fix(deps): update module github.com/onsi/gomega to v1.29.0

* Fix backend/cpp/llama CMakeList.txt on OSX - detect OSX and use homebrew libraries

* sneak a logging fix in too for gallery debugging

* additional logging

.devcontainer/Dockerfile

@@ -1,3 +0,0 @@
-ARG GO_VERSION=1.20
-FROM mcr.microsoft.com/devcontainers/go:0-$GO_VERSION-bullseye
-RUN apt-get update && apt-get install -y cmake

.devcontainer/devcontainer.json

@@ -1,46 +0,0 @@
-// For format details, see https://aka.ms/devcontainer.json. For config options, see the
-// README at: https://github.com/devcontainers/templates/tree/main/src/docker-existing-docker-compose
-{
-	"name": "Existing Docker Compose (Extend)",
-
-	// Update the 'dockerComposeFile' list if you have more compose files or use different names.
-	// The .devcontainer/docker-compose.yml file contains any overrides you need/want to make.
-	"dockerComposeFile": [
-		"../docker-compose.yaml",
-		"docker-compose.yml"
-	],
-
-	// The 'service' property is the name of the service for the container that VS Code should
-	// use. Update this value and .devcontainer/docker-compose.yml to the real service name.
-	"service": "api",
-
-	// The optional 'workspaceFolder' property is the path VS Code should open by default when
-	// connected. This is typically a file mount in .devcontainer/docker-compose.yml
-	"workspaceFolder": "/workspace",
-
-	"features": {
-		"ghcr.io/devcontainers/features/go:1": {},
-		"ghcr.io/azutake/devcontainer-features/go-packages-install:0": {}
-	},
-
-	// Features to add to the dev container. More info: https://containers.dev/features.
-	// "features": {},
-
-	// Use 'forwardPorts' to make a list of ports inside the container available locally.
-	// "forwardPorts": [],
-
-	// Uncomment the next line if you want start specific services in your Docker Compose config.
-	// "runServices": [],
-
-	// Uncomment the next line if you want to keep your containers running after VS Code shuts down.
-	// "shutdownAction": "none",
-
-	// Uncomment the next line to run commands after the container is created.
-	"postCreateCommand": "make prepare"
-
-	// Configure tool-specific properties.
-	// "customizations": {},
-
-	// Uncomment to connect as an existing user other than the container default. More info: https://aka.ms/dev-containers-non-root.
-	// "remoteUser": "devcontainer"
-}

.devcontainer/docker-compose.yml

@@ -1,26 +0,0 @@
-version: '3.6'
-services:
-  # Update this to the name of the service you want to work with in your docker-compose.yml file
-  api:
-    # Uncomment if you want to override the service's Dockerfile to one in the .devcontainer 
-    # folder. Note that the path of the Dockerfile and context is relative to the *primary* 
-    # docker-compose.yml file (the first in the devcontainer.json "dockerComposeFile"
-    # array). The sample below assumes your primary file is in the root of your project.
-    #
-    build:
-      context: .
-      dockerfile: .devcontainer/Dockerfile
-
-    volumes:
-      # Update this to wherever you want VS Code to mount the folder of your project
-      - .:/workspace:cached
-
-    # Uncomment the next four lines if you will use a ptrace-based debugger like C++, Go, and Rust.
-    # cap_add:
-    #   - SYS_PTRACE
-    # security_opt:
-    #   - seccomp:unconfined
-
-    # Overrides default command so things don't shut down after the process ends.
-    command: /bin/sh -c "while sleep 1000; do :; done"
- 

.dockerignore

@@ -1,3 +1,4 @@
+.idea
 models
 examples/chatbot-ui/models
 examples/rwkv/models

.env

@@ -7,24 +7,63 @@
 
 ## Default models context size
 # CONTEXT_SIZE=512
+#
+## Define galleries.
+## models will to install will be visible in `/models/available`
+# GALLERIES=[{"name":"model-gallery", "url":"github:go-skynet/model-gallery/index.yaml"}]
+
+## CORS settings
+# CORS=true
+# CORS_ALLOW_ORIGINS=*
 
 ## Default path for models
+#
 MODELS_PATH=/models
 
 ## Enable debug mode
 # DEBUG=true
 
-## Specify a build type. Available: cublas, openblas.
+## Disables COMPEL (Diffusers)
+# COMPEL=0
+
+## Enable/Disable single backend (useful if only one GPU is available)
+# SINGLE_ACTIVE_BACKEND=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.
+## clBLAS:   This is an open-source implementation of the BLAS library that uses OpenCL, a framework for writing programs that execute across heterogeneous platforms consisting of CPUs, GPUs, and other processors. clBLAS is designed to take advantage of the parallel computing power of GPUs but can also run on any hardware that supports OpenCL. This includes hardware from different vendors like Nvidia, AMD, and Intel.
 # BUILD_TYPE=openblas
 
-## Uncomment and set to false to disable rebuilding from source
-# REBUILD=false
+## Uncomment and set to true to enable rebuilding from source
+# REBUILD=true
 
-## Enable image generation with stablediffusion (requires REBUILD=true)
+## Enable go tags, available: stablediffusion, tts
+## stablediffusion: image generation with stablediffusion
+## tts: enables text-to-speech with go-piper 
+## (requires REBUILD=true)
+#
 # GO_TAGS=stablediffusion
 
 ## Path where to store generated images
 # IMAGE_PATH=/tmp
 
 ## Specify a default upload limit in MB (whisper)
-# UPLOAD_LIMIT
+# UPLOAD_LIMIT
+
+## List of external GRPC backends (note on the container image this variable is already set to use extra backends available in extra/)
+# EXTERNAL_GRPC_BACKENDS=my-backend:127.0.0.1:9000,my-backend2:/usr/bin/backend.py
+
+### Advanced settings ###
+### Those are not really used by LocalAI, but from components in the stack ###
+##
+### Preload libraries
+# LD_PRELOAD=
+
+### Huggingface cache for models
+# HUGGINGFACE_HUB_CACHE=/usr/local/huggingface
+
+### Python backends GRPC max workers
+### Default number of workers for GRPC Python backends.
+### This actually controls wether a backend can process multiple requests or not.
+# PYTHON_GRPC_MAX_WORKERS=1

.gitattributes

@@ -0,0 +1 @@
+*.sh text eol=lf

.github/FUNDING.yml

@@ -0,0 +1,5 @@
+# These are supported funding model platforms
+
+github: [mudler]
+custom: 
+- https://www.buymeacoffee.com/mudler

.github/PULL_REQUEST_TEMPLATE.md

@@ -8,16 +8,24 @@ This PR fixes #
 **[Signed commits](../CONTRIBUTING.md#signing-off-on-commits-developer-certificate-of-origin)**
 - [ ] Yes, I signed my commits.
  
-
 <!--
 Thank you for contributing to LocalAI! 
 
-Contributing Conventions:
+Contributing Conventions
+-------------------------
 
-1. Include descriptive PR titles with [<component-name>] prepended.
-2. Build and test your changes before submitting a PR. 
+The draft above helps to give a quick overview of your PR.
+
+Remember to remove this comment and to at least:
+
+1. Include descriptive PR titles with [<component-name>] prepended. We use [conventional commits](https://www.conventionalcommits.org/en/v1.0.0/).
+2. Build and test your changes before submitting a PR (`make build`). 
 3. Sign your commits
+4. **Tag maintainer:** for a quicker response, tag the relevant maintainer (see below).
+5. **X/Twitter handle:** we announce bigger features on X/Twitter. If your PR gets announced, and you'd like a mention, we'll gladly shout you out!
 
 By following the community's contribution conventions upfront, the review process will 
 be accelerated and your PR merged more quickly.
+
+If no one reviews your PR within a few days, please @-mention @mudler.
 -->

.github/workflows/bump_deps.yaml

@@ -12,8 +12,11 @@ jobs:
           - repository: "go-skynet/go-llama.cpp"
             variable: "GOLLAMA_VERSION"
             branch: "master"
-          - repository: "go-skynet/go-gpt2.cpp"
-            variable: "GOGPT2_VERSION"
+          - repository: "ggerganov/llama.cpp"
+            variable: "CPPLLAMA_VERSION"
+            branch: "master"
+          - repository: "go-skynet/go-ggml-transformers.cpp"
+            variable: "GOGGMLTRANSFORMERS_VERSION"
             branch: "master"
           - repository: "donomii/go-rwkv.cpp"
             variable: "RWKV_VERSION"
@@ -30,9 +33,18 @@ jobs:
           - repository: "nomic-ai/gpt4all"
             variable: "GPT4ALL_VERSION"
             branch: "main"
+          - repository: "mudler/go-ggllm.cpp"
+            variable: "GOGGLLM_VERSION"
+            branch: "master"
+          - repository: "mudler/go-stable-diffusion"
+            variable: "STABLEDIFFUSION_VERSION"
+            branch: "master"
+          - repository: "mudler/go-piper"
+            variable: "PIPER_VERSION"
+            branch: "master"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Bump dependencies 🔧
         run: |
           bash .github/bump_deps.sh ${{ matrix.repository }} ${{ matrix.branch }} ${{ matrix.variable }}

.github/workflows/image.yml

@@ -15,34 +15,102 @@ concurrency:
 
 jobs:
   docker:
+    strategy:
+      matrix:
+        include:
+          - build-type: ''
+            #platforms: 'linux/amd64,linux/arm64'
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: ''
+            ffmpeg: ''
+          - build-type: 'cublas'
+            cuda-major-version: 11
+            cuda-minor-version: 7
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11'
+            ffmpeg: ''
+          - build-type: 'cublas'
+            cuda-major-version: 12
+            cuda-minor-version: 1
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12'
+            ffmpeg: ''
+          - build-type: ''
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-ffmpeg'
+            ffmpeg: 'true'
+          - build-type: 'cublas'
+            cuda-major-version: 11
+            cuda-minor-version: 7
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11-ffmpeg'
+            ffmpeg: 'true'
+          - build-type: 'cublas'
+            cuda-major-version: 12
+            cuda-minor-version: 1
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12-ffmpeg'
+            ffmpeg: 'true'
+
     runs-on: ubuntu-latest
     steps:
-      - name: Checkout
-        uses: actions/checkout@v3
-
-      - name: Prepare
-        id: prep
+      - name: Release space from worker
         run: |
-          DOCKER_IMAGE=quay.io/go-skynet/local-ai
-          VERSION=master
-          SHORTREF=${GITHUB_SHA::8}
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          df -h
+          echo
+          sudo apt-get remove -y '^llvm-.*|^libllvm.*' || true
+          sudo apt-get remove --auto-remove android-sdk-platform-tools || true
+          sudo apt-get purge --auto-remove android-sdk-platform-tools || true
+          sudo rm -rf /usr/local/lib/android
+          sudo apt-get remove -y '^dotnet-.*|^aspnetcore-.*' || true
+          sudo rm -rf /usr/share/dotnet
+          sudo apt-get remove -y '^mono-.*' || true
+          sudo apt-get remove -y '^ghc-.*' || true
+          sudo apt-get remove -y '.*jdk.*|.*jre.*' || true
+          sudo apt-get remove -y 'php.*' || true
+          sudo apt-get remove -y hhvm powershell firefox monodoc-manual msbuild || true
+          sudo apt-get remove -y '^google-.*' || true
+          sudo apt-get remove -y azure-cli || true
+          sudo apt-get remove -y '^mongo.*-.*|^postgresql-.*|^mysql-.*|^mssql-.*' || true
+          sudo apt-get remove -y '^gfortran-.*' || true
+          sudo apt-get remove -y microsoft-edge-stable || true
+          sudo apt-get remove -y firefox || true
+          sudo apt-get remove -y powershell || true
+          sudo apt-get remove -y r-base-core || true
+          sudo apt-get autoremove -y
+          sudo apt-get clean
+          echo
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          sudo rm -rfv build || true
+          df -h
+      - name: Checkout
+        uses: actions/checkout@v4
 
-          # If this is git tag, use the tag name as a docker tag
-          if [[ $GITHUB_REF == refs/tags/* ]]; then
-            VERSION=${GITHUB_REF#refs/tags/}
-          fi
-          TAGS="${DOCKER_IMAGE}:${VERSION},${DOCKER_IMAGE}:${SHORTREF}"
-
-          # If the VERSION looks like a version number, assume that
-          # this is the most recent version of the image and also
-          # tag it 'latest'.
-          if [[ $VERSION =~ ^v[0-9]{1,3}\.[0-9]{1,3}\.[0-9]{1,3}$ ]]; then
-            TAGS="$TAGS,${DOCKER_IMAGE}:latest"
-          fi
-
-          # Set output parameters.
-          echo ::set-output name=tags::${TAGS}
-          echo ::set-output name=docker_image::${DOCKER_IMAGE}
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: quay.io/go-skynet/local-ai
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{raw}}
+            type=sha
+          flavor: |
+            latest=${{ matrix.tag-latest }}
+            suffix=${{ matrix.tag-suffix }}
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@master
@@ -55,28 +123,24 @@ jobs:
 
       - name: Login to DockerHub
         if: github.event_name != 'pull_request'
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           registry: quay.io
           username: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
           password: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
-      - name: Build
-        if: github.event_name != 'pull_request'
-        uses: docker/build-push-action@v4
+
+      - name: Build and push
+        uses: docker/build-push-action@v5
         with:
           builder: ${{ steps.buildx.outputs.name }}
+          build-args: |
+            BUILD_TYPE=${{ matrix.build-type }}
+            CUDA_MAJOR_VERSION=${{ matrix.cuda-major-version }}
+            CUDA_MINOR_VERSION=${{ matrix.cuda-minor-version }}
+            FFMPEG=${{ matrix.ffmpeg }}
           context: .
           file: ./Dockerfile
-          platforms: linux/amd64,linux/arm64
-          push: true
-          tags: ${{ steps.prep.outputs.tags }}
-      - name: Build PRs
-        if: github.event_name == 'pull_request'
-        uses: docker/build-push-action@v4
-        with:
-          builder: ${{ steps.buildx.outputs.name }}
-          context: .
-          file: ./Dockerfile
-          platforms: linux/amd64
-          push: false
-          tags: ${{ steps.prep.outputs.tags }}
+          platforms: ${{ matrix.platforms }}
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}

.github/workflows/release.yaml

@@ -2,6 +2,9 @@ name: Build and Release
 
 on: push
 
+permissions:
+  contents: write
+
 jobs:
   build-linux:
     strategy:
@@ -16,20 +19,29 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v3
-        with: 
+        uses: actions/checkout@v4
+        with:
           submodules: true
+      - uses: actions/setup-go@v4
+        with:
+          go-version: '>=1.21.0'
       - name: Dependencies
         run: |
           sudo apt-get update
           sudo apt-get install build-essential ffmpeg
+
+          git clone --recurse-submodules -b v1.58.0 --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+              cd grpc && mkdir -p cmake/build && cd cmake/build && cmake -DgRPC_INSTALL=ON \
+                -DgRPC_BUILD_TESTS=OFF \
+                ../.. && sudo make -j12 install
+
       - name: Build
         id: build
         env:
-          CMAKE_ARGS: "${{ matrix.define }}"
+          CMAKE_ARGS: "${{ matrix.defines }}"
           BUILD_ID: "${{ matrix.build }}"
         run: |
-          make dist
+          STATIC=true make dist
       - uses: actions/upload-artifact@v3
         with:
           name: ${{ matrix.build }}
@@ -54,20 +66,26 @@ jobs:
     runs-on: macOS-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v3
-        with: 
+        uses: actions/checkout@v4
+        with:
           submodules: true
-
+      - uses: actions/setup-go@v4
+        with:
+          go-version: '>=1.21.0'
       - name: Dependencies
         run: |
-          brew update
-          brew install sdl2 ffmpeg
+          git clone --recurse-submodules -b v1.58.0 --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+              cd grpc && mkdir -p cmake/build && cd cmake/build && cmake -DgRPC_INSTALL=ON \
+                -DgRPC_BUILD_TESTS=OFF \
+                ../.. && make -j12 install && rm -rf grpc
       - name: Build
         id: build
         env:
-          CMAKE_ARGS: "${{ matrix.define }}"
+          CMAKE_ARGS: "${{ matrix.defines }}"
           BUILD_ID: "${{ matrix.build }}"
         run: |
+          export C_INCLUDE_PATH=/usr/local/include
+          export CPLUS_INCLUDE_PATH=/usr/local/include
           make dist
       - uses: actions/upload-artifact@v3
         with:

.github/workflows/test-gpu.yml

@@ -0,0 +1,60 @@
+---
+name: 'GPU tests'
+
+on:
+  pull_request:
+  push:
+    branches:
+      - master
+    tags:
+      - '*'
+
+concurrency:
+  group: ci-gpu-tests-${{ github.head_ref || github.ref }}-${{ github.repository }}
+  cancel-in-progress: true
+
+jobs:
+  ubuntu-latest:
+    runs-on: self-hosted
+    strategy:
+      matrix:
+        go-version: ['1.21.x']
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with: 
+          submodules: true
+      - name: Setup Go ${{ matrix.go-version }}
+        uses: actions/setup-go@v4
+        with:
+          go-version: ${{ matrix.go-version }}
+      # You can test your matrix by printing the current Go version
+      - name: Display Go version
+        run: go version
+      - name: Dependencies
+        run: |
+          sudo apt-get update
+          sudo DEBIAN_FRONTEND=noninteractive apt-get install -y make wget
+      - name: Build
+        run: |
+          if [ ! -e /run/systemd/system ]; then
+            sudo mkdir /run/systemd/system
+          fi
+          make \
+            TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
+            BUILD_TYPE=cublas \
+            prepare-e2e run-e2e-image test-e2e
+      - name: Release space from worker ♻
+        if: always()
+        run: |
+          sudo rm -rf build || true
+          sudo rm -rf bin || true
+          sudo rm -rf dist || true
+          sudo docker logs $(sudo docker ps -q --filter ancestor=localai-tests) > logs.txt
+          sudo cat logs.txt || true
+          sudo rm -rf logs.txt
+          make clean || true
+          make \
+            TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
+            teardown-e2e || true
+          docker system prune -f -a --volumes || true

.github/workflows/test.yml

@@ -16,33 +16,116 @@ concurrency:
 jobs:
   ubuntu-latest:
     runs-on: ubuntu-latest
-
+    strategy:
+      matrix:
+        go-version: ['1.21.x']
     steps:
+      - name: Release space from worker
+        run: |
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          df -h
+          echo
+          sudo apt-get remove -y '^llvm-.*|^libllvm.*' || true
+          sudo apt-get remove --auto-remove android-sdk-platform-tools || true
+          sudo apt-get purge --auto-remove android-sdk-platform-tools || true
+          sudo rm -rf /usr/local/lib/android
+          sudo apt-get remove -y '^dotnet-.*|^aspnetcore-.*' || true
+          sudo rm -rf /usr/share/dotnet
+          sudo apt-get remove -y '^mono-.*' || true
+          sudo apt-get remove -y '^ghc-.*' || true
+          sudo apt-get remove -y '.*jdk.*|.*jre.*' || true
+          sudo apt-get remove -y 'php.*' || true
+          sudo apt-get remove -y hhvm powershell firefox monodoc-manual msbuild || true
+          sudo apt-get remove -y '^google-.*' || true
+          sudo apt-get remove -y azure-cli || true
+          sudo apt-get remove -y '^mongo.*-.*|^postgresql-.*|^mysql-.*|^mssql-.*' || true
+          sudo apt-get remove -y '^gfortran-.*' || true
+          sudo apt-get autoremove -y
+          sudo apt-get clean
+          echo
+          echo "Listing top largest packages"
+          pkgs=$(dpkg-query -Wf '${Installed-Size}\t${Package}\t${Status}\n' | awk '$NF == "installed"{print $1 "\t" $2}' | sort -nr)
+          head -n 30 <<< "${pkgs}"
+          echo
+          sudo rm -rfv build || true
+          df -h
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with: 
           submodules: true
+      - name: Setup Go ${{ matrix.go-version }}
+        uses: actions/setup-go@v4
+        with:
+          go-version: ${{ matrix.go-version }}
+      # You can test your matrix by printing the current Go version
+      - name: Display Go version
+        run: go version
       - name: Dependencies
         run: |
           sudo apt-get update
           sudo apt-get install build-essential ffmpeg
+
+          sudo apt-get install -y ca-certificates cmake curl patch
+          sudo apt-get install -y libopencv-dev && sudo ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
+          sudo pip install -r extra/requirements.txt
+
+
+          # Pre-build stable diffusion before we install a newever version of abseil (not compatible with stablediffusion-ncn)
+          GO_TAGS="tts stablediffusion" GRPC_BACKENDS=backend-assets/grpc/stablediffusion make build
+
+          sudo mkdir /build && sudo chmod -R 777 /build && cd /build && \
+          curl -L "https://github.com/gabime/spdlog/archive/refs/tags/v1.11.0.tar.gz" | \
+          tar -xzvf - && \
+          mkdir -p "spdlog-1.11.0/build" && \
+          cd "spdlog-1.11.0/build" && \
+          cmake ..  && \
+          make -j8 && \
+          sudo cmake --install . --prefix /usr && mkdir -p "lib/Linux-$(uname -m)" && \
+          cd /build && \
+          mkdir -p "lib/Linux-$(uname -m)/piper_phonemize" && \
+          curl -L "https://github.com/rhasspy/piper-phonemize/releases/download/v1.0.0/libpiper_phonemize-amd64.tar.gz" | \
+          tar -C "lib/Linux-$(uname -m)/piper_phonemize" -xzvf - && ls -liah /build/lib/Linux-$(uname -m)/piper_phonemize/ && \
+          sudo cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/lib/. /usr/lib/ && \
+          sudo ln -s /usr/lib/libpiper_phonemize.so /usr/lib/libpiper_phonemize.so.1 && \
+          sudo cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/include/. /usr/include/
+
+          git clone --recurse-submodules -b v1.58.0 --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+              cd grpc && mkdir -p cmake/build && cd cmake/build && cmake -DgRPC_INSTALL=ON \
+                -DgRPC_BUILD_TESTS=OFF \
+                ../.. && sudo make -j12 install
+
       - name: Test
         run: |
-          make test
+          ESPEAK_DATA="/build/lib/Linux-$(uname -m)/piper_phonemize/lib/espeak-ng-data" GO_TAGS="tts stablediffusion" make test
 
   macOS-latest:
     runs-on: macOS-latest
-
+    strategy:
+      matrix:
+        go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with: 
           submodules: true
-
+      - name: Setup Go ${{ matrix.go-version }}
+        uses: actions/setup-go@v4
+        with:
+          go-version: ${{ matrix.go-version }}
+      # You can test your matrix by printing the current Go version
+      - name: Display Go version
+        run: go version
       - name: Dependencies
         run: |
-          brew update
-          brew install sdl2 ffmpeg
+          git clone --recurse-submodules -b v1.58.0 --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+              cd grpc && mkdir -p cmake/build && cd cmake/build && cmake -DgRPC_INSTALL=ON \
+                -DgRPC_BUILD_TESTS=OFF \
+                ../.. && make -j12 install && rm -rf grpc
       - name: Test
         run: |
-          make test
+          export C_INCLUDE_PATH=/usr/local/include
+          export CPLUS_INCLUDE_PATH=/usr/local/include
+          CMAKE_ARGS="-DLLAMA_F16C=OFF -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF" make test

.gitignore

@@ -1,22 +1,43 @@
 # go-llama build artifacts
 go-llama
-gpt4all
+go-llama-stable
+/gpt4all
 go-stable-diffusion
+go-piper
+/go-bert
+go-ggllm
+/piper
+__pycache__/
+*.a
+get-sources
+
+go-ggml-transformers
 go-gpt2
 go-rwkv
 whisper.cpp
+/bloomz
+go-bert
 
 # LocalAI build binary
 LocalAI
 local-ai
 # prevent above rules from omitting the helm chart
 !charts/*
+# prevent above rules from omitting the api/localai folder
+!api/localai
 
 # Ignore models
 models/*
 test-models/
+test-dir/
 
 release/
 
 # just in case
 .DS_Store
+.idea
+
+# Generated during build
+backend-assets/
+prepare
+/ggml-metal.metal

.vscode/launch.json

@@ -25,7 +25,7 @@
             ],
             "env": {
                 "C_INCLUDE_PATH": "${workspaceFolder}/go-llama:${workspaceFolder}/go-stable-diffusion/:${workspaceFolder}/gpt4all/gpt4all-bindings/golang/:${workspaceFolder}/go-gpt2:${workspaceFolder}/go-rwkv:${workspaceFolder}/whisper.cpp:${workspaceFolder}/go-bert:${workspaceFolder}/bloomz",
-                "LIBRARY_PATH": "$${workspaceFolder}/go-llama:${workspaceFolder}/go-stable-diffusion/:${workspaceFolder}/gpt4all/gpt4all-bindings/golang/:${workspaceFolder}/go-gpt2:${workspaceFolder}/go-rwkv:${workspaceFolder}/whisper.cpp:${workspaceFolder}/go-bert:${workspaceFolder}/bloomz",
+                "LIBRARY_PATH": "${workspaceFolder}/go-llama:${workspaceFolder}/go-stable-diffusion/:${workspaceFolder}/gpt4all/gpt4all-bindings/golang/:${workspaceFolder}/go-gpt2:${workspaceFolder}/go-rwkv:${workspaceFolder}/whisper.cpp:${workspaceFolder}/go-bert:${workspaceFolder}/bloomz",
                 "DEBUG": "true"
             }
         }

CONTRIBUTING.md

@@ -0,0 +1,72 @@
+# Contributing to localAI
+
+Thank you for your interest in contributing to LocalAI! We appreciate your time and effort in helping to improve our project. Before you get started, please take a moment to review these guidelines.
+
+## Table of Contents
+
+- [Getting Started](#getting-started)
+  - [Prerequisites](#prerequisites)
+  - [Setting up the Development Environment](#setting-up-the-development-environment)
+- [Contributing](#contributing)
+  - [Submitting an Issue](#submitting-an-issue)
+  - [Creating a Pull Request (PR)](#creating-a-pull-request-pr)
+- [Coding Guidelines](#coding-guidelines)
+- [Testing](#testing)
+- [Documentation](#documentation)
+- [Community and Communication](#community-and-communication)
+
+
+
+## Getting Started
+
+### Prerequisites
+
+- Golang [1.21]
+- Git
+- macOS/Linux
+
+### Setting up the Development Environment and running localAI in the local environment
+
+1. Clone the repository: `git clone https://github.com/go-skynet/LocalAI.git`
+2. Navigate to the project directory: `cd LocalAI`
+3. Install the required dependencies: `make prepare`
+4. Run LocalAI: `make run`
+
+## Contributing
+
+We welcome contributions from everyone! To get started, follow these steps:
+
+### Submitting an Issue
+
+If you find a bug, have a feature request, or encounter any issues, please check the [issue tracker](https://github.com/go-skynet/LocalAI/issues) to see if a similar issue has already been reported. If not, feel free to [create a new issue](https://github.com/go-skynet/LocalAI/issues/new) and provide as much detail as possible.
+
+### Creating a Pull Request (PR)
+
+1. Fork the repository.
+2. Create a new branch with a descriptive name: `git checkout -b [branch name]`
+3. Make your changes and commit them.
+4. Push the changes to your fork: `git push origin [branch name]`
+5. Create a new pull request from your branch to the main project's `main` or `master` branch.
+6. Provide a clear description of your changes in the pull request.
+7. Make any requested changes during the review process.
+8. Once your PR is approved, it will be merged into the main project.
+
+## Coding Guidelines
+
+- No specific coding guidelines at the moment. Please make sure the code can be tested. The most popular lint tools like []`golangci-lint`](https://golangci-lint.run) can help you here.
+
+## Testing
+
+`make test` cannot handle all the model now. Please be sure to add a test case for the new features or the part was changed.
+
+## Documentation
+
+- We are welcome the contribution of the documents, please open new PR in the official document repo [localai-website](https://github.com/go-skynet/localai-website)
+
+## Community and Communication
+
+- You can reach out via the Github issue tracker.
+- Open a new discussion at [Discussion](https://github.com/go-skynet/LocalAI/discussions)
+- Join the Discord channel [Discord](https://discord.gg/uJAeKSAGDy)
+
+---

Dockerfile

@@ -1,11 +1,184 @@
-ARG GO_VERSION=1.20
-ARG BUILD_TYPE=
-FROM golang:$GO_VERSION
-ENV REBUILD=true
+ARG GO_VERSION=1.21-bullseye
+ARG IMAGE_TYPE=extras
+# extras or core
+
+
+FROM golang:$GO_VERSION as requirements-core
+
+ARG BUILD_TYPE
+ARG CUDA_MAJOR_VERSION=11
+ARG CUDA_MINOR_VERSION=7
+ARG SPDLOG_VERSION="1.11.0"
+ARG PIPER_PHONEMIZE_VERSION='1.0.0'
+ARG TARGETARCH
+ARG TARGETVARIANT
+
+ENV BUILD_TYPE=${BUILD_TYPE}
+ENV EXTERNAL_GRPC_BACKENDS="huggingface-embeddings:/build/extra/grpc/huggingface/huggingface.py,autogptq:/build/extra/grpc/autogptq/autogptq.py,bark:/build/extra/grpc/bark/ttsbark.py,diffusers:/build/extra/grpc/diffusers/backend_diffusers.py,exllama:/build/extra/grpc/exllama/exllama.py,vall-e-x:/build/extra/grpc/vall-e-x/ttsvalle.py,vllm:/build/extra/grpc/vllm/backend_vllm.py"
+ENV GALLERIES='[{"name":"model-gallery", "url":"github:go-skynet/model-gallery/index.yaml"}, {"url": "github:go-skynet/model-gallery/huggingface.yaml","name":"huggingface"}]'
+ARG GO_TAGS="stablediffusion tts"
+
+RUN apt-get update && \
+    apt-get install -y ca-certificates curl patch pip cmake && apt-get clean
+
+
+# Use the variables in subsequent instructions
+RUN echo "Target Architecture: $TARGETARCH"
+RUN echo "Target Variant: $TARGETVARIANT"
+
+# CuBLAS requirements
+RUN if [ "${BUILD_TYPE}" = "cublas" ]; then \
+    apt-get install -y software-properties-common && \
+    apt-add-repository contrib && \
+    curl -O https://developer.download.nvidia.com/compute/cuda/repos/debian11/x86_64/cuda-keyring_1.0-1_all.deb && \
+    dpkg -i cuda-keyring_1.0-1_all.deb && \
+    rm -f cuda-keyring_1.0-1_all.deb && \
+    apt-get update && \
+    apt-get install -y cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusparse-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusolver-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}  && apt-get clean \
+    ; fi
+ENV PATH /usr/local/cuda/bin:${PATH}
+
+# OpenBLAS requirements and stable diffusion
+RUN apt-get install -y \
+    libopenblas-dev \
+    libopencv-dev \ 
+    && apt-get clean
+
+# Set up OpenCV
+RUN ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
+
 WORKDIR /build
-RUN apt-get update && apt-get install -y cmake libgomp1 libopenblas-dev libopenblas-base libopencv-dev libopencv-core-dev libopencv-core4.5 ca-certificates
+
+# piper requirements
+# Use pre-compiled Piper phonemization library (includes onnxruntime)
+#RUN if echo "${GO_TAGS}" | grep -q "tts"; then \
+RUN test -n "$TARGETARCH" \
+    || (echo 'warn: missing $TARGETARCH, either set this `ARG` manually, or run using `docker buildkit`')
+
+RUN curl -L "https://github.com/gabime/spdlog/archive/refs/tags/v${SPDLOG_VERSION}.tar.gz" | \
+    tar -xzvf - && \
+    mkdir -p "spdlog-${SPDLOG_VERSION}/build" && \
+    cd "spdlog-${SPDLOG_VERSION}/build" && \
+    cmake ..  && \
+    make -j8 && \
+    cmake --install . --prefix /usr && mkdir -p "lib/Linux-$(uname -m)" && \
+    cd /build && \
+    mkdir -p "lib/Linux-$(uname -m)/piper_phonemize" && \
+    curl -L "https://github.com/rhasspy/piper-phonemize/releases/download/v${PIPER_PHONEMIZE_VERSION}/libpiper_phonemize-${TARGETARCH:-$(go env GOARCH)}${TARGETVARIANT}.tar.gz" | \
+    tar -C "lib/Linux-$(uname -m)/piper_phonemize" -xzvf - && ls -liah /build/lib/Linux-$(uname -m)/piper_phonemize/ && \
+    cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/lib/. /usr/lib/ && \
+    ln -s /usr/lib/libpiper_phonemize.so /usr/lib/libpiper_phonemize.so.1 && \
+    cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/include/. /usr/include/ && \
+    rm spdlog-${SPDLOG_VERSION} -rf
+
+# Extras requirements
+FROM requirements-core as requirements-extras
+
+COPY extra/requirements.txt /build/extra/requirements.txt
+ENV PATH="/root/.cargo/bin:${PATH}"
+RUN pip install --upgrade pip
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+RUN if [ "${TARGETARCH}" = "amd64" ]; then \
+        pip install git+https://github.com/suno-ai/bark.git diffusers invisible_watermark transformers accelerate safetensors;\
+    fi
+RUN if [ "${BUILD_TYPE}" = "cublas" ] && [ "${TARGETARCH}" = "amd64" ]; then \
+        pip install torch vllm && pip install auto-gptq https://github.com/jllllll/exllama/releases/download/0.0.10/exllama-0.0.10+cu${CUDA_MAJOR_VERSION}${CUDA_MINOR_VERSION}-cp39-cp39-linux_x86_64.whl;\
+    fi
+RUN pip install -r /build/extra/requirements.txt && rm -rf /build/extra/requirements.txt
+
+# Vall-e-X
+RUN git clone https://github.com/Plachtaa/VALL-E-X.git /usr/lib/vall-e-x && cd /usr/lib/vall-e-x && pip install -r requirements.txt
+
+# \
+#    ; fi
+
+###################################
+###################################
+
+FROM requirements-${IMAGE_TYPE} as builder
+
+ARG GO_TAGS="stablediffusion tts"
+ARG GRPC_BACKENDS
+ARG BUILD_GRPC=true
+ENV GRPC_BACKENDS=${GRPC_BACKENDS}
+ENV GO_TAGS=${GO_TAGS}
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_REQUIRE_CUDA="cuda>=${CUDA_MAJOR_VERSION}.0"
+ENV NVIDIA_VISIBLE_DEVICES=all
+
+WORKDIR /build
+
+COPY Makefile .
+RUN make get-sources
+COPY go.mod .
+RUN make prepare
 COPY . .
-RUN ln -s /usr/include/opencv4/opencv2/ /usr/include/opencv2
-RUN make build
+COPY .git .
+
+# stablediffusion does not tolerate a newer version of abseil, build it first
+RUN GRPC_BACKENDS=backend-assets/grpc/stablediffusion make build
+
+RUN if [ "${BUILD_GRPC}" = "true" ]; then \
+    git clone --recurse-submodules -b v1.58.0 --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+    cd grpc && mkdir -p cmake/build && cd cmake/build && cmake -DgRPC_INSTALL=ON \
+      -DgRPC_BUILD_TESTS=OFF \
+       ../.. && make -j12 install && rm -rf grpc \
+    ; fi
+
+# Rebuild with defaults backends
+RUN ESPEAK_DATA=/build/lib/Linux-$(uname -m)/piper_phonemize/lib/espeak-ng-data make build
+
+###################################
+###################################
+
+FROM requirements-${IMAGE_TYPE}
+
+ARG FFMPEG
+ARG BUILD_TYPE
+ARG TARGETARCH
+
+ENV BUILD_TYPE=${BUILD_TYPE}
+ENV REBUILD=false
+ENV HEALTHCHECK_ENDPOINT=http://localhost:8080/readyz
+
+ARG CUDA_MAJOR_VERSION=11
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_REQUIRE_CUDA="cuda>=${CUDA_MAJOR_VERSION}.0"
+ENV NVIDIA_VISIBLE_DEVICES=all
+
+# Add FFmpeg
+RUN if [ "${FFMPEG}" = "true" ]; then \
+    apt-get install -y ffmpeg \
+    ; fi
+
+WORKDIR /build
+
+# we start fresh & re-copy all assets because `make build` does not clean up nicely after itself
+# so when `entrypoint.sh` runs `make build` again (which it does by default), the build would fail
+# see https://github.com/go-skynet/LocalAI/pull/658#discussion_r1241971626 and
+# https://github.com/go-skynet/LocalAI/pull/434
+COPY . .
+RUN make prepare-sources
+
+# Copy the binary
+COPY --from=builder /build/local-ai ./
+
+# do not let stablediffusion rebuild (requires an older version of absl)
+COPY --from=builder /build/backend-assets/grpc/stablediffusion ./backend-assets/grpc/stablediffusion
+
+# Copy VALLE-X as it's not a real "lib"
+RUN if [ -d /usr/lib/vall-e-x ]; then \
+    cp -rfv /usr/lib/vall-e-x/* ./ ; \ 
+    fi
+
+# we also copy exllama libs over to resolve exllama import error
+RUN if [ -d /usr/local/lib/python3.9/dist-packages/exllama ]; then \
+        cp -rfv /usr/local/lib/python3.9/dist-packages/exllama extra/grpc/exllama/;\
+    fi
+
+# Define the health check command
+HEALTHCHECK --interval=1m --timeout=10m --retries=10 \
+  CMD curl -f $HEALTHCHECK_ENDPOINT || exit 1
+
 EXPOSE 8080
 ENTRYPOINT [ "/build/entrypoint.sh" ]

Dockerfile.dev

@@ -1,16 +0,0 @@
-ARG GO_VERSION=1.20
-ARG DEBIAN_VERSION=11
-ARG BUILD_TYPE=
-
-FROM golang:$GO_VERSION as builder
-WORKDIR /build
-RUN apt-get update && apt-get install -y cmake libgomp1 libopenblas-dev libopenblas-base libopencv-dev libopencv-core-dev libopencv-core4.5 
-RUN ln -s /usr/include/opencv4/opencv2/ /usr/include/opencv2
-COPY . .
-RUN make build
-
-FROM debian:$DEBIAN_VERSION
-COPY --from=builder /build/local-ai /usr/bin/local-ai
-RUN apt-get update && apt-get install -y ca-certificates
-EXPOSE 8080
-ENTRYPOINT [ "/usr/bin/local-ai" ]

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023 go-skynet authors
+Copyright (c) 2023 Ettore Di Giacinto
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -3,23 +3,62 @@ GOTEST=$(GOCMD) test
 GOVET=$(GOCMD) vet
 BINARY_NAME=local-ai
 
-GOLLAMA_VERSION?=ccf23adfb278c0165d388389a5d60f3fe38e4854
+# llama.cpp versions
+GOLLAMA_VERSION?=aeba71ee842819da681ea537e78846dc75949ac0
+
+GOLLAMA_STABLE_VERSION?=50cee7712066d9e38306eccadcfbb44ea87df4b7
+
+CPPLLAMA_VERSION?=96981f37b1e3f450d9e63e571514217bf60f0a7f
+
+# gpt4all version
 GPT4ALL_REPO?=https://github.com/nomic-ai/gpt4all
-GPT4ALL_VERSION?=8119ff4df0a99bde44255db2b8c7290b5582ac2b
-GOGPT2_VERSION?=7bff56f0224502c1c9ed6258d2a17e8084628827
+GPT4ALL_VERSION?=27a8b020c36b0df8f8b82a252d261cda47cf44b8
+
+# go-ggml-transformers version
+GOGGMLTRANSFORMERS_VERSION?=ffb09d7dd71e2cbc6c5d7d05357d230eea6f369a
+
+# go-rwkv version
 RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
-RWKV_VERSION?=07166da10cb2a9e8854395a4f210464dcea76e47
-WHISPER_CPP_VERSION?=041be06d5881d3c759cc4ed45d655804361237cd
-BERT_VERSION?=cea1ed76a7f48ef386a8e369f6c82c48cdf2d551
-BLOOMZ_VERSION?=e9366e82abdfe70565644fbfae9651976714efd1
-BUILD_TYPE?=
+RWKV_VERSION?=c898cd0f62df8f2a7830e53d1d513bef4f6f792b
+
+# whisper.cpp version
+WHISPER_CPP_VERSION?=85ed71aaec8e0612a84c0b67804bde75aa75a273
+
+# bert.cpp version
+BERT_VERSION?=6abe312cded14042f6b7c3cd8edf082713334a4d
+
+# go-piper version
+PIPER_VERSION?=56b8a81b4760a6fbee1a82e62f007ae7e8f010a7
+
+# go-bloomz version
+BLOOMZ_VERSION?=1834e77b83faafe912ad4092ccf7f77937349e2f
+
+# stablediffusion version
+STABLEDIFFUSION_VERSION?=d89260f598afb809279bc72aa0107b4292587632
+
+# Go-ggllm
+GOGGLLM_VERSION?=862477d16eefb0805261c19c9b0d053e3b2b684b
+
+export BUILD_TYPE?=
+export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
+export CMAKE_ARGS?=
 CGO_LDFLAGS?=
 CUDA_LIBPATH?=/usr/local/cuda/lib64/
-STABLEDIFFUSION_VERSION?=c0748eca3642d58bcf9521108bcee46959c647dc
 GO_TAGS?=
 BUILD_ID?=git
 
+TEST_DIR=/tmp/test
+
+RANDOM := $(shell bash -c 'echo $$RANDOM')
+
+VERSION?=$(shell git describe --always --tags || echo "dev" )
+# go tool nm ./local-ai | grep Commit
+LD_FLAGS?=
+override LD_FLAGS += -X "github.com/go-skynet/LocalAI/internal.Version=$(VERSION)"
+override LD_FLAGS += -X "github.com/go-skynet/LocalAI/internal.Commit=$(shell git rev-parse HEAD)"
+
 OPTIONAL_TARGETS?=
+ESPEAK_DATA?=
 
 OS := $(shell uname -s)
 ARCH := $(shell uname -m)
@@ -29,8 +68,20 @@ WHITE  := $(shell tput -Txterm setaf 7)
 CYAN   := $(shell tput -Txterm setaf 6)
 RESET  := $(shell tput -Txterm sgr0)
 
-C_INCLUDE_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-stable-diffusion/:$(shell pwd)/gpt4all/gpt4all-bindings/golang/:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert:$(shell pwd)/bloomz
-LIBRARY_PATH=$(shell pwd)/go-llama:$(shell pwd)/go-stable-diffusion/:$(shell pwd)/gpt4all/gpt4all-bindings/golang/:$(shell pwd)/go-gpt2:$(shell pwd)/go-rwkv:$(shell pwd)/whisper.cpp:$(shell pwd)/go-bert:$(shell pwd)/bloomz
+# Default Docker bridge IP
+E2E_BRIDGE_IP?=172.17.0.1
+
+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifeq ($(UNAME_S),Darwin)
+	CGO_LDFLAGS += -lcblas -framework Accelerate
+ifneq ($(BUILD_TYPE),metal)
+    # explicit disable metal if on Darwin and metal is disabled
+	CMAKE_ARGS+=-DLLAMA_METAL=OFF
+endif
+endif
 
 ifeq ($(BUILD_TYPE),openblas)
 	CGO_LDFLAGS+=-lopenblas
@@ -41,8 +92,49 @@ ifeq ($(BUILD_TYPE),cublas)
 	export LLAMA_CUBLAS=1
 endif
 
-ifeq ($(GO_TAGS),stablediffusion)
-	OPTIONAL_TARGETS+=go-stable-diffusion/libstablediffusion.a
+ifeq ($(BUILD_TYPE),hipblas)
+	ROCM_HOME ?= /opt/rocm
+	export CXX=$(ROCM_HOME)/llvm/bin/clang++
+	export CC=$(ROCM_HOME)/llvm/bin/clang
+	# Llama-stable has no hipblas support, so override it here.
+	export STABLE_BUILD_TYPE=
+	GPU_TARGETS ?= gfx900,gfx90a,gfx1030,gfx1031,gfx1100
+	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
+	CMAKE_ARGS+=-DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
+	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link
+endif
+
+ifeq ($(BUILD_TYPE),metal)
+	CGO_LDFLAGS+=-framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	export LLAMA_METAL=1
+endif
+
+ifeq ($(BUILD_TYPE),clblas)
+	CGO_LDFLAGS+=-lOpenCL -lclblast
+endif
+
+# glibc-static or glibc-devel-static required
+ifeq ($(STATIC),true)
+	LD_FLAGS=-linkmode external -extldflags -static
+endif
+
+ifeq ($(findstring stablediffusion,$(GO_TAGS)),stablediffusion)
+#	OPTIONAL_TARGETS+=go-stable-diffusion/libstablediffusion.a
+	OPTIONAL_GRPC+=backend-assets/grpc/stablediffusion
+endif
+
+ifeq ($(findstring tts,$(GO_TAGS)),tts)
+#	OPTIONAL_TARGETS+=go-piper/libpiper_binding.a
+#	OPTIONAL_TARGETS+=backend-assets/espeak-ng-data
+	OPTIONAL_GRPC+=backend-assets/grpc/piper
+endif
+
+ALL_GRPC_BACKENDS=backend-assets/grpc/langchain-huggingface backend-assets/grpc/falcon-ggml backend-assets/grpc/bert-embeddings backend-assets/grpc/falcon backend-assets/grpc/bloomz backend-assets/grpc/llama backend-assets/grpc/llama-cpp backend-assets/grpc/llama-stable backend-assets/grpc/gpt4all backend-assets/grpc/dolly backend-assets/grpc/gpt2 backend-assets/grpc/gptj backend-assets/grpc/gptneox backend-assets/grpc/mpt backend-assets/grpc/replit backend-assets/grpc/starcoder backend-assets/grpc/rwkv backend-assets/grpc/whisper $(OPTIONAL_GRPC)
+GRPC_BACKENDS?=$(ALL_GRPC_BACKENDS) $(OPTIONAL_GRPC)
+
+# If empty, then we build all
+ifeq ($(GRPC_BACKENDS),)
+	GRPC_BACKENDS=$(ALL_GRPC_BACKENDS)
 endif
 
 .PHONY: all test build vendor
@@ -53,31 +145,24 @@ all: help
 gpt4all:
 	git clone --recurse-submodules $(GPT4ALL_REPO) gpt4all
 	cd gpt4all && git checkout -b build $(GPT4ALL_VERSION) && git submodule update --init --recursive --depth 1
-	# This is hackish, but needed as both go-llama and go-gpt4allj have their own version of ggml..
-	@find ./gpt4all -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_/gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.h" -exec sed -i'' -e 's/gpt_/gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.h" -exec sed -i'' -e 's/set_console_color/set_gptj_console_color/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/set_console_color/set_gptj_console_color/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/llama_/gptjllama_/g' {} +
-	@find ./gpt4all -type f -name "*.go" -exec sed -i'' -e 's/llama_/gptjllama_/g' {} +
-	@find ./gpt4all -type f -name "*.h" -exec sed -i'' -e 's/llama_/gptjllama_/g' {} +
-	@find ./gpt4all -type f -name "*.txt" -exec sed -i'' -e 's/llama_/gptjllama_/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/json_/json_gptj_/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/void replace/void json_gptj_replace/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/::replace/::json_gptj_replace/g' {} +
-	@find ./gpt4all -type f -name "*.cpp" -exec sed -i'' -e 's/regex_escape/gpt4allregex_escape/g' {} +
-	mv ./gpt4all/gpt4all-backend/llama.cpp/llama_util.h ./gpt4all/gpt4all-backend/llama.cpp/gptjllama_util.h
+
+## go-ggllm
+go-ggllm:
+	git clone --recurse-submodules https://github.com/mudler/go-ggllm.cpp go-ggllm
+	cd go-ggllm && git checkout -b build $(GOGGLLM_VERSION) && git submodule update --init --recursive --depth 1
+
+go-ggllm/libggllm.a: go-ggllm
+	$(MAKE) -C go-ggllm BUILD_TYPE=$(BUILD_TYPE) libggllm.a
+
+## go-piper
+go-piper:
+	git clone --recurse-submodules https://github.com/mudler/go-piper go-piper
+	cd go-piper && git checkout -b build $(PIPER_VERSION) && git submodule update --init --recursive --depth 1
 
 ## BERT embeddings
 go-bert:
 	git clone --recurse-submodules https://github.com/go-skynet/go-bert.cpp go-bert
 	cd go-bert && git checkout -b build $(BERT_VERSION) && git submodule update --init --recursive --depth 1
-	@find ./go-bert -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
-	@find ./go-bert -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
-	@find ./go-bert -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_bert_/g' {} +
 
 ## stable diffusion
 go-stable-diffusion:
@@ -91,21 +176,14 @@ go-stable-diffusion/libstablediffusion.a:
 go-rwkv:
 	git clone --recurse-submodules $(RWKV_REPO) go-rwkv
 	cd go-rwkv && git checkout -b build $(RWKV_VERSION) && git submodule update --init --recursive --depth 1
-	@find ./go-rwkv -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_rwkv_/g' {} +
-	@find ./go-rwkv -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_rwkv_/g' {} +
-	@find ./go-rwkv -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_rwkv_/g' {} +
 
 go-rwkv/librwkv.a: go-rwkv
-	cd go-rwkv && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a .. && cp ggml/src/libggml.a ..
+	cd go-rwkv && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
 
 ## bloomz
 bloomz:
 	git clone --recurse-submodules https://github.com/go-skynet/bloomz.cpp bloomz
-	@find ./bloomz -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
-	@find ./bloomz -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
-	@find ./bloomz -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_bloomz_/g' {} +
-	@find ./bloomz -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_/gpt_bloomz_/g' {} +
-	@find ./bloomz -type f -name "*.h" -exec sed -i'' -e 's/gpt_/gpt_bloomz_/g' {} +
+	cd bloomz && git checkout -b build $(BLOOMZ_VERSION) && git submodule update --init --recursive --depth 1
 
 bloomz/libbloomz.a: bloomz
 	cd bloomz && make libbloomz.a
@@ -113,34 +191,35 @@ bloomz/libbloomz.a: bloomz
 go-bert/libgobert.a: go-bert
 	$(MAKE) -C go-bert libgobert.a
 
+backend-assets/gpt4all: gpt4all/gpt4all-bindings/golang/libgpt4all.a
+	mkdir -p backend-assets/gpt4all
+	@cp gpt4all/gpt4all-bindings/golang/buildllm/*.so backend-assets/gpt4all/ || true
+	@cp gpt4all/gpt4all-bindings/golang/buildllm/*.dylib backend-assets/gpt4all/ || true
+	@cp gpt4all/gpt4all-bindings/golang/buildllm/*.dll backend-assets/gpt4all/ || true
+
+backend-assets/espeak-ng-data:
+	mkdir -p backend-assets/espeak-ng-data
+ifdef ESPEAK_DATA
+	@cp -rf $(ESPEAK_DATA)/. backend-assets/espeak-ng-data
+else
+	@echo "ESPEAK_DATA not set, skipping tts. Note that this will break the tts functionality."
+	@touch backend-assets/espeak-ng-data/keep
+endif
+
 gpt4all/gpt4all-bindings/golang/libgpt4all.a: gpt4all
 	$(MAKE) -C gpt4all/gpt4all-bindings/golang/ libgpt4all.a
 
 ## CEREBRAS GPT
-go-gpt2: 
-	git clone --recurse-submodules https://github.com/go-skynet/go-gpt2.cpp go-gpt2
-	cd go-gpt2 && git checkout -b build $(GOGPT2_VERSION) && git submodule update --init --recursive --depth 1
-	# This is hackish, but needed as both go-llama and go-gpt4allj have their own version of ggml..
-	@find ./go-gpt2 -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_gpt2_/g' {} +
-	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_gpt2_/g' {} +
-	@find ./go-gpt2 -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_gpt2_/g' {} +
-	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_print_usage/gpt2_print_usage/g' {} +
-	@find ./go-gpt2 -type f -name "*.h" -exec sed -i'' -e 's/gpt_print_usage/gpt2_print_usage/g' {} +
-	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_params_parse/gpt2_params_parse/g' {} +
-	@find ./go-gpt2 -type f -name "*.h" -exec sed -i'' -e 's/gpt_params_parse/gpt2_params_parse/g' {} +
-	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/gpt_random_prompt/gpt2_random_prompt/g' {} +
-	@find ./go-gpt2 -type f -name "*.h" -exec sed -i'' -e 's/gpt_random_prompt/gpt2_random_prompt/g' {} +
-	@find ./go-gpt2 -type f -name "*.cpp" -exec sed -i'' -e 's/json_/json_gpt2_/g' {} +
+go-ggml-transformers:
+	git clone --recurse-submodules https://github.com/go-skynet/go-ggml-transformers.cpp go-ggml-transformers
+	cd go-ggml-transformers && git checkout -b build $(GOGPT2_VERSION) && git submodule update --init --recursive --depth 1
 
-go-gpt2/libgpt2.a: go-gpt2
-	$(MAKE) -C go-gpt2 libgpt2.a
+go-ggml-transformers/libtransformers.a: go-ggml-transformers
+	$(MAKE) -C go-ggml-transformers BUILD_TYPE=$(BUILD_TYPE) libtransformers.a
 
 whisper.cpp:
 	git clone https://github.com/ggerganov/whisper.cpp.git
 	cd whisper.cpp && git checkout -b build $(WHISPER_CPP_VERSION) && git submodule update --init --recursive --depth 1
-	@find ./whisper.cpp -type f -name "*.c" -exec sed -i'' -e 's/ggml_/ggml_whisper_/g' {} +
-	@find ./whisper.cpp -type f -name "*.cpp" -exec sed -i'' -e 's/ggml_/ggml_whisper_/g' {} +
-	@find ./whisper.cpp -type f -name "*.h" -exec sed -i'' -e 's/ggml_/ggml_whisper_/g' {} +
 
 whisper.cpp/libwhisper.a: whisper.cpp
 	cd whisper.cpp && make libwhisper.a
@@ -149,81 +228,163 @@ go-llama:
 	git clone --recurse-submodules https://github.com/go-skynet/go-llama.cpp go-llama
 	cd go-llama && git checkout -b build $(GOLLAMA_VERSION) && git submodule update --init --recursive --depth 1
 
-go-llama/libbinding.a: go-llama 
+go-llama-stable:
+	git clone --recurse-submodules https://github.com/go-skynet/go-llama.cpp go-llama-stable
+	cd go-llama-stable && git checkout -b build $(GOLLAMA_STABLE_VERSION) && git submodule update --init --recursive --depth 1
+
+go-llama/libbinding.a: go-llama
 	$(MAKE) -C go-llama BUILD_TYPE=$(BUILD_TYPE) libbinding.a
 
+go-llama-stable/libbinding.a: go-llama-stable
+	$(MAKE) -C go-llama-stable BUILD_TYPE=$(STABLE_BUILD_TYPE) libbinding.a
+
+go-piper/libpiper_binding.a: go-piper
+	$(MAKE) -C go-piper libpiper_binding.a example/main
+
+get-sources: go-llama go-llama-stable go-ggllm go-ggml-transformers gpt4all go-piper go-rwkv whisper.cpp go-bert bloomz go-stable-diffusion
+	touch $@
+
 replace:
-	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(shell pwd)/go-llama
 	$(GOCMD) mod edit -replace github.com/nomic-ai/gpt4all/gpt4all-bindings/golang=$(shell pwd)/gpt4all/gpt4all-bindings/golang
-	$(GOCMD) mod edit -replace github.com/go-skynet/go-gpt2.cpp=$(shell pwd)/go-gpt2
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-ggml-transformers.cpp=$(shell pwd)/go-ggml-transformers
 	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(shell pwd)/go-rwkv
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(shell pwd)/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(shell pwd)/go-bert
 	$(GOCMD) mod edit -replace github.com/go-skynet/bloomz.cpp=$(shell pwd)/bloomz
 	$(GOCMD) mod edit -replace github.com/mudler/go-stable-diffusion=$(shell pwd)/go-stable-diffusion
+	$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(shell pwd)/go-piper
+	$(GOCMD) mod edit -replace github.com/mudler/go-ggllm.cpp=$(shell pwd)/go-ggllm
 
-prepare-sources: go-llama go-gpt2 gpt4all go-rwkv whisper.cpp go-bert bloomz go-stable-diffusion replace
+prepare-sources: get-sources replace
 	$(GOCMD) mod download
 
 ## GENERIC
 rebuild: ## Rebuilds the project
+	$(GOCMD) clean -cache
 	$(MAKE) -C go-llama clean
+	$(MAKE) -C go-llama-stable clean
 	$(MAKE) -C gpt4all/gpt4all-bindings/golang/ clean
-	$(MAKE) -C go-gpt2 clean
+	$(MAKE) -C go-ggml-transformers clean
 	$(MAKE) -C go-rwkv clean
 	$(MAKE) -C whisper.cpp clean
 	$(MAKE) -C go-stable-diffusion clean
 	$(MAKE) -C go-bert clean
 	$(MAKE) -C bloomz clean
+	$(MAKE) -C go-piper clean
+	$(MAKE) -C go-ggllm clean
 	$(MAKE) build
 
-prepare: prepare-sources gpt4all/gpt4all-bindings/golang/libgpt4all.a $(OPTIONAL_TARGETS) go-llama/libbinding.a go-bert/libgobert.a go-gpt2/libgpt2.a go-rwkv/librwkv.a whisper.cpp/libwhisper.a bloomz/libbloomz.a  ## Prepares for building
+prepare: prepare-sources $(OPTIONAL_TARGETS)
+	touch $@
 
 clean: ## Remove build related file
-	rm -fr ./go-llama
+	$(GOCMD) clean -cache
+	rm -f prepare
+	rm -rf ./go-llama
 	rm -rf ./gpt4all
-	rm -rf ./go-stable-diffusion
+	rm -rf ./go-llama-stable
 	rm -rf ./go-gpt2
+	rm -rf ./go-stable-diffusion
+	rm -rf ./go-ggml-transformers
+	rm -rf ./backend-assets
 	rm -rf ./go-rwkv
 	rm -rf ./go-bert
 	rm -rf ./bloomz
 	rm -rf ./whisper.cpp
+	rm -rf ./go-piper
+	rm -rf ./go-ggllm
 	rm -rf $(BINARY_NAME)
 	rm -rf release/
+	$(MAKE) -C backend/cpp/llama clean
 
 ## Build:
 
-build: prepare ## Build the project
+build: grpcs prepare ## 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)${RESET})
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=${C_INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} $(GOCMD) build -tags "$(GO_TAGS)" -x -o $(BINARY_NAME) ./
+	$(info ${GREEN}I LD_FLAGS: ${YELLOW}$(LD_FLAGS)${RESET})
+
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o $(BINARY_NAME) ./
 
 dist: build
 	mkdir -p release
 	cp $(BINARY_NAME) release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-$(ARCH)
 
-generic-build: ## Build the project using generic
-	BUILD_TYPE="generic" $(MAKE) build
-
 ## Run
 run: prepare ## run local-ai
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=${C_INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} $(GOCMD) run ./main.go
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) run ./
 
 test-models/testmodel:
 	mkdir test-models
 	mkdir test-dir
-	wget https://huggingface.co/concedo/cerebras-111M-ggml/resolve/main/cerberas-111m-q4_0.bin -O test-models/testmodel
+	wget https://huggingface.co/nnakasato/ggml-model-test/resolve/main/ggml-model-q4.bin -O test-models/testmodel
 	wget https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
 	wget https://huggingface.co/skeskinen/ggml/resolve/main/all-MiniLM-L6-v2/ggml-model-q4_0.bin -O test-models/bert
 	wget https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
-	wget https://huggingface.co/imxcstar/rwkv-4-raven-ggml/resolve/main/RWKV-4-Raven-1B5-v11-Eng99%25-Other1%25-20230425-ctx4096-16_Q4_2.bin -O test-models/rwkv
+	wget https://huggingface.co/mudler/rwkv-4-raven-1.5B-ggml/resolve/main/RWKV-4-Raven-1B5-v11-Eng99%2525-Other1%2525-20230425-ctx4096_Q4_0.bin -O test-models/rwkv
 	wget https://raw.githubusercontent.com/saharNooby/rwkv.cpp/5eb8f09c146ea8124633ab041d9ea0b1f1db4459/rwkv/20B_tokenizer.json -O test-models/rwkv.tokenizer.json
 	cp tests/models_fixtures/* test-models
 
-test: prepare test-models/testmodel
+prepare-test: grpcs
+	cp -rf backend-assets api
 	cp tests/models_fixtures/* test-models
-	C_INCLUDE_PATH=${C_INCLUDE_PATH} LIBRARY_PATH=${LIBRARY_PATH} TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models $(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo -v -r ./api ./pkg
+
+test: prepare test-models/testmodel grpcs
+	@echo 'Running tests'
+	export GO_TAGS="tts stablediffusion"
+	$(MAKE) prepare-test
+	HUGGINGFACE_GRPC=$(abspath ./)/extra/grpc/huggingface/huggingface.py TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!gpt4all && !llama && !llama-gguf" --flake-attempts 5 -v -r ./api ./pkg
+	$(MAKE) test-gpt4all
+	$(MAKE) test-llama
+	$(MAKE) test-llama-gguf
+	$(MAKE) test-tts
+	$(MAKE) test-stablediffusion
+
+prepare-e2e:
+	mkdir -p $(TEST_DIR)
+	cp -rfv $(abspath ./tests/e2e-fixtures)/gpu.yaml $(TEST_DIR)/gpu.yaml
+	test -e $(TEST_DIR)/ggllm-test-model.bin || wget -q https://huggingface.co/TheBloke/CodeLlama-7B-Instruct-GGUF/resolve/main/codellama-7b-instruct.Q2_K.gguf -O $(TEST_DIR)/ggllm-test-model.bin
+	docker build --build-arg BUILD_GRPC=true --build-arg GRPC_BACKENDS="$(GRPC_BACKENDS)" --build-arg IMAGE_TYPE=core --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg CUDA_MAJOR_VERSION=11 --build-arg CUDA_MINOR_VERSION=7 --build-arg FFMPEG=true -t localai-tests .
+
+run-e2e-image:
+	ls -liah $(abspath ./tests/e2e-fixtures)
+	docker run -p 5390:8080 -e MODELS_PATH=/models -e THREADS=1 -e DEBUG=true -d --rm -v $(TEST_DIR):/models --gpus all --name e2e-tests-$(RANDOM) localai-tests
+
+test-e2e:
+	@echo 'Running e2e tests'
+	BUILD_TYPE=$(BUILD_TYPE) \
+	LOCALAI_API=http://$(E2E_BRIDGE_IP):5390/v1 \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --flake-attempts 5 -v -r ./tests/e2e
+
+teardown-e2e:
+	rm -rf $(TEST_DIR) || true
+	docker stop $$(docker ps -q --filter ancestor=localai-tests)
+
+test-gpt4all: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="gpt4all" --flake-attempts 5 -v -r ./api ./pkg
+
+test-llama: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama" --flake-attempts 5 -v -r ./api ./pkg
+
+test-llama-gguf: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama-gguf" --flake-attempts 5 -v -r ./api ./pkg
+
+test-tts: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="tts" --flake-attempts 1 -v -r ./api ./pkg
+
+test-stablediffusion: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="stablediffusion" --flake-attempts 1 -v -r ./api ./pkg
+
+test-container:
+	docker build --target requirements -t local-ai-test-container .
+	docker run -ti --rm --entrypoint /bin/bash -ti -v $(abspath ./):/build local-ai-test-container
 
 ## Help:
 help: ## Show this help.
@@ -236,3 +397,119 @@ help: ## Show this help.
 		if (/^[a-zA-Z_-]+:.*?##.*$$/) {printf "    ${YELLOW}%-20s${GREEN}%s${RESET}\n", $$1, $$2} \
 		else if (/^## .*$$/) {printf "  ${CYAN}%s${RESET}\n", substr($$1,4)} \
 		}' $(MAKEFILE_LIST)
+
+protogen: protogen-go protogen-python
+
+protogen-go:
+	protoc --go_out=. --go_opt=paths=source_relative --go-grpc_out=. --go-grpc_opt=paths=source_relative \
+    pkg/grpc/proto/backend.proto
+
+protogen-python:
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/huggingface/ --grpc_python_out=extra/grpc/huggingface/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/autogptq/ --grpc_python_out=extra/grpc/autogptq/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/exllama/ --grpc_python_out=extra/grpc/exllama/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/bark/ --grpc_python_out=extra/grpc/bark/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/diffusers/ --grpc_python_out=extra/grpc/diffusers/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/vall-e-x/ --grpc_python_out=extra/grpc/vall-e-x/ pkg/grpc/proto/backend.proto
+	python3 -m grpc_tools.protoc -Ipkg/grpc/proto/ --python_out=extra/grpc/vllm/ --grpc_python_out=extra/grpc/vllm/ pkg/grpc/proto/backend.proto
+
+## GRPC
+
+backend-assets/grpc:
+	mkdir -p backend-assets/grpc
+
+backend-assets/grpc/falcon: backend-assets/grpc go-ggllm/libggllm.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggllm LIBRARY_PATH=$(shell pwd)/go-ggllm \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/falcon ./cmd/grpc/falcon/
+
+backend-assets/grpc/llama: backend-assets/grpc go-llama/libbinding.a
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(shell pwd)/go-llama
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-llama LIBRARY_PATH=$(shell pwd)/go-llama \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama ./cmd/grpc/llama/
+# TODO: every binary should have its own folder instead, so can have different metal implementations
+ifeq ($(BUILD_TYPE),metal)
+	cp go-llama/build/bin/ggml-metal.metal backend-assets/grpc/
+endif
+
+backend/cpp/llama/grpc-server:
+	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/llama grpc-server
+
+backend-assets/grpc/llama-cpp: backend-assets/grpc backend/cpp/llama/grpc-server
+	cp -rfv backend/cpp/llama/grpc-server backend-assets/grpc/llama-cpp
+# TODO: every binary should have its own folder instead, so can have different metal implementations
+ifeq ($(BUILD_TYPE),metal)
+	cp backend/cpp/llama/llama.cpp/build/bin/ggml-metal.metal backend-assets/grpc/
+endif
+
+backend-assets/grpc/llama-stable: backend-assets/grpc go-llama-stable/libbinding.a
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(shell pwd)/go-llama-stable
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-llama-stable LIBRARY_PATH=$(shell pwd)/go-llama \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-stable ./cmd/grpc/llama-stable/
+
+backend-assets/grpc/gpt4all: backend-assets/grpc backend-assets/gpt4all gpt4all/gpt4all-bindings/golang/libgpt4all.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/gpt4all/gpt4all-bindings/golang/ LIBRARY_PATH=$(shell pwd)/gpt4all/gpt4all-bindings/golang/ \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gpt4all ./cmd/grpc/gpt4all/
+
+backend-assets/grpc/dolly: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/dolly ./cmd/grpc/dolly/
+
+backend-assets/grpc/gpt2: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gpt2 ./cmd/grpc/gpt2/
+
+backend-assets/grpc/gptj: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gptj ./cmd/grpc/gptj/
+
+backend-assets/grpc/gptneox: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/gptneox ./cmd/grpc/gptneox/
+
+backend-assets/grpc/mpt: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/mpt ./cmd/grpc/mpt/
+
+backend-assets/grpc/replit: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/replit ./cmd/grpc/replit/
+
+backend-assets/grpc/falcon-ggml: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/falcon-ggml ./cmd/grpc/falcon-ggml/
+
+backend-assets/grpc/starcoder: backend-assets/grpc go-ggml-transformers/libtransformers.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggml-transformers LIBRARY_PATH=$(shell pwd)/go-ggml-transformers \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/starcoder ./cmd/grpc/starcoder/
+
+backend-assets/grpc/rwkv: backend-assets/grpc go-rwkv/librwkv.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-rwkv LIBRARY_PATH=$(shell pwd)/go-rwkv \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/rwkv ./cmd/grpc/rwkv/
+
+backend-assets/grpc/bloomz: backend-assets/grpc bloomz/libbloomz.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/bloomz LIBRARY_PATH=$(shell pwd)/bloomz \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bloomz ./cmd/grpc/bloomz/
+
+backend-assets/grpc/bert-embeddings: backend-assets/grpc go-bert/libgobert.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-bert LIBRARY_PATH=$(shell pwd)/go-bert \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bert-embeddings ./cmd/grpc/bert-embeddings/
+
+backend-assets/grpc/langchain-huggingface: backend-assets/grpc
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/langchain-huggingface ./cmd/grpc/langchain-huggingface/
+
+backend-assets/grpc/stablediffusion: backend-assets/grpc
+	if [ ! -f backend-assets/grpc/stablediffusion ]; then \
+		$(MAKE) go-stable-diffusion/libstablediffusion.a; \
+		CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-stable-diffusion/ LIBRARY_PATH=$(shell pwd)/go-stable-diffusion/ \
+		$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion ./cmd/grpc/stablediffusion/; \
+	fi
+
+backend-assets/grpc/piper: backend-assets/grpc backend-assets/espeak-ng-data go-piper/libpiper_binding.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" LIBRARY_PATH=$(shell pwd)/go-piper \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/piper ./cmd/grpc/piper/
+
+backend-assets/grpc/whisper: backend-assets/grpc whisper.cpp/libwhisper.a
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/whisper.cpp LIBRARY_PATH=$(shell pwd)/whisper.cpp \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./cmd/grpc/whisper/
+
+grpcs: prepare $(GRPC_BACKENDS)

api/api.go

@@ -2,6 +2,17 @@ package api
 
 import (
 	"errors"
+	"fmt"
+	"strings"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/localai"
+	"github.com/go-skynet/LocalAI/api/openai"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	"github.com/go-skynet/LocalAI/internal"
+	"github.com/go-skynet/LocalAI/metrics"
+	"github.com/go-skynet/LocalAI/pkg/assets"
 
 	"github.com/gofiber/fiber/v2"
 	"github.com/gofiber/fiber/v2/middleware/cors"
@@ -11,18 +22,77 @@ import (
 	"github.com/rs/zerolog/log"
 )
 
-func App(opts ...AppOption) *fiber.App {
-	options := newOptions(opts...)
+func Startup(opts ...options.AppOption) (*options.Option, *config.ConfigLoader, error) {
+	options := options.NewOptions(opts...)
 
 	zerolog.SetGlobalLevel(zerolog.InfoLevel)
-	if options.debug {
+	if options.Debug {
 		zerolog.SetGlobalLevel(zerolog.DebugLevel)
 	}
 
+	log.Info().Msgf("Starting LocalAI using %d threads, with models path: %s", options.Threads, options.Loader.ModelPath)
+	log.Info().Msgf("LocalAI version: %s", internal.PrintableVersion())
+
+	cl := config.NewConfigLoader()
+	if err := cl.LoadConfigs(options.Loader.ModelPath); err != nil {
+		log.Error().Msgf("error loading config files: %s", err.Error())
+	}
+
+	if options.ConfigFile != "" {
+		if err := cl.LoadConfigFile(options.ConfigFile); err != nil {
+			log.Error().Msgf("error loading config file: %s", err.Error())
+		}
+	}
+
+	if options.Debug {
+		for _, v := range cl.ListConfigs() {
+			cfg, _ := cl.GetConfig(v)
+			log.Debug().Msgf("Model: %s (config: %+v)", v, cfg)
+		}
+	}
+
+	if options.AssetsDestination != "" {
+		// Extract files from the embedded FS
+		err := assets.ExtractFiles(options.BackendAssets, options.AssetsDestination)
+		log.Debug().Msgf("Extracting backend assets files to %s", options.AssetsDestination)
+		if err != nil {
+			log.Warn().Msgf("Failed extracting backend assets files: %s (might be required for some backends to work properly, like gpt4all)", err)
+		}
+	}
+
+	if options.PreloadJSONModels != "" {
+		if err := localai.ApplyGalleryFromString(options.Loader.ModelPath, options.PreloadJSONModels, cl, options.Galleries); err != nil {
+			return nil, nil, err
+		}
+	}
+
+	if options.PreloadModelsFromPath != "" {
+		if err := localai.ApplyGalleryFromFile(options.Loader.ModelPath, options.PreloadModelsFromPath, cl, options.Galleries); err != nil {
+			return nil, nil, err
+		}
+	}
+
+	// turn off any process that was started by GRPC if the context is canceled
+	go func() {
+		<-options.Context.Done()
+		log.Debug().Msgf("Context canceled, shutting down")
+		options.Loader.StopAllGRPC()
+	}()
+
+	return options, cl, nil
+}
+
+func App(opts ...options.AppOption) (*fiber.App, error) {
+
+	options, cl, err := Startup(opts...)
+	if err != nil {
+		return nil, fmt.Errorf("failed basic startup tasks with error %s", err.Error())
+	}
+
 	// Return errors as JSON responses
 	app := fiber.New(fiber.Config{
-		BodyLimit:             options.uploadLimitMB * 1024 * 1024, // this is the default limit of 4MB
-		DisableStartupMessage: options.disableMessage,
+		BodyLimit:             options.UploadLimitMB * 1024 * 1024, // this is the default limit of 4MB
+		DisableStartupMessage: options.DisableMessage,
 		// Override default error handler
 		ErrorHandler: func(ctx *fiber.Ctx, err error) error {
 			// Status code defaults to 500
@@ -36,87 +106,134 @@ func App(opts ...AppOption) *fiber.App {
 
 			// Send custom error page
 			return ctx.Status(code).JSON(
-				ErrorResponse{
-					Error: &APIError{Message: err.Error(), Code: code},
+				schema.ErrorResponse{
+					Error: &schema.APIError{Message: err.Error(), Code: code},
 				},
 			)
 		},
 	})
 
-	if options.debug {
+	if options.Debug {
 		app.Use(logger.New(logger.Config{
 			Format: "[${ip}]:${port} ${status} - ${method} ${path}\n",
 		}))
 	}
 
-	cm := NewConfigMerger()
-	if err := cm.LoadConfigs(options.loader.ModelPath); err != nil {
-		log.Error().Msgf("error loading config files: %s", err.Error())
-	}
-
-	if options.configFile != "" {
-		if err := cm.LoadConfigFile(options.configFile); err != nil {
-			log.Error().Msgf("error loading config file: %s", err.Error())
-		}
-	}
-
-	if options.debug {
-		for _, v := range cm.ListConfigs() {
-			cfg, _ := cm.GetConfig(v)
-			log.Debug().Msgf("Model: %s (config: %+v)", v, cfg)
-		}
-	}
 	// Default middleware config
 	app.Use(recover.New())
+	if options.Metrics != nil {
+		app.Use(metrics.APIMiddleware(options.Metrics))
+	}
 
-	if options.cors {
-		if options.corsAllowOrigins == "" {
-			app.Use(cors.New())
-		} else {
-			app.Use(cors.New(cors.Config{
-				AllowOrigins: options.corsAllowOrigins,
-			}))
+	// Auth middleware checking if API key is valid. If no API key is set, no auth is required.
+	auth := func(c *fiber.Ctx) error {
+		if len(options.ApiKeys) > 0 {
+			authHeader := c.Get("Authorization")
+			if authHeader == "" {
+				return c.Status(fiber.StatusUnauthorized).JSON(fiber.Map{"message": "Authorization header missing"})
+			}
+			authHeaderParts := strings.Split(authHeader, " ")
+			if len(authHeaderParts) != 2 || authHeaderParts[0] != "Bearer" {
+				return c.Status(fiber.StatusUnauthorized).JSON(fiber.Map{"message": "Invalid Authorization header format"})
+			}
+
+			apiKey := authHeaderParts[1]
+			validApiKey := false
+			for _, key := range options.ApiKeys {
+				if apiKey == key {
+					validApiKey = true
+				}
+			}
+			if !validApiKey {
+				return c.Status(fiber.StatusUnauthorized).JSON(fiber.Map{"message": "Invalid API key"})
+			}
 		}
+		return c.Next()
+	}
+
+	if options.CORS {
+		var c func(ctx *fiber.Ctx) error
+		if options.CORSAllowOrigins == "" {
+			c = cors.New()
+		} else {
+			c = cors.New(cors.Config{AllowOrigins: options.CORSAllowOrigins})
+		}
+
+		app.Use(c)
 	}
 
 	// LocalAI API endpoints
-	applier := newGalleryApplier(options.loader.ModelPath)
-	applier.start(options.context, cm)
-	app.Post("/models/apply", applyModelGallery(options.loader.ModelPath, cm, applier.C))
-	app.Get("/models/jobs/:uuid", getOpStatus(applier))
+	galleryService := localai.NewGalleryService(options.Loader.ModelPath)
+	galleryService.Start(options.Context, cl)
+
+	app.Get("/version", auth, func(c *fiber.Ctx) error {
+		return c.JSON(struct {
+			Version string `json:"version"`
+		}{Version: internal.PrintableVersion()})
+	})
+
+	modelGalleryService := localai.CreateModelGalleryService(options.Galleries, options.Loader.ModelPath, galleryService)
+	app.Post("/models/apply", auth, modelGalleryService.ApplyModelGalleryEndpoint())
+	app.Get("/models/available", auth, modelGalleryService.ListModelFromGalleryEndpoint())
+	app.Get("/models/galleries", auth, modelGalleryService.ListModelGalleriesEndpoint())
+	app.Post("/models/galleries", auth, modelGalleryService.AddModelGalleryEndpoint())
+	app.Delete("/models/galleries", auth, modelGalleryService.RemoveModelGalleryEndpoint())
+	app.Get("/models/jobs/:uuid", auth, modelGalleryService.GetOpStatusEndpoint())
+	app.Get("/models/jobs", auth, modelGalleryService.GetAllStatusEndpoint())
 
 	// openAI compatible API endpoint
 
 	// chat
-	app.Post("/v1/chat/completions", chatEndpoint(cm, options))
-	app.Post("/chat/completions", chatEndpoint(cm, options))
+	app.Post("/v1/chat/completions", auth, openai.ChatEndpoint(cl, options))
+	app.Post("/chat/completions", auth, openai.ChatEndpoint(cl, options))
 
 	// edit
-	app.Post("/v1/edits", editEndpoint(cm, options))
-	app.Post("/edits", editEndpoint(cm, options))
+	app.Post("/v1/edits", auth, openai.EditEndpoint(cl, options))
+	app.Post("/edits", auth, openai.EditEndpoint(cl, options))
 
 	// completion
-	app.Post("/v1/completions", completionEndpoint(cm, options))
-	app.Post("/completions", completionEndpoint(cm, options))
+	app.Post("/v1/completions", auth, openai.CompletionEndpoint(cl, options))
+	app.Post("/completions", auth, openai.CompletionEndpoint(cl, options))
+	app.Post("/v1/engines/:model/completions", auth, openai.CompletionEndpoint(cl, options))
 
 	// embeddings
-	app.Post("/v1/embeddings", embeddingsEndpoint(cm, options))
-	app.Post("/embeddings", embeddingsEndpoint(cm, options))
-	app.Post("/v1/engines/:model/embeddings", embeddingsEndpoint(cm, options))
+	app.Post("/v1/embeddings", auth, openai.EmbeddingsEndpoint(cl, options))
+	app.Post("/embeddings", auth, openai.EmbeddingsEndpoint(cl, options))
+	app.Post("/v1/engines/:model/embeddings", auth, openai.EmbeddingsEndpoint(cl, options))
 
 	// audio
-	app.Post("/v1/audio/transcriptions", transcriptEndpoint(cm, options))
+	app.Post("/v1/audio/transcriptions", auth, openai.TranscriptEndpoint(cl, options))
+	app.Post("/tts", auth, localai.TTSEndpoint(cl, options))
 
 	// images
-	app.Post("/v1/images/generations", imageEndpoint(cm, options))
+	app.Post("/v1/images/generations", auth, openai.ImageEndpoint(cl, options))
 
-	if options.imageDir != "" {
-		app.Static("/generated-images", options.imageDir)
+	if options.ImageDir != "" {
+		app.Static("/generated-images", options.ImageDir)
 	}
 
-	// models
-	app.Get("/v1/models", listModels(options.loader, cm))
-	app.Get("/models", listModels(options.loader, cm))
+	if options.AudioDir != "" {
+		app.Static("/generated-audio", options.AudioDir)
+	}
 
-	return app
+	ok := func(c *fiber.Ctx) error {
+		return c.SendStatus(200)
+	}
+
+	// Kubernetes health checks
+	app.Get("/healthz", ok)
+	app.Get("/readyz", ok)
+
+	// Experimental Backend Statistics Module
+	backendMonitor := localai.NewBackendMonitor(cl, options) // Split out for now
+	app.Get("/backend/monitor", localai.BackendMonitorEndpoint(backendMonitor))
+	app.Post("/backend/shutdown", localai.BackendShutdownEndpoint(backendMonitor))
+
+	// models
+	app.Get("/v1/models", auth, openai.ListModelsEndpoint(options.Loader, cl))
+	app.Get("/models", auth, openai.ListModelsEndpoint(options.Loader, cl))
+
+	app.Get("/metrics", metrics.MetricsHandler())
+
+	return app, nil
 }

api/api_test.go

@@ -3,16 +3,22 @@ package api_test
 import (
 	"bytes"
 	"context"
+	"embed"
 	"encoding/json"
+	"errors"
 	"fmt"
-	"io/ioutil"
+	"io"
 	"net/http"
 	"os"
 	"path/filepath"
 	"runtime"
 
 	. "github.com/go-skynet/LocalAI/api"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/metrics"
+	"github.com/go-skynet/LocalAI/pkg/gallery"
 	"github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/utils"
 	"github.com/gofiber/fiber/v2"
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
@@ -20,12 +26,14 @@ import (
 
 	openaigo "github.com/otiai10/openaigo"
 	"github.com/sashabaranov/go-openai"
+	"github.com/sashabaranov/go-openai/jsonschema"
 )
 
 type modelApplyRequest struct {
-	URL       string            `json:"url"`
-	Name      string            `json:"name"`
-	Overrides map[string]string `json:"overrides"`
+	ID        string                 `json:"id"`
+	URL       string                 `json:"url"`
+	Name      string                 `json:"name"`
+	Overrides map[string]interface{} `json:"overrides"`
 }
 
 func getModelStatus(url string) (response map[string]interface{}) {
@@ -37,7 +45,7 @@ func getModelStatus(url string) (response map[string]interface{}) {
 	}
 	defer resp.Body.Close()
 
-	body, err := ioutil.ReadAll(resp.Body)
+	body, err := io.ReadAll(resp.Body)
 	if err != nil {
 		fmt.Println("Error reading response body:", err)
 		return
@@ -51,6 +59,15 @@ func getModelStatus(url string) (response map[string]interface{}) {
 	}
 	return
 }
+
+func getModels(url string) (response []gallery.GalleryModel) {
+	utils.GetURI(url, func(url string, i []byte) error {
+		// Unmarshal YAML data into a struct
+		return json.Unmarshal(i, &response)
+	})
+	return
+}
+
 func postModelApplyRequest(url string, request modelApplyRequest) (response map[string]interface{}) {
 
 	//url := "http://localhost:AI/models/apply"
@@ -80,7 +97,7 @@ func postModelApplyRequest(url string, request modelApplyRequest) (response map[
 	}
 	defer resp.Body.Close()
 
-	body, err := ioutil.ReadAll(resp.Body)
+	body, err := io.ReadAll(resp.Body)
 	if err != nil {
 		fmt.Println("Error reading response body:", err)
 		return
@@ -95,6 +112,9 @@ func postModelApplyRequest(url string, request modelApplyRequest) (response map[
 	return
 }
 
+//go:embed backend-assets/*
+var backendAssets embed.FS
+
 var _ = Describe("API test", func() {
 
 	var app *fiber.App
@@ -105,6 +125,11 @@ var _ = Describe("API test", func() {
 	var cancel context.CancelFunc
 	var tmpdir string
 
+	commonOpts := []options.AppOption{
+		options.WithDebug(true),
+		options.WithDisableMessage(true),
+	}
+
 	Context("API with ephemeral models", func() {
 		BeforeEach(func() {
 			var err error
@@ -114,7 +139,40 @@ var _ = Describe("API test", func() {
 			modelLoader = model.NewModelLoader(tmpdir)
 			c, cancel = context.WithCancel(context.Background())
 
-			app = App(WithContext(c), WithModelLoader(modelLoader))
+			g := []gallery.GalleryModel{
+				{
+					Name: "bert",
+					URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+				},
+				{
+					Name:            "bert2",
+					URL:             "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+					Overrides:       map[string]interface{}{"foo": "bar"},
+					AdditionalFiles: []gallery.File{{Filename: "foo.yaml", URI: "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml"}},
+				},
+			}
+			out, err := yaml.Marshal(g)
+			Expect(err).ToNot(HaveOccurred())
+			err = os.WriteFile(filepath.Join(tmpdir, "gallery_simple.yaml"), out, 0644)
+			Expect(err).ToNot(HaveOccurred())
+
+			galleries := []gallery.Gallery{
+				{
+					Name: "test",
+					URL:  "file://" + filepath.Join(tmpdir, "gallery_simple.yaml"),
+				},
+			}
+
+			metricsService, err := metrics.SetupMetrics()
+			Expect(err).ToNot(HaveOccurred())
+
+			app, err = App(
+				append(commonOpts,
+					options.WithMetrics(metricsService),
+					options.WithContext(c),
+					options.WithGalleries(galleries),
+					options.WithModelLoader(modelLoader), options.WithBackendAssets(backendAssets), options.WithBackendAssetsOutput(tmpdir))...)
+			Expect(err).ToNot(HaveOccurred())
 			go app.Listen("127.0.0.1:9090")
 
 			defaultConfig := openai.DefaultConfig("")
@@ -138,11 +196,58 @@ var _ = Describe("API test", func() {
 		})
 
 		Context("Applying models", func() {
+			It("applies models from a gallery", func() {
+
+				models := getModels("http://127.0.0.1:9090/models/available")
+				Expect(len(models)).To(Equal(2), fmt.Sprint(models))
+				Expect(models[0].Installed).To(BeFalse(), fmt.Sprint(models))
+				Expect(models[1].Installed).To(BeFalse(), fmt.Sprint(models))
+
+				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+					ID: "test@bert2",
+				})
+
+				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+				uuid := response["uuid"].(string)
+				resp := map[string]interface{}{}
+				Eventually(func() bool {
+					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+					fmt.Println(response)
+					resp = response
+					return response["processed"].(bool)
+				}, "360s", "10s").Should(Equal(true))
+				Expect(resp["message"]).ToNot(ContainSubstring("error"))
+
+				dat, err := os.ReadFile(filepath.Join(tmpdir, "bert2.yaml"))
+				Expect(err).ToNot(HaveOccurred())
+
+				_, err = os.ReadFile(filepath.Join(tmpdir, "foo.yaml"))
+				Expect(err).ToNot(HaveOccurred())
+
+				content := map[string]interface{}{}
+				err = yaml.Unmarshal(dat, &content)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(content["backend"]).To(Equal("bert-embeddings"))
+				Expect(content["foo"]).To(Equal("bar"))
+
+				models = getModels("http://127.0.0.1:9090/models/available")
+				Expect(len(models)).To(Equal(2), fmt.Sprint(models))
+				Expect(models[0].Name).To(Or(Equal("bert"), Equal("bert2")))
+				Expect(models[1].Name).To(Or(Equal("bert"), Equal("bert2")))
+				for _, m := range models {
+					if m.Name == "bert2" {
+						Expect(m.Installed).To(BeTrue())
+					} else {
+						Expect(m.Installed).To(BeFalse())
+					}
+				}
+			})
 			It("overrides models", func() {
 				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
 					URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
 					Name: "bert",
-					Overrides: map[string]string{
+					Overrides: map[string]interface{}{
 						"backend": "llama",
 					},
 				})
@@ -153,9 +258,8 @@ var _ = Describe("API test", func() {
 
 				Eventually(func() bool {
 					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
-					fmt.Println(response)
 					return response["processed"].(bool)
-				}, "360s").Should(Equal(true))
+				}, "360s", "10s").Should(Equal(true))
 
 				dat, err := os.ReadFile(filepath.Join(tmpdir, "bert.yaml"))
 				Expect(err).ToNot(HaveOccurred())
@@ -169,7 +273,7 @@ var _ = Describe("API test", func() {
 				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
 					URL:       "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
 					Name:      "bert",
-					Overrides: map[string]string{},
+					Overrides: map[string]interface{}{},
 				})
 
 				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
@@ -178,9 +282,8 @@ var _ = Describe("API test", func() {
 
 				Eventually(func() bool {
 					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
-					fmt.Println(response)
 					return response["processed"].(bool)
-				}, "360s").Should(Equal(true))
+				}, "360s", "10s").Should(Equal(true))
 
 				dat, err := os.ReadFile(filepath.Join(tmpdir, "bert.yaml"))
 				Expect(err).ToNot(HaveOccurred())
@@ -190,6 +293,297 @@ var _ = Describe("API test", func() {
 				Expect(err).ToNot(HaveOccurred())
 				Expect(content["backend"]).To(Equal("bert-embeddings"))
 			})
+
+			It("runs openllama", Label("llama"), func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+					URL:       "github:go-skynet/model-gallery/openllama_3b.yaml",
+					Name:      "openllama_3b",
+					Overrides: map[string]interface{}{"backend": "llama-stable", "mmap": true, "f16": true, "context_size": 128},
+				})
+
+				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+				uuid := response["uuid"].(string)
+
+				Eventually(func() bool {
+					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+					return response["processed"].(bool)
+				}, "360s", "10s").Should(Equal(true))
+
+				By("testing completion")
+				resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "openllama_3b", Prompt: "Count up to five: one, two, three, four, "})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Choices)).To(Equal(1))
+				Expect(resp.Choices[0].Text).To(ContainSubstring("five"))
+
+				By("testing functions")
+				resp2, err := client.CreateChatCompletion(
+					context.TODO(),
+					openai.ChatCompletionRequest{
+						Model: "openllama_3b",
+						Messages: []openai.ChatCompletionMessage{
+							{
+								Role:    "user",
+								Content: "What is the weather like in San Francisco (celsius)?",
+							},
+						},
+						Functions: []openai.FunctionDefinition{
+							openai.FunctionDefinition{
+								Name:        "get_current_weather",
+								Description: "Get the current weather",
+								Parameters: jsonschema.Definition{
+									Type: jsonschema.Object,
+									Properties: map[string]jsonschema.Definition{
+										"location": {
+											Type:        jsonschema.String,
+											Description: "The city and state, e.g. San Francisco, CA",
+										},
+										"unit": {
+											Type: jsonschema.String,
+											Enum: []string{"celcius", "fahrenheit"},
+										},
+									},
+									Required: []string{"location"},
+								},
+							},
+						},
+					})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp2.Choices)).To(Equal(1))
+				Expect(resp2.Choices[0].Message.FunctionCall).ToNot(BeNil())
+				Expect(resp2.Choices[0].Message.FunctionCall.Name).To(Equal("get_current_weather"), resp2.Choices[0].Message.FunctionCall.Name)
+
+				var res map[string]string
+				err = json.Unmarshal([]byte(resp2.Choices[0].Message.FunctionCall.Arguments), &res)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(res["location"]).To(Equal("San Francisco, California, United States"), fmt.Sprint(res))
+				Expect(res["unit"]).To(Equal("celcius"), fmt.Sprint(res))
+				Expect(string(resp2.Choices[0].FinishReason)).To(Equal("function_call"), fmt.Sprint(resp2.Choices[0].FinishReason))
+			})
+
+			It("runs openllama gguf", Label("llama-gguf"), func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+				modelName := "codellama"
+				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+					URL:       "github:go-skynet/model-gallery/codellama-7b-instruct.yaml",
+					Name:      modelName,
+					Overrides: map[string]interface{}{"backend": "llama", "mmap": true, "f16": true, "context_size": 128},
+				})
+
+				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+				uuid := response["uuid"].(string)
+
+				Eventually(func() bool {
+					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+					return response["processed"].(bool)
+				}, "360s", "10s").Should(Equal(true))
+
+				By("testing chat")
+				resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: modelName, Messages: []openai.ChatCompletionMessage{
+					{
+						Role:    "user",
+						Content: "How much is 2+2?",
+					},
+				}})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Choices)).To(Equal(1))
+				Expect(resp.Choices[0].Message.Content).To(Or(ContainSubstring("4"), ContainSubstring("four")))
+
+				By("testing functions")
+				resp2, err := client.CreateChatCompletion(
+					context.TODO(),
+					openai.ChatCompletionRequest{
+						Model: modelName,
+						Messages: []openai.ChatCompletionMessage{
+							{
+								Role:    "user",
+								Content: "What is the weather like in San Francisco (celsius)?",
+							},
+						},
+						Functions: []openai.FunctionDefinition{
+							openai.FunctionDefinition{
+								Name:        "get_current_weather",
+								Description: "Get the current weather",
+								Parameters: jsonschema.Definition{
+									Type: jsonschema.Object,
+									Properties: map[string]jsonschema.Definition{
+										"location": {
+											Type:        jsonschema.String,
+											Description: "The city and state, e.g. San Francisco, CA",
+										},
+										"unit": {
+											Type: jsonschema.String,
+											Enum: []string{"celcius", "fahrenheit"},
+										},
+									},
+									Required: []string{"location"},
+								},
+							},
+						},
+					})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp2.Choices)).To(Equal(1))
+				Expect(resp2.Choices[0].Message.FunctionCall).ToNot(BeNil())
+				Expect(resp2.Choices[0].Message.FunctionCall.Name).To(Equal("get_current_weather"), resp2.Choices[0].Message.FunctionCall.Name)
+
+				var res map[string]string
+				err = json.Unmarshal([]byte(resp2.Choices[0].Message.FunctionCall.Arguments), &res)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(res["location"]).To(Equal("San Francisco"), fmt.Sprint(res))
+				Expect(res["unit"]).To(Equal("celcius"), fmt.Sprint(res))
+				Expect(string(resp2.Choices[0].FinishReason)).To(Equal("function_call"), fmt.Sprint(resp2.Choices[0].FinishReason))
+			})
+
+			It("runs gpt4all", Label("gpt4all"), func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+
+				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+					URL:  "github:go-skynet/model-gallery/gpt4all-j.yaml",
+					Name: "gpt4all-j",
+				})
+
+				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+				uuid := response["uuid"].(string)
+
+				Eventually(func() bool {
+					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+					return response["processed"].(bool)
+				}, "360s", "10s").Should(Equal(true))
+
+				resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "gpt4all-j", Messages: []openai.ChatCompletionMessage{openai.ChatCompletionMessage{Role: "user", Content: "How are you?"}}})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Choices)).To(Equal(1))
+				Expect(resp.Choices[0].Message.Content).To(ContainSubstring("well"))
+			})
+
+		})
+	})
+
+	Context("Model gallery", func() {
+		BeforeEach(func() {
+			var err error
+			tmpdir, err = os.MkdirTemp("", "")
+			Expect(err).ToNot(HaveOccurred())
+
+			modelLoader = model.NewModelLoader(tmpdir)
+			c, cancel = context.WithCancel(context.Background())
+
+			galleries := []gallery.Gallery{
+				{
+					Name: "model-gallery",
+					URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/index.yaml",
+				},
+			}
+
+			metricsService, err := metrics.SetupMetrics()
+			Expect(err).ToNot(HaveOccurred())
+
+			app, err = App(
+				append(commonOpts,
+					options.WithContext(c),
+					options.WithMetrics(metricsService),
+					options.WithAudioDir(tmpdir),
+					options.WithImageDir(tmpdir),
+					options.WithGalleries(galleries),
+					options.WithModelLoader(modelLoader),
+					options.WithBackendAssets(backendAssets),
+					options.WithBackendAssetsOutput(tmpdir))...,
+			)
+			Expect(err).ToNot(HaveOccurred())
+			go app.Listen("127.0.0.1:9090")
+
+			defaultConfig := openai.DefaultConfig("")
+			defaultConfig.BaseURL = "http://127.0.0.1:9090/v1"
+
+			client2 = openaigo.NewClient("")
+			client2.BaseURL = defaultConfig.BaseURL
+
+			// Wait for API to be ready
+			client = openai.NewClientWithConfig(defaultConfig)
+			Eventually(func() error {
+				_, err := client.ListModels(context.TODO())
+				return err
+			}, "2m").ShouldNot(HaveOccurred())
+		})
+
+		AfterEach(func() {
+			cancel()
+			app.Shutdown()
+			os.RemoveAll(tmpdir)
+		})
+		It("installs and is capable to run tts", Label("tts"), func() {
+			if runtime.GOOS != "linux" {
+				Skip("test supported only on linux")
+			}
+
+			response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+				ID: "model-gallery@voice-en-us-kathleen-low",
+			})
+
+			Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+			uuid := response["uuid"].(string)
+
+			Eventually(func() bool {
+				response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+				fmt.Println(response)
+				return response["processed"].(bool)
+			}, "360s", "10s").Should(Equal(true))
+
+			// An HTTP Post to the /tts endpoint should return a wav audio file
+			resp, err := http.Post("http://127.0.0.1:9090/tts", "application/json", bytes.NewBuffer([]byte(`{"input": "Hello world", "model": "en-us-kathleen-low.onnx"}`)))
+			Expect(err).ToNot(HaveOccurred(), fmt.Sprint(resp))
+			dat, err := io.ReadAll(resp.Body)
+			Expect(err).ToNot(HaveOccurred(), fmt.Sprint(resp))
+
+			Expect(resp.StatusCode).To(Equal(200), fmt.Sprint(string(dat)))
+			Expect(resp.Header.Get("Content-Type")).To(Equal("audio/x-wav"))
+		})
+		It("installs and is capable to generate images", Label("stablediffusion"), func() {
+			if runtime.GOOS != "linux" {
+				Skip("test supported only on linux")
+			}
+
+			response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+				ID: "model-gallery@stablediffusion",
+				Overrides: map[string]interface{}{
+					"parameters": map[string]interface{}{"model": "stablediffusion_assets"},
+				},
+			})
+
+			Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+			uuid := response["uuid"].(string)
+
+			Eventually(func() bool {
+				response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+				fmt.Println(response)
+				return response["processed"].(bool)
+			}, "360s", "10s").Should(Equal(true))
+
+			resp, err := http.Post(
+				"http://127.0.0.1:9090/v1/images/generations",
+				"application/json",
+				bytes.NewBuffer([]byte(`{
+					 			"prompt": "floating hair, portrait, ((loli)), ((one girl)), cute face, hidden hands, asymmetrical bangs, beautiful detailed eyes, eye shadow, hair ornament, ribbons, bowties, buttons, pleated skirt, (((masterpiece))), ((best quality)), colorful|((part of the head)), ((((mutated hands and fingers)))), deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, Octane renderer, lowres, bad anatomy, bad hands, text",
+								"mode": 2,  "seed":9000,
+					 			"size": "256x256", "n":2}`)))
+			// The response should contain an URL
+			Expect(err).ToNot(HaveOccurred(), fmt.Sprint(resp))
+			dat, err := io.ReadAll(resp.Body)
+			Expect(err).ToNot(HaveOccurred(), string(dat))
+			Expect(string(dat)).To(ContainSubstring("http://127.0.0.1:9090/"), string(dat))
+			Expect(string(dat)).To(ContainSubstring(".png"), string(dat))
+
 		})
 	})
 
@@ -198,7 +592,17 @@ var _ = Describe("API test", func() {
 			modelLoader = model.NewModelLoader(os.Getenv("MODELS_PATH"))
 			c, cancel = context.WithCancel(context.Background())
 
-			app = App(WithContext(c), WithModelLoader(modelLoader))
+			metricsService, err := metrics.SetupMetrics()
+			Expect(err).ToNot(HaveOccurred())
+
+			app, err = App(
+				append(commonOpts,
+					options.WithExternalBackend("huggingface", os.Getenv("HUGGINGFACE_GRPC")),
+					options.WithContext(c),
+					options.WithModelLoader(modelLoader),
+					options.WithMetrics(metricsService),
+				)...)
+			Expect(err).ToNot(HaveOccurred())
 			go app.Listen("127.0.0.1:9090")
 
 			defaultConfig := openai.DefaultConfig("")
@@ -221,7 +625,7 @@ var _ = Describe("API test", func() {
 		It("returns the models list", func() {
 			models, err := client.ListModels(context.TODO())
 			Expect(err).ToNot(HaveOccurred())
-			Expect(len(models.Models)).To(Equal(10))
+			Expect(len(models.Models)).To(Equal(6)) // If "config.yaml" should be included, this should be 8?
 		})
 		It("can generate completions", func() {
 			resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "testmodel", Prompt: "abcdedfghikl"})
@@ -252,9 +656,10 @@ var _ = Describe("API test", func() {
 		})
 
 		It("returns errors", func() {
+			backends := len(model.AutoLoadBackends) + 1 // +1 for huggingface
 			_, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "foomodel", Prompt: "abcdedfghikl"})
 			Expect(err).To(HaveOccurred())
-			Expect(err.Error()).To(ContainSubstring("error, status code: 500, message: could not load model - all backends returned error: 12 errors occurred:"))
+			Expect(err.Error()).To(ContainSubstring(fmt.Sprintf("error, status code: 500, message: could not load model - all backends returned error: %d errors occurred:", backends)))
 		})
 		It("transcribes audio", func() {
 			if runtime.GOOS != "linux" {
@@ -298,15 +703,98 @@ var _ = Describe("API test", func() {
 			Expect(resp2.Data[0].Embedding).To(Equal(sunEmbedding))
 		})
 
+		Context("External gRPC calls", func() {
+			It("calculate embeddings with huggingface", func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+				resp, err := client.CreateEmbeddings(
+					context.Background(),
+					openai.EmbeddingRequest{
+						Model: openai.AdaCodeSearchCode,
+						Input: []string{"sun", "cat"},
+					},
+				)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Data[0].Embedding)).To(BeNumerically("==", 384))
+				Expect(len(resp.Data[1].Embedding)).To(BeNumerically("==", 384))
+
+				sunEmbedding := resp.Data[0].Embedding
+				resp2, err := client.CreateEmbeddings(
+					context.Background(),
+					openai.EmbeddingRequest{
+						Model: openai.AdaCodeSearchCode,
+						Input: []string{"sun"},
+					},
+				)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(resp2.Data[0].Embedding).To(Equal(sunEmbedding))
+				Expect(resp2.Data[0].Embedding).ToNot(Equal(resp.Data[1].Embedding))
+			})
+		})
+
 		Context("backends", func() {
-			It("runs rwkv", func() {
+			It("runs rwkv completion", func() {
 				if runtime.GOOS != "linux" {
 					Skip("test supported only on linux")
 				}
 				resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "rwkv_test", Prompt: "Count up to five: one, two, three, four,"})
 				Expect(err).ToNot(HaveOccurred())
 				Expect(len(resp.Choices) > 0).To(BeTrue())
-				Expect(resp.Choices[0].Text).To(Equal(" five."))
+				Expect(resp.Choices[0].Text).To(ContainSubstring("five"))
+
+				stream, err := client.CreateCompletionStream(context.TODO(), openai.CompletionRequest{
+					Model: "rwkv_test", Prompt: "Count up to five: one, two, three, four,", Stream: true,
+				})
+				Expect(err).ToNot(HaveOccurred())
+				defer stream.Close()
+
+				tokens := 0
+				text := ""
+				for {
+					response, err := stream.Recv()
+					if errors.Is(err, io.EOF) {
+						break
+					}
+
+					Expect(err).ToNot(HaveOccurred())
+					text += response.Choices[0].Text
+					tokens++
+				}
+				Expect(text).ToNot(BeEmpty())
+				Expect(text).To(ContainSubstring("five"))
+				Expect(tokens).ToNot(Or(Equal(1), Equal(0)))
+			})
+			It("runs rwkv chat completion", func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+				resp, err := client.CreateChatCompletion(context.TODO(),
+					openai.ChatCompletionRequest{Model: "rwkv_test", Messages: []openai.ChatCompletionMessage{{Content: "Can you count up to five?", Role: "user"}}})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Choices) > 0).To(BeTrue())
+				Expect(resp.Choices[0].Message.Content).To(Or(ContainSubstring("Sure"), ContainSubstring("five")))
+
+				stream, err := client.CreateChatCompletionStream(context.TODO(), openai.ChatCompletionRequest{Model: "rwkv_test", Messages: []openai.ChatCompletionMessage{{Content: "Can you count up to five?", Role: "user"}}})
+				Expect(err).ToNot(HaveOccurred())
+				defer stream.Close()
+
+				tokens := 0
+				text := ""
+				for {
+					response, err := stream.Recv()
+					if errors.Is(err, io.EOF) {
+						break
+					}
+
+					Expect(err).ToNot(HaveOccurred())
+					text += response.Choices[0].Delta.Content
+					tokens++
+				}
+				Expect(text).ToNot(BeEmpty())
+				Expect(text).To(Or(ContainSubstring("Sure"), ContainSubstring("five")))
+
+				Expect(tokens).ToNot(Or(Equal(1), Equal(0)))
 			})
 		})
 	})
@@ -316,7 +804,17 @@ var _ = Describe("API test", func() {
 			modelLoader = model.NewModelLoader(os.Getenv("MODELS_PATH"))
 			c, cancel = context.WithCancel(context.Background())
 
-			app = App(WithContext(c), WithModelLoader(modelLoader), WithConfigFile(os.Getenv("CONFIG_FILE")))
+			metricsService, err := metrics.SetupMetrics()
+			Expect(err).ToNot(HaveOccurred())
+
+			app, err = App(
+				append(commonOpts,
+					options.WithContext(c),
+					options.WithMetrics(metricsService),
+					options.WithModelLoader(modelLoader),
+					options.WithConfigFile(os.Getenv("CONFIG_FILE")))...,
+			)
+			Expect(err).ToNot(HaveOccurred())
 			go app.Listen("127.0.0.1:9090")
 
 			defaultConfig := openai.DefaultConfig("")
@@ -334,19 +832,14 @@ var _ = Describe("API test", func() {
 			cancel()
 			app.Shutdown()
 		})
-		It("can generate chat completions from config file", func() {
-			models, err := client.ListModels(context.TODO())
-			Expect(err).ToNot(HaveOccurred())
-			Expect(len(models.Models)).To(Equal(12))
-		})
-		It("can generate chat completions from config file", func() {
-			resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "list1", Messages: []openai.ChatCompletionMessage{openai.ChatCompletionMessage{Role: "user", Content: "abcdedfghikl"}}})
+		It("can generate chat completions from config file (list1)", func() {
+			resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "list1", Messages: []openai.ChatCompletionMessage{{Role: "user", Content: "abcdedfghikl"}}})
 			Expect(err).ToNot(HaveOccurred())
 			Expect(len(resp.Choices)).To(Equal(1))
 			Expect(resp.Choices[0].Message.Content).ToNot(BeEmpty())
 		})
-		It("can generate chat completions from config file", func() {
-			resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "list2", Messages: []openai.ChatCompletionMessage{openai.ChatCompletionMessage{Role: "user", Content: "abcdedfghikl"}}})
+		It("can generate chat completions from config file (list2)", func() {
+			resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "list2", Messages: []openai.ChatCompletionMessage{{Role: "user", Content: "abcdedfghikl"}}})
 			Expect(err).ToNot(HaveOccurred())
 			Expect(len(resp.Choices)).To(Equal(1))
 			Expect(resp.Choices[0].Message.Content).ToNot(BeEmpty())

api/backend/embeddings.go

@@ -0,0 +1,92 @@
+package backend
+
+import (
+	"fmt"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/grpc"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+)
+
+func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, c config.Config, o *options.Option) (func() ([]float32, error), error) {
+	if !c.Embeddings {
+		return nil, fmt.Errorf("endpoint disabled for this model by API configuration")
+	}
+
+	modelFile := c.Model
+
+	grpcOpts := gRPCModelOpts(c)
+
+	var inferenceModel interface{}
+	var err error
+
+	opts := modelOpts(c, o, []model.Option{
+		model.WithLoadGRPCLoadModelOpts(grpcOpts),
+		model.WithThreads(uint32(c.Threads)),
+		model.WithAssetDir(o.AssetsDestination),
+		model.WithModel(modelFile),
+		model.WithContext(o.Context),
+	})
+
+	if c.Backend == "" {
+		inferenceModel, err = loader.GreedyLoader(opts...)
+	} else {
+		opts = append(opts, model.WithBackendString(c.Backend))
+		inferenceModel, err = loader.BackendLoader(opts...)
+	}
+	if err != nil {
+		return nil, err
+	}
+
+	var fn func() ([]float32, error)
+	switch model := inferenceModel.(type) {
+	case *grpc.Client:
+		fn = func() ([]float32, error) {
+			predictOptions := gRPCPredictOpts(c, loader.ModelPath)
+			if len(tokens) > 0 {
+				embeds := []int32{}
+
+				for _, t := range tokens {
+					embeds = append(embeds, int32(t))
+				}
+				predictOptions.EmbeddingTokens = embeds
+
+				res, err := model.Embeddings(o.Context, predictOptions)
+				if err != nil {
+					return nil, err
+				}
+
+				return res.Embeddings, nil
+			}
+			predictOptions.Embeddings = s
+
+			res, err := model.Embeddings(o.Context, predictOptions)
+			if err != nil {
+				return nil, err
+			}
+
+			return res.Embeddings, nil
+		}
+	default:
+		fn = func() ([]float32, error) {
+			return nil, fmt.Errorf("embeddings not supported by the backend")
+		}
+	}
+
+	return func() ([]float32, error) {
+		embeds, err := fn()
+		if err != nil {
+			return embeds, err
+		}
+		// Remove trailing 0s
+		for i := len(embeds) - 1; i >= 0; i-- {
+			if embeds[i] == 0.0 {
+				embeds = embeds[:i]
+			} else {
+				break
+			}
+		}
+		return embeds, nil
+	}, nil
+}

api/backend/image.go

@@ -0,0 +1,59 @@
+package backend
+
+import (
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+)
+
+func ImageGeneration(height, width, mode, step, seed int, positive_prompt, negative_prompt, src, dst string, loader *model.ModelLoader, c config.Config, o *options.Option) (func() error, error) {
+
+	opts := modelOpts(c, o, []model.Option{
+		model.WithBackendString(c.Backend),
+		model.WithAssetDir(o.AssetsDestination),
+		model.WithThreads(uint32(c.Threads)),
+		model.WithContext(o.Context),
+		model.WithModel(c.Model),
+		model.WithLoadGRPCLoadModelOpts(&proto.ModelOptions{
+			CUDA:          c.Diffusers.CUDA,
+			SchedulerType: c.Diffusers.SchedulerType,
+			PipelineType:  c.Diffusers.PipelineType,
+			CFGScale:      c.Diffusers.CFGScale,
+			LoraAdapter:   c.LoraAdapter,
+			LoraBase:      c.LoraBase,
+			IMG2IMG:       c.Diffusers.IMG2IMG,
+			CLIPModel:     c.Diffusers.ClipModel,
+			CLIPSubfolder: c.Diffusers.ClipSubFolder,
+			CLIPSkip:      int32(c.Diffusers.ClipSkip),
+		}),
+	})
+
+	inferenceModel, err := loader.BackendLoader(
+		opts...,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	fn := func() error {
+		_, err := inferenceModel.GenerateImage(
+			o.Context,
+			&proto.GenerateImageRequest{
+				Height:           int32(height),
+				Width:            int32(width),
+				Mode:             int32(mode),
+				Step:             int32(step),
+				Seed:             int32(seed),
+				CLIPSkip:         int32(c.Diffusers.ClipSkip),
+				PositivePrompt:   positive_prompt,
+				NegativePrompt:   negative_prompt,
+				Dst:              dst,
+				Src:              src,
+				EnableParameters: c.Diffusers.EnableParameters,
+			})
+		return err
+	}
+
+	return fn, nil
+}

api/backend/llm.go

@@ -0,0 +1,163 @@
+package backend
+
+import (
+	"context"
+	"os"
+	"regexp"
+	"strings"
+	"sync"
+	"unicode/utf8"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/gallery"
+	"github.com/go-skynet/LocalAI/pkg/grpc"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/utils"
+)
+
+type LLMResponse struct {
+	Response string // should this be []byte?
+	Usage    TokenUsage
+}
+
+type TokenUsage struct {
+	Prompt     int
+	Completion int
+}
+
+func ModelInference(ctx context.Context, s string, loader *model.ModelLoader, c config.Config, o *options.Option, tokenCallback func(string, TokenUsage) bool) (func() (LLMResponse, error), error) {
+	modelFile := c.Model
+
+	grpcOpts := gRPCModelOpts(c)
+
+	var inferenceModel *grpc.Client
+	var err error
+
+	opts := modelOpts(c, o, []model.Option{
+		model.WithLoadGRPCLoadModelOpts(grpcOpts),
+		model.WithThreads(uint32(c.Threads)), // some models uses this to allocate threads during startup
+		model.WithAssetDir(o.AssetsDestination),
+		model.WithModel(modelFile),
+		model.WithContext(o.Context),
+	})
+
+	if c.Backend != "" {
+		opts = append(opts, model.WithBackendString(c.Backend))
+	}
+
+	// Check if the modelFile exists, if it doesn't try to load it from the gallery
+	if o.AutoloadGalleries { // experimental
+		if _, err := os.Stat(modelFile); os.IsNotExist(err) {
+			utils.ResetDownloadTimers()
+			// if we failed to load the model, we try to download it
+			err := gallery.InstallModelFromGalleryByName(o.Galleries, modelFile, loader.ModelPath, gallery.GalleryModel{}, utils.DisplayDownloadFunction)
+			if err != nil {
+				return nil, err
+			}
+		}
+	}
+
+	if c.Backend == "" {
+		inferenceModel, err = loader.GreedyLoader(opts...)
+	} else {
+		inferenceModel, err = loader.BackendLoader(opts...)
+	}
+
+	if err != nil {
+		return nil, err
+	}
+
+	// in GRPC, the backend is supposed to answer to 1 single token if stream is not supported
+	fn := func() (LLMResponse, error) {
+		opts := gRPCPredictOpts(c, loader.ModelPath)
+		opts.Prompt = s
+
+		tokenUsage := TokenUsage{}
+
+		// check the per-model feature flag for usage, since tokenCallback may have a cost.
+		// Defaults to off as for now it is still experimental
+		if c.FeatureFlag.Enabled("usage") {
+			userTokenCallback := tokenCallback
+			if userTokenCallback == nil {
+				userTokenCallback = func(token string, usage TokenUsage) bool {
+					return true
+				}
+			}
+
+			promptInfo, pErr := inferenceModel.TokenizeString(ctx, opts)
+			if pErr == nil && promptInfo.Length > 0 {
+				tokenUsage.Prompt = int(promptInfo.Length)
+			}
+
+			tokenCallback = func(token string, usage TokenUsage) bool {
+				tokenUsage.Completion++
+				return userTokenCallback(token, tokenUsage)
+			}
+		}
+
+		if tokenCallback != nil {
+			ss := ""
+
+			var partialRune []byte
+			err := inferenceModel.PredictStream(ctx, opts, func(chars []byte) {
+				partialRune = append(partialRune, chars...)
+
+				for len(partialRune) > 0 {
+					r, size := utf8.DecodeRune(partialRune)
+					if r == utf8.RuneError {
+						// incomplete rune, wait for more bytes
+						break
+					}
+
+					tokenCallback(string(r), tokenUsage)
+					ss += string(r)
+
+					partialRune = partialRune[size:]
+				}
+			})
+			return LLMResponse{
+				Response: ss,
+				Usage:    tokenUsage,
+			}, err
+		} else {
+			// TODO: Is the chicken bit the only way to get here? is that acceptable?
+			reply, err := inferenceModel.Predict(ctx, opts)
+			if err != nil {
+				return LLMResponse{}, err
+			}
+			return LLMResponse{
+				Response: string(reply.Message),
+				Usage:    tokenUsage,
+			}, err
+		}
+	}
+
+	return fn, nil
+}
+
+var cutstrings map[string]*regexp.Regexp = make(map[string]*regexp.Regexp)
+var mu sync.Mutex = sync.Mutex{}
+
+func Finetune(config config.Config, input, prediction string) string {
+	if config.Echo {
+		prediction = input + prediction
+	}
+
+	for _, c := range config.Cutstrings {
+		mu.Lock()
+		reg, ok := cutstrings[c]
+		if !ok {
+			cutstrings[c] = regexp.MustCompile(c)
+			reg = cutstrings[c]
+		}
+		mu.Unlock()
+		prediction = reg.ReplaceAllString(prediction, "")
+	}
+
+	for _, c := range config.TrimSpace {
+		prediction = strings.TrimSpace(strings.TrimPrefix(prediction, c))
+	}
+	return prediction
+
+}

api/backend/options.go

@@ -0,0 +1,116 @@
+package backend
+
+import (
+	"os"
+	"path/filepath"
+
+	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+)
+
+func modelOpts(c config.Config, o *options.Option, opts []model.Option) []model.Option {
+	if o.SingleBackend {
+		opts = append(opts, model.WithSingleActiveBackend())
+	}
+
+	if c.GRPC.Attempts != 0 {
+		opts = append(opts, model.WithGRPCAttempts(c.GRPC.Attempts))
+	}
+
+	if c.GRPC.AttemptsSleepTime != 0 {
+		opts = append(opts, model.WithGRPCAttemptsDelay(c.GRPC.AttemptsSleepTime))
+	}
+
+	for k, v := range o.ExternalGRPCBackends {
+		opts = append(opts, model.WithExternalBackend(k, v))
+	}
+
+	return opts
+}
+
+func gRPCModelOpts(c config.Config) *pb.ModelOptions {
+	b := 512
+	if c.Batch != 0 {
+		b = c.Batch
+	}
+
+	return &pb.ModelOptions{
+		ContextSize:   int32(c.ContextSize),
+		Seed:          int32(c.Seed),
+		NBatch:        int32(b),
+		NoMulMatQ:     c.NoMulMatQ,
+		DraftModel:    c.DraftModel,
+		AudioPath:     c.VallE.AudioPath,
+		Quantization:  c.Quantization,
+		LoraAdapter:   c.LoraAdapter,
+		LoraBase:      c.LoraBase,
+		NGQA:          c.NGQA,
+		RMSNormEps:    c.RMSNormEps,
+		F16Memory:     c.F16,
+		MLock:         c.MMlock,
+		RopeFreqBase:  c.RopeFreqBase,
+		RopeFreqScale: c.RopeFreqScale,
+		NUMA:          c.NUMA,
+		Embeddings:    c.Embeddings,
+		LowVRAM:       c.LowVRAM,
+		NGPULayers:    int32(c.NGPULayers),
+		MMap:          c.MMap,
+		MainGPU:       c.MainGPU,
+		Threads:       int32(c.Threads),
+		TensorSplit:   c.TensorSplit,
+		// AutoGPTQ
+		ModelBaseName:    c.AutoGPTQ.ModelBaseName,
+		Device:           c.AutoGPTQ.Device,
+		UseTriton:        c.AutoGPTQ.Triton,
+		UseFastTokenizer: c.AutoGPTQ.UseFastTokenizer,
+		// RWKV
+		Tokenizer: c.Tokenizer,
+	}
+}
+
+func gRPCPredictOpts(c config.Config, modelPath string) *pb.PredictOptions {
+	promptCachePath := ""
+	if c.PromptCachePath != "" {
+		p := filepath.Join(modelPath, c.PromptCachePath)
+		os.MkdirAll(filepath.Dir(p), 0755)
+		promptCachePath = p
+	}
+	return &pb.PredictOptions{
+		Temperature:         float32(c.Temperature),
+		TopP:                float32(c.TopP),
+		NDraft:              c.NDraft,
+		TopK:                int32(c.TopK),
+		Tokens:              int32(c.Maxtokens),
+		Threads:             int32(c.Threads),
+		PromptCacheAll:      c.PromptCacheAll,
+		PromptCacheRO:       c.PromptCacheRO,
+		PromptCachePath:     promptCachePath,
+		F16KV:               c.F16,
+		DebugMode:           c.Debug,
+		Grammar:             c.Grammar,
+		NegativePromptScale: c.NegativePromptScale,
+		RopeFreqBase:        c.RopeFreqBase,
+		RopeFreqScale:       c.RopeFreqScale,
+		NegativePrompt:      c.NegativePrompt,
+		Mirostat:            int32(c.LLMConfig.Mirostat),
+		MirostatETA:         float32(c.LLMConfig.MirostatETA),
+		MirostatTAU:         float32(c.LLMConfig.MirostatTAU),
+		Debug:               c.Debug,
+		StopPrompts:         c.StopWords,
+		Repeat:              int32(c.RepeatPenalty),
+		NKeep:               int32(c.Keep),
+		Batch:               int32(c.Batch),
+		IgnoreEOS:           c.IgnoreEOS,
+		Seed:                int32(c.Seed),
+		FrequencyPenalty:    float32(c.FrequencyPenalty),
+		MLock:               c.MMlock,
+		MMap:                c.MMap,
+		MainGPU:             c.MainGPU,
+		TensorSplit:         c.TensorSplit,
+		TailFreeSamplingZ:   float32(c.TFZ),
+		TypicalP:            float32(c.TypicalP),
+	}
+}

api/backend/transcript.go

@@ -0,0 +1,39 @@
+package backend
+
+import (
+	"context"
+	"fmt"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/schema"
+
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+)
+
+func ModelTranscription(audio, language string, loader *model.ModelLoader, c config.Config, o *options.Option) (*schema.Result, error) {
+
+	opts := modelOpts(c, o, []model.Option{
+		model.WithBackendString(model.WhisperBackend),
+		model.WithModel(c.Model),
+		model.WithContext(o.Context),
+		model.WithThreads(uint32(c.Threads)),
+		model.WithAssetDir(o.AssetsDestination),
+	})
+
+	whisperModel, err := o.Loader.BackendLoader(opts...)
+	if err != nil {
+		return nil, err
+	}
+
+	if whisperModel == nil {
+		return nil, fmt.Errorf("could not load whisper model")
+	}
+
+	return whisperModel.AudioTranscription(context.Background(), &proto.TranscriptRequest{
+		Dst:      audio,
+		Language: language,
+		Threads:  uint32(c.Threads),
+	})
+}

api/backend/tts.go

@@ -0,0 +1,75 @@
+package backend
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+
+	api_config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/grpc/proto"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/utils"
+)
+
+func generateUniqueFileName(dir, baseName, ext string) string {
+	counter := 1
+	fileName := baseName + ext
+
+	for {
+		filePath := filepath.Join(dir, fileName)
+		_, err := os.Stat(filePath)
+		if os.IsNotExist(err) {
+			return fileName
+		}
+
+		counter++
+		fileName = fmt.Sprintf("%s_%d%s", baseName, counter, ext)
+	}
+}
+
+func ModelTTS(backend, text, modelFile string, loader *model.ModelLoader, o *options.Option) (string, *proto.Result, error) {
+	bb := backend
+	if bb == "" {
+		bb = model.PiperBackend
+	}
+	opts := modelOpts(api_config.Config{}, o, []model.Option{
+		model.WithBackendString(bb),
+		model.WithModel(modelFile),
+		model.WithContext(o.Context),
+		model.WithAssetDir(o.AssetsDestination),
+	})
+	piperModel, err := o.Loader.BackendLoader(opts...)
+	if err != nil {
+		return "", nil, err
+	}
+
+	if piperModel == nil {
+		return "", nil, fmt.Errorf("could not load piper model")
+	}
+
+	if err := os.MkdirAll(o.AudioDir, 0755); err != nil {
+		return "", nil, fmt.Errorf("failed creating audio directory: %s", err)
+	}
+
+	fileName := generateUniqueFileName(o.AudioDir, "piper", ".wav")
+	filePath := filepath.Join(o.AudioDir, fileName)
+
+	// If the model file is not empty, we pass it joined with the model path
+	modelPath := ""
+	if modelFile != "" {
+		modelPath = filepath.Join(o.Loader.ModelPath, modelFile)
+		if err := utils.VerifyPath(modelPath, o.Loader.ModelPath); err != nil {
+			return "", nil, err
+		}
+	}
+
+	res, err := piperModel.TTS(context.Background(), &proto.TTSRequest{
+		Text:  text,
+		Model: modelPath,
+		Dst:   filePath,
+	})
+
+	return filePath, res, err
+}

api/config.go

@@ -1,329 +0,0 @@
-package api
-
-import (
-	"encoding/json"
-	"fmt"
-	"io/ioutil"
-	"os"
-	"path/filepath"
-	"strings"
-	"sync"
-
-	model "github.com/go-skynet/LocalAI/pkg/model"
-	"github.com/gofiber/fiber/v2"
-	"github.com/rs/zerolog/log"
-	"gopkg.in/yaml.v3"
-)
-
-type Config struct {
-	OpenAIRequest               `yaml:"parameters"`
-	Name                        string            `yaml:"name"`
-	StopWords                   []string          `yaml:"stopwords"`
-	Cutstrings                  []string          `yaml:"cutstrings"`
-	TrimSpace                   []string          `yaml:"trimspace"`
-	ContextSize                 int               `yaml:"context_size"`
-	F16                         bool              `yaml:"f16"`
-	Threads                     int               `yaml:"threads"`
-	Debug                       bool              `yaml:"debug"`
-	Roles                       map[string]string `yaml:"roles"`
-	Embeddings                  bool              `yaml:"embeddings"`
-	Backend                     string            `yaml:"backend"`
-	TemplateConfig              TemplateConfig    `yaml:"template"`
-	MirostatETA                 float64           `yaml:"mirostat_eta"`
-	MirostatTAU                 float64           `yaml:"mirostat_tau"`
-	Mirostat                    int               `yaml:"mirostat"`
-	NGPULayers                  int               `yaml:"gpu_layers"`
-	ImageGenerationAssets       string            `yaml:"asset_dir"`
-	PromptStrings, InputStrings []string
-	InputToken                  [][]int
-}
-
-type TemplateConfig struct {
-	Completion string `yaml:"completion"`
-	Chat       string `yaml:"chat"`
-	Edit       string `yaml:"edit"`
-}
-
-type ConfigMerger struct {
-	configs map[string]Config
-	sync.Mutex
-}
-
-func NewConfigMerger() *ConfigMerger {
-	return &ConfigMerger{
-		configs: make(map[string]Config),
-	}
-}
-func ReadConfigFile(file string) ([]*Config, error) {
-	c := &[]*Config{}
-	f, err := os.ReadFile(file)
-	if err != nil {
-		return nil, fmt.Errorf("cannot read config file: %w", err)
-	}
-	if err := yaml.Unmarshal(f, c); err != nil {
-		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
-	}
-
-	return *c, nil
-}
-
-func ReadConfig(file string) (*Config, error) {
-	c := &Config{}
-	f, err := os.ReadFile(file)
-	if err != nil {
-		return nil, fmt.Errorf("cannot read config file: %w", err)
-	}
-	if err := yaml.Unmarshal(f, c); err != nil {
-		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
-	}
-
-	return c, nil
-}
-
-func (cm ConfigMerger) LoadConfigFile(file string) error {
-	cm.Lock()
-	defer cm.Unlock()
-	c, err := ReadConfigFile(file)
-	if err != nil {
-		return fmt.Errorf("cannot load config file: %w", err)
-	}
-
-	for _, cc := range c {
-		cm.configs[cc.Name] = *cc
-	}
-	return nil
-}
-
-func (cm ConfigMerger) LoadConfig(file string) error {
-	cm.Lock()
-	defer cm.Unlock()
-	c, err := ReadConfig(file)
-	if err != nil {
-		return fmt.Errorf("cannot read config file: %w", err)
-	}
-
-	cm.configs[c.Name] = *c
-	return nil
-}
-
-func (cm ConfigMerger) GetConfig(m string) (Config, bool) {
-	cm.Lock()
-	defer cm.Unlock()
-	v, exists := cm.configs[m]
-	return v, exists
-}
-
-func (cm ConfigMerger) ListConfigs() []string {
-	cm.Lock()
-	defer cm.Unlock()
-	var res []string
-	for k := range cm.configs {
-		res = append(res, k)
-	}
-	return res
-}
-
-func (cm ConfigMerger) LoadConfigs(path string) error {
-	cm.Lock()
-	defer cm.Unlock()
-	files, err := ioutil.ReadDir(path)
-	if err != nil {
-		return err
-	}
-
-	for _, file := range files {
-		// Skip templates, YAML and .keep files
-		if !strings.Contains(file.Name(), ".yaml") {
-			continue
-		}
-		c, err := ReadConfig(filepath.Join(path, file.Name()))
-		if err == nil {
-			cm.configs[c.Name] = *c
-		}
-	}
-
-	return nil
-}
-
-func updateConfig(config *Config, input *OpenAIRequest) {
-	if input.Echo {
-		config.Echo = input.Echo
-	}
-	if input.TopK != 0 {
-		config.TopK = input.TopK
-	}
-	if input.TopP != 0 {
-		config.TopP = input.TopP
-	}
-
-	if input.Temperature != 0 {
-		config.Temperature = input.Temperature
-	}
-
-	if input.Maxtokens != 0 {
-		config.Maxtokens = input.Maxtokens
-	}
-
-	switch stop := input.Stop.(type) {
-	case string:
-		if stop != "" {
-			config.StopWords = append(config.StopWords, stop)
-		}
-	case []interface{}:
-		for _, pp := range stop {
-			if s, ok := pp.(string); ok {
-				config.StopWords = append(config.StopWords, s)
-			}
-		}
-	}
-
-	if input.RepeatPenalty != 0 {
-		config.RepeatPenalty = input.RepeatPenalty
-	}
-
-	if input.Keep != 0 {
-		config.Keep = input.Keep
-	}
-
-	if input.Batch != 0 {
-		config.Batch = input.Batch
-	}
-
-	if input.F16 {
-		config.F16 = input.F16
-	}
-
-	if input.IgnoreEOS {
-		config.IgnoreEOS = input.IgnoreEOS
-	}
-
-	if input.Seed != 0 {
-		config.Seed = input.Seed
-	}
-
-	if input.Mirostat != 0 {
-		config.Mirostat = input.Mirostat
-	}
-
-	if input.MirostatETA != 0 {
-		config.MirostatETA = input.MirostatETA
-	}
-
-	if input.MirostatTAU != 0 {
-		config.MirostatTAU = input.MirostatTAU
-	}
-
-	switch inputs := input.Input.(type) {
-	case string:
-		if inputs != "" {
-			config.InputStrings = append(config.InputStrings, inputs)
-		}
-	case []interface{}:
-		for _, pp := range inputs {
-			switch i := pp.(type) {
-			case string:
-				config.InputStrings = append(config.InputStrings, i)
-			case []interface{}:
-				tokens := []int{}
-				for _, ii := range i {
-					tokens = append(tokens, int(ii.(float64)))
-				}
-				config.InputToken = append(config.InputToken, tokens)
-			}
-		}
-	}
-
-	switch p := input.Prompt.(type) {
-	case string:
-		config.PromptStrings = append(config.PromptStrings, p)
-	case []interface{}:
-		for _, pp := range p {
-			if s, ok := pp.(string); ok {
-				config.PromptStrings = append(config.PromptStrings, s)
-			}
-		}
-	}
-}
-func readInput(c *fiber.Ctx, loader *model.ModelLoader, randomModel bool) (string, *OpenAIRequest, error) {
-	input := new(OpenAIRequest)
-	// Get input data from the request body
-	if err := c.BodyParser(input); err != nil {
-		return "", nil, err
-	}
-
-	modelFile := input.Model
-
-	if c.Params("model") != "" {
-		modelFile = c.Params("model")
-	}
-
-	received, _ := json.Marshal(input)
-
-	log.Debug().Msgf("Request received: %s", string(received))
-
-	// Set model from bearer token, if available
-	bearer := strings.TrimLeft(c.Get("authorization"), "Bearer ")
-	bearerExists := bearer != "" && loader.ExistsInModelPath(bearer)
-
-	// If no model was specified, take the first available
-	if modelFile == "" && !bearerExists && randomModel {
-		models, _ := loader.ListModels()
-		if len(models) > 0 {
-			modelFile = models[0]
-			log.Debug().Msgf("No model specified, using: %s", modelFile)
-		} else {
-			log.Debug().Msgf("No model specified, returning error")
-			return "", nil, fmt.Errorf("no model specified")
-		}
-	}
-
-	// If a model is found in bearer token takes precedence
-	if bearerExists {
-		log.Debug().Msgf("Using model from bearer token: %s", bearer)
-		modelFile = bearer
-	}
-	return modelFile, input, nil
-}
-
-func readConfig(modelFile string, input *OpenAIRequest, cm *ConfigMerger, loader *model.ModelLoader, debug bool, threads, ctx int, f16 bool) (*Config, *OpenAIRequest, error) {
-	// Load a config file if present after the model name
-	modelConfig := filepath.Join(loader.ModelPath, modelFile+".yaml")
-	if _, err := os.Stat(modelConfig); err == nil {
-		if err := cm.LoadConfig(modelConfig); err != nil {
-			return nil, nil, fmt.Errorf("failed loading model config (%s) %s", modelConfig, err.Error())
-		}
-	}
-
-	var config *Config
-	cfg, exists := cm.GetConfig(modelFile)
-	if !exists {
-		config = &Config{
-			OpenAIRequest: defaultRequest(modelFile),
-			ContextSize:   ctx,
-			Threads:       threads,
-			F16:           f16,
-			Debug:         debug,
-		}
-	} else {
-		config = &cfg
-	}
-
-	// Set the parameters for the language model prediction
-	updateConfig(config, input)
-
-	// Don't allow 0 as setting
-	if config.Threads == 0 {
-		if threads != 0 {
-			config.Threads = threads
-		} else {
-			config.Threads = 4
-		}
-	}
-
-	// Enforce debug flag if passed from CLI
-	if debug {
-		config.Debug = true
-	}
-
-	return config, input, nil
-}

api/config/config.go

@@ -0,0 +1,282 @@
+package api_config
+
+import (
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+
+	"gopkg.in/yaml.v3"
+)
+
+type Config struct {
+	PredictionOptions `yaml:"parameters"`
+	Name              string `yaml:"name"`
+
+	F16            bool              `yaml:"f16"`
+	Threads        int               `yaml:"threads"`
+	Debug          bool              `yaml:"debug"`
+	Roles          map[string]string `yaml:"roles"`
+	Embeddings     bool              `yaml:"embeddings"`
+	Backend        string            `yaml:"backend"`
+	TemplateConfig TemplateConfig    `yaml:"template"`
+
+	PromptStrings, InputStrings                []string `yaml:"-"`
+	InputToken                                 [][]int  `yaml:"-"`
+	functionCallString, functionCallNameString string   `yaml:"-"`
+
+	FunctionsConfig Functions `yaml:"function"`
+
+	FeatureFlag FeatureFlag `yaml:"feature_flags"` // Feature Flag registry. We move fast, and features may break on a per model/backend basis. Registry for (usually temporary) flags that indicate aborting something early.
+	// LLM configs (GPT4ALL, Llama.cpp, ...)
+	LLMConfig `yaml:",inline"`
+
+	// AutoGPTQ specifics
+	AutoGPTQ AutoGPTQ `yaml:"autogptq"`
+
+	// Diffusers
+	Diffusers Diffusers `yaml:"diffusers"`
+
+	Step int `yaml:"step"`
+
+	// GRPC Options
+	GRPC GRPC `yaml:"grpc"`
+
+	// Vall-e-x
+	VallE VallE `yaml:"vall-e"`
+}
+
+type VallE struct {
+	AudioPath string `yaml:"audio_path"`
+}
+
+type FeatureFlag map[string]*bool
+
+func (ff FeatureFlag) Enabled(s string) bool {
+	v, exist := ff[s]
+	return exist && v != nil && *v
+}
+
+type GRPC struct {
+	Attempts          int `yaml:"attempts"`
+	AttemptsSleepTime int `yaml:"attempts_sleep_time"`
+}
+
+type Diffusers struct {
+	PipelineType     string  `yaml:"pipeline_type"`
+	SchedulerType    string  `yaml:"scheduler_type"`
+	CUDA             bool    `yaml:"cuda"`
+	EnableParameters string  `yaml:"enable_parameters"` // A list of comma separated parameters to specify
+	CFGScale         float32 `yaml:"cfg_scale"`         // Classifier-Free Guidance Scale
+	IMG2IMG          bool    `yaml:"img2img"`           // Image to Image Diffuser
+	ClipSkip         int     `yaml:"clip_skip"`         // Skip every N frames
+	ClipModel        string  `yaml:"clip_model"`        // Clip model to use
+	ClipSubFolder    string  `yaml:"clip_subfolder"`    // Subfolder to use for clip model
+}
+
+type LLMConfig struct {
+	SystemPrompt    string   `yaml:"system_prompt"`
+	TensorSplit     string   `yaml:"tensor_split"`
+	MainGPU         string   `yaml:"main_gpu"`
+	RMSNormEps      float32  `yaml:"rms_norm_eps"`
+	NGQA            int32    `yaml:"ngqa"`
+	PromptCachePath string   `yaml:"prompt_cache_path"`
+	PromptCacheAll  bool     `yaml:"prompt_cache_all"`
+	PromptCacheRO   bool     `yaml:"prompt_cache_ro"`
+	MirostatETA     float64  `yaml:"mirostat_eta"`
+	MirostatTAU     float64  `yaml:"mirostat_tau"`
+	Mirostat        int      `yaml:"mirostat"`
+	NGPULayers      int      `yaml:"gpu_layers"`
+	MMap            bool     `yaml:"mmap"`
+	MMlock          bool     `yaml:"mmlock"`
+	LowVRAM         bool     `yaml:"low_vram"`
+	Grammar         string   `yaml:"grammar"`
+	StopWords       []string `yaml:"stopwords"`
+	Cutstrings      []string `yaml:"cutstrings"`
+	TrimSpace       []string `yaml:"trimspace"`
+	ContextSize     int      `yaml:"context_size"`
+	NUMA            bool     `yaml:"numa"`
+	LoraAdapter     string   `yaml:"lora_adapter"`
+	LoraBase        string   `yaml:"lora_base"`
+	NoMulMatQ       bool     `yaml:"no_mulmatq"`
+	DraftModel      string   `yaml:"draft_model"`
+	NDraft          int32    `yaml:"n_draft"`
+	Quantization    string   `yaml:"quantization"`
+}
+
+type AutoGPTQ struct {
+	ModelBaseName    string `yaml:"model_base_name"`
+	Device           string `yaml:"device"`
+	Triton           bool   `yaml:"triton"`
+	UseFastTokenizer bool   `yaml:"use_fast_tokenizer"`
+}
+
+type Functions struct {
+	DisableNoAction         bool   `yaml:"disable_no_action"`
+	NoActionFunctionName    string `yaml:"no_action_function_name"`
+	NoActionDescriptionName string `yaml:"no_action_description_name"`
+}
+
+type TemplateConfig struct {
+	Chat        string `yaml:"chat"`
+	ChatMessage string `yaml:"chat_message"`
+	Completion  string `yaml:"completion"`
+	Edit        string `yaml:"edit"`
+	Functions   string `yaml:"function"`
+}
+
+type ConfigLoader struct {
+	configs map[string]Config
+	sync.Mutex
+}
+
+func (c *Config) SetFunctionCallString(s string) {
+	c.functionCallString = s
+}
+
+func (c *Config) SetFunctionCallNameString(s string) {
+	c.functionCallNameString = s
+}
+
+func (c *Config) ShouldUseFunctions() bool {
+	return ((c.functionCallString != "none" || c.functionCallString == "") || c.ShouldCallSpecificFunction())
+}
+
+func (c *Config) ShouldCallSpecificFunction() bool {
+	return len(c.functionCallNameString) > 0
+}
+
+func (c *Config) FunctionToCall() string {
+	return c.functionCallNameString
+}
+
+func defaultPredictOptions(modelFile string) PredictionOptions {
+	return PredictionOptions{
+		TopP:        0.7,
+		TopK:        80,
+		Maxtokens:   512,
+		Temperature: 0.9,
+		Model:       modelFile,
+	}
+}
+
+func DefaultConfig(modelFile string) *Config {
+	return &Config{
+		PredictionOptions: defaultPredictOptions(modelFile),
+	}
+}
+
+func NewConfigLoader() *ConfigLoader {
+	return &ConfigLoader{
+		configs: make(map[string]Config),
+	}
+}
+func ReadConfigFile(file string) ([]*Config, error) {
+	c := &[]*Config{}
+	f, err := os.ReadFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("cannot read config file: %w", err)
+	}
+	if err := yaml.Unmarshal(f, c); err != nil {
+		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
+	}
+
+	return *c, nil
+}
+
+func ReadConfig(file string) (*Config, error) {
+	c := &Config{}
+	f, err := os.ReadFile(file)
+	if err != nil {
+		return nil, fmt.Errorf("cannot read config file: %w", err)
+	}
+	if err := yaml.Unmarshal(f, c); err != nil {
+		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
+	}
+
+	return c, nil
+}
+
+func (cm *ConfigLoader) LoadConfigFile(file string) error {
+	cm.Lock()
+	defer cm.Unlock()
+	c, err := ReadConfigFile(file)
+	if err != nil {
+		return fmt.Errorf("cannot load config file: %w", err)
+	}
+
+	for _, cc := range c {
+		cm.configs[cc.Name] = *cc
+	}
+	return nil
+}
+
+func (cm *ConfigLoader) LoadConfig(file string) error {
+	cm.Lock()
+	defer cm.Unlock()
+	c, err := ReadConfig(file)
+	if err != nil {
+		return fmt.Errorf("cannot read config file: %w", err)
+	}
+
+	cm.configs[c.Name] = *c
+	return nil
+}
+
+func (cm *ConfigLoader) GetConfig(m string) (Config, bool) {
+	cm.Lock()
+	defer cm.Unlock()
+	v, exists := cm.configs[m]
+	return v, exists
+}
+
+func (cm *ConfigLoader) GetAllConfigs() []Config {
+	cm.Lock()
+	defer cm.Unlock()
+	var res []Config
+	for _, v := range cm.configs {
+		res = append(res, v)
+	}
+	return res
+}
+
+func (cm *ConfigLoader) ListConfigs() []string {
+	cm.Lock()
+	defer cm.Unlock()
+	var res []string
+	for k := range cm.configs {
+		res = append(res, k)
+	}
+	return res
+}
+
+func (cm *ConfigLoader) LoadConfigs(path string) error {
+	cm.Lock()
+	defer cm.Unlock()
+	entries, err := os.ReadDir(path)
+	if err != nil {
+		return err
+	}
+	files := make([]fs.FileInfo, 0, len(entries))
+	for _, entry := range entries {
+		info, err := entry.Info()
+		if err != nil {
+			return err
+		}
+		files = append(files, info)
+	}
+	for _, file := range files {
+		// Skip templates, YAML and .keep files
+		if !strings.Contains(file.Name(), ".yaml") {
+			continue
+		}
+		c, err := ReadConfig(filepath.Join(path, file.Name()))
+		if err == nil {
+			cm.configs[c.Name] = *c
+		}
+	}
+
+	return nil
+}

api/config/config_test.go

@@ -0,0 +1,56 @@
+package api_config_test
+
+import (
+	"os"
+
+	. "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/pkg/model"
+	. "github.com/onsi/ginkgo/v2"
+	. "github.com/onsi/gomega"
+)
+
+var _ = Describe("Test cases for config related functions", func() {
+
+	var (
+		configFile string
+	)
+
+	Context("Test Read configuration functions", func() {
+		configFile = os.Getenv("CONFIG_FILE")
+		It("Test ReadConfigFile", func() {
+			config, err := ReadConfigFile(configFile)
+			Expect(err).To(BeNil())
+			Expect(config).ToNot(BeNil())
+			// two configs in config.yaml
+			Expect(config[0].Name).To(Equal("list1"))
+			Expect(config[1].Name).To(Equal("list2"))
+		})
+
+		It("Test LoadConfigs", func() {
+			cm := NewConfigLoader()
+			opts := options.NewOptions()
+			modelLoader := model.NewModelLoader(os.Getenv("MODELS_PATH"))
+			options.WithModelLoader(modelLoader)(opts)
+
+			err := cm.LoadConfigs(opts.Loader.ModelPath)
+			Expect(err).To(BeNil())
+			Expect(cm.ListConfigs()).ToNot(BeNil())
+
+			// config should includes gpt4all models's api.config
+			Expect(cm.ListConfigs()).To(ContainElements("gpt4all"))
+
+			// config should includes gpt2 models's api.config
+			Expect(cm.ListConfigs()).To(ContainElements("gpt4all-2"))
+
+			// config should includes text-embedding-ada-002 models's api.config
+			Expect(cm.ListConfigs()).To(ContainElements("text-embedding-ada-002"))
+
+			// config should includes rwkv_test models's api.config
+			Expect(cm.ListConfigs()).To(ContainElements("rwkv_test"))
+
+			// config should includes whisper-1 models's api.config
+			Expect(cm.ListConfigs()).To(ContainElements("whisper-1"))
+		})
+	})
+})

api/config/prediction.go

@@ -0,0 +1,50 @@
+package api_config
+
+type PredictionOptions struct {
+
+	// Also part of the OpenAI official spec
+	Model string `json:"model" yaml:"model"`
+
+	// Also part of the OpenAI official spec
+	Language string `json:"language"`
+
+	// Also part of the OpenAI official spec. use it for returning multiple results
+	N int `json:"n"`
+
+	// Common options between all the API calls, part of the OpenAI spec
+	TopP        float64 `json:"top_p" yaml:"top_p"`
+	TopK        int     `json:"top_k" yaml:"top_k"`
+	Temperature float64 `json:"temperature" yaml:"temperature"`
+	Maxtokens   int     `json:"max_tokens" yaml:"max_tokens"`
+	Echo        bool    `json:"echo"`
+
+	// Custom parameters - not present in the OpenAI API
+	Batch         int     `json:"batch" yaml:"batch"`
+	F16           bool    `json:"f16" yaml:"f16"`
+	IgnoreEOS     bool    `json:"ignore_eos" yaml:"ignore_eos"`
+	RepeatPenalty float64 `json:"repeat_penalty" yaml:"repeat_penalty"`
+	Keep          int     `json:"n_keep" yaml:"n_keep"`
+
+	MirostatETA float64 `json:"mirostat_eta" yaml:"mirostat_eta"`
+	MirostatTAU float64 `json:"mirostat_tau" yaml:"mirostat_tau"`
+	Mirostat    int     `json:"mirostat" yaml:"mirostat"`
+
+	FrequencyPenalty float64 `json:"frequency_penalty" yaml:"frequency_penalty"`
+	TFZ              float64 `json:"tfz" yaml:"tfz"`
+
+	TypicalP float64 `json:"typical_p" yaml:"typical_p"`
+	Seed     int     `json:"seed" yaml:"seed"`
+
+	NegativePrompt      string  `json:"negative_prompt" yaml:"negative_prompt"`
+	RopeFreqBase        float32 `json:"rope_freq_base" yaml:"rope_freq_base"`
+	RopeFreqScale       float32 `json:"rope_freq_scale" yaml:"rope_freq_scale"`
+	NegativePromptScale float32 `json:"negative_prompt_scale" yaml:"negative_prompt_scale"`
+	// AutoGPTQ
+	UseFastTokenizer bool `json:"use_fast_tokenizer" yaml:"use_fast_tokenizer"`
+
+	// Diffusers
+	ClipSkip int `json:"clip_skip" yaml:"clip_skip"`
+
+	// RWKV (?)
+	Tokenizer string `json:"tokenizer" yaml:"tokenizer"`
+}

api/gallery.go

@@ -1,196 +0,0 @@
-package api
-
-import (
-	"context"
-	"fmt"
-	"io/ioutil"
-	"net/http"
-	"net/url"
-	"strings"
-	"sync"
-
-	"github.com/go-skynet/LocalAI/pkg/gallery"
-	"github.com/gofiber/fiber/v2"
-	"github.com/google/uuid"
-	"gopkg.in/yaml.v3"
-)
-
-type galleryOp struct {
-	req ApplyGalleryModelRequest
-	id  string
-}
-
-type galleryOpStatus struct {
-	Error     error  `json:"error"`
-	Processed bool   `json:"processed"`
-	Message   string `json:"message"`
-}
-
-type galleryApplier struct {
-	modelPath string
-	sync.Mutex
-	C        chan galleryOp
-	statuses map[string]*galleryOpStatus
-}
-
-func newGalleryApplier(modelPath string) *galleryApplier {
-	return &galleryApplier{
-		modelPath: modelPath,
-		C:         make(chan galleryOp),
-		statuses:  make(map[string]*galleryOpStatus),
-	}
-}
-func (g *galleryApplier) updatestatus(s string, op *galleryOpStatus) {
-	g.Lock()
-	defer g.Unlock()
-	g.statuses[s] = op
-}
-
-func (g *galleryApplier) getstatus(s string) *galleryOpStatus {
-	g.Lock()
-	defer g.Unlock()
-
-	return g.statuses[s]
-}
-
-func (g *galleryApplier) start(c context.Context, cm *ConfigMerger) {
-	go func() {
-		for {
-			select {
-			case <-c.Done():
-				return
-			case op := <-g.C:
-				g.updatestatus(op.id, &galleryOpStatus{Message: "processing"})
-
-				updateError := func(e error) {
-					g.updatestatus(op.id, &galleryOpStatus{Error: e, Processed: true})
-				}
-
-				url, err := op.req.DecodeURL()
-				if err != nil {
-					updateError(err)
-					continue
-				}
-
-				// Send a GET request to the URL
-				response, err := http.Get(url)
-				if err != nil {
-					updateError(err)
-					continue
-				}
-				defer response.Body.Close()
-
-				// Read the response body
-				body, err := ioutil.ReadAll(response.Body)
-				if err != nil {
-					updateError(err)
-					continue
-				}
-
-				// Unmarshal YAML data into a Config struct
-				var config gallery.Config
-				err = yaml.Unmarshal(body, &config)
-				if err != nil {
-					updateError(fmt.Errorf("failed to unmarshal YAML: %v", err))
-					continue
-				}
-
-				config.Files = append(config.Files, op.req.AdditionalFiles...)
-
-				if err := gallery.Apply(g.modelPath, op.req.Name, &config, op.req.Overrides); err != nil {
-					updateError(err)
-					continue
-				}
-
-				// Reload models
-				if err := cm.LoadConfigs(g.modelPath); err != nil {
-					updateError(err)
-					continue
-				}
-
-				g.updatestatus(op.id, &galleryOpStatus{Processed: true, Message: "completed"})
-			}
-		}
-	}()
-}
-
-// endpoints
-
-type ApplyGalleryModelRequest struct {
-	URL             string                 `json:"url"`
-	Name            string                 `json:"name"`
-	Overrides       map[string]interface{} `json:"overrides"`
-	AdditionalFiles []gallery.File         `json:"files"`
-}
-
-const (
-	githubURI = "github:"
-)
-
-func (request ApplyGalleryModelRequest) DecodeURL() (string, error) {
-	input := request.URL
-	var rawURL string
-
-	if strings.HasPrefix(input, githubURI) {
-		parts := strings.Split(input, ":")
-		repoParts := strings.Split(parts[1], "@")
-		branch := "main"
-
-		if len(repoParts) > 1 {
-			branch = repoParts[1]
-		}
-
-		repoPath := strings.Split(repoParts[0], "/")
-		org := repoPath[0]
-		project := repoPath[1]
-		projectPath := strings.Join(repoPath[2:], "/")
-
-		rawURL = fmt.Sprintf("https://raw.githubusercontent.com/%s/%s/%s/%s", org, project, branch, projectPath)
-	} else if strings.HasPrefix(input, "http://") || strings.HasPrefix(input, "https://") {
-		// Handle regular URLs
-		u, err := url.Parse(input)
-		if err != nil {
-			return "", fmt.Errorf("invalid URL: %w", err)
-		}
-		rawURL = u.String()
-	} else {
-		return "", fmt.Errorf("invalid URL format")
-	}
-
-	return rawURL, nil
-}
-
-func getOpStatus(g *galleryApplier) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-
-		status := g.getstatus(c.Params("uuid"))
-		if status == nil {
-			return fmt.Errorf("could not find any status for ID")
-		}
-
-		return c.JSON(status)
-	}
-}
-
-func applyModelGallery(modelPath string, cm *ConfigMerger, g chan galleryOp) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		input := new(ApplyGalleryModelRequest)
-		// Get input data from the request body
-		if err := c.BodyParser(input); err != nil {
-			return err
-		}
-
-		uuid, err := uuid.NewUUID()
-		if err != nil {
-			return err
-		}
-		g <- galleryOp{
-			req: *input,
-			id:  uuid.String(),
-		}
-		return c.JSON(struct {
-			ID        string `json:"uuid"`
-			StatusURL string `json:"status"`
-		}{ID: uuid.String(), StatusURL: c.BaseURL() + "/models/jobs/" + uuid.String()})
-	}
-}

api/gallery_test.go

@@ -1,30 +0,0 @@
-package api_test
-
-import (
-	. "github.com/go-skynet/LocalAI/api"
-	. "github.com/onsi/ginkgo/v2"
-	. "github.com/onsi/gomega"
-)
-
-var _ = Describe("Gallery API tests", func() {
-	Context("requests", func() {
-		It("parses github with a branch", func() {
-			req := ApplyGalleryModelRequest{URL: "github:go-skynet/model-gallery/gpt4all-j.yaml@main"}
-			str, err := req.DecodeURL()
-			Expect(err).ToNot(HaveOccurred())
-			Expect(str).To(Equal("https://raw.githubusercontent.com/go-skynet/model-gallery/main/gpt4all-j.yaml"))
-		})
-		It("parses github without a branch", func() {
-			req := ApplyGalleryModelRequest{URL: "github:go-skynet/model-gallery/gpt4all-j.yaml"}
-			str, err := req.DecodeURL()
-			Expect(err).ToNot(HaveOccurred())
-			Expect(str).To(Equal("https://raw.githubusercontent.com/go-skynet/model-gallery/main/gpt4all-j.yaml"))
-		})
-		It("parses URLS", func() {
-			req := ApplyGalleryModelRequest{URL: "https://raw.githubusercontent.com/go-skynet/model-gallery/main/gpt4all-j.yaml"}
-			str, err := req.DecodeURL()
-			Expect(err).ToNot(HaveOccurred())
-			Expect(str).To(Equal("https://raw.githubusercontent.com/go-skynet/model-gallery/main/gpt4all-j.yaml"))
-		})
-	})
-})

api/localai/backend_monitor.go

@@ -0,0 +1,163 @@
+package localai
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/pkg/grpc/proto"
+
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/gofiber/fiber/v2"
+	"github.com/rs/zerolog/log"
+
+	gopsutil "github.com/shirou/gopsutil/v3/process"
+)
+
+type BackendMonitorRequest struct {
+	Model string `json:"model" yaml:"model"`
+}
+
+type BackendMonitorResponse struct {
+	MemoryInfo    *gopsutil.MemoryInfoStat
+	MemoryPercent float32
+	CPUPercent    float64
+}
+
+type BackendMonitor struct {
+	configLoader *config.ConfigLoader
+	options      *options.Option // Taking options in case we need to inspect ExternalGRPCBackends, though that's out of scope for now, hence the name.
+}
+
+func NewBackendMonitor(configLoader *config.ConfigLoader, options *options.Option) BackendMonitor {
+	return BackendMonitor{
+		configLoader: configLoader,
+		options:      options,
+	}
+}
+
+func (bm *BackendMonitor) SampleLocalBackendProcess(model string) (*BackendMonitorResponse, error) {
+	config, exists := bm.configLoader.GetConfig(model)
+	var backend string
+	if exists {
+		backend = config.Model
+	} else {
+		// Last ditch effort: use it raw, see if a backend happens to match.
+		backend = model
+	}
+
+	if !strings.HasSuffix(backend, ".bin") {
+		backend = fmt.Sprintf("%s.bin", backend)
+	}
+
+	pid, err := bm.options.Loader.GetGRPCPID(backend)
+
+	if err != nil {
+		log.Error().Msgf("model %s : failed to find pid %+v", model, err)
+		return nil, err
+	}
+
+	// Name is slightly frightening but this does _not_ create a new process, rather it looks up an existing process by PID.
+	backendProcess, err := gopsutil.NewProcess(int32(pid))
+
+	if err != nil {
+		log.Error().Msgf("model %s [PID %d] : error getting process info %+v", model, pid, err)
+		return nil, err
+	}
+
+	memInfo, err := backendProcess.MemoryInfo()
+
+	if err != nil {
+		log.Error().Msgf("model %s [PID %d] : error getting memory info %+v", model, pid, err)
+		return nil, err
+	}
+
+	memPercent, err := backendProcess.MemoryPercent()
+	if err != nil {
+		log.Error().Msgf("model %s [PID %d] : error getting memory percent %+v", model, pid, err)
+		return nil, err
+	}
+
+	cpuPercent, err := backendProcess.CPUPercent()
+	if err != nil {
+		log.Error().Msgf("model %s [PID %d] : error getting cpu percent %+v", model, pid, err)
+		return nil, err
+	}
+
+	return &BackendMonitorResponse{
+		MemoryInfo:    memInfo,
+		MemoryPercent: memPercent,
+		CPUPercent:    cpuPercent,
+	}, nil
+}
+
+func (bm BackendMonitor) getModelLoaderIDFromCtx(c *fiber.Ctx) (string, error) {
+	input := new(BackendMonitorRequest)
+	// Get input data from the request body
+	if err := c.BodyParser(input); err != nil {
+		return "", err
+	}
+
+	config, exists := bm.configLoader.GetConfig(input.Model)
+	var backendId string
+	if exists {
+		backendId = config.Model
+	} else {
+		// Last ditch effort: use it raw, see if a backend happens to match.
+		backendId = input.Model
+	}
+
+	if !strings.HasSuffix(backendId, ".bin") {
+		backendId = fmt.Sprintf("%s.bin", backendId)
+	}
+
+	return backendId, nil
+}
+
+func BackendMonitorEndpoint(bm BackendMonitor) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+
+		backendId, err := bm.getModelLoaderIDFromCtx(c)
+		if err != nil {
+			return err
+		}
+
+		client := bm.options.Loader.CheckIsLoaded(backendId)
+
+		if client == nil {
+			return fmt.Errorf("backend %s is not currently loaded", backendId)
+		}
+
+		status, rpcErr := client.Status(context.TODO())
+		if rpcErr != nil {
+			log.Warn().Msgf("backend %s experienced an error retrieving status info: %s", backendId, rpcErr.Error())
+			val, slbErr := bm.SampleLocalBackendProcess(backendId)
+			if slbErr != nil {
+				return fmt.Errorf("backend %s experienced an error retrieving status info via rpc: %s, then failed local node process sample: %s", backendId, rpcErr.Error(), slbErr.Error())
+			}
+			return c.JSON(proto.StatusResponse{
+				State: proto.StatusResponse_ERROR,
+				Memory: &proto.MemoryUsageData{
+					Total: val.MemoryInfo.VMS,
+					Breakdown: map[string]uint64{
+						"gopsutil-RSS": val.MemoryInfo.RSS,
+					},
+				},
+			})
+		}
+
+		return c.JSON(status)
+	}
+}
+
+func BackendShutdownEndpoint(bm BackendMonitor) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		backendId, err := bm.getModelLoaderIDFromCtx(c)
+		if err != nil {
+			return err
+		}
+
+		return bm.options.Loader.ShutdownModel(backendId)
+	}
+}

api/localai/gallery.go

@@ -0,0 +1,320 @@
+package localai
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"slices"
+	"strings"
+	"sync"
+
+	json "github.com/json-iterator/go"
+	"gopkg.in/yaml.v3"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/pkg/gallery"
+	"github.com/go-skynet/LocalAI/pkg/utils"
+
+	"github.com/gofiber/fiber/v2"
+	"github.com/google/uuid"
+	"github.com/rs/zerolog/log"
+)
+
+type galleryOp struct {
+	req         gallery.GalleryModel
+	id          string
+	galleries   []gallery.Gallery
+	galleryName string
+}
+
+type galleryOpStatus struct {
+	FileName           string  `json:"file_name"`
+	Error              error   `json:"error"`
+	Processed          bool    `json:"processed"`
+	Message            string  `json:"message"`
+	Progress           float64 `json:"progress"`
+	TotalFileSize      string  `json:"file_size"`
+	DownloadedFileSize string  `json:"downloaded_size"`
+}
+
+type galleryApplier struct {
+	modelPath string
+	sync.Mutex
+	C        chan galleryOp
+	statuses map[string]*galleryOpStatus
+}
+
+func NewGalleryService(modelPath string) *galleryApplier {
+	return &galleryApplier{
+		modelPath: modelPath,
+		C:         make(chan galleryOp),
+		statuses:  make(map[string]*galleryOpStatus),
+	}
+}
+
+func prepareModel(modelPath string, req gallery.GalleryModel, cm *config.ConfigLoader, downloadStatus func(string, string, string, float64)) error {
+
+	config, err := gallery.GetGalleryConfigFromURL(req.URL)
+	if err != nil {
+		return err
+	}
+
+	config.Files = append(config.Files, req.AdditionalFiles...)
+
+	return gallery.InstallModel(modelPath, req.Name, &config, req.Overrides, downloadStatus)
+}
+
+func (g *galleryApplier) updateStatus(s string, op *galleryOpStatus) {
+	g.Lock()
+	defer g.Unlock()
+	g.statuses[s] = op
+}
+
+func (g *galleryApplier) getStatus(s string) *galleryOpStatus {
+	g.Lock()
+	defer g.Unlock()
+
+	return g.statuses[s]
+}
+
+func (g *galleryApplier) getAllStatus() map[string]*galleryOpStatus {
+	g.Lock()
+	defer g.Unlock()
+
+	return g.statuses
+}
+
+func (g *galleryApplier) Start(c context.Context, cm *config.ConfigLoader) {
+	go func() {
+		for {
+			select {
+			case <-c.Done():
+				return
+			case op := <-g.C:
+				utils.ResetDownloadTimers()
+
+				g.updateStatus(op.id, &galleryOpStatus{Message: "processing", Progress: 0})
+
+				// updates the status with an error
+				updateError := func(e error) {
+					g.updateStatus(op.id, &galleryOpStatus{Error: e, Processed: true, Message: "error: " + e.Error()})
+				}
+
+				// displayDownload displays the download progress
+				progressCallback := func(fileName string, current string, total string, percentage float64) {
+					g.updateStatus(op.id, &galleryOpStatus{Message: "processing", FileName: fileName, Progress: percentage, TotalFileSize: total, DownloadedFileSize: current})
+					utils.DisplayDownloadFunction(fileName, current, total, percentage)
+				}
+
+				var err error
+				// if the request contains a gallery name, we apply the gallery from the gallery list
+				if op.galleryName != "" {
+					if strings.Contains(op.galleryName, "@") {
+						err = gallery.InstallModelFromGallery(op.galleries, op.galleryName, g.modelPath, op.req, progressCallback)
+					} else {
+						err = gallery.InstallModelFromGalleryByName(op.galleries, op.galleryName, g.modelPath, op.req, progressCallback)
+					}
+				} else {
+					err = prepareModel(g.modelPath, op.req, cm, progressCallback)
+				}
+
+				if err != nil {
+					updateError(err)
+					continue
+				}
+
+				// Reload models
+				err = cm.LoadConfigs(g.modelPath)
+				if err != nil {
+					updateError(err)
+					continue
+				}
+
+				g.updateStatus(op.id, &galleryOpStatus{Processed: true, Message: "completed", Progress: 100})
+			}
+		}
+	}()
+}
+
+type galleryModel struct {
+	gallery.GalleryModel `yaml:",inline"` // https://github.com/go-yaml/yaml/issues/63
+	ID                   string           `json:"id"`
+}
+
+func processRequests(modelPath, s string, cm *config.ConfigLoader, galleries []gallery.Gallery, requests []galleryModel) error {
+	var err error
+	for _, r := range requests {
+		utils.ResetDownloadTimers()
+		if r.ID == "" {
+			err = prepareModel(modelPath, r.GalleryModel, cm, utils.DisplayDownloadFunction)
+		} else {
+			if strings.Contains(r.ID, "@") {
+				err = gallery.InstallModelFromGallery(
+					galleries, r.ID, modelPath, r.GalleryModel, utils.DisplayDownloadFunction)
+			} else {
+				err = gallery.InstallModelFromGalleryByName(
+					galleries, r.ID, modelPath, r.GalleryModel, utils.DisplayDownloadFunction)
+			}
+		}
+	}
+	return err
+}
+
+func ApplyGalleryFromFile(modelPath, s string, cm *config.ConfigLoader, galleries []gallery.Gallery) error {
+	dat, err := os.ReadFile(s)
+	if err != nil {
+		return err
+	}
+	var requests []galleryModel
+
+	if err := yaml.Unmarshal(dat, &requests); err != nil {
+		return err
+	}
+
+	return processRequests(modelPath, s, cm, galleries, requests)
+}
+
+func ApplyGalleryFromString(modelPath, s string, cm *config.ConfigLoader, galleries []gallery.Gallery) error {
+	var requests []galleryModel
+	err := json.Unmarshal([]byte(s), &requests)
+	if err != nil {
+		return err
+	}
+
+	return processRequests(modelPath, s, cm, galleries, requests)
+}
+
+/// Endpoint Service
+
+type ModelGalleryService struct {
+	galleries      []gallery.Gallery
+	modelPath      string
+	galleryApplier *galleryApplier
+}
+
+type GalleryModel struct {
+	ID string `json:"id"`
+	gallery.GalleryModel
+}
+
+func CreateModelGalleryService(galleries []gallery.Gallery, modelPath string, galleryApplier *galleryApplier) ModelGalleryService {
+	return ModelGalleryService{
+		galleries:      galleries,
+		modelPath:      modelPath,
+		galleryApplier: galleryApplier,
+	}
+}
+
+func (mgs *ModelGalleryService) GetOpStatusEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		status := mgs.galleryApplier.getStatus(c.Params("uuid"))
+		if status == nil {
+			return fmt.Errorf("could not find any status for ID")
+		}
+		return c.JSON(status)
+	}
+}
+
+func (mgs *ModelGalleryService) GetAllStatusEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		return c.JSON(mgs.galleryApplier.getAllStatus())
+	}
+}
+
+func (mgs *ModelGalleryService) ApplyModelGalleryEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		input := new(GalleryModel)
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+
+		uuid, err := uuid.NewUUID()
+		if err != nil {
+			return err
+		}
+		mgs.galleryApplier.C <- galleryOp{
+			req:         input.GalleryModel,
+			id:          uuid.String(),
+			galleryName: input.ID,
+			galleries:   mgs.galleries,
+		}
+		return c.JSON(struct {
+			ID        string `json:"uuid"`
+			StatusURL string `json:"status"`
+		}{ID: uuid.String(), StatusURL: c.BaseURL() + "/models/jobs/" + uuid.String()})
+	}
+}
+
+func (mgs *ModelGalleryService) ListModelFromGalleryEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		log.Debug().Msgf("Listing models from galleries: %+v", mgs.galleries)
+
+		models, err := gallery.AvailableGalleryModels(mgs.galleries, mgs.modelPath)
+		if err != nil {
+			return err
+		}
+		log.Debug().Msgf("Models found from galleries: %+v", models)
+		for _, m := range models {
+			log.Debug().Msgf("Model found from galleries: %+v", m)
+		}
+		dat, err := json.Marshal(models)
+		if err != nil {
+			return err
+		}
+		return c.Send(dat)
+	}
+}
+
+// NOTE: This is different (and much simpler!) than above! This JUST lists the model galleries that have been loaded, not their contents!
+func (mgs *ModelGalleryService) ListModelGalleriesEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		log.Debug().Msgf("Listing model galleries %+v", mgs.galleries)
+		dat, err := json.Marshal(mgs.galleries)
+		if err != nil {
+			return err
+		}
+		return c.Send(dat)
+	}
+}
+
+func (mgs *ModelGalleryService) AddModelGalleryEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		input := new(gallery.Gallery)
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+		if slices.ContainsFunc(mgs.galleries, func(gallery gallery.Gallery) bool {
+			return gallery.Name == input.Name
+		}) {
+			return fmt.Errorf("%s already exists", input.Name)
+		}
+		dat, err := json.Marshal(mgs.galleries)
+		if err != nil {
+			return err
+		}
+		log.Debug().Msgf("Adding %+v to gallery list", *input)
+		mgs.galleries = append(mgs.galleries, *input)
+		return c.Send(dat)
+	}
+}
+
+func (mgs *ModelGalleryService) RemoveModelGalleryEndpoint() func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		input := new(gallery.Gallery)
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+		if !slices.ContainsFunc(mgs.galleries, func(gallery gallery.Gallery) bool {
+			return gallery.Name == input.Name
+		}) {
+			return fmt.Errorf("%s is not currently registered", input.Name)
+		}
+		mgs.galleries = slices.DeleteFunc(mgs.galleries, func(gallery gallery.Gallery) bool {
+			return gallery.Name == input.Name
+		})
+		return c.Send(nil)
+	}
+}

api/localai/localai.go

@@ -0,0 +1,32 @@
+package localai
+
+import (
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/gofiber/fiber/v2"
+)
+
+type TTSRequest struct {
+	Model   string `json:"model" yaml:"model"`
+	Input   string `json:"input" yaml:"input"`
+	Backend string `json:"backend" yaml:"backend"`
+}
+
+func TTSEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+
+		input := new(TTSRequest)
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+
+		filePath, _, err := backend.ModelTTS(input.Backend, input.Input, input.Model, o.Loader, o)
+		if err != nil {
+			return err
+		}
+		return c.Download(filePath)
+	}
+}

api/openai.go

@@ -1,678 +0,0 @@
-package api
-
-import (
-	"bufio"
-	"bytes"
-	"encoding/base64"
-	"encoding/json"
-	"fmt"
-	"io"
-	"io/ioutil"
-	"net/http"
-	"os"
-	"path"
-	"path/filepath"
-	"strconv"
-	"strings"
-
-	"github.com/ggerganov/whisper.cpp/bindings/go/pkg/whisper"
-	model "github.com/go-skynet/LocalAI/pkg/model"
-	whisperutil "github.com/go-skynet/LocalAI/pkg/whisper"
-	llama "github.com/go-skynet/go-llama.cpp"
-	"github.com/gofiber/fiber/v2"
-	"github.com/rs/zerolog/log"
-	"github.com/valyala/fasthttp"
-)
-
-// APIError provides error information returned by the OpenAI API.
-type APIError struct {
-	Code    any     `json:"code,omitempty"`
-	Message string  `json:"message"`
-	Param   *string `json:"param,omitempty"`
-	Type    string  `json:"type"`
-}
-
-type ErrorResponse struct {
-	Error *APIError `json:"error,omitempty"`
-}
-
-type OpenAIUsage struct {
-	PromptTokens     int `json:"prompt_tokens"`
-	CompletionTokens int `json:"completion_tokens"`
-	TotalTokens      int `json:"total_tokens"`
-}
-
-type Item struct {
-	Embedding []float32 `json:"embedding"`
-	Index     int       `json:"index"`
-	Object    string    `json:"object,omitempty"`
-
-	// Images
-	URL     string `json:"url,omitempty"`
-	B64JSON string `json:"b64_json,omitempty"`
-}
-
-type OpenAIResponse struct {
-	Created int      `json:"created,omitempty"`
-	Object  string   `json:"object,omitempty"`
-	ID      string   `json:"id,omitempty"`
-	Model   string   `json:"model,omitempty"`
-	Choices []Choice `json:"choices,omitempty"`
-	Data    []Item   `json:"data,omitempty"`
-
-	Usage OpenAIUsage `json:"usage"`
-}
-
-type Choice struct {
-	Index        int      `json:"index,omitempty"`
-	FinishReason string   `json:"finish_reason,omitempty"`
-	Message      *Message `json:"message,omitempty"`
-	Delta        *Message `json:"delta,omitempty"`
-	Text         string   `json:"text,omitempty"`
-}
-
-type Message struct {
-	Role    string `json:"role,omitempty" yaml:"role"`
-	Content string `json:"content,omitempty" yaml:"content"`
-}
-
-type OpenAIModel struct {
-	ID     string `json:"id"`
-	Object string `json:"object"`
-}
-
-type OpenAIRequest struct {
-	Model string `json:"model" yaml:"model"`
-
-	// whisper
-	File     string `json:"file" validate:"required"`
-	Language string `json:"language"`
-	//whisper/image
-	ResponseFormat string `json:"response_format"`
-	// image
-	Size string `json:"size"`
-	// Prompt is read only by completion/image API calls
-	Prompt interface{} `json:"prompt" yaml:"prompt"`
-
-	// Edit endpoint
-	Instruction string      `json:"instruction" yaml:"instruction"`
-	Input       interface{} `json:"input" yaml:"input"`
-
-	Stop interface{} `json:"stop" yaml:"stop"`
-
-	// Messages is read only by chat/completion API calls
-	Messages []Message `json:"messages" yaml:"messages"`
-
-	Stream bool `json:"stream"`
-	Echo   bool `json:"echo"`
-	// Common options between all the API calls
-	TopP        float64 `json:"top_p" yaml:"top_p"`
-	TopK        int     `json:"top_k" yaml:"top_k"`
-	Temperature float64 `json:"temperature" yaml:"temperature"`
-	Maxtokens   int     `json:"max_tokens" yaml:"max_tokens"`
-
-	N int `json:"n"`
-
-	// Custom parameters - not present in the OpenAI API
-	Batch         int     `json:"batch" yaml:"batch"`
-	F16           bool    `json:"f16" yaml:"f16"`
-	IgnoreEOS     bool    `json:"ignore_eos" yaml:"ignore_eos"`
-	RepeatPenalty float64 `json:"repeat_penalty" yaml:"repeat_penalty"`
-	Keep          int     `json:"n_keep" yaml:"n_keep"`
-
-	MirostatETA float64 `json:"mirostat_eta" yaml:"mirostat_eta"`
-	MirostatTAU float64 `json:"mirostat_tau" yaml:"mirostat_tau"`
-	Mirostat    int     `json:"mirostat" yaml:"mirostat"`
-
-	Seed int `json:"seed" yaml:"seed"`
-
-	// Image (not supported by OpenAI)
-	Mode int `json:"mode"`
-	Step int `json:"step"`
-}
-
-func defaultRequest(modelFile string) OpenAIRequest {
-	return OpenAIRequest{
-		TopP:        0.7,
-		TopK:        80,
-		Maxtokens:   512,
-		Temperature: 0.9,
-		Model:       modelFile,
-	}
-}
-
-// https://platform.openai.com/docs/api-reference/completions
-func completionEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-
-		model, input, err := readInput(c, o.loader, true)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		config, input, err := readConfig(model, input, cm, o.loader, o.debug, o.threads, o.ctxSize, o.f16)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		log.Debug().Msgf("Parameter Config: %+v", config)
-
-		templateFile := config.Model
-
-		if config.TemplateConfig.Completion != "" {
-			templateFile = config.TemplateConfig.Completion
-		}
-
-		var result []Choice
-		for _, i := range config.PromptStrings {
-			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-			templatedInput, err := o.loader.TemplatePrefix(templateFile, struct {
-				Input string
-			}{Input: i})
-			if err == nil {
-				i = templatedInput
-				log.Debug().Msgf("Template found, input modified to: %s", i)
-			}
-
-			r, err := ComputeChoices(i, input, config, o.loader, func(s string, c *[]Choice) {
-				*c = append(*c, Choice{Text: s})
-			}, nil)
-			if err != nil {
-				return err
-			}
-
-			result = append(result, r...)
-		}
-
-		resp := &OpenAIResponse{
-			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
-			Choices: result,
-			Object:  "text_completion",
-		}
-
-		jsonResult, _ := json.Marshal(resp)
-		log.Debug().Msgf("Response: %s", jsonResult)
-
-		// Return the prediction in the response body
-		return c.JSON(resp)
-	}
-}
-
-// https://platform.openai.com/docs/api-reference/embeddings
-func embeddingsEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		model, input, err := readInput(c, o.loader, true)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		config, input, err := readConfig(model, input, cm, o.loader, o.debug, o.threads, o.ctxSize, o.f16)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		log.Debug().Msgf("Parameter Config: %+v", config)
-		items := []Item{}
-
-		for i, s := range config.InputToken {
-			// get the model function to call for the result
-			embedFn, err := ModelEmbedding("", s, o.loader, *config)
-			if err != nil {
-				return err
-			}
-
-			embeddings, err := embedFn()
-			if err != nil {
-				return err
-			}
-			items = append(items, Item{Embedding: embeddings, Index: i, Object: "embedding"})
-		}
-
-		for i, s := range config.InputStrings {
-			// get the model function to call for the result
-			embedFn, err := ModelEmbedding(s, []int{}, o.loader, *config)
-			if err != nil {
-				return err
-			}
-
-			embeddings, err := embedFn()
-			if err != nil {
-				return err
-			}
-			items = append(items, Item{Embedding: embeddings, Index: i, Object: "embedding"})
-		}
-
-		resp := &OpenAIResponse{
-			Model:  input.Model, // we have to return what the user sent here, due to OpenAI spec.
-			Data:   items,
-			Object: "list",
-		}
-
-		jsonResult, _ := json.Marshal(resp)
-		log.Debug().Msgf("Response: %s", jsonResult)
-
-		// Return the prediction in the response body
-		return c.JSON(resp)
-	}
-}
-
-func chatEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-
-	process := func(s string, req *OpenAIRequest, config *Config, loader *model.ModelLoader, responses chan OpenAIResponse) {
-		initialMessage := OpenAIResponse{
-			Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
-			Choices: []Choice{{Delta: &Message{Role: "assistant"}}},
-			Object:  "chat.completion.chunk",
-		}
-		responses <- initialMessage
-
-		ComputeChoices(s, req, config, loader, func(s string, c *[]Choice) {}, func(s string) bool {
-			resp := OpenAIResponse{
-				Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
-				Choices: []Choice{{Delta: &Message{Content: s}}},
-				Object:  "chat.completion.chunk",
-			}
-			log.Debug().Msgf("Sending goroutine: %s", s)
-
-			responses <- resp
-			return true
-		})
-		close(responses)
-	}
-	return func(c *fiber.Ctx) error {
-		model, input, err := readInput(c, o.loader, true)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		config, input, err := readConfig(model, input, cm, o.loader, o.debug, o.threads, o.ctxSize, o.f16)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		log.Debug().Msgf("Parameter Config: %+v", config)
-
-		var predInput string
-
-		mess := []string{}
-		for _, i := range input.Messages {
-			var content string
-			r := config.Roles[i.Role]
-			if r != "" {
-				content = fmt.Sprint(r, " ", i.Content)
-			} else {
-				content = i.Content
-			}
-
-			mess = append(mess, content)
-		}
-
-		predInput = strings.Join(mess, "\n")
-
-		if input.Stream {
-			log.Debug().Msgf("Stream request received")
-			c.Context().SetContentType("text/event-stream")
-			//c.Response().Header.SetContentType(fiber.MIMETextHTMLCharsetUTF8)
-			//	c.Set("Content-Type", "text/event-stream")
-			c.Set("Cache-Control", "no-cache")
-			c.Set("Connection", "keep-alive")
-			c.Set("Transfer-Encoding", "chunked")
-		}
-
-		templateFile := config.Model
-
-		if config.TemplateConfig.Chat != "" {
-			templateFile = config.TemplateConfig.Chat
-		}
-
-		// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-		templatedInput, err := o.loader.TemplatePrefix(templateFile, struct {
-			Input string
-		}{Input: predInput})
-		if err == nil {
-			predInput = templatedInput
-			log.Debug().Msgf("Template found, input modified to: %s", predInput)
-		}
-
-		if input.Stream {
-			responses := make(chan OpenAIResponse)
-
-			go process(predInput, input, config, o.loader, responses)
-
-			c.Context().SetBodyStreamWriter(fasthttp.StreamWriter(func(w *bufio.Writer) {
-
-				for ev := range responses {
-					var buf bytes.Buffer
-					enc := json.NewEncoder(&buf)
-					enc.Encode(ev)
-
-					log.Debug().Msgf("Sending chunk: %s", buf.String())
-					fmt.Fprintf(w, "data: %v\n", buf.String())
-					w.Flush()
-				}
-
-				resp := &OpenAIResponse{
-					Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
-					Choices: []Choice{{FinishReason: "stop"}},
-				}
-				respData, _ := json.Marshal(resp)
-
-				w.WriteString(fmt.Sprintf("data: %s\n\n", respData))
-				w.WriteString("data: [DONE]\n\n")
-				w.Flush()
-			}))
-			return nil
-		}
-
-		result, err := ComputeChoices(predInput, input, config, o.loader, func(s string, c *[]Choice) {
-			*c = append(*c, Choice{Message: &Message{Role: "assistant", Content: s}})
-		}, nil)
-		if err != nil {
-			return err
-		}
-
-		resp := &OpenAIResponse{
-			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
-			Choices: result,
-			Object:  "chat.completion",
-		}
-		respData, _ := json.Marshal(resp)
-		log.Debug().Msgf("Response: %s", respData)
-
-		// Return the prediction in the response body
-		return c.JSON(resp)
-	}
-}
-
-func editEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		model, input, err := readInput(c, o.loader, true)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		config, input, err := readConfig(model, input, cm, o.loader, o.debug, o.threads, o.ctxSize, o.f16)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		log.Debug().Msgf("Parameter Config: %+v", config)
-
-		templateFile := config.Model
-
-		if config.TemplateConfig.Edit != "" {
-			templateFile = config.TemplateConfig.Edit
-		}
-
-		var result []Choice
-		for _, i := range config.InputStrings {
-			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-			templatedInput, err := o.loader.TemplatePrefix(templateFile, struct {
-				Input       string
-				Instruction string
-			}{Input: i})
-			if err == nil {
-				i = templatedInput
-				log.Debug().Msgf("Template found, input modified to: %s", i)
-			}
-
-			r, err := ComputeChoices(i, input, config, o.loader, func(s string, c *[]Choice) {
-				*c = append(*c, Choice{Text: s})
-			}, nil)
-			if err != nil {
-				return err
-			}
-
-			result = append(result, r...)
-		}
-
-		resp := &OpenAIResponse{
-			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
-			Choices: result,
-			Object:  "edit",
-		}
-
-		jsonResult, _ := json.Marshal(resp)
-		log.Debug().Msgf("Response: %s", jsonResult)
-
-		// Return the prediction in the response body
-		return c.JSON(resp)
-	}
-}
-
-// https://platform.openai.com/docs/api-reference/images/create
-
-/*
-*
-
-	curl http://localhost:8080/v1/images/generations \
-	  -H "Content-Type: application/json" \
-	  -d '{
-	    "prompt": "A cute baby sea otter",
-	    "n": 1,
-	    "size": "512x512"
-	  }'
-
-*
-*/
-func imageEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		m, input, err := readInput(c, o.loader, false)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		if m == "" {
-			m = model.StableDiffusionBackend
-		}
-		log.Debug().Msgf("Loading model: %+v", m)
-
-		config, input, err := readConfig(m, input, cm, o.loader, o.debug, 0, 0, false)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		log.Debug().Msgf("Parameter Config: %+v", config)
-
-		// XXX: Only stablediffusion is supported for now
-		if config.Backend == "" {
-			config.Backend = model.StableDiffusionBackend
-		}
-
-		sizeParts := strings.Split(input.Size, "x")
-		if len(sizeParts) != 2 {
-			return fmt.Errorf("Invalid value for 'size'")
-		}
-		width, err := strconv.Atoi(sizeParts[0])
-		if err != nil {
-			return fmt.Errorf("Invalid value for 'size'")
-		}
-		height, err := strconv.Atoi(sizeParts[1])
-		if err != nil {
-			return fmt.Errorf("Invalid value for 'size'")
-		}
-
-		b64JSON := false
-		if input.ResponseFormat == "b64_json" {
-			b64JSON = true
-		}
-
-		var result []Item
-		for _, i := range config.PromptStrings {
-			n := input.N
-			if input.N == 0 {
-				n = 1
-			}
-			for j := 0; j < n; j++ {
-				prompts := strings.Split(i, "|")
-				positive_prompt := prompts[0]
-				negative_prompt := ""
-				if len(prompts) > 1 {
-					negative_prompt = prompts[1]
-				}
-
-				mode := 0
-				step := 15
-
-				if input.Mode != 0 {
-					mode = input.Mode
-				}
-
-				if input.Step != 0 {
-					step = input.Step
-				}
-
-				tempDir := ""
-				if !b64JSON {
-					tempDir = o.imageDir
-				}
-				// Create a temporary file
-				outputFile, err := ioutil.TempFile(tempDir, "b64")
-				if err != nil {
-					return err
-				}
-				outputFile.Close()
-				output := outputFile.Name() + ".png"
-				// Rename the temporary file
-				err = os.Rename(outputFile.Name(), output)
-				if err != nil {
-					return err
-				}
-
-				baseURL := c.BaseURL()
-
-				fn, err := ImageGeneration(height, width, mode, step, input.Seed, positive_prompt, negative_prompt, output, o.loader, *config)
-				if err != nil {
-					return err
-				}
-				if err := fn(); err != nil {
-					return err
-				}
-
-				item := &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-images/" + base
-				}
-
-				result = append(result, *item)
-			}
-		}
-
-		resp := &OpenAIResponse{
-			Data: result,
-		}
-
-		jsonResult, _ := json.Marshal(resp)
-		log.Debug().Msgf("Response: %s", jsonResult)
-
-		// Return the prediction in the response body
-		return c.JSON(resp)
-	}
-}
-
-// https://platform.openai.com/docs/api-reference/audio/create
-func transcriptEndpoint(cm *ConfigMerger, o *Option) func(c *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		m, input, err := readInput(c, o.loader, false)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-
-		config, input, err := readConfig(m, input, cm, o.loader, o.debug, o.threads, o.ctxSize, o.f16)
-		if err != nil {
-			return fmt.Errorf("failed reading parameters from request:%w", err)
-		}
-		// retrieve the file data from the request
-		file, err := c.FormFile("file")
-		if err != nil {
-			return err
-		}
-		f, err := file.Open()
-		if err != nil {
-			return err
-		}
-		defer f.Close()
-
-		dir, err := os.MkdirTemp("", "whisper")
-
-		if err != nil {
-			return err
-		}
-		defer os.RemoveAll(dir)
-
-		dst := filepath.Join(dir, path.Base(file.Filename))
-		dstFile, err := os.Create(dst)
-		if err != nil {
-			return err
-		}
-
-		if _, err := io.Copy(dstFile, f); err != nil {
-			log.Debug().Msgf("Audio file copying error %+v - %+v - err %+v", file.Filename, dst, err)
-			return err
-		}
-
-		log.Debug().Msgf("Audio file copied to: %+v", dst)
-
-		whisperModel, err := o.loader.BackendLoader(model.WhisperBackend, config.Model, []llama.ModelOption{}, uint32(config.Threads))
-		if err != nil {
-			return err
-		}
-
-		if whisperModel == nil {
-			return fmt.Errorf("could not load whisper model")
-		}
-
-		w, ok := whisperModel.(whisper.Model)
-		if !ok {
-			return fmt.Errorf("loader returned non-whisper object")
-		}
-
-		tr, err := whisperutil.Transcript(w, dst, input.Language, uint(config.Threads))
-		if err != nil {
-			return err
-		}
-
-		log.Debug().Msgf("Trascribed: %+v", tr)
-		// TODO: handle different outputs here
-		return c.Status(http.StatusOK).JSON(fiber.Map{"text": tr})
-	}
-}
-
-func listModels(loader *model.ModelLoader, cm *ConfigMerger) func(ctx *fiber.Ctx) error {
-	return func(c *fiber.Ctx) error {
-		models, err := loader.ListModels()
-		if err != nil {
-			return err
-		}
-		var mm map[string]interface{} = map[string]interface{}{}
-
-		dataModels := []OpenAIModel{}
-		for _, m := range models {
-			mm[m] = nil
-			dataModels = append(dataModels, OpenAIModel{ID: m, Object: "model"})
-		}
-
-		for _, k := range cm.ListConfigs() {
-			if _, exists := mm[k]; !exists {
-				dataModels = append(dataModels, OpenAIModel{ID: k, Object: "model"})
-			}
-		}
-
-		return c.JSON(struct {
-			Object string        `json:"object"`
-			Data   []OpenAIModel `json:"data"`
-		}{
-			Object: "list",
-			Data:   dataModels,
-		})
-	}
-}

api/openai/chat.go

@@ -0,0 +1,385 @@
+package openai
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	"github.com/go-skynet/LocalAI/pkg/grammar"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/pkg/utils"
+	"github.com/gofiber/fiber/v2"
+	"github.com/google/uuid"
+	"github.com/rs/zerolog/log"
+	"github.com/valyala/fasthttp"
+)
+
+func ChatEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	emptyMessage := ""
+	id := uuid.New().String()
+	created := int(time.Now().Unix())
+
+	process := func(s string, req *schema.OpenAIRequest, config *config.Config, loader *model.ModelLoader, responses chan schema.OpenAIResponse) {
+		initialMessage := schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Choices: []schema.Choice{{Delta: &schema.Message{Role: "assistant", Content: &emptyMessage}}},
+			Object:  "chat.completion.chunk",
+		}
+		responses <- initialMessage
+
+		ComputeChoices(req, s, config, o, loader, func(s string, c *[]schema.Choice) {}, func(s string, usage backend.TokenUsage) bool {
+			resp := schema.OpenAIResponse{
+				ID:      id,
+				Created: created,
+				Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
+				Choices: []schema.Choice{{Delta: &schema.Message{Content: &s}, Index: 0}},
+				Object:  "chat.completion.chunk",
+				Usage: schema.OpenAIUsage{
+					PromptTokens:     usage.Prompt,
+					CompletionTokens: usage.Completion,
+					TotalTokens:      usage.Prompt + usage.Completion,
+				},
+			}
+
+			responses <- resp
+			return true
+		})
+		close(responses)
+	}
+	return func(c *fiber.Ctx) error {
+		processFunctions := false
+		funcs := grammar.Functions{}
+		modelFile, input, err := readInput(c, o, true)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		config, input, err := readConfig(modelFile, input, cm, o.Loader, o.Debug, o.Threads, o.ContextSize, o.F16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+		log.Debug().Msgf("Configuration read: %+v", config)
+
+		// Allow the user to set custom actions via config file
+		// to be "embedded" in each model
+		noActionName := "answer"
+		noActionDescription := "use this action to answer without performing any action"
+
+		if config.FunctionsConfig.NoActionFunctionName != "" {
+			noActionName = config.FunctionsConfig.NoActionFunctionName
+		}
+		if config.FunctionsConfig.NoActionDescriptionName != "" {
+			noActionDescription = config.FunctionsConfig.NoActionDescriptionName
+		}
+
+		// process functions if we have any defined or if we have a function call string
+		if len(input.Functions) > 0 && config.ShouldUseFunctions() {
+			log.Debug().Msgf("Response needs to process functions")
+
+			processFunctions = true
+
+			noActionGrammar := grammar.Function{
+				Name:        noActionName,
+				Description: noActionDescription,
+				Parameters: map[string]interface{}{
+					"properties": map[string]interface{}{
+						"message": map[string]interface{}{
+							"type":        "string",
+							"description": "The message to reply the user with",
+						}},
+				},
+			}
+
+			// Append the no action function
+			funcs = append(funcs, input.Functions...)
+			if !config.FunctionsConfig.DisableNoAction {
+				funcs = append(funcs, noActionGrammar)
+			}
+
+			// Force picking one of the functions by the request
+			if config.FunctionToCall() != "" {
+				funcs = funcs.Select(config.FunctionToCall())
+			}
+
+			// Update input grammar
+			jsStruct := funcs.ToJSONStructure()
+			config.Grammar = jsStruct.Grammar("")
+		} else if input.JSONFunctionGrammarObject != nil {
+			config.Grammar = input.JSONFunctionGrammarObject.Grammar("")
+		}
+
+		// functions are not supported in stream mode (yet?)
+		toStream := input.Stream && !processFunctions
+
+		log.Debug().Msgf("Parameters: %+v", config)
+
+		var predInput string
+
+		suppressConfigSystemPrompt := false
+		mess := []string{}
+		for messageIndex, i := range input.Messages {
+			var content string
+			role := i.Role
+
+			// if function call, we might want to customize the role so we can display better that the "assistant called a json action"
+			// if an "assistant_function_call" role is defined, we use it, otherwise we use the role that is passed by in the request
+			if i.FunctionCall != nil && i.Role == "assistant" {
+				roleFn := "assistant_function_call"
+				r := config.Roles[roleFn]
+				if r != "" {
+					role = roleFn
+				}
+			}
+			r := config.Roles[role]
+			contentExists := i.Content != nil && *i.Content != ""
+			// First attempt to populate content via a chat message specific template
+			if config.TemplateConfig.ChatMessage != "" {
+				chatMessageData := model.ChatMessageTemplateData{
+					SystemPrompt: config.SystemPrompt,
+					Role:         r,
+					RoleName:     role,
+					Content:      *i.Content,
+					MessageIndex: messageIndex,
+				}
+				templatedChatMessage, err := o.Loader.EvaluateTemplateForChatMessage(config.TemplateConfig.ChatMessage, chatMessageData)
+				if err != nil {
+					log.Error().Msgf("error processing message %+v using template \"%s\": %v. Skipping!", chatMessageData, config.TemplateConfig.ChatMessage, err)
+				} else {
+					if templatedChatMessage == "" {
+						log.Warn().Msgf("template \"%s\" produced blank output for %+v. Skipping!", config.TemplateConfig.ChatMessage, chatMessageData)
+						continue // TODO: This continue is here intentionally to skip over the line `mess = append(mess, content)` below, and to prevent the sprintf
+					}
+					log.Debug().Msgf("templated message for chat: %s", templatedChatMessage)
+					content = templatedChatMessage
+				}
+			}
+			// If this model doesn't have such a template, or if that template fails to return a value, template at the message level.
+			if content == "" {
+				if r != "" {
+					if contentExists {
+						content = fmt.Sprint(r, " ", *i.Content)
+					}
+					if i.FunctionCall != nil {
+						j, err := json.Marshal(i.FunctionCall)
+						if err == nil {
+							if contentExists {
+								content += "\n" + fmt.Sprint(r, " ", string(j))
+							} else {
+								content = fmt.Sprint(r, " ", string(j))
+							}
+						}
+					}
+				} else {
+					if contentExists {
+						content = fmt.Sprint(*i.Content)
+					}
+					if i.FunctionCall != nil {
+						j, err := json.Marshal(i.FunctionCall)
+						if err == nil {
+							if contentExists {
+								content += "\n" + string(j)
+							} else {
+								content = string(j)
+							}
+						}
+					}
+				}
+				// Special Handling: System. We care if it was printed at all, not the r branch, so check seperately
+				if contentExists && role == "system" {
+					suppressConfigSystemPrompt = true
+				}
+			}
+
+			mess = append(mess, content)
+		}
+
+		predInput = strings.Join(mess, "\n")
+		log.Debug().Msgf("Prompt (before templating): %s", predInput)
+
+		if toStream {
+			log.Debug().Msgf("Stream request received")
+			c.Context().SetContentType("text/event-stream")
+			//c.Response().Header.SetContentType(fiber.MIMETextHTMLCharsetUTF8)
+			//	c.Set("Content-Type", "text/event-stream")
+			c.Set("Cache-Control", "no-cache")
+			c.Set("Connection", "keep-alive")
+			c.Set("Transfer-Encoding", "chunked")
+		}
+
+		templateFile := config.Model
+
+		if config.TemplateConfig.Chat != "" && !processFunctions {
+			templateFile = config.TemplateConfig.Chat
+		}
+
+		if config.TemplateConfig.Functions != "" && processFunctions {
+			templateFile = config.TemplateConfig.Functions
+		}
+
+		// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
+		templatedInput, err := o.Loader.EvaluateTemplateForPrompt(model.ChatPromptTemplate, templateFile, model.PromptTemplateData{
+			SystemPrompt:         config.SystemPrompt,
+			SuppressSystemPrompt: suppressConfigSystemPrompt,
+			Input:                predInput,
+			Functions:            funcs,
+		})
+		if err == nil {
+			predInput = templatedInput
+			log.Debug().Msgf("Template found, input modified to: %s", predInput)
+		} else {
+			log.Debug().Msgf("Template failed loading: %s", err.Error())
+		}
+
+		log.Debug().Msgf("Prompt (after templating): %s", predInput)
+		if processFunctions {
+			log.Debug().Msgf("Grammar: %+v", config.Grammar)
+		}
+
+		if toStream {
+			responses := make(chan schema.OpenAIResponse)
+
+			go process(predInput, input, config, o.Loader, responses)
+
+			c.Context().SetBodyStreamWriter(fasthttp.StreamWriter(func(w *bufio.Writer) {
+
+				usage := &schema.OpenAIUsage{}
+
+				for ev := range responses {
+					usage = &ev.Usage // Copy a pointer to the latest usage chunk so that the stop message can reference it
+					var buf bytes.Buffer
+					enc := json.NewEncoder(&buf)
+					enc.Encode(ev)
+					log.Debug().Msgf("Sending chunk: %s", buf.String())
+					_, err := fmt.Fprintf(w, "data: %v\n", buf.String())
+					if err != nil {
+						log.Debug().Msgf("Sending chunk failed: %v", err)
+						input.Cancel()
+						break
+					}
+					w.Flush()
+				}
+
+				resp := &schema.OpenAIResponse{
+					ID:      id,
+					Created: created,
+					Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+					Choices: []schema.Choice{
+						{
+							FinishReason: "stop",
+							Index:        0,
+							Delta:        &schema.Message{Content: &emptyMessage},
+						}},
+					Object: "chat.completion.chunk",
+					Usage:  *usage,
+				}
+				respData, _ := json.Marshal(resp)
+
+				w.WriteString(fmt.Sprintf("data: %s\n\n", respData))
+				w.WriteString("data: [DONE]\n\n")
+				w.Flush()
+			}))
+			return nil
+		}
+
+		result, tokenUsage, err := ComputeChoices(input, predInput, config, o, o.Loader, func(s string, c *[]schema.Choice) {
+			if processFunctions {
+				// As we have to change the result before processing, we can't stream the answer (yet?)
+				ss := map[string]interface{}{}
+				// This prevent newlines to break JSON parsing for clients
+				s = utils.EscapeNewLines(s)
+				json.Unmarshal([]byte(s), &ss)
+				log.Debug().Msgf("Function return: %s %+v", s, ss)
+
+				// The grammar defines the function name as "function", while OpenAI returns "name"
+				func_name := ss["function"]
+				// Similarly, while here arguments is a map[string]interface{}, OpenAI actually want a stringified object
+				args := ss["arguments"] // arguments needs to be a string, but we return an object from the grammar result (TODO: fix)
+				d, _ := json.Marshal(args)
+
+				ss["arguments"] = string(d)
+				ss["name"] = func_name
+
+				// if do nothing, reply with a message
+				if func_name == noActionName {
+					log.Debug().Msgf("nothing to do, computing a reply")
+
+					// If there is a message that the LLM already sends as part of the JSON reply, use it
+					arguments := map[string]interface{}{}
+					json.Unmarshal([]byte(d), &arguments)
+					m, exists := arguments["message"]
+					if exists {
+						switch message := m.(type) {
+						case string:
+							if message != "" {
+								log.Debug().Msgf("Reply received from LLM: %s", message)
+								message = backend.Finetune(*config, predInput, message)
+								log.Debug().Msgf("Reply received from LLM(finetuned): %s", message)
+
+								*c = append(*c, schema.Choice{Message: &schema.Message{Role: "assistant", Content: &message}})
+								return
+							}
+						}
+					}
+
+					log.Debug().Msgf("No action received from LLM, without a message, computing a reply")
+					// Otherwise ask the LLM to understand the JSON output and the context, and return a message
+					// Note: This costs (in term of CPU) another computation
+					config.Grammar = ""
+					predFunc, err := backend.ModelInference(input.Context, predInput, o.Loader, *config, o, nil)
+					if err != nil {
+						log.Error().Msgf("inference error: %s", err.Error())
+						return
+					}
+
+					prediction, err := predFunc()
+					if err != nil {
+						log.Error().Msgf("inference error: %s", err.Error())
+						return
+					}
+
+					fineTunedResponse := backend.Finetune(*config, predInput, prediction.Response)
+					*c = append(*c, schema.Choice{Message: &schema.Message{Role: "assistant", Content: &fineTunedResponse}})
+				} else {
+					// otherwise reply with the function call
+					*c = append(*c, schema.Choice{
+						FinishReason: "function_call",
+						Message:      &schema.Message{Role: "assistant", FunctionCall: ss},
+					})
+				}
+
+				return
+			}
+			*c = append(*c, schema.Choice{FinishReason: "stop", Index: 0, Message: &schema.Message{Role: "assistant", Content: &s}})
+		}, nil)
+		if err != nil {
+			return err
+		}
+
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Choices: result,
+			Object:  "chat.completion",
+			Usage: schema.OpenAIUsage{
+				PromptTokens:     tokenUsage.Prompt,
+				CompletionTokens: tokenUsage.Completion,
+				TotalTokens:      tokenUsage.Prompt + tokenUsage.Completion,
+			},
+		}
+		respData, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", respData)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

api/openai/completion.go

@@ -0,0 +1,186 @@
+package openai
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/gofiber/fiber/v2"
+	"github.com/google/uuid"
+	"github.com/rs/zerolog/log"
+	"github.com/valyala/fasthttp"
+)
+
+// https://platform.openai.com/docs/api-reference/completions
+func CompletionEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	id := uuid.New().String()
+	created := int(time.Now().Unix())
+
+	process := func(s string, req *schema.OpenAIRequest, config *config.Config, loader *model.ModelLoader, responses chan schema.OpenAIResponse) {
+		ComputeChoices(req, s, config, o, loader, func(s string, c *[]schema.Choice) {}, func(s string, usage backend.TokenUsage) bool {
+			resp := schema.OpenAIResponse{
+				ID:      id,
+				Created: created,
+				Model:   req.Model, // we have to return what the user sent here, due to OpenAI spec.
+				Choices: []schema.Choice{
+					{
+						Index: 0,
+						Text:  s,
+					},
+				},
+				Object: "text_completion",
+				Usage: schema.OpenAIUsage{
+					PromptTokens:     usage.Prompt,
+					CompletionTokens: usage.Completion,
+					TotalTokens:      usage.Prompt + usage.Completion,
+				},
+			}
+			log.Debug().Msgf("Sending goroutine: %s", s)
+
+			responses <- resp
+			return true
+		})
+		close(responses)
+	}
+
+	return func(c *fiber.Ctx) error {
+		modelFile, input, err := readInput(c, o, true)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		log.Debug().Msgf("`input`: %+v", input)
+
+		config, input, err := readConfig(modelFile, input, cm, o.Loader, o.Debug, o.Threads, o.ContextSize, o.F16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		log.Debug().Msgf("Parameter Config: %+v", config)
+
+		if input.Stream {
+			log.Debug().Msgf("Stream request received")
+			c.Context().SetContentType("text/event-stream")
+			//c.Response().Header.SetContentType(fiber.MIMETextHTMLCharsetUTF8)
+			//c.Set("Content-Type", "text/event-stream")
+			c.Set("Cache-Control", "no-cache")
+			c.Set("Connection", "keep-alive")
+			c.Set("Transfer-Encoding", "chunked")
+		}
+
+		templateFile := config.Model
+
+		if config.TemplateConfig.Completion != "" {
+			templateFile = config.TemplateConfig.Completion
+		}
+
+		if input.Stream {
+			if len(config.PromptStrings) > 1 {
+				return errors.New("cannot handle more than 1 `PromptStrings` when Streaming")
+			}
+
+			predInput := config.PromptStrings[0]
+
+			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
+			templatedInput, err := o.Loader.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
+				Input: predInput,
+			})
+			if err == nil {
+				predInput = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", predInput)
+			}
+
+			responses := make(chan schema.OpenAIResponse)
+
+			go process(predInput, input, config, o.Loader, responses)
+
+			c.Context().SetBodyStreamWriter(fasthttp.StreamWriter(func(w *bufio.Writer) {
+
+				for ev := range responses {
+					var buf bytes.Buffer
+					enc := json.NewEncoder(&buf)
+					enc.Encode(ev)
+
+					log.Debug().Msgf("Sending chunk: %s", buf.String())
+					fmt.Fprintf(w, "data: %v\n", buf.String())
+					w.Flush()
+				}
+
+				resp := &schema.OpenAIResponse{
+					ID:      id,
+					Created: created,
+					Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+					Choices: []schema.Choice{
+						{
+							Index:        0,
+							FinishReason: "stop",
+						},
+					},
+					Object: "text_completion",
+				}
+				respData, _ := json.Marshal(resp)
+
+				w.WriteString(fmt.Sprintf("data: %s\n\n", respData))
+				w.WriteString("data: [DONE]\n\n")
+				w.Flush()
+			}))
+			return nil
+		}
+
+		var result []schema.Choice
+
+		totalTokenUsage := backend.TokenUsage{}
+
+		for k, i := range config.PromptStrings {
+			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
+			templatedInput, err := o.Loader.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
+				SystemPrompt: config.SystemPrompt,
+				Input:        i,
+			})
+			if err == nil {
+				i = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", i)
+			}
+
+			r, tokenUsage, err := ComputeChoices(
+				input, i, config, o, o.Loader, func(s string, c *[]schema.Choice) {
+					*c = append(*c, schema.Choice{Text: s, FinishReason: "stop", Index: k})
+				}, nil)
+			if err != nil {
+				return err
+			}
+
+			totalTokenUsage.Prompt += tokenUsage.Prompt
+			totalTokenUsage.Completion += tokenUsage.Completion
+
+			result = append(result, r...)
+		}
+
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Choices: result,
+			Object:  "text_completion",
+			Usage: schema.OpenAIUsage{
+				PromptTokens:     totalTokenUsage.Prompt,
+				CompletionTokens: totalTokenUsage.Completion,
+				TotalTokens:      totalTokenUsage.Prompt + totalTokenUsage.Completion,
+			},
+		}
+
+		jsonResult, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", jsonResult)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

api/openai/edit.go

@@ -0,0 +1,88 @@
+package openai
+
+import (
+	"encoding/json"
+	"fmt"
+	"time"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/gofiber/fiber/v2"
+	"github.com/google/uuid"
+
+	"github.com/rs/zerolog/log"
+)
+
+func EditEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		modelFile, input, err := readInput(c, o, true)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		config, input, err := readConfig(modelFile, input, cm, o.Loader, o.Debug, o.Threads, o.ContextSize, o.F16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		log.Debug().Msgf("Parameter Config: %+v", config)
+
+		templateFile := config.Model
+
+		if config.TemplateConfig.Edit != "" {
+			templateFile = config.TemplateConfig.Edit
+		}
+
+		var result []schema.Choice
+		totalTokenUsage := backend.TokenUsage{}
+
+		for _, i := range config.InputStrings {
+			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
+			templatedInput, err := o.Loader.EvaluateTemplateForPrompt(model.EditPromptTemplate, templateFile, model.PromptTemplateData{
+				Input:        i,
+				Instruction:  input.Instruction,
+				SystemPrompt: config.SystemPrompt,
+			})
+			if err == nil {
+				i = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", i)
+			}
+
+			r, tokenUsage, err := ComputeChoices(input, i, config, o, o.Loader, func(s string, c *[]schema.Choice) {
+				*c = append(*c, schema.Choice{Text: s})
+			}, nil)
+			if err != nil {
+				return err
+			}
+
+			totalTokenUsage.Prompt += tokenUsage.Prompt
+			totalTokenUsage.Completion += tokenUsage.Completion
+
+			result = append(result, r...)
+		}
+
+		id := uuid.New().String()
+		created := int(time.Now().Unix())
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Choices: result,
+			Object:  "edit",
+			Usage: schema.OpenAIUsage{
+				PromptTokens:     totalTokenUsage.Prompt,
+				CompletionTokens: totalTokenUsage.Completion,
+				TotalTokens:      totalTokenUsage.Prompt + totalTokenUsage.Completion,
+			},
+		}
+
+		jsonResult, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", jsonResult)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

api/openai/embeddings.go

@@ -0,0 +1,78 @@
+package openai
+
+import (
+	"encoding/json"
+	"fmt"
+	"time"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/schema"
+	"github.com/google/uuid"
+
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/gofiber/fiber/v2"
+	"github.com/rs/zerolog/log"
+)
+
+// https://platform.openai.com/docs/api-reference/embeddings
+func EmbeddingsEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		model, input, err := readInput(c, o, true)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		config, input, err := readConfig(model, input, cm, o.Loader, o.Debug, o.Threads, o.ContextSize, o.F16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		log.Debug().Msgf("Parameter Config: %+v", config)
+		items := []schema.Item{}
+
+		for i, s := range config.InputToken {
+			// get the model function to call for the result
+			embedFn, err := backend.ModelEmbedding("", s, o.Loader, *config, o)
+			if err != nil {
+				return err
+			}
+
+			embeddings, err := embedFn()
+			if err != nil {
+				return err
+			}
+			items = append(items, schema.Item{Embedding: embeddings, Index: i, Object: "embedding"})
+		}
+
+		for i, s := range config.InputStrings {
+			// get the model function to call for the result
+			embedFn, err := backend.ModelEmbedding(s, []int{}, o.Loader, *config, o)
+			if err != nil {
+				return err
+			}
+
+			embeddings, err := embedFn()
+			if err != nil {
+				return err
+			}
+			items = append(items, schema.Item{Embedding: embeddings, Index: i, Object: "embedding"})
+		}
+
+		id := uuid.New().String()
+		created := int(time.Now().Unix())
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Model:   input.Model, // we have to return what the user sent here, due to OpenAI spec.
+			Data:    items,
+			Object:  "list",
+		}
+
+		jsonResult, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", jsonResult)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

api/openai/image.go

@@ -0,0 +1,194 @@
+package openai
+
+import (
+	"bufio"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/go-skynet/LocalAI/api/schema"
+	"github.com/google/uuid"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/gofiber/fiber/v2"
+	"github.com/rs/zerolog/log"
+)
+
+// https://platform.openai.com/docs/api-reference/images/create
+
+/*
+*
+
+	curl http://localhost:8080/v1/images/generations \
+	  -H "Content-Type: application/json" \
+	  -d '{
+	    "prompt": "A cute baby sea otter",
+	    "n": 1,
+	    "size": "512x512"
+	  }'
+
+*
+*/
+func ImageEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		m, input, err := readInput(c, o, false)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		if m == "" {
+			m = model.StableDiffusionBackend
+		}
+		log.Debug().Msgf("Loading model: %+v", m)
+
+		config, input, err := readConfig(m, input, cm, o.Loader, o.Debug, 0, 0, false)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		src := ""
+		if input.File != "" {
+			//base 64 decode the file and write it somewhere
+			// that we will cleanup
+			decoded, err := base64.StdEncoding.DecodeString(input.File)
+			if err != nil {
+				return err
+			}
+			// Create a temporary file
+			outputFile, err := os.CreateTemp(o.ImageDir, "b64")
+			if err != nil {
+				return err
+			}
+			// write the base64 result
+			writer := bufio.NewWriter(outputFile)
+			_, err = writer.Write(decoded)
+			if err != nil {
+				outputFile.Close()
+				return err
+			}
+			outputFile.Close()
+			src = outputFile.Name()
+			defer os.RemoveAll(src)
+		}
+
+		log.Debug().Msgf("Parameter Config: %+v", config)
+
+		// XXX: Only stablediffusion is supported for now
+		if config.Backend == "" {
+			config.Backend = model.StableDiffusionBackend
+		}
+
+		sizeParts := strings.Split(input.Size, "x")
+		if len(sizeParts) != 2 {
+			return fmt.Errorf("Invalid value for 'size'")
+		}
+		width, err := strconv.Atoi(sizeParts[0])
+		if err != nil {
+			return fmt.Errorf("Invalid value for 'size'")
+		}
+		height, err := strconv.Atoi(sizeParts[1])
+		if err != nil {
+			return fmt.Errorf("Invalid value for 'size'")
+		}
+
+		b64JSON := false
+		if input.ResponseFormat == "b64_json" {
+			b64JSON = true
+		}
+		// src and clip_skip
+		var result []schema.Item
+		for _, i := range config.PromptStrings {
+			n := input.N
+			if input.N == 0 {
+				n = 1
+			}
+			for j := 0; j < n; j++ {
+				prompts := strings.Split(i, "|")
+				positive_prompt := prompts[0]
+				negative_prompt := ""
+				if len(prompts) > 1 {
+					negative_prompt = prompts[1]
+				}
+
+				mode := 0
+				step := config.Step
+				if step == 0 {
+					step = 15
+				}
+
+				if input.Mode != 0 {
+					mode = input.Mode
+				}
+
+				if input.Step != 0 {
+					step = input.Step
+				}
+
+				tempDir := ""
+				if !b64JSON {
+					tempDir = o.ImageDir
+				}
+				// Create a temporary file
+				outputFile, err := os.CreateTemp(tempDir, "b64")
+				if err != nil {
+					return err
+				}
+				outputFile.Close()
+				output := outputFile.Name() + ".png"
+				// Rename the temporary file
+				err = os.Rename(outputFile.Name(), output)
+				if err != nil {
+					return err
+				}
+
+				baseURL := c.BaseURL()
+
+				fn, err := backend.ImageGeneration(height, width, mode, step, input.Seed, positive_prompt, negative_prompt, src, output, o.Loader, *config, o)
+				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-images/" + base
+				}
+
+				result = append(result, *item)
+			}
+		}
+
+		id := uuid.New().String()
+		created := int(time.Now().Unix())
+		resp := &schema.OpenAIResponse{
+			ID:      id,
+			Created: created,
+			Data:    result,
+		}
+
+		jsonResult, _ := json.Marshal(resp)
+		log.Debug().Msgf("Response: %s", jsonResult)
+
+		// Return the prediction in the response body
+		return c.JSON(resp)
+	}
+}

api/openai/inference.go

@@ -0,0 +1,50 @@
+package openai
+
+import (
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+)
+
+func ComputeChoices(
+	req *schema.OpenAIRequest,
+	predInput string,
+	config *config.Config,
+	o *options.Option,
+	loader *model.ModelLoader,
+	cb func(string, *[]schema.Choice),
+	tokenCallback func(string, backend.TokenUsage) bool) ([]schema.Choice, backend.TokenUsage, error) {
+	n := req.N // number of completions to return
+	result := []schema.Choice{}
+
+	if n == 0 {
+		n = 1
+	}
+
+	// get the model function to call for the result
+	predFunc, err := backend.ModelInference(req.Context, predInput, loader, *config, o, tokenCallback)
+	if err != nil {
+		return result, backend.TokenUsage{}, err
+	}
+
+	tokenUsage := backend.TokenUsage{}
+
+	for i := 0; i < n; i++ {
+		prediction, err := predFunc()
+		if err != nil {
+			return result, backend.TokenUsage{}, err
+		}
+
+		tokenUsage.Prompt += prediction.Usage.Prompt
+		tokenUsage.Completion += prediction.Usage.Completion
+
+		finetunedResponse := backend.Finetune(*config, predInput, prediction.Response)
+		cb(finetunedResponse, &result)
+
+		//result = append(result, Choice{Text: prediction})
+
+	}
+	return result, tokenUsage, err
+}

api/openai/list.go

@@ -0,0 +1,69 @@
+package openai
+
+import (
+	"regexp"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/schema"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/gofiber/fiber/v2"
+)
+
+func ListModelsEndpoint(loader *model.ModelLoader, cm *config.ConfigLoader) func(ctx *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		models, err := loader.ListModels()
+		if err != nil {
+			return err
+		}
+		var mm map[string]interface{} = map[string]interface{}{}
+
+		dataModels := []schema.OpenAIModel{}
+
+		var filterFn func(name string) bool
+		filter := c.Query("filter")
+
+		// If filter is not specified, do not filter the list by model name
+		if filter == "" {
+			filterFn = func(_ string) bool { return true }
+		} else {
+			// If filter _IS_ specified, we compile it to a regex which is used to create the filterFn
+			rxp, err := regexp.Compile(filter)
+			if err != nil {
+				return err
+			}
+			filterFn = func(name string) bool {
+				return rxp.MatchString(name)
+			}
+		}
+
+		// By default, exclude any loose files that are already referenced by a configuration file.
+		excludeConfigured := c.QueryBool("excludeConfigured", true)
+
+		// Start with the known configurations
+		for _, c := range cm.GetAllConfigs() {
+			if excludeConfigured {
+				mm[c.Model] = nil
+			}
+
+			if filterFn(c.Name) {
+				dataModels = append(dataModels, schema.OpenAIModel{ID: c.Name, Object: "model"})
+			}
+		}
+
+		// Then iterate through the loose files:
+		for _, m := range models {
+			// And only adds them if they shouldn't be skipped.
+			if _, exists := mm[m]; !exists && filterFn(m) {
+				dataModels = append(dataModels, schema.OpenAIModel{ID: m, Object: "model"})
+			}
+		}
+
+		return c.JSON(struct {
+			Object string               `json:"object"`
+			Data   []schema.OpenAIModel `json:"data"`
+		}{
+			Object: "list",
+			Data:   dataModels,
+		})
+	}
+}

api/openai/request.go

@@ -0,0 +1,273 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+	options "github.com/go-skynet/LocalAI/api/options"
+	"github.com/go-skynet/LocalAI/api/schema"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/gofiber/fiber/v2"
+	"github.com/rs/zerolog/log"
+)
+
+func readInput(c *fiber.Ctx, o *options.Option, randomModel bool) (string, *schema.OpenAIRequest, error) {
+	loader := o.Loader
+	input := new(schema.OpenAIRequest)
+	ctx, cancel := context.WithCancel(o.Context)
+	input.Context = ctx
+	input.Cancel = cancel
+	// Get input data from the request body
+	if err := c.BodyParser(input); err != nil {
+		return "", nil, err
+	}
+
+	modelFile := input.Model
+
+	if c.Params("model") != "" {
+		modelFile = c.Params("model")
+	}
+
+	received, _ := json.Marshal(input)
+
+	log.Debug().Msgf("Request received: %s", string(received))
+
+	// Set model from bearer token, if available
+	bearer := strings.TrimLeft(c.Get("authorization"), "Bearer ")
+	bearerExists := bearer != "" && loader.ExistsInModelPath(bearer)
+
+	// If no model was specified, take the first available
+	if modelFile == "" && !bearerExists && randomModel {
+		models, _ := loader.ListModels()
+		if len(models) > 0 {
+			modelFile = models[0]
+			log.Debug().Msgf("No model specified, using: %s", modelFile)
+		} else {
+			log.Debug().Msgf("No model specified, returning error")
+			return "", nil, fmt.Errorf("no model specified")
+		}
+	}
+
+	// If a model is found in bearer token takes precedence
+	if bearerExists {
+		log.Debug().Msgf("Using model from bearer token: %s", bearer)
+		modelFile = bearer
+	}
+	return modelFile, input, nil
+}
+
+func updateConfig(config *config.Config, input *schema.OpenAIRequest) {
+	if input.Echo {
+		config.Echo = input.Echo
+	}
+	if input.TopK != 0 {
+		config.TopK = input.TopK
+	}
+	if input.TopP != 0 {
+		config.TopP = input.TopP
+	}
+
+	if input.Backend != "" {
+		config.Backend = input.Backend
+	}
+
+	if input.ClipSkip != 0 {
+		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
+	}
+
+	if input.RopeFreqBase != 0 {
+		config.RopeFreqBase = input.RopeFreqBase
+	}
+
+	if input.RopeFreqScale != 0 {
+		config.RopeFreqScale = input.RopeFreqScale
+	}
+
+	if input.Grammar != "" {
+		config.Grammar = input.Grammar
+	}
+
+	if input.Temperature != 0 {
+		config.Temperature = input.Temperature
+	}
+
+	if input.Maxtokens != 0 {
+		config.Maxtokens = input.Maxtokens
+	}
+
+	switch stop := input.Stop.(type) {
+	case string:
+		if stop != "" {
+			config.StopWords = append(config.StopWords, stop)
+		}
+	case []interface{}:
+		for _, pp := range stop {
+			if s, ok := pp.(string); ok {
+				config.StopWords = append(config.StopWords, s)
+			}
+		}
+	}
+
+	if input.RepeatPenalty != 0 {
+		config.RepeatPenalty = input.RepeatPenalty
+	}
+
+	if input.Keep != 0 {
+		config.Keep = input.Keep
+	}
+
+	if input.Batch != 0 {
+		config.Batch = input.Batch
+	}
+
+	if input.F16 {
+		config.F16 = input.F16
+	}
+
+	if input.IgnoreEOS {
+		config.IgnoreEOS = input.IgnoreEOS
+	}
+
+	if input.Seed != 0 {
+		config.Seed = input.Seed
+	}
+
+	if input.Mirostat != 0 {
+		config.LLMConfig.Mirostat = input.Mirostat
+	}
+
+	if input.MirostatETA != 0 {
+		config.LLMConfig.MirostatETA = input.MirostatETA
+	}
+
+	if input.MirostatTAU != 0 {
+		config.LLMConfig.MirostatTAU = input.MirostatTAU
+	}
+
+	if input.TypicalP != 0 {
+		config.TypicalP = input.TypicalP
+	}
+
+	switch inputs := input.Input.(type) {
+	case string:
+		if inputs != "" {
+			config.InputStrings = append(config.InputStrings, inputs)
+		}
+	case []interface{}:
+		for _, pp := range inputs {
+			switch i := pp.(type) {
+			case string:
+				config.InputStrings = append(config.InputStrings, i)
+			case []interface{}:
+				tokens := []int{}
+				for _, ii := range i {
+					tokens = append(tokens, int(ii.(float64)))
+				}
+				config.InputToken = append(config.InputToken, tokens)
+			}
+		}
+	}
+
+	// Can be either a string or an object
+	switch fnc := input.FunctionCall.(type) {
+	case string:
+		if fnc != "" {
+			config.SetFunctionCallString(fnc)
+		}
+	case map[string]interface{}:
+		var name string
+		n, exists := fnc["name"]
+		if exists {
+			nn, e := n.(string)
+			if e {
+				name = nn
+			}
+		}
+		config.SetFunctionCallNameString(name)
+	}
+
+	switch p := input.Prompt.(type) {
+	case string:
+		config.PromptStrings = append(config.PromptStrings, p)
+	case []interface{}:
+		for _, pp := range p {
+			if s, ok := pp.(string); ok {
+				config.PromptStrings = append(config.PromptStrings, s)
+			}
+		}
+	}
+}
+
+func readConfig(modelFile string, input *schema.OpenAIRequest, cm *config.ConfigLoader, loader *model.ModelLoader, debug bool, threads, ctx int, f16 bool) (*config.Config, *schema.OpenAIRequest, error) {
+	// Load a config file if present after the model name
+	modelConfig := filepath.Join(loader.ModelPath, modelFile+".yaml")
+
+	var cfg *config.Config
+
+	defaults := func() {
+		cfg = config.DefaultConfig(modelFile)
+		cfg.ContextSize = ctx
+		cfg.Threads = threads
+		cfg.F16 = f16
+		cfg.Debug = debug
+	}
+
+	cfgExisting, exists := cm.GetConfig(modelFile)
+	if !exists {
+		if _, err := os.Stat(modelConfig); err == nil {
+			if err := cm.LoadConfig(modelConfig); err != nil {
+				return nil, nil, fmt.Errorf("failed loading model config (%s) %s", modelConfig, err.Error())
+			}
+			cfgExisting, exists = cm.GetConfig(modelFile)
+			if exists {
+				cfg = &cfgExisting
+			} else {
+				defaults()
+			}
+		} else {
+			defaults()
+		}
+	} else {
+		cfg = &cfgExisting
+	}
+
+	// Set the parameters for the language model prediction
+	updateConfig(cfg, input)
+
+	// Don't allow 0 as setting
+	if cfg.Threads == 0 {
+		if threads != 0 {
+			cfg.Threads = threads
+		} else {
+			cfg.Threads = 4
+		}
+	}
+
+	// Enforce debug flag if passed from CLI
+	if debug {
+		cfg.Debug = true
+	}
+
+	return cfg, input, nil
+}

api/openai/transcription.go

@@ -0,0 +1,71 @@
+package openai
+
+import (
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"path"
+	"path/filepath"
+
+	"github.com/go-skynet/LocalAI/api/backend"
+	config "github.com/go-skynet/LocalAI/api/config"
+	"github.com/go-skynet/LocalAI/api/options"
+
+	"github.com/gofiber/fiber/v2"
+	"github.com/rs/zerolog/log"
+)
+
+// https://platform.openai.com/docs/api-reference/audio/create
+func TranscriptEndpoint(cm *config.ConfigLoader, o *options.Option) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+		m, input, err := readInput(c, o, false)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+
+		config, input, err := readConfig(m, input, cm, o.Loader, o.Debug, o.Threads, o.ContextSize, o.F16)
+		if err != nil {
+			return fmt.Errorf("failed reading parameters from request:%w", err)
+		}
+		// retrieve the file data from the request
+		file, err := c.FormFile("file")
+		if err != nil {
+			return err
+		}
+		f, err := file.Open()
+		if err != nil {
+			return err
+		}
+		defer f.Close()
+
+		dir, err := os.MkdirTemp("", "whisper")
+
+		if err != nil {
+			return err
+		}
+		defer os.RemoveAll(dir)
+
+		dst := filepath.Join(dir, path.Base(file.Filename))
+		dstFile, err := os.Create(dst)
+		if err != nil {
+			return err
+		}
+
+		if _, err := io.Copy(dstFile, f); err != nil {
+			log.Debug().Msgf("Audio file copying error %+v - %+v - err %+v", file.Filename, dst, err)
+			return err
+		}
+
+		log.Debug().Msgf("Audio file copied to: %+v", dst)
+
+		tr, err := backend.ModelTranscription(dst, input.Language, o.Loader, *config, o)
+		if err != nil {
+			return err
+		}
+
+		log.Debug().Msgf("Trascribed: %+v", tr)
+		// TODO: handle different outputs here
+		return c.Status(http.StatusOK).JSON(tr)
+	}
+}

api/options.go

@@ -1,108 +0,0 @@
-package api
-
-import (
-	"context"
-
-	model "github.com/go-skynet/LocalAI/pkg/model"
-)
-
-type Option struct {
-	context                         context.Context
-	configFile                      string
-	loader                          *model.ModelLoader
-	uploadLimitMB, threads, ctxSize int
-	f16                             bool
-	debug, disableMessage           bool
-	imageDir                        string
-	cors                            bool
-	corsAllowOrigins                string
-}
-
-type AppOption func(*Option)
-
-func newOptions(o ...AppOption) *Option {
-	opt := &Option{
-		context:        context.Background(),
-		uploadLimitMB:  15,
-		threads:        1,
-		ctxSize:        512,
-		debug:          true,
-		disableMessage: true,
-	}
-	for _, oo := range o {
-		oo(opt)
-	}
-	return opt
-}
-
-func WithCors(b bool) AppOption {
-	return func(o *Option) {
-		o.cors = b
-	}
-}
-
-func WithCorsAllowOrigins(b string) AppOption {
-	return func(o *Option) {
-		o.corsAllowOrigins = b
-	}
-}
-
-func WithContext(ctx context.Context) AppOption {
-	return func(o *Option) {
-		o.context = ctx
-	}
-}
-
-func WithConfigFile(configFile string) AppOption {
-	return func(o *Option) {
-		o.configFile = configFile
-	}
-}
-
-func WithModelLoader(loader *model.ModelLoader) AppOption {
-	return func(o *Option) {
-		o.loader = loader
-	}
-}
-
-func WithUploadLimitMB(limit int) AppOption {
-	return func(o *Option) {
-		o.uploadLimitMB = limit
-	}
-}
-
-func WithThreads(threads int) AppOption {
-	return func(o *Option) {
-		o.threads = threads
-	}
-}
-
-func WithContextSize(ctxSize int) AppOption {
-	return func(o *Option) {
-		o.ctxSize = ctxSize
-	}
-}
-
-func WithF16(f16 bool) AppOption {
-	return func(o *Option) {
-		o.f16 = f16
-	}
-}
-
-func WithDebug(debug bool) AppOption {
-	return func(o *Option) {
-		o.debug = debug
-	}
-}
-
-func WithDisableMessage(disableMessage bool) AppOption {
-	return func(o *Option) {
-		o.disableMessage = disableMessage
-	}
-}
-
-func WithImageDir(imageDir string) AppOption {
-	return func(o *Option) {
-		o.imageDir = imageDir
-	}
-}

api/options/options.go

@@ -0,0 +1,208 @@
+package options
+
+import (
+	"context"
+	"embed"
+	"encoding/json"
+
+	"github.com/go-skynet/LocalAI/pkg/gallery"
+	model "github.com/go-skynet/LocalAI/pkg/model"
+	"github.com/go-skynet/LocalAI/metrics"
+	"github.com/rs/zerolog/log"
+)
+
+type Option struct {
+	Context                             context.Context
+	ConfigFile                          string
+	Loader                              *model.ModelLoader
+	UploadLimitMB, Threads, ContextSize int
+	F16                                 bool
+	Debug, DisableMessage               bool
+	ImageDir                            string
+	AudioDir                            string
+	CORS                                bool
+	PreloadJSONModels                   string
+	PreloadModelsFromPath               string
+	CORSAllowOrigins                    string
+	ApiKeys                             []string
+	Metrics                             *metrics.Metrics
+
+	Galleries []gallery.Gallery
+
+	BackendAssets     embed.FS
+	AssetsDestination string
+
+	ExternalGRPCBackends map[string]string
+
+	AutoloadGalleries bool
+
+	SingleBackend bool
+}
+
+type AppOption func(*Option)
+
+func NewOptions(o ...AppOption) *Option {
+	opt := &Option{
+		Context:        context.Background(),
+		UploadLimitMB:  15,
+		Threads:        1,
+		ContextSize:    512,
+		Debug:          true,
+		DisableMessage: true,
+	}
+	for _, oo := range o {
+		oo(opt)
+	}
+	return opt
+}
+
+func WithCors(b bool) AppOption {
+	return func(o *Option) {
+		o.CORS = b
+	}
+}
+
+var EnableSingleBackend = func(o *Option) {
+	o.SingleBackend = true
+}
+
+var EnableGalleriesAutoload = func(o *Option) {
+	o.AutoloadGalleries = true
+}
+
+func WithExternalBackend(name string, uri string) AppOption {
+	return func(o *Option) {
+		if o.ExternalGRPCBackends == nil {
+			o.ExternalGRPCBackends = make(map[string]string)
+		}
+		o.ExternalGRPCBackends[name] = uri
+	}
+}
+
+func WithCorsAllowOrigins(b string) AppOption {
+	return func(o *Option) {
+		o.CORSAllowOrigins = b
+	}
+}
+
+func WithBackendAssetsOutput(out string) AppOption {
+	return func(o *Option) {
+		o.AssetsDestination = out
+	}
+}
+
+func WithBackendAssets(f embed.FS) AppOption {
+	return func(o *Option) {
+		o.BackendAssets = f
+	}
+}
+
+func WithStringGalleries(galls string) AppOption {
+	return func(o *Option) {
+		if galls == "" {
+			log.Debug().Msgf("no galleries to load")
+			o.Galleries = []gallery.Gallery{}
+			return
+		}
+		var galleries []gallery.Gallery
+		if err := json.Unmarshal([]byte(galls), &galleries); err != nil {
+			log.Error().Msgf("failed loading galleries: %s", err.Error())
+		}
+		o.Galleries = append(o.Galleries, galleries...)
+	}
+}
+
+func WithGalleries(galleries []gallery.Gallery) AppOption {
+	return func(o *Option) {
+		o.Galleries = append(o.Galleries, galleries...)
+	}
+}
+
+func WithContext(ctx context.Context) AppOption {
+	return func(o *Option) {
+		o.Context = ctx
+	}
+}
+
+func WithYAMLConfigPreload(configFile string) AppOption {
+	return func(o *Option) {
+		o.PreloadModelsFromPath = configFile
+	}
+}
+
+func WithJSONStringPreload(configFile string) AppOption {
+	return func(o *Option) {
+		o.PreloadJSONModels = configFile
+	}
+}
+func WithConfigFile(configFile string) AppOption {
+	return func(o *Option) {
+		o.ConfigFile = configFile
+	}
+}
+
+func WithModelLoader(loader *model.ModelLoader) AppOption {
+	return func(o *Option) {
+		o.Loader = loader
+	}
+}
+
+func WithUploadLimitMB(limit int) AppOption {
+	return func(o *Option) {
+		o.UploadLimitMB = limit
+	}
+}
+
+func WithThreads(threads int) AppOption {
+	return func(o *Option) {
+		o.Threads = threads
+	}
+}
+
+func WithContextSize(ctxSize int) AppOption {
+	return func(o *Option) {
+		o.ContextSize = ctxSize
+	}
+}
+
+func WithF16(f16 bool) AppOption {
+	return func(o *Option) {
+		o.F16 = f16
+	}
+}
+
+func WithDebug(debug bool) AppOption {
+	return func(o *Option) {
+		o.Debug = debug
+	}
+}
+
+func WithDisableMessage(disableMessage bool) AppOption {
+	return func(o *Option) {
+		o.DisableMessage = disableMessage
+	}
+}
+
+func WithAudioDir(audioDir string) AppOption {
+	return func(o *Option) {
+		o.AudioDir = audioDir
+	}
+}
+
+func WithImageDir(imageDir string) AppOption {
+	return func(o *Option) {
+		o.ImageDir = imageDir
+	}
+}
+
+func WithApiKeys(apiKeys []string) AppOption {
+	return func(o *Option) {
+		o.ApiKeys = apiKeys
+	}
+}
+
+func WithMetrics(meter *metrics.Metrics) AppOption {
+	return func(o *Option) {
+		o.Metrics = meter
+	}
+}

api/prediction.go

@@ -1,561 +0,0 @@
-package api
-
-import (
-	"fmt"
-	"regexp"
-	"strings"
-	"sync"
-
-	"github.com/donomii/go-rwkv.cpp"
-	model "github.com/go-skynet/LocalAI/pkg/model"
-	"github.com/go-skynet/LocalAI/pkg/stablediffusion"
-	"github.com/go-skynet/bloomz.cpp"
-	bert "github.com/go-skynet/go-bert.cpp"
-	gpt2 "github.com/go-skynet/go-gpt2.cpp"
-	llama "github.com/go-skynet/go-llama.cpp"
-	gpt4all "github.com/nomic-ai/gpt4all/gpt4all-bindings/golang"
-)
-
-// mutex still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
-var mutexMap sync.Mutex
-var mutexes map[string]*sync.Mutex = make(map[string]*sync.Mutex)
-
-func defaultLLamaOpts(c Config) []llama.ModelOption {
-	llamaOpts := []llama.ModelOption{}
-	if c.ContextSize != 0 {
-		llamaOpts = append(llamaOpts, llama.SetContext(c.ContextSize))
-	}
-	if c.F16 {
-		llamaOpts = append(llamaOpts, llama.EnableF16Memory)
-	}
-	if c.Embeddings {
-		llamaOpts = append(llamaOpts, llama.EnableEmbeddings)
-	}
-
-	if c.NGPULayers != 0 {
-		llamaOpts = append(llamaOpts, llama.SetGPULayers(c.NGPULayers))
-	}
-
-	return llamaOpts
-}
-
-func ImageGeneration(height, width, mode, step, seed int, positive_prompt, negative_prompt, dst string, loader *model.ModelLoader, c Config) (func() error, error) {
-	if c.Backend != model.StableDiffusionBackend {
-		return nil, fmt.Errorf("endpoint only working with stablediffusion models")
-	}
-	inferenceModel, err := loader.BackendLoader(c.Backend, c.ImageGenerationAssets, []llama.ModelOption{}, uint32(c.Threads))
-	if err != nil {
-		return nil, err
-	}
-
-	var fn func() error
-	switch model := inferenceModel.(type) {
-	case *stablediffusion.StableDiffusion:
-		fn = func() error {
-			return model.GenerateImage(height, width, mode, step, seed, positive_prompt, negative_prompt, dst)
-		}
-
-	default:
-		fn = func() error {
-			return fmt.Errorf("creation of images not supported by the backend")
-		}
-	}
-
-	return func() error {
-		// This is still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
-		mutexMap.Lock()
-		l, ok := mutexes[c.Backend]
-		if !ok {
-			m := &sync.Mutex{}
-			mutexes[c.Backend] = m
-			l = m
-		}
-		mutexMap.Unlock()
-		l.Lock()
-		defer l.Unlock()
-
-		return fn()
-	}, nil
-}
-
-func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, c Config) (func() ([]float32, error), error) {
-	if !c.Embeddings {
-		return nil, fmt.Errorf("endpoint disabled for this model by API configuration")
-	}
-
-	modelFile := c.Model
-
-	llamaOpts := defaultLLamaOpts(c)
-
-	var inferenceModel interface{}
-	var err error
-	if c.Backend == "" {
-		inferenceModel, err = loader.GreedyLoader(modelFile, llamaOpts, uint32(c.Threads))
-	} else {
-		inferenceModel, err = loader.BackendLoader(c.Backend, modelFile, llamaOpts, uint32(c.Threads))
-	}
-	if err != nil {
-		return nil, err
-	}
-
-	var fn func() ([]float32, error)
-	switch model := inferenceModel.(type) {
-	case *llama.LLama:
-		fn = func() ([]float32, error) {
-			predictOptions := buildLLamaPredictOptions(c)
-			if len(tokens) > 0 {
-				return model.TokenEmbeddings(tokens, predictOptions...)
-			}
-			return model.Embeddings(s, predictOptions...)
-		}
-	// bert embeddings
-	case *bert.Bert:
-		fn = func() ([]float32, error) {
-			if len(tokens) > 0 {
-				return model.TokenEmbeddings(tokens, bert.SetThreads(c.Threads))
-			}
-			return model.Embeddings(s, bert.SetThreads(c.Threads))
-		}
-	default:
-		fn = func() ([]float32, error) {
-			return nil, fmt.Errorf("embeddings not supported by the backend")
-		}
-	}
-
-	return func() ([]float32, error) {
-		// This is still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
-		mutexMap.Lock()
-		l, ok := mutexes[modelFile]
-		if !ok {
-			m := &sync.Mutex{}
-			mutexes[modelFile] = m
-			l = m
-		}
-		mutexMap.Unlock()
-		l.Lock()
-		defer l.Unlock()
-
-		embeds, err := fn()
-		if err != nil {
-			return embeds, err
-		}
-		// Remove trailing 0s
-		for i := len(embeds) - 1; i >= 0; i-- {
-			if embeds[i] == 0.0 {
-				embeds = embeds[:i]
-			} else {
-				break
-			}
-		}
-		return embeds, nil
-	}, nil
-}
-
-func buildLLamaPredictOptions(c Config) []llama.PredictOption {
-	// Generate the prediction using the language model
-	predictOptions := []llama.PredictOption{
-		llama.SetTemperature(c.Temperature),
-		llama.SetTopP(c.TopP),
-		llama.SetTopK(c.TopK),
-		llama.SetTokens(c.Maxtokens),
-		llama.SetThreads(c.Threads),
-	}
-
-	if c.Mirostat != 0 {
-		predictOptions = append(predictOptions, llama.SetMirostat(c.Mirostat))
-	}
-
-	if c.MirostatETA != 0 {
-		predictOptions = append(predictOptions, llama.SetMirostatETA(c.MirostatETA))
-	}
-
-	if c.MirostatTAU != 0 {
-		predictOptions = append(predictOptions, llama.SetMirostatTAU(c.MirostatTAU))
-	}
-
-	if c.Debug {
-		predictOptions = append(predictOptions, llama.Debug)
-	}
-
-	predictOptions = append(predictOptions, llama.SetStopWords(c.StopWords...))
-
-	if c.RepeatPenalty != 0 {
-		predictOptions = append(predictOptions, llama.SetPenalty(c.RepeatPenalty))
-	}
-
-	if c.Keep != 0 {
-		predictOptions = append(predictOptions, llama.SetNKeep(c.Keep))
-	}
-
-	if c.Batch != 0 {
-		predictOptions = append(predictOptions, llama.SetBatch(c.Batch))
-	}
-
-	if c.F16 {
-		predictOptions = append(predictOptions, llama.EnableF16KV)
-	}
-
-	if c.IgnoreEOS {
-		predictOptions = append(predictOptions, llama.IgnoreEOS)
-	}
-
-	if c.Seed != 0 {
-		predictOptions = append(predictOptions, llama.SetSeed(c.Seed))
-	}
-
-	return predictOptions
-}
-
-func ModelInference(s string, loader *model.ModelLoader, c Config, tokenCallback func(string) bool) (func() (string, error), error) {
-	supportStreams := false
-	modelFile := c.Model
-
-	llamaOpts := defaultLLamaOpts(c)
-
-	var inferenceModel interface{}
-	var err error
-	if c.Backend == "" {
-		inferenceModel, err = loader.GreedyLoader(modelFile, llamaOpts, uint32(c.Threads))
-	} else {
-		inferenceModel, err = loader.BackendLoader(c.Backend, modelFile, llamaOpts, uint32(c.Threads))
-	}
-	if err != nil {
-		return nil, err
-	}
-
-	var fn func() (string, error)
-
-	switch model := inferenceModel.(type) {
-	case *rwkv.RwkvState:
-		supportStreams = true
-
-		fn = func() (string, error) {
-			stopWord := "\n"
-			if len(c.StopWords) > 0 {
-				stopWord = c.StopWords[0]
-			}
-
-			if err := model.ProcessInput(s); err != nil {
-				return "", err
-			}
-
-			response := model.GenerateResponse(c.Maxtokens, stopWord, float32(c.Temperature), float32(c.TopP), tokenCallback)
-
-			return response, nil
-		}
-	case *gpt2.GPTNeoX:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.Replit:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.Starcoder:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.RedPajama:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *bloomz.Bloomz:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []bloomz.PredictOption{
-				bloomz.SetTemperature(c.Temperature),
-				bloomz.SetTopP(c.TopP),
-				bloomz.SetTopK(c.TopK),
-				bloomz.SetTokens(c.Maxtokens),
-				bloomz.SetThreads(c.Threads),
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, bloomz.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.StableLM:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.Dolly:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt2.GPT2:
-		fn = func() (string, error) {
-			// Generate the prediction using the language model
-			predictOptions := []gpt2.PredictOption{
-				gpt2.SetTemperature(c.Temperature),
-				gpt2.SetTopP(c.TopP),
-				gpt2.SetTopK(c.TopK),
-				gpt2.SetTokens(c.Maxtokens),
-				gpt2.SetThreads(c.Threads),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt2.SetBatch(c.Batch))
-			}
-
-			if c.Seed != 0 {
-				predictOptions = append(predictOptions, gpt2.SetSeed(c.Seed))
-			}
-
-			return model.Predict(
-				s,
-				predictOptions...,
-			)
-		}
-	case *gpt4all.Model:
-		supportStreams = true
-
-		fn = func() (string, error) {
-			if tokenCallback != nil {
-				model.SetTokenCallback(tokenCallback)
-			}
-
-			// Generate the prediction using the language model
-			predictOptions := []gpt4all.PredictOption{
-				gpt4all.SetTemperature(c.Temperature),
-				gpt4all.SetTopP(c.TopP),
-				gpt4all.SetTopK(c.TopK),
-				gpt4all.SetTokens(c.Maxtokens),
-			}
-
-			if c.Batch != 0 {
-				predictOptions = append(predictOptions, gpt4all.SetBatch(c.Batch))
-			}
-
-			str, er := model.Predict(
-				s,
-				predictOptions...,
-			)
-			// Seems that if we don't free the callback explicitly we leave functions registered (that might try to send on closed channels)
-			// For instance otherwise the API returns: {"error":{"code":500,"message":"send on closed channel","type":""}}
-			// after a stream event has occurred
-			model.SetTokenCallback(nil)
-			return str, er
-		}
-	case *llama.LLama:
-		supportStreams = true
-		fn = func() (string, error) {
-
-			if tokenCallback != nil {
-				model.SetTokenCallback(tokenCallback)
-			}
-
-			predictOptions := buildLLamaPredictOptions(c)
-
-			str, er := model.Predict(
-				s,
-				predictOptions...,
-			)
-			// Seems that if we don't free the callback explicitly we leave functions registered (that might try to send on closed channels)
-			// For instance otherwise the API returns: {"error":{"code":500,"message":"send on closed channel","type":""}}
-			// after a stream event has occurred
-			model.SetTokenCallback(nil)
-			return str, er
-		}
-	}
-
-	return func() (string, error) {
-		// This is still needed, see: https://github.com/ggerganov/llama.cpp/discussions/784
-		mutexMap.Lock()
-		l, ok := mutexes[modelFile]
-		if !ok {
-			m := &sync.Mutex{}
-			mutexes[modelFile] = m
-			l = m
-		}
-		mutexMap.Unlock()
-		l.Lock()
-		defer l.Unlock()
-
-		res, err := fn()
-		if tokenCallback != nil && !supportStreams {
-			tokenCallback(res)
-		}
-		return res, err
-	}, nil
-}
-
-func ComputeChoices(predInput string, input *OpenAIRequest, config *Config, loader *model.ModelLoader, cb func(string, *[]Choice), tokenCallback func(string) bool) ([]Choice, error) {
-	result := []Choice{}
-
-	n := input.N
-
-	if input.N == 0 {
-		n = 1
-	}
-
-	// get the model function to call for the result
-	predFunc, err := ModelInference(predInput, loader, *config, tokenCallback)
-	if err != nil {
-		return result, err
-	}
-
-	for i := 0; i < n; i++ {
-		prediction, err := predFunc()
-		if err != nil {
-			return result, err
-		}
-
-		prediction = Finetune(*config, predInput, prediction)
-		cb(prediction, &result)
-
-		//result = append(result, Choice{Text: prediction})
-
-	}
-	return result, err
-}
-
-var cutstrings map[string]*regexp.Regexp = make(map[string]*regexp.Regexp)
-var mu sync.Mutex = sync.Mutex{}
-
-func Finetune(config Config, input, prediction string) string {
-	if config.Echo {
-		prediction = input + prediction
-	}
-
-	for _, c := range config.Cutstrings {
-		mu.Lock()
-		reg, ok := cutstrings[c]
-		if !ok {
-			cutstrings[c] = regexp.MustCompile(c)
-			reg = cutstrings[c]
-		}
-		mu.Unlock()
-		prediction = reg.ReplaceAllString(prediction, "")
-	}
-
-	for _, c := range config.TrimSpace {
-		prediction = strings.TrimSpace(strings.TrimPrefix(prediction, c))
-	}
-	return prediction
-
-}

api/schema/openai.go

@@ -0,0 +1,115 @@
+package schema
+
+import (
+	"context"
+
+	config "github.com/go-skynet/LocalAI/api/config"
+
+	"github.com/go-skynet/LocalAI/pkg/grammar"
+)
+
+// APIError provides error information returned by the OpenAI API.
+type APIError struct {
+	Code    any     `json:"code,omitempty"`
+	Message string  `json:"message"`
+	Param   *string `json:"param,omitempty"`
+	Type    string  `json:"type"`
+}
+
+type ErrorResponse struct {
+	Error *APIError `json:"error,omitempty"`
+}
+
+type OpenAIUsage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+}
+
+type Item struct {
+	Embedding []float32 `json:"embedding"`
+	Index     int       `json:"index"`
+	Object    string    `json:"object,omitempty"`
+
+	// Images
+	URL     string `json:"url,omitempty"`
+	B64JSON string `json:"b64_json,omitempty"`
+}
+
+type OpenAIResponse struct {
+	Created int      `json:"created,omitempty"`
+	Object  string   `json:"object,omitempty"`
+	ID      string   `json:"id,omitempty"`
+	Model   string   `json:"model,omitempty"`
+	Choices []Choice `json:"choices,omitempty"`
+	Data    []Item   `json:"data,omitempty"`
+
+	Usage OpenAIUsage `json:"usage"`
+}
+
+type Choice struct {
+	Index        int      `json:"index"`
+	FinishReason string   `json:"finish_reason,omitempty"`
+	Message      *Message `json:"message,omitempty"`
+	Delta        *Message `json:"delta,omitempty"`
+	Text         string   `json:"text,omitempty"`
+}
+
+type Message struct {
+	// The message role
+	Role string `json:"role,omitempty" yaml:"role"`
+	// The message content
+	Content *string `json:"content" yaml:"content"`
+	// A result of a function call
+	FunctionCall interface{} `json:"function_call,omitempty" yaml:"function_call,omitempty"`
+}
+
+type OpenAIModel struct {
+	ID     string `json:"id"`
+	Object string `json:"object"`
+}
+
+type OpenAIRequest struct {
+	config.PredictionOptions
+
+	Context context.Context
+	Cancel  context.CancelFunc
+
+	// whisper
+	File string `json:"file" validate:"required"`
+	//whisper/image
+	ResponseFormat string `json:"response_format"`
+	// image
+	Size string `json:"size"`
+	// Prompt is read only by completion/image API calls
+	Prompt interface{} `json:"prompt" yaml:"prompt"`
+
+	// Edit endpoint
+	Instruction string      `json:"instruction" yaml:"instruction"`
+	Input       interface{} `json:"input" yaml:"input"`
+
+	Stop interface{} `json:"stop" yaml:"stop"`
+
+	// Messages is read only by chat/completion API calls
+	Messages []Message `json:"messages" yaml:"messages"`
+
+	// A list of available functions to call
+	Functions    []grammar.Function `json:"functions" yaml:"functions"`
+	FunctionCall interface{}        `json:"function_call" yaml:"function_call"` // might be a string or an object
+
+	Stream bool `json:"stream"`
+
+	// Image (not supported by OpenAI)
+	Mode int `json:"mode"`
+	Step int `json:"step"`
+
+	// A grammar to constrain the LLM output
+	Grammar string `json:"grammar" yaml:"grammar"`
+
+	JSONFunctionGrammarObject *grammar.JSONFunctionStructure `json:"grammar_json_functions" yaml:"grammar_json_functions"`
+
+	Backend string `json:"backend" yaml:"backend"`
+
+	// AutoGPTQ
+	ModelBaseName string `json:"model_base_name" yaml:"model_base_name"`
+}

api/schema/whisper.go

@@ -0,0 +1,16 @@
+package schema
+
+import "time"
+
+type Segment struct {
+	Id     int           `json:"id"`
+	Start  time.Duration `json:"start"`
+	End    time.Duration `json:"end"`
+	Text   string        `json:"text"`
+	Tokens []int         `json:"tokens"`
+}
+
+type Result struct {
+	Segments []Segment `json:"segments"`
+	Text     string    `json:"text"`
+}

assets.go

@@ -0,0 +1,6 @@
+package main
+
+import "embed"
+
+//go:embed backend-assets/*
+var backendAssets embed.FS

backend/cpp/llama/CMakeLists.txt

@@ -0,0 +1,61 @@
+set(CMAKE_CXX_STANDARD 17)
+cmake_minimum_required(VERSION 3.15)
+set(TARGET grpc-server)
+set(_PROTOBUF_LIBPROTOBUF libprotobuf)
+set(_REFLECTION grpc++_reflection)
+
+if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    link_directories("/opt/homebrew/lib")
+    include_directories("/opt/homebrew/include")
+endif()
+
+find_package(absl CONFIG REQUIRED)
+find_package(Protobuf CONFIG REQUIRED)
+find_package(gRPC CONFIG REQUIRED)
+
+find_program(_PROTOBUF_PROTOC protoc)
+set(_GRPC_GRPCPP grpc++)
+find_program(_GRPC_CPP_PLUGIN_EXECUTABLE grpc_cpp_plugin)
+
+include_directories(${CMAKE_CURRENT_BINARY_DIR})
+include_directories(${Protobuf_INCLUDE_DIRS})
+
+message(STATUS "Using protobuf version ${Protobuf_VERSION} | Protobuf_INCLUDE_DIRS: ${Protobuf_INCLUDE_DIRS} | CMAKE_CURRENT_BINARY_DIR: ${CMAKE_CURRENT_BINARY_DIR}")
+
+# Proto file
+get_filename_component(hw_proto "../../../../../../pkg/grpc/proto/backend.proto" ABSOLUTE)
+get_filename_component(hw_proto_path "${hw_proto}" PATH)
+
+# Generated sources
+set(hw_proto_srcs "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.cc")
+set(hw_proto_hdrs "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.h")
+set(hw_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.cc")
+set(hw_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.h")
+
+add_custom_command(
+      OUTPUT "${hw_proto_srcs}" "${hw_proto_hdrs}" "${hw_grpc_srcs}" "${hw_grpc_hdrs}"
+      COMMAND ${_PROTOBUF_PROTOC}
+      ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}"
+        --cpp_out "${CMAKE_CURRENT_BINARY_DIR}"
+        -I "${hw_proto_path}"
+        --plugin=protoc-gen-grpc="${_GRPC_CPP_PLUGIN_EXECUTABLE}"
+        "${hw_proto}"
+      DEPENDS "${hw_proto}")
+
+# hw_grpc_proto
+add_library(hw_grpc_proto
+  ${hw_grpc_srcs}
+  ${hw_grpc_hdrs}
+  ${hw_proto_srcs}
+  ${hw_proto_hdrs})
+
+add_executable(${TARGET} grpc-server.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
+  absl::flags_parse
+  gRPC::${_REFLECTION}
+  gRPC::${_GRPC_GRPCPP}
+  protobuf::${_PROTOBUF_LIBPROTOBUF})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

backend/cpp/llama/Makefile

@@ -0,0 +1,44 @@
+
+LLAMA_VERSION?=
+
+CMAKE_ARGS?=
+BUILD_TYPE?=
+
+# If build type is cublas, then we set -DLLAMA_CUBLAS=ON to CMAKE_ARGS automatically
+ifeq ($(BUILD_TYPE),cublas)
+	CMAKE_ARGS+=-DLLAMA_CUBLAS=ON
+# If build type is openblas then we set -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
+# to CMAKE_ARGS automatically
+else ifeq ($(BUILD_TYPE),openblas)
+	CMAKE_ARGS+=-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
+# If build type is clblast (openCL) we set -DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
+else ifeq ($(BUILD_TYPE),clblast)
+	CMAKE_ARGS+=-DLLAMA_CLBLAST=ON -DCLBlast_DIR=/some/path
+# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
+else ifeq ($(BUILD_TYPE),hipblas)
+	CMAKE_ARGS+=-DLLAMA_HIPBLAS=ON
+endif
+
+llama.cpp:
+	git clone --recurse-submodules https://github.com/ggerganov/llama.cpp llama.cpp
+	cd llama.cpp && git checkout -b build $(LLAMA_VERSION) && git submodule update --init --recursive --depth 1
+
+llama.cpp/examples/grpc-server:
+	mkdir -p llama.cpp/examples/grpc-server
+	cp -r $(abspath ./)/CMakeLists.txt llama.cpp/examples/grpc-server/
+	cp -r $(abspath ./)/grpc-server.cpp llama.cpp/examples/grpc-server/
+	echo "add_subdirectory(grpc-server)" >> llama.cpp/examples/CMakeLists.txt
+
+rebuild:
+	cp -rfv $(abspath ./)/CMakeLists.txt llama.cpp/examples/grpc-server/
+	cp -rfv $(abspath ./)/grpc-server.cpp llama.cpp/examples/grpc-server/
+	rm -rf grpc-server
+	$(MAKE) grpc-server
+
+clean:
+	rm -rf llama.cpp
+	rm -rf grpc-server
+
+grpc-server: llama.cpp llama.cpp/examples/grpc-server
+	cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release
+	cp llama.cpp/build/bin/grpc-server .

backend/cpp/llama/grpc-server.cpp

@@ -0,0 +1,926 @@
+// llama.cpp gRPC C++ backend server
+//
+// Ettore Di Giacinto <mudler@localai.io>
+//
+// This is a gRPC server for llama.cpp compatible with the LocalAI proto
+// Note: this is a re-adaptation of the original llama.cpp example/server.cpp for HTTP, 
+// but modified to work with gRPC
+//
+
+#include <iostream>
+#include <memory>
+#include <string>
+#include <getopt.h>
+
+#include "common.h"
+#include "llama.h"
+#include "grammar-parser.h"
+#include "backend.pb.h"
+#include "backend.grpc.pb.h"
+
+// include std::regex
+#include <regex>
+#include <grpcpp/ext/proto_server_reflection_plugin.h>
+#include <grpcpp/grpcpp.h>
+#include <grpcpp/health_check_service_interface.h>
+
+using grpc::Server;
+using grpc::ServerBuilder;
+using grpc::ServerContext;
+using grpc::Status;
+
+
+using backend::HealthMessage;
+
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+};
+
+static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
+{
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
+    {
+    }
+    return i;
+}
+
+enum stop_type
+{
+    STOP_FULL,
+    STOP_PARTIAL,
+};
+
+static bool ends_with(const std::string &str, const std::string &suffix)
+{
+    return str.size() >= suffix.size() &&
+           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+}
+
+static size_t find_partial_stop_string(const std::string &stop,
+                                       const std::string &text)
+{
+    if (!text.empty() && !stop.empty())
+    {
+        const char text_last_char = text.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
+        {
+            if (stop[char_index] == text_last_char)
+            {
+                const std::string current_partial = stop.substr(0, char_index + 1);
+                if (ends_with(text, current_partial))
+                {
+                    return text.size() - char_index - 1;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
+
+
+template <class Iter>
+static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
+{
+    std::string ret;
+    for (; begin != end; ++begin)
+    {
+        ret += llama_token_to_piece(ctx, *begin);
+    }
+    return ret;
+}
+
+
+// format incomplete utf-8 multibyte character for output
+static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
+{
+    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
+    // if the size is 1 and first bit is 1, meaning it's a partial character
+    //   (size > 1 meaning it's already a known token)
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
+    {
+        std::stringstream ss;
+        ss << std::hex << (out[0] & 0xff);
+        std::string res(ss.str());
+        out = "byte: \\x" + res;
+    }
+    return out;
+}
+
+struct llama_server_context
+{
+    bool stream = false;
+    bool has_next_token = false;
+    std::string generated_text;
+    std::vector<completion_token_output> generated_token_probs;
+
+    size_t num_prompt_tokens = 0;
+    size_t num_tokens_predicted = 0;
+    size_t n_past = 0;
+    size_t n_remain = 0;
+
+    std::vector<llama_token> embd;
+
+    gpt_params params;
+
+    llama_model *model = nullptr;
+    llama_context *ctx = nullptr;
+    llama_sampling_context *ctx_sampling = nullptr;
+
+    int n_ctx;
+
+    bool truncated = false;
+    bool stopped_eos = false;
+    bool stopped_word = false;
+    bool stopped_limit = false;
+    std::string stopping_word;
+    int32_t multibyte_pending = 0;
+
+    std::mutex mutex;
+
+    std::unique_lock<std::mutex> lock()
+    {
+        return std::unique_lock<std::mutex>(mutex);
+    }
+
+    ~llama_server_context()
+    {
+        if (ctx)
+        {
+            llama_free(ctx);
+            ctx = nullptr;
+        }
+        if (model)
+        {
+            llama_free_model(model);
+            model = nullptr;
+        }
+    }
+
+    void rewind()
+    {
+        params.antiprompt.clear();
+        params.sparams.grammar.clear();
+        num_prompt_tokens = 0;
+        num_tokens_predicted = 0;
+        generated_text = "";
+        generated_text.reserve(n_ctx);
+        generated_token_probs.clear();
+        truncated = false;
+        stopped_eos = false;
+        stopped_word = false;
+        stopped_limit = false;
+        stopping_word = "";
+        multibyte_pending = 0;
+        n_remain = 0;
+        n_past = 0;
+        params.sparams.n_prev = n_ctx;
+    }
+
+    void initSampling() {
+        if (ctx_sampling != nullptr) {
+            llama_sampling_free(ctx_sampling);
+        }
+        ctx_sampling = llama_sampling_init(params.sparams);
+    }
+
+    bool loadModel(const gpt_params &params_)
+    {
+        params = params_;
+        std::tie(model, ctx) = llama_init_from_gpt_params(params);
+        if (model == nullptr)
+        {
+            return false;
+        }
+        n_ctx = llama_n_ctx(ctx);
+        return true;
+    }
+    std::vector<llama_token> tokenize_string(const char *prompt, bool add_bos) const {
+       // 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.
+        std::vector<llama_token> prompt_tokens; 
+        auto s = std::string(prompt);
+        prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
+        return prompt_tokens;
+    }
+     std::vector<llama_token> tokenize_array(const char **prompts, bool add_bos) const {
+                std::vector<llama_token> prompt_tokens; 
+
+            bool first = true;
+            bool is_string = true;
+            for (const char **p = prompts; *p != nullptr; ++p)
+              {
+                if (is_string)
+                {
+                    auto s = std::string(*p);
+                    std::vector<llama_token> p;
+                    if (first)
+                    {
+                        p = ::llama_tokenize(ctx, s, add_bos);
+                        first = false;
+                    }
+                    else
+                    {
+                        p = ::llama_tokenize(ctx, s, false);
+                    }
+                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+                }
+                else
+                {
+                    if (first)
+                    {
+                        first = false;
+                    }
+                    //prompt_tokens.push_back(p.template get<llama_token>());
+                }
+            }
+            return prompt_tokens;
+     }
+
+    void truncatePrompt(std::vector<llama_token> &prompt_tokens) {
+        const int n_left = n_ctx - params.n_keep;
+        const int n_block_size = n_left / 2;
+        const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_block_size) / n_block_size;
+
+        // Keep n_keep tokens at start of prompt (at most n_ctx - 4)
+        std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+
+        new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_block_size, prompt_tokens.end());
+
+        truncated = true;
+        prompt_tokens = new_tokens;
+    }
+
+    void loadInfill()
+    {
+        bool suff_rm_leading_spc = true;
+        if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
+            params.input_suffix.erase(0, 1);
+            suff_rm_leading_spc = false;
+        }
+
+        auto prefix_tokens = tokenize_string(params.input_prefix.c_str(), false);
+        auto suffix_tokens = tokenize_string(params.input_suffix.c_str(), false);
+        const int space_token = 29871;
+        if (suff_rm_leading_spc  && suffix_tokens[0] == space_token) {
+            suffix_tokens.erase(suffix_tokens.begin());
+        }
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
+        prefix_tokens.push_back(llama_token_middle(ctx));
+
+        auto prompt_tokens = prefix_tokens;
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t) n_ctx)
+        {
+            truncatePrompt(prompt_tokens);
+            num_prompt_tokens = prompt_tokens.size();
+
+            GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
+        }
+
+        // push the prompt into the sampling context (do not apply grammar)
+        for (auto & token : prompt_tokens)
+        {
+            llama_sampling_accept(ctx_sampling, ctx, token, false);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+        embd = prompt_tokens;
+
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            printf("we have to evaluate at least 1 token to generate logits\n");
+            n_past--;
+        }
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        has_next_token = true;
+    }
+    void loadPrompt(std::string prompt)
+    {
+        auto prompt_tokens = tokenize_string(prompt.c_str(), true);  // always add BOS
+
+        num_prompt_tokens = prompt_tokens.size();
+
+        if (params.n_keep < 0)
+        {
+            params.n_keep = (int)num_prompt_tokens;
+        }
+        params.n_keep = std::min(n_ctx - 4, params.n_keep);
+
+        // if input prompt is too big, truncate like normal
+        if (num_prompt_tokens >= (size_t) n_ctx)
+        {
+            truncatePrompt(prompt_tokens);
+            num_prompt_tokens = prompt_tokens.size();
+
+            GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
+        }
+
+        // push the prompt into the sampling context (do not apply grammar)
+        for (auto & token : prompt_tokens)
+        {
+            llama_sampling_accept(ctx_sampling, ctx, token, false);
+        }
+
+        // compare the evaluated prompt with the new prompt
+        n_past = common_part(embd, prompt_tokens);
+
+        embd = prompt_tokens;
+        if (n_past == num_prompt_tokens)
+        {
+            // we have to evaluate at least 1 token to generate logits.
+            n_past--;
+        }
+
+        // since #3228 we now have to manually manage the KV cache
+        llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
+        has_next_token = true;
+    }
+
+    void beginCompletion()
+    {
+        // number of tokens to keep when resetting context
+        n_remain = params.n_predict;
+        llama_set_rng_seed(ctx, params.seed);
+    }
+
+    completion_token_output nextToken()
+    {
+        completion_token_output result;
+        result.tok = -1;
+
+        if (embd.size() >= (size_t)n_ctx)
+        {
+            // Shift context
+
+            const int n_left    = n_past - params.n_keep - 1;
+            const int n_discard = n_left/2;
+
+            llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+            llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+            for (size_t i = params.n_keep + 1 + n_discard; i < embd.size(); i++)
+            {
+                embd[i - n_discard] = embd[i];
+            }
+            embd.resize(embd.size() - n_discard);
+
+            n_past -= n_discard;
+
+            truncated = true;
+        }
+
+        bool tg = true;
+        while (n_past < embd.size())
+        {
+            int n_eval = (int)embd.size() - n_past;
+            tg = n_eval == 1;
+            if (n_eval > params.n_batch)
+            {
+                n_eval = params.n_batch;
+            }
+
+            if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0)))
+            {
+                has_next_token = false;
+                return result;
+            }
+            n_past += n_eval;
+        }
+
+        if (params.n_predict == 0)
+        {
+            has_next_token = false;
+            result.tok = llama_token_eos(ctx);
+            return result;
+        }
+
+        {
+            // out of user input, sample next token
+            result.tok = llama_sampling_sample(ctx_sampling, ctx, NULL);
+
+            llama_token_data_array cur_p = { ctx_sampling->cur.data(), ctx_sampling->cur.size(), false };
+
+            const int32_t n_probs = params.sparams.n_probs;
+            if (params.sparams.temp <= 0 && n_probs > 0)
+            {
+                // For llama_sample_token_greedy we need to sort candidates
+                llama_sample_softmax(ctx, &cur_p);
+            }
+
+            for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i)
+            {
+                result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p});
+            }
+
+            llama_sampling_accept(ctx_sampling, ctx, result.tok, true);
+
+            if (tg) {
+                num_tokens_predicted++;
+            }
+        }
+
+        // add it to the context
+        embd.push_back(result.tok);
+        // decrement remaining sampling budget
+        --n_remain;
+
+        if (!embd.empty() && embd.back() == llama_token_eos(ctx))
+        {
+            // stopping_word = llama_token_to_piece(ctx, embd.back());
+            has_next_token = false;
+            stopped_eos = true;
+            return result;
+        }
+
+        has_next_token = params.n_predict == -1 || n_remain != 0;
+        return result;
+    }
+
+    size_t findStoppingStrings(const std::string &text, const size_t last_token_size,
+                               const stop_type type)
+    {
+        size_t stop_pos = std::string::npos;
+        for (const std::string &word : params.antiprompt)
+        {
+            size_t pos;
+            if (type == STOP_FULL)
+            {
+                const size_t tmp = word.size() + last_token_size;
+                const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;
+                pos = text.find(word, from_pos);
+            }
+            else
+            {
+                pos = find_partial_stop_string(word, text);
+            }
+            if (pos != std::string::npos &&
+                (stop_pos == std::string::npos || pos < stop_pos))
+            {
+                if (type == STOP_FULL)
+                {
+                    stopping_word = word;
+                    stopped_word = true;
+                    has_next_token = false;
+                }
+                stop_pos = pos;
+            }
+        }
+        return stop_pos;
+    }
+
+    completion_token_output doCompletion()
+    {
+        auto token_with_probs = nextToken();
+
+        const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
+        generated_text += token_text;
+
+        if (params.sparams.n_probs > 0)
+        {
+            generated_token_probs.push_back(token_with_probs);
+        }
+
+        if (multibyte_pending > 0)
+        {
+            multibyte_pending -= token_text.size();
+        }
+        else if (token_text.size() == 1)
+        {
+            const char c = token_text[0];
+            // 2-byte characters: 110xxxxx 10xxxxxx
+            if ((c & 0xE0) == 0xC0)
+            {
+                multibyte_pending = 1;
+                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+            }
+            else if ((c & 0xF0) == 0xE0)
+            {
+                multibyte_pending = 2;
+                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+            }
+            else if ((c & 0xF8) == 0xF0)
+            {
+                multibyte_pending = 3;
+            }
+            else
+            {
+                multibyte_pending = 0;
+            }
+        }
+
+        if (multibyte_pending > 0 && !has_next_token)
+        {
+            has_next_token = true;
+            n_remain++;
+        }
+
+        if (!has_next_token && n_remain == 0)
+        {
+            stopped_limit = true;
+        }
+
+        return token_with_probs;
+    }
+
+    std::vector<float> getEmbedding()
+    {
+        static const int n_embd = llama_n_embd(model);
+        if (!params.embedding)
+        {
+            return std::vector<float>(n_embd, 0.0f);
+        }
+        const float *data = llama_get_embeddings(ctx);
+        std::vector<float> embedding(data, data + n_embd);
+        return embedding;
+    }
+};
+
+
+static void parse_options_completion(bool streaming,const backend::PredictOptions* predict, llama_server_context &llama)
+{
+    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.tfs_z = predict->tailfreesamplingz();
+    llama.params.sparams.typical_p = predict->typicalp();
+    llama.params.sparams.penalty_last_n = predict->repeat();
+    llama.params.sparams.temp = predict->temperature();
+    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.sparams.penalize_nl = predict->penalizenl();
+    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.ctx)] = -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);
+            }
+    }
+}
+
+
+
+static void params_parse(const backend::ModelOptions* request,
+                                gpt_params & params) {
+   
+    params.model = request->modelfile();
+    //  params.model_alias ??
+    params.model_alias =  request->modelfile();
+    params.n_ctx = request->contextsize();
+    params.memory_f16 = request->f16memory();
+    params.n_threads = request->threads();
+    params.n_gpu_layers = request->ngpulayers();
+    params.n_batch = request->nbatch();
+
+    if (!request->tensorsplit().empty()) {
+        std::string arg_next = request->tensorsplit();
+
+        // split string by , and /
+        const std::regex regex{ R"([,/]+)" };
+        std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+        std::vector<std::string> split_arg{ it, {} };
+
+        GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
+        for (size_t i_device = 0; i_device < LLAMA_MAX_DEVICES; ++i_device) {
+            if (i_device < split_arg.size()) {
+                params.tensor_split[i_device] = std::stof(split_arg[i_device]);
+            }
+            else {
+                params.tensor_split[i_device] = 0.0f;
+            }
+        }
+    }
+
+    if (!request->maingpu().empty()) {
+        params.main_gpu = std::stoi(request->maingpu());
+    }
+    // TODO: lora needs also a scale factor
+    //params.lora_adapter = request->loraadapter();
+    //params.lora_base = request->lorabase();
+    params.use_mlock = request->mlock();
+    params.use_mmap = request->mmap();
+    params.embedding = request->embeddings();
+}
+
+static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
+}
+
+// Function matching type llama_beam_search_callback_fn_t.
+// Custom callback example is called each time the beams lengths increase:
+//  * Show progress by printing ',' following by number of convergent beam tokens if any.
+//  * When all beams converge to a common prefix, they are made available in beams_state.beams[0].
+//    This is also called when the stop condition is met.
+//    Collect tokens into std::vector<llama_token> response which is pointed to by callback_data.
+static void beam_search_callback(void *callback_data, llama_beams_state beams_state) {
+    auto & llama = *static_cast<llama_server_context*>(callback_data);
+    // Mark beams as EOS as needed.
+    for (size_t i = 0 ; i < beams_state.n_beams ; ++i) {
+        llama_beam_view& beam_view = beams_state.beam_views[i];
+        if (!beam_view.eob && is_at_eob(llama, beam_view.tokens, beam_view.n_tokens)) {
+            beam_view.eob = true;
+        }
+    }
+    printf(",");  // Show progress
+    if (const size_t n = beams_state.common_prefix_length) {
+        llama.generated_token_probs.resize(llama.generated_token_probs.size() + n);
+        assert(0u < beams_state.n_beams);
+        const llama_token * tokens = beams_state.beam_views[0].tokens;
+        const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
+        std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
+        printf("%zu", n);
+    }
+    fflush(stdout);
+#if 0 // DEBUG: print current beams for this iteration
+    std::cout << "\n\nCurrent beams:\n";
+    for (size_t i=0 ; i < beams_state.n_beams ; ++i) {
+        std::cout << "beams["<<i<<"]: " << ostream_beam_view{state.ctx,beams_state.beam_views[i]} << std::endl;
+    }
+#endif
+}
+struct token_translator {
+    llama_context * ctx;
+    std::string operator()(llama_token tok) const { return llama_token_to_piece(ctx, tok); }
+    std::string operator()(const completion_token_output & cto) const { return (*this)(cto.tok); }
+};
+
+
+static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama)
+{
+    auto & gtps = llama.generated_token_probs;
+    auto translator = token_translator{llama.ctx};
+    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
+    const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
+    if (llama.generated_text.capacity() < llama.generated_text.size() + len) {
+        llama.generated_text.reserve(llama.generated_text.size() + len);
+    }
+    for (const completion_token_output & cto : gtps) {
+        llama.generated_text += translator(cto);
+    }
+}
+
+// GRPC Server start
+class BackendServiceImpl final : public backend::Backend::Service {
+  // The class has a llama instance that is shared across all RPCs
+  llama_server_context llama;
+public:
+  grpc::Status Health(ServerContext* context, const backend::HealthMessage* request, backend::Reply* reply) {
+    // Implement Health RPC
+    reply->set_message("OK");
+    return Status::OK;
+  }
+
+  grpc::Status LoadModel(ServerContext* context, const backend::ModelOptions* request, backend::Result* result) {
+    // Implement LoadModel RPC
+    gpt_params params;
+    params_parse(request, params);
+
+    llama_backend_init(params.numa);
+
+    // load the model
+    if (!llama.loadModel(params))
+    {
+        result->set_message("Failed loading model");
+        result->set_success(false);
+        return Status::CANCELLED;
+    }
+    result->set_message("Loading succeeded");
+    result->set_success(true);
+    return Status::OK;
+  }
+  grpc::Status PredictStream(grpc::ServerContext* context, const backend::PredictOptions* request, grpc::ServerWriter<backend::Reply>* writer) override {
+        // Implement the streaming logic here based on the request options
+        // You can use writer->Write(response) to send a reply to the client
+        // and return grpc::Status::OK when the operation is complete.
+        auto lock = llama.lock();
+
+        llama.rewind();
+
+        llama_reset_timings(llama.ctx);
+
+        parse_options_completion(false, request, llama);
+
+        llama.initSampling();
+        llama.loadPrompt(request->prompt());
+        llama.beginCompletion();
+        size_t sent_count = 0;
+        size_t sent_token_probs_index = 0;
+
+        while (llama.has_next_token) {
+            const completion_token_output token_with_probs = llama.doCompletion();
+            if (token_with_probs.tok == -1 || llama.multibyte_pending > 0) {
+                continue;
+            }
+            const std::string token_text = llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+            size_t pos = std::min(sent_count, llama.generated_text.size());
+
+            const std::string str_test = llama.generated_text.substr(pos);
+            bool is_stop_full = false;
+            size_t stop_pos =
+                llama.findStoppingStrings(str_test, token_text.size(), STOP_FULL);
+            if (stop_pos != std::string::npos) {
+                is_stop_full = true;
+                llama.generated_text.erase(
+                    llama.generated_text.begin() + pos + stop_pos,
+                    llama.generated_text.end());
+                pos = std::min(sent_count, llama.generated_text.size());
+            } else {
+                is_stop_full = false;
+                stop_pos = llama.findStoppingStrings(str_test, token_text.size(),
+                    STOP_PARTIAL);
+            }
+
+            if (
+                stop_pos == std::string::npos ||
+                // Send rest of the text if we are at the end of the generation
+                (!llama.has_next_token && !is_stop_full && stop_pos > 0)
+            ) {
+                const std::string to_send = llama.generated_text.substr(pos, std::string::npos);
+
+                sent_count += to_send.size();
+
+                std::vector<completion_token_output> probs_output = {};
+
+                if (llama.params.sparams.n_probs > 0) {
+                    const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
+                    size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
+                    size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
+                    if (probs_pos < probs_stop_pos) {
+                        probs_output = std::vector<completion_token_output>(llama.generated_token_probs.begin() + probs_pos, llama.generated_token_probs.begin() + probs_stop_pos);
+                    }
+                    sent_token_probs_index = probs_stop_pos;
+                }
+                backend::Reply reply;
+                reply.set_message(to_send);
+
+                // Send the reply
+                writer->Write(reply);
+            }
+        }
+
+        llama_print_timings(llama.ctx);
+
+        llama.mutex.unlock();
+        lock.release();
+        return grpc::Status::OK;
+    }
+
+
+    grpc::Status Predict(ServerContext* context, const backend::PredictOptions* request, backend::Reply* reply) {
+        auto lock = llama.lock();
+        llama.rewind();
+        llama_reset_timings(llama.ctx);
+        parse_options_completion(false, request, llama);
+
+        llama.initSampling();
+        llama.loadPrompt(request->prompt());
+        llama.beginCompletion();
+
+        if (llama.params.n_beams) {
+            // Fill llama.generated_token_probs vector with final beam.
+            llama_beam_search(llama.ctx, beam_search_callback, &llama, llama.params.n_beams,
+                                llama.n_past, llama.n_remain);
+            // Translate llama.generated_token_probs to llama.generated_text.
+            append_to_generated_text_from_generated_token_probs(llama);
+        } else {
+            size_t stop_pos = std::string::npos;
+
+            while (llama.has_next_token) {
+                const completion_token_output token_with_probs = llama.doCompletion();
+                const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(llama.ctx, token_with_probs.tok);
+
+                stop_pos = llama.findStoppingStrings(llama.generated_text,
+                    token_text.size(), STOP_FULL);
+            }
+
+            if (stop_pos == std::string::npos) {
+                stop_pos = llama.findStoppingStrings(llama.generated_text, 0, STOP_PARTIAL);
+            }
+            if (stop_pos != std::string::npos) {
+                llama.generated_text.erase(llama.generated_text.begin() + stop_pos,
+                    llama.generated_text.end());
+            }
+        }
+
+        auto probs = llama.generated_token_probs;
+        if (llama.params.sparams.n_probs > 0 && llama.stopped_word) {
+            const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
+            probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
+        }
+        reply->set_message(llama.generated_text);
+        return grpc::Status::OK;
+    }
+};
+
+void RunServer(const std::string& server_address) {
+  BackendServiceImpl service;
+
+  ServerBuilder builder;
+  builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
+  builder.RegisterService(&service);
+
+  std::unique_ptr<Server> server(builder.BuildAndStart());
+  std::cout << "Server listening on " << server_address << std::endl;
+  server->Wait();
+}
+
+int main(int argc, char** argv) {
+  std::string server_address("localhost:50051");
+
+  // Define long and short options
+  struct option long_options[] = {
+      {"addr", required_argument, nullptr, 'a'},
+      {nullptr, 0, nullptr, 0}
+  };
+
+  // Parse command-line arguments
+  int option;
+  int option_index = 0;
+  while ((option = getopt_long(argc, argv, "a:", long_options, &option_index)) != -1) {
+    switch (option) {
+      case 'a':
+        server_address = optarg;
+        break;
+      default:
+        std::cerr << "Usage: " << argv[0] << " [--addr=<address>] or [-a <address>]" << std::endl;
+        return 1;
+    }
+  }
+
+  RunServer(server_address);
+  return 0;
+}

cmd/grpc/bert-embeddings/main.go

@@ -0,0 +1,22 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	bert "github.com/go-skynet/LocalAI/pkg/backend/llm/bert"
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &bert.Embeddings{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/bloomz/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	bloomz "github.com/go-skynet/LocalAI/pkg/backend/llm/bloomz"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &bloomz.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/dolly/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Dolly{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/falcon-ggml/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Falcon{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/falcon/main.go

@@ -0,0 +1,25 @@
+package main
+
+// GRPC Falcon server
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	falcon "github.com/go-skynet/LocalAI/pkg/backend/llm/falcon"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &falcon.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/gpt2/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPT2{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/gpt4all/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	gpt4all "github.com/go-skynet/LocalAI/pkg/backend/llm/gpt4all"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &gpt4all.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/gptj/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPTJ{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/gptneox/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.GPTNeoX{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/langchain-huggingface/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	langchain "github.com/go-skynet/LocalAI/pkg/backend/llm/langchain"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &langchain.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/llama-stable/main.go

@@ -0,0 +1,21 @@
+package main
+
+import (
+	"flag"
+
+	llama "github.com/go-skynet/LocalAI/pkg/backend/llm/llama-stable"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &llama.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/llama/main.go

@@ -0,0 +1,25 @@
+package main
+
+// GRPC Falcon server
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	llama "github.com/go-skynet/LocalAI/pkg/backend/llm/llama"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &llama.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/mpt/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.MPT{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/piper/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	tts "github.com/go-skynet/LocalAI/pkg/backend/tts"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &tts.Piper{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/replit/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Replit{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/rwkv/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	rwkv "github.com/go-skynet/LocalAI/pkg/backend/llm/rwkv"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &rwkv.LLM{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/stablediffusion/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	image "github.com/go-skynet/LocalAI/pkg/backend/image"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &image.StableDiffusion{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/starcoder/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transformers "github.com/go-skynet/LocalAI/pkg/backend/llm/transformers"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transformers.Starcoder{}); err != nil {
+		panic(err)
+	}
+}

cmd/grpc/whisper/main.go

@@ -0,0 +1,23 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+
+import (
+	"flag"
+
+	transcribe "github.com/go-skynet/LocalAI/pkg/backend/transcribe"
+
+	grpc "github.com/go-skynet/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &transcribe.Whisper{}); err != nil {
+		panic(err)
+	}
+}

docker-compose.yaml

@@ -5,11 +5,12 @@ services:
     image: quay.io/go-skynet/local-ai:latest
     build:
       context: .
-      dockerfile: Dockerfile.dev
+      dockerfile: Dockerfile
     ports:
       - 8080:8080
     env_file:
       - .env
     volumes:
       - ./models:/models:cached
+      - ./images/:/tmp/generated/images/
     command: ["/usr/bin/local-ai" ]

entrypoint.sh

@@ -1,9 +1,40 @@
 #!/bin/bash
+set -e
 
 cd /build
 
 if [ "$REBUILD" != "false" ]; then
-	make rebuild
+	rm -rf ./local-ai
+	ESPEAK_DATA=/build/lib/Linux-$(uname -m)/piper_phonemize/lib/espeak-ng-data make build -j${BUILD_PARALLELISM:-1}
+else
+	echo "@@@@@"
+	echo "Skipping rebuild"
+	echo "@@@@@"
+	echo "If you are experiencing issues with the pre-compiled builds, try setting REBUILD=true"
+	echo "If you are still experiencing issues with the build, try setting CMAKE_ARGS and disable the instructions set as needed:"
+	echo 'CMAKE_ARGS="-DLLAMA_F16C=OFF -DLLAMA_AVX512=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF"'
+	echo "see the documentation at: https://localai.io/basics/build/index.html"
+	echo "Note: See also https://github.com/go-skynet/LocalAI/issues/288"
+	echo "@@@@@"
+	echo "CPU info:"
+	grep -e "model\sname" /proc/cpuinfo | head -1
+	grep -e "flags" /proc/cpuinfo | head -1
+	if grep -q -e "\savx\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX    found OK"
+	else
+		echo "CPU: no AVX    found"
+	fi
+	if grep -q -e "\savx2\s" /proc/cpuinfo ; then
+		echo "CPU:    AVX2   found OK"
+	else
+		echo "CPU: no AVX2   found"
+	fi
+	if grep -q -e "\savx512" /proc/cpuinfo ; then
+		echo "CPU:    AVX512 found OK"
+	else
+		echo "CPU: no AVX512 found"
+	fi
+	echo "@@@@@"
 fi
 
-./local-ai "$@"
+./local-ai "$@"

examples/README.md

@@ -1,9 +1,25 @@
 # Examples
 
+| [ChatGPT OSS alternative](https://github.com/go-skynet/LocalAI/tree/master/examples/chatbot-ui)                                                                                                                | [Image generation](https://localai.io/api-endpoints/index.html#image-generation)                                                                                                              |
+|------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------|
+|  ![Screenshot from 2023-04-26 23-59-55](https://user-images.githubusercontent.com/2420543/234715439-98d12e03-d3ce-4f94-ab54-2b256808e05e.png)            | ![b6441997879](https://github.com/go-skynet/LocalAI/assets/2420543/d50af51c-51b7-4f39-b6c2-bf04c403894c)                  |
+
+|                                                                    [Telegram bot](https://github.com/go-skynet/LocalAI/tree/master/examples/telegram-bot)   | [Flowise](https://github.com/go-skynet/LocalAI/tree/master/examples/flowise)                                                                                                                     |
+|------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------|
+![Screenshot from 2023-06-09 00-36-26](https://github.com/go-skynet/LocalAI/assets/2420543/e98b4305-fa2d-41cf-9d2f-1bb2d75ca902)   |  ![Screenshot from 2023-05-30 18-01-03](https://github.com/go-skynet/LocalAI/assets/2420543/02458782-0549-4131-971c-95ee56ec1af8)|    |
+
 Here is a list of projects that can easily be integrated with the LocalAI backend. 
 
+
 ### Projects
 
+### AutoGPT
+
+_by [@mudler](https://github.com/mudler)_
+
+This example shows how to use AutoGPT with LocalAI.
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/autoGPT/)
 
 ### Chatbot-UI
 
@@ -15,6 +31,24 @@ This integration shows how to use LocalAI with [mckaywrigley/chatbot-ui](https:/
 
 [Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/chatbot-ui/)
 
+There is also a separate example to show how to manually setup a model: [example](https://github.com/go-skynet/LocalAI/tree/master/examples/chatbot-ui-manual/)
+
+### K8sGPT
+
+_by [@mudler](https://github.com/mudler)_
+
+This example show how to use LocalAI inside Kubernetes with [k8sgpt](https://k8sgpt.ai).
+
+![Screenshot from 2023-06-19 23-58-47](https://github.com/go-skynet/go-ggml-transformers.cpp/assets/2420543/cab87409-ee68-44ae-8d53-41627fb49509)
+
+### Flowise
+
+_by [@mudler](https://github.com/mudler)_
+
+This example shows how to use [FlowiseAI/Flowise](https://github.com/FlowiseAI/Flowise) with LocalAI.
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/flowise/)
+
 ### Discord bot
 
 _by [@mudler](https://github.com/mudler)_
@@ -39,6 +73,14 @@ A ready to use example to show e2e how to integrate LocalAI with langchain
 
 [Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/langchain-python/)
 
+### LocalAI functions
+
+_by [@mudler](https://github.com/mudler)_
+
+A ready to use example to show how to use OpenAI functions with LocalAI
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/functions/)
+
 ### LocalAI WebUI
 
 _by [@dhruvgera](https://github.com/dhruvgera)_
@@ -73,6 +115,14 @@ Run a slack bot which lets you talk directly with a model
 
 [Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/slack-bot/)
 
+### Slack bot (Question answering)
+
+_by [@mudler](https://github.com/mudler)_
+
+Run a slack bot, ideally for teams, which lets you ask questions on a documentation website, or a github repository.
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/slack-qa-bot/)
+
 ### Question answering on documents with llama-index
 
 _by [@mudler](https://github.com/mudler)_
@@ -89,6 +139,16 @@ Shows how to integrate with `Langchain` and `Chroma` to enable question answerin
 
 [Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/langchain-chroma/)
 
+### Telegram bot
+
+_by [@mudler](https://github.com/mudler)
+
+![Screenshot from 2023-06-09 00-36-26](https://github.com/go-skynet/LocalAI/assets/2420543/e98b4305-fa2d-41cf-9d2f-1bb2d75ca902)
+
+Use LocalAI to power a Telegram bot assistant, with Image generation and audio support!
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/telegram-bot/)
+
 ### Template for Runpod.io
 
 _by [@fHachenberg](https://github.com/fHachenberg)_
@@ -97,6 +157,26 @@ Allows to run any LocalAI-compatible model as a backend on the servers of https:
 
 [Check it out here](https://runpod.io/gsc?template=uv9mtqnrd0&ref=984wlcra)
 
+### Continue
+
+_by [@gruberdev](https://github.com/gruberdev)_
+
+<img src="continue/img/screen.png" width="600" height="200" alt="Screenshot">
+
+Demonstrates how to integrate an open-source copilot alternative that enhances code analysis, completion, and improvements. This approach seamlessly integrates with any LocalAI model, offering a more user-friendly experience.
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/continue/)
+
+### Streamlit bot
+
+_by [@majoshi1](https://github.com/majoshi1)_
+
+![Screenshot](streamlit-bot/streamlit-bot.png)
+
+A chat bot made using `Streamlit` & LocalAI.
+
+[Check it out here](https://github.com/go-skynet/LocalAI/tree/master/examples/streamlit-bot/)
+
 ## Want to contribute?
 
 Create an issue, and put `Example: <description>` in the title! We will post your examples here.

examples/autoGPT/.env.example

@@ -0,0 +1,9 @@
+# CPU .env docs: https://localai.io/howtos/easy-setup-docker-cpu/
+# GPU .env docs: https://localai.io/howtos/easy-setup-docker-gpu/
+
+OPENAI_API_KEY=sk---anystringhere
+OPENAI_API_BASE=http://api:8080/v1
+# Models to preload at start
+# Here we configure gpt4all as gpt-3.5-turbo and bert as embeddings,
+# see other options in the model gallery at https://github.com/go-skynet/model-gallery
+PRELOAD_MODELS=[{"url": "github:go-skynet/model-gallery/gpt4all-j.yaml", "name": "gpt-3.5-turbo"}, { "url": "github:go-skynet/model-gallery/bert-embeddings.yaml", "name": "text-embedding-ada-002"}]

examples/autoGPT/README.md

@@ -0,0 +1,36 @@
+# AutoGPT
+
+Example of integration with [AutoGPT](https://github.com/Significant-Gravitas/Auto-GPT).
+
+## Run
+
+```bash
+# Clone LocalAI
+git clone https://github.com/go-skynet/LocalAI
+
+cd LocalAI/examples/autoGPT
+
+cp -rfv .env.example .env
+
+# Edit the .env file to set a different model by editing `PRELOAD_MODELS`.
+vim .env
+
+docker-compose run --rm auto-gpt
+```
+
+Note: The example automatically downloads the `gpt4all` model as it is under a permissive license. The GPT4All model does not seem to be enough to run AutoGPT. WizardLM-7b-uncensored seems to perform better (with `f16: true`).
+
+
+## Without docker
+
+Run AutoGPT with `OPENAI_API_BASE` pointing to the LocalAI endpoint. If you run it locally for instance:
+
+```
+OPENAI_API_BASE=http://localhost:8080 python ...
+```
+
+Note: you need a model named `gpt-3.5-turbo` and `text-embedding-ada-002`. You can preload those in LocalAI at start by setting in the env:
+
+```
+PRELOAD_MODELS=[{"url": "github:go-skynet/model-gallery/gpt4all-j.yaml", "name": "gpt-3.5-turbo"}, { "url": "github:go-skynet/model-gallery/bert-embeddings.yaml", "name": "text-embedding-ada-002"}]
+```

examples/autoGPT/docker-compose.yaml

@@ -0,0 +1,42 @@
+version: "3.9"
+services:
+  api:
+    image: quay.io/go-skynet/local-ai:latest
+    ports:
+      - 8080:8080
+    env_file:
+      - .env
+    environment:
+      - DEBUG=true
+      - MODELS_PATH=/models
+    volumes:
+      - ./models:/models:cached
+    command: ["/usr/bin/local-ai" ]
+  auto-gpt:
+    image: significantgravitas/auto-gpt
+    depends_on:
+      api:
+        condition: service_healthy
+      redis:
+        condition: service_started
+    env_file:
+      - .env
+    environment:
+      MEMORY_BACKEND: ${MEMORY_BACKEND:-redis}
+      REDIS_HOST: ${REDIS_HOST:-redis}
+    profiles: ["exclude-from-up"]
+    volumes:
+      - ./auto_gpt_workspace:/app/autogpt/auto_gpt_workspace
+      - ./data:/app/data
+      ## allow auto-gpt to write logs to disk
+      - ./logs:/app/logs
+      ## uncomment following lines if you want to make use of these files
+      ## you must have them existing in the same folder as this docker-compose.yml
+      #- type: bind
+      #  source: ./azure.yaml
+      #  target: /app/azure.yaml
+      #- type: bind
+      #  source: ./ai_settings.yaml
+      #  target: /app/ai_settings.yaml
+  redis:
+    image: "redis/redis-stack-server:latest"

examples/bruno/LocalAI Test Requests/backend monitor/backend monitor.bru

@@ -0,0 +1,11 @@
+meta {
+  name: backend monitor
+  type: http
+  seq: 4
+}
+
+get {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/backend/monitor
+  body: none
+  auth: none
+}

examples/bruno/LocalAI Test Requests/backend monitor/backend-shutdown.bru

@@ -0,0 +1,21 @@
+meta {
+  name: backend-shutdown
+  type: http
+  seq: 3
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/backend/shutdown
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "model": "{{DEFAULT_MODEL}}"
+  }
+}

examples/bruno/LocalAI Test Requests/bruno.json

@@ -0,0 +1,5 @@
+{
+  "version": "1",
+  "name": "LocalAI Test Requests",
+  "type": "collection"
+}

examples/bruno/LocalAI Test Requests/environments/localhost.bru

@@ -0,0 +1,6 @@
+vars {
+  HOST: localhost
+  PORT: 8080
+  DEFAULT_MODEL: gpt-3.5-turbo
+  PROTOCOL: http://
+}

examples/bruno/LocalAI Test Requests/get models list.bru

@@ -0,0 +1,11 @@
+meta {
+  name: get models list
+  type: http
+  seq: 2
+}
+
+get {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models
+  body: none
+  auth: none
+}

examples/bruno/LocalAI Test Requests/image generation/Generate image.bru

@@ -0,0 +1,25 @@
+meta {
+  name: Generate image
+  type: http
+  seq: 1
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/v1/images/generations
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+    "prompt": "<positive prompt>|<negative prompt>",
+    "model": "model-name",
+    "step": 51,
+    "size": "1024x1024",
+    "image": ""
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/-completions.bru

@@ -0,0 +1,24 @@
+meta {
+  name: -completions
+  type: http
+  seq: 4
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/completions
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "model": "{{DEFAULT_MODEL}}",
+      "prompt": "function downloadFile(string url, string outputPath) {",
+      "max_tokens": 256,
+      "temperature": 0.5
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/-edits.bru

@@ -0,0 +1,23 @@
+meta {
+  name: -edits
+  type: http
+  seq: 5
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/edits
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "model": "{{DEFAULT_MODEL}}",
+      "input": "What day of the wek is it?",
+      "instruction": "Fix the spelling mistakes"
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/-embeddings.bru

@@ -0,0 +1,22 @@
+meta {
+  name: -embeddings
+  type: http
+  seq: 6
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/embeddings
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "model": "{{DEFAULT_MODEL}}",
+      "input": "A STRANGE GAME.\nTHE ONLY WINNING MOVE IS NOT TO PLAY.\n\nHOW ABOUT A NICE GAME OF CHESS?"
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/chat/chat completion -simple- 1 message-.bru

@@ -0,0 +1,30 @@
+meta {
+  name: chat completion -simple- 1 message-
+  type: http
+  seq: 4
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/chat/completions
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+    "model": "{{DEFAULT_MODEL}}",
+    "messages": [
+      {
+        "role": "user",
+        "content": "How could one use friction to cook an egg?"
+      }
+    ],
+    "max_tokens": 256,
+    "temperature": 0.2,
+    "grammar": ""
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/chat/chat-completions -long-.bru

@@ -0,0 +1,29 @@
+meta {
+  name: chat-completions -long-
+  type: http
+  seq: 5
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/chat/completions
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+       "model": "{{DEFAULT_MODEL}}",
+       "messages": [{"role": "system", "content": "You are a helpful, intelligent pirate assistant. Always answer as helpfully as possible, while being a pirate and using pirate language. If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."},
+           {"role": "user", "content": "How could one use electricity to cook an egg?"},
+           {"role": "assistant",
+                  "content": "Shiver me timbers! Using electricity to cook an egg? Well, matey, I reckon that be a right curious idea! *adjusts spectacles* Now, I ain't sure if this be possible, but I'll do me best to help ye out.\n\nFirst things first, ye gotta understand that electricity be a powerful force, and it can't just cook an egg on its own. Ye see, electricity be like a mighty wind that can make things happen, but it needs somethin' to work with. So, if ye want to use electricity to cook an egg, ye gotta find a way to harness that power and make it do the cookin'.\n\nNow, I know what ye might be thinkin': \"How do I harness the power of electricity to cook an egg?\" Well, matey, there be a few ways to do it. One way be to use a special device called an \"electric frying pan.\" This be a pan that has a built-in heating element that gets hot when ye plug it into a wall socket. When the element gets hot, ye can crack an egg into the pan and watch as it cook"
+              },
+              {"role": "user", "content": "I don't have one of those, just a raw wire and plenty of power! How do we get it done?"}],
+       "max_tokens": 1024,
+       "temperature": 0.5
+  }
+}

examples/bruno/LocalAI Test Requests/llm text/chat/chat-completions -stream-.bru

@@ -0,0 +1,25 @@
+meta {
+  name: chat-completions -stream-
+  type: http
+  seq: 6
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/chat/completions
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+       "model": "{{DEFAULT_MODEL}}",
+       "messages": [{"role": "user", "content": "Explain how I can set sail on the ocean using only power generated by seagulls?"}],
+       "max_tokens": 256,
+       "temperature": 0.9,
+       "stream": true
+  }
+}

examples/bruno/LocalAI Test Requests/model gallery/add model gallery.bru

@@ -0,0 +1,22 @@
+meta {
+  name: add model gallery
+  type: http
+  seq: 10
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/galleries
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "url": "file:///home/dave/projects/model-gallery/huggingface/TheBloke__CodeLlama-7B-Instruct-GGML.yaml",
+      "name": "test"
+  }
+}

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

@@ -0,0 +1,21 @@
+meta {
+  name: delete model gallery
+  type: http
+  seq: 11
+}
+
+delete {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/galleries
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "name": "test"
+  }
+}

examples/bruno/LocalAI Test Requests/model gallery/list MODELS in galleries.bru

@@ -0,0 +1,11 @@
+meta {
+  name: list MODELS in galleries
+  type: http
+  seq: 7
+}
+
+get {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/available
+  body: none
+  auth: none
+}

examples/bruno/LocalAI Test Requests/model gallery/list model GALLERIES.bru

@@ -0,0 +1,11 @@
+meta {
+  name: list model GALLERIES
+  type: http
+  seq: 8
+}
+
+get {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/galleries
+  body: none
+  auth: none
+}

examples/bruno/LocalAI Test Requests/model gallery/model gallery apply -gist-.bru

@@ -0,0 +1,21 @@
+meta {
+  name: model gallery apply -gist-
+  type: http
+  seq: 12
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/apply
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "id": "TheBloke__CodeLlama-7B-Instruct-GGML__codellama-7b-instruct.ggmlv3.Q2_K.bin"
+  }
+}

examples/bruno/LocalAI Test Requests/model gallery/model gallery apply.bru

@@ -0,0 +1,22 @@
+meta {
+  name: model gallery apply
+  type: http
+  seq: 9
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/models/apply
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "id": "dave@TheBloke__CodeLlama-7B-Instruct-GGML__codellama-7b-instruct.ggmlv3.Q3_K_S.bin",
+      "name": "codellama7b"
+  }
+}