request.Model, device=device, compute_type="float16")
+            self.model = SentenceTransformer(model_name, trust_remote_code=request.TrustRemoteCode)
        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 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)
+    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)

-        return backend_pb2.TranscriptResult(segments=resultSegments, text=text)

 def serve(address):
    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
--- a/backend/python/sentencetransformers/install.sh
+++ b/backend/python/sentencetransformers/install.sh
--- a/backend/python/sentencetransformers/requirements-cpu.txt
+++ b/backend/python/sentencetransformers/requirements-cpu.txt
@@ -0,0 +1,6 @@
+torch==2.4.1
+accelerate
+transformers
+bitsandbytes
+sentence-transformers==3.3.1
+transformers
--- a/backend/python/sentencetransformers/requirements-cublas11.txt
+++ b/backend/python/sentencetransformers/requirements-cublas11.txt
@@ -1,9 +1,5 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.4.1+cu118
-faster-whisper
-opencv-python
 accelerate
-compel
-peft
-sentencepiece
-optimum-quanto
+sentence-transformers==3.3.1
+transformers
--- a/backend/python/sentencetransformers/requirements-cublas12.txt
+++ b/backend/python/sentencetransformers/requirements-cublas12.txt
@@ -0,0 +1,4 @@
+torch==2.4.1
+accelerate
+sentence-transformers==3.3.1
+transformers
--- a/backend/python/sentencetransformers/requirements-hipblas.txt
+++ b/backend/python/sentencetransformers/requirements-hipblas.txt
@@ -1,3 +1,5 @@
 --extra-index-url https://download.pytorch.org/whl/rocm6.0
 torch==2.4.1+rocm6.0
+accelerate
+sentence-transformers==3.3.1
 transformers
--- a/Show More
+++ b/Show More