wip

⬆️ Update docs version mudler/LocalAI

Signed-off-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: mudler <2420543+mudler@users.noreply.github.com>

.env

@@ -29,6 +29,9 @@
 ## Enable/Disable single backend (useful if only one GPU is available)
 # LOCALAI_SINGLE_ACTIVE_BACKEND=true
 
+# Forces shutdown of the backends if busy (only if LOCALAI_SINGLE_ACTIVE_BACKEND is set)
+# LOCALAI_FORCE_BACKEND_SHUTDOWN=true
+
 ## Specify a build type. Available: cublas, openblas, clblas.
 ## cuBLAS: This is a GPU-accelerated version of the complete standard BLAS (Basic Linear Algebra Subprograms) library. It's provided by Nvidia and is part of their CUDA toolkit.
 ## OpenBLAS: This is an open-source implementation of the BLAS library that aims to provide highly optimized code for various platforms. It includes support for multi-threading and can be compiled to use hardware-specific features for additional performance. OpenBLAS can run on many kinds of hardware, including CPUs from Intel, AMD, and ARM.
@@ -73,7 +76,7 @@
 
 ### Define a list of GRPC Servers for llama-cpp workers to distribute the load
 # https://github.com/ggerganov/llama.cpp/pull/6829
-# https://github.com/ggerganov/llama.cpp/blob/master/examples/rpc/README.md
+# https://github.com/ggerganov/llama.cpp/blob/master/tools/rpc/README.md
 # LLAMACPP_GRPC_SERVERS=""
 
 ### Enable to run parallel requests

.github/dependabot.yml

@@ -29,10 +29,6 @@ updates:
     schedule:
       # Check for updates to GitHub Actions every weekday
       interval: "weekly"
-  - package-ecosystem: "pip"
-    directory: "/backend/python/autogptq"
-    schedule:
-      interval: "weekly"
   - package-ecosystem: "pip"
     directory: "/backend/python/bark"
     schedule:

.github/workflows/bump_deps.yaml

@@ -12,7 +12,7 @@ jobs:
           - repository: "ggml-org/llama.cpp"
             variable: "CPPLLAMA_VERSION"
             branch: "master"
-          - repository: "ggerganov/whisper.cpp"
+          - repository: "ggml-org/whisper.cpp"
             variable: "WHISPER_CPP_VERSION"
             branch: "master"
           - repository: "PABannier/bark.cpp"

.github/workflows/dependabot_auto.yml

@@ -14,7 +14,7 @@ jobs:
     steps:
       - name: Dependabot metadata
         id: metadata
-        uses: dependabot/fetch-metadata@v2.3.0
+        uses: dependabot/fetch-metadata@v2.4.0
         with:
           github-token: "${{ secrets.GITHUB_TOKEN }}"
           skip-commit-verification: true

.github/workflows/deploy-explorer.yaml

@@ -42,7 +42,7 @@ jobs:
             script: |
                 sudo rm -rf local-ai/ || true
       - name: copy file via ssh
-        uses: appleboy/scp-action@v0.1.7
+        uses: appleboy/scp-action@v1.0.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}

.github/workflows/generate_intel_image.yaml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - base-image: intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04
+          - base-image: intel/oneapi-basekit:2025.1.0-0-devel-ubuntu22.04
             runs-on: 'ubuntu-latest'
             platforms: 'linux/amd64'
     runs-on: ${{matrix.runs-on}}

.github/workflows/image-pr.yml

@@ -33,6 +33,7 @@ jobs:
       # Pushing with all jobs in parallel
       # eats the bandwidth of all the nodes
       max-parallel: ${{ github.event_name != 'pull_request' && 4 || 8 }}
+      fail-fast: false
       matrix:
         include:
           # This is basically covered by the AIO test
@@ -56,26 +57,35 @@ jobs:
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             makeflags: "--jobs=3 --output-sync=target"
-          # - build-type: 'hipblas'
+          - build-type: 'hipblas'
-          #   platforms: 'linux/amd64'
+            platforms: 'linux/amd64'
-          #   tag-latest: 'false'
+            tag-latest: 'false'
-          #   tag-suffix: '-hipblas'
+            tag-suffix: '-hipblas'
-          #   ffmpeg: 'false'
+            ffmpeg: 'false'
-          #   image-type: 'extras'
+            image-type: 'extras'
-          #   base-image: "rocm/dev-ubuntu-22.04:6.1"
+            base-image: "rocm/dev-ubuntu-22.04:6.1"
-          #   grpc-base-image: "ubuntu:22.04"
+            grpc-base-image: "ubuntu:22.04"
-          #   runs-on: 'arc-runner-set'
+            runs-on: 'arc-runner-set'
-          #   makeflags: "--jobs=3 --output-sync=target"
+            makeflags: "--jobs=3 --output-sync=target"
-          # - build-type: 'sycl_f16'
+          - build-type: 'sycl_f16'
-          #   platforms: 'linux/amd64'
+            platforms: 'linux/amd64'
-          #   tag-latest: 'false'
+            tag-latest: 'false'
-          #   base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
+            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
-          #   grpc-base-image: "ubuntu:22.04"
+            grpc-base-image: "ubuntu:22.04"
-          #   tag-suffix: 'sycl-f16-ffmpeg'
+            tag-suffix: 'sycl-f16-ffmpeg'
-          #   ffmpeg: 'true'
+            ffmpeg: 'true'
-          #   image-type: 'extras'
+            image-type: 'extras'
-          #   runs-on: 'arc-runner-set'
+            runs-on: 'arc-runner-set'
-          #   makeflags: "--jobs=3 --output-sync=target"
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'vulkan'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-vulkan-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'ubuntu-latest'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=4 --output-sync=target"
   # core-image-build:
   #   uses: ./.github/workflows/image_build.yml
   #   with:

.github/workflows/image.yml

@@ -45,13 +45,13 @@ jobs:
           - build-type: 'hipblas'
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-hipblas-ffmpeg'
+            tag-suffix: '-hipblas-extras'
             ffmpeg: 'true'
             image-type: 'extras'
             aio: "-aio-gpu-hipblas"
             base-image: "rocm/dev-ubuntu-22.04:6.1"
             grpc-base-image: "ubuntu:22.04"
-            latest-image: 'latest-gpu-hipblas'
+            latest-image: 'latest-gpu-hipblas-extras'
             latest-image-aio: 'latest-aio-gpu-hipblas'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
@@ -59,32 +59,13 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'false'
             tag-suffix: '-hipblas'
-            ffmpeg: 'false'
-            image-type: 'extras'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
-            grpc-base-image: "ubuntu:22.04"
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'hipblas'
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-hipblas-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "rocm/dev-ubuntu-22.04:6.1"
             grpc-base-image: "ubuntu:22.04"
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'hipblas'
+            latest-image: 'latest-gpu-hipblas'
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-hipblas-core'
-            ffmpeg: 'false'
-            image-type: 'core'
-            base-image: "rocm/dev-ubuntu-22.04:6.1"
-            grpc-base-image: "ubuntu:22.04"
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=3 --output-sync=target"
   self-hosted-jobs:
     uses: ./.github/workflows/image_build.yml
     with:
@@ -114,110 +95,58 @@ jobs:
       max-parallel: ${{ github.event_name != 'pull_request' && 5 || 8 }}
       matrix:
         include:
-          # Extra images
-          - build-type: ''
-            #platforms: 'linux/amd64,linux/arm64'
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: ''
-            ffmpeg: ''
-            image-type: 'extras'
-            runs-on: 'arc-runner-set'
-            base-image: "ubuntu:22.04"
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: ''
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-ffmpeg'
-            ffmpeg: 'true'
-            image-type: 'extras'
-            runs-on: 'arc-runner-set'
-            base-image: "ubuntu:22.04"
-            makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11'
+            tag-suffix: '-cublas-cuda11-extras'
-            ffmpeg: ''
-            image-type: 'extras'
-            runs-on: 'arc-runner-set'
-            base-image: "ubuntu:22.04"
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'cublas'
-            cuda-major-version: "12"
-            cuda-minor-version: "0"
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12'
-            ffmpeg: ''
-            image-type: 'extras'
-            runs-on: 'arc-runner-set'
-            base-image: "ubuntu:22.04"
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'cublas'
-            cuda-major-version: "11"
-            cuda-minor-version: "7"
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: '-cublas-cuda11-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             aio: "-aio-gpu-nvidia-cuda-11"
-            latest-image: 'latest-gpu-nvidia-cuda-11'
+            latest-image: 'latest-gpu-nvidia-cuda-11-extras'
             latest-image-aio: 'latest-aio-gpu-nvidia-cuda-11'
             makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
-            tag-latest: 'auto'
+            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12-ffmpeg'
+            tag-suffix: '-cublas-cuda12-extras'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             aio: "-aio-gpu-nvidia-cuda-12"
-            latest-image: 'latest-gpu-nvidia-cuda-12'
+            latest-image: 'latest-gpu-nvidia-cuda-12-extras'
             latest-image-aio: 'latest-aio-gpu-nvidia-cuda-12'
             makeflags: "--jobs=3 --output-sync=target"
-          - build-type: ''
-            #platforms: 'linux/amd64,linux/arm64'
-            platforms: 'linux/amd64'
-            tag-latest: 'auto'
-            tag-suffix: ''
-            ffmpeg: ''
-            image-type: 'extras'
-            base-image: "ubuntu:22.04"
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'sycl_f16'
             platforms: 'linux/amd64'
-            tag-latest: 'auto'
+            tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f16-ffmpeg'
+            tag-suffix: '-sycl-f16-extras'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             aio: "-aio-gpu-intel-f16"
-            latest-image: 'latest-gpu-intel-f16'
+            latest-image: 'latest-gpu-intel-f16-extras'
             latest-image-aio: 'latest-aio-gpu-intel-f16'
             makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'sycl_f32'
             platforms: 'linux/amd64'
-            tag-latest: 'auto'
+            tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f32-ffmpeg'
+            tag-suffix: '-sycl-f32-extras'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             aio: "-aio-gpu-intel-f32"
-            latest-image: 'latest-gpu-intel-f32'
+            latest-image: 'latest-gpu-intel-f32-extras'
             latest-image-aio: 'latest-aio-gpu-intel-f32'
             makeflags: "--jobs=3 --output-sync=target"
           # Core images
@@ -226,41 +155,23 @@ jobs:
             tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f16-core'
+            tag-suffix: '-sycl-f16'
-            ffmpeg: 'false'
+            ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
+            latest-image: 'latest-gpu-intel-f16'
           - build-type: 'sycl_f32'
             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-f32-core'
+            tag-suffix: '-sycl-f32'
-            ffmpeg: 'false'
-            image-type: 'core'
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'sycl_f16'
-            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-ffmpeg-core'
-            ffmpeg: 'true'
-            image-type: 'core'
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=3 --output-sync=target"
-          - build-type: 'sycl_f32'
-            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-f32-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
+            latest-image: 'latest-gpu-intel-f32'
 
   core-image-build:
     uses: ./.github/workflows/image_build.yml
@@ -293,7 +204,7 @@ jobs:
           - build-type: ''
             platforms: 'linux/amd64,linux/arm64'
             tag-latest: 'auto'
-            tag-suffix: '-ffmpeg-core'
+            tag-suffix: ''
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "ubuntu:22.04"
@@ -308,60 +219,38 @@ jobs:
             cuda-minor-version: "7"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11-core'
+            tag-suffix: '-cublas-cuda11'
-            ffmpeg: ''
-            image-type: 'core'
-            base-image: "ubuntu:22.04"
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=4 --output-sync=target"
-            skip-drivers: 'false'
-          - build-type: 'cublas'
-            cuda-major-version: "12"
-            cuda-minor-version: "0"
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12-core'
-            ffmpeg: ''
-            image-type: 'core'
-            base-image: "ubuntu:22.04"
-            runs-on: 'arc-runner-set'
-            makeflags: "--jobs=4 --output-sync=target"
-            skip-drivers: 'false'
-          - build-type: 'cublas'
-            cuda-major-version: "11"
-            cuda-minor-version: "7"
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             makeflags: "--jobs=4 --output-sync=target"
             skip-drivers: 'false'
+            latest-image: 'latest-gpu-nvidia-cuda-12'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12-ffmpeg-core'
+            tag-suffix: '-cublas-cuda12'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
+            latest-image: 'latest-gpu-nvidia-cuda-12'
           - build-type: 'vulkan'
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-vulkan-ffmpeg-core'
+            tag-suffix: '-vulkan'
-            latest-image: 'latest-vulkan-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
+            latest-image: 'latest-gpu-vulkan'
   gh-runner:
     uses: ./.github/workflows/image_build.yml
     with:
@@ -394,8 +283,8 @@ jobs:
             cuda-minor-version: "0"
             platforms: 'linux/arm64'
             tag-latest: 'false'
-            tag-suffix: '-nvidia-l4t-arm64-core'
+            tag-suffix: '-nvidia-l4t-arm64'
-            latest-image: 'latest-nvidia-l4t-arm64-core'
+            latest-image: 'latest-nvidia-l4t-arm64'
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"

.github/workflows/notify-models.yaml

@@ -8,7 +8,7 @@ jobs:
   notify-discord:
     if: ${{ (github.event.pull_request.merged == true) && (contains(github.event.pull_request.labels.*.name, 'area/ai-model')) }}
     env:
-        MODEL_NAME: hermes-2-theta-llama-3-8b
+        MODEL_NAME: gemma-3-12b-it
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
@@ -16,7 +16,7 @@ jobs:
         fetch-depth: 0 # needed to checkout all branches for this Action to work
     - uses: mudler/localai-github-action@v1
       with:
-        model: 'hermes-2-theta-llama-3-8b' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
+        model: 'gemma-3-12b-it' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
         # Check the PR diff using the current branch and the base branch of the PR
     - uses: GrantBirki/git-diff-action@v2.8.0
       id: git-diff-action
@@ -79,7 +79,7 @@ jobs:
         args: ${{ steps.summarize.outputs.message }}
     - name: Setup tmate session if fails
       if: ${{ failure() }}
-      uses: mxschmitt/action-tmate@v3.19
+      uses: mxschmitt/action-tmate@v3.22
       with:
         detached: true
         connect-timeout-seconds: 180
@@ -87,7 +87,7 @@ jobs:
   notify-twitter:
     if: ${{ (github.event.pull_request.merged == true) && (contains(github.event.pull_request.labels.*.name, 'area/ai-model')) }}
     env:
-        MODEL_NAME: hermes-2-theta-llama-3-8b
+        MODEL_NAME: gemma-3-12b-it
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
@@ -161,7 +161,7 @@ jobs:
         TWITTER_ACCESS_TOKEN_SECRET: ${{ secrets.TWITTER_ACCESS_TOKEN_SECRET }}
     - name: Setup tmate session if fails
       if: ${{ failure() }}
-      uses: mxschmitt/action-tmate@v3.19
+      uses: mxschmitt/action-tmate@v3.22
       with:
         detached: true
         connect-timeout-seconds: 180

.github/workflows/notify-releases.yaml

@@ -14,7 +14,7 @@ jobs:
     steps:
     - uses: mudler/localai-github-action@v1
       with:
-        model: 'hermes-2-theta-llama-3-8b' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
+        model: 'gemma-3-12b-it' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
     - name: Summarize
       id: summarize
       run: |
@@ -60,4 +60,4 @@ jobs:
         DISCORD_AVATAR: "https://avatars.githubusercontent.com/u/139863280?v=4"
       uses: Ilshidur/action-discord@master
       with:
-        args: ${{ steps.summarize.outputs.message }}
+        args: ${{ steps.summarize.outputs.message }}

.github/workflows/release.yaml

@@ -36,6 +36,7 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential ffmpeg protobuf-compiler ccache upx-ucl gawk
           sudo apt-get install -qy binutils-aarch64-linux-gnu gcc-aarch64-linux-gnu g++-aarch64-linux-gnu libgmock-dev
+          make install-go-tools
       - name: Install CUDA Dependencies
         run: |
           curl -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/cross-linux-aarch64/cuda-keyring_1.1-1_all.deb
@@ -123,7 +124,7 @@ jobs:
             release/*
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180
@@ -151,6 +152,7 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install -y wget curl build-essential ffmpeg protobuf-compiler ccache upx-ucl gawk cmake libgmock-dev
+          make install-go-tools
       - name: Intel Dependencies
         run: |
           wget -O- https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB | gpg --dearmor | sudo tee /usr/share/keyrings/oneapi-archive-keyring.gpg > /dev/null
@@ -232,7 +234,7 @@ jobs:
             release/*
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180
@@ -253,8 +255,7 @@ jobs:
       - name: Dependencies
         run: |
           brew install protobuf grpc
-          go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@8ba23be9613c672d40ae261d2a1335d639bdd59b
+          make install-go-tools
-          go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.0
       - name: Build
         id: build
         run: |
@@ -275,7 +276,7 @@ jobs:
             release/*
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180
@@ -295,8 +296,7 @@ jobs:
       - name: Dependencies
         run: |
           brew install protobuf grpc libomp llvm
-          go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+          make install-go-tools
-          go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
       - name: Build
         id: build
         run: |
@@ -317,7 +317,7 @@ jobs:
             release/*
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180

.github/workflows/secscan.yaml

@@ -18,7 +18,7 @@ jobs:
         if: ${{ github.actor != 'dependabot[bot]' }}
       - name: Run Gosec Security Scanner
         if: ${{ github.actor != 'dependabot[bot]' }}
-        uses: securego/gosec@v2.22.0
+        uses: securego/gosec@v2.22.4
         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

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

.github/workflows/test.yml

@@ -71,7 +71,7 @@ jobs:
         run: |
           sudo apt-get update
           sudo apt-get install build-essential ccache upx-ucl curl ffmpeg
-          sudo apt-get install -y libgmock-dev
+          sudo apt-get install -y libgmock-dev clang
           curl https://repo.anaconda.com/pkgs/misc/gpgkeys/anaconda.asc | gpg --dearmor > conda.gpg && \
              sudo install -o root -g root -m 644 conda.gpg /usr/share/keyrings/conda-archive-keyring.gpg && \
              gpg --keyring /usr/share/keyrings/conda-archive-keyring.gpg --no-default-keyring --fingerprint 34161F5BF5EB1D4BFBBB8F0A8AEB4F8B29D82806 && \
@@ -96,6 +96,7 @@ jobs:
 
           go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
           go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+          go install github.com/GeertJohan/go.rice/rice@latest
 
           # The python3-grpc-tools package in 22.04 is too old
           pip install --user grpcio-tools
@@ -130,7 +131,7 @@ jobs:
           PATH="$PATH:/root/go/bin" GO_TAGS="tts" make --jobs 5 --output-sync=target test
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180
@@ -183,6 +184,7 @@ jobs:
           rm protoc.zip
           go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
           go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+          go install github.com/GeertJohan/go.rice/rice@latest
           PATH="$PATH:$HOME/go/bin" make protogen-go
       - name: Build images
         run: |
@@ -194,7 +196,7 @@ jobs:
             make run-e2e-aio
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180
@@ -222,6 +224,7 @@ jobs:
         run: |
           brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm
           pip install --user --no-cache-dir grpcio-tools
+          go install github.com/GeertJohan/go.rice/rice@latest
       - name: Test
         run: |
           export C_INCLUDE_PATH=/usr/local/include
@@ -232,7 +235,7 @@ jobs:
           BUILD_TYPE="GITHUB_CI_HAS_BROKEN_METAL" CMAKE_ARGS="-DGGML_F16C=OFF -DGGML_AVX512=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF" make --jobs 4 --output-sync=target test
       - name: Setup tmate session if tests fail
         if: ${{ failure() }}
-        uses: mxschmitt/action-tmate@v3.19
+        uses: mxschmitt/action-tmate@v3.22
         with:
           detached: true
           connect-timeout-seconds: 180

Dockerfile

@@ -15,7 +15,7 @@ ARG TARGETARCH
 ARG TARGETVARIANT
 
 ENV DEBIAN_FRONTEND=noninteractive
-ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,transformers:/build/backend/python/transformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,autogptq:/build/backend/python/autogptq/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,faster-whisper:/build/backend/python/faster-whisper/run.sh,kokoro:/build/backend/python/kokoro/run.sh,vllm:/build/backend/python/vllm/run.sh,exllama2:/build/backend/python/exllama2/run.sh"
+ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,transformers:/build/backend/python/transformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,faster-whisper:/build/backend/python/faster-whisper/run.sh,kokoro:/build/backend/python/kokoro/run.sh,vllm:/build/backend/python/vllm/run.sh,exllama2:/build/backend/python/exllama2/run.sh"
 
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
@@ -46,9 +46,10 @@ EOT
 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
 
-# Install grpc compilers
+# Install grpc compilers and rice
 RUN go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2 && \
-    go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+    go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af && \
+    go install github.com/GeertJohan/go.rice/rice@latest
 
 COPY --chmod=644 custom-ca-certs/* /usr/local/share/ca-certificates/
 RUN update-ca-certificates
@@ -300,10 +301,9 @@ COPY .git .
 RUN make prepare
 
 ## Build the binary
-## If it's CUDA or hipblas, we want to skip some of the llama-compat backends to save space
+## If we're on arm64 AND using cublas/hipblas, skip some of the llama-compat backends to save space
-## We only leave the most CPU-optimized variant and the fallback for the cublas/hipblas build
+## Otherwise just run the normal build
-## (both will use CUDA or hipblas for the actual computation)
+RUN if [ "${TARGETARCH}" = "arm64" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
-RUN if [ "${BUILD_TYPE}" = "cublas" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
         SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx512 backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
     else \
         make build; \
@@ -431,9 +431,6 @@ RUN if [[ ( "${EXTRA_BACKENDS}" =~ "kokoro" || -z "${EXTRA_BACKENDS}" ) && "$IMA
 RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vllm" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/vllm \
     ; fi && \
-    if [[ ( "${EXTRA_BACKENDS}" =~ "autogptq" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
-        make -C backend/python/autogptq \
-    ; fi && \
     if [[ ( "${EXTRA_BACKENDS}" =~ "bark" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/bark \
     ; fi && \

Makefile

@@ -6,11 +6,11 @@ BINARY_NAME=local-ai
 DETECT_LIBS?=true
 
 # llama.cpp versions
-CPPLLAMA_VERSION?=4663bd353c61c1136cd8a97b9908755e4ab30cec
+CPPLLAMA_VERSION?=e5c834f718a32b7584f142799bbf508fddb9021c
 
 # whisper.cpp version
-WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
+WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
+WHISPER_CPP_VERSION?=e41bc5c61ae66af6be2bd7011769bb821a83e8ae
 
 # go-piper version
 PIPER_REPO?=https://github.com/mudler/go-piper
@@ -21,8 +21,11 @@ BARKCPP_REPO?=https://github.com/PABannier/bark.cpp.git
 BARKCPP_VERSION?=v1.0.0
 
 # stablediffusion.cpp (ggml)
-STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
+STABLEDIFFUSION_GGML_REPO?=https://github.com/richiejp/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=19d876ee300a055629926ff836489901f734f2b7
+STABLEDIFFUSION_GGML_VERSION?=53e3b17eb3d0b5760ced06a1f98320b68b34aaae
+
+# ONEAPI variables for SYCL
+export ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 
 ONNX_VERSION?=1.20.0
 ONNX_ARCH?=x64
@@ -30,8 +33,12 @@ ONNX_OS?=linux
 
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
-export CMAKE_ARGS?=
+export CMAKE_ARGS?=-DBUILD_SHARED_LIBS=OFF
+export WHISPER_CMAKE_ARGS?=-DBUILD_SHARED_LIBS=OFF
 export BACKEND_LIBS?=
+export WHISPER_DIR=$(abspath ./sources/whisper.cpp)
+export WHISPER_INCLUDE_PATH=$(WHISPER_DIR)/include:$(WHISPER_DIR)/ggml/include
+export WHISPER_LIBRARY_PATH=$(WHISPER_DIR)/build/src/:$(WHISPER_DIR)/build/ggml/src
 
 CGO_LDFLAGS?=
 CGO_LDFLAGS_WHISPER?=
@@ -81,6 +88,7 @@ endif
 # IF native is false, we add -DGGML_NATIVE=OFF to CMAKE_ARGS
 ifeq ($(NATIVE),false)
 	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+	WHISPER_CMAKE_ARGS+=-DGGML_NATIVE=OFF
 endif
 
 # Detect if we are running on arm64
@@ -108,13 +116,31 @@ ifeq ($(OS),Darwin)
 	# disable metal if on Darwin and any other value is explicitly passed.
 	else ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
+		WHISPER_CMAKE_ARGS+=-DGGML_METAL=OFF
 		export GGML_NO_ACCELERATE=1
 		export GGML_NO_METAL=1
+		GO_LDFLAGS_WHISPER+=-lggml-blas
+		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-blas
 	endif
 
 	ifeq ($(BUILD_TYPE),metal)
-#			-lcblas 	removed: it seems to always be listed as a duplicate flag.
 		CGO_LDFLAGS += -framework Accelerate
+		CGO_LDFLAGS_WHISPER+=-lggml-metal -lggml-blas
+		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL_USE_BF16=ON
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		CMAKE_ARGS+=-DGGML_OPENMP=OFF
+		WHISPER_CMAKE_ARGS+=-DGGML_METAL=ON
+		WHISPER_CMAKE_ARGS+=-DGGML_METAL_USE_BF16=ON
+		WHISPER_CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_EXAMPLES=OFF
+		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_TESTS=OFF
+		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_SERVER=OFF
+		WHISPER_CMAKE_ARGS+=-DGGML_OPENMP=OFF
+		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-metal/:$(WHISPER_DIR)/build/ggml/src/ggml-blas
+	else
+		CGO_LDFLAGS_WHISPER+=-lggml-blas
+		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-blas
 	endif
 else
 CGO_LDFLAGS_WHISPER+=-lgomp
@@ -126,21 +152,29 @@ ifeq ($(BUILD_TYPE),openblas)
 endif
 
 ifeq ($(BUILD_TYPE),cublas)
-	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH)
+	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH) -L$(CUDA_LIBPATH)/stubs/ -lcuda
 	export GGML_CUDA=1
-	CGO_LDFLAGS_WHISPER+=-L$(CUDA_LIBPATH)/stubs/ -lcuda -lcufft
+	CMAKE_ARGS+=-DGGML_CUDA=ON
+	WHISPER_CMAKE_ARGS+=-DGGML_CUDA=ON
+	CGO_LDFLAGS_WHISPER+=-lcufft -lggml-cuda
+	export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-cuda/
 endif
 
 ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=1
+	WHISPER_CMAKE_ARGS+=-DGGML_VULKAN=1
+	CGO_LDFLAGS_WHISPER+=-lggml-vulkan -lvulkan
+	export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-vulkan/
 endif
 
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	export GGML_SYCL=1
+	CMAKE_ARGS+=-DGGML_SYCL=ON
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)
 	export GGML_SYCL_F16=1
+	CMAKE_ARGS+=-DGGML_SYCL_F16=ON
 endif
 
 ifeq ($(BUILD_TYPE),hipblas)
@@ -151,7 +185,7 @@ ifeq ($(BUILD_TYPE),hipblas)
 	export CC=$(ROCM_HOME)/llvm/bin/clang
 	export STABLE_BUILD_TYPE=
 	export GGML_HIP=1
-	GPU_TARGETS ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
+	GPU_TARGETS ?= gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102
 	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
 	CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
 	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link -L${ROCM_HOME}/lib/llvm/lib
@@ -260,11 +294,7 @@ backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp
 	$(MAKE) -C backend/go/image/stablediffusion-ggml libsd.a
 
 backend-assets/grpc/stablediffusion-ggml: backend/go/image/stablediffusion-ggml/libsd.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ LIBRARY_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ \
+	$(MAKE) -C backend/go/image/stablediffusion-ggml CGO_LDFLAGS="$(CGO_LDFLAGS)" stablediffusion-ggml
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion-ggml ./backend/go/image/stablediffusion-ggml/
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/stablediffusion-ggml
-endif
 
 sources/onnxruntime:
 	mkdir -p sources/onnxruntime
@@ -290,8 +320,9 @@ sources/whisper.cpp:
 	git checkout $(WHISPER_CPP_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-sources/whisper.cpp/libwhisper.a: sources/whisper.cpp
+sources/whisper.cpp/build/src/libwhisper.a: sources/whisper.cpp
-	cd sources/whisper.cpp && $(MAKE) libwhisper.a libggml.a
+	cd sources/whisper.cpp && cmake $(WHISPER_CMAKE_ARGS) . -B ./build
+	cd sources/whisper.cpp/build && cmake --build . --config Release
 
 get-sources: sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
 
@@ -341,8 +372,14 @@ clean-tests:
 clean-dc: clean
 	cp -r /build/backend-assets /workspace/backend-assets
 
+## Install Go tools
+install-go-tools:
+	go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+	go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
+	go install github.com/GeertJohan/go.rice/rice@latest
+
 ## Build:
-build: prepare backend-assets grpcs ## Build the project
+build: prepare backend-assets grpcs install-go-tools ## Build the project
 	$(info ${GREEN}I local-ai build info:${RESET})
 	$(info ${GREEN}I BUILD_TYPE: ${YELLOW}$(BUILD_TYPE)${RESET})
 	$(info ${GREEN}I GO_TAGS: ${YELLOW}$(GO_TAGS)${RESET})
@@ -352,7 +389,9 @@ ifneq ($(BACKEND_LIBS),)
 	$(MAKE) backend-assets/lib
 	cp -f $(BACKEND_LIBS) backend-assets/lib/
 endif
+	rm -rf $(BINARY_NAME) || true
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o $(BINARY_NAME) ./
+	rice append --exec $(BINARY_NAME)
 
 build-minimal:
 	BUILD_GRPC_FOR_BACKEND_LLAMA=true GRPC_BACKENDS="backend-assets/grpc/llama-cpp-avx2" GO_TAGS=p2p $(MAKE) build
@@ -424,6 +463,7 @@ prepare-test: grpcs
 	cp -rf backend-assets core/http
 	cp tests/models_fixtures/* test-models
 
+## Test targets
 test: prepare test-models/testmodel.ggml grpcs
 	@echo 'Running tests'
 	export GO_TAGS="tts debug"
@@ -498,7 +538,7 @@ protogen: protogen-go protogen-python
 protogen-clean: protogen-go-clean protogen-python-clean
 
 .PHONY: protogen-go
-protogen-go:
+protogen-go: install-go-tools
 	mkdir -p pkg/grpc/proto
 	protoc --experimental_allow_proto3_optional -Ibackend/ --go_out=pkg/grpc/proto/ --go_opt=paths=source_relative --go-grpc_out=pkg/grpc/proto/ --go-grpc_opt=paths=source_relative \
     backend/backend.proto
@@ -509,18 +549,10 @@ protogen-go-clean:
 	$(RM) bin/*
 
 .PHONY: protogen-python
-protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
+protogen-python: bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
 
 .PHONY: protogen-python-clean
-protogen-python-clean: autogptq-protogen-clean bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean rerankers-protogen-clean transformers-protogen-clean kokoro-protogen-clean vllm-protogen-clean faster-whisper-protogen-clean
+protogen-python-clean: bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean rerankers-protogen-clean transformers-protogen-clean kokoro-protogen-clean vllm-protogen-clean faster-whisper-protogen-clean
-
-.PHONY: autogptq-protogen
-autogptq-protogen:
-	$(MAKE) -C backend/python/autogptq protogen
-
-.PHONY: autogptq-protogen-clean
-autogptq-protogen-clean:
-	$(MAKE) -C backend/python/autogptq protogen-clean
 
 .PHONY: bark-protogen
 bark-protogen:
@@ -597,7 +629,6 @@ vllm-protogen-clean:
 ## GRPC
 # Note: it is duplicated in the Dockerfile
 prepare-extra-conda-environments: protogen-python
-	$(MAKE) -C backend/python/autogptq
 	$(MAKE) -C backend/python/bark
 	$(MAKE) -C backend/python/coqui
 	$(MAKE) -C backend/python/diffusers
@@ -611,10 +642,12 @@ prepare-extra-conda-environments: protogen-python
 prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/transformers
 	$(MAKE) -C backend/python/diffusers
+	$(MAKE) -C backend/python/vllm
 
 test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/transformers test
 	$(MAKE) -C backend/python/diffusers test
+	$(MAKE) -C backend/python/vllm test
 
 backend-assets:
 	mkdir -p backend-assets
@@ -756,8 +789,8 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/silero-vad
 endif
 
-backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/libwhisper.a backend-assets/grpc
+backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/build/src/libwhisper.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="$(CURDIR)/sources/whisper.cpp/include:$(CURDIR)/sources/whisper.cpp/ggml/include" LIBRARY_PATH=$(CURDIR)/sources/whisper.cpp \
+	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="${WHISPER_INCLUDE_PATH}" LIBRARY_PATH="${WHISPER_LIBRARY_PATH}" LD_LIBRARY_PATH="${WHISPER_LIBRARY_PATH}" \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./backend/go/transcribe/whisper
 ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/whisper
@@ -809,7 +842,8 @@ docker-aio-all:
 
 docker-image-intel:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
+		--progress plain \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.1.0-0-devel-ubuntu24.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
@@ -817,7 +851,7 @@ docker-image-intel:
 
 docker-image-intel-xpu:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.1.0-0-devel-ubuntu22.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \

README.md

@@ -1,7 +1,6 @@
 <h1 align="center">
   <br>
-  <img height="300" src="https://github.com/go-skynet/LocalAI/assets/2420543/0966aa2a-166e-4f99-a3e5-6c915fc997dd"> <br>
+  <img height="300" src="./core/http/static/logo.png"> <br>
-    LocalAI
 <br>
 </h1>
 
@@ -31,7 +30,7 @@
 
 <p align="center">
 <a href="https://twitter.com/LocalAI_API" target="blank">
-<img src="https://img.shields.io/twitter/follow/LocalAI_API?label=Follow: LocalAI_API&style=social" alt="Follow LocalAI_API"/>
+<img src="https://img.shields.io/badge/X-%23000000.svg?style=for-the-badge&logo=X&logoColor=white&label=LocalAI_API" alt="Follow LocalAI_API"/>
 </a>
 <a href="https://discord.gg/uJAeKSAGDy" target="blank">
 <img src="https://dcbadge.vercel.app/api/server/uJAeKSAGDy?style=flat-square&theme=default-inverted" alt="Join LocalAI Discord Community"/>
@@ -44,32 +43,89 @@
 
 > :bulb: Get help - [❓FAQ](https://localai.io/faq/) [💭Discussions](https://github.com/go-skynet/LocalAI/discussions) [:speech_balloon: Discord](https://discord.gg/uJAeKSAGDy) [:book: Documentation website](https://localai.io/)
 >
-> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) 
+> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) Try on 
+[![Telegram](https://img.shields.io/badge/Telegram-2CA5E0?style=for-the-badge&logo=telegram&logoColor=white)](https://t.me/localaiofficial_bot)
 
 [![tests](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml)[![Build and Release](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml)[![build container images](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml)[![Bump dependencies](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml)[![Artifact Hub](https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/localai)](https://artifacthub.io/packages/search?repo=localai)
 
-**LocalAI** is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that’s compatible with OpenAI (Elevenlabs, Anthropic... ) API specifications for local AI inferencing. It allows you to run LLMs, generate images, audio (and not only) locally or on-prem with consumer grade hardware, supporting multiple model families. Does not require GPU. It is created and maintained by [Ettore Di Giacinto](https://github.com/mudler).
+**LocalAI** is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that's compatible with OpenAI (Elevenlabs, Anthropic... ) API specifications for local AI inferencing. It allows you to run LLMs, generate images, audio (and not only) locally or on-prem with consumer grade hardware, supporting multiple model families. Does not require GPU. It is created and maintained by [Ettore Di Giacinto](https://github.com/mudler).
 
-![screen](https://github.com/mudler/LocalAI/assets/2420543/20b5ccd2-8393-44f0-aaf6-87a23806381e)
+
+## 📚🆕 Local Stack Family
+
+🆕 LocalAI is now part of a comprehensive suite of AI tools designed to work together:
+
+<table>
+  <tr>
+    <td width="50%" valign="top">
+      <a href="https://github.com/mudler/LocalAGI">
+        <img src="https://raw.githubusercontent.com/mudler/LocalAGI/refs/heads/main/webui/react-ui/public/logo_2.png" width="300" alt="LocalAGI Logo">
+      </a>
+    </td>
+    <td width="50%" valign="top">
+      <h3><a href="https://github.com/mudler/LocalAGI">LocalAGI</a></h3>
+      <p>A powerful Local AI agent management platform that serves as a drop-in replacement for OpenAI's Responses API, enhanced with advanced agentic capabilities.</p>
+    </td>
+  </tr>
+  <tr>
+    <td width="50%" valign="top">
+      <a href="https://github.com/mudler/LocalRecall">
+        <img src="https://raw.githubusercontent.com/mudler/LocalRecall/refs/heads/main/static/localrecall_horizontal.png" width="300" alt="LocalRecall Logo">
+      </a>
+    </td>
+    <td width="50%" valign="top">
+      <h3><a href="https://github.com/mudler/LocalRecall">LocalRecall</a></h3>
+      <p>A REST-ful API and knowledge base management system that provides persistent memory and storage capabilities for AI agents.</p>
+    </td>
+  </tr>
+</table>
+
+## Screenshots
+
+
+| Talk Interface | Generate Audio |
+| --- | --- |
+| ![Screenshot 2025-03-31 at 12-01-36 LocalAI - Talk](./docs/assets/images/screenshots/screenshot_tts.png) | ![Screenshot 2025-03-31 at 12-01-29 LocalAI - Generate audio with voice-en-us-ryan-low](./docs/assets/images/screenshots/screenshot_tts.png) |
+
+| Models Overview | Generate Images |
+| --- | --- |
+| ![Screenshot 2025-03-31 at 12-01-20 LocalAI - Models](./docs/assets/images/screenshots/screenshot_gallery.png) | ![Screenshot 2025-03-31 at 12-31-41 LocalAI - Generate images with flux 1-dev](./docs/assets/images/screenshots/screenshot_image.png) |
+
+| Chat Interface | Home |
+| --- | --- |
+| ![Screenshot 2025-03-31 at 11-57-44 LocalAI - Chat with localai-functioncall-qwen2 5-7b-v0 5](./docs/assets/images/screenshots/screenshot_chat.png) | ![Screenshot 2025-03-31 at 11-57-23 LocalAI API - c2a39e3 (c2a39e3639227cfd94ffffe9f5691239acc275a8)](./docs/assets/images/screenshots/screenshot_home.png) |
+
+| Login | Swarm |
+| --- | --- |
+|![Screenshot 2025-03-31 at 12-09-59 ](./docs/assets/images/screenshots/screenshot_login.png) | ![Screenshot 2025-03-31 at 12-10-39 LocalAI - P2P dashboard](./docs/assets/images/screenshots/screenshot_p2p.png) |
+
+## 💻 Quickstart
 
 Run the installer script:
 
 ```bash
+# Basic installation
 curl https://localai.io/install.sh | sh
 ```
 
+For more installation options, see [Installer Options](https://localai.io/docs/advanced/installer/).
+
 Or run with docker:
+
+### CPU only image:
 ```bash
-# CPU only image:
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-cpu
-
+```
-# Nvidia GPU:
+### Nvidia GPU:
+```bash
 docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-gpu-nvidia-cuda-12
-
+```
-# CPU and GPU image (bigger size):
+### CPU and GPU image (bigger size):
+```bash
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest
-
+```
-# AIO images (it will pre-download a set of models ready for use, see https://localai.io/basics/container/)
+### AIO images (it will pre-download a set of models ready for use, see https://localai.io/basics/container/)
+```bash
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-cpu
 ```
 
@@ -88,10 +144,13 @@ local-ai run https://gist.githubusercontent.com/.../phi-2.yaml
 local-ai run oci://localai/phi-2:latest
 ```
 
-[💻 Getting started](https://localai.io/basics/getting_started/index.html)
+For more information, see [💻 Getting started](https://localai.io/basics/getting_started/index.html)
 
 ## 📰 Latest project news
 
+- Apr 2025: [LocalAGI](https://github.com/mudler/LocalAGI) and [LocalRecall](https://github.com/mudler/LocalRecall) join the LocalAI family stack.
+- Apr 2025: WebUI overhaul, AIO images updates
+- Feb 2025: Backend cleanup, Breaking changes, new backends (kokoro, OutelTTS, faster-whisper), Nvidia L4T images
 - Jan 2025: LocalAI model release: https://huggingface.co/mudler/LocalAI-functioncall-phi-4-v0.3, SANA support in diffusers: https://github.com/mudler/LocalAI/pull/4603
 - Dec 2024: stablediffusion.cpp backend (ggml) added ( https://github.com/mudler/LocalAI/pull/4289 )
 - Nov 2024: Bark.cpp backend added ( https://github.com/mudler/LocalAI/pull/4287 )
@@ -105,19 +164,6 @@ local-ai run oci://localai/phi-2:latest
 
 Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap)
 
-## 🔥🔥 Hot topics (looking for help):
-
-- Multimodal with vLLM and Video understanding: https://github.com/mudler/LocalAI/pull/3729
-- Realtime API https://github.com/mudler/LocalAI/issues/3714
-- WebUI improvements: https://github.com/mudler/LocalAI/issues/2156
-- Backends v2: https://github.com/mudler/LocalAI/issues/1126
-- Improving UX v2: https://github.com/mudler/LocalAI/issues/1373
-- Assistant API: https://github.com/mudler/LocalAI/issues/1273
-- Vulkan: https://github.com/mudler/LocalAI/issues/1647
-- Anthropic API: https://github.com/mudler/LocalAI/issues/1808
-
-If you want to help and contribute, issues up for grabs: https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3A%22up+for+grabs%22
-
 ## 🚀 [Features](https://localai.io/features/)
 
 - 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `transformers`, `vllm` ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
@@ -131,12 +177,10 @@ If you want to help and contribute, issues up for grabs: https://github.com/mudl
 - 🥽 [Vision API](https://localai.io/features/gpt-vision/)
 - 📈 [Reranker API](https://localai.io/features/reranker/)
 - 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
+- [Agentic capabilities](https://github.com/mudler/LocalAGI)
 - 🔊 Voice activity detection (Silero-VAD support)
 - 🌍 Integrated WebUI!
 
-## 💻 Usage
-
-Check out the [Getting started](https://localai.io/basics/getting_started/index.html) section in our documentation.
 
 ### 🔗 Community and integrations
 

aio/gpu-8g/text-to-text.yaml

@@ -48,6 +48,6 @@ template:
     <|im_start|>assistant
 
 download_files:
-- filename: localai-functioncall-phi-4-v0.3-q4_k_m.gguf
+- filename: localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf
-  sha256: 23fee048ded2a6e2e1a7b6bbefa6cbf83068f194caa9552aecbaa00fec8a16d5
+  sha256: 4e7b7fe1d54b881f1ef90799219dc6cc285d29db24f559c8998d1addb35713d4
-  uri: huggingface://mudler/LocalAI-functioncall-phi-4-v0.3-Q4_K_M-GGUF/localai-functioncall-phi-4-v0.3-q4_k_m.gguf
+  uri: huggingface://mudler/LocalAI-functioncall-qwen2.5-7b-v0.5-Q4_K_M-GGUF/localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf

assets.go

@@ -1,6 +1,15 @@
 package main
 
-import "embed"
+import (
+	rice "github.com/GeertJohan/go.rice"
+)
 
-//go:embed backend-assets/*
+var backendAssets *rice.Box
-var backendAssets embed.FS
+
+func init() {
+	var err error
+	backendAssets, err = rice.FindBox("backend-assets")
+	if err != nil {
+		panic(err)
+	}
+}

backend/backend.proto

@@ -14,6 +14,7 @@ service Backend {
   rpc PredictStream(PredictOptions) returns (stream Reply) {}
   rpc Embedding(PredictOptions) returns (EmbeddingResult) {}
   rpc GenerateImage(GenerateImageRequest) returns (Result) {}
+  rpc GenerateVideo(GenerateVideoRequest) returns (Result) {}
   rpc AudioTranscription(TranscriptRequest) returns (TranscriptResult) {}
   rpc TTS(TTSRequest) returns (Result) {}
   rpc SoundGeneration(SoundGenerationRequest) returns (Result) {}
@@ -190,11 +191,7 @@ message ModelOptions {
   int32 NGQA = 20;
   string ModelFile = 21;
 
-  // AutoGPTQ
+
-  string Device = 22;
-  bool UseTriton = 23;
-  string ModelBaseName = 24;
-  bool UseFastTokenizer = 25;
 
   // Diffusers
   string PipelineType = 26;
@@ -305,6 +302,19 @@ message GenerateImageRequest {
   int32 CLIPSkip = 11;
 }
 
+message GenerateVideoRequest {
+  string prompt = 1;
+  string start_image = 2;  // Path or base64 encoded image for the start frame
+  string end_image = 3;    // Path or base64 encoded image for the end frame
+  int32 width = 4;
+  int32 height = 5;
+  int32 num_frames = 6;    // Number of frames to generate
+  int32 fps = 7;          // Frames per second
+  int32 seed = 8;
+  float cfg_scale = 9;    // Classifier-free guidance scale
+  string dst = 10;        // Output path for the generated video
+}
+
 message TTSRequest {
   string text = 1;
   string model = 2;

backend/cpp/llama/CMakeLists.txt

@@ -1,17 +1,17 @@
 
 ## XXX: In some versions of CMake clip wasn't being built before llama.
 ## This is an hack for now, but it should be fixed in the future.
-set(TARGET myclip)
+# set(TARGET myclip)
-add_library(${TARGET} clip.cpp clip.h llava.cpp llava.h)
+# add_library(${TARGET} clip.cpp clip.h clip-impl.h llava.cpp llava.h)
-install(TARGETS ${TARGET} LIBRARY)
+# install(TARGETS ${TARGET} LIBRARY)
-target_include_directories(myclip PUBLIC .)
+# target_include_directories(myclip PUBLIC .)
-target_include_directories(myclip PUBLIC ../..)
+# target_include_directories(myclip PUBLIC ../..)
-target_include_directories(myclip PUBLIC ../../common)
+# target_include_directories(myclip PUBLIC ../../common)
-target_link_libraries(${TARGET} PRIVATE common ggml llama ${CMAKE_THREAD_LIBS_INIT})
+# target_link_libraries(${TARGET} PRIVATE common ggml llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+# target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if (NOT MSVC)
+# if (NOT MSVC)
-    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
+#     target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
-endif()
+# endif()
 # END CLIP hack
 
 
@@ -75,7 +75,11 @@ add_library(hw_grpc_proto
   ${hw_proto_hdrs} )
 
 add_executable(${TARGET} grpc-server.cpp utils.hpp json.hpp)
-target_link_libraries(${TARGET} PRIVATE common llama myclip ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
+
+target_include_directories(${TARGET} PRIVATE ../llava)
+target_include_directories(${TARGET} PRIVATE ${CMAKE_SOURCE_DIR})
+
+target_link_libraries(${TARGET} PRIVATE common llama mtmd ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
   absl::flags_parse
   gRPC::${_REFLECTION}
   gRPC::${_GRPC_GRPCPP}

backend/cpp/llama/Makefile

@@ -8,7 +8,7 @@ ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 TARGET?=--target grpc-server
 
 # 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+=-DBUILD_SHARED_LIBS=OFF -DLLAMA_CURL=OFF
 
 # If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
@@ -36,11 +36,18 @@ else ifeq ($(OS),Darwin)
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)
-	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DCMAKE_CXX_FLAGS="-fsycl" \
+		-DGGML_SYCL_F16=ON
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f32)
-	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DCMAKE_CXX_FLAGS="-fsycl"
 endif
 
 llama.cpp:
@@ -52,8 +59,8 @@ llama.cpp:
 	git checkout -b build $(LLAMA_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-llama.cpp/examples/grpc-server: llama.cpp
+llama.cpp/tools/grpc-server: llama.cpp
-	mkdir -p llama.cpp/examples/grpc-server
+	mkdir -p llama.cpp/tools/grpc-server
 	bash prepare.sh
 
 rebuild:
@@ -63,13 +70,13 @@ rebuild:
 
 purge:
 	rm -rf llama.cpp/build
-	rm -rf llama.cpp/examples/grpc-server
+	rm -rf llama.cpp/tools/grpc-server
 	rm -rf grpc-server
 
 clean: purge
 	rm -rf llama.cpp
 
-grpc-server: llama.cpp llama.cpp/examples/grpc-server
+grpc-server: llama.cpp llama.cpp/tools/grpc-server
 	@echo "Building grpc-server with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	+bash -c "source $(ONEAPI_VARS); \
@@ -77,4 +84,4 @@ ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 else
 	+cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release $(TARGET)
 endif
-	cp llama.cpp/build/bin/grpc-server .
+	cp llama.cpp/build/bin/grpc-server .

backend/cpp/llama/grpc-server.cpp

@@ -11,8 +11,7 @@
 #include <memory>
 #include <string>
 #include <getopt.h>
-#include "clip.h"
+#include "mtmd.h"
-#include "llava.h"
 #include "log.h"
 #include "stb_image.h"
 #include "common.h"
@@ -52,7 +51,7 @@ struct server_params
 {
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
-    std::string public_path = "examples/server/public";
+    std::string public_path = "tools/server/public";
     std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -210,6 +209,8 @@ struct llama_client_slot
     int32_t num_prompt_tokens_processed = 0;
 
     json prompt;
+    json data;
+
     std::string generated_text;
     llama_token sampled;
     std::vector<llama_token> cache_tokens;
@@ -239,7 +240,7 @@ struct llama_client_slot
     int32_t n_past_se = 0; // self-extend
 
     // multimodal
-    std::vector<slot_image> images;
+    mtmd_context * mctx = nullptr;
 
     // stats
     size_t sent_count = 0;
@@ -270,17 +271,6 @@ struct llama_client_slot
         n_past_se              = 0;
 
         generated_token_probs.clear();
-
-        for (slot_image & img : images)
-        {
-            free(img.image_embedding);
-            if (img.img_data) {
-                clip_image_u8_free(img.img_data);
-            }
-            img.prefix_prompt = "";
-        }
-
-        images.clear();
     }
 
     bool has_budget(common_params &global_params) {
@@ -456,6 +446,9 @@ struct llama_server_context
     llama_context *ctx = nullptr;
     const llama_vocab * vocab = nullptr;
 
+    // multimodal
+    mtmd_context * mctx = nullptr;
+
     clip_ctx *clp_ctx = nullptr;
 
     common_params params;
@@ -494,6 +487,10 @@ struct llama_server_context
 
     ~llama_server_context()
     {
+        if (mctx) {
+            mtmd_free(mctx);
+            mctx = nullptr;
+        }
         if (ctx)
         {
             llama_free(ctx);
@@ -509,15 +506,17 @@ struct llama_server_context
     bool load_model(const common_params &params_)
     {
         params = params_;
-        if (!params.mmproj.empty()) {
+        if (!params.mmproj.path.empty()) {
             multimodal = true;
             LOG_INFO("Multi Modal Mode Enabled", {});
-            clp_ctx = clip_init(params.mmproj.c_str(), clip_context_params {
+            mtmd_context_params mparams = mtmd_context_params_default();
-                /* use_gpu */ has_gpu,
+            mparams.use_gpu       = has_gpu;
-                /*verbosity=*/ 1,
+            mparams.print_timings = false;
-            });
+            mparams.n_threads     = params.cpuparams.n_threads;
-            if(clp_ctx == nullptr) {
+            mparams.verbosity     = GGML_LOG_LEVEL_INFO;
-                LOG_ERR("unable to load clip model: %s", params.mmproj.c_str());
+            mctx = mtmd_init_from_file(params.mmproj.path.c_str(), model, mparams);
+            if (mctx == nullptr) {
+                LOG_ERR("failed to load multimodal model, '%s'\n", params.mmproj.path.c_str());
                 return false;
             }
 
@@ -531,7 +530,7 @@ struct llama_server_context
         ctx = common_init.context.release();
         if (model == nullptr)
         {
-            LOG_ERR("unable to load model: %s", params.model.c_str());
+            LOG_ERR("unable to load model: %s", params.model.path.c_str());
             return false;
         }
 
@@ -579,6 +578,8 @@ struct llama_server_context
             slot.id = i;
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
+            slot.mctx = mctx;
+            //slot.cache_tokens.has_mtmd = mctx != nullptr;
 
             LOG_INFO("new slot", {
                 {"slot_id",    slot.id},
@@ -616,54 +617,61 @@ struct llama_server_context
         batch = llama_batch_init(n_ctx, 0, params.n_parallel);
     }
 
-    std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
+    std::vector<server_tokens> tokenize(json &data, const json & json_prompt, bool add_bos) const
     {
-        // TODO: currently, we tokenize using special tokens by default
+        mtmd::bitmaps bitmaps;
-        //       this is not always correct (see https://github.com/ggerganov/llama.cpp/pull/4160#issuecomment-1824826216)
+        std::vector<server_tokens> inputs;
-        //       but it's better compared to completely ignoring ChatML and other chat templates
-        const bool TMP_FORCE_SPECIAL = true;
 
-        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+        if (mctx != nullptr)
-        // or the first element of the json_prompt array is a string.
-        std::vector<llama_token> prompt_tokens;
-
-        if (json_prompt.is_array())
         {
-            bool first = true;
+            const auto &images_data = data.find("image_data");
-            for (const auto& p : json_prompt)
+            if (images_data != data.end() && images_data->is_array())
             {
-                if (p.is_string())
+                for (const auto &img : *images_data)
                 {
-                    auto s = p.template get<std::string>();
+                    const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
-                    std::vector<llama_token> p;
+
-                    if (first)
+                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(image_buffer.data(), image_buffer.size()));
-                    {
+                    if (!bmp.ptr) {
-                        p = common_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
+                            throw std::runtime_error("Failed to load image");
-                        first = false;
                     }
-                    else
+                    // calculate bitmap hash (for KV caching)
-                    {
+                    std::string hash = fnv_hash(bmp.data(), bmp.nx()*bmp.ny()*3);
-                        p = common_tokenize(ctx, s, false, TMP_FORCE_SPECIAL);
+                    bmp.set_id(hash.c_str());
-                    }
+                    bitmaps.entries.push_back(std::move(bmp));
-                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
-                }
-                else
-                {
-                    if (first)
-                    {
-                        first = false;
-                    }
-                    prompt_tokens.push_back(p.template get<llama_token>());
                 }
             }
-        }
+
-        else
+            // multimodal
-        {
+            std::string prompt_str = json_prompt.template get<std::string>();
-            auto s = json_prompt.template get<std::string>();
+            mtmd_input_text inp_txt = {
-            prompt_tokens = common_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
+                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(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));
+        } else {
+            // non-multimodal version
+            auto tokenized_prompts = tokenize_input_prompts(vocab, json_prompt, true, true);
+            for (auto & p : tokenized_prompts) {
+                auto tmp = server_tokens(p, mctx != nullptr);
+                inputs.push_back(std::move(tmp));
+            }
         }
 
-        return prompt_tokens;
+        return inputs;
     }
 
     llama_client_slot* get_slot(int id) {
@@ -716,6 +724,8 @@ struct llama_server_context
         slot->sparams.grammar_triggers = grammar_triggers;
         slot->sparams.grammar_lazy = grammar_lazy;
 
+        slot->data = data;
+
         if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
             // Might be better to reject the request with a 400 ?
             LOG_WARNING("Max tokens to predict exceeds server configuration", {
@@ -757,43 +767,7 @@ struct llama_server_context
         if (json_value(data, "ignore_eos", false) && has_eos_token) {
                 slot->sparams.logit_bias.push_back({llama_vocab_eos(vocab), -INFINITY});
         }
-        /*
+        
-        slot->sparams.penalty_prompt_tokens.clear();
-        slot->sparams.use_penalty_prompt_tokens = false;
-        const auto &penalty_prompt = data.find("penalty_prompt");
-        if (penalty_prompt != data.end())
-        {
-            if (penalty_prompt->is_string())
-            {
-                const auto penalty_prompt_string = penalty_prompt->get<std::string>();
-                auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false);
-                slot->sparams.penalty_prompt_tokens.swap(penalty_tokens);
-                if (slot->params.n_predict > 0)
-                {
-                    slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict);
-                }
-                slot->sparams.use_penalty_prompt_tokens = true;
-            }
-            else if (penalty_prompt->is_array())
-            {
-                const auto n_tokens = penalty_prompt->size();
-                slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict));
-                const int n_vocab = llama_n_vocab(model);
-                for (const auto &penalty_token : *penalty_prompt)
-                {
-                    if (penalty_token.is_number_integer())
-                    {
-                        const auto tok = penalty_token.get<llama_token>();
-                        if (tok >= 0 && tok < n_vocab)
-                        {
-                            slot->sparams.penalty_prompt_tokens.push_back(tok);
-                        }
-                    }
-                }
-                slot->sparams.use_penalty_prompt_tokens = true;
-            }
-        }
-      */
         slot->sparams.logit_bias.clear();
 
         const auto &logit_bias = data.find("logit_bias");
@@ -869,79 +843,6 @@ struct llama_server_context
         }
         
 
-        if (multimodal)
-        {
-            const auto &images_data = data.find("image_data");
-            if (images_data != data.end() && images_data->is_array())
-            {
-                for (const auto &img : *images_data)
-                {
-                    const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
-
-                    slot_image img_sl;
-                    img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
-                    img_sl.img_data = clip_image_u8_init();
-                    if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
-                    {
-                        LOG_ERR("%s: failed to load image, slot_id: %d, img_sl_id: %d", 
-                             __func__,
-                             slot->id,
-                             img_sl.id
-                        );
-                        return false;
-                    }
-                    LOG_VERBOSE("image loaded", {
-                        {"slot_id",   slot->id},
-                        {"img_sl_id", img_sl.id}
-                    });
-                    img_sl.request_encode_image = true;
-                    slot->images.push_back(img_sl);
-                }
-                // process prompt
-                // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]}
-                if (slot->images.size() > 0 && !slot->prompt.is_array())
-                {
-                    std::string prompt = slot->prompt.get<std::string>();
-                    size_t pos = 0, begin_prefix = 0;
-                    std::string pattern = "[img-";
-                    while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
-                        size_t end_prefix = pos;
-                        pos += pattern.length();
-                        size_t end_pos = prompt.find(']', pos);
-                        if (end_pos != std::string::npos)
-                        {
-                            std::string image_id = prompt.substr(pos, end_pos - pos);
-                            try
-                            {
-                                int img_id = std::stoi(image_id);
-                                bool found = false;
-                                for (slot_image &img : slot->images)
-                                {
-                                    if (img.id == img_id) {
-                                        found = true;
-                                        img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix);
-                                        begin_prefix = end_pos + 1;
-                                        break;
-                                    }
-                                }
-                                if (!found) {
-                                    LOG("ERROR: Image with id: %i, not found.\n", img_id);
-                                    slot->images.clear();
-                                    return false;
-                                }
-                            } catch (const std::invalid_argument& e) {
-                                LOG("Invalid image number id in prompt\n");
-                                slot->images.clear();
-                                return false;
-                            }
-                        }
-                    }
-                    slot->prompt = "";
-                    slot->params.input_suffix = prompt.substr(begin_prefix);
-                    slot->params.cache_prompt = false; // multimodal doesn't support cache prompt
-                }
-            }
-        }
 
         if (slot->ctx_sampling != nullptr)
         {
@@ -1189,26 +1090,6 @@ struct llama_server_context
         return slot.has_next_token; // continue
     }
 
-    bool process_images(llama_client_slot &slot) const
-    {
-        for (slot_image &img : slot.images)
-        {
-            if (!img.request_encode_image)
-            {
-                continue;
-            }
-
-            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
-                LOG("Error processing the given image");
-                return false;
-            }
-
-            img.request_encode_image = false;
-        }
-
-        return slot.images.size() > 0;
-    }
-
     void send_error(task_server& task, const std::string &error)
     {
         LOG("task %i - error: %s\n", task.id, error.c_str());
@@ -1451,74 +1332,6 @@ struct llama_server_context
         }
     }
 
-    // for multiple images processing
-    bool ingest_images(llama_client_slot &slot, int n_batch)
-    {
-        int image_idx = 0;
-
-        while (image_idx < (int) slot.images.size())
-        {
-            slot_image &img = slot.images[image_idx];
-
-            // process prefix prompt
-            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
-            {
-                const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
-                llama_batch batch_view = {
-                    n_tokens,
-                    batch.token    + i,
-                    nullptr,
-                    batch.pos      + i,
-                    batch.n_seq_id + i,
-                    batch.seq_id   + i,
-                    batch.logits   + i,
-                };
-                if (llama_decode(ctx, batch_view))
-                {
-                    LOG("%s : failed to eval\n", __func__);
-                    return false;
-                }
-            }
-
-            // process image with llm
-            for (int i = 0; i < img.image_tokens; i += n_batch)
-            {
-                int n_eval = img.image_tokens - i;
-                if (n_eval > n_batch)
-                {
-                    n_eval = n_batch;
-                }
-
-                const int n_embd = llama_model_n_embd(model);
-                float * embd = img.image_embedding + i * n_embd;
-                llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, slot.n_past, 0);
-                if (llama_decode(ctx, llava_batch.batch))
-                {
-                    LOG("%s : failed to eval image\n", __func__);
-                    return false;
-                }
-                slot.n_past += n_eval;
-            }
-            image_idx++;
-
-            common_batch_clear(batch);
-
-            // append prefix of next image
-            const auto json_prompt = (image_idx >= (int) slot.images.size()) ?
-                slot.params.input_suffix : // no more images, then process suffix prompt
-                (json)(slot.images[image_idx].prefix_prompt);
-
-            std::vector<llama_token> append_tokens = tokenize(json_prompt, false); // has next image
-            for (int i = 0; i < (int) append_tokens.size(); ++i)
-            {
-                common_batch_add(batch, append_tokens[i], system_tokens.size() + slot.n_past, { slot.id }, true);
-                slot.n_past += 1;
-            }
-        }
-
-        return true;
-    }
-
     void request_cancel(int task_id)
     {
         task_server task;
@@ -1733,7 +1546,7 @@ struct llama_server_context
         {
             for (auto & slot : slots)
             {
-                const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
+                const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty());
 
                 // empty prompt passed -> release the slot and send empty response
                 // note: infill mode allows empty prompt
@@ -1750,7 +1563,7 @@ struct llama_server_context
                 {
                     slot.state = PROCESSING;
                     slot.command = NONE;
-                    std::vector<llama_token> prompt_tokens;
+                    std::vector<server_tokens> prompt_tokens;
                     slot.t_start_process_prompt = ggml_time_us();
                     slot.t_start_genereration = 0;
 
@@ -1762,24 +1575,41 @@ struct llama_server_context
                             params.input_suffix.erase(0, 1);
                             suff_rm_leading_spc = false;
                         }
-                        auto prefix_tokens = tokenize(slot.params.input_prefix, false);
+                        auto prefix_tokens = tokenize(slot.data, slot.params.input_prefix, false);
-                        auto suffix_tokens = tokenize(slot.params.input_suffix, false);
+                        auto suffix_tokens = tokenize(slot.data, slot.params.input_suffix, false);
 
                         const int space_token = 29871; // TODO: this should not be hardcoded
-                        if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
+                        if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0][0] == space_token) {
                             suffix_tokens.erase(suffix_tokens.begin());
                         }
 
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_vocab_fim_pre(vocab));
+                        // Create llama_tokens vectors for the special tokens
-                        prefix_tokens.insert(prefix_tokens.begin(), llama_vocab_bos(vocab)); // always add BOS
+                        llama_tokens fim_pre_tokens;
-                        prefix_tokens.insert(prefix_tokens.end(),   llama_vocab_fim_suf(vocab));
+                        fim_pre_tokens.push_back(llama_vocab_fim_pre(vocab));
-                        prefix_tokens.insert(prefix_tokens.end(),   suffix_tokens.begin(), suffix_tokens.end());
+                        llama_tokens bos_tokens;
-                        prefix_tokens.push_back(llama_vocab_fim_mid(vocab));
+                        bos_tokens.push_back(llama_vocab_bos(vocab));
+                        llama_tokens fim_suf_tokens;
+                        fim_suf_tokens.push_back(llama_vocab_fim_suf(vocab));
+                        llama_tokens fim_mid_tokens;
+                        fim_mid_tokens.push_back(llama_vocab_fim_mid(vocab));
+
+                        // Create server_tokens objects
+                        server_tokens fim_pre_token(fim_pre_tokens, mctx != nullptr);
+                        server_tokens bos_token(bos_tokens, mctx != nullptr);
+                        server_tokens fim_suf_token(fim_suf_tokens, mctx != nullptr);
+                        server_tokens fim_mid_token(fim_mid_tokens, mctx != nullptr);
+
+                        // Insert tokens in the correct order
+                        prefix_tokens.insert(prefix_tokens.begin(), fim_pre_token);
+                        prefix_tokens.insert(prefix_tokens.begin(), bos_token); // always add BOS
+                        prefix_tokens.insert(prefix_tokens.end(), fim_suf_token);
+                        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+                        prefix_tokens.push_back(fim_mid_token);
                         prompt_tokens = prefix_tokens;
                     }
                     else
                     {
-                        prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token);  // add BOS if there isn't system prompt
+                        prompt_tokens = tokenize(slot.data, slot.prompt, system_prompt.empty() && add_bos_token);  // add BOS if there isn't system prompt
                     }
 
                     slot.num_prompt_tokens = prompt_tokens.size();
@@ -1807,7 +1637,12 @@ struct llama_server_context
                             {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
                         });
                         slot.truncated = true;
-                        prompt_tokens = new_tokens;
+                        
+                        // Convert new_tokens to server_tokens
+                        std::vector<server_tokens> new_prompt_tokens;
+                        server_tokens new_server_tokens(new_tokens, mctx != nullptr);
+                        new_prompt_tokens.push_back(std::move(new_server_tokens));
+                        prompt_tokens = std::move(new_prompt_tokens);
 
                         slot.num_prompt_tokens = prompt_tokens.size();
                         GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx);
@@ -1827,10 +1662,17 @@ struct llama_server_context
                         // push the prompt into the sampling context (do not apply grammar)
                         for (auto &token : prompt_tokens)
                         {
-                            common_sampler_accept(slot.ctx_sampling, token, false);
+                            // Convert server_tokens to llama_token for sampling
+                            llama_token tok = token[0];  // Get first token
+                            common_sampler_accept(slot.ctx_sampling, tok, false);
                         }
 
-                        slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
+                        // Convert server_tokens to llama_tokens for comparison
+                        std::vector<llama_token> prompt_llama_tokens;
+                        for (const auto &token : prompt_tokens) {
+                            prompt_llama_tokens.push_back(token[0]);
+                        }
+                        slot.n_past = common_part(slot.cache_tokens, prompt_llama_tokens);
 
                         // the last token of the cache is not in the KV cache until the next call to llama_decode
                         // (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
@@ -1868,7 +1710,12 @@ struct llama_server_context
                         });
                     }
 
-                    slot.cache_tokens = prompt_tokens;
+                    // Convert server_tokens to llama_tokens for cache
+                    std::vector<llama_token> cache_llama_tokens;
+                    for (const auto &token : prompt_tokens) {
+                        cache_llama_tokens.push_back(token[0]);
+                    }
+                    slot.cache_tokens = cache_llama_tokens;
 
                     if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
                     {
@@ -1892,18 +1739,36 @@ struct llama_server_context
                     });
                     llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
 
+
+                    // process the prefix of first image
+                    std::vector<server_tokens> prefix_tokens = prompt_tokens;
+
+                    int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
+
+                    // check if we should process the image
+                    if (slot.n_past < slot.n_prompt_tokens
+                            && slot.prompt_tokens[slot.n_past] == LLAMA_TOKEN_NULL) {
+                        // process the image
+                        int32_t new_n_past;
+                        int32_t res = prompt_tokens.process_chunk(ctx, mctx, slot.n_past, slot.id, new_n_past);
+                        int32_t n_pos = new_n_past - slot.n_past;
+                        if (res != 0) {
+                            slot.release();
+                            LOG_ERR("failed to process image, res = %d\n", res);
+                            continue;
+                        }
+
+
+                        slot.n_past                    += n_pos;
+                       // slot.n_prompt_tokens_processed += n_pos;
+                    }
+
                     LOG_VERBOSE("prompt ingested", {
                                                     {"n_past",  slot.n_past},
                                                     {"cached",  tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)},
                                                     {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())},
                                                 });
 
-                    const bool has_images = process_images(slot);
-
-                    // process the prefix of first image
-                    std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
-
-                    int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
 
                     int32_t ga_i = slot.ga_i;
                     int32_t ga_n = slot.ga_n;
@@ -1923,19 +1788,6 @@ struct llama_server_context
                         slot_npast++;
                     }
 
-                    if (has_images && !ingest_images(slot, n_batch))
-                    {
-                        LOG_ERR("%s: failed processing images Slot id : %d, Task id: %d", 
-                            __func__,
-                            slot.id,
-                            slot.task_id
-                        );
-                        // FIXME @phymbert: to be properly tested
-                        //  early returning without changing the slot state will block the slot for ever
-                        // no one at the moment is checking the return value
-                        return false;
-                    }
-
                     // extract the logits only for the last token
                     if (batch.n_tokens > 0)
                     {
@@ -2164,26 +2016,6 @@ static void start_llama_server() {
 json parse_options(bool streaming, const backend::PredictOptions* predict, llama_server_context &llama)
 {
     
-    // This is for example a slot data from the json data
-    //     slot->params.stream           = json_value(data, "stream",            false);
-    //     slot->params.cache_prompt     = json_value(data, "cache_prompt",      false);
-    //     slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
-    //     slot->sparams.top_k           = json_value(data, "top_k",             default_sparams.top_k);
-    //     slot->sparams.top_p           = json_value(data, "top_p",             default_sparams.top_p);
-    //     slot->sparams.typical_p       = json_value(data, "typical_p",         default_sparams.typical_p);
-    //     slot->sparams.temp            = json_value(data, "temperature",       default_sparams.temp);
-    //     slot->sparams.penalty_last_n  = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
-    //     slot->sparams.penalty_repeat  = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
-    //     slot->sparams.penalty_freq    = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
-    //     slot->sparams.penalty_present = json_value(data, "presence_penalty",  default_sparams.penalty_present);
-    //     slot->sparams.mirostat        = json_value(data, "mirostat",          default_sparams.mirostat);
-    //     slot->sparams.mirostat_tau    = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
-    //     slot->sparams.mirostat_eta    = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-    //     slot->params.n_keep           = json_value(data, "n_keep",            slot->params.n_keep);
-    //     slot->params.seed             = json_value(data, "seed",              default_params.seed);
-    //     slot->sparams.grammar         = json_value(data, "grammar",           default_sparams.grammar);
-    //     slot->sparams.n_probs         = json_value(data, "n_probs",           default_sparams.n_probs);
-
     // Create now a json data from the prediction options instead
     //
     json data;
@@ -2228,69 +2060,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     return data;
 }
 
-// static void parse_options_completion(bool streaming,const backend::PredictOptions* predict, llama_server_context &llama)
-// {
-//     // https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L673
-//     gpt_params default_params;
-
-//     llama.stream = streaming;
-//     llama.params.n_predict = predict->tokens() == 0 ? -1 : predict->tokens();
-//     llama.params.sparams.top_k = predict->topk();
-//     llama.params.sparams.top_p = predict->topp();
-//     llama.params.sparams.typical_p = predict->typicalp();
-//     llama.params.sparams.penalty_last_n = predict->repeat();
-//     llama.params.sparams.temp = predict->temperature();
-//     llama.params.sparams.penalty_repeat = predict->penalty();
-//     llama.params.sparams.penalty_present = predict->presencepenalty();
-//     llama.params.sparams.penalty_freq = predict->frequencypenalty();
-//     llama.params.sparams.mirostat = predict->mirostat();
-//     llama.params.sparams.mirostat_tau = predict->mirostattau();
-//     llama.params.sparams.mirostat_eta = predict->mirostateta();
-//     llama.params.n_keep = predict->nkeep();
-//     llama.params.seed = predict->seed();
-//     llama.params.sparams.grammar = predict->grammar();
-//     // llama.params.n_probs = predict->
-//     llama.params.prompt = predict->prompt();
-
-//     llama.params.sparams.logit_bias.clear();
-
-//     if (predict->ignoreeos())
-//     {
-//         llama.params.sparams.logit_bias[llama_token_eos(llama.model)] = -INFINITY;
-//     }
-
-//     // const auto &logit_bias = body.find("logit_bias");
-//     // if (logit_bias != body.end() && logit_bias->is_array())
-//     // {
-//     //     const int n_vocab = llama_n_vocab(llama.model);
-//     //     for (const auto &el : *logit_bias)
-//     //     {
-//     //         if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
-//     //         {
-//     //             llama_token tok = el[0].get<llama_token>();
-//     //             if (tok >= 0 && tok < n_vocab)
-//     //             {
-//     //                 if (el[1].is_number())
-//     //                 {
-//     //                     llama.params.logit_bias[tok] = el[1].get<float>();
-//     //                 }
-//     //                 else if (el[1].is_boolean() && !el[1].get<bool>())
-//     //                 {
-//     //                     llama.params.logit_bias[tok] = -INFINITY;
-//     //                 }
-//     //             }
-//     //         }
-//     //     }
-//     // }
-
-//     llama.params.antiprompt.clear();
-//     for (const std::string& stopPrompt : predict->stopprompts()) {
-//     if (!stopPrompt.empty())
-//             {
-//                 llama.params.antiprompt.push_back(stopPrompt);
-//             }
-//     }
-// }
 
 const std::vector<ggml_type> kv_cache_types = {
     GGML_TYPE_F32,
@@ -2326,11 +2095,11 @@ static void params_parse(const backend::ModelOptions* request,
    
     // this is comparable to: https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L1809
 
-    params.model = request->modelfile();
+    params.model.path = request->modelfile();
     if (!request->mmproj().empty()) {
     // get the directory of modelfile
-      std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
+      std::string model_dir = params.model.path.substr(0, params.model.path.find_last_of("/\\"));
-      params.mmproj = model_dir + "/"+ request->mmproj();
+      params.mmproj.path = model_dir + "/"+ request->mmproj();
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
@@ -2405,7 +2174,7 @@ static void params_parse(const backend::ModelOptions* request,
         scale_factor = request->lorascale();
      }
      // get the directory of modelfile
-     std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
+     std::string model_dir = params.model.path.substr(0, params.model.path.find_last_of("/\\"));
      params.lora_adapters.push_back({ model_dir + "/"+request->loraadapter(), scale_factor });
     }
     params.use_mlock = request->mlock();
@@ -2603,10 +2372,10 @@ public:
     grpc::Status TokenizeString(ServerContext* context, const backend::PredictOptions* request, backend::TokenizationResponse* response){
          json data = parse_options(false, request, llama);
 
-         std::vector<llama_token> tokens = llama.tokenize(data["prompt"],false);
+         std::vector<server_tokens> tokens = llama.tokenize(data, data["prompt"],false);
 
          for (int i=0 ; i< tokens.size(); i++){
-            response->add_tokens(tokens[i]);
+            response->add_tokens(tokens[i].llama_token);
          }
 
         return grpc::Status::OK;
@@ -2644,7 +2413,9 @@ void RunServer(const std::string& server_address) {
   ServerBuilder builder;
   builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
   builder.RegisterService(&service);
-
+  builder.SetMaxMessageSize(50 * 1024 * 1024); // 50MB
+  builder.SetMaxSendMessageSize(50 * 1024 * 1024); // 50MB
+  builder.SetMaxReceiveMessageSize(50 * 1024 * 1024); // 50MB
   std::unique_ptr<Server> server(builder.BuildAndStart());
   std::cout << "Server listening on " << server_address << std::endl;
   server->Wait();

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

@@ -1,7 +1,7 @@
-diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
 index 3cd0d2fa..6c5e811a 100644
---- a/examples/llava/clip.cpp
+--- a/tools/mtmd/clip.cpp
-+++ b/examples/llava/clip.cpp
++++ b/tools/mtmd/clip.cpp
 @@ -2608,7 +2608,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
                  struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
                  int* patches_data = (int*)malloc(ggml_nbytes(patches));

backend/cpp/llama/prepare.sh

@@ -7,21 +7,22 @@ for patch in $(ls patches); do
     patch -d llama.cpp/ -p1 < patches/$patch
 done 
 
-cp -r CMakeLists.txt llama.cpp/examples/grpc-server/
+cp -r CMakeLists.txt llama.cpp/tools/grpc-server/
-cp -r grpc-server.cpp llama.cpp/examples/grpc-server/
+cp -r grpc-server.cpp llama.cpp/tools/grpc-server/
-cp -rfv json.hpp llama.cpp/examples/grpc-server/
+cp -rfv json.hpp llama.cpp/tools/grpc-server/
-cp -rfv utils.hpp llama.cpp/examples/grpc-server/
+cp -rfv utils.hpp llama.cpp/tools/grpc-server/
     
-if grep -q "grpc-server" llama.cpp/examples/CMakeLists.txt; then
+if grep -q "grpc-server" llama.cpp/tools/CMakeLists.txt; then
     echo "grpc-server already added"
 else
-    echo "add_subdirectory(grpc-server)" >> llama.cpp/examples/CMakeLists.txt
+    echo "add_subdirectory(grpc-server)" >> llama.cpp/tools/CMakeLists.txt
 fi
 
 ## XXX: In some versions of CMake clip wasn't being built before llama.
 ## This is an hack for now, but it should be fixed in the future.
-cp -rfv llama.cpp/examples/llava/clip.h llama.cpp/examples/grpc-server/clip.h
+# cp -rfv llama.cpp/tools/mtmd/clip.h llama.cpp/tools/grpc-server/clip.h
-cp -rfv llama.cpp/examples/llava/llava.cpp llama.cpp/examples/grpc-server/llava.cpp
+# cp -rfv llama.cpp/tools/mtmd/clip-impl.h llama.cpp/tools/grpc-server/clip-impl.h
-echo '#include "llama.h"' > llama.cpp/examples/grpc-server/llava.h
+# cp -rfv llama.cpp/tools/mtmd/llava.cpp llama.cpp/tools/grpc-server/llava.cpp
-cat llama.cpp/examples/llava/llava.h >> llama.cpp/examples/grpc-server/llava.h
+# echo '#include "llama.h"' > llama.cpp/tools/grpc-server/llava.h
-cp -rfv llama.cpp/examples/llava/clip.cpp llama.cpp/examples/grpc-server/clip.cpp
+# cat llama.cpp/tools/mtmd/llava.h >> llama.cpp/tools/grpc-server/llava.h
+# cp -rfv llama.cpp/tools/mtmd/clip.cpp llama.cpp/tools/grpc-server/clip.cpp

backend/cpp/llama/utils.hpp

@@ -1,4 +1,4 @@
-// https://github.com/ggerganov/llama.cpp/blob/master/examples/server/utils.hpp
+// https://github.com/ggerganov/llama.cpp/blob/master/tools/server/utils.hpp
 
 #pragma once
 
@@ -11,7 +11,7 @@
 
 #include "json.hpp"
 
-#include "../llava/clip.h"
+#include "../mtmd/clip.h"
 
 using json = nlohmann::json;
 
@@ -480,4 +480,431 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
     }
 
     return ret;
+
+}
+
+
+
+//
+// tokenizer and input processing utils
+//
+
+static bool json_is_array_of_numbers(const json & data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (!e.is_number_integer()) {
+                return false;
+            }
+        }
+        return true;
+    }
+    return false;
+}
+
+// is array having BOTH numbers & strings?
+static bool json_is_array_of_mixed_numbers_strings(const json & data) {
+    bool seen_string = false;
+    bool seen_number = false;
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            seen_string |= e.is_string();
+            seen_number |= e.is_number_integer();
+            if (seen_number && seen_string) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+// get value by path(key1 / key2)
+static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
+    json result = json::object();
+
+    for (const std::string & path : paths) {
+        json current = js;
+        const auto keys = string_split<std::string>(path, /*separator*/ '/');
+        bool valid_path = true;
+        for (const std::string & k : keys) {
+            if (valid_path && current.is_object() && current.contains(k)) {
+                current = current[k];
+            } else {
+                valid_path = false;
+            }
+        }
+        if (valid_path) {
+            result[path] = current;
+        }
+    }
+    return result;
+}
+
+
+/**
+ * this handles 2 cases:
+ * - only string, example: "string"
+ * - mixed string and tokens, example: [12, 34, "string", 56, 78]
+ */
+static llama_tokens tokenize_mixed(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
+    // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+    // or the first element of the json_prompt array is a string.
+    llama_tokens prompt_tokens;
+
+    if (json_prompt.is_array()) {
+        bool first = true;
+        for (const auto & p : json_prompt) {
+            if (p.is_string()) {
+                auto s = p.template get<std::string>();
+
+                llama_tokens p;
+                if (first) {
+                    p = common_tokenize(vocab, s, add_special, parse_special);
+                    first = false;
+                } else {
+                    p = common_tokenize(vocab, s, false, parse_special);
+                }
+
+                prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+            } else {
+                if (first) {
+                    first = false;
+                }
+
+                prompt_tokens.push_back(p.template get<llama_token>());
+            }
+        }
+    } else {
+        auto s = json_prompt.template get<std::string>();
+        prompt_tokens = common_tokenize(vocab, s, add_special, parse_special);
+    }
+
+    return prompt_tokens;
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * and multiple prompts (multi-tasks):
+ * - "prompt": ["string1", "string2"]
+ * - "prompt": ["string1", [12, 34, 56]]
+ * - "prompt": [[12, 34, 56], [78, 90, 12]]
+ * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
+ */
+static std::vector<llama_tokens> tokenize_input_prompts(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
+    std::vector<llama_tokens> result;
+    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
+        // string or mixed
+        result.push_back(tokenize_mixed(vocab, json_prompt, add_special, parse_special));
+    } else if (json_is_array_of_numbers(json_prompt)) {
+        // array of tokens
+        result.push_back(json_prompt.get<llama_tokens>());
+    } else if (json_prompt.is_array()) {
+        // array of prompts
+        result.reserve(json_prompt.size());
+        for (const auto & p : json_prompt) {
+            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
+                result.push_back(tokenize_mixed(vocab, p, add_special, parse_special));
+            } else if (json_is_array_of_numbers(p)) {
+                // array of tokens
+                result.push_back(p.get<llama_tokens>());
+            } else {
+                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
+            }
+        }
+    } else {
+        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
+    }
+    if (result.empty()) {
+        throw std::runtime_error("\"prompt\" must not be empty");
+    }
+    return result;
+}
+
+
+
+
+//
+// utils for interacting with libmtmd
+// (may need to refactor in near future)
+//
+
+/**
+ * server_tokens is a helper to manage the input tokens and image for the server.
+ * it is made this way to simplify the logic of KV cache management.
+ */
+struct server_tokens {
+    bool has_mtmd = false;
+
+private: // disallow accessing these members directly, risking out-of-sync
+
+    // map a **start** position in tokens to the image chunk
+    std::unordered_map<llama_pos, mtmd::input_chunk_ptr> map_pos_to_image;
+
+    // list of tokens
+    // it can include LLAMA_TOKEN_NULL, which is used to indicate a token that is not a text token
+    // a mtmd_input_chunk can occupy multiple tokens, one llama_token per **position**
+    // important: for models using mrope, an image can contain multiple tokens but will use only one **position**
+    llama_tokens tokens;
+
+    // for ex. with input of 5 text tokens and 2 images:
+    //      [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1]
+    // pos  0   1   2   3   4   5      6      7      8      9
+    // map_pos_to_image will contain: {5, img0}, {8, img1}
+
+public:
+    server_tokens() = default;
+    ~server_tokens() = default;
+
+    // Prevent copying
+    server_tokens(const server_tokens&) = delete;
+    server_tokens& operator=(const server_tokens&) = delete;
+
+    // Allow moving (usually implicitly generated if members are movable)
+    server_tokens(server_tokens&&) = default;
+    server_tokens& operator=(server_tokens&&) = default;
+
+    // Allow accessing elements using [] operator
+    llama_token operator[](size_t index) { return tokens[index]; }
+    const llama_token& operator[](size_t index) const { return tokens[index]; }
+
+    server_tokens(mtmd::input_chunks & mtmd_chunks, bool has_mtmd) : has_mtmd(has_mtmd) {
+        for (size_t i = 0; i < mtmd_chunks.size(); ++i) {
+            push_back(mtmd_chunks[i]);
+        }
+    }
+
+    server_tokens(llama_tokens & tokens, bool has_mtmd) : has_mtmd(has_mtmd), tokens(tokens) {}
+
+    // for debugging
+    std::string str() const {
+        std::ostringstream oss;
+        oss << "tokens: ";
+        for (const auto & t : tokens) {
+            if (t == LLAMA_TOKEN_NULL) {
+                oss << "<embd> ";
+            } else {
+                oss << t << " ";
+            }
+        }
+        oss << "\n";
+        oss << "image pos: ";
+        for (const auto & it : map_pos_to_image) {
+            oss << it.first << ", ";
+        }
+        return oss.str();
+    }
+
+    const mtmd::input_chunk_ptr & find_chunk(llama_pos pos) const {
+        auto it = map_pos_to_image.find(pos);
+        if (it != map_pos_to_image.end()) {
+            return it->second;
+        } else {
+            throw std::runtime_error("Chunk not found");
+        }
+    }
+
+    void push_back(llama_token tok) {
+        if (tok == LLAMA_TOKEN_NULL) {
+            throw std::runtime_error("Invalid token");
+        }
+        tokens.emplace_back(tok);
+    }
+
+    // will create a copy of the chunk if it contains non-text data
+    void push_back(const mtmd_input_chunk * chunk) {
+        auto type = mtmd_input_chunk_get_type(chunk);
+        if (type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
+            GGML_ASSERT(has_mtmd);
+            auto img_tokens = mtmd_input_chunk_get_tokens_image(chunk);
+            const int n_pos = mtmd_image_tokens_get_n_pos(img_tokens);
+            llama_pos start_pos = tokens.size();
+            for (int i = 0; i < n_pos; ++i) {
+                tokens.emplace_back(LLAMA_TOKEN_NULL);
+            }
+            mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_copy(chunk));
+            map_pos_to_image[start_pos] = std::move(new_chunk);
+        } else if (type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
+            size_t n_tokens;
+            auto text_tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
+            for (size_t i = 0; i < n_tokens; ++i) {
+                push_back(text_tokens[i]);
+            }
+        } else {
+            GGML_ABORT("Invalid chunk type");
+        }
+    }
+
+    // for compatibility with context shift and prompt truncation
+    void insert(const llama_tokens & inp_tokens) {
+        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
+        tokens.insert(tokens.end(), inp_tokens.begin(), inp_tokens.end());
+    }
+
+    // for compatibility with speculative decoding, ctx shift, slot save/load
+    const llama_tokens & get_text_tokens() const {
+        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
+        return tokens;
+    }
+
+    // for compatibility with speculative decoding
+    void set_token(llama_pos pos, llama_token id) {
+        GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
+        tokens[pos] = id;
+    }
+
+    size_t size() const {
+        return tokens.size();
+    }
+
+    bool empty() const {
+        return tokens.empty();
+    }
+
+    void clear() {
+        tokens.clear();
+    }
+
+    void resize(size_t n) {
+        GGML_ASSERT(n <= tokens.size());
+        if (has_mtmd) {
+            // we throw an error if we try to remove a token in the middle of an image
+            // for ex. with input of 5 text tokens and 2 images:
+            //    [0] [1] [2] [3] [4] [img0] [img0] [img0] [img1] [img1]
+            // n  1   2   3   4   5   6      7      8      9      10
+            // allowed to resize      ^                    ^
+            // disallowed to resize          ^      ^             ^
+            if (n > 0) {
+                llama_token last_token = tokens[n - 1];
+                // make sure we never remove tokens in the middle of an image
+                if (last_token == LLAMA_TOKEN_NULL) {
+                    find_chunk(n - 1); // will throw an error if the token is not begin-of-chunk
+                }
+            }
+            // remove all image chunks that are not used anymore
+            for (auto it = map_pos_to_image.begin(); it != map_pos_to_image.end(); ) {
+                llama_pos pos = it->first;
+                if (pos >= (llama_pos)n) {
+                    it = map_pos_to_image.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+        }
+        tokens.resize(n);
+    }
+
+    std::string detokenize(const llama_context * ctx, bool special) const {
+        llama_tokens text_tokens;
+        text_tokens.reserve(tokens.size());
+        for (const auto & t : tokens) {
+            if (t != LLAMA_TOKEN_NULL) {
+                text_tokens.push_back(t);
+            }
+        }
+        return common_detokenize(ctx, text_tokens, special);
+    }
+
+    size_t get_common_prefix(const server_tokens & b) const {
+        size_t max_idx = std::min(tokens.size(), b.tokens.size());
+        for (size_t i = 0; i < max_idx; ++i) {
+            auto & ai =   tokens[i];
+            auto & bi = b.tokens[i];
+
+            if (ai == LLAMA_TOKEN_NULL && bi == LLAMA_TOKEN_NULL) {
+                GGML_ASSERT(has_mtmd);
+                const auto & a_chunk =   find_chunk(i);
+                const auto & b_chunk = b.find_chunk(i);
+                GGML_ASSERT(a_chunk && b_chunk);
+                const auto * a_img = mtmd_input_chunk_get_tokens_image(a_chunk.get());
+                const auto * b_img = mtmd_input_chunk_get_tokens_image(b_chunk.get());
+                std::string ai_id  = mtmd_image_tokens_get_id(a_img);
+                std::string bi_id  = mtmd_image_tokens_get_id(b_img);
+                size_t a_pos       = mtmd_image_tokens_get_n_pos(a_img);
+                size_t b_pos       = mtmd_image_tokens_get_n_pos(b_img);
+                if (ai_id == bi_id && a_pos == b_pos) {
+                    GGML_ASSERT(a_pos > 0 && "Invalid image token"); // should never happen
+                    i += a_pos - 1; // will be +1 by the for loop
+                    continue;
+                } else {
+                    return i;
+                }
+            } else if (ai == bi) {
+                continue;
+            } else {
+                return i;
+            }
+        }
+        return max_idx; // all tokens are equal
+    }
+
+    // make sure all text tokens are within the vocab range
+    bool validate(const struct llama_context * ctx) const {
+        const llama_model * model = llama_get_model(ctx);
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const int32_t n_vocab = llama_vocab_n_tokens(vocab);
+
+        for (size_t i = 0; i < tokens.size(); ++i) {
+            auto & t = tokens[i];
+            if (t == LLAMA_TOKEN_NULL) {
+                try {
+                    const auto & chunk = find_chunk(i);
+                    const auto * img_tokens = mtmd_input_chunk_get_tokens_image(chunk.get());
+                    size_t n_pos = mtmd_image_tokens_get_n_pos(img_tokens);
+                    i += n_pos - 1; // will be +1 by the for loop
+                } catch (const std::exception & e) {
+                    return false;
+                }
+            } else if (t < 0 || t >= n_vocab) {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    // encode and decode the image chunk
+    int32_t process_chunk(
+                llama_context * ctx,
+                mtmd_context * mctx,
+                llama_pos n_past,
+                int32_t seq_id,
+                llama_pos & n_pos_out) {
+        auto it = map_pos_to_image.find(n_past);
+        if (it == map_pos_to_image.end()) {
+            throw std::runtime_error("Chunk not found");
+        }
+     //   SRV_INF("%s\n", "processing image...");
+        int32_t n_batch = llama_n_batch(ctx);
+        int64_t t0 = ggml_time_ms();
+        llama_pos new_n_past = n_past;
+        int32_t result = mtmd_helper_eval_chunk_single(mctx, ctx,
+            it->second.get(), // chunk
+            n_past,
+            seq_id,
+            n_batch,
+            true, // logits last
+            &new_n_past);
+        //SRV_INF("image processed in %" PRId64 " ms\n", ggml_time_ms() - t0);
+        if (result != 0) {
+            LOG_ERR("mtmd_helper_eval failed with status %d", result);
+            n_pos_out = n_past;
+            return result;
+        }
+        n_pos_out = new_n_past;
+        return 0;
+    }
+};
+
+// Computes FNV-1a hash of the data
+static std::string fnv_hash(const uint8_t * data, size_t len) {
+    const uint64_t fnv_prime = 0x100000001b3ULL;
+    uint64_t hash = 0xcbf29ce484222325ULL;
+
+    for (size_t i = 0; i < len; ++i) {
+        hash ^= data[i];
+        hash *= fnv_prime;
+    }
+    return std::to_string(hash);
 }

backend/go/image/stablediffusion-ggml/Makefile

@@ -8,12 +8,19 @@ 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?=
+
 # Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
 
 # If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
-	CMAKE_ARGS+=-DGGML_CUDA=ON
+	CMAKE_ARGS+=-DSD_CUDA=ON
 # If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
 # to CMAKE_ARGS automatically
 else ifeq ($(BUILD_TYPE),openblas)
@@ -21,31 +28,50 @@ 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 ($(BUILD_TYPE),hipblas)
-	CMAKE_ARGS+=-DGGML_HIP=ON
+	CMAKE_ARGS+=-DSD_HIPBLAS=ON
 # If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
 # But if it's OSX without metal, disable it here
 else ifeq ($(OS),Darwin)
 	ifneq ($(BUILD_TYPE),metal)
-		CMAKE_ARGS+=-DGGML_METAL=OFF
+		CMAKE_ARGS+=-DSD_METAL=OFF
 	else
-		CMAKE_ARGS+=-DGGML_METAL=ON
+		CMAKE_ARGS+=-DSD_METAL=ON
 		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
 		TARGET+=--target ggml-metal
 	endif
 endif
 
-# ifeq ($(BUILD_TYPE),sycl_f16)
+ifeq ($(BUILD_TYPE),sycl_f16)
-# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DSD_SYCL=ON -DGGML_SYCL_F16=ON
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
-# endif
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DSD_SYCL=ON \
+		-DGGML_SYCL_F16=ON
+	CC=icx
+	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 ($(BUILD_TYPE),sycl_f32)
+ifeq ($(BUILD_TYPE),sycl_f32)
-# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON
+	CMAKE_ARGS+=-DGGML_SYCL=ON \
-# endif
+		-DCMAKE_C_COMPILER=icx \
+		-DCMAKE_CXX_COMPILER=icpx \
+		-DSD_SYCL=ON
+	CC=icx
+	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
+# 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)
@@ -86,11 +112,24 @@ 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
 
 libsd.a: gosd.o
 	cp $(INCLUDE_PATH)/build/libstable-diffusion.a ./libsd.a
 	$(AR) rcs libsd.a gosd.o
 
+stablediffusion-ggml:
+	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 "$(GO_TAGS)" -o ../../../../backend-assets/grpc/stablediffusion-ggml ./
+ifneq ($(UPX),)
+	$(UPX) ../../../../backend-assets/grpc/stablediffusion-ggml
+endif
+
 clean:
-	rm -rf gosd.o libsd.a build $(COMBINED_LIB)
+	rm -rf gosd.o libsd.a build $(COMBINED_LIB)

backend/go/transcribe/whisper/whisper.go

@@ -74,7 +74,7 @@ func (sd *Whisper) AudioTranscription(opts *pb.TranscriptRequest) (pb.Transcript
 		context.SetTranslate(true)
 	}
 
-	if err := context.Process(data, nil, nil); err != nil {
+	if err := context.Process(data, nil, nil, nil); err != nil {
 		return pb.TranscriptResult{}, err
 	}
 

backend/python/autogptq/Makefile

@@ -1,17 +0,0 @@
-.PHONY: autogptq
-autogptq: protogen
-	bash install.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../.. --python_out=. --grpc_python_out=. backend.proto
-
-.PHONY: clean
-clean: protogen-clean
-	rm -rf venv __pycache__

backend/python/autogptq/README.md

@@ -1,5 +0,0 @@
-# Creating a separate environment for the autogptq project
-
-```
-make autogptq
-```

backend/python/autogptq/backend.py

@@ -1,153 +0,0 @@
-#!/usr/bin/env python3
-from concurrent import futures
-import argparse
-import signal
-import sys
-import os
-import time
-import base64
-
-import grpc
-import backend_pb2
-import backend_pb2_grpc
-
-from auto_gptq import AutoGPTQForCausalLM
-from transformers import AutoTokenizer, AutoModelForCausalLM
-from transformers import TextGenerationPipeline
-
-_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):
-    def Health(self, request, context):
-        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
-    def LoadModel(self, request, context):
-        try:
-            device = "cuda:0"
-            if request.Device != "":
-                device = request.Device
-
-            # support loading local model files
-            model_path = os.path.join(os.environ.get('MODELS_PATH', './'), request.Model)
-            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True, trust_remote_code=request.TrustRemoteCode)
-
-            # support model `Qwen/Qwen-VL-Chat-Int4`
-            if "qwen-vl" in request.Model.lower():
-                self.model_name = "Qwen-VL-Chat"
-                model = AutoModelForCausalLM.from_pretrained(model_path, 
-                    trust_remote_code=request.TrustRemoteCode,
-                    device_map="auto").eval()
-            else:
-                model = AutoGPTQForCausalLM.from_quantized(model_path,
-                    model_basename=request.ModelBaseName,
-                    use_safetensors=True,
-                    trust_remote_code=request.TrustRemoteCode,
-                    device=device,
-                    use_triton=request.UseTriton,
-                    quantize_config=None)
-            
-            self.model = model
-            self.tokenizer = tokenizer
-        except Exception as err:
-            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
-        return backend_pb2.Result(message="Model loaded successfully", success=True)
-
-    def Predict(self, request, context):
-        penalty = 1.0
-        if request.Penalty != 0.0:
-            penalty = request.Penalty
-        tokens = 512
-        if request.Tokens != 0:
-            tokens = request.Tokens
-        top_p = 0.95
-        if request.TopP != 0.0:
-            top_p = request.TopP
-
-        
-        prompt_images = self.recompile_vl_prompt(request)
-        compiled_prompt = prompt_images[0]
-        print(f"Prompt: {compiled_prompt}", file=sys.stderr)
-
-        # Implement Predict RPC
-        pipeline = TextGenerationPipeline(
-            model=self.model, 
-            tokenizer=self.tokenizer,
-            max_new_tokens=tokens,
-            temperature=request.Temperature,
-            top_p=top_p,
-            repetition_penalty=penalty,
-            )
-        t = pipeline(compiled_prompt)[0]["generated_text"]
-        print(f"generated_text: {t}", file=sys.stderr)
-        
-        if compiled_prompt in t:
-            t = t.replace(compiled_prompt, "")
-        # house keeping. Remove the image files from /tmp folder
-        for img_path in prompt_images[1]:
-            try:
-                os.remove(img_path)
-            except Exception as e:
-                print(f"Error removing image file: {img_path}, {e}", file=sys.stderr)
-
-        return backend_pb2.Result(message=bytes(t, encoding='utf-8'))
-
-    def PredictStream(self, request, context):
-        # Implement PredictStream RPC
-        #for reply in some_data_generator():
-        #    yield reply
-        # Not implemented yet
-        return self.Predict(request, context)
-
-    def recompile_vl_prompt(self, request):
-        prompt = request.Prompt
-        image_paths = []
-
-        if "qwen-vl" in self.model_name.lower():
-            # request.Images is an array which contains base64 encoded images. Iterate the request.Images array, decode and save each image to /tmp folder with a random filename.
-            # Then, save the image file paths to an array "image_paths".
-            # read "request.Prompt", replace "[img-%d]" with the image file paths in the order they appear in "image_paths". Save the new prompt to "prompt".
-            for i, img in enumerate(request.Images):
-                timestamp = str(int(time.time() * 1000))  # Generate timestamp
-                img_path = f"/tmp/vl-{timestamp}.jpg"  # Use timestamp in filename
-                with open(img_path, "wb") as f:
-                    f.write(base64.b64decode(img))
-                image_paths.append(img_path)
-                prompt = prompt.replace(f"[img-{i}]", "<img>" + img_path + "</img>,")
-        else:
-            prompt = request.Prompt
-        return (prompt, image_paths)
-
-def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
-    backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
-    server.add_insecure_port(address)
-    server.start()
-    print("Server started. Listening on: " + address, file=sys.stderr)
-
-    # Define the signal handler function
-    def signal_handler(sig, frame):
-        print("Received termination signal. Shutting down...")
-        server.stop(0)
-        sys.exit(0)
-
-    # Set the signal handlers for SIGINT and SIGTERM
-    signal.signal(signal.SIGINT, signal_handler)
-    signal.signal(signal.SIGTERM, signal_handler)
-
-    try:
-        while True:
-            time.sleep(_ONE_DAY_IN_SECONDS)
-    except KeyboardInterrupt:
-        server.stop(0)
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="Run the gRPC server.")
-    parser.add_argument(
-        "--addr", default="localhost:50051", help="The address to bind the server to."
-    )
-    args = parser.parse_args()
-
-    serve(args.addr)

backend/python/autogptq/install.sh

@@ -1,14 +0,0 @@
-#!/bin/bash
-set -e
-
-source $(dirname $0)/../common/libbackend.sh
-
-# This is here because the Intel pip index is broken and returns 200 status codes for every package name, it just doesn't return any package links.
-# This makes uv think that the package exists in the Intel pip index, and by default it stops looking at other pip indexes once it finds a match.
-# We need uv to continue falling through to the pypi default index to find optimum[openvino] in the pypi index
-# the --upgrade actually allows us to *downgrade* torch to the version provided in the Intel pip index
-if [ "x${BUILD_PROFILE}" == "xintel" ]; then
-    EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
-fi
-
-installRequirements

backend/python/autogptq/requirements-cublas11.txt

@@ -1,2 +0,0 @@
---extra-index-url https://download.pytorch.org/whl/cu118
-torch==2.4.1+cu118

backend/python/autogptq/requirements-cublas12.txt

@@ -1 +0,0 @@
-torch==2.4.1

backend/python/autogptq/requirements-hipblas.txt

@@ -1,2 +0,0 @@
---extra-index-url https://download.pytorch.org/whl/rocm6.0
-torch==2.4.1+rocm6.0

backend/python/autogptq/requirements-intel.txt

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

backend/python/autogptq/requirements.txt

@@ -1,6 +0,0 @@
-accelerate
-auto-gptq==0.7.1
-grpcio==1.71.0
-protobuf
-certifi
-transformers

backend/python/autogptq/run.sh

@@ -1,4 +0,0 @@
-#!/bin/bash
-source $(dirname $0)/../common/libbackend.sh
-
-startBackend $@

backend/python/autogptq/test.sh

@@ -1,6 +0,0 @@
-#!/bin/bash
-set -e
-
-source $(dirname $0)/../common/libbackend.sh
-
-runUnittests

backend/python/bark/backend.py

@@ -61,7 +61,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 grpcio-tools

backend/python/coqui/backend.py

@@ -86,7 +86,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/backend.py

@@ -19,7 +19,7 @@ import grpc
 
 from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
     EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
-from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline
+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
 from compel import Compel, ReturnedEmbeddingsType
@@ -168,9 +168,13 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             # We are storing all the options in a dict so we can use it later when
             # generating the images
             for opt in options:
+                if ":" not in opt:
+                    continue
                 key, value = opt.split(":")
                 self.options[key] = value
 
+            print(f"Options: {self.options}", file=sys.stderr)
+
             local = False
             modelFile = request.Model
 
@@ -287,6 +291,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
                     if request.LowVRAM:
                         self.pipe.enable_model_cpu_offload()
+            elif request.PipelineType == "Lumina2Text2ImgPipeline":
+                self.pipe = Lumina2Text2ImgPipeline.from_pretrained(
+                    request.Model,
+                    torch_dtype=torch.bfloat16)
+                if request.LowVRAM:
+                    self.pipe.enable_model_cpu_offload()
             elif request.PipelineType == "SanaPipeline":
                 self.pipe = SanaPipeline.from_pretrained(
                     request.Model,
@@ -516,7 +526,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.71.0
+grpcio==1.72.0
 pillow
 protobuf
 certifi

backend/python/exllama2/backend.py

@@ -105,7 +105,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi
 wheel

backend/python/faster-whisper/backend.py

@@ -62,7 +62,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.TranscriptResult(segments=resultSegments, text=text)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/faster-whisper/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 grpcio-tools

backend/python/kokoro/backend.py

@@ -99,7 +99,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/kokoro/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 phonemizer
 scipy

backend/python/rerankers/backend.py

@@ -91,7 +91,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.RerankResult(usage=usage, results=results)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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()

backend/python/rerankers/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi

backend/python/transformers/backend.py

@@ -559,7 +559,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 async def serve(address):
     # Start asyncio gRPC server
-    server = grpc.aio.server(migration_thread_pool=futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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

backend/python/transformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi
 setuptools

backend/python/vllm/backend.py

@@ -194,27 +194,40 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             await iterations.aclose()
 
     async def _predict(self, request, context, streaming=False):
+        # Build the sampling parameters
+        # NOTE: this must stay in sync with the vllm backend
+        request_to_sampling_params = {
+            "N": "n",
+            "PresencePenalty": "presence_penalty",
+            "FrequencyPenalty": "frequency_penalty",
+            "RepetitionPenalty": "repetition_penalty",
+            "Temperature": "temperature",
+            "TopP": "top_p",
+            "TopK": "top_k",
+            "MinP": "min_p",
+            "Seed": "seed",
+            "StopPrompts": "stop",
+            "StopTokenIds": "stop_token_ids",
+            "BadWords": "bad_words",
+            "IncludeStopStrInOutput": "include_stop_str_in_output",
+            "IgnoreEOS": "ignore_eos",
+            "Tokens": "max_tokens",
+            "MinTokens": "min_tokens",
+            "Logprobs": "logprobs",
+            "PromptLogprobs": "prompt_logprobs",
+            "SkipSpecialTokens": "skip_special_tokens",
+            "SpacesBetweenSpecialTokens": "spaces_between_special_tokens",
+            "TruncatePromptTokens": "truncate_prompt_tokens",
+            "GuidedDecoding": "guided_decoding",
+        }
 
-        # Build sampling parameters
         sampling_params = SamplingParams(top_p=0.9, max_tokens=200)
-        if request.TopP != 0:
+
-            sampling_params.top_p = request.TopP
+        for request_field, param_field in request_to_sampling_params.items():
-        if request.Tokens > 0:
+            if hasattr(request, request_field):
-            sampling_params.max_tokens = request.Tokens
+                value = getattr(request, request_field)
-        if request.Temperature != 0:
+                if value not in (None, 0, [], False):
-            sampling_params.temperature = request.Temperature
+                    setattr(sampling_params, param_field, value)
-        if request.TopK != 0:
-            sampling_params.top_k = request.TopK
-        if request.PresencePenalty != 0:
-            sampling_params.presence_penalty = request.PresencePenalty
-        if request.FrequencyPenalty != 0:
-            sampling_params.frequency_penalty = request.FrequencyPenalty
-        if request.StopPrompts:
-            sampling_params.stop = request.StopPrompts
-        if request.IgnoreEOS:
-            sampling_params.ignore_eos = request.IgnoreEOS
-        if request.Seed != 0:
-            sampling_params.seed = request.Seed
 
         # Extract image paths and process images
         prompt = request.Prompt
@@ -320,7 +333,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 async def serve(address):
     # Start asyncio gRPC server
-    server = grpc.aio.server(migration_thread_pool=futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    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

backend/python/vllm/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.72.0
 protobuf
 certifi
 setuptools

backend/python/vllm/test.py

@@ -75,6 +75,53 @@ class TestBackendServicer(unittest.TestCase):
         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

core/application/application.go

@@ -16,7 +16,7 @@ type Application struct {
 func newApplication(appConfig *config.ApplicationConfig) *Application {
 	return &Application{
 		backendLoader:      config.NewBackendConfigLoader(appConfig.ModelPath),
-		modelLoader:        model.NewModelLoader(appConfig.ModelPath),
+		modelLoader:        model.NewModelLoader(appConfig.ModelPath, appConfig.SingleBackend),
 		applicationConfig:  appConfig,
 		templatesEvaluator: templates.NewEvaluator(appConfig.ModelPath),
 	}

core/application/startup.go

@@ -43,18 +43,12 @@ func New(opts ...config.AppOption) (*Application, error) {
 	if err != nil {
 		return nil, fmt.Errorf("unable to create ModelPath: %q", err)
 	}
-	if options.ImageDir != "" {
+	if options.GeneratedContentDir != "" {
-		err := os.MkdirAll(options.ImageDir, 0750)
+		err := os.MkdirAll(options.GeneratedContentDir, 0750)
 		if err != nil {
 			return nil, fmt.Errorf("unable to create ImageDir: %q", err)
 		}
 	}
-	if options.AudioDir != "" {
-		err := os.MkdirAll(options.AudioDir, 0750)
-		if err != nil {
-			return nil, fmt.Errorf("unable to create AudioDir: %q", err)
-		}
-	}
 	if options.UploadDir != "" {
 		err := os.MkdirAll(options.UploadDir, 0750)
 		if err != nil {
@@ -143,7 +137,7 @@ func New(opts ...config.AppOption) (*Application, error) {
 		}()
 	}
 
-	if options.LoadToMemory != nil {
+	if options.LoadToMemory != nil && !options.SingleBackend {
 		for _, m := range options.LoadToMemory {
 			cfg, err := application.BackendLoader().LoadBackendConfigFileByNameDefaultOptions(m, options)
 			if err != nil {

core/backend/embeddings.go

@@ -17,6 +17,7 @@ func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendCo
 	if err != nil {
 		return nil, err
 	}
+	defer loader.Close()
 
 	var fn func() ([]float32, error)
 	switch model := inferenceModel.(type) {

core/backend/image.go

@@ -16,6 +16,7 @@ func ImageGeneration(height, width, mode, step, seed int, positive_prompt, negat
 	if err != nil {
 		return nil, err
 	}
+	defer loader.Close()
 
 	fn := func() error {
 		_, err := inferenceModel.GenerateImage(

core/backend/llm.go

@@ -53,6 +53,7 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 	if err != nil {
 		return nil, err
 	}
+	defer loader.Close()
 
 	var protoMessages []*proto.Message
 	// if we are using the tokenizer template, we need to convert the messages to proto messages

core/backend/options.go

@@ -40,10 +40,6 @@ func ModelOptions(c config.BackendConfig, so *config.ApplicationConfig, opts ...
 	grpcOpts := grpcModelOpts(c)
 	defOpts = append(defOpts, model.WithLoadGRPCLoadModelOpts(grpcOpts))
 
-	if so.SingleBackend {
-		defOpts = append(defOpts, model.WithSingleActiveBackend())
-	}
-
 	if so.ParallelBackendRequests {
 		defOpts = append(defOpts, model.EnableParallelRequests)
 	}
@@ -103,7 +99,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		mmap = *c.MMap
 	}
 
-	ctxSize := 1024
+	ctxSize := 4096
 	if c.ContextSize != nil {
 		ctxSize = *c.ContextSize
 	}
@@ -121,7 +117,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 	triggers := make([]*pb.GrammarTrigger, 0)
 	for _, t := range c.FunctionsConfig.GrammarConfig.GrammarTriggers {
 		triggers = append(triggers, &pb.GrammarTrigger{
-			Word:    t.Word,
+			Word: t.Word,
 		})
 
 	}
@@ -161,38 +157,33 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		DisableLogStatus:     c.DisableLogStatus,
 		DType:                c.DType,
 		// LimitMMPerPrompt vLLM
-		LimitImagePerPrompt:  int32(c.LimitMMPerPrompt.LimitImagePerPrompt),
+		LimitImagePerPrompt: int32(c.LimitMMPerPrompt.LimitImagePerPrompt),
-		LimitVideoPerPrompt:  int32(c.LimitMMPerPrompt.LimitVideoPerPrompt),
+		LimitVideoPerPrompt: int32(c.LimitMMPerPrompt.LimitVideoPerPrompt),
-		LimitAudioPerPrompt:  int32(c.LimitMMPerPrompt.LimitAudioPerPrompt),
+		LimitAudioPerPrompt: int32(c.LimitMMPerPrompt.LimitAudioPerPrompt),
-		MMProj:               c.MMProj,
+		MMProj:              c.MMProj,
-		FlashAttention:       c.FlashAttention,
+		FlashAttention:      c.FlashAttention,
-		CacheTypeKey:         c.CacheTypeK,
+		CacheTypeKey:        c.CacheTypeK,
-		CacheTypeValue:       c.CacheTypeV,
+		CacheTypeValue:      c.CacheTypeV,
-		NoKVOffload:          c.NoKVOffloading,
+		NoKVOffload:         c.NoKVOffloading,
-		YarnExtFactor:        c.YarnExtFactor,
+		YarnExtFactor:       c.YarnExtFactor,
-		YarnAttnFactor:       c.YarnAttnFactor,
+		YarnAttnFactor:      c.YarnAttnFactor,
-		YarnBetaFast:         c.YarnBetaFast,
+		YarnBetaFast:        c.YarnBetaFast,
-		YarnBetaSlow:         c.YarnBetaSlow,
+		YarnBetaSlow:        c.YarnBetaSlow,
-		NGQA:                 c.NGQA,
+		NGQA:                c.NGQA,
-		RMSNormEps:           c.RMSNormEps,
+		RMSNormEps:          c.RMSNormEps,
-		MLock:                mmlock,
+		MLock:               mmlock,
-		RopeFreqBase:         c.RopeFreqBase,
+		RopeFreqBase:        c.RopeFreqBase,
-		RopeScaling:          c.RopeScaling,
+		RopeScaling:         c.RopeScaling,
-		Type:                 c.ModelType,
+		Type:                c.ModelType,
-		RopeFreqScale:        c.RopeFreqScale,
+		RopeFreqScale:       c.RopeFreqScale,
-		NUMA:                 c.NUMA,
+		NUMA:                c.NUMA,
-		Embeddings:           embeddings,
+		Embeddings:          embeddings,
-		LowVRAM:              lowVRAM,
+		LowVRAM:             lowVRAM,
-		NGPULayers:           int32(nGPULayers),
+		NGPULayers:          int32(nGPULayers),
-		MMap:                 mmap,
+		MMap:                mmap,
-		MainGPU:              c.MainGPU,
+		MainGPU:             c.MainGPU,
-		Threads:              int32(*c.Threads),
+		Threads:             int32(*c.Threads),
-		TensorSplit:          c.TensorSplit,
+		TensorSplit:         c.TensorSplit,
-		// AutoGPTQ
-		ModelBaseName:    c.AutoGPTQ.ModelBaseName,
-		Device:           c.AutoGPTQ.Device,
-		UseTriton:        c.AutoGPTQ.Triton,
-		UseFastTokenizer: c.AutoGPTQ.UseFastTokenizer,
 		// RWKV
 		Tokenizer: c.Tokenizer,
 	}

core/backend/rerank.go

@@ -12,10 +12,10 @@ import (
 func Rerank(request *proto.RerankRequest, loader *model.ModelLoader, appConfig *config.ApplicationConfig, backendConfig config.BackendConfig) (*proto.RerankResult, error) {
 	opts := ModelOptions(backendConfig, appConfig)
 	rerankModel, err := loader.Load(opts...)
-
 	if err != nil {
 		return nil, err
 	}
+	defer loader.Close()
 
 	if rerankModel == nil {
 		return nil, fmt.Errorf("could not load rerank model")

core/backend/soundgeneration.go

@@ -26,21 +26,26 @@ func SoundGeneration(
 
 	opts := ModelOptions(backendConfig, appConfig)
 	soundGenModel, err := loader.Load(opts...)
-
 	if err != nil {
 		return "", nil, err
 	}
+	defer loader.Close()
 
 	if soundGenModel == nil {
 		return "", nil, fmt.Errorf("could not load sound generation model")
 	}
 
-	if err := os.MkdirAll(appConfig.AudioDir, 0750); err != nil {
+	if err := os.MkdirAll(appConfig.GeneratedContentDir, 0750); err != nil {
 		return "", nil, fmt.Errorf("failed creating audio directory: %s", err)
 	}
 
-	fileName := utils.GenerateUniqueFileName(appConfig.AudioDir, "sound_generation", ".wav")
+	audioDir := filepath.Join(appConfig.GeneratedContentDir, "audio")
-	filePath := filepath.Join(appConfig.AudioDir, fileName)
+	if err := os.MkdirAll(audioDir, 0750); err != nil {
+		return "", nil, fmt.Errorf("failed creating audio directory: %s", err)
+	}
+
+	fileName := utils.GenerateUniqueFileName(audioDir, "sound_generation", ".wav")
+	filePath := filepath.Join(audioDir, fileName)
 
 	res, err := soundGenModel.SoundGeneration(context.Background(), &proto.SoundGenerationRequest{
 		Text:        text,

core/backend/token_metrics.go

@@ -20,6 +20,7 @@ func TokenMetrics(
 	if err != nil {
 		return nil, err
 	}
+	defer loader.Close()
 
 	if model == nil {
 		return nil, fmt.Errorf("could not loadmodel model")

core/backend/tokenize.go

@@ -14,10 +14,10 @@ func ModelTokenize(s string, loader *model.ModelLoader, backendConfig config.Bac
 
 	opts := ModelOptions(backendConfig, appConfig)
 	inferenceModel, err = loader.Load(opts...)
-
 	if err != nil {
 		return schema.TokenizeResponse{}, err
 	}
+	defer loader.Close()
 
 	predictOptions := gRPCPredictOpts(backendConfig, loader.ModelPath)
 	predictOptions.Prompt = s

core/backend/transcript.go

@@ -24,6 +24,7 @@ func ModelTranscription(audio, language string, translate bool, ml *model.ModelL
 	if err != nil {
 		return nil, err
 	}
+	defer ml.Close()
 
 	if transcriptionModel == nil {
 		return nil, fmt.Errorf("could not load transcription model")

core/backend/tts.go

@@ -23,21 +23,22 @@ func ModelTTS(
 ) (string, *proto.Result, error) {
 	opts := ModelOptions(backendConfig, appConfig, model.WithDefaultBackendString(model.PiperBackend))
 	ttsModel, err := loader.Load(opts...)
-
 	if err != nil {
 		return "", nil, err
 	}
+	defer loader.Close()
 
 	if ttsModel == nil {
 		return "", nil, fmt.Errorf("could not load tts model %q", backendConfig.Model)
 	}
 
-	if err := os.MkdirAll(appConfig.AudioDir, 0750); err != nil {
+	audioDir := filepath.Join(appConfig.GeneratedContentDir, "audio")
+	if err := os.MkdirAll(audioDir, 0750); err != nil {
 		return "", nil, fmt.Errorf("failed creating audio directory: %s", err)
 	}
 
-	fileName := utils.GenerateUniqueFileName(appConfig.AudioDir, "tts", ".wav")
+	fileName := utils.GenerateUniqueFileName(audioDir, "tts", ".wav")
-	filePath := filepath.Join(appConfig.AudioDir, fileName)
+	filePath := filepath.Join(audioDir, fileName)
 
 	// We join the model name to the model path here. This seems to only be done for TTS and is HIGHLY suspect.
 	// This should be addressed in a follow up PR soon.

core/backend/vad.go

@@ -19,6 +19,8 @@ func VAD(request *schema.VADRequest,
 	if err != nil {
 		return nil, err
 	}
+	defer ml.Close()
+
 	req := proto.VADRequest{
 		Audio: request.Audio,
 	}

core/backend/video.go

@@ -0,0 +1,36 @@
+package backend
+
+import (
+	"github.com/mudler/LocalAI/core/config"
+
+	"github.com/mudler/LocalAI/pkg/grpc/proto"
+	model "github.com/mudler/LocalAI/pkg/model"
+)
+
+func VideoGeneration(height, width int32, prompt, startImage, endImage, dst string, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (func() error, error) {
+
+	opts := ModelOptions(backendConfig, appConfig)
+	inferenceModel, err := loader.Load(
+		opts...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	defer loader.Close()
+
+	fn := func() error {
+		_, err := inferenceModel.GenerateVideo(
+			appConfig.Context,
+			&proto.GenerateVideoRequest{
+				Height:     height,
+				Width:      width,
+				Prompt:     prompt,
+				StartImage: startImage,
+				EndImage:   endImage,
+				Dst:        dst,
+			})
+		return err
+	}
+
+	return fn, nil
+}

core/cli/context/context.go

@@ -1,11 +1,13 @@
 package cliContext
 
-import "embed"
+import (
+	rice "github.com/GeertJohan/go.rice"
+)
 
 type Context struct {
 	Debug    bool    `env:"LOCALAI_DEBUG,DEBUG" default:"false" hidden:"" help:"DEPRECATED, use --log-level=debug instead. Enable debug logging"`
 	LogLevel *string `env:"LOCALAI_LOG_LEVEL" enum:"error,warn,info,debug,trace" help:"Set the level of logs to output [${enum}]"`
 
 	// This field is not a command line argument/flag, the struct tag excludes it from the parsed CLI
-	BackendAssets embed.FS `kong:"-"`
+	BackendAssets *rice.Box `kong:"-"`
 }

core/cli/run.go

@@ -21,8 +21,7 @@ type RunCMD struct {
 
 	ModelsPath                   string        `env:"LOCALAI_MODELS_PATH,MODELS_PATH" type:"path" default:"${basepath}/models" help:"Path containing models used for inferencing" group:"storage"`
 	BackendAssetsPath            string        `env:"LOCALAI_BACKEND_ASSETS_PATH,BACKEND_ASSETS_PATH" type:"path" default:"/tmp/localai/backend_data" help:"Path used to extract libraries that are required by some of the backends in runtime" group:"storage"`
-	ImagePath                    string        `env:"LOCALAI_IMAGE_PATH,IMAGE_PATH" type:"path" default:"/tmp/generated/images" help:"Location for images generated by backends (e.g. stablediffusion)" group:"storage"`
+	GeneratedContentPath         string        `env:"LOCALAI_GENERATED_CONTENT_PATH,GENERATED_CONTENT_PATH" type:"path" default:"/tmp/generated/content" help:"Location for generated content (e.g. images, audio, videos)" group:"storage"`
-	AudioPath                    string        `env:"LOCALAI_AUDIO_PATH,AUDIO_PATH" type:"path" default:"/tmp/generated/audio" help:"Location for audio generated by backends (e.g. piper)" group:"storage"`
 	UploadPath                   string        `env:"LOCALAI_UPLOAD_PATH,UPLOAD_PATH" type:"path" default:"/tmp/localai/upload" help:"Path to store uploads from files api" group:"storage"`
 	ConfigPath                   string        `env:"LOCALAI_CONFIG_PATH,CONFIG_PATH" default:"/tmp/localai/config" group:"storage"`
 	LocalaiConfigDir             string        `env:"LOCALAI_CONFIG_DIR" type:"path" default:"${basepath}/configuration" help:"Directory for dynamic loading of certain configuration files (currently api_keys.json and external_backends.json)" group:"storage"`
@@ -47,7 +46,7 @@ type RunCMD struct {
 	CSRF                               bool     `env:"LOCALAI_CSRF" help:"Enables fiber CSRF middleware" group:"api"`
 	UploadLimit                        int      `env:"LOCALAI_UPLOAD_LIMIT,UPLOAD_LIMIT" default:"15" help:"Default upload-limit in MB" group:"api"`
 	APIKeys                            []string `env:"LOCALAI_API_KEY,API_KEY" help:"List of API Keys to enable API authentication. When this is set, all the requests must be authenticated with one of these API keys" group:"api"`
-	DisableWebUI                       bool     `env:"LOCALAI_DISABLE_WEBUI,DISABLE_WEBUI" default:"false" help:"Disable webui" group:"api"`
+	DisableWebUI                       bool     `env:"LOCALAI_DISABLE_WEBUI,DISABLE_WEBUI" default:"false" help:"Disables the web user interface. When set to true, the server will only expose API endpoints without serving the web interface" group:"api"`
 	DisablePredownloadScan             bool     `env:"LOCALAI_DISABLE_PREDOWNLOAD_SCAN" help:"If true, disables the best-effort security scanner before downloading any files." group:"hardening" default:"false"`
 	OpaqueErrors                       bool     `env:"LOCALAI_OPAQUE_ERRORS" default:"false" help:"If true, all error responses are replaced with blank 500 errors. This is intended only for hardening against information leaks and is normally not recommended." group:"hardening"`
 	UseSubtleKeyComparison             bool     `env:"LOCALAI_SUBTLE_KEY_COMPARISON" default:"false" help:"If true, API Key validation comparisons will be performed using constant-time comparisons rather than simple equality. This trades off performance on each request for resiliancy against timing attacks." group:"hardening"`
@@ -81,8 +80,7 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 		config.WithModelPath(r.ModelsPath),
 		config.WithContextSize(r.ContextSize),
 		config.WithDebug(zerolog.GlobalLevel() <= zerolog.DebugLevel),
-		config.WithImageDir(r.ImagePath),
+		config.WithGeneratedContentDir(r.GeneratedContentPath),
-		config.WithAudioDir(r.AudioPath),
 		config.WithUploadDir(r.UploadPath),
 		config.WithConfigsDir(r.ConfigPath),
 		config.WithDynamicConfigDir(r.LocalaiConfigDir),

core/cli/soundgeneration.go

@@ -70,11 +70,11 @@ func (t *SoundGenerationCMD) Run(ctx *cliContext.Context) error {
 	opts := &config.ApplicationConfig{
 		ModelPath:            t.ModelsPath,
 		Context:              context.Background(),
-		AudioDir:             outputDir,
+		GeneratedContentDir:  outputDir,
 		AssetsDestination:    t.BackendAssetsPath,
 		ExternalGRPCBackends: externalBackends,
 	}
-	ml := model.NewModelLoader(opts.ModelPath)
+	ml := model.NewModelLoader(opts.ModelPath, opts.SingleBackend)
 
 	defer func() {
 		err := ml.StopAllGRPC()

core/cli/transcript.go

@@ -32,7 +32,7 @@ func (t *TranscriptCMD) Run(ctx *cliContext.Context) error {
 	}
 
 	cl := config.NewBackendConfigLoader(t.ModelsPath)
-	ml := model.NewModelLoader(opts.ModelPath)
+	ml := model.NewModelLoader(opts.ModelPath, opts.SingleBackend)
 	if err := cl.LoadBackendConfigsFromPath(t.ModelsPath); err != nil {
 		return err
 	}

core/cli/tts.go

@@ -36,12 +36,12 @@ func (t *TTSCMD) Run(ctx *cliContext.Context) error {
 	text := strings.Join(t.Text, " ")
 
 	opts := &config.ApplicationConfig{
-		ModelPath:         t.ModelsPath,
+		ModelPath:           t.ModelsPath,
-		Context:           context.Background(),
+		Context:             context.Background(),
-		AudioDir:          outputDir,
+		GeneratedContentDir: outputDir,
-		AssetsDestination: t.BackendAssetsPath,
+		AssetsDestination:   t.BackendAssetsPath,
 	}
-	ml := model.NewModelLoader(opts.ModelPath)
+	ml := model.NewModelLoader(opts.ModelPath, opts.SingleBackend)
 
 	defer func() {
 		err := ml.StopAllGRPC()

core/cli/util.go

@@ -7,11 +7,11 @@ import (
 
 	"github.com/rs/zerolog/log"
 
+	gguf "github.com/gpustack/gguf-parser-go"
 	cliContext "github.com/mudler/LocalAI/core/cli/context"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery"
 	"github.com/mudler/LocalAI/pkg/downloader"
-	gguf "github.com/thxcode/gguf-parser-go"
 )
 
 type UtilCMD struct {
@@ -51,7 +51,7 @@ func (u *GGUFInfoCMD) Run(ctx *cliContext.Context) error {
 	log.Info().
 		Any("eosTokenID", f.Tokenizer().EOSTokenID).
 		Any("bosTokenID", f.Tokenizer().BOSTokenID).
-		Any("modelName", f.Model().Name).
+		Any("modelName", f.Metadata().Name).
 		Any("architecture", f.Architecture().Architecture).Msgf("GGUF file loaded: %s", u.Args[0])
 
 	log.Info().Any("tokenizer", fmt.Sprintf("%+v", f.Tokenizer())).Msg("Tokenizer")

core/config/application_config.go

@@ -2,11 +2,11 @@ package config
 
 import (
 	"context"
-	"embed"
 	"encoding/json"
 	"regexp"
 	"time"
 
+	rice "github.com/GeertJohan/go.rice"
 	"github.com/mudler/LocalAI/pkg/xsysinfo"
 	"github.com/rs/zerolog/log"
 )
@@ -19,20 +19,21 @@ type ApplicationConfig struct {
 	UploadLimitMB, Threads, ContextSize int
 	F16                                 bool
 	Debug                               bool
-	ImageDir                            string
+	GeneratedContentDir                 string
-	AudioDir                            string
+
-	UploadDir                           string
+	ConfigsDir string
-	ConfigsDir                          string
+	UploadDir  string
-	DynamicConfigsDir                   string
+
-	DynamicConfigsDirPollInterval       time.Duration
+	DynamicConfigsDir             string
-	CORS                                bool
+	DynamicConfigsDirPollInterval time.Duration
-	CSRF                                bool
+	CORS                          bool
-	PreloadJSONModels                   string
+	CSRF                          bool
-	PreloadModelsFromPath               string
+	PreloadJSONModels             string
-	CORSAllowOrigins                    string
+	PreloadModelsFromPath         string
-	ApiKeys                             []string
+	CORSAllowOrigins              string
-	P2PToken                            string
+	ApiKeys                       []string
-	P2PNetworkID                        string
+	P2PToken                      string
+	P2PNetworkID                  string
 
 	DisableWebUI                       bool
 	EnforcePredownloadScans            bool
@@ -46,7 +47,7 @@ type ApplicationConfig struct {
 
 	Galleries []Gallery
 
-	BackendAssets     embed.FS
+	BackendAssets     *rice.Box
 	AssetsDestination string
 
 	ExternalGRPCBackends map[string]string
@@ -197,7 +198,7 @@ func WithBackendAssetsOutput(out string) AppOption {
 	}
 }
 
-func WithBackendAssets(f embed.FS) AppOption {
+func WithBackendAssets(f *rice.Box) AppOption {
 	return func(o *ApplicationConfig) {
 		o.BackendAssets = f
 	}
@@ -279,15 +280,9 @@ func WithDebug(debug bool) AppOption {
 	}
 }
 
-func WithAudioDir(audioDir string) AppOption {
+func WithGeneratedContentDir(generatedContentDir string) AppOption {
 	return func(o *ApplicationConfig) {
-		o.AudioDir = audioDir
+		o.GeneratedContentDir = generatedContentDir
-	}
-}
-
-func WithImageDir(imageDir string) AppOption {
-	return func(o *ApplicationConfig) {
-		o.ImageDir = imageDir
 	}
 }
 

core/config/backend_config.go

@@ -50,9 +50,6 @@ type BackendConfig struct {
 	// LLM configs (GPT4ALL, Llama.cpp, ...)
 	LLMConfig `yaml:",inline"`
 
-	// AutoGPTQ specifics
-	AutoGPTQ AutoGPTQ `yaml:"autogptq"`
-
 	// Diffusers
 	Diffusers Diffusers `yaml:"diffusers"`
 	Step      int       `yaml:"step"`
@@ -176,14 +173,6 @@ type LimitMMPerPrompt struct {
 	LimitAudioPerPrompt int `yaml:"audio"`
 }
 
-// AutoGPTQ is a struct that holds the configuration specific to the AutoGPTQ backend
-type AutoGPTQ struct {
-	ModelBaseName    string `yaml:"model_base_name"`
-	Device           string `yaml:"device"`
-	Triton           bool   `yaml:"triton"`
-	UseFastTokenizer bool   `yaml:"use_fast_tokenizer"`
-}
-
 // TemplateConfig is a struct that holds the configuration of the templating system
 type TemplateConfig struct {
 	// Chat is the template used in the chat completion endpoint
@@ -315,9 +304,6 @@ func (cfg *BackendConfig) SetDefaults(opts ...ConfigLoaderOption) {
 	defaultTFZ := 1.0
 	defaultZero := 0
 
-	// Try to offload all GPU layers (if GPU is found)
-	defaultHigh := 99999999
-
 	trueV := true
 	falseV := false
 
@@ -377,9 +363,6 @@ func (cfg *BackendConfig) SetDefaults(opts ...ConfigLoaderOption) {
 	if cfg.MirostatTAU == nil {
 		cfg.MirostatTAU = &defaultMirostatTAU
 	}
-	if cfg.NGPULayers == nil {
-		cfg.NGPULayers = &defaultHigh
-	}
 
 	if cfg.LowVRAM == nil {
 		cfg.LowVRAM = &falseV
@@ -447,18 +430,19 @@ func (c *BackendConfig) HasTemplate() bool {
 type BackendConfigUsecases int
 
 const (
-	FLAG_ANY              BackendConfigUsecases = 0b00000000000
+	FLAG_ANY              BackendConfigUsecases = 0b000000000000
-	FLAG_CHAT             BackendConfigUsecases = 0b00000000001
+	FLAG_CHAT             BackendConfigUsecases = 0b000000000001
-	FLAG_COMPLETION       BackendConfigUsecases = 0b00000000010
+	FLAG_COMPLETION       BackendConfigUsecases = 0b000000000010
-	FLAG_EDIT             BackendConfigUsecases = 0b00000000100
+	FLAG_EDIT             BackendConfigUsecases = 0b000000000100
-	FLAG_EMBEDDINGS       BackendConfigUsecases = 0b00000001000
+	FLAG_EMBEDDINGS       BackendConfigUsecases = 0b000000001000
-	FLAG_RERANK           BackendConfigUsecases = 0b00000010000
+	FLAG_RERANK           BackendConfigUsecases = 0b000000010000
-	FLAG_IMAGE            BackendConfigUsecases = 0b00000100000
+	FLAG_IMAGE            BackendConfigUsecases = 0b000000100000
-	FLAG_TRANSCRIPT       BackendConfigUsecases = 0b00001000000
+	FLAG_TRANSCRIPT       BackendConfigUsecases = 0b000001000000
-	FLAG_TTS              BackendConfigUsecases = 0b00010000000
+	FLAG_TTS              BackendConfigUsecases = 0b000010000000
-	FLAG_SOUND_GENERATION BackendConfigUsecases = 0b00100000000
+	FLAG_SOUND_GENERATION BackendConfigUsecases = 0b000100000000
-	FLAG_TOKENIZE         BackendConfigUsecases = 0b01000000000
+	FLAG_TOKENIZE         BackendConfigUsecases = 0b001000000000
-	FLAG_VAD              BackendConfigUsecases = 0b10000000000
+	FLAG_VAD              BackendConfigUsecases = 0b010000000000
+	FLAG_VIDEO            BackendConfigUsecases = 0b100000000000
 
 	// Common Subsets
 	FLAG_LLM BackendConfigUsecases = FLAG_CHAT | FLAG_COMPLETION | FLAG_EDIT
@@ -479,6 +463,7 @@ func GetAllBackendConfigUsecases() map[string]BackendConfigUsecases {
 		"FLAG_TOKENIZE":         FLAG_TOKENIZE,
 		"FLAG_VAD":              FLAG_VAD,
 		"FLAG_LLM":              FLAG_LLM,
+		"FLAG_VIDEO":            FLAG_VIDEO,
 	}
 }
 
@@ -543,6 +528,17 @@ func (c *BackendConfig) GuessUsecases(u BackendConfigUsecases) bool {
 			return false
 		}
 
+	}
+	if (u & FLAG_VIDEO) == FLAG_VIDEO {
+		videoBackends := []string{"diffusers", "stablediffusion"}
+		if !slices.Contains(videoBackends, c.Backend) {
+			return false
+		}
+
+		if c.Backend == "diffusers" && c.Diffusers.PipelineType == "" {
+			return false
+		}
+
 	}
 	if (u & FLAG_RERANK) == FLAG_RERANK {
 		if c.Backend != "rerankers" {
@@ -555,7 +551,7 @@ func (c *BackendConfig) GuessUsecases(u BackendConfigUsecases) bool {
 		}
 	}
 	if (u & FLAG_TTS) == FLAG_TTS {
-		ttsBackends := []string{"piper", "transformers-musicgen", "parler-tts"}
+		ttsBackends := []string{"bark-cpp", "parler-tts", "piper", "transformers-musicgen"}
 		if !slices.Contains(ttsBackends, c.Backend) {
 			return false
 		}

core/config/gguf.go

@@ -3,9 +3,10 @@ package config
 import (
 	"strings"
 
+	"github.com/mudler/LocalAI/pkg/xsysinfo"
 	"github.com/rs/zerolog/log"
 
-	gguf "github.com/thxcode/gguf-parser-go"
+	gguf "github.com/gpustack/gguf-parser-go"
 )
 
 type familyType uint8
@@ -23,6 +24,7 @@ const (
 
 const (
 	defaultContextSize = 1024
+	defaultNGPULayers  = 99999999
 )
 
 type settingsConfig struct {
@@ -147,7 +149,7 @@ var knownTemplates = map[string]familyType{
 func guessGGUFFromFile(cfg *BackendConfig, f *gguf.GGUFFile, defaultCtx int) {
 
 	if defaultCtx == 0 && cfg.ContextSize == nil {
-		ctxSize := f.EstimateLLaMACppUsage().ContextSize
+		ctxSize := f.EstimateLLaMACppRun().ContextSize
 		if ctxSize > 0 {
 			cSize := int(ctxSize)
 			cfg.ContextSize = &cSize
@@ -157,6 +159,46 @@ func guessGGUFFromFile(cfg *BackendConfig, f *gguf.GGUFFile, defaultCtx int) {
 		}
 	}
 
+	// GPU options
+	if cfg.Options == nil {
+		if xsysinfo.HasGPU("nvidia") || xsysinfo.HasGPU("amd") {
+			cfg.Options = []string{"gpu"}
+		}
+	}
+
+	// vram estimation
+	vram, err := xsysinfo.TotalAvailableVRAM()
+	if err != nil {
+		log.Error().Msgf("guessDefaultsFromFile(TotalAvailableVRAM): %s", err)
+	} else if vram > 0 {
+		estimate, err := xsysinfo.EstimateGGUFVRAMUsage(f, vram)
+		if err != nil {
+			log.Error().Msgf("guessDefaultsFromFile(EstimateGGUFVRAMUsage): %s", err)
+		} else {
+			if estimate.IsFullOffload {
+				log.Warn().Msgf("guessDefaultsFromFile: %s", "full offload is recommended")
+			}
+
+			if estimate.EstimatedVRAM > vram {
+				log.Warn().Msgf("guessDefaultsFromFile: %s", "estimated VRAM usage is greater than available VRAM")
+			}
+
+			if cfg.NGPULayers == nil && estimate.EstimatedLayers > 0 {
+				log.Debug().Msgf("guessDefaultsFromFile: %d layers estimated", estimate.EstimatedLayers)
+				cfg.NGPULayers = &estimate.EstimatedLayers
+			}
+		}
+	}
+
+	if cfg.NGPULayers == nil {
+		// we assume we want to offload all layers
+		defaultHigh := defaultNGPULayers
+		cfg.NGPULayers = &defaultHigh
+	}
+
+	log.Debug().Any("NGPULayers", cfg.NGPULayers).Msgf("guessDefaultsFromFile: %s", "NGPULayers set")
+
+	// template estimations
 	if cfg.HasTemplate() {
 		// nothing to guess here
 		log.Debug().Any("name", cfg.Name).Msgf("guessDefaultsFromFile: %s", "template already set")
@@ -166,12 +208,12 @@ func guessGGUFFromFile(cfg *BackendConfig, f *gguf.GGUFFile, defaultCtx int) {
 	log.Debug().
 		Any("eosTokenID", f.Tokenizer().EOSTokenID).
 		Any("bosTokenID", f.Tokenizer().BOSTokenID).
-		Any("modelName", f.Model().Name).
+		Any("modelName", f.Metadata().Name).
 		Any("architecture", f.Architecture().Architecture).Msgf("Model file loaded: %s", cfg.ModelFileName())
 
 	// guess the name
 	if cfg.Name == "" {
-		cfg.Name = f.Model().Name
+		cfg.Name = f.Metadata().Name
 	}
 
 	family := identifyFamily(f)
@@ -207,6 +249,7 @@ func guessGGUFFromFile(cfg *BackendConfig, f *gguf.GGUFFile, defaultCtx int) {
 		cfg.TemplateConfig.JinjaTemplate = true
 		cfg.TemplateConfig.ChatMessage = chatTemplate.ValueString()
 	}
+
 }
 
 func identifyFamily(f *gguf.GGUFFile) familyType {
@@ -231,7 +274,7 @@ func identifyFamily(f *gguf.GGUFFile) familyType {
 	commandR := arch == "command-r" && eosTokenID == 255001
 	qwen2 := arch == "qwen2"
 	phi3 := arch == "phi-3"
-	gemma := strings.HasPrefix(arch, "gemma") || strings.Contains(strings.ToLower(f.Model().Name), "gemma")
+	gemma := strings.HasPrefix(arch, "gemma") || strings.Contains(strings.ToLower(f.Metadata().Name), "gemma")
 	deepseek2 := arch == "deepseek2"
 
 	switch {

core/config/guesser.go

@@ -4,8 +4,8 @@ import (
 	"os"
 	"path/filepath"
 
+	gguf "github.com/gpustack/gguf-parser-go"
 	"github.com/rs/zerolog/log"
-	gguf "github.com/thxcode/gguf-parser-go"
 )
 
 func guessDefaultsFromFile(cfg *BackendConfig, modelPath string, defaultCtx int) {

core/http/app.go

@@ -5,6 +5,8 @@ import (
 	"errors"
 	"fmt"
 	"net/http"
+	"os"
+	"path/filepath"
 
 	"github.com/dave-gray101/v2keyauth"
 	"github.com/mudler/LocalAI/pkg/utils"
@@ -142,9 +144,9 @@ func API(application *application.Application) (*fiber.App, error) {
 	httpFS := http.FS(embedDirStatic)
 
 	router.Use(favicon.New(favicon.Config{
-		URL:        "/favicon.ico",
+		URL:        "/favicon.svg",
 		FileSystem: httpFS,
-		File:       "static/favicon.ico",
+		File:       "static/favicon.svg",
 	}))
 
 	router.Use("/static", filesystem.New(filesystem.Config{
@@ -153,12 +155,19 @@ func API(application *application.Application) (*fiber.App, error) {
 		Browse:     true,
 	}))
 
-	if application.ApplicationConfig().ImageDir != "" {
+	if application.ApplicationConfig().GeneratedContentDir != "" {
-		router.Static("/generated-images", application.ApplicationConfig().ImageDir)
+		os.MkdirAll(application.ApplicationConfig().GeneratedContentDir, 0750)
-	}
+		audioPath := filepath.Join(application.ApplicationConfig().GeneratedContentDir, "audio")
+		imagePath := filepath.Join(application.ApplicationConfig().GeneratedContentDir, "images")
+		videoPath := filepath.Join(application.ApplicationConfig().GeneratedContentDir, "videos")
 
-	if application.ApplicationConfig().AudioDir != "" {
+		os.MkdirAll(audioPath, 0750)
-		router.Static("/generated-audio", application.ApplicationConfig().AudioDir)
+		os.MkdirAll(imagePath, 0750)
+		os.MkdirAll(videoPath, 0750)
+
+		router.Static("/generated-audio", audioPath)
+		router.Static("/generated-images", imagePath)
+		router.Static("/generated-videos", videoPath)
 	}
 
 	// Auth is applied to _all_ endpoints. No exceptions. Filtering out endpoints to bypass is the role of the Filter property of the KeyAuth Configuration

core/http/app_test.go

@@ -3,7 +3,6 @@ package http_test
 import (
 	"bytes"
 	"context"
-	"embed"
 	"encoding/json"
 	"fmt"
 	"io"
@@ -24,6 +23,7 @@ import (
 	. "github.com/onsi/gomega"
 	"gopkg.in/yaml.v3"
 
+	rice "github.com/GeertJohan/go.rice"
 	openaigo "github.com/otiai10/openaigo"
 	"github.com/sashabaranov/go-openai"
 	"github.com/sashabaranov/go-openai/jsonschema"
@@ -264,8 +264,15 @@ func getRequest(url string, header http.Header) (error, int, []byte) {
 
 const bertEmbeddingsURL = `https://gist.githubusercontent.com/mudler/0a080b166b87640e8644b09c2aee6e3b/raw/f0e8c26bb72edc16d9fbafbfd6638072126ff225/bert-embeddings-gallery.yaml`
 
-//go:embed backend-assets/*
+var backendAssets *rice.Box
-var backendAssets embed.FS
+
+func init() {
+	var err error
+	backendAssets, err = rice.FindBox("backend-assets")
+	if err != nil {
+		panic(err)
+	}
+}
 
 var _ = Describe("API test", func() {
 
@@ -629,8 +636,7 @@ var _ = Describe("API test", func() {
 			application, err := application.New(
 				append(commonOpts,
 					config.WithContext(c),
-					config.WithAudioDir(tmpdir),
+					config.WithGeneratedContentDir(tmpdir),
-					config.WithImageDir(tmpdir),
 					config.WithGalleries(galleries),
 					config.WithModelPath(modelDir),
 					config.WithBackendAssets(backendAssets),

core/http/elements/gallery.go

@@ -122,15 +122,15 @@ func modelModal(m *gallery.GalleryModel) elem.Node {
 			"id":          modalName(m),
 			"tabindex":    "-1",
 			"aria-hidden": "true",
-			"class":       "hidden overflow-y-auto overflow-x-hidden fixed top-0 right-0 left-0 z-50 justify-center items-center w-full md:inset-0 h-[calc(100%-1rem)] max-h-full",
+			"class":       "hidden fixed top-0 right-0 left-0 z-50 justify-center items-center w-full md:inset-0 h-full max-h-full bg-gray-900/50",
 		},
 		elem.Div(
 			attrs.Props{
-				"class": "relative p-4 w-full max-w-2xl max-h-full",
+				"class": "relative p-4 w-full max-w-2xl h-[90vh] mx-auto mt-[5vh]",
 			},
 			elem.Div(
 				attrs.Props{
-					"class": "relative p-4 w-full max-w-2xl max-h-full bg-white rounded-lg shadow dark:bg-gray-700",
+					"class": "relative bg-white rounded-lg shadow dark:bg-gray-700 h-full flex flex-col",
 				},
 				// header
 				elem.Div(
@@ -164,14 +164,13 @@ func modelModal(m *gallery.GalleryModel) elem.Node {
 				// body
 				elem.Div(
 					attrs.Props{
-						"class": "p-4 md:p-5 space-y-4",
+						"class": "p-4 md:p-5 space-y-4 overflow-y-auto flex-1 min-h-0",
 					},
 					elem.Div(
 						attrs.Props{
 							"class": "flex justify-center items-center",
 						},
 						elem.Img(attrs.Props{
-							//	"class": "rounded-t-lg object-fit object-center h-96",
 							"class":   "lazy rounded-t-lg max-h-48 max-w-96 object-cover mt-3 entered loaded",
 							"src":     m.Icon,
 							"loading": "lazy",
@@ -232,7 +231,6 @@ func modelModal(m *gallery.GalleryModel) elem.Node {
 			),
 		),
 	)
-
 }
 
 func modelDescription(m *gallery.GalleryModel) elem.Node {

core/http/endpoints/localai/stores.go

@@ -21,6 +21,7 @@ func StoresSetEndpoint(sl *model.ModelLoader, appConfig *config.ApplicationConfi
 		if err != nil {
 			return err
 		}
+		defer sl.Close()
 
 		vals := make([][]byte, len(input.Values))
 		for i, v := range input.Values {
@@ -48,6 +49,7 @@ func StoresDeleteEndpoint(sl *model.ModelLoader, appConfig *config.ApplicationCo
 		if err != nil {
 			return err
 		}
+		defer sl.Close()
 
 		if err := store.DeleteCols(c.Context(), sb, input.Keys); err != nil {
 			return err
@@ -69,6 +71,7 @@ func StoresGetEndpoint(sl *model.ModelLoader, appConfig *config.ApplicationConfi
 		if err != nil {
 			return err
 		}
+		defer sl.Close()
 
 		keys, vals, err := store.GetCols(c.Context(), sb, input.Keys)
 		if err != nil {
@@ -100,6 +103,7 @@ func StoresFindEndpoint(sl *model.ModelLoader, appConfig *config.ApplicationConf
 		if err != nil {
 			return err
 		}
+		defer sl.Close()
 
 		keys, vals, similarities, err := store.Find(c.Context(), sb, input.Key, input.Topk)
 		if err != nil {

core/http/endpoints/localai/video.go

@@ -0,0 +1,205 @@
+package localai
+
+import (
+	"bufio"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/middleware"
+	"github.com/mudler/LocalAI/core/schema"
+
+	"github.com/mudler/LocalAI/core/backend"
+
+	"github.com/gofiber/fiber/v2"
+	model "github.com/mudler/LocalAI/pkg/model"
+	"github.com/rs/zerolog/log"
+)
+
+func downloadFile(url string) (string, error) {
+	// Get the data
+	resp, err := http.Get(url)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+
+	// Create the file
+	out, err := os.CreateTemp("", "video")
+	if err != nil {
+		return "", err
+	}
+	defer out.Close()
+
+	// Write the body to file
+	_, err = io.Copy(out, resp.Body)
+	return out.Name(), err
+}
+
+//
+
+/*
+*
+
+	curl http://localhost:8080/v1/images/generations \
+	  -H "Content-Type: application/json" \
+	  -d '{
+	    "prompt": "A cute baby sea otter",
+	    "n": 1,
+	    "size": "512x512"
+	  }'
+
+*
+*/
+// VideoEndpoint
+// @Summary Creates a video given a prompt.
+// @Param request body schema.OpenAIRequest true "query params"
+// @Success 200 {object} schema.OpenAIResponse "Response"
+// @Router /video [post]
+func VideoEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		input, ok := c.Locals(middleware.CONTEXT_LOCALS_KEY_LOCALAI_REQUEST).(*schema.VideoRequest)
+		if !ok || input.Model == "" {
+			log.Error().Msg("Video Endpoint - Invalid Input")
+			return fiber.ErrBadRequest
+		}
+
+		config, ok := c.Locals(middleware.CONTEXT_LOCALS_KEY_MODEL_CONFIG).(*config.BackendConfig)
+		if !ok || config == nil {
+			log.Error().Msg("Video Endpoint - Invalid Config")
+			return fiber.ErrBadRequest
+		}
+
+		src := ""
+		if input.StartImage != "" {
+
+			var fileData []byte
+			var err error
+			// check if input.File is an URL, if so download it and save it
+			// to a temporary file
+			if strings.HasPrefix(input.StartImage, "http://") || strings.HasPrefix(input.StartImage, "https://") {
+				out, err := downloadFile(input.StartImage)
+				if err != nil {
+					return fmt.Errorf("failed downloading file:%w", err)
+				}
+				defer os.RemoveAll(out)
+
+				fileData, err = os.ReadFile(out)
+				if err != nil {
+					return fmt.Errorf("failed reading file:%w", err)
+				}
+
+			} else {
+				// base 64 decode the file and write it somewhere
+				// that we will cleanup
+				fileData, err = base64.StdEncoding.DecodeString(input.StartImage)
+				if err != nil {
+					return err
+				}
+			}
+
+			// Create a temporary file
+			outputFile, err := os.CreateTemp(appConfig.GeneratedContentDir, "b64")
+			if err != nil {
+				return err
+			}
+			// write the base64 result
+			writer := bufio.NewWriter(outputFile)
+			_, err = writer.Write(fileData)
+			if err != nil {
+				outputFile.Close()
+				return err
+			}
+			outputFile.Close()
+			src = outputFile.Name()
+			defer os.RemoveAll(src)
+		}
+
+		log.Debug().Msgf("Parameter Config: %+v", config)
+
+		switch config.Backend {
+		case "stablediffusion":
+			config.Backend = model.StableDiffusionGGMLBackend
+		case "":
+			config.Backend = model.StableDiffusionGGMLBackend
+		}
+
+		width := input.Width
+		height := input.Height
+
+		if width == 0 {
+			width = 512
+		}
+		if height == 0 {
+			height = 512
+		}
+
+		b64JSON := input.ResponseFormat == "b64_json"
+
+		tempDir := ""
+		if !b64JSON {
+			tempDir = filepath.Join(appConfig.GeneratedContentDir, "videos")
+		}
+		// Create a temporary file
+		outputFile, err := os.CreateTemp(tempDir, "b64")
+		if err != nil {
+			return err
+		}
+		outputFile.Close()
+
+		// TODO: use mime type to determine the extension
+		output := outputFile.Name() + ".mp4"
+
+		// Rename the temporary file
+		err = os.Rename(outputFile.Name(), output)
+		if err != nil {
+			return err
+		}
+
+		baseURL := c.BaseURL()
+
+		fn, err := backend.VideoGeneration(height, width, input.Prompt, src, input.EndImage, output, ml, *config, appConfig)
+		if err != nil {
+			return err
+		}
+		if err := fn(); err != nil {
+			return err
+		}
+
+		item := &schema.Item{}
+
+		if b64JSON {
+			defer os.RemoveAll(output)
+			data, err := os.ReadFile(output)
+			if err != nil {
+				return err
+			}
+			item.B64JSON = base64.StdEncoding.EncodeToString(data)
+		} else {
+			base := filepath.Base(output)
+			item.URL = baseURL + "/generated-videos/" + base
+		}
+
+		id := uuid.New().String()
+		created := int(time.Now().Unix())
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Data:    []schema.Item{*item},
+		}
+
+		jsonResult, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", jsonResult)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

core/http/endpoints/openai/assistant_test.go

@@ -40,7 +40,7 @@ func TestAssistantEndpoints(t *testing.T) {
 	cl := &config.BackendConfigLoader{}
 	//configsDir := "/tmp/localai/configs"
 	modelPath := "/tmp/localai/model"
-	var ml = model.NewModelLoader(modelPath)
+	var ml = model.NewModelLoader(modelPath, false)
 
 	appConfig := &config.ApplicationConfig{
 		ConfigsDir:    configsDir,

core/http/endpoints/openai/image.go

@@ -72,7 +72,7 @@ func ImageEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appCon
 			log.Error().Msg("Image Endpoint - Invalid Input")
 			return fiber.ErrBadRequest
 		}
-		
+
 		config, ok := c.Locals(middleware.CONTEXT_LOCALS_KEY_MODEL_CONFIG).(*config.BackendConfig)
 		if !ok || config == nil {
 			log.Error().Msg("Image Endpoint - Invalid Config")
@@ -108,7 +108,7 @@ func ImageEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appCon
 			}
 
 			// Create a temporary file
-			outputFile, err := os.CreateTemp(appConfig.ImageDir, "b64")
+			outputFile, err := os.CreateTemp(appConfig.GeneratedContentDir, "b64")
 			if err != nil {
 				return err
 			}
@@ -184,7 +184,7 @@ func ImageEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appCon
 
 				tempDir := ""
 				if !b64JSON {
-					tempDir = appConfig.ImageDir
+					tempDir = filepath.Join(appConfig.GeneratedContentDir, "images")
 				}
 				// Create a temporary file
 				outputFile, err := os.CreateTemp(tempDir, "b64")
@@ -192,6 +192,7 @@ func ImageEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appCon
 					return err
 				}
 				outputFile.Close()
+
 				output := outputFile.Name() + ".png"
 				// Rename the temporary file
 				err = os.Rename(outputFile.Name(), output)

core/http/explorer.go

@@ -29,9 +29,9 @@ func Explorer(db *explorer.Database) *fiber.App {
 	httpFS := http.FS(embedDirStatic)
 
 	app.Use(favicon.New(favicon.Config{
-		URL:        "/favicon.ico",
+		URL:        "/favicon.svg",
 		FileSystem: httpFS,
-		File:       "static/favicon.ico",
+		File:       "static/favicon.svg",
 	}))
 
 	app.Use("/static", filesystem.New(filesystem.Config{

core/http/middleware/request.go

@@ -203,18 +203,10 @@ func mergeOpenAIRequestAndBackendConfig(config *config.BackendConfig, input *sch
 		config.Diffusers.ClipSkip = input.ClipSkip
 	}
 
-	if input.ModelBaseName != "" {
-		config.AutoGPTQ.ModelBaseName = input.ModelBaseName
-	}
-
 	if input.NegativePromptScale != 0 {
 		config.NegativePromptScale = input.NegativePromptScale
 	}
 
-	if input.UseFastTokenizer {
-		config.UseFastTokenizer = input.UseFastTokenizer
-	}
-
 	if input.NegativePrompt != "" {
 		config.NegativePrompt = input.NegativePrompt
 	}

core/http/routes/localai.go

@@ -50,16 +50,20 @@ func RegisterLocalAIRoutes(router *fiber.App,
 	router.Post("/v1/vad", vadChain...)
 
 	// Stores
-	sl := model.NewModelLoader("")
+	router.Post("/stores/set", localai.StoresSetEndpoint(ml, appConfig))
-	router.Post("/stores/set", localai.StoresSetEndpoint(sl, appConfig))
+	router.Post("/stores/delete", localai.StoresDeleteEndpoint(ml, appConfig))
-	router.Post("/stores/delete", localai.StoresDeleteEndpoint(sl, appConfig))
+	router.Post("/stores/get", localai.StoresGetEndpoint(ml, appConfig))
-	router.Post("/stores/get", localai.StoresGetEndpoint(sl, appConfig))
+	router.Post("/stores/find", localai.StoresFindEndpoint(ml, appConfig))
-	router.Post("/stores/find", localai.StoresFindEndpoint(sl, appConfig))
 
 	if !appConfig.DisableMetrics {
 		router.Get("/metrics", localai.LocalAIMetricsEndpoint())
 	}
 
+	router.Post("/video",
+		requestExtractor.BuildFilteredFirstAvailableDefaultModel(config.BuildUsecaseFilterFn(config.FLAG_VIDEO)),
+		requestExtractor.SetModelAndConfig(func() schema.LocalAIRequest { return new(schema.VideoRequest) }),
+		localai.VideoEndpoint(cl, ml, appConfig))
+
 	// Backend Statistics Module
 	// TODO: Should these use standard middlewares? Refactor later, they are extremely simple.
 	backendMonitorService := services.NewBackendMonitorService(ml, cl, appConfig) // Split out for now

core/http/static/talk.js

@@ -115,6 +115,7 @@ async function sendTextToChatGPT(text) {
 
     const response = await fetch('v1/chat/completions', {
         method: 'POST',
+        headers: { "Content-Type": "application/json" },
         body: JSON.stringify({
             model: getModel(),
             messages: conversationHistory

core/http/views/login.html

@@ -12,7 +12,7 @@
         <div class="max-w-md w-full bg-gray-800/90 border border-gray-700/50 rounded-xl overflow-hidden shadow-xl">
             <div class="animation-container">
                 <div class="text-overlay">
-                <!--    <i class="fas fa-circle-nodes text-5xl text-blue-400 mb-2"></i> -->
+                    <img src="static/logo.png" alt="LocalAI Logo" class="h-32">
                 </div>
             </div>
             

core/http/views/partials/head.html

@@ -3,7 +3,7 @@
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
   <title>{{.Title}}</title>
   <base href="{{.BaseURL}}" />
-  <link rel="icon" type="image/x-icon" href="favicon.ico" />
+  <link rel="shortcut icon" href="static/favicon.svg" type="image/svg">
   <link rel="stylesheet" href="static/assets/highlightjs.css" />
   <script defer src="static/assets/highlightjs.js"></script>
   <script defer src="static/assets/alpine.js"></script>