Trigger CI

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

.github/workflows/backend.yml

@@ -2,6 +2,7 @@
 name: 'build backend container images'
 
 on:
+  pull_request:
   push:
     branches:
       - master
@@ -63,6 +64,18 @@ jobs:
             backend: "llama-cpp"
             dockerfile: "./backend/Dockerfile.llama-cpp"
             context: "./"
+          - build-type: ''
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-cpu-transformers'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'true'
+            backend: "transformers"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
@@ -87,6 +100,30 @@ jobs:
             backend: "diffusers"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
+          - build-type: 'l4t'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/arm64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-nvidia-l4t-diffusers'
+            runs-on: 'ubuntu-24.04-arm'
+            base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+            skip-drivers: 'true'
+            backend: "diffusers"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: ''
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-cpu-diffusers'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'true'
+            backend: "diffusers"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
           # CUDA 11 additional backends
           - build-type: 'cublas'
             cuda-major-version: "11"
@@ -179,7 +216,7 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-nvidia-cuda-12-vllm'
-            runs-on: 'ubuntu-latest'
+            runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             backend: "vllm"
@@ -278,7 +315,7 @@ jobs:
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-rerankers'
             runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "rerankers"
             dockerfile: "./backend/Dockerfile.python"
@@ -290,7 +327,7 @@ jobs:
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-llama-cpp'
             runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "llama-cpp"
             dockerfile: "./backend/Dockerfile.llama-cpp"
@@ -301,8 +338,8 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-vllm'
-            runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            runs-on: 'arc-runner-set'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "vllm"
             dockerfile: "./backend/Dockerfile.python"
@@ -313,8 +350,8 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-transformers'
-            runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            runs-on: 'arc-runner-set'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "transformers"
             dockerfile: "./backend/Dockerfile.python"
@@ -325,8 +362,8 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-diffusers'
-            runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            runs-on: 'arc-runner-set'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "diffusers"
             dockerfile: "./backend/Dockerfile.python"
@@ -338,8 +375,8 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-kokoro'
-            runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            runs-on: 'arc-runner-set'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "kokoro"
             dockerfile: "./backend/Dockerfile.python"
@@ -351,7 +388,7 @@ jobs:
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-faster-whisper'
             runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "faster-whisper"
             dockerfile: "./backend/Dockerfile.python"
@@ -363,7 +400,7 @@ jobs:
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-coqui'
             runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "coqui"
             dockerfile: "./backend/Dockerfile.python"
@@ -374,31 +411,19 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-rocm-hipblas-bark'
-            runs-on: 'ubuntu-latest'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            runs-on: 'arc-runner-set'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             skip-drivers: 'false'
             backend: "bark"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
           # sycl builds
-          - build-type: 'sycl_f32'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-rerankers'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "rerankers"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f16'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-rerankers'
+            tag-suffix: '-gpu-intel-rerankers'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
@@ -429,60 +454,36 @@ jobs:
             backend: "llama-cpp"
             dockerfile: "./backend/Dockerfile.llama-cpp"
             context: "./"
-          - build-type: 'sycl_f32'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-vllm'
-            runs-on: 'ubuntu-latest'
+            tag-suffix: '-gpu-intel-vllm'
+            runs-on: 'arc-runner-set'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
             backend: "vllm"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-vllm'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "vllm"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f32'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-transformers'
+            tag-suffix: '-gpu-intel-transformers'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
             backend: "transformers"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-transformers'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "transformers"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f32'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-diffusers'
+            tag-suffix: '-gpu-intel-diffusers'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
@@ -490,96 +491,48 @@ jobs:
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
           # SYCL additional backends
-          - build-type: 'sycl_f32'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-kokoro'
+            tag-suffix: '-gpu-intel-kokoro'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
             backend: "kokoro"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-kokoro'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "kokoro"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f32'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-faster-whisper'
+            tag-suffix: '-gpu-intel-faster-whisper'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
             backend: "faster-whisper"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-faster-whisper'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "faster-whisper"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f32'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-coqui'
+            tag-suffix: '-gpu-intel-coqui'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
             backend: "coqui"
             dockerfile: "./backend/Dockerfile.python"
             context: "./backend"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             cuda-major-version: ""
             cuda-minor-version: ""
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-coqui'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "coqui"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f32'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f32-bark'
-            runs-on: 'ubuntu-latest'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            skip-drivers: 'false'
-            backend: "bark"
-            dockerfile: "./backend/Dockerfile.python"
-            context: "./backend"
-          - build-type: 'sycl_f16'
-            cuda-major-version: ""
-            cuda-minor-version: ""
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-gpu-intel-sycl-f16-bark'
+            tag-suffix: '-gpu-intel-bark'
             runs-on: 'ubuntu-latest'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             skip-drivers: 'false'
@@ -824,7 +777,7 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-hipblas-whisper'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             runs-on: 'ubuntu-latest'
             skip-drivers: 'false'
             backend: "whisper"
@@ -868,7 +821,209 @@ jobs:
             skip-drivers: 'false'
             backend: "huggingface"
             dockerfile: "./backend/Dockerfile.golang"
-            context: "./"  
+            context: "./"
+          # rfdetr
+          - build-type: ''
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64,linux/arm64'
+            tag-latest: 'auto'
+            tag-suffix: '-cpu-rfdetr'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "rfdetr"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-nvidia-cuda-12-rfdetr'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "rfdetr"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'cublas'
+            cuda-major-version: "11"
+            cuda-minor-version: "7"
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-nvidia-cuda-11-rfdetr'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "rfdetr"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'intel'
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-intel-rfdetr'
+            runs-on: 'ubuntu-latest'
+            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
+            skip-drivers: 'false'
+            backend: "rfdetr"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/arm64'
+            skip-drivers: 'true'
+            tag-latest: 'auto'
+            tag-suffix: '-nvidia-l4t-arm64-rfdetr'
+            base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"
+            runs-on: 'ubuntu-24.04-arm'
+            backend: "rfdetr"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          # exllama2
+          - build-type: ''
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-cpu-exllama2'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "exllama2"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-nvidia-cuda-12-exllama2'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "exllama2"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'cublas'
+            cuda-major-version: "11"
+            cuda-minor-version: "7"
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-nvidia-cuda-11-exllama2'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "exllama2"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'intel'
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-intel-exllama2'
+            runs-on: 'ubuntu-latest'
+            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
+            skip-drivers: 'false'
+            backend: "exllama2"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          - build-type: 'hipblas'
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64'
+            skip-drivers: 'true'
+            tag-latest: 'auto'
+            tag-suffix: '-gpu-hipblas-exllama2'
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
+            runs-on: 'ubuntu-latest'
+            backend: "exllama2"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+          # runs out of space on the runner
+          # - build-type: 'hipblas'
+          #   cuda-major-version: ""
+          #   cuda-minor-version: ""
+          #   platforms: 'linux/amd64'
+          #   tag-latest: 'auto'
+          #   tag-suffix: '-gpu-hipblas-rfdetr'
+          #   base-image: "rocm/dev-ubuntu-22.04:6.4.3"
+          #   runs-on: 'ubuntu-latest'
+          #   skip-drivers: 'false'
+          #   backend: "rfdetr"
+          #   dockerfile: "./backend/Dockerfile.python"
+          #   context: "./backend"
+          # kitten-tts
+          - build-type: ''
+            cuda-major-version: ""
+            cuda-minor-version: ""
+            platforms: 'linux/amd64,linux/arm64'
+            tag-latest: 'auto'
+            tag-suffix: '-kitten-tts'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
+            backend: "kitten-tts"
+            dockerfile: "./backend/Dockerfile.python"
+            context: "./backend"
+  transformers-darwin:
+    uses: ./.github/workflows/backend_build_darwin.yml
+    with:
+      backend: "transformers"
+      build-type: "mps"
+      go-version: "1.24.x"
+      tag-suffix: "-metal-darwin-arm64-transformers"
+      use-pip: true
+      runs-on: "macOS-14"
+    secrets:
+      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+  diffusers-darwin:
+    uses: ./.github/workflows/backend_build_darwin.yml
+    with:
+      backend: "diffusers"
+      build-type: "mps"
+      go-version: "1.24.x"
+      tag-suffix: "-metal-darwin-arm64-diffusers"
+      use-pip: true
+      runs-on: "macOS-14"
+    secrets:
+      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+  mlx-darwin:
+    uses: ./.github/workflows/backend_build_darwin.yml
+    with:
+      backend: "mlx"
+      build-type: "mps"
+      go-version: "1.24.x"
+      tag-suffix: "-metal-darwin-arm64-mlx"
+      runs-on: "macOS-14"
+    secrets:
+      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+  mlx-vlm-darwin:
+    uses: ./.github/workflows/backend_build_darwin.yml
+    with:
+      backend: "mlx-vlm"
+      build-type: "mps"
+      go-version: "1.24.x"
+      tag-suffix: "-metal-darwin-arm64-mlx-vlm"
+      runs-on: "macOS-14"
+    secrets:
+      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
   llama-cpp-darwin:
     runs-on: macOS-14
     strategy:
@@ -876,7 +1031,7 @@ jobs:
         go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -893,21 +1048,19 @@ jobs:
       - name: Build llama-cpp-darwin
         run: |
           make protogen-go
-          make build
-          bash scripts/build-llama-cpp-darwin.sh
-          ls -la build/darwin.tar
-          mv build/darwin.tar build/llama-cpp.tar
+          make backends/llama-cpp-darwin
       - name: Upload llama-cpp.tar
         uses: actions/upload-artifact@v4
         with:
           name: llama-cpp-tar
-          path: build/llama-cpp.tar
+          path: backend-images/llama-cpp.tar
   llama-cpp-darwin-publish:
     needs: llama-cpp-darwin
+    if: github.event_name != 'pull_request'
     runs-on: ubuntu-latest
     steps:
       - name: Download llama-cpp.tar
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v5
         with:
           name: llama-cpp-tar
           path: .
@@ -964,7 +1117,7 @@ jobs:
         go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -983,20 +1136,19 @@ jobs:
           make protogen-go
           make build
           export PLATFORMARCH=darwin/amd64
-          bash scripts/build-llama-cpp-darwin.sh
-          ls -la build/darwin.tar
-          mv build/darwin.tar build/llama-cpp.tar
+          make backends/llama-cpp-darwin
       - name: Upload llama-cpp.tar
         uses: actions/upload-artifact@v4
         with:
           name: llama-cpp-tar-x86
-          path: build/llama-cpp.tar
+          path: backend-images/llama-cpp.tar
   llama-cpp-darwin-x86-publish:
+    if: github.event_name != 'pull_request'
     needs: llama-cpp-darwin-x86
     runs-on: ubuntu-latest
     steps:
       - name: Download llama-cpp.tar
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v5
         with:
           name: llama-cpp-tar-x86
           path: .
@@ -1045,4 +1197,4 @@ jobs:
         run: |
           for tag in $(echo "${{ steps.quaymeta.outputs.tags }}" | tr ',' '\n'); do
             crane push llama-cpp.tar $tag
-          done
+          done

.github/workflows/backend_build.yml

@@ -55,9 +55,9 @@ on:
         type: string
     secrets:
       dockerUsername:
-        required: true
+        required: false
       dockerPassword:
-        required: true
+        required: false
       quayUsername:
         required: true
       quayPassword:
@@ -66,6 +66,8 @@ on:
 jobs:
   backend-build:
     runs-on: ${{ inputs.runs-on }}
+    env:
+        quay_username: ${{ secrets.quayUsername }}
     steps:
 
 
@@ -95,7 +97,7 @@ jobs:
           && sudo apt-get install -y git
 
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Release space from worker
         if: inputs.runs-on == 'ubuntu-latest'
@@ -187,7 +189,7 @@ jobs:
           password: ${{ secrets.dockerPassword }}
 
       - name: Login to Quay.io
-        # if: github.event_name != 'pull_request'
+        if: ${{ env.quay_username != '' }}
         uses: docker/login-action@v3
         with:
           registry: quay.io
@@ -230,7 +232,7 @@ jobs:
           file: ${{ inputs.dockerfile }}
           cache-from: type=gha
           platforms: ${{ inputs.platforms }}
-          push: true
+          push: ${{ env.quay_username != '' }}
           tags: ${{ steps.meta_pull_request.outputs.tags }}
           labels: ${{ steps.meta_pull_request.outputs.labels }}
 
@@ -238,4 +240,4 @@ jobs:
 
       - name: job summary
         run: |
-          echo "Built image: ${{ steps.meta.outputs.labels }}" >> $GITHUB_STEP_SUMMARY 
+          echo "Built image: ${{ steps.meta.outputs.labels }}" >> $GITHUB_STEP_SUMMARY

.github/workflows/backend_build_darwin.yml

@@ -0,0 +1,140 @@
+---
+name: 'build darwin python backend container images (reusable)'
+
+on:
+  workflow_call:
+    inputs:
+      backend:
+        description: 'Backend to build'
+        required: true
+        type: string
+      build-type:
+        description: 'Build type (e.g., mps)'
+        default: ''
+        type: string
+      use-pip:
+        description: 'Use pip to install dependencies'
+        default: false
+        type: boolean
+      go-version:
+        description: 'Go version to use'
+        default: '1.24.x'
+        type: string
+      tag-suffix:
+        description: 'Tag suffix for the built image'
+        required: true
+        type: string
+      runs-on:
+        description: 'Runner to use'
+        default: 'macOS-14'
+        type: string
+    secrets:
+      dockerUsername:
+        required: false
+      dockerPassword:
+        required: false
+      quayUsername:
+        required: true
+      quayPassword:
+        required: true
+
+jobs:
+  darwin-backend-build:
+    runs-on: ${{ inputs.runs-on }}
+    strategy:
+      matrix:
+        go-version: ['${{ inputs.go-version }}']
+    steps:
+      - name: Clone
+        uses: actions/checkout@v5
+        with:
+          submodules: true
+          
+      - name: Setup Go ${{ matrix.go-version }}
+        uses: actions/setup-go@v5
+        with:
+          go-version: ${{ matrix.go-version }}
+          cache: false
+          
+      # You can test your matrix by printing the current Go version
+      - name: Display Go version
+        run: go version
+        
+      - name: Dependencies
+        run: |
+          brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm
+          
+      - name: Build ${{ inputs.backend }}-darwin
+        run: |
+          make protogen-go
+          BACKEND=${{ inputs.backend }} BUILD_TYPE=${{ inputs.build-type }} USE_PIP=${{ inputs.use-pip }} make build-darwin-python-backend
+          
+      - name: Upload ${{ inputs.backend }}.tar
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ inputs.backend }}-tar
+          path: backend-images/${{ inputs.backend }}.tar
+
+  darwin-backend-publish:
+    needs: darwin-backend-build
+    if: github.event_name != 'pull_request'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Download ${{ inputs.backend }}.tar
+        uses: actions/download-artifact@v5
+        with:
+          name: ${{ inputs.backend }}-tar
+          path: .
+          
+      - name: Install crane
+        run: |
+          curl -L https://github.com/google/go-containerregistry/releases/latest/download/go-containerregistry_Linux_x86_64.tar.gz | tar -xz
+          sudo mv crane /usr/local/bin/
+          
+      - name: Log in to DockerHub
+        run: |
+          echo "${{ secrets.dockerPassword }}" | crane auth login docker.io -u "${{ secrets.dockerUsername }}" --password-stdin
+          
+      - name: Log in to quay.io
+        run: |
+          echo "${{ secrets.quayPassword }}" | crane auth login quay.io -u "${{ secrets.quayUsername }}" --password-stdin
+          
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            localai/localai-backends
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{raw}}
+            type=sha
+          flavor: |
+            latest=auto
+            suffix=${{ inputs.tag-suffix }},onlatest=true
+            
+      - name: Docker meta
+        id: quaymeta
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            quay.io/go-skynet/local-ai-backends
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{raw}}
+            type=sha
+          flavor: |
+            latest=auto
+            suffix=${{ inputs.tag-suffix }},onlatest=true
+            
+      - name: Push Docker image (DockerHub)
+        run: |
+          for tag in $(echo "${{ steps.meta.outputs.tags }}" | tr ',' '\n'); do
+            crane push ${{ inputs.backend }}.tar $tag
+          done
+          
+      - name: Push Docker image (Quay)
+        run: |
+          for tag in $(echo "${{ steps.quaymeta.outputs.tags }}" | tr ',' '\n'); do
+            crane push ${{ inputs.backend }}.tar $tag
+          done

.github/workflows/backend_pr.yml

@@ -0,0 +1,58 @@
+name: 'build backend container images (PR-filtered)'
+
+on:
+  pull_request:
+
+concurrency:
+  group: ci-backends-pr-${{ github.head_ref || github.ref }}-${{ github.repository }}
+  cancel-in-progress: true
+
+jobs:
+  generate-matrix:
+    runs-on: ubuntu-latest
+    outputs:
+      matrix: ${{ steps.set-matrix.outputs.matrix }}
+      has-backends: ${{ steps.set-matrix.outputs.has-backends }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v5
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install dependencies
+        run: |
+          bun add js-yaml
+          bun add @octokit/core
+
+      # filters the matrix in backend.yml
+      - name: Filter matrix for changed backends
+        id: set-matrix
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          GITHUB_EVENT_PATH: ${{ github.event_path }}
+        run: bun run scripts/changed-backends.js
+
+  backend-jobs:
+    needs: generate-matrix
+    uses: ./.github/workflows/backend_build.yml
+    if: needs.generate-matrix.outputs.has-backends == 'true'
+    with:
+      tag-latest: ${{ matrix.tag-latest }}
+      tag-suffix: ${{ matrix.tag-suffix }}
+      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 }}
+      base-image: ${{ matrix.base-image }}
+      backend: ${{ matrix.backend }}
+      dockerfile: ${{ matrix.dockerfile }}
+      skip-drivers: ${{ matrix.skip-drivers }}
+      context: ${{ matrix.context }}
+    secrets:
+      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+    strategy:
+      fail-fast: true
+      matrix: ${{ fromJson(needs.generate-matrix.outputs.matrix) }}

.github/workflows/build-test.yaml

@@ -11,7 +11,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           fetch-depth: 0
       - name: Set up Go

.github/workflows/bump_deps.yaml

@@ -31,7 +31,7 @@ jobs:
             file: "backend/go/piper/Makefile"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Bump dependencies 🔧
         id: bump
         run: |

.github/workflows/bump_docs.yaml

@@ -12,7 +12,7 @@ jobs:
           - repository: "mudler/LocalAI"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Bump dependencies 🔧
         run: |
           bash .github/bump_docs.sh ${{ matrix.repository }}

.github/workflows/checksum_checker.yaml

@@ -15,7 +15,7 @@ jobs:
           && sudo add-apt-repository -y ppa:git-core/ppa \
           && sudo apt-get update \
           && sudo apt-get install -y git
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - name: Install dependencies
         run: |
           sudo apt-get update

.github/workflows/dependabot_auto.yml

@@ -20,7 +20,7 @@ jobs:
           skip-commit-verification: true
 
       - name: Checkout repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Approve a PR if not already approved
         run: |

.github/workflows/deploy-explorer.yaml

@@ -15,7 +15,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - uses: actions/setup-go@v5

.github/workflows/generate_grpc_cache.yaml

@@ -73,7 +73,7 @@ jobs:
         uses: docker/setup-buildx-action@master
 
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Cache GRPC
         uses: docker/build-push-action@v6

.github/workflows/generate_intel_image.yaml

@@ -43,7 +43,7 @@ jobs:
         uses: docker/setup-buildx-action@master
 
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Cache Intel images
         uses: docker/build-push-action@v6

.github/workflows/image-pr.yml

@@ -39,7 +39,7 @@ jobs:
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-gpu-nvidia-cuda12'
+            tag-suffix: '-gpu-nvidia-cuda-12'
             runs-on: 'ubuntu-latest'
             base-image: "ubuntu:22.04"
             makeflags: "--jobs=3 --output-sync=target"
@@ -47,16 +47,16 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'false'
             tag-suffix: '-hipblas'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             grpc-base-image: "ubuntu:22.04"
             runs-on: 'ubuntu-latest'
             makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'sycl_f16'
+          - build-type: 'sycl'
             platforms: 'linux/amd64'
             tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: 'sycl-f16'
+            tag-suffix: 'sycl'
             runs-on: 'ubuntu-latest'
             makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'vulkan'

.github/workflows/image.yml

@@ -39,7 +39,7 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             tag-suffix: '-gpu-hipblas'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.4.3"
             grpc-base-image: "ubuntu:22.04"
             runs-on: 'ubuntu-latest'
             makeflags: "--jobs=3 --output-sync=target"
@@ -83,7 +83,7 @@ jobs:
             cuda-minor-version: "7"
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-nvidia-cuda11'
+            tag-suffix: '-gpu-nvidia-cuda-11'
             runs-on: 'ubuntu-latest'
             base-image: "ubuntu:22.04"
             makeflags: "--jobs=4 --output-sync=target"
@@ -94,7 +94,7 @@ jobs:
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-gpu-nvidia-cuda12'
+            tag-suffix: '-gpu-nvidia-cuda-12'
             runs-on: 'ubuntu-latest'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
@@ -103,30 +103,21 @@ jobs:
           - build-type: 'vulkan'
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-vulkan'
+            tag-suffix: '-gpu-vulkan'
             runs-on: 'ubuntu-latest'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
             aio: "-aio-gpu-vulkan"
-          - build-type: 'sycl_f16'
+          - build-type: 'intel'
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-gpu-intel-f16'
+            tag-suffix: '-gpu-intel'
             runs-on: 'ubuntu-latest'
             makeflags: "--jobs=3 --output-sync=target"
-            aio: "-aio-gpu-intel-f16"
-          - build-type: 'sycl_f32'
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-            grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-gpu-intel-f32'
-            runs-on: 'ubuntu-latest'
-            makeflags: "--jobs=3 --output-sync=target"
-            aio: "-aio-gpu-intel-f32"
+            aio: "-aio-gpu-intel"
 
   gh-runner:
     uses: ./.github/workflows/image_build.yml

.github/workflows/image_build.yml

@@ -94,7 +94,7 @@ jobs:
           && sudo apt-get update \
           && sudo apt-get install -y git
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Release space from worker
         if: inputs.runs-on == 'ubuntu-latest'

.github/workflows/localaibot_automerge.yml

@@ -13,7 +13,7 @@ jobs:
     if: ${{ github.actor == 'localai-bot' }}
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
 
       - name: Approve a PR if not already approved
         run: |

.github/workflows/notify-models.yaml

@@ -11,7 +11,7 @@ jobs:
         MODEL_NAME: gemma-3-12b-it
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v5
       with:
         fetch-depth: 0 # needed to checkout all branches for this Action to work
     - uses: mudler/localai-github-action@v1
@@ -90,7 +90,7 @@ jobs:
         MODEL_NAME: gemma-3-12b-it
     runs-on: ubuntu-latest
     steps:
-    - uses: actions/checkout@v4
+    - uses: actions/checkout@v5
       with:
         fetch-depth: 0 # needed to checkout all branches for this Action to work
     - name: Start LocalAI

.github/workflows/release.yaml

@@ -10,7 +10,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           fetch-depth: 0
       - name: Set up Go

.github/workflows/secscan.yaml

@@ -14,11 +14,11 @@ jobs:
       GO111MODULE: on
     steps:
       - name: Checkout Source
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         if: ${{ github.actor != 'dependabot[bot]' }}
       - name: Run Gosec Security Scanner
         if: ${{ github.actor != 'dependabot[bot]' }}
-        uses: securego/gosec@v2.22.7
+        uses: securego/gosec@v2.22.8
         with:
           # we let the report trigger content trigger a failure using the GitHub Security features.
           args: '-no-fail -fmt sarif -out results.sarif ./...'

.github/workflows/test-extra.yml

@@ -19,7 +19,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v5
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -40,7 +40,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Dependencies
@@ -61,7 +61,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Dependencies
@@ -83,7 +83,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Dependencies
@@ -104,7 +104,7 @@ jobs:
   #  runs-on: ubuntu-latest
   #  steps:
   #    - name: Clone
-  #      uses: actions/checkout@v4
+  #      uses: actions/checkout@v5
   #      with:
   #        submodules: true
   #    - name: Dependencies
@@ -124,7 +124,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v5
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -186,7 +186,7 @@ jobs:
   #           sudo rm -rf "$AGENT_TOOLSDIRECTORY" || true
   #           df -h
   #     - name: Clone
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v5
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -211,7 +211,7 @@ jobs:
   #   runs-on: ubuntu-latest
   #   steps:
   #     - name: Clone
-  #       uses: actions/checkout@v4
+  #       uses: actions/checkout@v5
   #       with:
   #         submodules: true
   #     - name: Dependencies
@@ -232,7 +232,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Dependencies

.github/workflows/test.yml

@@ -23,6 +23,20 @@ jobs:
       matrix:
         go-version: ['1.21.x']
     steps:
+      - name: Free Disk Space (Ubuntu)
+        uses: jlumbroso/free-disk-space@main
+        with:
+          # this might remove tools that are actually needed,
+          # if set to "true" but frees about 6 GB
+          tool-cache: true
+          # all of these default to true, but feel free to set to
+          # "false" if necessary for your workflow
+          android: true
+          dotnet: true
+          haskell: true
+          large-packages: true
+          docker-images: true
+          swap-storage: true
       - name: Release space from worker
         run: |
           echo "Listing top largest packages"
@@ -56,7 +70,7 @@ jobs:
           sudo rm -rfv build || true
           df -h
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -152,7 +166,7 @@ jobs:
           sudo rm -rfv build || true
           df -h
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Dependencies
@@ -182,7 +196,7 @@ jobs:
         go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v4
+        uses: actions/checkout@v5
         with:
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -200,11 +214,7 @@ jobs:
       - name: Build llama-cpp-darwin
         run: |
           make protogen-go
-          make build
-          bash scripts/build-llama-cpp-darwin.sh
-          ls -la build/darwin.tar
-          mv build/darwin.tar build/llama-cpp.tar
-          ./local-ai backends install "ocifile://$PWD/build/llama-cpp.tar"
+          make backends/llama-cpp-darwin
       - name: Test
         run: |
           export C_INCLUDE_PATH=/usr/local/include

.github/workflows/update_swagger.yaml

@@ -9,7 +9,7 @@ jobs:
       fail-fast: false
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v5
       - uses: actions/setup-go@v5
         with:
           go-version: 'stable'

Dockerfile

@@ -9,7 +9,7 @@ ENV DEBIAN_FRONTEND=noninteractive
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
         ca-certificates curl wget espeak-ng libgomp1 \
-        python3 python-is-python3 ffmpeg && \
+        ffmpeg libopenblas-base libopenblas-dev && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
 
@@ -72,6 +72,12 @@ RUN <<EOT bash
     fi
 EOT
 
+RUN <<EOT bash
+    if [ "${BUILD_TYPE}" = "cublas" ] && [ "${TARGETARCH}" = "arm64" ]; then
+        echo "nvidia-l4t" > /run/localai/capability
+    fi
+EOT
+
 # If we are building with clblas support, we need the libraries for the builds
 RUN if [ "${BUILD_TYPE}" = "clblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
         apt-get update && \
@@ -94,6 +100,8 @@ RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         ldconfig \
     ; fi
 
+RUN expr "${BUILD_TYPE}" = intel && echo "intel" > /run/localai/capability || echo "not intel"
+
 # Cuda
 ENV PATH=/usr/local/cuda/bin:${PATH}
 

Makefile

@@ -5,8 +5,6 @@ BINARY_NAME=local-ai
 
 GORELEASER?=
 
-ONEAPI_VERSION?=2025.2
-
 export BUILD_TYPE?=
 
 GO_TAGS?=
@@ -134,27 +132,6 @@ test: test-models/testmodel.ggml protogen-go
 	$(MAKE) test-tts
 	$(MAKE) test-stablediffusion
 
-backends/llama-cpp: docker-build-llama-cpp docker-save-llama-cpp build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/llama-cpp.tar)"
-
-backends/piper: docker-build-piper docker-save-piper build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/piper.tar)"
-
-backends/stablediffusion-ggml: docker-build-stablediffusion-ggml docker-save-stablediffusion-ggml build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/stablediffusion-ggml.tar)"
-
-backends/whisper: docker-build-whisper docker-save-whisper build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/whisper.tar)"
-
-backends/silero-vad: docker-build-silero-vad docker-save-silero-vad build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/silero-vad.tar)"
-
-backends/local-store: docker-build-local-store docker-save-local-store build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/local-store.tar)"
-
-backends/huggingface: docker-build-huggingface docker-save-huggingface build
-	./local-ai backends install "ocifile://$(abspath ./backend-images/huggingface.tar)"
-
 ########################################################
 ## AIO tests
 ########################################################
@@ -322,7 +299,7 @@ docker-cuda11:
 		--build-arg GO_TAGS="$(GO_TAGS)" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
 		--build-arg BUILD_TYPE=$(BUILD_TYPE) \
-		-t $(DOCKER_IMAGE)-cuda11 .
+		-t $(DOCKER_IMAGE)-cuda-11 .
 
 docker-aio:
 	@echo "Building AIO image with base $(BASE_IMAGE) as $(DOCKER_AIO_IMAGE)"
@@ -337,24 +314,69 @@ docker-aio-all:
 
 docker-image-intel:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:${ONEAPI_VERSION}.0-0-devel-ubuntu24.04 \
+		--build-arg BASE_IMAGE=quay.io/go-skynet/intel-oneapi-base:latest \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="$(GO_TAGS)" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
-		--build-arg BUILD_TYPE=sycl_f32 -t $(DOCKER_IMAGE) .
-
-docker-image-intel-xpu:
-	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:${ONEAPI_VERSION}.0-0-devel-ubuntu22.04 \
-		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
-		--build-arg GO_TAGS="$(GO_TAGS)" \
-		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
-		--build-arg BUILD_TYPE=sycl_f32 -t $(DOCKER_IMAGE) .
+		--build-arg BUILD_TYPE=intel -t $(DOCKER_IMAGE) .
 
 ########################################################
 ## Backends
 ########################################################
 
+
+backends/diffusers: docker-build-diffusers docker-save-diffusers build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/diffusers.tar)"
+
+backends/llama-cpp: docker-build-llama-cpp docker-save-llama-cpp build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/llama-cpp.tar)"
+
+backends/piper: docker-build-piper docker-save-piper build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/piper.tar)"
+
+backends/stablediffusion-ggml: docker-build-stablediffusion-ggml docker-save-stablediffusion-ggml build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/stablediffusion-ggml.tar)"
+
+backends/whisper: docker-build-whisper docker-save-whisper build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/whisper.tar)"
+
+backends/silero-vad: docker-build-silero-vad docker-save-silero-vad build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/silero-vad.tar)"
+
+backends/local-store: docker-build-local-store docker-save-local-store build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/local-store.tar)"
+
+backends/huggingface: docker-build-huggingface docker-save-huggingface build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/huggingface.tar)"
+
+backends/rfdetr: docker-build-rfdetr docker-save-rfdetr build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/rfdetr.tar)"
+
+backends/kitten-tts: docker-build-kitten-tts docker-save-kitten-tts build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/kitten-tts.tar)"
+
+backends/kokoro: docker-build-kokoro docker-save-kokoro build
+	./local-ai backends install "ocifile://$(abspath ./backend-images/kokoro.tar)"
+
+backends/llama-cpp-darwin: build
+	bash ./scripts/build/llama-cpp-darwin.sh
+	./local-ai backends install "ocifile://$(abspath ./backend-images/llama-cpp.tar)"
+
+build-darwin-python-backend: build
+	bash ./scripts/build/python-darwin.sh
+
+backends/mlx:
+	BACKEND=mlx $(MAKE) build-darwin-python-backend
+	./local-ai backends install "ocifile://$(abspath ./backend-images/mlx.tar)"
+
+backends/diffuser-darwin:
+	BACKEND=diffusers $(MAKE) build-darwin-python-backend
+	./local-ai backends install "ocifile://$(abspath ./backend-images/diffusers.tar)"
+
+backends/mlx-vlm:
+	BACKEND=mlx-vlm $(MAKE) build-darwin-python-backend
+	./local-ai backends install "ocifile://$(abspath ./backend-images/mlx-vlm.tar)"
+
 backend-images:
 	mkdir -p backend-images
 
@@ -373,6 +395,24 @@ docker-build-local-store:
 docker-build-huggingface:
 	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:huggingface -f backend/Dockerfile.golang --build-arg BACKEND=huggingface .
 
+docker-build-rfdetr:
+	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:rfdetr -f backend/Dockerfile.python --build-arg BACKEND=rfdetr ./backend
+
+docker-build-kitten-tts:
+	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:kitten-tts -f backend/Dockerfile.python --build-arg BACKEND=kitten-tts ./backend
+
+docker-save-kitten-tts: backend-images
+	docker save local-ai-backend:kitten-tts -o backend-images/kitten-tts.tar
+
+docker-build-kokoro:
+	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:kokoro -f backend/Dockerfile.python --build-arg BACKEND=kokoro ./backend
+
+docker-save-kokoro: backend-images
+	docker save local-ai-backend:kokoro -o backend-images/kokoro.tar
+
+docker-save-rfdetr: backend-images
+	docker save local-ai-backend:rfdetr -o backend-images/rfdetr.tar
+
 docker-save-huggingface: backend-images
 	docker save local-ai-backend:huggingface -o backend-images/huggingface.tar
 
@@ -410,10 +450,10 @@ docker-build-transformers:
 	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:transformers -f backend/Dockerfile.python --build-arg BACKEND=transformers .
 
 docker-build-diffusers:
-	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:diffusers -f backend/Dockerfile.python --build-arg BACKEND=diffusers .
+	docker build --progress=plain --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:diffusers -f backend/Dockerfile.python --build-arg BACKEND=diffusers ./backend
 
-docker-build-kokoro:
-	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:kokoro -f backend/Dockerfile.python --build-arg BACKEND=kokoro .
+docker-save-diffusers: backend-images
+	docker save local-ai-backend:diffusers -o backend-images/diffusers.tar
 
 docker-build-whisper:
 	docker build --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg BASE_IMAGE=$(BASE_IMAGE) -t local-ai-backend:whisper -f backend/Dockerfile.golang --build-arg BACKEND=whisper  .

README.md

@@ -140,11 +140,7 @@ docker run -ti --name local-ai -p 8080:8080 --device=/dev/kfd --device=/dev/dri
 ### Intel GPU Images (oneAPI):
 
 ```bash
-# Intel GPU with FP16 support
-docker run -ti --name local-ai -p 8080:8080 --device=/dev/dri/card1 --device=/dev/dri/renderD128 localai/localai:latest-gpu-intel-f16
-
-# Intel GPU with FP32 support
-docker run -ti --name local-ai -p 8080:8080 --device=/dev/dri/card1 --device=/dev/dri/renderD128 localai/localai:latest-gpu-intel-f32
+docker run -ti --name local-ai -p 8080:8080 --device=/dev/dri/card1 --device=/dev/dri/renderD128 localai/localai:latest-gpu-intel
 ```
 
 ### Vulkan GPU Images:
@@ -166,7 +162,7 @@ docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-ai
 docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-aio-gpu-nvidia-cuda-11
 
 # Intel GPU version
-docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-gpu-intel-f16
+docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-gpu-intel
 
 # AMD GPU version
 docker run -ti --name local-ai -p 8080:8080 --device=/dev/kfd --device=/dev/dri --group-add=video localai/localai:latest-aio-gpu-hipblas
@@ -189,10 +185,14 @@ local-ai run https://gist.githubusercontent.com/.../phi-2.yaml
 local-ai run oci://localai/phi-2:latest
 ```
 
+> ⚡ **Automatic Backend Detection**: When you install models from the gallery or YAML files, LocalAI automatically detects your system's GPU capabilities (NVIDIA, AMD, Intel) and downloads the appropriate backend. For advanced configuration options, see [GPU Acceleration](https://localai.io/features/gpu-acceleration/#automatic-backend-detection).
+
 For more information, see [💻 Getting started](https://localai.io/basics/getting_started/index.html)
 
 ## 📰 Latest project news
 
+- August 2025: MLX, MLX-VLM, Diffusers and llama.cpp are now supported on Mac M1/M2/M3+ chips ( with `development` suffix in the gallery ): https://github.com/mudler/LocalAI/pull/6049 https://github.com/mudler/LocalAI/pull/6119 https://github.com/mudler/LocalAI/pull/6121 https://github.com/mudler/LocalAI/pull/6060
+- July/August 2025: 🔍 [Object Detection](https://localai.io/features/object-detection/) added to the API featuring [rf-detr](https://github.com/roboflow/rf-detr)
 - July 2025: All backends migrated outside of the main binary. LocalAI is now more lightweight, small, and automatically downloads the required backend to run the model. [Read the release notes](https://github.com/mudler/LocalAI/releases/tag/v3.2.0)
 - June 2025: [Backend management](https://github.com/mudler/LocalAI/pull/5607) has been added. Attention: extras images are going to be deprecated from the next release! Read [the backend management PR](https://github.com/mudler/LocalAI/pull/5607).
 - May 2025: [Audio input](https://github.com/mudler/LocalAI/pull/5466) and [Reranking](https://github.com/mudler/LocalAI/pull/5396) in llama.cpp backend, [Realtime API](https://github.com/mudler/LocalAI/pull/5392),  Support to Gemma, SmollVLM, and more multimodal models (available in the gallery).
@@ -226,6 +226,7 @@ Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3A
 - ✍️ [Constrained grammars](https://localai.io/features/constrained_grammars/)
 - 🖼️ [Download Models directly from Huggingface ](https://localai.io/models/)
 - 🥽 [Vision API](https://localai.io/features/gpt-vision/)
+- 🔍 [Object Detection](https://localai.io/features/object-detection/)
 - 📈 [Reranker API](https://localai.io/features/reranker/)
 - 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
 - [Agentic capabilities](https://github.com/mudler/LocalAGI)

backend/Dockerfile.golang

@@ -96,17 +96,6 @@ RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         ldconfig \
     ; fi
 
-# Intel oneAPI requirements
-RUN <<EOT bash
-    if [[ "${BUILD_TYPE}" == sycl* ]] && [ "${SKIP_DRIVERS}" = "false" ]; then
-        apt-get update && \
-        apt-get install -y --no-install-recommends \
-            intel-oneapi-runtime-libs && \
-        apt-get clean && \
-        rm -rf /var/lib/apt/lists/*
-    fi
-EOT
-
 # Install Go
 RUN curl -L -s https://go.dev/dl/go${GO_VERSION}.linux-${TARGETARCH}.tar.gz | tar -C /usr/local -xz
 ENV PATH=$PATH:/root/go/bin:/usr/local/go/bin:/usr/local/bin

backend/Dockerfile.python

@@ -23,7 +23,7 @@ RUN apt-get update && \
         libssl-dev \
         git \
         git-lfs \
-        unzip \
+        unzip clang \
         upx-ucl \
         curl python3-pip \
         python-is-python3 \
@@ -116,7 +116,7 @@ COPY python/${BACKEND} /${BACKEND}
 COPY backend.proto /${BACKEND}/backend.proto
 COPY python/common/ /${BACKEND}/common
 
-RUN cd /${BACKEND} && make
+RUN cd /${BACKEND} && PORTABLE_PYTHON=true make
 
 FROM scratch
 ARG BACKEND=rerankers

backend/backend.proto

@@ -20,6 +20,7 @@ service Backend {
   rpc SoundGeneration(SoundGenerationRequest) returns (Result) {}
   rpc TokenizeString(PredictOptions) returns (TokenizationResponse) {}
   rpc Status(HealthMessage) returns (StatusResponse) {}
+  rpc Detect(DetectOptions) returns (DetectResponse) {}
 
   rpc StoresSet(StoresSetOptions) returns (Result) {}
   rpc StoresDelete(StoresDeleteOptions) returns (Result) {}
@@ -304,6 +305,9 @@ message GenerateImageRequest {
   // Diffusers
   string EnableParameters = 10;
   int32 CLIPSkip = 11;
+  
+  // Reference images for models that support them (e.g., Flux Kontext)
+  repeated string ref_images = 12;
 }
 
 message GenerateVideoRequest {
@@ -376,3 +380,20 @@ message Message {
   string role = 1;
   string content = 2;
 }
+
+message DetectOptions {
+  string src = 1;
+}
+
+message Detection {
+  float x = 1;
+  float y = 2;
+  float width = 3;
+  float height = 4;
+  float confidence = 5;
+  string class_name = 6;
+}
+
+message DetectResponse {
+  repeated Detection Detections = 1;
+}

backend/cpp/llama-cpp/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=3f4fc97f1d745f1d5d3c853949503136d419e6de
+LLAMA_VERSION?=710dfc465a68f7443b87d9f792cffba00ed739fe
 LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
 
 CMAKE_ARGS?=
@@ -26,16 +26,14 @@ else ifeq ($(BUILD_TYPE),openblas)
 # If build type is clblas (openCL) we set -DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
 else ifeq ($(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++ 
+# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++
 else ifeq ($(BUILD_TYPE),hipblas)
 	ROCM_HOME ?= /opt/rocm
 	ROCM_PATH ?= /opt/rocm
 	export CXX=$(ROCM_HOME)/llvm/bin/clang++
 	export CC=$(ROCM_HOME)/llvm/bin/clang
-#	GPU_TARGETS ?= gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102
-#	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
-	CMAKE_ARGS+=-DGGML_HIP=ON
-#	CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
+	AMDGPU_TARGETS?=gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102,gfx1200,gfx1201
+	CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS=$(AMDGPU_TARGETS)
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=1
 else ifeq ($(OS),Darwin)

backend/cpp/llama-cpp/grpc-server.cpp

@@ -53,9 +53,9 @@ static void start_llama_server(server_context& ctx_server) {
     LOG_INF("%s: model loaded\n", __func__);
 
     // print sample chat example to make it clear which template is used
-    LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
-        common_chat_templates_source(ctx_server.chat_templates.get()),
-        common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str());
+    // LOG_INF("%s: chat template, chat_template: %s, example_format: '%s'\n", __func__,
+    //     common_chat_templates_source(ctx_server.chat_templates.get()),
+    //     common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params_base.use_jinja).c_str(), ctx_server.params_base.default_template_kwargs);
 
     // Reset the chat templates
     // TODO: We should make this configurable by respecting the option that is already present in LocalAI for vLLM
@@ -313,9 +313,11 @@ static void params_parse(const backend::ModelOptions* request,
         params.pooling_type = LLAMA_POOLING_TYPE_RANK;
     }
 
+    
     if (request->ropescaling() == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
     else if (request->ropescaling() == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
-    else { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
+    else if (request->ropescaling() == "linear")   {  params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
+
     if ( request->yarnextfactor() != 0.0f ) {
         params.yarn_ext_factor = request->yarnextfactor();
     }
@@ -435,24 +437,7 @@ public:
                 }
             }
 
-            // process files
-            mtmd::bitmaps bitmaps;
             const bool has_mtmd = ctx_server.mctx != nullptr;
-            {
-                if (!has_mtmd && !files.empty()) {
-                    throw std::runtime_error("This server does not support multimodal");
-                }
-                for (auto & file : files) {
-                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(ctx_server.mctx, file.data(), file.size()));
-                    if (!bmp.ptr) {
-                        throw std::runtime_error("Failed to load image/audio");
-                    }
-                    // calculate bitmap hash (for KV caching)
-                    std::string hash = fnv_hash(bmp.data(), bmp.n_bytes());
-                    bmp.set_id(hash.c_str());
-                    bitmaps.entries.push_back(std::move(bmp));
-                }
-            }
 
             // process prompt
             std::vector<server_tokens> inputs;
@@ -462,32 +447,10 @@ public:
 
             if (has_mtmd) {
                 // multimodal
-                std::string prompt_str = prompt.get<std::string>();
-                mtmd_input_text inp_txt = {
-                    prompt_str.c_str(),
-                    /* add_special */   true,
-                    /* parse_special */ true,
-                };
-                mtmd::input_chunks chunks(mtmd_input_chunks_init());
-                auto bitmaps_c_ptr = bitmaps.c_ptr();
-                int32_t tokenized = mtmd_tokenize(ctx_server.mctx,
-                                                    chunks.ptr.get(),
-                                                    &inp_txt,
-                                                    bitmaps_c_ptr.data(),
-                                                    bitmaps_c_ptr.size());
-                if (tokenized != 0) {
-                    throw std::runtime_error("Failed to tokenize prompt");
-                }
-
-                server_tokens tmp(chunks, true);
-                inputs.push_back(std::move(tmp));
+                inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files));
             } else {
-                // non-multimodal version
-                auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
-                for (auto & p : tokenized_prompts) {
-                    auto tmp = server_tokens(p, ctx_server.mctx != nullptr);
-                    inputs.push_back(std::move(tmp));
-                }
+                 // Everything else, including multimodal completions.
+                inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
             }
 
             tasks.reserve(inputs.size());
@@ -628,23 +591,7 @@ public:
             }
 
             // process files
-            mtmd::bitmaps bitmaps;
             const bool has_mtmd = ctx_server.mctx != nullptr;
-            {
-                if (!has_mtmd && !files.empty()) {
-                    throw std::runtime_error("This server does not support multimodal");
-                }
-                for (auto & file : files) {
-                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(ctx_server.mctx, file.data(), file.size()));
-                    if (!bmp.ptr) {
-                        throw std::runtime_error("Failed to load image/audio");
-                    }
-                    // calculate bitmap hash (for KV caching)
-                    std::string hash = fnv_hash(bmp.data(), bmp.n_bytes());
-                    bmp.set_id(hash.c_str());
-                    bitmaps.entries.push_back(std::move(bmp));
-                }
-            }
 
             // process prompt
             std::vector<server_tokens> inputs;
@@ -655,33 +602,10 @@ public:
 
             if (has_mtmd) {
                 // multimodal
-                std::string prompt_str = prompt.get<std::string>();
-                mtmd_input_text inp_txt = {
-                    prompt_str.c_str(),
-                    /* add_special */   true,
-                    /* parse_special */ true,
-                };
-                mtmd::input_chunks chunks(mtmd_input_chunks_init());
-                auto bitmaps_c_ptr = bitmaps.c_ptr();
-                int32_t tokenized = mtmd_tokenize(ctx_server.mctx,
-                                                    chunks.ptr.get(),
-                                                    &inp_txt,
-                                                    bitmaps_c_ptr.data(),
-                                                    bitmaps_c_ptr.size());
-                if (tokenized != 0) {
-                    std::cout << "[PREDICT] Failed to tokenize prompt" << std::endl;
-                    throw std::runtime_error("Failed to tokenize prompt");
-                }
-
-                server_tokens tmp(chunks, true);
-                inputs.push_back(std::move(tmp));
+                inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files));
             } else {
-                // non-multimodal version
-                auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
-                for (auto & p : tokenized_prompts) {
-                    auto tmp = server_tokens(p, ctx_server.mctx != nullptr);
-                    inputs.push_back(std::move(tmp));
-                }
+                 // Everything else, including multimodal completions.
+                inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
             }
 
             tasks.reserve(inputs.size());
@@ -772,7 +696,7 @@ public:
         json prompt = body.at("prompt");
 
 
-        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
+        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
             if (tokens.empty()) {
@@ -791,7 +715,7 @@ public:
 
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = server_tokens(tokenized_prompts[i], ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tokenized_prompts[i]);
 
                 // OAI-compat
                 task.params.oaicompat = OAICOMPAT_TYPE_EMBEDDING;
@@ -847,8 +771,10 @@ public:
         }
 
         // Tokenize the query
-        llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.vocab, request->query(), /* add_special */ false, true)[0];
-
+        auto tokenized_query = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, request->query(), /* add_special */ false, true);
+        if (tokenized_query.size() != 1) {
+            return grpc::Status(grpc::StatusCode::INVALID_ARGUMENT, "\"query\" must contain only a single prompt");
+        }
         // Create and queue the task
         json responses = json::array();
         bool error = false;
@@ -860,14 +786,14 @@ public:
                 documents.push_back(request->documents(i));
             }
             
-            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
+            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, documents, /* add_special */ false, true);
             tasks.reserve(tokenized_docs.size());
             for (size_t i = 0; i < tokenized_docs.size(); i++) {
-                auto tmp = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
+                auto tmp = format_rerank(ctx_server.vocab, tokenized_query[0], tokenized_docs[i]);
                 server_task task = server_task(SERVER_TASK_TYPE_RERANK);
                 task.id = ctx_server.queue_tasks.get_new_id();
                 task.index = i;
-                task.prompt_tokens = server_tokens(tmp, ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tmp);
                 tasks.push_back(std::move(task));
             }
 

backend/cpp/llama-cpp/run.sh

@@ -42,7 +42,8 @@ fi
  
 # Extend ld library path with the dir where this script is located/lib
 if [ "$(uname)" == "Darwin" ]; then
-	DYLD_FALLBACK_LIBRARY_PATH=$CURDIR/lib:$DYLD_FALLBACK_LIBRARY_PATH
+	export DYLD_LIBRARY_PATH=$CURDIR/lib:$DYLD_LIBRARY_PATH
+	#export DYLD_FALLBACK_LIBRARY_PATH=$CURDIR/lib:$DYLD_FALLBACK_LIBRARY_PATH
 else
 	export LD_LIBRARY_PATH=$CURDIR/lib:$LD_LIBRARY_PATH
 fi
@@ -57,5 +58,5 @@ fi
 echo "Using binary: $BINARY"
 exec $CURDIR/$BINARY "$@"
 
-# In case we fail execing, just run fallback
+# We should never reach this point, however just in case we do, run fallback
 exec $CURDIR/llama-cpp-fallback "$@"

backend/go/stablediffusion-ggml/CMakeLists.txt

@@ -0,0 +1,16 @@
+cmake_minimum_required(VERSION 3.12)
+project(gosd LANGUAGES C CXX)
+set(CMAKE_POSITION_INDEPENDENT_CODE ON)
+
+add_subdirectory(./sources/stablediffusion-ggml.cpp)
+
+add_library(gosd MODULE gosd.cpp)
+target_link_libraries(gosd PRIVATE stable-diffusion ggml stdc++fs)
+
+target_include_directories(gosd PUBLIC
+    stable-diffusion.cpp
+    stable-diffusion.cpp/thirdparty
+)
+
+set_property(TARGET gosd PROPERTY CXX_STANDARD 17)
+set_target_properties(gosd PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})

backend/go/stablediffusion-ggml/Makefile

@@ -1,28 +1,16 @@
-INCLUDE_PATH := $(abspath ./)
-LIBRARY_PATH := $(abspath ./)
-
-AR?=ar
 CMAKE_ARGS?=
 BUILD_TYPE?=
 NATIVE?=false
-CUDA_LIBPATH?=/usr/local/cuda/lib64/
-ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
-# keep standard at C11 and C++11
-CXXFLAGS = -I. -I$(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
 
 GOCMD?=go
-CGO_LDFLAGS?=
-# Avoid parent make file overwriting CGO_LDFLAGS which is needed for hipblas
-CGO_LDFLAGS_SYCL=
 GO_TAGS?=
-LD_FLAGS?=
+JOBS?=$(shell nproc --ignore=1)
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=eed97a5e1d054f9c1e7ac01982ae480411d4157e
+STABLEDIFFUSION_GGML_VERSION?=5900ef6605c6fbf7934239f795c13c97bc993853
 
-# Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
-CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
+CMAKE_ARGS+=-DGGML_MAX_NAME=128
 
 ifeq ($(NATIVE),false)
 	CMAKE_ARGS+=-DGGML_NATIVE=OFF
@@ -31,7 +19,6 @@ endif
 # If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
 	CMAKE_ARGS+=-DSD_CUDA=ON -DGGML_CUDA=ON
-	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH) -L$(CUDA_LIBPATH)/stubs/ -lcuda
 # If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
 # to CMAKE_ARGS automatically
 else ifeq ($(BUILD_TYPE),openblas)
@@ -46,14 +33,12 @@ else ifeq ($(BUILD_TYPE),hipblas)
 # But if it's OSX without metal, disable it here
 else ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DSD_VULKAN=ON -DGGML_VULKAN=ON
-	CGO_LDFLAGS+=-lvulkan
 else ifeq ($(OS),Darwin)
 	ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DSD_METAL=OFF -DGGML_METAL=OFF
 	else
 		CMAKE_ARGS+=-DSD_METAL=ON -DGGML_METAL=ON
 		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
-		TARGET+=--target ggml-metal
 	endif
 endif
 
@@ -63,12 +48,6 @@ ifeq ($(BUILD_TYPE),sycl_f16)
 		-DCMAKE_CXX_COMPILER=icpx \
 		-DSD_SYCL=ON \
 		-DGGML_SYCL_F16=ON
-	export CC=icx
-	export CXX=icpx
-	CGO_LDFLAGS_SYCL += -fsycl -L${DNNLROOT}/lib -ldnnl ${MKLROOT}/lib/intel64/libmkl_sycl.a -fiopenmp -fopenmp-targets=spir64 -lOpenCL
-	CGO_LDFLAGS_SYCL += $(shell pkg-config --libs mkl-static-lp64-gomp)
-	CGO_CXXFLAGS += -fiopenmp -fopenmp-targets=spir64
-	CGO_CXXFLAGS += $(shell pkg-config --cflags mkl-static-lp64-gomp )
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f32)
@@ -76,73 +55,24 @@ ifeq ($(BUILD_TYPE),sycl_f32)
 		-DCMAKE_C_COMPILER=icx \
 		-DCMAKE_CXX_COMPILER=icpx \
 		-DSD_SYCL=ON
-	export CC=icx
-	export CXX=icpx
-	CGO_LDFLAGS_SYCL += -fsycl -L${DNNLROOT}/lib -ldnnl ${MKLROOT}/lib/intel64/libmkl_sycl.a -fiopenmp -fopenmp-targets=spir64 -lOpenCL
-	CGO_LDFLAGS_SYCL += $(shell pkg-config --libs mkl-static-lp64-gomp)
-	CGO_CXXFLAGS += -fiopenmp -fopenmp-targets=spir64
-	CGO_CXXFLAGS += $(shell pkg-config --cflags mkl-static-lp64-gomp )
 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')
-ALL_OBJS := $(shell find $(GGML_ARCHIVE_DIR) -type f -name '*.o')
-
-# Name of the single merged library
-COMBINED_LIB := libggmlall.a
-
-# Instead of using the archives generated by GGML, use the object files directly to avoid overwriting objects with the same base name
-$(COMBINED_LIB): $(ALL_ARCHIVES)
-	@echo "Merging all .o into $(COMBINED_LIB): $(ALL_OBJS)"
-	rm -f $@
-	ar -qc $@ $(ALL_OBJS)
-	# Ensure we have a proper index
-	ranlib $@
-
-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) $(COMBINED_LIB)
-
-gosd.o:
-ifneq (,$(findstring sycl,$(BUILD_TYPE)))
-	+bash -c "source $(ONEAPI_VARS); \
-	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c"
-else
-	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c
-endif
-
-## stablediffusion (ggml)
 sources/stablediffusion-ggml.cpp:
 	git clone --recursive $(STABLEDIFFUSION_GGML_REPO) sources/stablediffusion-ggml.cpp && \
 	cd sources/stablediffusion-ggml.cpp && \
 	git checkout $(STABLEDIFFUSION_GGML_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-libsd.a: sources/stablediffusion-ggml.cpp build/libstable-diffusion.a gosd.o
-	cp $(INCLUDE_PATH)/build/libstable-diffusion.a ./libsd.a
-	$(AR) rcs libsd.a gosd.o
+libgosd.so: sources/stablediffusion-ggml.cpp CMakeLists.txt gosd.cpp gosd.h
+	mkdir -p build && \
+	cd build && \
+	cmake .. $(CMAKE_ARGS) && \
+	cmake --build . --config Release -j$(JOBS) && \
+	cd .. && \
+	mv build/libgosd.so ./
 
-stablediffusion-ggml: libsd.a
-	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_SYCL)" C_INCLUDE_PATH="$(INCLUDE_PATH)" LIBRARY_PATH="$(LIBRARY_PATH)" \
-	CC="$(CC)" CXX="$(CXX)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o stablediffusion-ggml ./
+stablediffusion-ggml: main.go gosd.go libgosd.so
+	CGO_ENABLED=0 $(GOCMD) build -tags "$(GO_TAGS)" -o stablediffusion-ggml ./
 
 package:
 	bash package.sh
@@ -150,4 +80,4 @@ package:
 build: stablediffusion-ggml package
 
 clean:
-	rm -rf gosd.o libsd.a build $(COMBINED_LIB)
+	rm -rf libgosd.o build stablediffusion-ggml

backend/go/stablediffusion-ggml/gosd.cpp

@@ -1,3 +1,6 @@
+#include <cstdint>
+#define GGML_MAX_NAME 128
+
 #include <stdio.h>
 #include <string.h>
 #include <time.h>
@@ -5,6 +8,7 @@
 #include <random>
 #include <string>
 #include <vector>
+#include <filesystem>
 #include "gosd.h"
 
 // #include "preprocessing.hpp"
@@ -54,7 +58,7 @@ sd_ctx_t* sd_c;
 sample_method_t sample_method;
 
 // Copied from the upstream CLI
-void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
+static void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     //SDParams* params = (SDParams*)data;
     const char* level_str;
 
@@ -85,31 +89,33 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     fflush(stderr);
 }
 
-int load_model(char *model, char* options[], int threads, int diff) {
-    fprintf (stderr, "Loading model!\n");
+int load_model(const char *model, char *model_path, char* options[], int threads, int diff) {
+    fprintf (stderr, "Loading model: %p=%s\n", model, model);
 
     sd_set_log_callback(sd_log_cb, NULL);
 
-    char *stableDiffusionModel = "";
+    const char *stableDiffusionModel = "";
     if (diff == 1 ) {
         stableDiffusionModel = model;
         model = "";
     }
 
     // decode options. Options are in form optname:optvale, or if booleans only optname.
-    char *clip_l_path  = "";
-    char *clip_g_path  = "";
-    char *t5xxl_path  = "";
-    char *vae_path  = "";
-    char *scheduler = "";
-    char *sampler = "";
+    const char *clip_l_path  = "";
+    const char *clip_g_path  = "";
+    const char *t5xxl_path  = "";
+    const char *vae_path  = "";
+    const char *scheduler = "";
+    const char *sampler = "";
+    char *lora_dir = model_path;
+    bool lora_dir_allocated = false;
 
-    fprintf(stderr, "parsing options\n");
+    fprintf(stderr, "parsing options: %p\n", options);
 
     // If options is not NULL, parse options
     for (int i = 0; options[i] != NULL; i++) {
-        char *optname = strtok(options[i], ":");
-        char *optval = strtok(NULL, ":");
+        const char *optname = strtok(options[i], ":");
+        const char *optval = strtok(NULL, ":");
         if (optval == NULL) {
             optval = "true";
         }
@@ -132,6 +138,21 @@ int load_model(char *model, char* options[], int threads, int diff) {
         if (!strcmp(optname, "sampler")) {
             sampler = optval;
         }
+        if (!strcmp(optname, "lora_dir")) {
+            // Path join with model dir
+            if (model_path && strlen(model_path) > 0) {
+                std::filesystem::path model_path_str(model_path);
+                std::filesystem::path lora_path(optval);
+                std::filesystem::path full_lora_path = model_path_str / lora_path;
+                lora_dir = strdup(full_lora_path.string().c_str());
+                lora_dir_allocated = true;
+                fprintf(stderr, "Lora dir resolved to: %s\n", lora_dir);
+            } else {
+                lora_dir = strdup(optval);
+                lora_dir_allocated = true;
+                fprintf(stderr, "No model path provided, using lora dir as-is: %s\n", lora_dir);
+            }
+        }
     }
 
     fprintf(stderr, "parsed options\n");
@@ -176,7 +197,7 @@ int load_model(char *model, char* options[], int threads, int diff) {
     ctx_params.vae_path = vae_path;
     ctx_params.taesd_path = "";
     ctx_params.control_net_path = "";
-    ctx_params.lora_model_dir = "";
+    ctx_params.lora_model_dir = lora_dir;
     ctx_params.embedding_dir = "";
     ctx_params.stacked_id_embed_dir = "";
     ctx_params.vae_decode_only = false;
@@ -189,16 +210,25 @@ int load_model(char *model, char* options[], int threads, int diff) {
 
     if (sd_ctx == NULL) {
         fprintf (stderr, "failed loading model (generic error)\n");
+        // Clean up allocated memory
+        if (lora_dir_allocated && lora_dir) {
+            free(lora_dir);
+        }
         return 1;
     }
     fprintf (stderr, "Created context: OK\n");
 
     sd_c = sd_ctx;
 
+    // Clean up allocated memory
+    if (lora_dir_allocated && lora_dir) {
+        free(lora_dir);
+    }
+
     return 0;
 }
 
-int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed , char *dst, float cfg_scale) {
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int64_t seed, char *dst, float cfg_scale, char *src_image, float strength, char *mask_image, char **ref_images, int ref_images_count) {
 
     sd_image_t* results;
 
@@ -221,15 +251,187 @@ int gen_image(char *text, char *negativeText, int width, int height, int steps,
     p.seed = seed;
     p.input_id_images_path = "";
 
+    // Handle input image for img2img
+    bool has_input_image = (src_image != NULL && strlen(src_image) > 0);
+    bool has_mask_image = (mask_image != NULL && strlen(mask_image) > 0);
+
+    uint8_t* input_image_buffer = NULL;
+    uint8_t* mask_image_buffer = NULL;
+    std::vector<uint8_t> default_mask_image_vec;
+
+    if (has_input_image) {
+        fprintf(stderr, "Loading input image: %s\n", src_image);
+
+        int c = 0;
+        int img_width = 0;
+        int img_height = 0;
+        input_image_buffer = stbi_load(src_image, &img_width, &img_height, &c, 3);
+        if (input_image_buffer == NULL) {
+            fprintf(stderr, "Failed to load input image from '%s'\n", src_image);
+            return 1;
+        }
+        if (c < 3) {
+            fprintf(stderr, "Input image must have at least 3 channels, got %d\n", c);
+            free(input_image_buffer);
+            return 1;
+        }
+
+        // Resize input image if dimensions don't match
+        if (img_width != width || img_height != height) {
+            fprintf(stderr, "Resizing input image from %dx%d to %dx%d\n", img_width, img_height, width, height);
+
+            uint8_t* resized_image_buffer = (uint8_t*)malloc(height * width * 3);
+            if (resized_image_buffer == NULL) {
+                fprintf(stderr, "Failed to allocate memory for resized image\n");
+                free(input_image_buffer);
+                return 1;
+            }
+
+            stbir_resize(input_image_buffer, img_width, img_height, 0,
+                         resized_image_buffer, width, height, 0, STBIR_TYPE_UINT8,
+                         3, STBIR_ALPHA_CHANNEL_NONE, 0,
+                         STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
+                         STBIR_FILTER_BOX, STBIR_FILTER_BOX,
+                         STBIR_COLORSPACE_SRGB, nullptr);
+
+            free(input_image_buffer);
+            input_image_buffer = resized_image_buffer;
+        }
+
+        p.init_image = {(uint32_t)width, (uint32_t)height, 3, input_image_buffer};
+        p.strength = strength;
+        fprintf(stderr, "Using img2img with strength: %.2f\n", strength);
+    } else {
+        // No input image, use empty image for text-to-image
+        p.init_image = {(uint32_t)width, (uint32_t)height, 3, NULL};
+        p.strength = 0.0f;
+    }
+
+    // Handle mask image for inpainting
+    if (has_mask_image) {
+        fprintf(stderr, "Loading mask image: %s\n", mask_image);
+
+        int c = 0;
+        int mask_width = 0;
+        int mask_height = 0;
+        mask_image_buffer = stbi_load(mask_image, &mask_width, &mask_height, &c, 1);
+        if (mask_image_buffer == NULL) {
+            fprintf(stderr, "Failed to load mask image from '%s'\n", mask_image);
+            if (input_image_buffer) free(input_image_buffer);
+            return 1;
+        }
+
+        // Resize mask if dimensions don't match
+        if (mask_width != width || mask_height != height) {
+            fprintf(stderr, "Resizing mask image from %dx%d to %dx%d\n", mask_width, mask_height, width, height);
+
+            uint8_t* resized_mask_buffer = (uint8_t*)malloc(height * width);
+            if (resized_mask_buffer == NULL) {
+                fprintf(stderr, "Failed to allocate memory for resized mask\n");
+                free(mask_image_buffer);
+                if (input_image_buffer) free(input_image_buffer);
+                return 1;
+            }
+
+            stbir_resize(mask_image_buffer, mask_width, mask_height, 0,
+                         resized_mask_buffer, width, height, 0, STBIR_TYPE_UINT8,
+                         1, STBIR_ALPHA_CHANNEL_NONE, 0,
+                         STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
+                         STBIR_FILTER_BOX, STBIR_FILTER_BOX,
+                         STBIR_COLORSPACE_SRGB, nullptr);
+
+            free(mask_image_buffer);
+            mask_image_buffer = resized_mask_buffer;
+        }
+
+        p.mask_image = {(uint32_t)width, (uint32_t)height, 1, mask_image_buffer};
+        fprintf(stderr, "Using inpainting with mask\n");
+    } else {
+        // No mask image, create default full mask
+        default_mask_image_vec.resize(width * height, 255);
+        p.mask_image = {(uint32_t)width, (uint32_t)height, 1, default_mask_image_vec.data()};
+    }
+
+    // Handle reference images
+    std::vector<sd_image_t> ref_images_vec;
+    std::vector<uint8_t*> ref_image_buffers;
+
+    if (ref_images_count > 0 && ref_images != NULL) {
+        fprintf(stderr, "Loading %d reference images\n", ref_images_count);
+
+        for (int i = 0; i < ref_images_count; i++) {
+            if (ref_images[i] == NULL || strlen(ref_images[i]) == 0) {
+                continue;
+            }
+
+            fprintf(stderr, "Loading reference image %d: %s\n", i + 1, ref_images[i]);
+
+            int c = 0;
+            int ref_width = 0;
+            int ref_height = 0;
+            uint8_t* ref_image_buffer = stbi_load(ref_images[i], &ref_width, &ref_height, &c, 3);
+            if (ref_image_buffer == NULL) {
+                fprintf(stderr, "Failed to load reference image from '%s'\n", ref_images[i]);
+                continue;
+            }
+            if (c < 3) {
+                fprintf(stderr, "Reference image must have at least 3 channels, got %d\n", c);
+                free(ref_image_buffer);
+                continue;
+            }
+
+            // Resize reference image if dimensions don't match
+            if (ref_width != width || ref_height != height) {
+                fprintf(stderr, "Resizing reference image from %dx%d to %dx%d\n", ref_width, ref_height, width, height);
+
+                uint8_t* resized_ref_buffer = (uint8_t*)malloc(height * width * 3);
+                if (resized_ref_buffer == NULL) {
+                    fprintf(stderr, "Failed to allocate memory for resized reference image\n");
+                    free(ref_image_buffer);
+                    continue;
+                }
+
+                stbir_resize(ref_image_buffer, ref_width, ref_height, 0,
+                             resized_ref_buffer, width, height, 0, STBIR_TYPE_UINT8,
+                             3, STBIR_ALPHA_CHANNEL_NONE, 0,
+                             STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
+                             STBIR_FILTER_BOX, STBIR_FILTER_BOX,
+                             STBIR_COLORSPACE_SRGB, nullptr);
+
+                free(ref_image_buffer);
+                ref_image_buffer = resized_ref_buffer;
+            }
+
+            ref_image_buffers.push_back(ref_image_buffer);
+            ref_images_vec.push_back({(uint32_t)width, (uint32_t)height, 3, ref_image_buffer});
+        }
+
+        if (!ref_images_vec.empty()) {
+            p.ref_images = ref_images_vec.data();
+            p.ref_images_count = ref_images_vec.size();
+            fprintf(stderr, "Using %zu reference images\n", ref_images_vec.size());
+        }
+    }
+
     results = generate_image(sd_c, &p);
 
     if (results == NULL) {
         fprintf (stderr, "NO results\n");
+        if (input_image_buffer) free(input_image_buffer);
+        if (mask_image_buffer) free(mask_image_buffer);
+        for (auto buffer : ref_image_buffers) {
+            if (buffer) free(buffer);
+        }
         return 1;
     }
 
     if (results[0].data == NULL) {
         fprintf (stderr, "Results with no data\n");
+        if (input_image_buffer) free(input_image_buffer);
+        if (mask_image_buffer) free(mask_image_buffer);
+        for (auto buffer : ref_image_buffers) {
+            if (buffer) free(buffer);
+        }
         return 1;
     }
 
@@ -245,17 +447,21 @@ int gen_image(char *text, char *negativeText, int width, int height, int steps,
                        results[0].data, 0, NULL);
     fprintf (stderr, "Saved resulting image to '%s'\n", dst);
 
-    // TODO: free results. Why does it crash?
-
+    // Clean up
     free(results[0].data);
     results[0].data = NULL;
     free(results);
-    fprintf (stderr, "gen_image is done", dst);
+    if (input_image_buffer) free(input_image_buffer);
+    if (mask_image_buffer) free(mask_image_buffer);
+    for (auto buffer : ref_image_buffers) {
+        if (buffer) free(buffer);
+    }
+    fprintf (stderr, "gen_image is done: %s", dst);
 
     return 0;
 }
 
 int unload() {
     free_sd_ctx(sd_c);
+    return 0;
 }
-

backend/go/stablediffusion-ggml/gosd.go

@@ -1,15 +1,10 @@
 package main
 
-// #cgo CXXFLAGS: -I${SRCDIR}/sources/stablediffusion-ggml.cpp/thirdparty -I${SRCDIR}/sources/stablediffusion-ggml.cpp -I${SRCDIR}/sources/stablediffusion-ggml.cpp/ggml/include
-// #cgo LDFLAGS: -L${SRCDIR}/ -lsd -lstdc++ -lm -lggmlall -lgomp
-// #include <gosd.h>
-// #include <stdlib.h>
-import "C"
-
 import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"runtime"
 	"strings"
 	"unsafe"
 
@@ -25,20 +20,34 @@ type SDGGML struct {
 	cfgScale     float32
 }
 
+var (
+	LoadModel func(model, model_apth string, options []uintptr, threads int32, diff int) int
+	GenImage  func(text, negativeText string, width, height, steps int, seed int64, dst string, cfgScale float32, srcImage string, strength float32, maskImage string, refImages []string, refImagesCount int) int
+)
+
+// Copied from Purego internal/strings
+// TODO: We should upstream sending []string
+func hasSuffix(s, suffix string) bool {
+	return len(s) >= len(suffix) && s[len(s)-len(suffix):] == suffix
+}
+
+func CString(name string) *byte {
+	if hasSuffix(name, "\x00") {
+		return &(*(*[]byte)(unsafe.Pointer(&name)))[0]
+	}
+	b := make([]byte, len(name)+1)
+	copy(b, name)
+	return &b[0]
+}
+
 func (sd *SDGGML) Load(opts *pb.ModelOptions) error {
 
 	sd.threads = int(opts.Threads)
 
-	modelFile := C.CString(opts.ModelFile)
-	defer C.free(unsafe.Pointer(modelFile))
+	modelPath := opts.ModelPath
 
-	var options **C.char
-	// prepare the options array to pass to C
-
-	size := C.size_t(unsafe.Sizeof((*C.char)(nil)))
-	length := C.size_t(len(opts.Options))
-	options = (**C.char)(C.malloc((length + 1) * size))
-	view := (*[1 << 30]*C.char)(unsafe.Pointer(options))[0:len(opts.Options) + 1:len(opts.Options) + 1]
+	modelFile := opts.ModelFile
+	modelPathC := modelPath
 
 	var diffusionModel int
 
@@ -63,32 +72,55 @@ func (sd *SDGGML) Load(opts *pb.ModelOptions) error {
 
 	fmt.Fprintf(os.Stderr, "Options: %+v\n", oo)
 
-	for i, x := range oo {
-		view[i] = C.CString(x)
+	// At the time of writing Purego doesn't recurse into slices and convert Go strings to pointers so we need to do that
+	var keepAlive []any
+	options := make([]uintptr, len(oo), len(oo)+1)
+	for i, op := range oo {
+		bytep := CString(op)
+		options[i] = uintptr(unsafe.Pointer(bytep))
+		keepAlive = append(keepAlive, bytep)
 	}
-	view[len(oo)] = nil
 
 	sd.cfgScale = opts.CFGScale
 
-	ret := C.load_model(modelFile, options, C.int(opts.Threads), C.int(diffusionModel))
+	ret := LoadModel(modelFile, modelPathC, options, opts.Threads, diffusionModel)
 	if ret != 0 {
 		return fmt.Errorf("could not load model")
 	}
 
+	runtime.KeepAlive(keepAlive)
+
 	return nil
 }
 
 func (sd *SDGGML) GenerateImage(opts *pb.GenerateImageRequest) error {
-	t := C.CString(opts.PositivePrompt)
-	defer C.free(unsafe.Pointer(t))
+	t := opts.PositivePrompt
+	dst := opts.Dst
+	negative := opts.NegativePrompt
+	srcImage := opts.Src
 
-	dst := C.CString(opts.Dst)
-	defer C.free(unsafe.Pointer(dst))
+	var maskImage string
+	if opts.EnableParameters != "" {
+		if strings.Contains(opts.EnableParameters, "mask:") {
+			parts := strings.Split(opts.EnableParameters, "mask:")
+			if len(parts) > 1 {
+				maskPath := strings.TrimSpace(parts[1])
+				if maskPath != "" {
+					maskImage = maskPath
+				}
+			}
+		}
+	}
 
-	negative := C.CString(opts.NegativePrompt)
-	defer C.free(unsafe.Pointer(negative))
+	refImagesCount := len(opts.RefImages)
+	refImages := make([]string, refImagesCount, refImagesCount+1)
+	copy(refImages, opts.RefImages)
+	*(*uintptr)(unsafe.Add(unsafe.Pointer(&refImages), refImagesCount)) = 0
 
-	ret := C.gen_image(t, negative, C.int(opts.Width), C.int(opts.Height), C.int(opts.Step), C.int(opts.Seed), dst, C.float(sd.cfgScale))
+	// Default strength for img2img (0.75 is a good default)
+	strength := float32(0.75)
+
+	ret := GenImage(t, negative, int(opts.Width), int(opts.Height), int(opts.Step), int64(opts.Seed), dst, sd.cfgScale, srcImage, strength, maskImage, refImages, refImagesCount)
 	if ret != 0 {
 		return fmt.Errorf("inference failed")
 	}

backend/go/stablediffusion-ggml/gosd.h

@@ -1,8 +1,8 @@
 #ifdef __cplusplus
 extern "C" {
 #endif
-int load_model(char *model, char* options[], int threads, int diffusionModel);
-int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed, char *dst, float cfg_scale);
+int load_model(const char *model, char *model_path, char* options[], int threads, int diffusionModel);
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int64_t seed, char *dst, float cfg_scale, char *src_image, float strength, char *mask_image, char **ref_images, int ref_images_count);
 #ifdef __cplusplus
 }
-#endif
+#endif

backend/go/stablediffusion-ggml/main.go

@@ -1,9 +1,9 @@
 package main
 
-// Note: this is started internally by LocalAI and a server is allocated for each model
 import (
 	"flag"
 
+	"github.com/ebitengine/purego"
 	grpc "github.com/mudler/LocalAI/pkg/grpc"
 )
 
@@ -12,6 +12,14 @@ var (
 )
 
 func main() {
+	gosd, err := purego.Dlopen("./libgosd.so", purego.RTLD_NOW|purego.RTLD_GLOBAL)
+	if err != nil {
+		panic(err)
+	}
+
+	purego.RegisterLibFunc(&LoadModel, gosd, "load_model")
+	purego.RegisterLibFunc(&GenImage, gosd, "gen_image")
+
 	flag.Parse()
 
 	if err := grpc.StartServer(*addr, &SDGGML{}); err != nil {

backend/go/stablediffusion-ggml/package.sh

@@ -10,6 +10,7 @@ CURDIR=$(dirname "$(realpath $0)")
 # Create lib directory
 mkdir -p $CURDIR/package/lib
 
+cp -avrf $CURDIR/libgosd.so $CURDIR/package/
 cp -avrf $CURDIR/stablediffusion-ggml $CURDIR/package/
 cp -rfv $CURDIR/run.sh $CURDIR/package/
 
@@ -47,6 +48,6 @@ else
     exit 1
 fi
 
-echo "Packaging completed successfully" 
+echo "Packaging completed successfully"
 ls -liah $CURDIR/package/
-ls -liah $CURDIR/package/lib/
+ls -liah $CURDIR/package/lib/

backend/go/whisper/Makefile

@@ -6,7 +6,7 @@ CMAKE_ARGS?=
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=7de8dd783f7b2eab56bff6bbc5d3369e34f0e77f
+WHISPER_CPP_VERSION?=fc45bb86251f774ef817e89878bb4c2636c8a58f
 
 export WHISPER_CMAKE_ARGS?=-DBUILD_SHARED_LIBS=OFF
 export WHISPER_DIR=$(abspath ./sources/whisper.cpp)

backend/python/bark/Makefile

@@ -1,29 +1,23 @@
 .PHONY: ttsbark
-ttsbark: protogen
+ttsbark:
 	bash install.sh
 
 .PHONY: run
-run: protogen
+run: ttsbark
 	@echo "Running bark..."
 	bash run.sh
 	@echo "bark run."
 
 .PHONY: test
-test: protogen
+test: ttsbark
 	@echo "Testing bark..."
 	bash test.sh
 	@echo "bark 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../.. -I./ --python_out=. --grpc_python_out=. backend.proto
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/bark/requirements-intel.txt

@@ -1,5 +1,5 @@
 --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==2.8.10+xpu
 torch==2.3.1+cxx11.abi
 torchaudio==2.3.1+cxx11.abi
 oneccl_bind_pt==2.3.100+xpu

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.71.0
+grpcio==1.74.0
 protobuf
 certifi

backend/python/chatterbox/Makefile

@@ -1,29 +1,23 @@
-.PHONY: coqui
-coqui: protogen
+.PHONY: chatterbox
+chatterbox:
 	bash install.sh
 
 .PHONY: run
-run: protogen
+run: chatterbox
 	@echo "Running coqui..."
 	bash run.sh
 	@echo "coqui run."
 
 .PHONY: test
-test: protogen
+test: chatterbox
 	@echo "Testing coqui..."
 	bash test.sh
 	@echo "coqui 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../.. -I./ --python_out=. --grpc_python_out=. backend.proto
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/chatterbox/backend.py

@@ -41,7 +41,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         else:
             print("CUDA is not available", file=sys.stderr)
             device = "cpu"
-
+        mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+        if mps_available:
+            device = "mps"
         if not torch.cuda.is_available() and request.CUDA:
             return backend_pb2.Result(success=False, message="CUDA is not available")
 

backend/python/common/libbackend.sh

@@ -1,6 +1,6 @@
+#!/usr/bin/env bash
+set -euo pipefail
 
-
-# init handles the setup of the library
 # 
 # use the library by adding the following line to a script:
 # source $(dirname $0)/../common/libbackend.sh
@@ -17,29 +17,182 @@
 # LIMIT_TARGETS="cublas12"
 # source $(dirname $0)/../common/libbackend.sh
 #
+# You can switch between uv (conda-like) and pip installation methods by setting USE_PIP:
+# USE_PIP=true source $(dirname $0)/../common/libbackend.sh
+#
+# ===================== user-configurable defaults =====================
+PYTHON_VERSION="${PYTHON_VERSION:-3.10}"      # e.g. 3.10 / 3.11 / 3.12 / 3.13
+PYTHON_PATCH="${PYTHON_PATCH:-18}"            # e.g. 18 -> 3.10.18 ; 13 -> 3.11.13
+PY_STANDALONE_TAG="${PY_STANDALONE_TAG:-20250818}"  # release tag date
+# Enable/disable bundling of a portable Python build
+PORTABLE_PYTHON="${PORTABLE_PYTHON:-false}"
 
-PYTHON_VERSION="3.10"
+# If you want to fully pin the filename (including tuned CPU targets), set:
+# PORTABLE_PY_FILENAME="cpython-3.10.18+20250818-x86_64_v3-unknown-linux-gnu-install_only.tar.gz"
+: "${PORTABLE_PY_FILENAME:=}"
+: "${PORTABLE_PY_SHA256:=}"  # optional; if set we verify the download
+# =====================================================================
 
+# Default to uv if USE_PIP is not set
+if [ "x${USE_PIP:-}" == "x" ]; then
+    USE_PIP=false
+fi
+
+# ----------------------- helpers -----------------------
+function _is_musl() {
+    # detect musl (Alpine, etc)
+    if command -v ldd >/dev/null 2>&1; then
+        ldd --version 2>&1 | grep -qi musl && return 0
+    fi
+    # busybox-ish fallback
+    if command -v getconf >/dev/null 2>&1; then
+        getconf GNU_LIBC_VERSION >/dev/null 2>&1 || return 0
+    fi
+    return 1
+}
+
+function _triple() {
+    local os="" arch="" libc="gnu"
+    case "$(uname -s)" in
+        Linux*)  os="unknown-linux" ;;
+        Darwin*) os="apple-darwin" ;;
+        MINGW*|MSYS*|CYGWIN*) os="pc-windows-msvc" ;;  # best-effort for Git Bash
+        *) echo "Unsupported OS $(uname -s)"; exit 1;;
+    esac
+
+    case "$(uname -m)" in
+        x86_64) arch="x86_64" ;;
+        aarch64|arm64) arch="aarch64" ;;
+        armv7l) arch="armv7" ;;
+        i686|i386) arch="i686" ;;
+        ppc64le) arch="ppc64le" ;;
+        s390x) arch="s390x" ;;
+        riscv64) arch="riscv64" ;;
+        *) echo "Unsupported arch $(uname -m)"; exit 1;;
+    esac
+
+    if [[ "$os" == "unknown-linux" ]]; then
+        if _is_musl; then
+            libc="musl"
+        else
+            libc="gnu"
+        fi
+        echo "${arch}-${os}-${libc}"
+    else
+        echo "${arch}-${os}"
+    fi
+}
+
+function _portable_dir() {
+    echo "${EDIR}/python"
+}
+
+function _portable_bin() {
+    # python-build-standalone puts python in ./bin
+    echo "$(_portable_dir)/bin"
+}
+
+function _portable_python() {
+    if [ -x "$(_portable_bin)/python3" ]; then
+        echo "$(_portable_bin)/python3"
+    else
+        echo "$(_portable_bin)/python"
+    fi
+}
+
+
+# macOS loader env for the portable CPython
+_macosPortableEnv() {
+  if [ "$(uname -s)" = "Darwin" ]; then
+    export DYLD_LIBRARY_PATH="$(_portable_dir)/lib${DYLD_LIBRARY_PATH:+:${DYLD_LIBRARY_PATH}}"
+    export DYLD_FALLBACK_LIBRARY_PATH="$(_portable_dir)/lib${DYLD_FALLBACK_LIBRARY_PATH:+:${DYLD_FALLBACK_LIBRARY_PATH}}"
+  fi
+}
+
+# Good hygiene on macOS for downloaded/extracted trees
+_unquarantinePortablePython() {
+  if [ "$(uname -s)" = "Darwin" ]; then
+    command -v xattr >/dev/null 2>&1 && xattr -dr com.apple.quarantine "$(_portable_dir)" || true
+  fi
+}
+
+# ------------------ ### PORTABLE PYTHON ------------------
+function ensurePortablePython() {
+    local pdir="$(_portable_dir)"
+    local pbin="$(_portable_bin)"
+    local pyexe
+
+    if [ -x "${pbin}/python3" ] || [ -x "${pbin}/python" ]; then
+        _macosPortableEnv
+        return 0
+    fi
+
+    mkdir -p "${pdir}"
+    local triple="$(_triple)"
+
+    local full_ver="${PYTHON_VERSION}.${PYTHON_PATCH}"
+    local fn=""
+    if [ -n "${PORTABLE_PY_FILENAME}" ]; then
+        fn="${PORTABLE_PY_FILENAME}"
+    else
+        # generic asset name: cpython-<full_ver>+<tag>-<triple>-install_only.tar.gz
+        fn="cpython-${full_ver}+${PY_STANDALONE_TAG}-${triple}-install_only.tar.gz"
+    fi
+
+    local url="https://github.com/astral-sh/python-build-standalone/releases/download/${PY_STANDALONE_TAG}/${fn}"
+    local tmp="${pdir}/${fn}"
+    echo "Downloading portable Python: ${fn}"
+    # curl with retries; fall back to wget if needed
+    if command -v curl >/dev/null 2>&1; then
+        curl -L --fail --retry 3 --retry-delay 1 -o "${tmp}" "${url}"
+    else
+        wget -O "${tmp}" "${url}"
+    fi
+
+    if [ -n "${PORTABLE_PY_SHA256}" ]; then
+        echo "${PORTABLE_PY_SHA256}  ${tmp}" | sha256sum -c -
+    fi
+
+    echo "Extracting ${fn} -> ${pdir}"
+    # always a .tar.gz (we purposely choose install_only)
+    tar -xzf "${tmp}" -C "${pdir}"
+    rm -f "${tmp}"
+
+    # Some archives nest a directory; if so, flatten to ${pdir}
+    # Find the first dir with a 'bin/python*'
+    local inner
+    inner="$(find "${pdir}" -type f -path "*/bin/python*" -maxdepth 3 2>/dev/null | head -n1 || true)"
+    if [ -n "${inner}" ]; then
+        local inner_root
+        inner_root="$(dirname "$(dirname "${inner}")")" # .../bin -> root
+        if [ "${inner_root}" != "${pdir}" ]; then
+            # move contents up one level
+            shopt -s dotglob
+            mv "${inner_root}/"* "${pdir}/"
+            rm -rf "${inner_root}"
+            shopt -u dotglob
+        fi
+    fi
+
+    _unquarantinePortablePython
+    _macosPortableEnv
+    # Make sure it's runnable
+    pyexe="$(_portable_python)"
+    "${pyexe}" -V
+}
+
+# init handles the setup of the library
 function init() {
-    # Name of the backend (directory name)
     BACKEND_NAME=${PWD##*/}
-
-    # Path where all backends files are
-    MY_DIR=$(realpath `dirname $0`)
-
-    # Build type
+    MY_DIR=$(realpath "$(dirname "$0")")
     BUILD_PROFILE=$(getBuildProfile)
 
-    # Environment directory
     EDIR=${MY_DIR}
-
-    # Allow to specify a custom env dir for shared environments
-    if [ "x${ENV_DIR}" != "x" ]; then
+    if [ "x${ENV_DIR:-}" != "x" ]; then
         EDIR=${ENV_DIR}
     fi
 
-    # If a backend has defined a list of valid build profiles...
-    if [ ! -z "${LIMIT_TARGETS}" ]; then
+    if [ ! -z "${LIMIT_TARGETS:-}" ]; then
         isValidTarget=$(checkTargets ${LIMIT_TARGETS})
         if [ ${isValidTarget} != true ]; then
             echo "${BACKEND_NAME} can only be used on the following targets: ${LIMIT_TARGETS}"
@@ -50,6 +203,7 @@ function init() {
     echo "Initializing libbackend for ${BACKEND_NAME}"
 }
 
+
 # getBuildProfile will inspect the system to determine which build profile is appropriate:
 # returns one of the following:
 # - cublas11
@@ -57,53 +211,139 @@ function init() {
 # - hipblas
 # - intel
 function getBuildProfile() {
-    # First check if we are a cublas build, and if so report the correct build profile
-    if [ x"${BUILD_TYPE}" == "xcublas" ]; then
-        if [ ! -z ${CUDA_MAJOR_VERSION} ]; then
-            # If we have been given a CUDA version, we trust it
+    if [ x"${BUILD_TYPE:-}" == "xcublas" ]; then
+        if [ ! -z "${CUDA_MAJOR_VERSION:-}" ]; then
             echo ${BUILD_TYPE}${CUDA_MAJOR_VERSION}
         else
-            # We don't know what version of cuda we are, so we report ourselves as a generic cublas
             echo ${BUILD_TYPE}
         fi
         return 0
     fi
 
-    # If /opt/intel exists, then we are doing an intel/ARC build
     if [ -d "/opt/intel" ]; then
         echo "intel"
         return 0
     fi
 
-    # If for any other values of BUILD_TYPE, we don't need any special handling/discovery
-    if [ ! -z ${BUILD_TYPE} ]; then
+    if [ -n "${BUILD_TYPE:-}" ]; then
         echo ${BUILD_TYPE}
         return 0
     fi
 
-    # If there is no BUILD_TYPE set at all, set a build-profile value of CPU, we aren't building for any GPU targets
     echo "cpu"
 }
 
+
+# Make the venv relocatable:
+# - rewrite venv/bin/python{,3} to relative symlinks into $(_portable_dir)
+# - normalize entrypoint shebangs to /usr/bin/env python3
+_makeVenvPortable() {
+    local venv_dir="${EDIR}/venv"
+    local vbin="${venv_dir}/bin"
+
+    [ -d "${vbin}" ] || return 0
+
+    # 1) Replace python symlinks with relative ones to ../../python/bin/python3
+    #    (venv/bin -> venv -> EDIR -> python/bin)
+    local rel_py='../../python/bin/python3'
+
+    for name in python3 python; do
+        if [ -e "${vbin}/${name}" ] || [ -L "${vbin}/${name}" ]; then
+            rm -f "${vbin}/${name}"
+        fi
+    done
+    ln -s "${rel_py}" "${vbin}/python3"
+    ln -s "python3" "${vbin}/python"
+
+    # 2) Rewrite shebangs of entry points to use env, so the venv is relocatable
+    #    Only touch text files that start with #! and reference the current venv.
+    local ve_abs="${vbin}/python"
+    local sed_i=(sed -i)
+    # macOS/BSD sed needs a backup suffix; GNU sed doesn't. Make it portable:
+    if sed --version >/dev/null 2>&1; then
+        sed_i=(sed -i)
+    else
+        sed_i=(sed -i '')
+    fi
+
+    for f in "${vbin}"/*; do
+        [ -f "$f" ] || continue
+        # Fast path: check first two bytes (#!)
+        head -c2 "$f" 2>/dev/null | grep -q '^#!' || continue
+        # Only rewrite if the shebang mentions the (absolute) venv python
+        if head -n1 "$f" | grep -Fq "${ve_abs}"; then
+            "${sed_i[@]}" '1s|^#!.*$|#!/usr/bin/env python3|' "$f"
+            chmod +x "$f" 2>/dev/null || true
+        fi
+    done
+}
+
+
 # ensureVenv makes sure that the venv for the backend both exists, and is activated.
 #
 # This function is idempotent, so you can call it as many times as you want and it will
 # always result in an activated virtual environment
 function ensureVenv() {
+    local interpreter=""
+
+    if [ "x${PORTABLE_PYTHON}" == "xtrue" ]; then
+        ensurePortablePython
+        interpreter="$(_portable_python)"
+    else
+        # Prefer system python${PYTHON_VERSION}, else python3, else fall back to bundled
+        if command -v python${PYTHON_VERSION} >/dev/null 2>&1; then
+            interpreter="python${PYTHON_VERSION}"
+        elif command -v python3 >/dev/null 2>&1; then
+            interpreter="python3"
+        else
+            echo "No suitable system Python found, bootstrapping portable build..."
+            ensurePortablePython
+            interpreter="$(_portable_python)"
+        fi
+    fi
+
     if [ ! -d "${EDIR}/venv" ]; then
-        uv venv --python ${PYTHON_VERSION} ${EDIR}/venv
-        echo "virtualenv created"
+        if [ "x${USE_PIP}" == "xtrue" ]; then
+            "${interpreter}" -m venv --copies "${EDIR}/venv"
+            source "${EDIR}/venv/bin/activate"
+            "${interpreter}" -m pip install --upgrade pip
+        else
+            if [ "x${PORTABLE_PYTHON}" == "xtrue" ]; then
+                uv venv --python "${interpreter}" "${EDIR}/venv"
+            else
+                uv venv --python "${PYTHON_VERSION}" "${EDIR}/venv"
+            fi
+        fi
+        if [ "x${PORTABLE_PYTHON}" == "xtrue" ]; then
+            _makeVenvPortable
+        fi
     fi
 
-    # Source if we are not already in a Virtual env
-    if [ "x${VIRTUAL_ENV}" != "x${EDIR}/venv" ]; then
-        source ${EDIR}/venv/bin/activate
-        echo "virtualenv activated"
+    # We call it here to make sure that when we source a venv we can still use python as expected
+    if [ -x "$(_portable_python)" ]; then
+        _macosPortableEnv
     fi
 
-    echo "activated virtualenv has been ensured"
+    if [ "x${VIRTUAL_ENV:-}" != "x${EDIR}/venv" ]; then
+        source "${EDIR}/venv/bin/activate"
+    fi
 }
 
+
+function runProtogen() {
+    ensureVenv
+    if [ "x${USE_PIP}" == "xtrue" ]; then
+        pip install grpcio-tools
+    else
+        uv pip install grpcio-tools
+    fi
+    pushd "${EDIR}" >/dev/null
+        # use the venv python (ensures correct interpreter & sys.path)
+        python -m grpc_tools.protoc -I../../ -I./ --python_out=. --grpc_python_out=. backend.proto
+    popd >/dev/null
+}
+
+
 # installRequirements looks for several requirements files and if they exist runs the install for them in order
 #
 #  - requirements-install.txt
@@ -111,7 +351,7 @@ function ensureVenv() {
 #  - requirements-${BUILD_TYPE}.txt
 #  - requirements-${BUILD_PROFILE}.txt
 #
-# BUILD_PROFILE is a pore specific version of BUILD_TYPE, ex: cuda11 or cuda12
+# BUILD_PROFILE is a more specific version of BUILD_TYPE, ex: cuda-11 or cuda-12
 # it can also include some options that we do not have BUILD_TYPES for, ex: intel
 #
 # NOTE: for BUILD_PROFILE==intel, this function does NOT automatically use the Intel python package index.
@@ -127,36 +367,36 @@ function ensureVenv() {
 # installRequirements
 function installRequirements() {
     ensureVenv
-
-    # These are the requirements files we will attempt to install, in order
     declare -a requirementFiles=(
         "${EDIR}/requirements-install.txt"
         "${EDIR}/requirements.txt"
-        "${EDIR}/requirements-${BUILD_TYPE}.txt"
+        "${EDIR}/requirements-${BUILD_TYPE:-}.txt"
     )
 
-    if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
+    if [ "x${BUILD_TYPE:-}" != "x${BUILD_PROFILE}" ]; then
         requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}.txt")
     fi
-
-    # if BUILD_TYPE is empty, we are a CPU build, so we should try to install the CPU requirements
-    if [ "x${BUILD_TYPE}" == "x" ]; then
+    if [ "x${BUILD_TYPE:-}" == "x" ]; then
         requirementFiles+=("${EDIR}/requirements-cpu.txt")
     fi
-
     requirementFiles+=("${EDIR}/requirements-after.txt")
-
-    if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
+    if [ "x${BUILD_TYPE:-}" != "x${BUILD_PROFILE}" ]; then
         requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}-after.txt")
     fi
 
     for reqFile in ${requirementFiles[@]}; do
-        if [ -f ${reqFile} ]; then
+        if [ -f "${reqFile}" ]; then
             echo "starting requirements install for ${reqFile}"
-            uv pip install ${EXTRA_PIP_INSTALL_FLAGS} --requirement ${reqFile}
+            if [ "x${USE_PIP}" == "xtrue" ]; then
+                pip install ${EXTRA_PIP_INSTALL_FLAGS:-} --requirement "${reqFile}"
+            else
+                uv pip install ${EXTRA_PIP_INSTALL_FLAGS:-} --requirement "${reqFile}"
+            fi
             echo "finished requirements install for ${reqFile}"
         fi
     done
+
+    runProtogen
 }
 
 # startBackend discovers and runs the backend GRPC server
@@ -174,18 +414,18 @@ function installRequirements() {
 #  - ${BACKEND_NAME}.py
 function startBackend() {
     ensureVenv
-
-    if [ ! -z ${BACKEND_FILE} ]; then
-        exec ${EDIR}/venv/bin/python ${BACKEND_FILE} $@
+    if [ ! -z "${BACKEND_FILE:-}" ]; then
+        exec "${EDIR}/venv/bin/python" "${BACKEND_FILE}" "$@"
     elif [ -e "${MY_DIR}/server.py" ]; then
-        exec ${EDIR}/venv/bin/python ${MY_DIR}/server.py $@
+        exec "${EDIR}/venv/bin/python" "${MY_DIR}/server.py" "$@"
     elif [ -e "${MY_DIR}/backend.py" ]; then
-        exec ${EDIR}/venv/bin/python ${MY_DIR}/backend.py $@
+        exec "${EDIR}/venv/bin/python" "${MY_DIR}/backend.py" "$@"
     elif [ -e "${MY_DIR}/${BACKEND_NAME}.py" ]; then
-        exec ${EDIR}/venv/bin/python ${MY_DIR}/${BACKEND_NAME}.py $@
+        exec "${EDIR}/venv/bin/python" "${MY_DIR}/${BACKEND_NAME}.py" "$@"
     fi
 }
 
+
 # runUnittests discovers and runs python unittests
 #
 # You can specify a specific test file to use by setting TEST_FILE before calling runUnittests.
@@ -198,41 +438,36 @@ function startBackend() {
 # be default a file named test.py in the backends directory will be used
 function runUnittests() {
     ensureVenv
-
-    if [ ! -z ${TEST_FILE} ]; then
-        testDir=$(dirname `realpath ${TEST_FILE}`)
-        testFile=$(basename ${TEST_FILE})
-        pushd ${testDir}
-        python -m unittest ${testFile}
-        popd
+    if [ ! -z "${TEST_FILE:-}" ]; then
+        testDir=$(dirname "$(realpath "${TEST_FILE}")")
+        testFile=$(basename "${TEST_FILE}")
+        pushd "${testDir}" >/dev/null
+        python -m unittest "${testFile}"
+        popd >/dev/null
     elif [ -f "${MY_DIR}/test.py" ]; then
-        pushd ${MY_DIR}
+        pushd "${MY_DIR}" >/dev/null
         python -m unittest test.py
-        popd
+        popd >/dev/null
     else
         echo "no tests defined for ${BACKEND_NAME}"
     fi
 }
 
+
 ##################################################################################
 # Below here are helper functions not intended to be used outside of the library #
 ##################################################################################
 
 # checkTargets determines if the current BUILD_TYPE or BUILD_PROFILE is in a list of valid targets
 function checkTargets() {
-    # Collect all provided targets into a variable and...
     targets=$@
-    # ...convert it into an array
     declare -a targets=($targets)
-
     for target in ${targets[@]}; do
-        if [ "x${BUILD_TYPE}" == "x${target}" ]; then
-            echo true
-            return 0
+        if [ "x${BUILD_TYPE:-}" == "x${target}" ]; then
+            echo true; return 0
         fi
         if [ "x${BUILD_PROFILE}" == "x${target}" ]; then
-            echo true
-            return 0
+            echo true; return 0
         fi
     done
     echo false

backend/python/common/template/Makefile

@@ -3,18 +3,11 @@
 .PHONY: install
 install:
 	bash install.sh
-	$(MAKE) protogen
-
-.PHONY: protogen
-protogen: backend_pb2_grpc.py backend_pb2.py
 
 .PHONY: protogen-clean
 protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
-backend_pb2_grpc.py backend_pb2.py:
-	bash protogen.sh
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/common/template/protogen.sh

@@ -8,4 +8,4 @@ else
     source $backend_dir/../common/libbackend.sh
 fi
 
-python3 -m grpc_tools.protoc -I../.. -I./ --python_out=. --grpc_python_out=. backend.proto
+runProtogen

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

@@ -1,5 +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==2.8.10+xpu
+torch==2.8.0
+oneccl_bind_pt==2.8.0+xpu
 optimum[openvino]

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.74.0
 protobuf
 grpcio-tools

backend/python/coqui/Makefile

@@ -1,29 +1,23 @@
 .PHONY: coqui
-coqui: protogen
+coqui:
 	bash install.sh
 
 .PHONY: run
-run: protogen
+run: coqui
 	@echo "Running coqui..."
 	bash run.sh
 	@echo "coqui run."
 
 .PHONY: test
-test: protogen
+test: coqui
 	@echo "Testing coqui..."
 	bash test.sh
 	@echo "coqui 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../.. -I./ --python_out=. --grpc_python_out=. backend.proto
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/coqui/backend.py

@@ -40,7 +40,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         else:
             print("CUDA is not available", file=sys.stderr)
             device = "cpu"
-
+        mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+        if mps_available:
+            device = "mps"
         if not torch.cuda.is_available() and request.CUDA:
             return backend_pb2.Result(success=False, message="CUDA is not available")
 

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.74.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/Makefile

@@ -12,28 +12,22 @@ export SKIP_CONDA=1
 endif
 
 .PHONY: diffusers
-diffusers: protogen
+diffusers:
 	bash install.sh
 
 .PHONY: run
-run: protogen
+run: diffusers
 	@echo "Running diffusers..."
 	bash run.sh
 	@echo "Diffusers run."
 
-test: protogen
+test: diffusers
 	bash test.sh
 
-.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../.. -I./ --python_out=. --grpc_python_out=. backend.proto
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/diffusers/backend.py

@@ -18,7 +18,7 @@ import backend_pb2_grpc
 import grpc
 
 from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
-    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
+    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel, QwenImageEditPipeline
 from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline, Lumina2Text2ImgPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
 from diffusers.utils import load_image, export_to_video
@@ -65,6 +65,19 @@ from diffusers.schedulers import (
     UniPCMultistepScheduler,
 )
 
+def is_float(s):
+    try:
+        float(s)
+        return True
+    except ValueError:
+        return False
+
+def is_int(s):
+    try:
+        int(s)
+        return True
+    except ValueError:
+        return False
 
 # The scheduler list mapping was taken from here: https://github.com/neggles/animatediff-cli/blob/6f336f5f4b5e38e85d7f06f1744ef42d0a45f2a7/src/animatediff/schedulers.py#L39
 # Credits to https://github.com/neggles
@@ -169,8 +182,24 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 if ":" not in opt:
                     continue
                 key, value = opt.split(":")
+                # if value is a number, convert it to the appropriate type
+                if is_float(value):
+                    value = float(value)
+                elif is_int(value):
+                    value = int(value)
                 self.options[key] = value
 
+            # From options, extract if present "torch_dtype" and set it to the appropriate type
+            if "torch_dtype" in self.options:
+                if self.options["torch_dtype"] == "fp16":
+                    torchType = torch.float16
+                elif self.options["torch_dtype"] == "bf16":
+                    torchType = torch.bfloat16
+                elif self.options["torch_dtype"] == "fp32":
+                    torchType = torch.float32
+                # remove it from options
+                del self.options["torch_dtype"]
+
             print(f"Options: {self.options}", file=sys.stderr)
 
             local = False
@@ -234,6 +263,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             elif request.PipelineType == "DiffusionPipeline":
                 self.pipe = DiffusionPipeline.from_pretrained(request.Model,
                                                               torch_dtype=torchType)
+            elif request.PipelineType == "QwenImageEditPipeline":
+                self.pipe = QwenImageEditPipeline.from_pretrained(request.Model,
+                                                                 torch_dtype=torchType)
             elif request.PipelineType == "VideoDiffusionPipeline":
                 self.txt2vid = True
                 self.pipe = DiffusionPipeline.from_pretrained(request.Model,
@@ -336,6 +368,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             device = "cpu" if not request.CUDA else "cuda"
             if XPU:
                 device = "xpu"
+            mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+            if mps_available:
+                device = "mps"
             self.device = device
             if request.LoraAdapter:
                 # Check if its a local file and not a directory ( we load lora differently for a safetensor file )

backend/python/diffusers/requirements-cpu.txt

@@ -1,9 +1,11 @@
-diffusers
+--extra-index-url https://download.pytorch.org/whl/cpu
+git+https://github.com/huggingface/diffusers
 opencv-python
 transformers
+torchvision==0.22.1
 accelerate
 compel
 peft
 sentencepiece
-torch==2.4.1
+torch==2.7.1
 optimum-quanto

backend/python/diffusers/requirements-cublas11.txt

@@ -1,6 +1,7 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
-torch==2.4.1+cu118
-diffusers
+torch==2.7.1+cu118
+torchvision==0.22.1+cu118
+git+https://github.com/huggingface/diffusers
 opencv-python
 transformers
 accelerate

backend/python/diffusers/requirements-cublas12.txt

@@ -1,5 +1,6 @@
-torch==2.4.1
-diffusers
+torch==2.7.1
+torchvision==0.22.1
+git+https://github.com/huggingface/diffusers
 opencv-python
 transformers
 accelerate

backend/python/diffusers/requirements-hipblas.txt

@@ -1,7 +1,7 @@
---extra-index-url https://download.pytorch.org/whl/rocm6.0
-torch==2.3.1+rocm6.0
-torchvision==0.18.1+rocm6.0
-diffusers
+--extra-index-url https://download.pytorch.org/whl/rocm6.3
+torch==2.7.1+rocm6.3
+torchvision==0.22.1+rocm6.3
+git+https://github.com/huggingface/diffusers
 opencv-python
 transformers
 accelerate

backend/python/diffusers/requirements-intel.txt

@@ -1,11 +1,11 @@
 --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
+torch==2.5.1+cxx11.abi
+torchvision==0.20.1+cxx11.abi
+oneccl_bind_pt==2.8.0+xpu
 optimum[openvino]
 setuptools
-diffusers
+git+https://github.com/huggingface/diffusers
 opencv-python
 transformers
 accelerate

backend/python/diffusers/requirements-l4t.txt

@@ -0,0 +1,11 @@
+--extra-index-url https://pypi.jetson-ai-lab.io/jp6/cu126/
+torch
+diffusers
+transformers
+accelerate
+compel
+peft
+optimum-quanto
+numpy<2
+sentencepiece
+torchvision

backend/python/diffusers/requirements-mps.txt

@@ -0,0 +1,10 @@
+torch==2.7.1
+torchvision==0.22.1
+git+https://github.com/huggingface/diffusers
+opencv-python
+transformers
+accelerate
+compel
+peft
+sentencepiece
+optimum-quanto

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.71.0
+grpcio==1.74.0
 pillow
 protobuf
 certifi

backend/python/diffusers/run.sh

@@ -12,4 +12,6 @@ if [ -d "/opt/intel" ]; then
     export XPU=1
 fi
 
+export PYTORCH_ENABLE_MPS_FALLBACK=1
+
 startBackend $@

backend/python/exllama2/Makefile

@@ -1,23 +1,17 @@
 .PHONY: exllama2
-exllama2: protogen
+exllama2:
 	bash install.sh
 
 .PHONY: run
-run: protogen
+run: exllama2
 	@echo "Running exllama2..."
 	bash run.sh
 	@echo "exllama2 run."
 
-.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../.. -I./ --python_out=. --grpc_python_out=. backend.proto
-
 .PHONY: clean
 clean: protogen-clean
 	$(RM) -r venv source __pycache__

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.74.0
 protobuf
 certifi
 wheel

backend/python/faster-whisper/Makefile

@@ -3,18 +3,11 @@
 .PHONY: install
 install:
 	bash install.sh
-	$(MAKE) protogen
-
-.PHONY: protogen
-protogen: backend_pb2_grpc.py backend_pb2.py
 
 .PHONY: protogen-clean
 protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
-backend_pb2_grpc.py backend_pb2.py:
-	bash protogen.sh
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/faster-whisper/backend.py

@@ -10,7 +10,7 @@ import sys
 import os
 import backend_pb2
 import backend_pb2_grpc
-
+import torch
 from faster_whisper import WhisperModel
 
 import grpc
@@ -35,7 +35,9 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # device = "cuda" if request.CUDA else "cpu"
         if request.CUDA:
             device = "cuda"
-
+        mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
+        if mps_available:
+            device = "mps"
         try:
             print("Preparing models, please wait", file=sys.stderr)
             self.model = WhisperModel(request.Model, device=device, compute_type="float16")

backend/python/kitten-tts/Makefile

@@ -0,0 +1,23 @@
+.PHONY: kitten-tts
+kitten-tts:
+	bash install.sh
+
+.PHONY: run
+run: kitten-tts
+	@echo "Running kitten-tts..."
+	bash run.sh
+	@echo "kitten-tts run."
+
+.PHONY: test
+test: kitten-tts
+	@echo "Testing kitten-tts..."
+	bash test.sh
+	@echo "kitten-tts tested."
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__

backend/python/kitten-tts/backend.py

@@ -0,0 +1,109 @@
+#!/usr/bin/env python3
+"""
+This is an extra gRPC server of LocalAI for Kitten TTS
+"""
+from concurrent import futures
+import time
+import argparse
+import signal
+import sys
+import os
+import backend_pb2
+import backend_pb2_grpc
+
+import torch
+from kittentts import KittenTTS
+import soundfile as sf
+
+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'))
+KITTEN_LANGUAGE = os.environ.get('KITTEN_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
+    """
+    def Health(self, request, context):
+        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
+    def LoadModel(self, request, context):
+
+        self.AudioPath = None
+        # List available KittenTTS models
+        print("Available KittenTTS voices: expr-voice-2-m, expr-voice-2-f, expr-voice-3-m, expr-voice-3-f, expr-voice-4-m, expr-voice-4-f, expr-voice-5-m, expr-voice-5-f")
+        if os.path.isabs(request.AudioPath):
+            self.AudioPath = request.AudioPath
+        elif request.AudioPath and request.ModelFile != "" and not os.path.isabs(request.AudioPath):
+            # get base path of modelFile
+            modelFileBase = os.path.dirname(request.ModelFile)
+            # modify LoraAdapter to be relative to modelFileBase
+            self.AudioPath = os.path.join(modelFileBase, request.AudioPath)
+
+        try:
+            print("Preparing KittenTTS model, please wait", file=sys.stderr)
+            # Use the model name from request.Model, defaulting to "KittenML/kitten-tts-nano-0.1" if not specified
+            model_name = request.Model if request.Model else "KittenML/kitten-tts-nano-0.1"
+            self.tts = KittenTTS(model_name)
+        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 TTS(self, request, context):
+        try:
+            # KittenTTS doesn't use language parameter like TTS, so we ignore it
+            # For multi-speaker models, use voice parameter
+            voice = request.voice if request.voice else "expr-voice-2-f"
+            
+            # Generate audio using KittenTTS
+            audio = self.tts.generate(request.text, voice=voice)
+            
+            # Save the audio using soundfile
+            sf.write(request.dst, audio, 24000)
+            
+        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),
+        options=[
+            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
+        ])
+    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)

backend/python/kitten-tts/install.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+set -e
+
+backend_dir=$(dirname $0)
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+# This is here because the Intel pip index is broken and returns 200 status codes for every package name, it just doesn't return any package links.
+# This makes uv think that the package exists in the Intel pip index, and by default it stops looking at other pip indexes once it finds a match.
+# We need uv to continue falling through to the pypi default index to find optimum[openvino] in the pypi index
+# the --upgrade actually allows us to *downgrade* torch to the version provided in the Intel pip index
+if [ "x${BUILD_PROFILE}" == "xintel" ]; then
+    EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
+fi
+
+installRequirements

backend/python/kitten-tts/requirements.txt

@@ -0,0 +1,5 @@
+grpcio==1.71.0
+protobuf
+certifi
+packaging==24.1
+https://github.com/KittenML/KittenTTS/releases/download/0.1/kittentts-0.1.0-py3-none-any.whl

backend/python/kitten-tts/run.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+backend_dir=$(dirname $0)
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+startBackend $@

backend/python/kitten-tts/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="tts_models/en/vctk/vits"))
+                print(response)
+                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="tts_models/en/vctk/vits"))
+                self.assertTrue(response.success)
+                tts_request = backend_pb2.TTSRequest(text="80s TV news production music hit for tonight's biggest story")
+                tts_response = stub.TTS(tts_request)
+                self.assertIsNotNone(tts_response)
+        except Exception as err:
+            print(err)
+            self.fail("TTS service failed")
+        finally:
+            self.tearDown()

backend/python/kitten-tts/test.sh

@@ -8,4 +8,4 @@ else
     source $backend_dir/../common/libbackend.sh
 fi
 
-python3 -m grpc_tools.protoc -I../.. -I./ --python_out=. --grpc_python_out=. backend.proto
+runUnittests

backend/python/kokoro/Makefile

@@ -1,20 +1,23 @@
-.DEFAULT_GOAL := install
-
-.PHONY: install
-install:
+.PHONY: kokoro
+kokoro:
 	bash install.sh
-	$(MAKE) protogen
 
-.PHONY: protogen
-protogen: backend_pb2_grpc.py backend_pb2.py
+.PHONY: run
+run: kokoro
+	@echo "Running kokoro..."
+	bash run.sh
+	@echo "kokoro run."
+
+.PHONY: test
+test: kokoro
+	@echo "Testing kokoro..."
+	bash test.sh
+	@echo "kokoro tested."
 
 .PHONY: protogen-clean
 protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
-backend_pb2_grpc.py backend_pb2.py:
-	bash protogen.sh
-
 .PHONY: clean
 clean: protogen-clean
 	rm -rf venv __pycache__

backend/python/kokoro/README.md

@@ -0,0 +1,23 @@
+# Kokoro TTS Backend for LocalAI
+
+This is a gRPC server backend for LocalAI that uses the Kokoro TTS pipeline.
+
+## Creating a separate environment for kokoro project
+
+```bash
+make kokoro
+```
+
+## Testing the gRPC server
+
+```bash
+make test
+```
+
+## Features
+
+- Lightweight TTS model with 82 million parameters
+- Apache-licensed weights
+- Fast and cost-efficient
+- Multi-language support
+- Multiple voice options

backend/python/kokoro/backend.py

@@ -1,101 +1,81 @@
 #!/usr/bin/env python3
 """
-Extra gRPC server for Kokoro models.
+This is an extra gRPC server of LocalAI for Kokoro TTS
 """
 from concurrent import futures
-
+import time
 import argparse
 import signal
 import sys
 import os
-import time
 import backend_pb2
 import backend_pb2_grpc
+
+import torch
+from kokoro import KPipeline
 import soundfile as sf
+
 import grpc
 
-from models import build_model
-from kokoro import generate
-import torch
 
-SAMPLE_RATE = 22050
 _ONE_DAY_IN_SECONDS = 60 * 60 * 24
 
 # If MAX_WORKERS are specified in the environment use it, otherwise default to 1
 MAX_WORKERS = int(os.environ.get('PYTHON_GRPC_MAX_WORKERS', '1'))
+KOKORO_LANG_CODE = os.environ.get('KOKORO_LANG_CODE', 'a')
 
 # Implement the BackendServicer class with the service methods
 class BackendServicer(backend_pb2_grpc.BackendServicer):
     """
-    A gRPC servicer for the backend service.
-
-    This class implements the gRPC methods for the backend service, including Health, LoadModel, and Embedding.
+    BackendServicer is the class that implements the gRPC service
     """
     def Health(self, request, context):
-        """
-        A gRPC method that returns the health status of the backend service.
-
-        Args:
-            request: A HealthRequest object that contains the request parameters.
-            context: A grpc.ServicerContext object that provides information about the RPC.
-
-        Returns:
-            A Reply object that contains the health status of the backend service.
-        """
         return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
-
+    
     def LoadModel(self, request, context):
-        """
-        A gRPC method that loads a model into memory.
-
-        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:
-            device = "cuda:0" if torch.cuda.is_available() else "cpu"
-            self.MODEL = build_model(request.ModelFile, device)
+            print("Preparing Kokoro TTS pipeline, please wait", file=sys.stderr)
+            # empty dict
+            self.options = {}
             options = request.Options
-            # Find the voice from the options, options are a list of strings in this form optname:optvalue:
-            VOICE_NAME = None
+            # The options are a list of strings in this form optname:optvalue
+            # We are storing all the options in a dict so we can use it later when
+            # generating the images
             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)
+                if ":" not in opt:
+                    continue
+                key, value = opt.split(":")
+                self.options[key] = value
 
-            self.VOICE_NAME = VOICE_NAME
-
-            print(f'Loaded voice: {VOICE_NAME}')
+            # Initialize Kokoro pipeline with language code
+            lang_code = self.options.get("lang_code", KOKORO_LANG_CODE)
+            self.pipeline = KPipeline(lang_code=lang_code)
+            print(f"Kokoro TTS pipeline loaded with language code: {lang_code}", file=sys.stderr)
         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)
+        
+        return backend_pb2.Result(message="Kokoro TTS pipeline 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:
-            audio, out_ps = generate(self.MODEL, request.text, self.VOICEPACK, lang=self.VOICE_NAME)
-            print(out_ps)
-            sf.write(request.dst, audio, SAMPLE_RATE)
+            # Get voice from request, default to 'af_heart' if not specified
+            voice = request.voice if request.voice else 'af_heart'
+            
+            # Generate audio using Kokoro pipeline
+            generator = self.pipeline(request.text, voice=voice)
+            
+            # Get the first (and typically only) audio segment
+            for i, (gs, ps, audio) in enumerate(generator):
+                # Save audio to the destination file
+                sf.write(request.dst, audio, 24000)
+                print(f"Generated audio segment {i}: gs={gs}, ps={ps}", file=sys.stderr)
+                # For now, we only process the first segment
+                # If you need to handle multiple segments, you might want to modify this
+                break
+                
         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):
@@ -108,11 +88,11 @@ def serve(address):
     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("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("Received termination signal. Shutting down...")
         server.stop(0)
         sys.exit(0)
 
@@ -132,5 +112,5 @@ 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)
+
     serve(args.addr)

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 integer part n can be ignored
-        # because 2 * np.pi * n doesn't affect phase
-        rad_values = (f0_values / self.sampling_rate) % 1
-
-        # initial phase noise (no noise for fundamental component)
-        rand_ini = torch.rand(f0_values.shape[0], f0_values.shape[2], \
-                              device=f0_values.device)
-        rand_ini[:, 0] = 0
-        rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
-
-        # instantanouse phase sine[t] = sin(2*pi \sum_i=1 ^{t} rad)
-        if not self.flag_for_pulse:
-#             # for normal case
-
-#             # To prevent torch.cumsum numerical overflow,
-#             # it is necessary to add -1 whenever \sum_k=1^n rad_value_k > 1.
-#             # Buffer tmp_over_one_idx indicates the time step to add -1.
-#             # This will not change F0 of sine because (x-1) * 2*pi = x * 2*pi
-#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
-#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
-#             cumsum_shift = torch.zeros_like(rad_values)
-#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
-
-#             phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
-            rad_values = torch.nn.functional.interpolate(rad_values.transpose(1, 2), 
-                                                         scale_factor=1/self.upsample_scale, 
-                                                         mode="linear").transpose(1, 2)
-    
-#             tmp_over_one = torch.cumsum(rad_values, 1) % 1
-#             tmp_over_one_idx = (padDiff(tmp_over_one)) < 0
-#             cumsum_shift = torch.zeros_like(rad_values)
-#             cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
-    
-            phase = torch.cumsum(rad_values, dim=1) * 2 * np.pi
-            phase = torch.nn.functional.interpolate(phase.transpose(1, 2) * self.upsample_scale, 
-                                                    scale_factor=self.upsample_scale, mode="linear").transpose(1, 2)
-            sines = torch.sin(phase)
-            
-        else:
-            # If necessary, make sure that the first time step of every
-            # voiced segments is sin(pi) or cos(0)
-            # This is used for pulse-train generation
-
-            # identify the last time step in unvoiced segments
-            uv = self._f02uv(f0_values)
-            uv_1 = torch.roll(uv, shifts=-1, dims=1)
-            uv_1[:, -1, :] = 1
-            u_loc = (uv < 1) * (uv_1 > 0)
-
-            # get the instantanouse phase
-            tmp_cumsum = torch.cumsum(rad_values, dim=1)
-            # different batch needs to be processed differently
-            for idx in range(f0_values.shape[0]):
-                temp_sum = tmp_cumsum[idx, u_loc[idx, :, 0], :]
-                temp_sum[1:, :] = temp_sum[1:, :] - temp_sum[0:-1, :]
-                # stores the accumulation of i.phase within
-                # each voiced segments
-                tmp_cumsum[idx, :, :] = 0
-                tmp_cumsum[idx, u_loc[idx, :, 0], :] = temp_sum
-
-            # rad_values - tmp_cumsum: remove the accumulation of i.phase
-            # within the previous voiced segment.
-            i_phase = torch.cumsum(rad_values - tmp_cumsum, dim=1)
-
-            # get the sines
-            sines = torch.cos(i_phase * 2 * np.pi)
-        return sines
-
-    def forward(self, f0):
-        """ sine_tensor, uv = forward(f0)
-        input F0: tensor(batchsize=1, length, dim=1)
-                  f0 for unvoiced steps should be 0
-        output sine_tensor: tensor(batchsize=1, length, dim)
-        output uv: tensor(batchsize=1, length, 1)
-        """
-        f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim,
-                             device=f0.device)
-        # fundamental component
-        fn = torch.multiply(f0, torch.FloatTensor([[range(1, self.harmonic_num + 2)]]).to(f0.device))
-
-        # generate sine waveforms
-        sine_waves = self._f02sine(fn) * self.sine_amp
-
-        # generate uv signal
-        # uv = torch.ones(f0.shape)
-        # uv = uv * (f0 > self.voiced_threshold)
-        uv = self._f02uv(f0)
-
-        # noise: for unvoiced should be similar to sine_amp
-        #        std = self.sine_amp/3 -> max value ~ self.sine_amp
-        # .       for voiced regions is self.noise_std
-        noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
-        noise = noise_amp * torch.randn_like(sine_waves)
-
-        # first: set the unvoiced part to 0 by uv
-        # then: additive noise
-        sine_waves = sine_waves * uv + noise
-        return sine_waves, uv, noise
-
-
-class SourceModuleHnNSF(torch.nn.Module):
-    """ SourceModule for hn-nsf
-    SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
-                 add_noise_std=0.003, voiced_threshod=0)
-    sampling_rate: sampling_rate in Hz
-    harmonic_num: number of harmonic above F0 (default: 0)
-    sine_amp: amplitude of sine source signal (default: 0.1)
-    add_noise_std: std of additive Gaussian noise (default: 0.003)
-        note that amplitude of noise in unvoiced is decided
-        by sine_amp
-    voiced_threshold: threhold to set U/V given F0 (default: 0)
-    Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
-    F0_sampled (batchsize, length, 1)
-    Sine_source (batchsize, length, 1)
-    noise_source (batchsize, length 1)
-    uv (batchsize, length, 1)
-    """
-
-    def __init__(self, sampling_rate, upsample_scale, harmonic_num=0, sine_amp=0.1,
-                 add_noise_std=0.003, voiced_threshod=0):
-        super(SourceModuleHnNSF, self).__init__()
-
-        self.sine_amp = sine_amp
-        self.noise_std = add_noise_std
-
-        # to produce sine waveforms
-        self.l_sin_gen = SineGen(sampling_rate, upsample_scale, harmonic_num,
-                                 sine_amp, add_noise_std, voiced_threshod)
-
-        # to merge source harmonics into a single excitation
-        self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
-        self.l_tanh = torch.nn.Tanh()
-
-    def forward(self, x):
-        """
-        Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
-        F0_sampled (batchsize, length, 1)
-        Sine_source (batchsize, length, 1)
-        noise_source (batchsize, length 1)
-        """
-        # source for harmonic branch
-        with torch.no_grad():
-            sine_wavs, uv, _ = self.l_sin_gen(x)
-        sine_merge = self.l_tanh(self.l_linear(sine_wavs))
-
-        # source for noise branch, in the same shape as uv
-        noise = torch.randn_like(uv) * self.sine_amp / 3
-        return sine_merge, noise, uv
-def padDiff(x):
-    return F.pad(F.pad(x, (0,0,-1,1), 'constant', 0) - x, (0,0,0,-1), 'constant', 0)
-
-    
-class Generator(torch.nn.Module):
-    def __init__(self, style_dim, resblock_kernel_sizes, upsample_rates, upsample_initial_channel, resblock_dilation_sizes, upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size):
-        super(Generator, self).__init__()
-
-        self.num_kernels = len(resblock_kernel_sizes)
-        self.num_upsamples = len(upsample_rates)
-        resblock = AdaINResBlock1
-
-        self.m_source = SourceModuleHnNSF(
-                    sampling_rate=24000,
-                    upsample_scale=np.prod(upsample_rates) * gen_istft_hop_size,
-                    harmonic_num=8, voiced_threshod=10)
-        self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates) * gen_istft_hop_size)
-        self.noise_convs = nn.ModuleList()
-        self.noise_res = nn.ModuleList()
-        
-        self.ups = nn.ModuleList()
-        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
-            self.ups.append(weight_norm(
-                ConvTranspose1d(upsample_initial_channel//(2**i), upsample_initial_channel//(2**(i+1)),
-                                k, u, padding=(k-u)//2)))
-
-        self.resblocks = nn.ModuleList()
-        for i in range(len(self.ups)):
-            ch = upsample_initial_channel//(2**(i+1))
-            for j, (k, d) in enumerate(zip(resblock_kernel_sizes,resblock_dilation_sizes)):
-                self.resblocks.append(resblock(ch, k, d, style_dim))
-                
-            c_cur = upsample_initial_channel // (2 ** (i + 1))
-            
-            if i + 1 < len(upsample_rates):  #
-                stride_f0 = np.prod(upsample_rates[i + 1:])
-                self.noise_convs.append(Conv1d(
-                    gen_istft_n_fft + 2, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=(stride_f0+1) // 2))
-                self.noise_res.append(resblock(c_cur, 7, [1,3,5], style_dim))
-            else:
-                self.noise_convs.append(Conv1d(gen_istft_n_fft + 2, c_cur, kernel_size=1))
-                self.noise_res.append(resblock(c_cur, 11, [1,3,5], style_dim))
-                
-                
-        self.post_n_fft = gen_istft_n_fft
-        self.conv_post = weight_norm(Conv1d(ch, self.post_n_fft + 2, 7, 1, padding=3))
-        self.ups.apply(init_weights)
-        self.conv_post.apply(init_weights)
-        self.reflection_pad = torch.nn.ReflectionPad1d((1, 0))
-        self.stft = TorchSTFT(filter_length=gen_istft_n_fft, hop_length=gen_istft_hop_size, win_length=gen_istft_n_fft)
-        
-        
-    def forward(self, x, s, f0):
-        with torch.no_grad():
-            f0 = self.f0_upsamp(f0[:, None]).transpose(1, 2)  # bs,n,t
-
-            har_source, noi_source, uv = self.m_source(f0)
-            har_source = har_source.transpose(1, 2).squeeze(1)
-            har_spec, har_phase = self.stft.transform(har_source)
-            har = torch.cat([har_spec, har_phase], dim=1)
-        
-        for i in range(self.num_upsamples):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            x_source = self.noise_convs[i](har)
-            x_source = self.noise_res[i](x_source, s)
-
-            x = self.ups[i](x)
-            if i == self.num_upsamples - 1:
-                x = self.reflection_pad(x)
-
-            x = x + x_source
-            xs = None
-            for j in range(self.num_kernels):
-                if xs is None:
-                    xs = self.resblocks[i*self.num_kernels+j](x, s)
-                else:
-                    xs += self.resblocks[i*self.num_kernels+j](x, s)
-            x = xs / self.num_kernels
-        x = F.leaky_relu(x)
-        x = self.conv_post(x)
-        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
-        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
-        return self.stft.inverse(spec, phase)
-    
-    def fw_phase(self, x, s):
-        for i in range(self.num_upsamples):
-            x = F.leaky_relu(x, LRELU_SLOPE)
-            x = self.ups[i](x)
-            xs = None
-            for j in range(self.num_kernels):
-                if xs is None:
-                    xs = self.resblocks[i*self.num_kernels+j](x, s)
-                else:
-                    xs += self.resblocks[i*self.num_kernels+j](x, s)
-            x = xs / self.num_kernels
-        x = F.leaky_relu(x)
-        x = self.reflection_pad(x)
-        x = self.conv_post(x)
-        spec = torch.exp(x[:,:self.post_n_fft // 2 + 1, :])
-        phase = torch.sin(x[:, self.post_n_fft // 2 + 1:, :])
-        return spec, phase
-
-    def remove_weight_norm(self):
-        print('Removing weight norm...')
-        for l in self.ups:
-            remove_weight_norm(l)
-        for l in self.resblocks:
-            l.remove_weight_norm()
-        remove_weight_norm(self.conv_pre)
-        remove_weight_norm(self.conv_post)
-
-        
-class AdainResBlk1d(nn.Module):
-    def __init__(self, dim_in, dim_out, style_dim=64, actv=nn.LeakyReLU(0.2),
-                 upsample='none', dropout_p=0.0):
-        super().__init__()
-        self.actv = actv
-        self.upsample_type = upsample
-        self.upsample = UpSample1d(upsample)
-        self.learned_sc = dim_in != dim_out
-        self._build_weights(dim_in, dim_out, style_dim)
-        self.dropout = nn.Dropout(dropout_p)
-        
-        if upsample == 'none':
-            self.pool = nn.Identity()
-        else:
-            self.pool = weight_norm(nn.ConvTranspose1d(dim_in, dim_in, kernel_size=3, stride=2, groups=dim_in, padding=1, output_padding=1))
-        
-        
-    def _build_weights(self, dim_in, dim_out, style_dim):
-        self.conv1 = weight_norm(nn.Conv1d(dim_in, dim_out, 3, 1, 1))
-        self.conv2 = weight_norm(nn.Conv1d(dim_out, dim_out, 3, 1, 1))
-        self.norm1 = AdaIN1d(style_dim, dim_in)
-        self.norm2 = AdaIN1d(style_dim, dim_out)
-        if self.learned_sc:
-            self.conv1x1 = weight_norm(nn.Conv1d(dim_in, dim_out, 1, 1, 0, bias=False))
-
-    def _shortcut(self, x):
-        x = self.upsample(x)
-        if self.learned_sc:
-            x = self.conv1x1(x)
-        return x
-
-    def _residual(self, x, s):
-        x = self.norm1(x, s)
-        x = self.actv(x)
-        x = self.pool(x)
-        x = self.conv1(self.dropout(x))
-        x = self.norm2(x, s)
-        x = self.actv(x)
-        x = self.conv2(self.dropout(x))
-        return x
-
-    def forward(self, x, s):
-        out = self._residual(x, s)
-        out = (out + self._shortcut(x)) / np.sqrt(2)
-        return out
-    
-class UpSample1d(nn.Module):
-    def __init__(self, layer_type):
-        super().__init__()
-        self.layer_type = layer_type
-
-    def forward(self, x):
-        if self.layer_type == 'none':
-            return x
-        else:
-            return F.interpolate(x, scale_factor=2, mode='nearest')
-
-class Decoder(nn.Module):
-    def __init__(self, dim_in=512, F0_channel=512, style_dim=64, dim_out=80, 
-                resblock_kernel_sizes = [3,7,11],
-                upsample_rates = [10, 6],
-                upsample_initial_channel=512,
-                resblock_dilation_sizes=[[1,3,5], [1,3,5], [1,3,5]],
-                upsample_kernel_sizes=[20, 12], 
-                gen_istft_n_fft=20, gen_istft_hop_size=5):
-        super().__init__()
-        
-        self.decode = nn.ModuleList()
-        
-        self.encode = AdainResBlk1d(dim_in + 2, 1024, style_dim)
-        
-        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
-        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
-        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 1024, style_dim))
-        self.decode.append(AdainResBlk1d(1024 + 2 + 64, 512, style_dim, upsample=True))
-
-        self.F0_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
-        
-        self.N_conv = weight_norm(nn.Conv1d(1, 1, kernel_size=3, stride=2, groups=1, padding=1))
-        
-        self.asr_res = nn.Sequential(
-            weight_norm(nn.Conv1d(512, 64, kernel_size=1)),
-        )
-        
-        
-        self.generator = Generator(style_dim, resblock_kernel_sizes, upsample_rates, 
-                                   upsample_initial_channel, resblock_dilation_sizes, 
-                                   upsample_kernel_sizes, gen_istft_n_fft, gen_istft_hop_size)
-        
-    def forward(self, asr, F0_curve, N, s):
-        F0 = self.F0_conv(F0_curve.unsqueeze(1))
-        N = self.N_conv(N.unsqueeze(1))
-        
-        x = torch.cat([asr, F0, N], axis=1)
-        x = self.encode(x, s)
-        
-        asr_res = self.asr_res(asr)
-        
-        res = True
-        for block in self.decode:
-            if res:
-                x = torch.cat([x, asr_res, F0, N], axis=1)
-            x = block(x, s)
-            if block.upsample_type != "none":
-                res = False
-                
-        x = self.generator(x, s, F0_curve)
-        return x

backend/python/kokoro/kokoro.py

@@ -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

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

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

backend/python/kokoro/requirements-cpu.txt

@@ -1,2 +1,6 @@
-torch==2.4.1
-transformers
+--extra-index-url https://download.pytorch.org/whl/cpu
+transformers
+accelerate
+torch
+kokoro
+soundfile

backend/python/kokoro/requirements-cublas11.txt

@@ -1,3 +1,7 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
-torch==2.4.1+cu118
-transformers
+torch==2.7.1+cu118
+torchaudio==2.7.1+cu118
+transformers
+accelerate
+kokoro
+soundfile

backend/python/kokoro/requirements-cublas12.txt

@@ -1,2 +1,6 @@
-torch==2.4.1
-transformers
+torch==2.7.1
+torchaudio==2.7.1
+transformers
+accelerate
+kokoro
+soundfile

backend/python/kokoro/requirements-hipblas.txt

@@ -1,3 +1,7 @@
---extra-index-url https://download.pytorch.org/whl/rocm6.0
-torch==2.4.1+rocm6.0
-transformers
+--extra-index-url https://download.pytorch.org/whl/rocm6.3
+torch==2.7.1+rocm6.3
+torchaudio==2.7.1+rocm6.3
+transformers
+accelerate
+kokoro
+soundfile

backend/python/kokoro/requirements-intel.txt

@@ -1,5 +1,11 @@
 --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
+intel-extension-for-pytorch==2.8.10+xpu
+torch==2.5.1+cxx11.abi
+oneccl_bind_pt==2.8.0+xpu
+torchaudio==2.5.1+cxx11.abi
+optimum[openvino]
+setuptools
+transformers==4.48.3
+accelerate
+kokoro
+soundfile

backend/python/kokoro/requirements.txt

@@ -1,7 +1,6 @@
 grpcio==1.71.0
 protobuf
-phonemizer
-scipy
-munch
-setuptools
-soundfile
+certifi
+packaging==24.1
+pip
+chardet

backend/python/kokoro/test.py

@@ -0,0 +1,87 @@
+"""
+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 Kokoro pipeline 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(language="a"))
+                print(response)
+                self.assertTrue(response.success)
+                self.assertEqual(response.message, "Kokoro TTS pipeline 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 TTS generation works successfully
+        """
+        try:
+            self.setUp()
+            with grpc.insecure_channel("localhost:50051") as channel:
+                stub = backend_pb2_grpc.BackendStub(channel)
+                response = stub.LoadModel(backend_pb2.ModelOptions(language="a"))
+                self.assertTrue(response.success)
+                tts_request = backend_pb2.TTSRequest(
+                    text="Kokoro is an open-weight TTS model with 82 million parameters.",
+                    voice="af_heart",
+                    dst="test_output.wav"
+                )
+                tts_response = stub.TTS(tts_request)
+                self.assertIsNotNone(tts_response)
+                self.assertTrue(tts_response.success)
+        except Exception as err:
+            print(err)
+            self.fail("TTS service failed")
+        finally:
+            self.tearDown()

backend/python/mlx-vlm/Makefile

@@ -0,0 +1,23 @@
+.PHONY: mlx-vlm
+mlx-vlm:
+	bash install.sh
+
+.PHONY: run
+run: mlx-vlm
+	@echo "Running mlx-vlm..."
+	bash run.sh
+	@echo "mlx run."
+
+.PHONY: test
+test: mlx-vlm
+	@echo "Testing mlx-vlm..."
+	bash test.sh
+	@echo "mlx tested."
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__

backend/python/mlx-vlm/backend.py

@@ -0,0 +1,477 @@
+#!/usr/bin/env python3
+import asyncio
+from concurrent import futures
+import argparse
+import signal
+import sys
+import os
+from typing import List
+import time
+
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+from mlx_vlm import load, generate, stream_generate
+from mlx_vlm.prompt_utils import apply_chat_template
+from mlx_vlm.utils import load_config, load_image
+import mlx.core as mx
+import base64
+import io
+from PIL import Image
+import tempfile
+
+_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 that implements the Backend service defined in backend.proto.
+    """
+
+    def _is_float(self, s):
+        """Check if a string can be converted to float."""
+        try:
+            float(s)
+            return True
+        except ValueError:
+            return False
+
+    def _is_int(self, s):
+        """Check if a string can be converted to int."""
+        try:
+            int(s)
+            return True
+        except ValueError:
+            return False
+
+    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'))
+
+    async def LoadModel(self, request, context):
+        """
+        Loads a multimodal vision-language model using MLX-VLM.
+
+        Args:
+            request: The load model request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Result: The load model result.
+        """
+        try:
+            print(f"Loading MLX-VLM model: {request.Model}", file=sys.stderr)
+            print(f"Request: {request}", file=sys.stderr)
+            
+            # Parse options like in the diffusers backend
+            options = request.Options
+            self.options = {}
+            
+            # The options are a list of strings in this form optname:optvalue
+            # We store all the options in a dict for later use
+            for opt in options:
+                if ":" not in opt:
+                    continue
+                key, value = opt.split(":", 1)  # Split only on first colon to handle values with colons
+                
+                # Convert numeric values to appropriate types
+                if self._is_float(value):
+                    value = float(value)
+                elif self._is_int(value):
+                    value = int(value)
+                elif value.lower() in ["true", "false"]:
+                    value = value.lower() == "true"
+                    
+                self.options[key] = value
+            
+            print(f"Options: {self.options}", file=sys.stderr)
+            
+            # Load model and processor using MLX-VLM
+            # mlx-vlm load function returns (model, processor) instead of (model, tokenizer)
+            self.model, self.processor = load(request.Model)
+            
+            # Load model config for chat template support
+            self.config = load_config(request.Model)
+                
+        except Exception as err:
+            print(f"Error loading MLX-VLM model {err=}, {type(err)=}", file=sys.stderr)
+            return backend_pb2.Result(success=False, message=f"Error loading MLX-VLM model: {err}")
+
+        print("MLX-VLM model loaded successfully", file=sys.stderr)
+        return backend_pb2.Result(message="MLX-VLM model loaded successfully", success=True)
+
+    async def Predict(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters using MLX-VLM with multimodal support.
+
+        Args:
+            request: The predict request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Reply: The predict result.
+        """
+        temp_files = []
+        try:
+            # Process images and audios from request
+            image_paths = []
+            audio_paths = []
+            
+            # Process images
+            if request.Images:
+                for img_data in request.Images:
+                    img_path = self.load_image_from_base64(img_data)
+                    if img_path:
+                        image_paths.append(img_path)
+                        temp_files.append(img_path)
+            
+            # Process audios
+            if request.Audios:
+                for audio_data in request.Audios:
+                    audio_path = self.load_audio_from_base64(audio_data)
+                    if audio_path:
+                        audio_paths.append(audio_path)
+                        temp_files.append(audio_path)
+            
+            # Prepare the prompt with multimodal information
+            prompt = self._prepare_prompt(request, num_images=len(image_paths), num_audios=len(audio_paths))
+            
+            # Build generation parameters using request attributes and options
+            max_tokens, generation_params = self._build_generation_params(request)
+            
+            print(f"Generating text with MLX-VLM - max_tokens: {max_tokens}, params: {generation_params}", file=sys.stderr)
+            print(f"Images: {len(image_paths)}, Audios: {len(audio_paths)}", file=sys.stderr)
+            
+            # Generate text using MLX-VLM with multimodal inputs
+            response = generate(
+                model=self.model,
+                processor=self.processor,
+                prompt=prompt,
+                image=image_paths if image_paths else None,
+                audio=audio_paths if audio_paths else None,
+                max_tokens=max_tokens,
+                temperature=generation_params.get('temp', 0.6),
+                top_p=generation_params.get('top_p', 1.0),
+                verbose=False
+            )
+            
+            return backend_pb2.Reply(message=bytes(response, encoding='utf-8'))
+            
+        except Exception as e:
+            print(f"Error in MLX-VLM Predict: {e}", file=sys.stderr)
+            context.set_code(grpc.StatusCode.INTERNAL)
+            context.set_details(f"Generation failed: {str(e)}")
+            return backend_pb2.Reply(message=bytes("", encoding='utf-8'))
+        finally:
+            # Clean up temporary files
+            self.cleanup_temp_files(temp_files)
+
+    def Embedding(self, request, context):
+        """
+        A gRPC method that calculates embeddings for a given sentence.
+        
+        Note: MLX-VLM doesn't support embeddings directly. This method returns an error.
+
+        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.
+        """
+        print("Embeddings not supported in MLX-VLM backend", file=sys.stderr)
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details("Embeddings are not supported in the MLX-VLM backend.")
+        return backend_pb2.EmbeddingResult()
+
+    async def PredictStream(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters, and streams the results using MLX-VLM with multimodal support.
+
+        Args:
+            request: The predict stream request.
+            context: The gRPC context.
+
+        Yields:
+            backend_pb2.Reply: Streaming predict results.
+        """
+        temp_files = []
+        try:
+            # Process images and audios from request
+            image_paths = []
+            audio_paths = []
+            
+            # Process images
+            if request.Images:
+                for img_data in request.Images:
+                    img_path = self.load_image_from_base64(img_data)
+                    if img_path:
+                        image_paths.append(img_path)
+                        temp_files.append(img_path)
+            
+            # Process audios
+            if request.Audios:
+                for audio_data in request.Audios:
+                    audio_path = self.load_audio_from_base64(audio_data)
+                    if audio_path:
+                        audio_paths.append(audio_path)
+                        temp_files.append(audio_path)
+            
+            # Prepare the prompt with multimodal information
+            prompt = self._prepare_prompt(request, num_images=len(image_paths), num_audios=len(audio_paths))
+            
+            # Build generation parameters using request attributes and options
+            max_tokens, generation_params = self._build_generation_params(request, default_max_tokens=512)
+            
+            print(f"Streaming text with MLX-VLM - max_tokens: {max_tokens}, params: {generation_params}", file=sys.stderr)
+            print(f"Images: {len(image_paths)}, Audios: {len(audio_paths)}", file=sys.stderr)
+            
+            # Stream text generation using MLX-VLM with multimodal inputs
+            for response in stream_generate(
+                model=self.model,
+                processor=self.processor,
+                prompt=prompt,
+                image=image_paths if image_paths else None,
+                audio=audio_paths if audio_paths else None,
+                max_tokens=max_tokens,
+                temperature=generation_params.get('temp', 0.6),
+                top_p=generation_params.get('top_p', 1.0),
+            ):
+                yield backend_pb2.Reply(message=bytes(response.text, encoding='utf-8'))
+                
+        except Exception as e:
+            print(f"Error in MLX-VLM PredictStream: {e}", file=sys.stderr)
+            context.set_code(grpc.StatusCode.INTERNAL)
+            context.set_details(f"Streaming generation failed: {str(e)}")
+            yield backend_pb2.Reply(message=bytes("", encoding='utf-8'))
+        finally:
+            # Clean up temporary files
+            self.cleanup_temp_files(temp_files)
+
+    def _prepare_prompt(self, request, num_images=0, num_audios=0):
+        """
+        Prepare the prompt for MLX-VLM generation, handling chat templates and multimodal inputs.
+
+        Args:
+            request: The gRPC request containing prompt and message information.
+            num_images: Number of images in the request.
+            num_audios: Number of audio files in the request.
+
+        Returns:
+            str: The prepared prompt.
+        """
+        # If tokenizer template is enabled and messages are provided instead of prompt, apply the tokenizer template
+        if not request.Prompt and request.UseTokenizerTemplate and request.Messages:
+            # Convert gRPC messages to the format expected by apply_chat_template
+            messages = []
+            for msg in request.Messages:
+                messages.append({"role": msg.role, "content": msg.content})
+            
+            # Use mlx-vlm's apply_chat_template which handles multimodal inputs
+            prompt = apply_chat_template(
+                self.processor,
+                self.config, 
+                messages,
+                num_images=num_images,
+                num_audios=num_audios
+            )
+            return prompt
+        elif request.Prompt:
+            # If we have a direct prompt but also have images/audio, we need to format it properly
+            if num_images > 0 or num_audios > 0:
+                # Create a simple message structure for multimodal prompt
+                messages = [{"role": "user", "content": request.Prompt}]
+                prompt = apply_chat_template(
+                    self.processor,
+                    self.config, 
+                    messages,
+                    num_images=num_images,
+                    num_audios=num_audios
+                )
+                return prompt
+            else:
+                return request.Prompt
+        else:
+            # Fallback to empty prompt with multimodal template if we have media
+            if num_images > 0 or num_audios > 0:
+                messages = [{"role": "user", "content": ""}]
+                prompt = apply_chat_template(
+                    self.processor,
+                    self.config, 
+                    messages,
+                    num_images=num_images,
+                    num_audios=num_audios
+                )
+                return prompt
+            else:
+                return ""
+
+
+
+
+
+    def _build_generation_params(self, request, default_max_tokens=200):
+        """
+        Build generation parameters from request attributes and options for MLX-VLM.
+
+        Args:
+            request: The gRPC request.
+            default_max_tokens: Default max_tokens if not specified.
+
+        Returns:
+            tuple: (max_tokens, generation_params dict)
+        """
+        # Extract max_tokens
+        max_tokens = getattr(request, 'Tokens', default_max_tokens)
+        if max_tokens == 0:
+            max_tokens = default_max_tokens
+        
+        # Extract generation parameters from request attributes
+        temp = getattr(request, 'Temperature', 0.0)
+        if temp == 0.0:
+            temp = 0.6  # Default temperature
+        
+        top_p = getattr(request, 'TopP', 0.0)
+        if top_p == 0.0:
+            top_p = 1.0  # Default top_p
+        
+        # Initialize generation parameters for MLX-VLM
+        generation_params = {
+            'temp': temp,
+            'top_p': top_p,
+        }
+        
+        # Add seed if specified
+        seed = getattr(request, 'Seed', 0)
+        if seed != 0:
+            mx.random.seed(seed)
+        
+        # Override with options if available
+        if hasattr(self, 'options'):
+            # Max tokens from options
+            if 'max_tokens' in self.options:
+                max_tokens = self.options['max_tokens']
+            
+            # Generation parameters from options
+            param_option_mapping = {
+                'temp': 'temp',
+                'temperature': 'temp',  # alias
+                'top_p': 'top_p', 
+            }
+            
+            for option_key, param_key in param_option_mapping.items():
+                if option_key in self.options:
+                    generation_params[param_key] = self.options[option_key]
+            
+            # Handle seed from options
+            if 'seed' in self.options:
+                mx.random.seed(self.options['seed'])
+        
+        return max_tokens, generation_params
+
+    def load_image_from_base64(self, image_data: str):
+        """
+        Load an image from base64 encoded data.
+
+        Args:
+            image_data (str): Base64 encoded image data.
+
+        Returns:
+            PIL.Image or str: The loaded image or path to the image.
+        """
+        try:
+            decoded_data = base64.b64decode(image_data)
+            image = Image.open(io.BytesIO(decoded_data))
+            
+            # Save to temporary file for mlx-vlm
+            with tempfile.NamedTemporaryFile(delete=False, suffix='.jpg') as tmp_file:
+                image.save(tmp_file.name, format='JPEG')
+                return tmp_file.name
+                
+        except Exception as e:
+            print(f"Error loading image from base64: {e}", file=sys.stderr)
+            return None
+
+    def load_audio_from_base64(self, audio_data: str):
+        """
+        Load audio from base64 encoded data.
+
+        Args:
+            audio_data (str): Base64 encoded audio data.
+
+        Returns:
+            str: Path to the loaded audio file.
+        """
+        try:
+            decoded_data = base64.b64decode(audio_data)
+            
+            # Save to temporary file for mlx-vlm
+            with tempfile.NamedTemporaryFile(delete=False, suffix='.wav') as tmp_file:
+                tmp_file.write(decoded_data)
+                return tmp_file.name
+                
+        except Exception as e:
+            print(f"Error loading audio from base64: {e}", file=sys.stderr)
+            return None
+
+    def cleanup_temp_files(self, file_paths: List[str]):
+        """
+        Clean up temporary files.
+
+        Args:
+            file_paths (List[str]): List of file paths to clean up.
+        """
+        for file_path in file_paths:
+            try:
+                if file_path and os.path.exists(file_path):
+                    os.remove(file_path)
+            except Exception as e:
+                print(f"Error removing temporary file {file_path}: {e}", file=sys.stderr)
+
+async def serve(address):
+    # Start asyncio gRPC server
+    server = grpc.aio.server(migration_thread_pool=futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
+        options=[
+            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
+        ])
+    # Add the servicer to the server
+    backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
+    # Bind the server to the address
+    server.add_insecure_port(address)
+
+    # Gracefully shutdown the server on SIGTERM or SIGINT
+    loop = asyncio.get_event_loop()
+    for sig in (signal.SIGINT, signal.SIGTERM):
+        loop.add_signal_handler(
+            sig, lambda: asyncio.ensure_future(server.stop(5))
+        )
+
+    # Start the server
+    await server.start()
+    print("Server started. Listening on: " + address, file=sys.stderr)
+    # Wait for the server to be terminated
+    await server.wait_for_termination()
+
+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()
+
+    asyncio.run(serve(args.addr))

backend/python/mlx-vlm/install.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+set -e
+
+USE_PIP=true
+
+backend_dir=$(dirname $0)
+
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+installRequirements

backend/python/mlx-vlm/requirements-mps.txt

@@ -0,0 +1 @@
+git+https://github.com/Blaizzy/mlx-vlm

backend/python/mlx-vlm/requirements.txt

@@ -0,0 +1,4 @@
+grpcio==1.71.0
+protobuf
+certifi
+setuptools

backend/python/mlx-vlm/run.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+
+backend_dir=$(dirname $0)
+
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+startBackend $@

backend/python/mlx-vlm/test.py

@@ -0,0 +1,146 @@
+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()
+
+    def test_sampling_params(self):
+        """
+        This method tests if all sampling parameters are correctly processed
+        NOTE: this does NOT test for correctness, just that we received a compatible response
+        """
+        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",
+                    TopP=0.8,
+                    Tokens=50,
+                    Temperature=0.7,
+                    TopK=40,
+                    PresencePenalty=0.1,
+                    FrequencyPenalty=0.2,
+                    RepetitionPenalty=1.1,
+                    MinP=0.05,
+                    Seed=42,
+                    StopPrompts=["\n"],
+                    StopTokenIds=[50256],
+                    BadWords=["badword"],
+                    IncludeStopStrInOutput=True,
+                    IgnoreEOS=True,
+                    MinTokens=5,
+                    Logprobs=5,
+                    PromptLogprobs=5,
+                    SkipSpecialTokens=True,
+                    SpacesBetweenSpecialTokens=True,
+                    TruncatePromptTokens=10,
+                    GuidedDecoding=True,
+                    N=2,
+                )
+                resp = stub.Predict(req)
+                self.assertIsNotNone(resp.message)
+                self.assertIsNotNone(resp.logprobs)
+        except Exception as err:
+            print(err)
+            self.fail("sampling params service failed")
+        finally:
+            self.tearDown()
+
+
+    def test_embedding(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="intfloat/e5-mistral-7b-instruct"))
+                self.assertTrue(response.success)
+                embedding_request = backend_pb2.PredictOptions(Embeddings="This is a test sentence.")
+                embedding_response = stub.Embedding(embedding_request)
+                self.assertIsNotNone(embedding_response.embeddings)
+                # assert that is a list of floats
+                self.assertIsInstance(embedding_response.embeddings, list)
+                # assert that the list is not empty
+                self.assertTrue(len(embedding_response.embeddings) > 0)
+        except Exception as err:
+            print(err)
+            self.fail("Embedding service failed")
+        finally:
+            self.tearDown()

backend/python/mlx-vlm/test.sh

@@ -0,0 +1,12 @@
+#!/bin/bash
+set -e
+
+backend_dir=$(dirname $0)
+
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+runUnittests

backend/python/mlx/Makefile

@@ -0,0 +1,23 @@
+.PHONY: mlx
+mlx:
+	bash install.sh
+
+.PHONY: run
+run:
+	@echo "Running mlx..."
+	bash run.sh
+	@echo "mlx run."
+
+.PHONY: test
+test:
+	@echo "Testing mlx..."
+	bash test.sh
+	@echo "mlx tested."
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__

backend/python/mlx/backend.py

@@ -0,0 +1,376 @@
+#!/usr/bin/env python3
+import asyncio
+from concurrent import futures
+import argparse
+import signal
+import sys
+import os
+from typing import List
+import time
+
+import backend_pb2
+import backend_pb2_grpc
+
+import grpc
+from mlx_lm import load, generate, stream_generate
+from mlx_lm.sample_utils import make_sampler
+from mlx_lm.models.cache import make_prompt_cache
+import mlx.core as mx
+import base64
+import io
+
+_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 that implements the Backend service defined in backend.proto.
+    """
+
+    def _is_float(self, s):
+        """Check if a string can be converted to float."""
+        try:
+            float(s)
+            return True
+        except ValueError:
+            return False
+
+    def _is_int(self, s):
+        """Check if a string can be converted to int."""
+        try:
+            int(s)
+            return True
+        except ValueError:
+            return False
+
+    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'))
+
+    async def LoadModel(self, request, context):
+        """
+        Loads a language model using MLX.
+
+        Args:
+            request: The load model request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Result: The load model result.
+        """
+        try:
+            print(f"Loading MLX model: {request.Model}", file=sys.stderr)
+            print(f"Request: {request}", file=sys.stderr)
+            
+            # Parse options like in the diffusers backend
+            options = request.Options
+            self.options = {}
+            
+            # The options are a list of strings in this form optname:optvalue
+            # We store all the options in a dict for later use
+            for opt in options:
+                if ":" not in opt:
+                    continue
+                key, value = opt.split(":", 1)  # Split only on first colon to handle values with colons
+                
+                # Convert numeric values to appropriate types
+                if self._is_float(value):
+                    value = float(value)
+                elif self._is_int(value):
+                    value = int(value)
+                elif value.lower() in ["true", "false"]:
+                    value = value.lower() == "true"
+                    
+                self.options[key] = value
+            
+            print(f"Options: {self.options}", file=sys.stderr)
+            
+            # Build tokenizer config for MLX using options
+            tokenizer_config = {}
+            
+            # Handle trust_remote_code from request or options
+            if request.TrustRemoteCode or self.options.get("trust_remote_code", False):
+                tokenizer_config["trust_remote_code"] = True
+            
+            # Handle EOS token from options
+            if "eos_token" in self.options:
+                tokenizer_config["eos_token"] = self.options["eos_token"]
+            
+            # Handle other tokenizer config options
+            for key in ["pad_token", "bos_token", "unk_token", "sep_token", "cls_token", "mask_token"]:
+                if key in self.options:
+                    tokenizer_config[key] = self.options[key]
+            
+            # Load model and tokenizer using MLX
+            if tokenizer_config:
+                print(f"Loading with tokenizer_config: {tokenizer_config}", file=sys.stderr)
+                self.model, self.tokenizer = load(request.Model, tokenizer_config=tokenizer_config)
+            else:
+                self.model, self.tokenizer = load(request.Model)
+            
+            # Initialize prompt cache for efficient generation
+            max_kv_size = self.options.get("max_kv_size", None)
+            self.prompt_cache = make_prompt_cache(self.model, max_kv_size)
+                
+        except Exception as err:
+            print(f"Error loading MLX model {err=}, {type(err)=}", file=sys.stderr)
+            return backend_pb2.Result(success=False, message=f"Error loading MLX model: {err}")
+
+        print("MLX model loaded successfully", file=sys.stderr)
+        return backend_pb2.Result(message="MLX model loaded successfully", success=True)
+
+    async def Predict(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters using MLX.
+
+        Args:
+            request: The predict request.
+            context: The gRPC context.
+
+        Returns:
+            backend_pb2.Reply: The predict result.
+        """
+        try:
+            # Prepare the prompt
+            prompt = self._prepare_prompt(request)
+            
+            # Build generation parameters using request attributes and options
+            max_tokens, sampler_params = self._build_generation_params(request)
+            
+            print(f"Generating text with MLX - max_tokens: {max_tokens}, sampler_params: {sampler_params}", file=sys.stderr)
+            
+            # Create sampler with parameters
+            sampler = make_sampler(**sampler_params)
+            
+            # Generate text using MLX with proper parameters
+            response = generate(
+                self.model,
+                self.tokenizer,
+                prompt=prompt,
+                max_tokens=max_tokens,
+                sampler=sampler,
+                prompt_cache=self.prompt_cache,
+                verbose=False
+            )
+            
+            return backend_pb2.Reply(message=bytes(response, encoding='utf-8'))
+            
+        except Exception as e:
+            print(f"Error in MLX Predict: {e}", file=sys.stderr)
+            context.set_code(grpc.StatusCode.INTERNAL)
+            context.set_details(f"Generation failed: {str(e)}")
+            return backend_pb2.Reply(message=bytes("", encoding='utf-8'))
+
+    def Embedding(self, request, context):
+        """
+        A gRPC method that calculates embeddings for a given sentence.
+        
+        Note: MLX-LM doesn't support embeddings directly. This method returns an error.
+
+        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.
+        """
+        print("Embeddings not supported in MLX backend", file=sys.stderr)
+        context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+        context.set_details("Embeddings are not supported in the MLX backend.")
+        return backend_pb2.EmbeddingResult()
+
+    async def PredictStream(self, request, context):
+        """
+        Generates text based on the given prompt and sampling parameters, and streams the results using MLX.
+
+        Args:
+            request: The predict stream request.
+            context: The gRPC context.
+
+        Yields:
+            backend_pb2.Reply: Streaming predict results.
+        """
+        try:
+            # Prepare the prompt
+            prompt = self._prepare_prompt(request)
+            
+            # Build generation parameters using request attributes and options
+            max_tokens, sampler_params = self._build_generation_params(request, default_max_tokens=512)
+            
+            print(f"Streaming text with MLX - max_tokens: {max_tokens}, sampler_params: {sampler_params}", file=sys.stderr)
+            
+            # Create sampler with parameters
+            sampler = make_sampler(**sampler_params)
+            
+            # Stream text generation using MLX with proper parameters
+            for response in stream_generate(
+                self.model,
+                self.tokenizer,
+                prompt=prompt,
+                max_tokens=max_tokens,
+                sampler=sampler,
+                prompt_cache=self.prompt_cache,
+            ):
+                yield backend_pb2.Reply(message=bytes(response.text, encoding='utf-8'))
+                
+        except Exception as e:
+            print(f"Error in MLX PredictStream: {e}", file=sys.stderr)
+            context.set_code(grpc.StatusCode.INTERNAL)
+            context.set_details(f"Streaming generation failed: {str(e)}")
+            yield backend_pb2.Reply(message=bytes("", encoding='utf-8'))
+
+    def _prepare_prompt(self, request):
+        """
+        Prepare the prompt for MLX generation, handling chat templates if needed.
+
+        Args:
+            request: The gRPC request containing prompt and message information.
+
+        Returns:
+            str: The prepared prompt.
+        """
+        # If tokenizer template is enabled and messages are provided instead of prompt, apply the tokenizer template
+        if not request.Prompt and request.UseTokenizerTemplate and request.Messages:
+            # Convert gRPC messages to the format expected by apply_chat_template
+            messages = []
+            for msg in request.Messages:
+                messages.append({"role": msg.role, "content": msg.content})
+            
+            prompt = self.tokenizer.apply_chat_template(
+                messages, 
+                tokenize=False, 
+                add_generation_prompt=True
+            )
+            return prompt
+        else:
+            return request.Prompt
+
+
+
+
+
+    def _build_generation_params(self, request, default_max_tokens=200):
+        """
+        Build generation parameters from request attributes and options.
+
+        Args:
+            request: The gRPC request.
+            default_max_tokens: Default max_tokens if not specified.
+
+        Returns:
+            tuple: (max_tokens, sampler_params dict)
+        """
+        # Extract max_tokens
+        max_tokens = getattr(request, 'Tokens', default_max_tokens)
+        if max_tokens == 0:
+            max_tokens = default_max_tokens
+        
+        # Extract sampler parameters from request attributes
+        temp = getattr(request, 'Temperature', 0.0)
+        if temp == 0.0:
+            temp = 0.6  # Default temperature
+        
+        top_p = getattr(request, 'TopP', 0.0)
+        if top_p == 0.0:
+            top_p = 1.0  # Default top_p
+        
+        # Initialize sampler parameters
+        sampler_params = {
+            'temp': temp,
+            'top_p': top_p,
+            'xtc_threshold': 0.0,
+            'xtc_probability': 0.0,
+        }
+        
+        # Add seed if specified
+        seed = getattr(request, 'Seed', 0)
+        if seed != 0:
+            mx.random.seed(seed)
+        
+        # Override with options if available
+        if hasattr(self, 'options'):
+            # Max tokens from options
+            if 'max_tokens' in self.options:
+                max_tokens = self.options['max_tokens']
+            
+            # Sampler parameters from options
+            sampler_option_mapping = {
+                'temp': 'temp',
+                'temperature': 'temp',  # alias
+                'top_p': 'top_p', 
+                'xtc_threshold': 'xtc_threshold',
+                'xtc_probability': 'xtc_probability',
+            }
+            
+            for option_key, param_key in sampler_option_mapping.items():
+                if option_key in self.options:
+                    sampler_params[param_key] = self.options[option_key]
+            
+            # Handle seed from options
+            if 'seed' in self.options:
+                mx.random.seed(self.options['seed'])
+        
+        # Special tokens for XTC sampling (if tokenizer has eos_token_ids)
+        xtc_special_tokens = []
+        if hasattr(self.tokenizer, 'eos_token_ids') and self.tokenizer.eos_token_ids:
+            xtc_special_tokens = list(self.tokenizer.eos_token_ids)
+        elif hasattr(self.tokenizer, 'eos_token_id') and self.tokenizer.eos_token_id is not None:
+            xtc_special_tokens = [self.tokenizer.eos_token_id]
+        
+        # Add newline token if available
+        try:
+            newline_tokens = self.tokenizer.encode("\n")
+            xtc_special_tokens.extend(newline_tokens)
+        except:
+            pass  # Skip if encoding fails
+        
+        sampler_params['xtc_special_tokens'] = xtc_special_tokens
+        
+        return max_tokens, sampler_params
+
+async def serve(address):
+    # Start asyncio gRPC server
+    server = grpc.aio.server(migration_thread_pool=futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
+        options=[
+            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
+            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
+        ])
+    # Add the servicer to the server
+    backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
+    # Bind the server to the address
+    server.add_insecure_port(address)
+
+    # Gracefully shutdown the server on SIGTERM or SIGINT
+    loop = asyncio.get_event_loop()
+    for sig in (signal.SIGINT, signal.SIGTERM):
+        loop.add_signal_handler(
+            sig, lambda: asyncio.ensure_future(server.stop(5))
+        )
+
+    # Start the server
+    await server.start()
+    print("Server started. Listening on: " + address, file=sys.stderr)
+    # Wait for the server to be terminated
+    await server.wait_for_termination()
+
+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()
+
+    asyncio.run(serve(args.addr))

backend/python/mlx/install.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+set -e
+
+USE_PIP=true
+PYTHON_VERSION=""
+
+backend_dir=$(dirname $0)
+
+if [ -d $backend_dir/common ]; then
+    source $backend_dir/common/libbackend.sh
+else
+    source $backend_dir/../common/libbackend.sh
+fi
+
+installRequirements