WIP speculative

.bruno/LocalAI Test Requests/Sound Generation/musicgen.bru

@@ -0,0 +1,23 @@
+meta {
+  name: musicgen
+  type: http
+  seq: 1
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/v1/sound-generation
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "model_id": "facebook/musicgen-small",
+      "text": "Exciting 80s Newscast Interstitial",
+      "duration_seconds": 8
+  }
+}

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

@@ -0,0 +1,17 @@
+meta {
+  name: backend monitor
+  type: http
+  seq: 4
+}
+
+get {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/backend/monitor
+  body: json
+  auth: none
+}
+
+body:json {
+  {
+    "model": "{{DEFAULT_MODEL}}"
+  }
+}

.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}}"
+  }
+}

.bruno/LocalAI Test Requests/bruno.json

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

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

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

.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
+}

.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": ""
+  }
+}

.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
+  }
+}

.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"
+  }
+}

.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?"
+  }
+}

.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": ""
+  }
+}

.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
+  }
+}

.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
+  }
+}

.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"
+  }
+}

.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"
+  }
+}

.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
+}

.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
+}

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

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

.bruno/LocalAI Test Requests/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"
+  }
+}

.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"
+  }
+}

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

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

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

@@ -0,0 +1,22 @@
+meta {
+  name: -tts
+  type: http
+  seq: 2
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/tts
+  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?"
+  }
+}

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

@@ -0,0 +1,23 @@
+meta {
+  name: musicgen
+  type: http
+  seq: 2
+}
+
+post {
+  url: {{PROTOCOL}}{{HOST}}:{{PORT}}/tts
+  body: json
+  auth: none
+}
+
+headers {
+  Content-Type: application/json
+}
+
+body:json {
+  {
+      "backend": "transformers",
+      "model": "facebook/musicgen-small",
+      "input": "80s Synths playing Jazz"
+  }
+}

.env

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

.github/dependabot.yml

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

.github/labeler.yml

@@ -1,4 +1,4 @@
-enhancement:
+enhancements:
  - head-branch: ['^feature', 'feature']
 
 dependencies:

.github/workflows/bump_deps.yaml

@@ -9,10 +9,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - repository: "ggml-org/llama.cpp"
+          - repository: "ggerganov/llama.cpp"
             variable: "CPPLLAMA_VERSION"
             branch: "master"
-          - repository: "ggml-org/whisper.cpp"
+          - repository: "ggerganov/whisper.cpp"
             variable: "WHISPER_CPP_VERSION"
             branch: "master"
           - repository: "PABannier/bark.cpp"

.github/workflows/dependabot_auto.yml

@@ -14,7 +14,7 @@ jobs:
     steps:
       - name: Dependabot metadata
         id: metadata
-        uses: dependabot/fetch-metadata@v2.4.0
+        uses: dependabot/fetch-metadata@v2.2.0
         with:
           github-token: "${{ secrets.GITHUB_TOKEN }}"
           skip-commit-verification: true

.github/workflows/deploy-explorer.yaml

@@ -33,7 +33,7 @@ jobs:
         run: |
           CGO_ENABLED=0 make build-api
       - name: rm
-        uses: appleboy/ssh-action@v1.2.2
+        uses: appleboy/ssh-action@v1.2.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}
@@ -42,7 +42,7 @@ jobs:
             script: |
                 sudo rm -rf local-ai/ || true
       - name: copy file via ssh
-        uses: appleboy/scp-action@v1.0.0
+        uses: appleboy/scp-action@v0.1.7
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}
@@ -53,7 +53,7 @@ jobs:
             rm: true
             target: ./local-ai
       - name: restarting
-        uses: appleboy/ssh-action@v1.2.2
+        uses: appleboy/ssh-action@v1.2.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}

.github/workflows/generate_grpc_cache.yaml

@@ -2,10 +2,9 @@ name: 'generate and publish GRPC docker caches'
 
 on:
   workflow_dispatch:
-
-  schedule:
-    # daily at midnight
-    - cron: '0 0 * * *'
+  push:
+    branches:
+      - master
 
 concurrency:
   group: grpc-cache-${{ github.head_ref || github.ref }}-${{ github.repository }}
@@ -17,7 +16,7 @@ jobs:
       matrix:
         include:
           - grpc-base-image: ubuntu:22.04
-            runs-on: 'arc-runner-set'
+            runs-on: 'ubuntu-latest'
             platforms: 'linux/amd64,linux/arm64'
     runs-on: ${{matrix.runs-on}}
     steps:

.github/workflows/generate_intel_image.yaml

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

.github/workflows/image-pr.yml

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

.github/workflows/image.yml

@@ -45,13 +45,13 @@ jobs:
           - build-type: 'hipblas'
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-hipblas-extras'
+            tag-suffix: '-hipblas-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             aio: "-aio-gpu-hipblas"
             base-image: "rocm/dev-ubuntu-22.04:6.1"
             grpc-base-image: "ubuntu:22.04"
-            latest-image: 'latest-gpu-hipblas-extras'
+            latest-image: 'latest-gpu-hipblas'
             latest-image-aio: 'latest-aio-gpu-hipblas'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
@@ -59,13 +59,32 @@ jobs:
             platforms: 'linux/amd64'
             tag-latest: 'false'
             tag-suffix: '-hipblas'
+            ffmpeg: 'false'
+            image-type: 'extras'
+            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            grpc-base-image: "ubuntu:22.04"
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'hipblas'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-hipblas-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "rocm/dev-ubuntu-22.04:6.1"
             grpc-base-image: "ubuntu:22.04"
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
-            latest-image: 'latest-gpu-hipblas'
+          - build-type: 'hipblas'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-hipblas-core'
+            ffmpeg: 'false'
+            image-type: 'core'
+            base-image: "rocm/dev-ubuntu-22.04:6.1"
+            grpc-base-image: "ubuntu:22.04"
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=3 --output-sync=target"
   self-hosted-jobs:
     uses: ./.github/workflows/image_build.yml
     with:
@@ -95,18 +114,60 @@ jobs:
       max-parallel: ${{ github.event_name != 'pull_request' && 5 || 8 }}
       matrix:
         include:
+          # Extra images
+          - build-type: ''
+            #platforms: 'linux/amd64,linux/arm64'
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: ''
+            ffmpeg: ''
+            image-type: 'extras'
+            runs-on: 'arc-runner-set'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: ''
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: '-ffmpeg'
+            ffmpeg: 'true'
+            image-type: 'extras'
+            runs-on: 'arc-runner-set'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'cublas'
+            cuda-major-version: "11"
+            cuda-minor-version: "7"
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11'
+            ffmpeg: ''
+            image-type: 'extras'
+            runs-on: 'arc-runner-set'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12'
+            ffmpeg: ''
+            image-type: 'extras'
+            runs-on: 'arc-runner-set'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-cublas-cuda11-extras'
+            tag-suffix: '-cublas-cuda11-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             aio: "-aio-gpu-nvidia-cuda-11"
-            latest-image: 'latest-gpu-nvidia-cuda-11-extras'
+            latest-image: 'latest-gpu-nvidia-cuda-11'
             latest-image-aio: 'latest-aio-gpu-nvidia-cuda-11'
             makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'cublas'
@@ -114,26 +175,36 @@ jobs:
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
             tag-latest: 'auto'
-            tag-suffix: '-cublas-cuda12-extras'
+            tag-suffix: '-cublas-cuda12-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             aio: "-aio-gpu-nvidia-cuda-12"
-            latest-image: 'latest-gpu-nvidia-cuda-12-extras'
+            latest-image: 'latest-gpu-nvidia-cuda-12'
             latest-image-aio: 'latest-aio-gpu-nvidia-cuda-12'
             makeflags: "--jobs=3 --output-sync=target"
+          - build-type: ''
+            #platforms: 'linux/amd64,linux/arm64'
+            platforms: 'linux/amd64'
+            tag-latest: 'auto'
+            tag-suffix: ''
+            ffmpeg: ''
+            image-type: 'extras'
+            base-image: "ubuntu:22.04"
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'sycl_f16'
             platforms: 'linux/amd64'
             tag-latest: 'auto'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f16-extras'
+            tag-suffix: '-sycl-f16-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             aio: "-aio-gpu-intel-f16"
-            latest-image: 'latest-gpu-intel-f16-extras'
+            latest-image: 'latest-gpu-intel-f16'
             latest-image-aio: 'latest-aio-gpu-intel-f16'
             makeflags: "--jobs=3 --output-sync=target"
           - build-type: 'sycl_f32'
@@ -141,12 +212,12 @@ jobs:
             tag-latest: 'auto'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f32-extras'
+            tag-suffix: '-sycl-f32-ffmpeg'
             ffmpeg: 'true'
             image-type: 'extras'
             runs-on: 'arc-runner-set'
             aio: "-aio-gpu-intel-f32"
-            latest-image: 'latest-gpu-intel-f32-extras'
+            latest-image: 'latest-gpu-intel-f32'
             latest-image-aio: 'latest-aio-gpu-intel-f32'
             makeflags: "--jobs=3 --output-sync=target"
           # Core images
@@ -155,23 +226,41 @@ jobs:
             tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f16'
-            ffmpeg: 'true'
+            tag-suffix: '-sycl-f16-core'
+            ffmpeg: 'false'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
-            latest-image: 'latest-gpu-intel-f16'
           - build-type: 'sycl_f32'
             platforms: 'linux/amd64'
             tag-latest: 'false'
             base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
             grpc-base-image: "ubuntu:22.04"
-            tag-suffix: '-sycl-f32'
+            tag-suffix: '-sycl-f32-core'
+            ffmpeg: 'false'
+            image-type: 'core'
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'sycl_f16'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
+            grpc-base-image: "ubuntu:22.04"
+            tag-suffix: '-sycl-f16-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=3 --output-sync=target"
+          - build-type: 'sycl_f32'
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            base-image: "quay.io/go-skynet/intel-oneapi-base:latest"
+            grpc-base-image: "ubuntu:22.04"
+            tag-suffix: '-sycl-f32-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=3 --output-sync=target"
-            latest-image: 'latest-gpu-intel-f32'
 
   core-image-build:
     uses: ./.github/workflows/image_build.yml
@@ -204,7 +293,7 @@ jobs:
           - build-type: ''
             platforms: 'linux/amd64,linux/arm64'
             tag-latest: 'auto'
-            tag-suffix: ''
+            tag-suffix: '-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "ubuntu:22.04"
@@ -219,38 +308,60 @@ jobs:
             cuda-minor-version: "7"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11'
-            ffmpeg: 'true'
+            tag-suffix: '-cublas-cuda11-core'
+            ffmpeg: ''
             image-type: 'core'
-            runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
+            runs-on: 'arc-runner-set'
             makeflags: "--jobs=4 --output-sync=target"
             skip-drivers: 'false'
-            latest-image: 'latest-gpu-nvidia-cuda-12'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12'
+            tag-suffix: '-cublas-cuda12-core'
+            ffmpeg: ''
+            image-type: 'core'
+            base-image: "ubuntu:22.04"
+            runs-on: 'arc-runner-set'
+            makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
+          - build-type: 'cublas'
+            cuda-major-version: "11"
+            cuda-minor-version: "7"
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11-ffmpeg-core'
+            ffmpeg: 'true'
+            image-type: 'core'
+            runs-on: 'arc-runner-set'
+            base-image: "ubuntu:22.04"
+            makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
+          - build-type: 'cublas'
+            cuda-major-version: "12"
+            cuda-minor-version: "0"
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
-            latest-image: 'latest-gpu-nvidia-cuda-12'
           - build-type: 'vulkan'
             platforms: 'linux/amd64'
             tag-latest: 'false'
-            tag-suffix: '-vulkan'
+            tag-suffix: '-vulkan-ffmpeg-core'
+            latest-image: 'latest-vulkan-ffmpeg-core'
             ffmpeg: 'true'
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
             skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
-            latest-image: 'latest-gpu-vulkan'
   gh-runner:
     uses: ./.github/workflows/image_build.yml
     with:
@@ -283,8 +394,8 @@ jobs:
             cuda-minor-version: "0"
             platforms: 'linux/arm64'
             tag-latest: 'false'
-            tag-suffix: '-nvidia-l4t-arm64'
-            latest-image: 'latest-nvidia-l4t-arm64'
+            tag-suffix: '-nvidia-l4t-arm64-core'
+            latest-image: 'latest-nvidia-l4t-arm64-core'
             ffmpeg: 'true'
             image-type: 'core'
             base-image: "nvcr.io/nvidia/l4t-jetpack:r36.4.0"

.github/workflows/image_build.yml

@@ -310,11 +310,6 @@ jobs:
           tags: ${{ steps.meta_aio_dockerhub.outputs.tags }}
           labels: ${{ steps.meta_aio_dockerhub.outputs.labels }}
 
-      - name: Cleanup
-        run: |
-          docker builder prune -f
-          docker system prune --force --volumes --all
-
       - name: Latest tag
         # run this on branches, when it is a tag and there is a latest-image defined
         if: github.event_name != 'pull_request' && inputs.latest-image != ''  && github.ref_type == 'tag'

.github/workflows/notify-models.yaml

@@ -8,7 +8,7 @@ jobs:
   notify-discord:
     if: ${{ (github.event.pull_request.merged == true) && (contains(github.event.pull_request.labels.*.name, 'area/ai-model')) }}
     env:
-        MODEL_NAME: gemma-3-12b-it
+        MODEL_NAME: hermes-2-theta-llama-3-8b
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
@@ -16,9 +16,9 @@ jobs:
         fetch-depth: 0 # needed to checkout all branches for this Action to work
     - uses: mudler/localai-github-action@v1
       with:
-        model: 'gemma-3-12b-it' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
+        model: 'hermes-2-theta-llama-3-8b' # Any from models.localai.io, or from huggingface.com with: "huggingface://<repository>/file"
         # Check the PR diff using the current branch and the base branch of the PR
-    - uses: GrantBirki/git-diff-action@v2.8.0
+    - uses: GrantBirki/git-diff-action@v2.7.0
       id: git-diff-action
       with:
             json_diff_file_output: diff.json
@@ -79,7 +79,7 @@ jobs:
         args: ${{ steps.summarize.outputs.message }}
     - name: Setup tmate session if fails
       if: ${{ failure() }}
-      uses: mxschmitt/action-tmate@v3.22
+      uses: mxschmitt/action-tmate@v3.19
       with:
         detached: true
         connect-timeout-seconds: 180
@@ -87,7 +87,7 @@ jobs:
   notify-twitter:
     if: ${{ (github.event.pull_request.merged == true) && (contains(github.event.pull_request.labels.*.name, 'area/ai-model')) }}
     env:
-        MODEL_NAME: gemma-3-12b-it
+        MODEL_NAME: hermes-2-theta-llama-3-8b
     runs-on: ubuntu-latest
     steps:
     - uses: actions/checkout@v4
@@ -99,7 +99,7 @@ jobs:
         docker run -e -ti -d --name local-ai -p 8080:8080 localai/localai:master-ffmpeg-core run --debug $MODEL_NAME
         until [ "`docker inspect -f {{.State.Health.Status}} local-ai`" == "healthy" ]; do echo "Waiting for container to be ready";  docker logs --tail 10 local-ai; sleep 2; done
       # Check the PR diff using the current branch and the base branch of the PR
-    - uses: GrantBirki/git-diff-action@v2.8.0
+    - uses: GrantBirki/git-diff-action@v2.7.0
       id: git-diff-action
       with:
             json_diff_file_output: diff.json
@@ -161,7 +161,7 @@ jobs:
         TWITTER_ACCESS_TOKEN_SECRET: ${{ secrets.TWITTER_ACCESS_TOKEN_SECRET }}
     - name: Setup tmate session if fails
       if: ${{ failure() }}
-      uses: mxschmitt/action-tmate@v3.22
+      uses: mxschmitt/action-tmate@v3.19
       with:
         detached: true
         connect-timeout-seconds: 180

.github/workflows/notify-releases.yaml

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

.github/workflows/release.yaml

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

.github/workflows/secscan.yaml

@@ -18,7 +18,7 @@ jobs:
         if: ${{ github.actor != 'dependabot[bot]' }}
       - name: Run Gosec Security Scanner
         if: ${{ github.actor != 'dependabot[bot]' }}
-        uses: securego/gosec@v2.22.4
+        uses: securego/gosec@v2.22.0
         with:
           # we let the report trigger content trigger a failure using the GitHub Security features.
           args: '-no-fail -fmt sarif -out results.sarif ./...'

.github/workflows/test-extra.yml

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

.github/workflows/test.yml

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

Dockerfile

@@ -15,7 +15,7 @@ ARG TARGETARCH
 ARG TARGETVARIANT
 
 ENV DEBIAN_FRONTEND=noninteractive
-ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,transformers:/build/backend/python/transformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,faster-whisper:/build/backend/python/faster-whisper/run.sh,kokoro:/build/backend/python/kokoro/run.sh,vllm:/build/backend/python/vllm/run.sh,exllama2:/build/backend/python/exllama2/run.sh"
+ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,transformers:/build/backend/python/transformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,autogptq:/build/backend/python/autogptq/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,faster-whisper:/build/backend/python/faster-whisper/run.sh,kokoro:/build/backend/python/kokoro/run.sh,vllm:/build/backend/python/vllm/run.sh,exllama2:/build/backend/python/exllama2/run.sh"
 
 RUN apt-get update && \
     apt-get install -y --no-install-recommends \
@@ -24,7 +24,6 @@ RUN apt-get update && \
         ca-certificates \
         curl libssl-dev \
         git \
-        git-lfs \
         unzip upx-ucl && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/*
@@ -46,10 +45,9 @@ EOT
 RUN curl -L -s https://go.dev/dl/go${GO_VERSION}.linux-${TARGETARCH}.tar.gz | tar -C /usr/local -xz
 ENV PATH=$PATH:/root/go/bin:/usr/local/go/bin
 
-# Install grpc compilers and rice
+# Install grpc compilers
 RUN go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2 && \
-    go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af && \
-    go install github.com/GeertJohan/go.rice/rice@latest
+    go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
 
 COPY --chmod=644 custom-ca-certs/* /usr/local/share/ca-certificates/
 RUN update-ca-certificates
@@ -301,9 +299,10 @@ COPY .git .
 RUN make prepare
 
 ## Build the binary
-## If we're on arm64 AND using cublas/hipblas, skip some of the llama-compat backends to save space
-## Otherwise just run the normal build
-RUN if [ "${TARGETARCH}" = "arm64" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
+## If it's CUDA or hipblas, we want to skip some of the llama-compat backends to save space
+## We only leave the most CPU-optimized variant and the fallback for the cublas/hipblas build
+## (both will use CUDA or hipblas for the actual computation)
+RUN if [ "${BUILD_TYPE}" = "cublas" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
         SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx512 backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
     else \
         make build; \
@@ -431,6 +430,9 @@ RUN if [[ ( "${EXTRA_BACKENDS}" =~ "kokoro" || -z "${EXTRA_BACKENDS}" ) && "$IMA
 RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vllm" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/vllm \
     ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "autogptq" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/autogptq \
+    ; fi && \
     if [[ ( "${EXTRA_BACKENDS}" =~ "bark" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
         make -C backend/python/bark \
     ; fi && \

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023-2025 Ettore Di Giacinto (mudler@localai.io)
+Copyright (c) 2023-2024 Ettore Di Giacinto (mudler@localai.io)
 
 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

@@ -6,11 +6,13 @@ BINARY_NAME=local-ai
 DETECT_LIBS?=true
 
 # llama.cpp versions
-CPPLLAMA_VERSION?=9a390c4829cd3058d26a2e2c09d16e3fd12bf1b1
+GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
+GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
+CPPLLAMA_VERSION?=6152129d05870cb38162c422c6ba80434e021e9f
 
 # whisper.cpp version
-WHISPER_REPO?=https://github.com/ggml-org/whisper.cpp
-WHISPER_CPP_VERSION?=2e310b841e0b4e7cf00890b53411dd9f8578f243
+WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
+WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
 
 # go-piper version
 PIPER_REPO?=https://github.com/mudler/go-piper
@@ -21,11 +23,8 @@ BARKCPP_REPO?=https://github.com/PABannier/bark.cpp.git
 BARKCPP_VERSION?=v1.0.0
 
 # stablediffusion.cpp (ggml)
-STABLEDIFFUSION_GGML_REPO?=https://github.com/richiejp/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=53e3b17eb3d0b5760ced06a1f98320b68b34aaae
-
-# ONEAPI variables for SYCL
-export ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
+STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
+STABLEDIFFUSION_GGML_VERSION?=5eb15ef4d022bef4a391de4f5f6556e81fbb5024
 
 ONNX_VERSION?=1.20.0
 ONNX_ARCH?=x64
@@ -33,12 +32,8 @@ ONNX_OS?=linux
 
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
-export CMAKE_ARGS?=-DBUILD_SHARED_LIBS=OFF
-export WHISPER_CMAKE_ARGS?=-DBUILD_SHARED_LIBS=OFF
+export CMAKE_ARGS?=
 export BACKEND_LIBS?=
-export WHISPER_DIR=$(abspath ./sources/whisper.cpp)
-export WHISPER_INCLUDE_PATH=$(WHISPER_DIR)/include:$(WHISPER_DIR)/ggml/include
-export WHISPER_LIBRARY_PATH=$(WHISPER_DIR)/build/src/:$(WHISPER_DIR)/build/ggml/src
 
 CGO_LDFLAGS?=
 CGO_LDFLAGS_WHISPER?=
@@ -88,7 +83,6 @@ endif
 # IF native is false, we add -DGGML_NATIVE=OFF to CMAKE_ARGS
 ifeq ($(NATIVE),false)
 	CMAKE_ARGS+=-DGGML_NATIVE=OFF
-	WHISPER_CMAKE_ARGS+=-DGGML_NATIVE=OFF
 endif
 
 # Detect if we are running on arm64
@@ -116,31 +110,13 @@ ifeq ($(OS),Darwin)
 	# disable metal if on Darwin and any other value is explicitly passed.
 	else ifneq ($(BUILD_TYPE),metal)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
-		WHISPER_CMAKE_ARGS+=-DGGML_METAL=OFF
 		export GGML_NO_ACCELERATE=1
 		export GGML_NO_METAL=1
-		GO_LDFLAGS_WHISPER+=-lggml-blas
-		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-blas
 	endif
 
 	ifeq ($(BUILD_TYPE),metal)
+#			-lcblas 	removed: it seems to always be listed as a duplicate flag.
 		CGO_LDFLAGS += -framework Accelerate
-		CGO_LDFLAGS_WHISPER+=-lggml-metal -lggml-blas
-		CMAKE_ARGS+=-DGGML_METAL=ON
-		CMAKE_ARGS+=-DGGML_METAL_USE_BF16=ON
-		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
-		CMAKE_ARGS+=-DGGML_OPENMP=OFF
-		WHISPER_CMAKE_ARGS+=-DGGML_METAL=ON
-		WHISPER_CMAKE_ARGS+=-DGGML_METAL_USE_BF16=ON
-		WHISPER_CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
-		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_EXAMPLES=OFF
-		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_TESTS=OFF
-		WHISPER_CMAKE_ARGS+=-DWHISPER_BUILD_SERVER=OFF
-		WHISPER_CMAKE_ARGS+=-DGGML_OPENMP=OFF
-		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-metal/:$(WHISPER_DIR)/build/ggml/src/ggml-blas
-	else
-		CGO_LDFLAGS_WHISPER+=-lggml-blas
-		export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-blas
 	endif
 else
 CGO_LDFLAGS_WHISPER+=-lgomp
@@ -152,29 +128,21 @@ ifeq ($(BUILD_TYPE),openblas)
 endif
 
 ifeq ($(BUILD_TYPE),cublas)
-	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH) -L$(CUDA_LIBPATH)/stubs/ -lcuda
+	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH)
 	export GGML_CUDA=1
-	CMAKE_ARGS+=-DGGML_CUDA=ON
-	WHISPER_CMAKE_ARGS+=-DGGML_CUDA=ON
-	CGO_LDFLAGS_WHISPER+=-lcufft -lggml-cuda
-	export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-cuda/
+	CGO_LDFLAGS_WHISPER+=-L$(CUDA_LIBPATH)/stubs/ -lcuda -lcufft
 endif
 
 ifeq ($(BUILD_TYPE),vulkan)
 	CMAKE_ARGS+=-DGGML_VULKAN=1
-	WHISPER_CMAKE_ARGS+=-DGGML_VULKAN=1
-	CGO_LDFLAGS_WHISPER+=-lggml-vulkan -lvulkan
-	export WHISPER_LIBRARY_PATH:=$(WHISPER_LIBRARY_PATH):$(WHISPER_DIR)/build/ggml/src/ggml-vulkan/
 endif
 
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	export GGML_SYCL=1
-	CMAKE_ARGS+=-DGGML_SYCL=ON
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)
 	export GGML_SYCL_F16=1
-	CMAKE_ARGS+=-DGGML_SYCL_F16=ON
 endif
 
 ifeq ($(BUILD_TYPE),hipblas)
@@ -183,9 +151,10 @@ ifeq ($(BUILD_TYPE),hipblas)
 	LD_LIBRARY_PATH ?= /opt/rocm/lib:/opt/rocm/llvm/lib
 	export CXX=$(ROCM_HOME)/llvm/bin/clang++
 	export CC=$(ROCM_HOME)/llvm/bin/clang
+	# llama-ggml has no hipblas support, so override it here.
 	export STABLE_BUILD_TYPE=
 	export GGML_HIP=1
-	GPU_TARGETS ?= gfx803,gfx900,gfx906,gfx908,gfx90a,gfx942,gfx1010,gfx1030,gfx1032,gfx1100,gfx1101,gfx1102
+	GPU_TARGETS ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
 	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
 	CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(GPU_TARGETS)"
 	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link -L${ROCM_HOME}/lib/llvm/lib
@@ -219,6 +188,7 @@ ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx2
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-avx512
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-fallback
+ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-ggml
 ALL_GRPC_BACKENDS+=backend-assets/grpc/llama-cpp-grpc
 ALL_GRPC_BACKENDS+=backend-assets/util/llama-cpp-rpc-server
 ALL_GRPC_BACKENDS+=backend-assets/grpc/whisper
@@ -252,6 +222,19 @@ endif
 
 all: help
 
+## go-llama.cpp
+sources/go-llama.cpp:
+	mkdir -p sources/go-llama.cpp
+	cd sources/go-llama.cpp && \
+	git init && \
+	git remote add origin $(GOLLAMA_REPO) && \
+	git fetch origin && \
+	git checkout $(GOLLAMA_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-llama.cpp/libbinding.a: sources/go-llama.cpp
+	$(MAKE) -C sources/go-llama.cpp BUILD_TYPE=$(STABLE_BUILD_TYPE) libbinding.a
+
 ## bark.cpp
 sources/bark.cpp:
 	git clone --recursive $(BARKCPP_REPO) sources/bark.cpp && \
@@ -294,7 +277,11 @@ backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp
 	$(MAKE) -C backend/go/image/stablediffusion-ggml libsd.a
 
 backend-assets/grpc/stablediffusion-ggml: backend/go/image/stablediffusion-ggml/libsd.a backend-assets/grpc
-	$(MAKE) -C backend/go/image/stablediffusion-ggml CGO_LDFLAGS="$(CGO_LDFLAGS)" stablediffusion-ggml
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ LIBRARY_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion-ggml ./backend/go/image/stablediffusion-ggml/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/stablediffusion-ggml
+endif
 
 sources/onnxruntime:
 	mkdir -p sources/onnxruntime
@@ -320,21 +307,22 @@ sources/whisper.cpp:
 	git checkout $(WHISPER_CPP_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-sources/whisper.cpp/build/src/libwhisper.a: sources/whisper.cpp
-	cd sources/whisper.cpp && cmake $(WHISPER_CMAKE_ARGS) . -B ./build
-	cd sources/whisper.cpp/build && cmake --build . --config Release
+sources/whisper.cpp/libwhisper.a: sources/whisper.cpp
+	cd sources/whisper.cpp && $(MAKE) libwhisper.a libggml.a
 
-get-sources: sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
+get-sources: sources/go-llama.cpp sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp backend/cpp/llama/llama.cpp
 
 replace:
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(CURDIR)/sources/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp/bindings/go=$(CURDIR)/sources/whisper.cpp/bindings/go
 	$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(CURDIR)/sources/go-piper
+	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(CURDIR)/sources/go-llama.cpp
 
 dropreplace:
 	$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp
 	$(GOCMD) mod edit -dropreplace github.com/ggerganov/whisper.cpp/bindings/go
 	$(GOCMD) mod edit -dropreplace github.com/mudler/go-piper
+	$(GOCMD) mod edit -dropreplace github.com/go-skynet/go-llama.cpp
 
 prepare-sources: get-sources replace
 	$(GOCMD) mod download
@@ -342,6 +330,7 @@ prepare-sources: get-sources replace
 ## GENERIC
 rebuild: ## Rebuilds the project
 	$(GOCMD) clean -cache
+	$(MAKE) -C sources/go-llama.cpp clean
 	$(MAKE) -C sources/whisper.cpp clean
 	$(MAKE) -C sources/go-piper clean
 	$(MAKE) build
@@ -372,14 +361,8 @@ clean-tests:
 clean-dc: clean
 	cp -r /build/backend-assets /workspace/backend-assets
 
-## Install Go tools
-install-go-tools:
-	go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
-	go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2
-	go install github.com/GeertJohan/go.rice/rice@latest
-
 ## Build:
-build: prepare backend-assets grpcs install-go-tools ## Build the project
+build: prepare backend-assets grpcs ## 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)${RESET})
@@ -389,9 +372,7 @@ ifneq ($(BACKEND_LIBS),)
 	$(MAKE) backend-assets/lib
 	cp -f $(BACKEND_LIBS) backend-assets/lib/
 endif
-	rm -rf $(BINARY_NAME) || true
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o $(BINARY_NAME) ./
-	rice append --exec $(BINARY_NAME)
 
 build-minimal:
 	BUILD_GRPC_FOR_BACKEND_LLAMA=true GRPC_BACKENDS="backend-assets/grpc/llama-cpp-avx2" GO_TAGS=p2p $(MAKE) build
@@ -453,7 +434,7 @@ run: prepare ## run local-ai
 test-models/testmodel.ggml:
 	mkdir test-models
 	mkdir test-dir
-	wget -q https://huggingface.co/RichardErkhov/Qwen_-_Qwen2-1.5B-Instruct-gguf/resolve/main/Qwen2-1.5B-Instruct.Q2_K.gguf -O test-models/testmodel.ggml
+	wget -q https://huggingface.co/TheBloke/orca_mini_3B-GGML/resolve/main/orca-mini-3b.ggmlv3.q4_0.bin -O test-models/testmodel.ggml
 	wget -q https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
 	wget -q https://huggingface.co/mudler/all-MiniLM-L6-v2/resolve/main/ggml-model-q4_0.bin -O test-models/bert
 	wget -q https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
@@ -463,13 +444,13 @@ prepare-test: grpcs
 	cp -rf backend-assets core/http
 	cp tests/models_fixtures/* test-models
 
-## Test targets
 test: prepare test-models/testmodel.ggml grpcs
 	@echo 'Running tests'
 	export GO_TAGS="tts debug"
 	$(MAKE) prepare-test
 	HUGGINGFACE_GRPC=$(abspath ./)/backend/python/transformers/run.sh TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
-	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama-gguf"  --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama && !llama-gguf"  --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
+	$(MAKE) test-llama
 	$(MAKE) test-llama-gguf
 	$(MAKE) test-tts
 	$(MAKE) test-stablediffusion
@@ -498,6 +479,10 @@ teardown-e2e:
 	rm -rf $(TEST_DIR) || true
 	docker stop $$(docker ps -q --filter ancestor=localai-tests)
 
+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 $(TEST_FLAKES) -v -r $(TEST_PATHS)
+
 test-llama-gguf: prepare-test
 	TEST_DIR=$(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 $(TEST_FLAKES) -v -r $(TEST_PATHS)
@@ -538,7 +523,7 @@ protogen: protogen-go protogen-python
 protogen-clean: protogen-go-clean protogen-python-clean
 
 .PHONY: protogen-go
-protogen-go: install-go-tools
+protogen-go:
 	mkdir -p pkg/grpc/proto
 	protoc --experimental_allow_proto3_optional -Ibackend/ --go_out=pkg/grpc/proto/ --go_opt=paths=source_relative --go-grpc_out=pkg/grpc/proto/ --go-grpc_opt=paths=source_relative \
     backend/backend.proto
@@ -549,10 +534,18 @@ protogen-go-clean:
 	$(RM) bin/*
 
 .PHONY: protogen-python
-protogen-python: bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
+protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen rerankers-protogen transformers-protogen kokoro-protogen vllm-protogen faster-whisper-protogen
 
 .PHONY: protogen-python-clean
-protogen-python-clean: bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean rerankers-protogen-clean transformers-protogen-clean kokoro-protogen-clean vllm-protogen-clean faster-whisper-protogen-clean
+protogen-python-clean: autogptq-protogen-clean bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean rerankers-protogen-clean transformers-protogen-clean kokoro-protogen-clean vllm-protogen-clean faster-whisper-protogen-clean
+
+.PHONY: autogptq-protogen
+autogptq-protogen:
+	$(MAKE) -C backend/python/autogptq protogen
+
+.PHONY: autogptq-protogen-clean
+autogptq-protogen-clean:
+	$(MAKE) -C backend/python/autogptq protogen-clean
 
 .PHONY: bark-protogen
 bark-protogen:
@@ -629,6 +622,7 @@ vllm-protogen-clean:
 ## GRPC
 # Note: it is duplicated in the Dockerfile
 prepare-extra-conda-environments: protogen-python
+	$(MAKE) -C backend/python/autogptq
 	$(MAKE) -C backend/python/bark
 	$(MAKE) -C backend/python/coqui
 	$(MAKE) -C backend/python/diffusers
@@ -642,12 +636,10 @@ prepare-extra-conda-environments: protogen-python
 prepare-test-extra: protogen-python
 	$(MAKE) -C backend/python/transformers
 	$(MAKE) -C backend/python/diffusers
-	$(MAKE) -C backend/python/vllm
 
 test-extra: prepare-test-extra
 	$(MAKE) -C backend/python/transformers test
 	$(MAKE) -C backend/python/diffusers test
-	$(MAKE) -C backend/python/vllm test
 
 backend-assets:
 	mkdir -p backend-assets
@@ -768,6 +760,13 @@ backend-assets/util/llama-cpp-rpc-server: backend-assets/grpc/llama-cpp-grpc
 	mkdir -p backend-assets/util/
 	cp -rf backend/cpp/llama-grpc/llama.cpp/build/bin/rpc-server backend-assets/util/llama-cpp-rpc-server
 
+backend-assets/grpc/llama-ggml: sources/go-llama.cpp sources/go-llama.cpp/libbinding.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-llama.cpp LIBRARY_PATH=$(CURDIR)/sources/go-llama.cpp \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-ggml ./backend/go/llm/llama-ggml/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/llama-ggml
+endif
+
 backend-assets/grpc/bark-cpp: backend/go/bark/libbark.a backend-assets/grpc
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/bark/ LIBRARY_PATH=$(CURDIR)/backend/go/bark/ \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bark-cpp ./backend/go/bark/
@@ -789,8 +788,8 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/silero-vad
 endif
 
-backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/build/src/libwhisper.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="${WHISPER_INCLUDE_PATH}" LIBRARY_PATH="${WHISPER_LIBRARY_PATH}" LD_LIBRARY_PATH="${WHISPER_LIBRARY_PATH}" \
+backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/libwhisper.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="$(CURDIR)/sources/whisper.cpp/include:$(CURDIR)/sources/whisper.cpp/ggml/include" LIBRARY_PATH=$(CURDIR)/sources/whisper.cpp \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./backend/go/transcribe/whisper
 ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/whisper
@@ -842,8 +841,7 @@ docker-aio-all:
 
 docker-image-intel:
 	docker build \
-		--progress plain \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.1.0-0-devel-ubuntu24.04 \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
@@ -851,7 +849,7 @@ docker-image-intel:
 
 docker-image-intel-xpu:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.1.0-0-devel-ubuntu22.04 \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2025.0.0-0-devel-ubuntu22.04 \
 		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
 		--build-arg GO_TAGS="none" \
 		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
@@ -863,7 +861,7 @@ swagger:
 
 .PHONY: gen-assets
 gen-assets:
-	$(GOCMD) run core/dependencies_manager/manager.go webui_static.yaml core/http/static/assets
+	$(GOCMD) run core/dependencies_manager/manager.go embedded/webui_static.yaml core/http/static/assets
 
 ## Documentation
 docs/layouts/_default:

README.md

@@ -1,6 +1,7 @@
 <h1 align="center">
   <br>
-  <img height="300" src="./core/http/static/logo.png"> <br>
+  <img height="300" src="https://github.com/go-skynet/LocalAI/assets/2420543/0966aa2a-166e-4f99-a3e5-6c915fc997dd"> <br>
+    LocalAI
 <br>
 </h1>
 
@@ -30,7 +31,7 @@
 
 <p align="center">
 <a href="https://twitter.com/LocalAI_API" target="blank">
-<img src="https://img.shields.io/badge/X-%23000000.svg?style=for-the-badge&logo=X&logoColor=white&label=LocalAI_API" alt="Follow LocalAI_API"/>
+<img src="https://img.shields.io/twitter/follow/LocalAI_API?label=Follow: LocalAI_API&style=social" alt="Follow LocalAI_API"/>
 </a>
 <a href="https://discord.gg/uJAeKSAGDy" target="blank">
 <img src="https://dcbadge.vercel.app/api/server/uJAeKSAGDy?style=flat-square&theme=default-inverted" alt="Join LocalAI Discord Community"/>
@@ -43,89 +44,32 @@
 
 > :bulb: Get help - [❓FAQ](https://localai.io/faq/) [💭Discussions](https://github.com/go-skynet/LocalAI/discussions) [:speech_balloon: Discord](https://discord.gg/uJAeKSAGDy) [:book: Documentation website](https://localai.io/)
 >
-> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) Try on 
-[![Telegram](https://img.shields.io/badge/Telegram-2CA5E0?style=for-the-badge&logo=telegram&logoColor=white)](https://t.me/localaiofficial_bot)
+> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) 
 
 [![tests](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml)[![Build and Release](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml)[![build container images](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml)[![Bump dependencies](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml)[![Artifact Hub](https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/localai)](https://artifacthub.io/packages/search?repo=localai)
 
-**LocalAI** is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that's compatible with OpenAI (Elevenlabs, Anthropic... ) API specifications for local AI inferencing. It allows you to run LLMs, generate images, audio (and not only) locally or on-prem with consumer grade hardware, supporting multiple model families. Does not require GPU. It is created and maintained by [Ettore Di Giacinto](https://github.com/mudler).
+**LocalAI** is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that’s compatible with OpenAI (Elevenlabs, Anthropic... ) API specifications for local AI inferencing. It allows you to run LLMs, generate images, audio (and not only) locally or on-prem with consumer grade hardware, supporting multiple model families. Does not require GPU. It is created and maintained by [Ettore Di Giacinto](https://github.com/mudler).
 
-
-## 📚🆕 Local Stack Family
-
-🆕 LocalAI is now part of a comprehensive suite of AI tools designed to work together:
-
-<table>
-  <tr>
-    <td width="50%" valign="top">
-      <a href="https://github.com/mudler/LocalAGI">
-        <img src="https://raw.githubusercontent.com/mudler/LocalAGI/refs/heads/main/webui/react-ui/public/logo_2.png" width="300" alt="LocalAGI Logo">
-      </a>
-    </td>
-    <td width="50%" valign="top">
-      <h3><a href="https://github.com/mudler/LocalAGI">LocalAGI</a></h3>
-      <p>A powerful Local AI agent management platform that serves as a drop-in replacement for OpenAI's Responses API, enhanced with advanced agentic capabilities.</p>
-    </td>
-  </tr>
-  <tr>
-    <td width="50%" valign="top">
-      <a href="https://github.com/mudler/LocalRecall">
-        <img src="https://raw.githubusercontent.com/mudler/LocalRecall/refs/heads/main/static/localrecall_horizontal.png" width="300" alt="LocalRecall Logo">
-      </a>
-    </td>
-    <td width="50%" valign="top">
-      <h3><a href="https://github.com/mudler/LocalRecall">LocalRecall</a></h3>
-      <p>A REST-ful API and knowledge base management system that provides persistent memory and storage capabilities for AI agents.</p>
-    </td>
-  </tr>
-</table>
-
-## Screenshots
-
-
-| Talk Interface | Generate Audio |
-| --- | --- |
-| ![Screenshot 2025-03-31 at 12-01-36 LocalAI - Talk](./docs/assets/images/screenshots/screenshot_tts.png) | ![Screenshot 2025-03-31 at 12-01-29 LocalAI - Generate audio with voice-en-us-ryan-low](./docs/assets/images/screenshots/screenshot_tts.png) |
-
-| Models Overview | Generate Images |
-| --- | --- |
-| ![Screenshot 2025-03-31 at 12-01-20 LocalAI - Models](./docs/assets/images/screenshots/screenshot_gallery.png) | ![Screenshot 2025-03-31 at 12-31-41 LocalAI - Generate images with flux 1-dev](./docs/assets/images/screenshots/screenshot_image.png) |
-
-| Chat Interface | Home |
-| --- | --- |
-| ![Screenshot 2025-03-31 at 11-57-44 LocalAI - Chat with localai-functioncall-qwen2 5-7b-v0 5](./docs/assets/images/screenshots/screenshot_chat.png) | ![Screenshot 2025-03-31 at 11-57-23 LocalAI API - c2a39e3 (c2a39e3639227cfd94ffffe9f5691239acc275a8)](./docs/assets/images/screenshots/screenshot_home.png) |
-
-| Login | Swarm |
-| --- | --- |
-|![Screenshot 2025-03-31 at 12-09-59 ](./docs/assets/images/screenshots/screenshot_login.png) | ![Screenshot 2025-03-31 at 12-10-39 LocalAI - P2P dashboard](./docs/assets/images/screenshots/screenshot_p2p.png) |
-
-## 💻 Quickstart
+![screen](https://github.com/mudler/LocalAI/assets/2420543/20b5ccd2-8393-44f0-aaf6-87a23806381e)
 
 Run the installer script:
 
 ```bash
-# Basic installation
 curl https://localai.io/install.sh | sh
 ```
 
-For more installation options, see [Installer Options](https://localai.io/docs/advanced/installer/).
-
 Or run with docker:
-
-### CPU only image:
 ```bash
+# CPU only image:
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-cpu
-```
-### Nvidia GPU:
-```bash
+
+# Nvidia GPU:
 docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-gpu-nvidia-cuda-12
-```
-### CPU and GPU image (bigger size):
-```bash
+
+# CPU and GPU image (bigger size):
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest
-```
-### AIO images (it will pre-download a set of models ready for use, see https://localai.io/basics/container/)
-```bash
+
+# AIO images (it will pre-download a set of models ready for use, see https://localai.io/basics/container/)
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-cpu
 ```
 
@@ -144,13 +88,10 @@ local-ai run https://gist.githubusercontent.com/.../phi-2.yaml
 local-ai run oci://localai/phi-2:latest
 ```
 
-For more information, see [💻 Getting started](https://localai.io/basics/getting_started/index.html)
+[💻 Getting started](https://localai.io/basics/getting_started/index.html)
 
 ## 📰 Latest project news
 
-- Apr 2025: [LocalAGI](https://github.com/mudler/LocalAGI) and [LocalRecall](https://github.com/mudler/LocalRecall) join the LocalAI family stack.
-- Apr 2025: WebUI overhaul, AIO images updates
-- Feb 2025: Backend cleanup, Breaking changes, new backends (kokoro, OutelTTS, faster-whisper), Nvidia L4T images
 - Jan 2025: LocalAI model release: https://huggingface.co/mudler/LocalAI-functioncall-phi-4-v0.3, SANA support in diffusers: https://github.com/mudler/LocalAI/pull/4603
 - Dec 2024: stablediffusion.cpp backend (ggml) added ( https://github.com/mudler/LocalAI/pull/4289 )
 - Nov 2024: Bark.cpp backend added ( https://github.com/mudler/LocalAI/pull/4287 )
@@ -164,6 +105,19 @@ For more information, see [💻 Getting started](https://localai.io/basics/getti
 
 Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap)
 
+## 🔥🔥 Hot topics (looking for help):
+
+- Multimodal with vLLM and Video understanding: https://github.com/mudler/LocalAI/pull/3729
+- Realtime API https://github.com/mudler/LocalAI/issues/3714
+- WebUI improvements: https://github.com/mudler/LocalAI/issues/2156
+- Backends v2: https://github.com/mudler/LocalAI/issues/1126
+- Improving UX v2: https://github.com/mudler/LocalAI/issues/1373
+- Assistant API: https://github.com/mudler/LocalAI/issues/1273
+- Vulkan: https://github.com/mudler/LocalAI/issues/1647
+- Anthropic API: https://github.com/mudler/LocalAI/issues/1808
+
+If you want to help and contribute, issues up for grabs: https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3A%22up+for+grabs%22
+
 ## 🚀 [Features](https://localai.io/features/)
 
 - 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `transformers`, `vllm` ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
@@ -177,10 +131,12 @@ Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3A
 - 🥽 [Vision API](https://localai.io/features/gpt-vision/)
 - 📈 [Reranker API](https://localai.io/features/reranker/)
 - 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
-- [Agentic capabilities](https://github.com/mudler/LocalAGI)
 - 🔊 Voice activity detection (Silero-VAD support)
 - 🌍 Integrated WebUI!
 
+## 💻 Usage
+
+Check out the [Getting started](https://localai.io/basics/getting_started/index.html) section in our documentation.
 
 ### 🔗 Community and integrations
 
@@ -256,7 +212,7 @@ A huge thank you to our generous sponsors who support this project covering CI e
 
 <p align="center">
   <a href="https://www.spectrocloud.com/" target="blank">
-    <img height="200" src="https://github.com/user-attachments/assets/72eab1dd-8b93-4fc0-9ade-84db49f24962">
+    <img height="200" src="https://github.com/go-skynet/LocalAI/assets/2420543/68a6f3cb-8a65-4a4d-99b5-6417a8905512">
   </a>
   <a href="https://www.premai.io/" target="blank">
     <img height="200" src="https://github.com/mudler/LocalAI/assets/2420543/42e4ca83-661e-4f79-8e46-ae43689683d6"> <br>

aio/cpu/embeddings.yaml

@@ -1,7 +1,7 @@
-embeddings: true
 name: text-embedding-ada-002
+embeddings: true
 parameters:
-  model: huggingface://bartowski/granite-embedding-107m-multilingual-GGUF/granite-embedding-107m-multilingual-f16.gguf
+  model: huggingface://hugging-quants/Llama-3.2-1B-Instruct-Q4_K_M-GGUF/llama-3.2-1b-instruct-q4_k_m.gguf
 
 usage: |
     You can test this model with curl like this:

aio/cpu/text-to-text.yaml

@@ -1,57 +1,101 @@
-context_size: 8192
-f16: true
-function:
-  grammar:
-    no_mixed_free_string: true
-    schema_type: llama3.1 # or JSON is supported too (json)
-  response_regex:
-  - <function=(?P<name>\w+)>(?P<arguments>.*)</function>
-mmap: true
 name: gpt-4
+mmap: true
 parameters:
-  model: Hermes-3-Llama-3.2-3B-Q4_K_M.gguf
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+context_size: 8192
+
 stopwords:
-- <|im_end|>
-- <dummy32000>
-- <|eot_id|>
-- <|end_of_text|>
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
 template:
   chat: |
-    <|begin_of_text|><|start_header_id|>system<|end_header_id|>
-    You are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>
-    {{.Input }}
-    <|start_header_id|>assistant<|end_header_id|>
+    {{.Input -}}
+    <|im_start|>assistant
   chat_message: |
-    <|start_header_id|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}<|end_header_id|>
-    {{ if .FunctionCall -}}
-    {{ else if eq .RoleName "tool" -}}
-    The Function was executed and the response was:
-    {{ end -}}
-    {{ if .Content -}}
-    {{.Content -}}
-    {{ else if .FunctionCall -}}
-    {{ range .FunctionCall }}
-    [{{.FunctionCall.Name}}({{.FunctionCall.Arguments}})]
-    {{ end }}
-    {{ end -}}
-    <|eot_id|>
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
+    {{.Content }}
+    {{- end }}
+    {{- if .FunctionCall}}
+    {{toJson .FunctionCall}}
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
   completion: |
     {{.Input}}
-  function: |
-    <|start_header_id|>system<|end_header_id|>
-    You are an expert in composing functions. You are given a question and a set of possible functions.
-    Based on the question, you will need to make one or more function/tool calls to achieve the purpose.
-    If none of the functions can be used, point it out. If the given question lacks the parameters required by the function, also point it out. You should only return the function call in tools call sections.
-    If you decide to invoke any of the function(s), you MUST put it in the format as follows:
-    [func_name1(params_name1=params_value1,params_name2=params_value2,...),func_name2(params_name1=params_value1,params_name2=params_value2,...)]
-    You SHOULD NOT include any other text in the response.
-    Here is a list of functions in JSON format that you can invoke.
-    {{toJson .Functions}}
-    <|eot_id|><|start_header_id|>user<|end_header_id|>
-    {{.Input}}
-    <|eot_id|><|start_header_id|>assistant<|end_header_id|>
-
-download_files:
-- filename: Hermes-3-Llama-3.2-3B-Q4_K_M.gguf
-  sha256: 2e220a14ba4328fee38cf36c2c068261560f999fadb5725ce5c6d977cb5126b5
-  uri: huggingface://bartowski/Hermes-3-Llama-3.2-3B-GGUF/Hermes-3-Llama-3.2-3B-Q4_K_M.gguf
+  function: |-
+    <|im_start|>system
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
+    {{range .Functions}}
+    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
+    {{end}}
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
+    {{.Input -}}
+    <|im_start|>assistant

aio/cpu/vad.yaml

@@ -1,8 +0,0 @@
-backend: silero-vad
-name: silero-vad
-parameters:
-  model: silero-vad.onnx
-download_files:
-- filename: silero-vad.onnx
-  uri: https://huggingface.co/onnx-community/silero-vad/resolve/main/onnx/model.onnx
-  sha256: a4a068cd6cf1ea8355b84327595838ca748ec29a25bc91fc82e6c299ccdc5808

aio/cpu/vision.yaml

@@ -1,49 +1,31 @@
+backend: llama-cpp
 context_size: 4096
 f16: true
 mmap: true
-mmproj: minicpm-v-2_6-mmproj-f16.gguf
 name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: bakllava-mmproj.gguf
 parameters:
-  model: minicpm-v-2_6-Q4_K_M.gguf
-stopwords:
-- <|im_end|>
-- <dummy32000>
-- </s>
-- <|endoftext|>
+  model: bakllava.gguf
+
 template:
   chat: |
-    {{.Input -}}
-    <|im_start|>assistant
-  chat_message: |
-    <|im_start|>{{ .RoleName }}
-    {{ if .FunctionCall -}}
-    Function call:
-    {{ else if eq .RoleName "tool" -}}
-    Function response:
-    {{ end -}}
-    {{ if .Content -}}
-    {{.Content }}
-    {{ end -}}
-    {{ if .FunctionCall -}}
-    {{toJson .FunctionCall}}
-    {{ end -}}<|im_end|>
-  completion: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
     {{.Input}}
-  function: |
-    <|im_start|>system
-    You are a function calling AI model. You are provided with functions to execute. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. Here are the available tools:
-    {{range .Functions}}
-    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
-    {{end}}
-    For each function call return a json object with function name and arguments
-    <|im_end|>
-    {{.Input -}}
-    <|im_start|>assistant
+    ASSISTANT:
 
 download_files:
-- filename: minicpm-v-2_6-Q4_K_M.gguf
-  sha256: 3a4078d53b46f22989adbf998ce5a3fd090b6541f112d7e936eb4204a04100b1
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/ggml-model-Q4_K_M.gguf
-- filename: minicpm-v-2_6-mmproj-f16.gguf
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/mmproj-model-f16.gguf
-  sha256: 4485f68a0f1aa404c391e788ea88ea653c100d8e98fe572698f701e5809711fd
+- filename: bakllava.gguf
+  uri: huggingface://mys/ggml_bakllava-1/ggml-model-q4_k.gguf
+- filename: bakllava-mmproj.gguf
+  uri: huggingface://mys/ggml_bakllava-1/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

aio/entrypoint.sh

@@ -129,7 +129,7 @@ detect_gpu
 detect_gpu_size
 
 PROFILE="${PROFILE:-$GPU_SIZE}" # default to cpu
-export MODELS="${MODELS:-/aio/${PROFILE}/embeddings.yaml,/aio/${PROFILE}/rerank.yaml,/aio/${PROFILE}/text-to-speech.yaml,/aio/${PROFILE}/image-gen.yaml,/aio/${PROFILE}/text-to-text.yaml,/aio/${PROFILE}/speech-to-text.yaml,/aio/${PROFILE}/vad.yaml,/aio/${PROFILE}/vision.yaml}"
+export MODELS="${MODELS:-/aio/${PROFILE}/embeddings.yaml,/aio/${PROFILE}/rerank.yaml,/aio/${PROFILE}/text-to-speech.yaml,/aio/${PROFILE}/image-gen.yaml,/aio/${PROFILE}/text-to-text.yaml,/aio/${PROFILE}/speech-to-text.yaml,/aio/${PROFILE}/vision.yaml}"
 
 check_vars
 

aio/gpu-8g/embeddings.yaml

@@ -1,7 +1,7 @@
-embeddings: true
 name: text-embedding-ada-002
+backend: sentencetransformers
 parameters:
-  model: huggingface://bartowski/granite-embedding-107m-multilingual-GGUF/granite-embedding-107m-multilingual-f16.gguf
+  model: all-MiniLM-L6-v2
 
 usage: |
     You can test this model with curl like this:

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

@@ -1,53 +1,101 @@
-context_size: 4096
-f16: true
-function:
-  capture_llm_results:
-  - (?s)<Thought>(.*?)</Thought>
-  grammar:
-    properties_order: name,arguments
-  json_regex_match:
-  - (?s)<Output>(.*?)</Output>
-  replace_llm_results:
-  - key: (?s)<Thought>(.*?)</Thought>
-    value: ""
-mmap: true
 name: gpt-4
+mmap: true
 parameters:
-  model: localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+context_size: 8192
+
 stopwords:
-- <|im_end|>
-- <dummy32000>
-- </s>
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
 template:
   chat: |
     {{.Input -}}
     <|im_start|>assistant
   chat_message: |
-    <|im_start|>{{ .RoleName }}
-    {{ if .FunctionCall -}}
-    Function call:
-    {{ else if eq .RoleName "tool" -}}
-    Function response:
-    {{ end -}}
-    {{ if .Content -}}
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
     {{.Content }}
-    {{ end -}}
-    {{ if .FunctionCall -}}
+    {{- end }}
+    {{- if .FunctionCall}}
     {{toJson .FunctionCall}}
-    {{ end -}}<|im_end|>
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
   completion: |
     {{.Input}}
-  function: |
+  function: |-
     <|im_start|>system
-    You are an AI assistant that executes function calls, and these are the tools at your disposal:
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
     {{range .Functions}}
     {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
     {{end}}
-    <|im_end|>
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
     {{.Input -}}
-    <|im_start|>assistant
-
-download_files:
-- filename: localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf
-  sha256: 4e7b7fe1d54b881f1ef90799219dc6cc285d29db24f559c8998d1addb35713d4
-  uri: huggingface://mudler/LocalAI-functioncall-qwen2.5-7b-v0.5-Q4_K_M-GGUF/localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf
+    <|im_start|>assistant

aio/gpu-8g/vad.yaml

@@ -1,8 +0,0 @@
-backend: silero-vad
-name: silero-vad
-parameters:
-  model: silero-vad.onnx
-download_files:
-- filename: silero-vad.onnx
-  uri: https://huggingface.co/onnx-community/silero-vad/resolve/main/onnx/model.onnx
-  sha256: a4a068cd6cf1ea8355b84327595838ca748ec29a25bc91fc82e6c299ccdc5808

aio/gpu-8g/vision.yaml

@@ -1,49 +1,35 @@
+backend: llama-cpp
 context_size: 4096
 f16: true
 mmap: true
-mmproj: minicpm-v-2_6-mmproj-f16.gguf
 name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: llava-v1.6-7b-mmproj-f16.gguf
 parameters:
-  model: minicpm-v-2_6-Q4_K_M.gguf
-stopwords:
-- <|im_end|>
-- <dummy32000>
-- </s>
-- <|endoftext|>
+  model: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  temperature: 0.2
+  top_k: 40
+  top_p: 0.95
+  seed: -1
+
 template:
   chat: |
-    {{.Input -}}
-    <|im_start|>assistant
-  chat_message: |
-    <|im_start|>{{ .RoleName }}
-    {{ if .FunctionCall -}}
-    Function call:
-    {{ else if eq .RoleName "tool" -}}
-    Function response:
-    {{ end -}}
-    {{ if .Content -}}
-    {{.Content }}
-    {{ end -}}
-    {{ if .FunctionCall -}}
-    {{toJson .FunctionCall}}
-    {{ end -}}<|im_end|>
-  completion: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
     {{.Input}}
-  function: |
-    <|im_start|>system
-    You are a function calling AI model. You are provided with functions to execute. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. Here are the available tools:
-    {{range .Functions}}
-    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
-    {{end}}
-    For each function call return a json object with function name and arguments
-    <|im_end|>
-    {{.Input -}}
-    <|im_start|>assistant
+    ASSISTANT:
 
 download_files:
-- filename: minicpm-v-2_6-Q4_K_M.gguf
-  sha256: 3a4078d53b46f22989adbf998ce5a3fd090b6541f112d7e936eb4204a04100b1
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/ggml-model-Q4_K_M.gguf
-- filename: minicpm-v-2_6-mmproj-f16.gguf
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/mmproj-model-f16.gguf
-  sha256: 4485f68a0f1aa404c391e788ea88ea653c100d8e98fe572698f701e5809711fd
+- filename: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/llava-v1.6-mistral-7b.Q5_K_M.gguf
+- filename: llava-v1.6-7b-mmproj-f16.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

aio/intel/embeddings.yaml

@@ -1,7 +1,7 @@
-embeddings: true
 name: text-embedding-ada-002
+backend: sentencetransformers
 parameters:
-  model: huggingface://bartowski/granite-embedding-107m-multilingual-GGUF/granite-embedding-107m-multilingual-f16.gguf
+  model: all-MiniLM-L6-v2
 
 usage: |
     You can test this model with curl like this:

aio/intel/text-to-text.yaml

@@ -1,53 +1,103 @@
-context_size: 4096
-f16: true
-function:
-  capture_llm_results:
-  - (?s)<Thought>(.*?)</Thought>
-  grammar:
-    properties_order: name,arguments
-  json_regex_match:
-  - (?s)<Output>(.*?)</Output>
-  replace_llm_results:
-  - key: (?s)<Thought>(.*?)</Thought>
-    value: ""
-mmap: true
 name: gpt-4
+mmap: false
+context_size: 8192
+
+f16: false
 parameters:
-  model: localai-functioncall-qwen2.5-7b-v0.5-q4_k_m.gguf
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+
 stopwords:
-- <|im_end|>
-- <dummy32000>
-- </s>
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
 template:
   chat: |
     {{.Input -}}
     <|im_start|>assistant
   chat_message: |
-    <|im_start|>{{ .RoleName }}
-    {{ if .FunctionCall -}}
-    Function call:
-    {{ else if eq .RoleName "tool" -}}
-    Function response:
-    {{ end -}}
-    {{ if .Content -}}
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
     {{.Content }}
-    {{ end -}}
-    {{ if .FunctionCall -}}
+    {{- end }}
+    {{- if .FunctionCall}}
     {{toJson .FunctionCall}}
-    {{ end -}}<|im_end|>
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
   completion: |
     {{.Input}}
-  function: |
+  function: |-
     <|im_start|>system
-    You are an AI assistant that executes function calls, and these are the tools at your disposal:
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
     {{range .Functions}}
     {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
     {{end}}
-    <|im_end|>
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
     {{.Input -}}
     <|im_start|>assistant
-
-download_files:
-- filename: localai-functioncall-phi-4-v0.3-q4_k_m.gguf
-  sha256: 23fee048ded2a6e2e1a7b6bbefa6cbf83068f194caa9552aecbaa00fec8a16d5
-  uri: huggingface://mudler/LocalAI-functioncall-phi-4-v0.3-Q4_K_M-GGUF/localai-functioncall-phi-4-v0.3-q4_k_m.gguf

aio/intel/vad.yaml

@@ -1,8 +0,0 @@
-backend: silero-vad
-name: silero-vad
-parameters:
-  model: silero-vad.onnx
-download_files:
-- filename: silero-vad.onnx
-  uri: https://huggingface.co/onnx-community/silero-vad/resolve/main/onnx/model.onnx
-  sha256: a4a068cd6cf1ea8355b84327595838ca748ec29a25bc91fc82e6c299ccdc5808

aio/intel/vision.yaml

@@ -1,50 +1,35 @@
+backend: llama-cpp
 context_size: 4096
-f16: true
-mmap: true
-mmproj: minicpm-v-2_6-mmproj-f16.gguf
+mmap: false
+f16: false
 name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: llava-v1.6-7b-mmproj-f16.gguf
 parameters:
-  model: minicpm-v-2_6-Q4_K_M.gguf
-stopwords:
-- <|im_end|>
-- <dummy32000>
-- </s>
-- <|endoftext|>
+  model: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  temperature: 0.2
+  top_k: 40
+  top_p: 0.95
+  seed: -1
+
 template:
   chat: |
-    {{.Input -}}
-    <|im_start|>assistant
-  chat_message: |
-    <|im_start|>{{ .RoleName }}
-    {{ if .FunctionCall -}}
-    Function call:
-    {{ else if eq .RoleName "tool" -}}
-    Function response:
-    {{ end -}}
-    {{ if .Content -}}
-    {{.Content }}
-    {{ end -}}
-    {{ if .FunctionCall -}}
-    {{toJson .FunctionCall}}
-    {{ end -}}<|im_end|>
-  completion: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
     {{.Input}}
-  function: |
-    <|im_start|>system
-    You are a function calling AI model. You are provided with functions to execute. You may call one or more functions to assist with the user query. Don't make assumptions about what values to plug into functions. Here are the available tools:
-    {{range .Functions}}
-    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
-    {{end}}
-    For each function call return a json object with function name and arguments
-    <|im_end|>
-    {{.Input -}}
-    <|im_start|>assistant
-
+    ASSISTANT:
 
 download_files:
-- filename: minicpm-v-2_6-Q4_K_M.gguf
-  sha256: 3a4078d53b46f22989adbf998ce5a3fd090b6541f112d7e936eb4204a04100b1
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/ggml-model-Q4_K_M.gguf
-- filename: minicpm-v-2_6-mmproj-f16.gguf
-  uri: huggingface://openbmb/MiniCPM-V-2_6-gguf/mmproj-model-f16.gguf
-  sha256: 4485f68a0f1aa404c391e788ea88ea653c100d8e98fe572698f701e5809711fd
+- filename: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/llava-v1.6-mistral-7b.Q5_K_M.gguf
+- filename: llava-v1.6-7b-mmproj-f16.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

assets.go

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

backend/backend.proto

@@ -14,7 +14,6 @@ service Backend {
   rpc PredictStream(PredictOptions) returns (stream Reply) {}
   rpc Embedding(PredictOptions) returns (EmbeddingResult) {}
   rpc GenerateImage(GenerateImageRequest) returns (Result) {}
-  rpc GenerateVideo(GenerateVideoRequest) returns (Result) {}
   rpc AudioTranscription(TranscriptRequest) returns (TranscriptResult) {}
   rpc TTS(TTSRequest) returns (Result) {}
   rpc SoundGeneration(SoundGenerationRequest) returns (Result) {}
@@ -164,10 +163,6 @@ message Reply {
   double timing_token_generation = 5;
 }
 
-message GrammarTrigger {
-  string word = 1;
-}
-
 message ModelOptions {
   string Model = 1;
   int32 ContextSize = 2;
@@ -191,7 +186,11 @@ message ModelOptions {
   int32 NGQA = 20;
   string ModelFile = 21;
 
-
+  // AutoGPTQ
+  string Device = 22;
+  bool UseTriton = 23;
+  string ModelBaseName = 24;
+  bool UseFastTokenizer = 25;
 
   // Diffusers
   string PipelineType = 26;
@@ -225,11 +224,6 @@ message ModelOptions {
   int32  MaxModelLen = 54;
   int32  TensorParallelSize = 55;
   string LoadFormat = 58;
-  bool   DisableLogStatus = 66;
-  string DType = 67;
-  int32  LimitImagePerPrompt = 68;
-  int32  LimitVideoPerPrompt = 69;
-  int32  LimitAudioPerPrompt = 70;
 
   string MMProj = 41;
 
@@ -253,8 +247,6 @@ message ModelOptions {
 
   string CacheTypeKey = 63;
   string CacheTypeValue = 64;
-
-  repeated GrammarTrigger GrammarTriggers = 65;
 }
 
 message Result {
@@ -302,19 +294,6 @@ message GenerateImageRequest {
   int32 CLIPSkip = 11;
 }
 
-message GenerateVideoRequest {
-  string prompt = 1;
-  string start_image = 2;  // Path or base64 encoded image for the start frame
-  string end_image = 3;    // Path or base64 encoded image for the end frame
-  int32 width = 4;
-  int32 height = 5;
-  int32 num_frames = 6;    // Number of frames to generate
-  int32 fps = 7;          // Frames per second
-  int32 seed = 8;
-  float cfg_scale = 9;    // Classifier-free guidance scale
-  string dst = 10;        // Output path for the generated video
-}
-
 message TTSRequest {
   string text = 1;
   string model = 2;

backend/cpp/llama/CMakeLists.txt

@@ -2,7 +2,7 @@
 ## XXX: In some versions of CMake clip wasn't being built before llama.
 ## This is an hack for now, but it should be fixed in the future.
 set(TARGET myclip)
-add_library(${TARGET} clip.cpp clip.h clip-impl.h llava.cpp llava.h)
+add_library(${TARGET} clip.cpp clip.h llava.cpp llava.h)
 install(TARGETS ${TARGET} LIBRARY)
 target_include_directories(myclip PUBLIC .)
 target_include_directories(myclip PUBLIC ../..)

backend/cpp/llama/Makefile

@@ -8,7 +8,7 @@ ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 TARGET?=--target grpc-server
 
 # Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
-CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF -DLLAMA_CURL=OFF
+CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
 
 # If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
@@ -36,18 +36,11 @@ else ifeq ($(OS),Darwin)
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f16)
-	CMAKE_ARGS+=-DGGML_SYCL=ON \
-		-DCMAKE_C_COMPILER=icx \
-		-DCMAKE_CXX_COMPILER=icpx \
-		-DCMAKE_CXX_FLAGS="-fsycl" \
-		-DGGML_SYCL_F16=ON
+	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
 endif
 
 ifeq ($(BUILD_TYPE),sycl_f32)
-	CMAKE_ARGS+=-DGGML_SYCL=ON \
-		-DCMAKE_C_COMPILER=icx \
-		-DCMAKE_CXX_COMPILER=icpx \
-		-DCMAKE_CXX_FLAGS="-fsycl"
+	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 endif
 
 llama.cpp:
@@ -59,8 +52,8 @@ llama.cpp:
 	git checkout -b build $(LLAMA_VERSION) && \
 	git submodule update --init --recursive --depth 1 --single-branch
 
-llama.cpp/tools/grpc-server: llama.cpp
-	mkdir -p llama.cpp/tools/grpc-server
+llama.cpp/examples/grpc-server: llama.cpp
+	mkdir -p llama.cpp/examples/grpc-server
 	bash prepare.sh
 
 rebuild:
@@ -70,13 +63,13 @@ rebuild:
 
 purge:
 	rm -rf llama.cpp/build
-	rm -rf llama.cpp/tools/grpc-server
+	rm -rf llama.cpp/examples/grpc-server
 	rm -rf grpc-server
 
 clean: purge
 	rm -rf llama.cpp
 
-grpc-server: llama.cpp llama.cpp/tools/grpc-server
+grpc-server: llama.cpp llama.cpp/examples/grpc-server
 	@echo "Building grpc-server with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	+bash -c "source $(ONEAPI_VARS); \
@@ -84,4 +77,4 @@ ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 else
 	+cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release $(TARGET)
 endif
-	cp llama.cpp/build/bin/grpc-server .
+	cp llama.cpp/build/bin/grpc-server .

backend/cpp/llama/grpc-server.cpp

@@ -22,6 +22,7 @@
 #include "backend.grpc.pb.h"
 #include "utils.hpp"
 #include "sampling.h"
+#include "speculative.h"
 // include std::regex
 #include <cstddef>
 #include <thread>
@@ -52,7 +53,7 @@ struct server_params
 {
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
-    std::string public_path = "tools/server/public";
+    std::string public_path = "examples/server/public";
     std::string chat_template = "";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -185,12 +186,45 @@ static json probs_vector_to_json(const llama_context *ctx, const std::vector<com
     return out;
 }
 
+struct llama_slot_params {
+    uint32_t seed      = -1; // RNG seed
+    bool stream        = true;
+    bool cache_prompt  = true; // remember the prompt to avoid reprocessing all prompt
+    bool return_tokens = false;
+
+    int32_t n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
+    int32_t n_predict = -1; // new tokens to predict
+    int32_t n_indent  =  0; // mininum line indentation for the generated text in number of whitespace characters
+
+    int64_t t_max_prompt_ms  = -1; // TODO: implement
+    int64_t t_max_predict_ms = -1; // if positive, limit the generation phase to this time limit
+
+    std::vector<common_adapter_lora_info> lora;
+
+    std::vector<std::string> antiprompt;
+    std::vector<std::string> response_fields;
+    bool timings_per_token = false;
+    bool post_sampling_probs = false;
+    bool ignore_eos = false;
+
+    json input_prefix;
+    json input_suffix;
+
+    struct common_params_sampling sampling;
+    struct common_params_speculative speculative;
+};
+
+
 struct llama_client_slot
 {
     int id;
     int task_id = -1;
 
-    struct slot_params params;
+    struct llama_slot_params params;
+    common_speculative * spec = nullptr;
+    llama_batch batch_spec = {};
+
 
     slot_state state = IDLE;
     slot_command command = NONE;
@@ -283,6 +317,7 @@ struct llama_client_slot
         images.clear();
     }
 
+
     bool has_budget(common_params &global_params) {
         if (params.n_predict == -1 && global_params.n_predict == -1)
         {
@@ -454,6 +489,10 @@ struct llama_server_context
 {
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
+    common_init_result llama_init_dft;
+    llama_context * ctx_dft = nullptr;
+    llama_model * model_dft = nullptr;
+    llama_context_params cparams_dft;
     const llama_vocab * vocab = nullptr;
 
     clip_ctx *clp_ctx = nullptr;
@@ -467,10 +506,6 @@ struct llama_server_context
     bool all_slots_are_idle = false;
     bool add_bos_token      = true;
     bool has_eos_token      = true;
-    bool has_gpu = false;
-
-    bool grammar_lazy = false;
-    std::vector<common_grammar_trigger> grammar_triggers;
 
     int32_t n_ctx;  // total context for all clients / slots
 
@@ -506,18 +541,16 @@ struct llama_server_context
         }
     }
 
+
     bool load_model(const common_params &params_)
     {
         params = params_;
-        if (!params.mmproj.path.empty()) {
+        if (!params.mmproj.empty()) {
             multimodal = true;
             LOG_INFO("Multi Modal Mode Enabled", {});
-            clp_ctx = clip_init(params.mmproj.path.c_str(), clip_context_params {
-                /* use_gpu */ has_gpu,
-                /*verbosity=*/ GGML_LOG_LEVEL_INFO,
-            });
+            clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
             if(clp_ctx == nullptr) {
-                LOG_ERR("unable to load clip model: %s", params.mmproj.path.c_str());
+                LOG_ERR("unable to load clip model: %s", params.mmproj.c_str());
                 return false;
             }
 
@@ -531,7 +564,7 @@ struct llama_server_context
         ctx = common_init.context.release();
         if (model == nullptr)
         {
-            LOG_ERR("unable to load model: %s", params.model.path.c_str());
+            LOG_ERR("unable to load model: %s", params.model.c_str());
             return false;
         }
 
@@ -552,6 +585,45 @@ struct llama_server_context
         add_bos_token = llama_vocab_get_add_bos(vocab);
         has_eos_token = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
 
+        if (!params.speculative.model.empty()) {
+            LOG("loading draft model '%s'\n", params.speculative.model.c_str());
+
+            auto params_dft = params;
+
+            params_dft.devices      = params.speculative.devices;
+            params_dft.model        = params.speculative.model;
+            params_dft.n_ctx        = params.speculative.n_ctx == 0 ? params.n_ctx / params.n_parallel : params.speculative.n_ctx;
+            params_dft.n_gpu_layers = params.speculative.n_gpu_layers;
+            params_dft.n_parallel   = 1;
+
+            llama_init_dft = common_init_from_params(params_dft);
+
+            model_dft = llama_init_dft.model.get();
+
+            if (model_dft == nullptr) {
+                LOG("failed to load draft model, '%s'\n", params.speculative.model.c_str());
+                return false;
+            }
+
+            if (!common_speculative_are_compatible(ctx, llama_init_dft.context.get())) {
+                LOG("the draft model '%s' is not compatible with the target model '%s'\n", params.speculative.model.c_str(), params.model.c_str());
+
+                return false;
+            }
+
+            const int n_ctx_dft = llama_n_ctx(llama_init_dft.context.get());
+
+            cparams_dft = common_context_params_to_llama(params_dft);
+            cparams_dft.n_batch = n_ctx_dft;
+
+            // force F16 KV cache for the draft model for extra performance
+            cparams_dft.type_k = GGML_TYPE_F16;
+            cparams_dft.type_v = GGML_TYPE_F16;
+
+            // the context is not needed - we will create one for each slot
+            llama_init_dft.context.reset();
+        }
+
         return true;
     }
 
@@ -580,6 +652,22 @@ struct llama_server_context
             slot.n_ctx = n_ctx_slot;
             slot.n_predict = params.n_predict;
 
+            if (model_dft) {
+                slot.batch_spec = llama_batch_init(params.speculative.n_max + 1, 0, 1);
+
+                ctx_dft = llama_init_from_model(model_dft, cparams_dft);
+                if (ctx_dft == nullptr) {
+                    LOG("%s", "failed to create draft context\n");
+                    return;
+                }
+
+                slot.spec = common_speculative_init(ctx_dft);
+                if (slot.spec == nullptr) {
+                    LOG("%s", "failed to create speculator\n");
+                    return;
+                }
+            }
+
             LOG_INFO("new slot", {
                 {"slot_id",    slot.id},
                 {"n_ctx_slot", slot.n_ctx}
@@ -688,9 +776,11 @@ struct llama_server_context
     }
 
     bool launch_slot_with_data(llama_client_slot* &slot, json data) {
-        slot_params default_params;
+        llama_slot_params default_params;
         common_params_sampling default_sparams;
- 
+        
+        default_sparams.speculative = params_base.speculative;
+
         slot->params.stream             = json_value(data, "stream",            false);
         slot->params.cache_prompt       = json_value(data, "cache_prompt",      false);
         slot->params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
@@ -713,8 +803,15 @@ struct llama_server_context
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
         slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
         slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
-        slot->sparams.grammar_triggers = grammar_triggers;
-        slot->sparams.grammar_lazy = grammar_lazy;
+
+
+        slot->sparams.speculative.n_min = json_value(data, "speculative.n_min", defaults.speculative.n_min);
+        slot->sparams.speculative.n_max = json_value(data, "speculative.n_max", defaults.speculative.n_max);
+        slot->sparams.speculative.p_min = json_value(data, "speculative.p_min", defaults.speculative.p_min);
+
+        slot->sparams.speculative.n_min = std::min(params.speculative.n_max, params.speculative.n_min);
+        slot->sparams.speculative.n_min = std::max(params.speculative.n_min, 2);
+        slot->sparams.speculative.n_max = std::max(params.speculative.n_max, 0);
 
         if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
             // Might be better to reject the request with a 400 ?
@@ -1159,14 +1256,6 @@ struct llama_server_context
             slot.has_next_token = false;
         }
 
-        if (slot.n_past >= slot.n_ctx) {
-            slot.truncated      = true;
-            slot.stopped_limit = true;
-            slot.has_next_token = false;
-
-            LOG_VERBOSE("stopped due to running out of context capacity", {});
-        }
-
         if (result.tok == llama_vocab_eos(vocab) || llama_vocab_is_eog(vocab, result.tok))
         {
             slot.stopped_eos = true;
@@ -1354,7 +1443,7 @@ struct llama_server_context
         queue_results.send(res);
     }
 
-    void send_embedding(llama_client_slot &slot, const llama_batch & batch)
+    void send_embedding(llama_client_slot &slot)
     {
         task_result res;
         res.id = slot.task_id;
@@ -1376,38 +1465,10 @@ struct llama_server_context
         else
         {
             const float *data = llama_get_embeddings(ctx);
-            std::vector<float> embd_res(n_embd, 0.0f);
-            std::vector<std::vector<float>> embedding;
-            for (int i = 0; i < batch.n_tokens; ++i) {
-                if (!batch.logits[i] || batch.seq_id[i][0] != slot.id) {
-                    continue;
-                }
-
-                const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-                if (embd == NULL) {
-                    embd = llama_get_embeddings_ith(ctx, i);
-                }
-
-                if (embd == NULL) {
-                    LOG("failed to get embeddings");
-
-                    continue;
-                }
-
-                // normalize only when there is pooling
-                // TODO: configurable
-                if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_NONE) {
-                    common_embd_normalize(embd, embd_res.data(), n_embd, 2);
-                    embedding.push_back(embd_res);
-                } else {
-                    embedding.push_back({ embd, embd + n_embd });
-                }
-            }
-
-            // OAI compat
+            std::vector<float> embedding(data, data + n_embd);
             res.result_json = json
             {
-                {"embedding", embedding[0] },
+                {"embedding", embedding },
             };
         }
         queue_results.send(res);
@@ -1667,17 +1728,17 @@ struct llama_server_context
             {
                 if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
                 {
-                    // this check is redundant (for good)
-                    // we should never get here, because generation should already stopped in process_token()
-
                     // START LOCALAI changes
                     // Temporary disable context-shifting as it can lead to infinite loops (issue: https://github.com/ggerganov/llama.cpp/issues/3969)
                     // See: https://github.com/mudler/LocalAI/issues/1333
                     // Context is exhausted, release the slot
                     slot.release();
                     send_final_response(slot);
-                    slot.has_next_token = false;
-                    LOG_ERROR("context is exhausted, release the slot", {});
+                    slot.cache_tokens.clear();
+                    slot.n_past = 0;
+                    slot.truncated = false;
+                    slot.has_next_token = true;
+                    LOG("Context exhausted. Slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
 
                     continue;
                     // END LOCALAI changes
@@ -2028,7 +2089,7 @@ struct llama_server_context
                 // prompt evaluated for embedding
                 if (slot.embedding)
                 {
-                    send_embedding(slot, batch_view);
+                    send_embedding(slot);
                     slot.release();
                     slot.i_batch = -1;
                     continue;
@@ -2069,6 +2130,97 @@ struct llama_server_context
             }
         }
 
+        // do speculative decoding
+        for (auto & slot : slots) {
+            if (!slot.is_processing() || !(ctx_dft && params.speculative.n_max > 0)) {
+                continue;
+            }
+
+            if (slot.state != PROCESSING) {
+                continue;
+            }
+
+            // determine the max draft that fits the current slot state
+            int n_draft_max = slot.params.speculative.n_max;
+
+            // note: n_past is not yet increased for the `id` token sampled above
+            //       also, need to leave space for 1 extra token to allow context shifts
+            n_draft_max = std::min(n_draft_max, slot.n_ctx - slot.n_past - 2);
+
+            if (slot.n_remaining > 0) {
+                n_draft_max = std::min(n_draft_max, slot.n_remaining - 1);
+            }
+
+            LOG("max possible draft: %d\n", n_draft_max);
+
+            if (n_draft_max < slot.params.speculative.n_min) {
+                LOG("the max possible draft is too small: %d < %d - skipping speculative decoding\n", n_draft_max, slot.params.speculative.n_min);
+
+                continue;
+            }
+
+            llama_token id = slot.sampled;
+
+            struct common_speculative_params params_spec;
+            params_spec.n_draft   = n_draft_max;
+            params_spec.n_reuse   = llama_n_ctx(ctx_dft) - slot.params.speculative.n_max;
+            params_spec.p_min     = slot.params.speculative.p_min;
+
+            llama_tokens draft = common_speculative_gen_draft(slot.spec, params_spec, slot.cache_tokens, id);
+
+            // ignore small drafts
+            if (slot.params.speculative.n_min > (int) draft.size()) {
+                LOG("ignoring small draft: %d < %d\n", (int) draft.size(), slot.params.speculative.n_min);
+
+                continue;
+            }
+
+            // construct the speculation batch
+            common_batch_clear(slot.batch_spec);
+            common_batch_add  (slot.batch_spec, id, slot.n_past, { slot.id }, true);
+
+            for (size_t i = 0; i < draft.size(); ++i) {
+                common_batch_add(slot.batch_spec, draft[i], slot.n_past + 1 + i, { slot.id }, true);
+            }
+
+            LOG("decoding speculative batch, size = %d\n", slot.batch_spec.n_tokens);
+
+            llama_decode(ctx, slot.batch_spec);
+
+            // the accepted tokens from the speculation
+            const auto ids = common_sampler_sample_and_accept_n(slot.ctx_sampling, ctx, draft);
+
+            slot.n_past    += ids.size();
+            slot.n_decoded += ids.size();
+
+            slot.cache_tokens.push_back(id);
+            slot.cache_tokens.insert(slot.cache_tokens.end(), ids.begin(), ids.end() - 1);
+
+            llama_kv_cache_seq_rm(ctx, slot.id, slot.n_past, -1);
+
+            for (size_t i = 0; i < ids.size(); ++i) {
+                completion_token_output result;
+
+                result.tok          = ids[i];
+                result.text_to_send = common_token_to_piece(ctx, result.tok, params.special);
+                //result.prob         = 1.0f; // set later
+
+                // TODO: set result.probs
+
+                if (!process_token(result, slot)) {
+                    // release slot because of stop condition
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    metrics.on_prediction(slot);
+                    break;
+                }
+            }
+
+        LOG("accepted %d/%d draft tokens, new n_past = %d\n", (int) ids.size() - 1, (int) draft.size(), slot.n_past);
+        }
+    
+
         LOG_VERBOSE("slots updated", {});
         return true;
     }
@@ -2122,11 +2274,7 @@ static void append_to_generated_text_from_generated_token_probs(llama_server_con
 }
 
 std::function<void(int)> shutdown_handler;
-
-inline void signal_handler(int signal) {
-    exit(1);
-}
-
+inline void signal_handler(int signal) { shutdown_handler(signal); }
 
 /////////////////////////////////
 ////////////////////////////////
@@ -2322,15 +2470,15 @@ static std::string get_all_kv_cache_types() {
 }
 
 static void params_parse(const backend::ModelOptions* request,
-                                common_params & params, llama_server_context &llama) {
+                                common_params & params) {
    
     // this is comparable to: https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L1809
 
-    params.model.path = request->modelfile();
+    params.model = request->modelfile();
     if (!request->mmproj().empty()) {
     // get the directory of modelfile
-      std::string model_dir = params.model.path.substr(0, params.model.path.find_last_of("/\\"));
-      params.mmproj.path = model_dir + "/"+ request->mmproj();
+      std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
+      params.mmproj = model_dir + "/"+ request->mmproj();
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
@@ -2345,6 +2493,30 @@ static void params_parse(const backend::ModelOptions* request,
     params.cpuparams.n_threads = request->threads();
     params.n_gpu_layers = request->ngpulayers();
     params.n_batch = request->nbatch();
+    params.speculative.model = request->draftmodel();
+
+    // If options is not NULL, parse options
+    for (int i = 0; request->options()[i] != NULL; i++) {
+        char *optname = strtok(request->options()[i], ":");
+        char *optval = strtok(NULL, ":");
+        if (optval == NULL) {
+            optval = "true";
+        }
+
+        if (!strcmp(optname, "speculative.n_gpu_layers")) {
+            params.speculative.n_gpu_layers = std::stoi(optval);
+        }
+        if (!strcmp(optname, "speculative.n_ctx")) {
+            params.speculative.n_ctx = std::stoi(optval);
+        }
+    }
+
+    if params.speculative.n_gpu_layers == 0 {
+        params.speculative.n_gpu_layers = params.n_gpu_layers;
+    }
+    if params.speculative.n_ctx == 0 {
+        params.speculative.n_ctx = params.n_ctx;
+    }
     // Set params.n_parallel by environment variable (LLAMA_PARALLEL), defaults to 1
     //params.n_parallel = 1;
     const char *env_parallel = std::getenv("LLAMACPP_PARALLEL");
@@ -2360,20 +2532,6 @@ static void params_parse(const backend::ModelOptions* request,
         add_rpc_devices(std::string(llama_grpc_servers));
     }
     
-     // decode options. Options are in form optname:optvale, or if booleans only optname.
-    for (int i = 0; i < request->options_size(); i++) {
-        std::string opt = request->options(i);
-        char *optname = strtok(&opt[0], ":");
-        char *optval = strtok(NULL, ":");
-        if (optval == NULL) {
-            optval = "true";
-        }
-
-        if (!strcmp(optname, "gpu")) {
-            llama.has_gpu = true;
-        }
-    }
-
     // TODO: Add yarn
 
     if (!request->tensorsplit().empty()) {
@@ -2405,7 +2563,7 @@ static void params_parse(const backend::ModelOptions* request,
         scale_factor = request->lorascale();
      }
      // get the directory of modelfile
-     std::string model_dir = params.model.path.substr(0, params.model.path.find_last_of("/\\"));
+     std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
      params.lora_adapters.push_back({ model_dir + "/"+request->loraadapter(), scale_factor });
     }
     params.use_mlock = request->mlock();
@@ -2437,21 +2595,6 @@ static void params_parse(const backend::ModelOptions* request,
     if ( request->ropefreqscale() != 0.0f ) {
         params.rope_freq_scale = request->ropefreqscale();
     }
-
-    if (request->grammartriggers_size() > 0) {
-        LOG_INFO("configuring grammar triggers", {});
-        llama.grammar_lazy = true;
-        for (int i = 0; i < request->grammartriggers_size(); i++) {
-            common_grammar_trigger trigger;
-	    trigger.type = COMMON_GRAMMAR_TRIGGER_TYPE_WORD;
-            trigger.value = request->grammartriggers(i).word();
-	    // trigger.at_start = request->grammartriggers(i).at_start();
-            llama.grammar_triggers.push_back(trigger);
-            LOG_INFO("grammar trigger", {
-                { "word", trigger.value },
-            });
-        }
-    }
 }
 
 
@@ -2467,7 +2610,7 @@ public:
   grpc::Status LoadModel(ServerContext* context, const backend::ModelOptions* request, backend::Result* result) {
     // Implement LoadModel RPC
     common_params params;
-    params_parse(request, params, llama);
+    params_parse(request, params);
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -2600,18 +2743,6 @@ public:
         return grpc::Status::OK;
     }
 
-    grpc::Status TokenizeString(ServerContext* context, const backend::PredictOptions* request, backend::TokenizationResponse* response){
-         json data = parse_options(false, request, llama);
-
-         std::vector<llama_token> tokens = llama.tokenize(data["prompt"],false);
-
-         for (int i=0 ; i< tokens.size(); i++){
-            response->add_tokens(tokens[i]);
-         }
-
-        return grpc::Status::OK;
-    }
-
     grpc::Status GetMetrics(ServerContext* context, const backend::MetricsRequest* request, backend::MetricsResponse* response) {
         llama_client_slot* active_slot = llama.get_active_slot();
 
@@ -2644,9 +2775,7 @@ void RunServer(const std::string& server_address) {
   ServerBuilder builder;
   builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
   builder.RegisterService(&service);
-  builder.SetMaxMessageSize(50 * 1024 * 1024); // 50MB
-  builder.SetMaxSendMessageSize(50 * 1024 * 1024); // 50MB
-  builder.SetMaxReceiveMessageSize(50 * 1024 * 1024); // 50MB
+
   std::unique_ptr<Server> server(builder.BuildAndStart());
   std::cout << "Server listening on " << server_address << std::endl;
   server->Wait();
@@ -2655,20 +2784,6 @@ void RunServer(const std::string& server_address) {
 int main(int argc, char** argv) {
   std::string server_address("localhost:50051");
 
-#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
-    struct sigaction sigint_action;
-    sigint_action.sa_handler = signal_handler;
-    sigemptyset (&sigint_action.sa_mask);
-    sigint_action.sa_flags = 0;
-    sigaction(SIGINT, &sigint_action, NULL);
-    sigaction(SIGTERM, &sigint_action, NULL);
-#elif defined (_WIN32)
-    auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
-        return (ctrl_type == CTRL_C_EVENT) ? (signal_handler(SIGINT), true) : false;
-    };
-    SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
-#endif
-
   // Define long and short options
   struct option long_options[] = {
       {"addr", required_argument, nullptr, 'a'},

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

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

backend/cpp/llama/prepare.sh

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

backend/cpp/llama/utils.hpp

@@ -1,4 +1,4 @@
-// https://github.com/ggerganov/llama.cpp/blob/master/tools/server/utils.hpp
+// https://github.com/ggerganov/llama.cpp/blob/master/examples/server/utils.hpp
 
 #pragma once
 
@@ -11,7 +11,7 @@
 
 #include "json.hpp"
 
-#include "../mtmd/clip.h"
+#include "../llava/clip.h"
 
 using json = nlohmann::json;
 

backend/go/image/stablediffusion-ggml/Makefile

@@ -8,19 +8,12 @@ ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 # keep standard at C11 and C++11
 CXXFLAGS = -I. -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/thirdparty -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/ggml/include -I$(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp -O3 -DNDEBUG -std=c++17 -fPIC
 
-GOCMD?=go
-CGO_LDFLAGS?=
-# Avoid parent make file overwriting CGO_LDFLAGS which is needed for hipblas
-CGO_LDFLAGS_SYCL=
-GO_TAGS?=
-LD_FLAGS?=
-
 # Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
 CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
 
 # If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
 ifeq ($(BUILD_TYPE),cublas)
-	CMAKE_ARGS+=-DSD_CUDA=ON
+	CMAKE_ARGS+=-DGGML_CUDA=ON
 # If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
 # to CMAKE_ARGS automatically
 else ifeq ($(BUILD_TYPE),openblas)
@@ -28,50 +21,31 @@ else ifeq ($(BUILD_TYPE),openblas)
 # If build type is clblas (openCL) we set -DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
 else ifeq ($(BUILD_TYPE),clblas)
 	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
-# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++
+# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
 else ifeq ($(BUILD_TYPE),hipblas)
-	CMAKE_ARGS+=-DSD_HIPBLAS=ON
+	CMAKE_ARGS+=-DGGML_HIP=ON
 # If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
 # But if it's OSX without metal, disable it here
 else ifeq ($(OS),Darwin)
 	ifneq ($(BUILD_TYPE),metal)
-		CMAKE_ARGS+=-DSD_METAL=OFF
+		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
-		CMAKE_ARGS+=-DSD_METAL=ON
+		CMAKE_ARGS+=-DGGML_METAL=ON
 		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
 		TARGET+=--target ggml-metal
 	endif
 endif
 
-ifeq ($(BUILD_TYPE),sycl_f16)
-	CMAKE_ARGS+=-DGGML_SYCL=ON \
-		-DCMAKE_C_COMPILER=icx \
-		-DCMAKE_CXX_COMPILER=icpx \
-		-DSD_SYCL=ON \
-		-DGGML_SYCL_F16=ON
-	CC=icx
-	CXX=icpx
-	CGO_LDFLAGS_SYCL += -fsycl -L${DNNLROOT}/lib -ldnnl ${MKLROOT}/lib/intel64/libmkl_sycl.a -fiopenmp -fopenmp-targets=spir64 -lOpenCL
-	CGO_LDFLAGS_SYCL += $(shell pkg-config --libs mkl-static-lp64-gomp)
-	CGO_CXXFLAGS += -fiopenmp -fopenmp-targets=spir64
-	CGO_CXXFLAGS += $(shell pkg-config --cflags mkl-static-lp64-gomp )
-endif
+# ifeq ($(BUILD_TYPE),sycl_f16)
+# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON -DSD_SYCL=ON -DGGML_SYCL_F16=ON
+# endif
 
-ifeq ($(BUILD_TYPE),sycl_f32)
-	CMAKE_ARGS+=-DGGML_SYCL=ON \
-		-DCMAKE_C_COMPILER=icx \
-		-DCMAKE_CXX_COMPILER=icpx \
-		-DSD_SYCL=ON
-	CC=icx
-	CXX=icpx
-	CGO_LDFLAGS_SYCL += -fsycl -L${DNNLROOT}/lib -ldnnl ${MKLROOT}/lib/intel64/libmkl_sycl.a -fiopenmp -fopenmp-targets=spir64 -lOpenCL
-	CGO_LDFLAGS_SYCL += $(shell pkg-config --libs mkl-static-lp64-gomp)
-	CGO_CXXFLAGS += -fiopenmp -fopenmp-targets=spir64
-	CGO_CXXFLAGS += $(shell pkg-config --cflags mkl-static-lp64-gomp )
-endif
+# ifeq ($(BUILD_TYPE),sycl_f32)
+# 	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON
+# endif
 
 # warnings
-# CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
 
 # Find all .a archives in ARCHIVE_DIR
 # (ggml can have different backends cpu, cuda, etc., each backend generates a .a archive)
@@ -112,24 +86,11 @@ endif
 	$(MAKE) $(COMBINED_LIB)
 
 gosd.o:
-ifneq (,$(findstring sycl,$(BUILD_TYPE)))
-	+bash -c "source $(ONEAPI_VARS); \
-	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c"
-else
 	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c
-endif
 
 libsd.a: gosd.o
 	cp $(INCLUDE_PATH)/build/libstable-diffusion.a ./libsd.a
 	$(AR) rcs libsd.a gosd.o
 
-stablediffusion-ggml:
-	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_SYCL)" C_INCLUDE_PATH="$(INCLUDE_PATH)" LIBRARY_PATH="$(LIBRARY_PATH)" \
-	CC="$(CC)" CXX="$(CXX)" CGO_CXXFLAGS="$(CGO_CXXFLAGS)" \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o ../../../../backend-assets/grpc/stablediffusion-ggml ./
-ifneq ($(UPX),)
-	$(UPX) ../../../../backend-assets/grpc/stablediffusion-ggml
-endif
-
 clean:
-	rm -rf gosd.o libsd.a build $(COMBINED_LIB)
+	rm -rf gosd.o libsd.a build $(COMBINED_LIB)

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

@@ -35,8 +35,6 @@ const char* sample_method_str[] = {
     "ipndm",
     "ipndm_v",
     "lcm",
-    "ddim_trailing",
-    "tcd",
 };
 
 // Names of the sigma schedule overrides, same order as sample_schedule in stable-diffusion.h
@@ -175,7 +173,6 @@ int gen_image(char *text, char *negativeText, int width, int height, int steps,
                             -1, //clip_skip
                             cfg_scale, // sfg_scale
                             3.5f,
-			    0, // eta
                             width,
                             height,
                             sample_method, 

backend/go/llm/llama-ggml/llama.go

@@ -0,0 +1,204 @@
+package main
+
+// This is a wrapper to statisfy the GRPC service interface
+// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
+import (
+	"fmt"
+
+	"github.com/go-skynet/go-llama.cpp"
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+type LLM struct {
+	base.SingleThread
+
+	llama *llama.LLama
+}
+
+func (llm *LLM) Load(opts *pb.ModelOptions) error {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+
+	llamaOpts := []llama.ModelOption{
+		llama.WithRopeFreqBase(ropeFreqBase),
+		llama.WithRopeFreqScale(ropeFreqScale),
+	}
+
+	if opts.NGQA != 0 {
+		llamaOpts = append(llamaOpts, llama.WithGQA(int(opts.NGQA)))
+	}
+
+	if opts.RMSNormEps != 0 {
+		llamaOpts = append(llamaOpts, llama.WithRMSNormEPS(opts.RMSNormEps))
+	}
+
+	if opts.ContextSize != 0 {
+		llamaOpts = append(llamaOpts, llama.SetContext(int(opts.ContextSize)))
+	}
+	if opts.F16Memory {
+		llamaOpts = append(llamaOpts, llama.EnableF16Memory)
+	}
+	if opts.Embeddings {
+		llamaOpts = append(llamaOpts, llama.EnableEmbeddings)
+	}
+	if opts.NGPULayers != 0 {
+		llamaOpts = append(llamaOpts, llama.SetGPULayers(int(opts.NGPULayers)))
+	}
+
+	llamaOpts = append(llamaOpts, llama.SetMMap(opts.MMap))
+	llamaOpts = append(llamaOpts, llama.SetMainGPU(opts.MainGPU))
+	llamaOpts = append(llamaOpts, llama.SetTensorSplit(opts.TensorSplit))
+	if opts.NBatch != 0 {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(int(opts.NBatch)))
+	} else {
+		llamaOpts = append(llamaOpts, llama.SetNBatch(512))
+	}
+
+	if opts.NUMA {
+		llamaOpts = append(llamaOpts, llama.EnableNUMA)
+	}
+
+	if opts.LowVRAM {
+		llamaOpts = append(llamaOpts, llama.EnabelLowVRAM)
+	}
+
+	model, err := llama.New(opts.ModelFile, llamaOpts...)
+	llm.llama = model
+
+	return err
+}
+
+func buildPredictOptions(opts *pb.PredictOptions) []llama.PredictOption {
+	ropeFreqBase := float32(10000)
+	ropeFreqScale := float32(1)
+
+	if opts.RopeFreqBase != 0 {
+		ropeFreqBase = opts.RopeFreqBase
+	}
+	if opts.RopeFreqScale != 0 {
+		ropeFreqScale = opts.RopeFreqScale
+	}
+	predictOptions := []llama.PredictOption{
+		llama.SetTemperature(opts.Temperature),
+		llama.SetTopP(opts.TopP),
+		llama.SetTopK(int(opts.TopK)),
+		llama.SetTokens(int(opts.Tokens)),
+		llama.SetThreads(int(opts.Threads)),
+		llama.WithGrammar(opts.Grammar),
+		llama.SetRopeFreqBase(ropeFreqBase),
+		llama.SetRopeFreqScale(ropeFreqScale),
+		llama.SetNegativePromptScale(opts.NegativePromptScale),
+		llama.SetNegativePrompt(opts.NegativePrompt),
+	}
+
+	if opts.PromptCacheAll {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheAll)
+	}
+
+	if opts.PromptCacheRO {
+		predictOptions = append(predictOptions, llama.EnablePromptCacheRO)
+	}
+
+	// Expected absolute path
+	if opts.PromptCachePath != "" {
+		predictOptions = append(predictOptions, llama.SetPathPromptCache(opts.PromptCachePath))
+	}
+
+	if opts.Mirostat != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostat(int(opts.Mirostat)))
+	}
+
+	if opts.MirostatETA != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatETA(opts.MirostatETA))
+	}
+
+	if opts.MirostatTAU != 0 {
+		predictOptions = append(predictOptions, llama.SetMirostatTAU(opts.MirostatTAU))
+	}
+
+	if opts.Debug {
+		predictOptions = append(predictOptions, llama.Debug)
+	}
+
+	predictOptions = append(predictOptions, llama.SetStopWords(opts.StopPrompts...))
+
+	if opts.PresencePenalty != 0 {
+		predictOptions = append(predictOptions, llama.SetPenalty(opts.PresencePenalty))
+	}
+
+	if opts.NKeep != 0 {
+		predictOptions = append(predictOptions, llama.SetNKeep(int(opts.NKeep)))
+	}
+
+	if opts.Batch != 0 {
+		predictOptions = append(predictOptions, llama.SetBatch(int(opts.Batch)))
+	}
+
+	if opts.F16KV {
+		predictOptions = append(predictOptions, llama.EnableF16KV)
+	}
+
+	if opts.IgnoreEOS {
+		predictOptions = append(predictOptions, llama.IgnoreEOS)
+	}
+
+	if opts.Seed != 0 {
+		predictOptions = append(predictOptions, llama.SetSeed(int(opts.Seed)))
+	}
+
+	//predictOptions = append(predictOptions, llama.SetLogitBias(c.Seed))
+
+	predictOptions = append(predictOptions, llama.SetFrequencyPenalty(opts.FrequencyPenalty))
+	predictOptions = append(predictOptions, llama.SetMlock(opts.MLock))
+	predictOptions = append(predictOptions, llama.SetMemoryMap(opts.MMap))
+	predictOptions = append(predictOptions, llama.SetPredictionMainGPU(opts.MainGPU))
+	predictOptions = append(predictOptions, llama.SetPredictionTensorSplit(opts.TensorSplit))
+	predictOptions = append(predictOptions, llama.SetTailFreeSamplingZ(opts.TailFreeSamplingZ))
+	predictOptions = append(predictOptions, llama.SetTypicalP(opts.TypicalP))
+	return predictOptions
+}
+
+func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
+	return llm.llama.Predict(opts.Prompt, buildPredictOptions(opts)...)
+}
+
+func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
+	predictOptions := buildPredictOptions(opts)
+
+	predictOptions = append(predictOptions, llama.SetTokenCallback(func(token string) bool {
+		results <- token
+		return true
+	}))
+
+	go func() {
+		_, err := llm.llama.Predict(opts.Prompt, predictOptions...)
+		if err != nil {
+			fmt.Println("err: ", err)
+		}
+		close(results)
+	}()
+
+	return nil
+}
+
+func (llm *LLM) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
+	predictOptions := buildPredictOptions(opts)
+
+	if len(opts.EmbeddingTokens) > 0 {
+		tokens := []int{}
+		for _, t := range opts.EmbeddingTokens {
+			tokens = append(tokens, int(t))
+		}
+		return llm.llama.TokenEmbeddings(tokens, predictOptions...)
+	}
+
+	return llm.llama.Embeddings(opts.Embeddings, predictOptions...)
+}

backend/go/llm/llama-ggml/main.go

@@ -0,0 +1,19 @@
+package main
+
+import (
+	"flag"
+
+	grpc "github.com/mudler/LocalAI/pkg/grpc"
+)
+
+var (
+	addr = flag.String("addr", "localhost:50051", "the address to connect to")
+)
+
+func main() {
+	flag.Parse()
+
+	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+		panic(err)
+	}
+}

backend/go/transcribe/whisper/whisper.go

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

backend/python/autogptq/Makefile

@@ -0,0 +1,17 @@
+.PHONY: autogptq
+autogptq: protogen
+	bash install.sh
+
+.PHONY: protogen
+protogen: backend_pb2_grpc.py backend_pb2.py
+
+.PHONY: protogen-clean
+protogen-clean:
+	$(RM) backend_pb2_grpc.py backend_pb2.py
+
+backend_pb2_grpc.py backend_pb2.py:
+	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+
+.PHONY: clean
+clean: protogen-clean
+	rm -rf venv __pycache__

backend/python/autogptq/README.md

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

backend/python/autogptq/backend.py

@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+from concurrent import futures
+import argparse
+import signal
+import sys
+import os
+import time
+import base64
+
+import grpc
+import backend_pb2
+import backend_pb2_grpc
+
+from auto_gptq import AutoGPTQForCausalLM
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from transformers import TextGenerationPipeline
+
+_ONE_DAY_IN_SECONDS = 60 * 60 * 24
+
+# If MAX_WORKERS are specified in the environment use it, otherwise default to 1
+MAX_WORKERS = int(os.environ.get('PYTHON_GRPC_MAX_WORKERS', '1'))
+
+# Implement the BackendServicer class with the service methods
+class BackendServicer(backend_pb2_grpc.BackendServicer):
+    def Health(self, request, context):
+        return backend_pb2.Reply(message=bytes("OK", 'utf-8'))
+    def LoadModel(self, request, context):
+        try:
+            device = "cuda:0"
+            if request.Device != "":
+                device = request.Device
+
+            # support loading local model files
+            model_path = os.path.join(os.environ.get('MODELS_PATH', './'), request.Model)
+            tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True, trust_remote_code=request.TrustRemoteCode)
+
+            # support model `Qwen/Qwen-VL-Chat-Int4`
+            if "qwen-vl" in request.Model.lower():
+                self.model_name = "Qwen-VL-Chat"
+                model = AutoModelForCausalLM.from_pretrained(model_path, 
+                    trust_remote_code=request.TrustRemoteCode,
+                    device_map="auto").eval()
+            else:
+                model = AutoGPTQForCausalLM.from_quantized(model_path,
+                    model_basename=request.ModelBaseName,
+                    use_safetensors=True,
+                    trust_remote_code=request.TrustRemoteCode,
+                    device=device,
+                    use_triton=request.UseTriton,
+                    quantize_config=None)
+            
+            self.model = model
+            self.tokenizer = tokenizer
+        except Exception as err:
+            return backend_pb2.Result(success=False, message=f"Unexpected {err=}, {type(err)=}")
+        return backend_pb2.Result(message="Model loaded successfully", success=True)
+
+    def Predict(self, request, context):
+        penalty = 1.0
+        if request.Penalty != 0.0:
+            penalty = request.Penalty
+        tokens = 512
+        if request.Tokens != 0:
+            tokens = request.Tokens
+        top_p = 0.95
+        if request.TopP != 0.0:
+            top_p = request.TopP
+
+        
+        prompt_images = self.recompile_vl_prompt(request)
+        compiled_prompt = prompt_images[0]
+        print(f"Prompt: {compiled_prompt}", file=sys.stderr)
+
+        # Implement Predict RPC
+        pipeline = TextGenerationPipeline(
+            model=self.model, 
+            tokenizer=self.tokenizer,
+            max_new_tokens=tokens,
+            temperature=request.Temperature,
+            top_p=top_p,
+            repetition_penalty=penalty,
+            )
+        t = pipeline(compiled_prompt)[0]["generated_text"]
+        print(f"generated_text: {t}", file=sys.stderr)
+        
+        if compiled_prompt in t:
+            t = t.replace(compiled_prompt, "")
+        # house keeping. Remove the image files from /tmp folder
+        for img_path in prompt_images[1]:
+            try:
+                os.remove(img_path)
+            except Exception as e:
+                print(f"Error removing image file: {img_path}, {e}", file=sys.stderr)
+
+        return backend_pb2.Result(message=bytes(t, encoding='utf-8'))
+
+    def PredictStream(self, request, context):
+        # Implement PredictStream RPC
+        #for reply in some_data_generator():
+        #    yield reply
+        # Not implemented yet
+        return self.Predict(request, context)
+
+    def recompile_vl_prompt(self, request):
+        prompt = request.Prompt
+        image_paths = []
+
+        if "qwen-vl" in self.model_name.lower():
+            # request.Images is an array which contains base64 encoded images. Iterate the request.Images array, decode and save each image to /tmp folder with a random filename.
+            # Then, save the image file paths to an array "image_paths".
+            # read "request.Prompt", replace "[img-%d]" with the image file paths in the order they appear in "image_paths". Save the new prompt to "prompt".
+            for i, img in enumerate(request.Images):
+                timestamp = str(int(time.time() * 1000))  # Generate timestamp
+                img_path = f"/tmp/vl-{timestamp}.jpg"  # Use timestamp in filename
+                with open(img_path, "wb") as f:
+                    f.write(base64.b64decode(img))
+                image_paths.append(img_path)
+                prompt = prompt.replace(f"[img-{i}]", "<img>" + img_path + "</img>,")
+        else:
+            prompt = request.Prompt
+        return (prompt, image_paths)
+
+def serve(address):
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
+    backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
+    server.add_insecure_port(address)
+    server.start()
+    print("Server started. Listening on: " + address, file=sys.stderr)
+
+    # Define the signal handler function
+    def signal_handler(sig, frame):
+        print("Received termination signal. Shutting down...")
+        server.stop(0)
+        sys.exit(0)
+
+    # Set the signal handlers for SIGINT and SIGTERM
+    signal.signal(signal.SIGINT, signal_handler)
+    signal.signal(signal.SIGTERM, signal_handler)
+
+    try:
+        while True:
+            time.sleep(_ONE_DAY_IN_SECONDS)
+    except KeyboardInterrupt:
+        server.stop(0)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the gRPC server.")
+    parser.add_argument(
+        "--addr", default="localhost:50051", help="The address to bind the server to."
+    )
+    args = parser.parse_args()
+
+    serve(args.addr)

backend/python/autogptq/install.sh

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

backend/python/autogptq/requirements-cublas11.txt

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

backend/python/autogptq/requirements-cublas12.txt

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

backend/python/autogptq/requirements-hipblas.txt

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

backend/python/autogptq/requirements-intel.txt

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

backend/python/autogptq/requirements.txt

@@ -0,0 +1,6 @@
+accelerate
+auto-gptq==0.7.1
+grpcio==1.69.0
+protobuf
+certifi
+transformers

backend/python/autogptq/run.sh

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

backend/python/autogptq/test.sh

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

backend/python/bark/backend.py

@@ -61,12 +61,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 certifi

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 grpcio-tools

backend/python/coqui/backend.py

@@ -86,12 +86,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/coqui/requirements-cpu.txt

@@ -1,4 +1,4 @@
-transformers==4.48.3
+transformers
 accelerate
 torch==2.4.1
 coqui-tts

backend/python/coqui/requirements-cublas11.txt

@@ -1,6 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.4.1+cu118
 torchaudio==2.4.1+cu118
-transformers==4.48.3
+transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements-cublas12.txt

@@ -1,5 +1,5 @@
 torch==2.4.1
 torchaudio==2.4.1
-transformers==4.48.3
+transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements-hipblas.txt

@@ -1,6 +1,6 @@
 --extra-index-url https://download.pytorch.org/whl/rocm6.0
 torch==2.4.1+rocm6.0
 torchaudio==2.4.1+rocm6.0
-transformers==4.48.3
+transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements-intel.txt

@@ -5,6 +5,6 @@ torchaudio==2.3.1+cxx11.abi
 oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
 setuptools
-transformers==4.48.3
+transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/backend.py

@@ -19,7 +19,7 @@ import grpc
 
 from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
     EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
-from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline, Lumina2Text2ImgPipeline
+from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
 from diffusers.utils import load_image, export_to_video
 from compel import Compel, ReturnedEmbeddingsType
@@ -159,22 +159,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 torchType = torch.float16
                 variant = "fp16"
 
-            options = request.Options
-
-            # empty dict
-            self.options = {}
-
-            # The options are a list of strings in this form optname:optvalue
-            # We are storing all the options in a dict so we can use it later when
-            # generating the images
-            for opt in options:
-                if ":" not in opt:
-                    continue
-                key, value = opt.split(":")
-                self.options[key] = value
-
-            print(f"Options: {self.options}", file=sys.stderr)
-
             local = False
             modelFile = request.Model
 
@@ -291,12 +275,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
                     if request.LowVRAM:
                         self.pipe.enable_model_cpu_offload()
-            elif request.PipelineType == "Lumina2Text2ImgPipeline":
-                self.pipe = Lumina2Text2ImgPipeline.from_pretrained(
-                    request.Model,
-                    torch_dtype=torch.bfloat16)
-                if request.LowVRAM:
-                    self.pipe.enable_model_cpu_offload()
             elif request.PipelineType == "SanaPipeline":
                 self.pipe = SanaPipeline.from_pretrained(
                     request.Model,
@@ -463,9 +441,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # create a dictionary of parameters by using the keys from EnableParameters and the values from defaults
         kwargs = {key: options.get(key) for key in keys if key in options}
 
-        # populate kwargs from self.options.
-        kwargs.update(self.options)
-
         # Set seed
         if request.seed > 0:
             kwargs["generator"] = torch.Generator(device=self.device).manual_seed(
@@ -526,12 +501,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.72.0
+grpcio==1.69.0
 pillow
 protobuf
 certifi

backend/python/exllama2/backend.py

@@ -105,12 +105,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 certifi
 wheel

backend/python/faster-whisper/backend.py

@@ -62,12 +62,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.TranscriptResult(segments=resultSegments, text=text)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/faster-whisper/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 grpcio-tools

backend/python/kokoro/backend.py

@@ -99,12 +99,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         return backend_pb2.Result(success=True)
 
 def serve(address):
-    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),
-        options=[
-            ('grpc.max_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_send_message_length', 50 * 1024 * 1024),  # 50MB
-            ('grpc.max_receive_message_length', 50 * 1024 * 1024),  # 50MB
-        ])
+    server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS))
     backend_pb2_grpc.add_BackendServicer_to_server(BackendServicer(), server)
     server.add_insecure_port(address)
     server.start()

backend/python/kokoro/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.72.0
+grpcio==1.69.0
 protobuf
 phonemizer
 scipy