feat(l4t): add support for extras images

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

.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"
+  }
+}

.github/labeler.yml

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

.github/workflows/bump_deps.yaml

@@ -9,7 +9,7 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - repository: "ggml-org/llama.cpp"
+          - repository: "ggerganov/llama.cpp"
             variable: "CPPLLAMA_VERSION"
             branch: "master"
           - repository: "ggerganov/whisper.cpp"

.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 }}
@@ -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/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'

Dockerfile

@@ -354,12 +354,14 @@ FROM requirements-drivers
 
 ARG FFMPEG
 ARG BUILD_TYPE
+ARG BUILD_PLATFORM
 ARG TARGETARCH
 ARG IMAGE_TYPE=extras
 ARG EXTRA_BACKENDS
 ARG MAKEFLAGS
 
 ENV BUILD_TYPE=${BUILD_TYPE}
+ENV BUILD_PLATFORM=${BUILD_PLATFORM}
 ENV REBUILD=false
 ENV HEALTHCHECK_ENDPOINT=http://localhost:8080/readyz
 ENV MAKEFLAGS=${MAKEFLAGS}

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

@@ -6,7 +6,9 @@ BINARY_NAME=local-ai
 DETECT_LIBS?=true
 
 # llama.cpp versions
-CPPLLAMA_VERSION?=10f2e81809bbb69ecfe64fc8b4686285f84b0c07
+GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
+GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
+CPPLLAMA_VERSION?=5598f475be3e31430fbe17ebb85654ec90dc201e
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
@@ -22,7 +24,7 @@ BARKCPP_VERSION?=v1.0.0
 
 # stablediffusion.cpp (ggml)
 STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
-STABLEDIFFUSION_GGML_VERSION?=19d876ee300a055629926ff836489901f734f2b7
+STABLEDIFFUSION_GGML_VERSION?=5eb15ef4d022bef4a391de4f5f6556e81fbb5024
 
 ONNX_VERSION?=1.20.0
 ONNX_ARCH?=x64
@@ -149,6 +151,7 @@ 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 ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
@@ -185,6 +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
@@ -218,6 +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 && \
@@ -293,17 +310,19 @@ sources/whisper.cpp:
 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
@@ -311,6 +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
@@ -414,7 +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
@@ -429,7 +449,8 @@ test: prepare test-models/testmodel.ggml grpcs
 	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
@@ -458,6 +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)
@@ -735,6 +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/

README.md

@@ -212,7 +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/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/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/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/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

backend/backend.proto

@@ -165,6 +165,7 @@ message Reply {
 
 message GrammarTrigger {
   string word = 1;
+  bool at_start = 2; 
 }
 
 message ModelOptions {
@@ -228,11 +229,6 @@ 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;
 

backend/cpp/llama/grpc-server.cpp

@@ -469,7 +469,7 @@ struct llama_server_context
     bool has_eos_token      = true;
 
     bool grammar_lazy = false;
-    std::vector<common_grammar_trigger> grammar_triggers;
+    std::vector<common_grammar_trigger> grammar_trigger_words;
 
     int32_t n_ctx;  // total context for all clients / slots
 
@@ -709,7 +709,7 @@ 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_trigger_words = grammar_trigger_words;
         slot->sparams.grammar_lazy = grammar_lazy;
 
         if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
@@ -1155,14 +1155,6 @@ 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;
@@ -1350,7 +1342,7 @@ 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;
@@ -1372,38 +1364,10 @@ 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);
@@ -1663,17 +1627,17 @@ 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
@@ -2024,7 +1988,7 @@ 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;
@@ -2421,12 +2385,12 @@ static void params_parse(const backend::ModelOptions* request,
         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);
+            trigger.word = request->grammartriggers(i).word();
+            trigger.at_start = request->grammartriggers(i).at_start();
+            llama.grammar_trigger_words.push_back(trigger);
             LOG_INFO("grammar trigger", {
-                { "word", trigger.value },
+                { "word", trigger.word },
+                { "at_start", trigger.at_start }
             });
         }
     }

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

@@ -1,13 +1,13 @@
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 7f892beb..0517e529 100644
+index 3cd0d2fa..6c5e811a 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -2766,7 +2766,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
-                 int patch_offset = ctx->has_class_embedding ? 1 : 0;
+@@ -2608,7 +2608,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
+                 struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
                  int* patches_data = (int*)malloc(ggml_nbytes(patches));
                  for (int i = 0; i < num_patches; i++) {
--                    patches_data[i] = i + patch_offset;
-+                    patches_data[i] = i + 1;
+-                    patches_data[i] = i + 1;
++                    patches_data[i] = i;
                  }
                  ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
                  free(patches_data);

backend/cpp/llama/prepare.sh

@@ -1,7 +1,5 @@
 #!/bin/bash
 
-set -e
-
 ## Patches
 ## Apply patches from the `patches` directory
 for patch in $(ls patches); do

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/python/autogptq/requirements-l4t.txt

@@ -0,0 +1,2 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch

backend/python/autogptq/requirements.txt

@@ -1,6 +1,6 @@
 accelerate
 auto-gptq==0.7.1
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi
 transformers

backend/python/bark/requirements-l4t.txt

@@ -0,0 +1,5 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+torchaudio
+transformers
+accelerate

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi

backend/python/common/libbackend.sh

@@ -132,11 +132,16 @@ function installRequirements() {
     declare -a requirementFiles=(
         "${EDIR}/requirements-install.txt"
         "${EDIR}/requirements.txt"
-        "${EDIR}/requirements-${BUILD_TYPE}.txt"
     )
 
-    if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
-        requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}.txt")
+    if [ -n "${BUILD_PLATFORM}" ]; then
+        requirementFiles+=("${EDIR}/requirements-${BUILD_PLATFORM}.txt")
+    else
+        requirementFiles+=("${EDIR}/requirements-${BUILD_TYPE}.txt")
+
+        if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
+            requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}.txt")
+        fi
     fi
 
     # if BUILD_TYPE is empty, we are a CPU build, so we should try to install the CPU requirements
@@ -146,8 +151,14 @@ function installRequirements() {
 
     requirementFiles+=("${EDIR}/requirements-after.txt")
 
-    if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
-        requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}-after.txt")
+    if [ -n "${BUILD_PLATFORM}" ]; then
+        requirementFiles+=("${EDIR}/requirements-${BUILD_PLATFORM}-after.txt")
+    else
+        if [ "x${BUILD_TYPE}" != "x${BUILD_PROFILE}" ]; then
+            requirementFiles+=("${EDIR}/requirements-${BUILD_PROFILE}-after.txt")
+        else
+            requirementFiles+=("${EDIR}/requirements-${BUILD_TYPE}-after.txt")
+        fi
     fi
 
     for reqFile in ${requirementFiles[@]}; do

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 grpcio-tools

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-l4t.txt

@@ -0,0 +1,6 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+torchaudio
+transformers
+accelerate
+coqui-tts

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/backend.py

@@ -159,18 +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:
-                key, value = opt.split(":")
-                self.options[key] = value
-
             local = False
             modelFile = request.Model
 
@@ -453,9 +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(

backend/python/diffusers/requirements-l4t.txt

@@ -0,0 +1,10 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+diffusers
+opencv-python
+transformers
+accelerate
+compel
+peft
+sentencepiece
+optimum-quanto

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.71.0
+grpcio==1.70.0
 pillow
 protobuf
 certifi

backend/python/exllama2/requirements-l4t.txt

@@ -0,0 +1,4 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+transformers
+accelerate

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi
 wheel

backend/python/faster-whisper/requirements-l4t.txt

@@ -0,0 +1,9 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+faster-whisper
+opencv-python
+accelerate
+compel
+peft
+sentencepiece
+optimum-quanto

backend/python/faster-whisper/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 grpcio-tools

backend/python/kokoro/requirements-l4t.txt

@@ -0,0 +1,3 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+transformers

backend/python/kokoro/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 phonemizer
 scipy

backend/python/rerankers/requirements-l4t.txt

@@ -0,0 +1,5 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+transformers
+accelerate
+torch
+rerankers[transformers]

backend/python/rerankers/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi

backend/python/transformers/requirements-l4t.txt

@@ -0,0 +1,9 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+torch
+accelerate
+llvmlite==0.43.0
+numba==0.60.0
+transformers
+bitsandbytes
+outetts
+sentence-transformers==3.4.1

backend/python/transformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi
 setuptools

backend/python/vllm/backend.py

@@ -109,17 +109,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             engine_args.swap_space = request.SwapSpace
         if request.MaxModelLen != 0:
             engine_args.max_model_len = request.MaxModelLen
-        if request.DisableLogStatus:
-            engine_args.disable_log_status = request.DisableLogStatus
-        if request.DType != "":
-            engine_args.dtype = request.DType
-        if request.LimitImagePerPrompt != 0 or request.LimitVideoPerPrompt != 0 or request.LimitAudioPerPrompt != 0:
-            # limit-mm-per-prompt defaults to 1 per modality, based on vLLM docs
-            engine_args.limit_mm_per_prompt = {
-                "image": max(request.LimitImagePerPrompt, 1),
-                "video": max(request.LimitVideoPerPrompt, 1),
-                "audio": max(request.LimitAudioPerPrompt, 1)
-            }
 
         try:
             self.llm = AsyncLLMEngine.from_engine_args(engine_args)
@@ -280,7 +269,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
     def load_image(self, image_path: str):
         """
         Load an image from the given file path or base64 encoded data.
-
+        
         Args:
             image_path (str): The path to the image file or base64 encoded data.
 
@@ -299,7 +288,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
     def load_video(self, video_path: str):
         """
         Load a video from the given file path.
-
+        
         Args:
             video_path (str): The path to the image file.
 
@@ -346,4 +335,4 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    asyncio.run(serve(args.addr))
+    asyncio.run(serve(args.addr))

backend/python/vllm/requirements-after.txt

@@ -1 +0,0 @@
-vllm

backend/python/vllm/requirements-cpu.txt

@@ -1,3 +1,4 @@
 accelerate
 torch==2.4.1
-transformers
+transformers
+vllm

backend/python/vllm/requirements-cublas11.txt

@@ -2,4 +2,5 @@
 accelerate
 torch==2.4.1+cu118
 transformers
-bitsandbytes
+bitsandbytes
+vllm

backend/python/vllm/requirements-cublas12.txt

@@ -1,4 +1,5 @@
 accelerate
 torch==2.4.1
 transformers
-bitsandbytes
+bitsandbytes
+vllm

backend/python/vllm/requirements-hipblas.txt

@@ -2,4 +2,5 @@
 accelerate
 torch==2.4.1+rocm6.0
 transformers
-bitsandbytes
+bitsandbytes
+vllm

backend/python/vllm/requirements-intel.txt

@@ -6,4 +6,5 @@ transformers
 optimum[openvino]
 setuptools
 bitsandbytes
-oneccl_bind_pt==2.3.100+xpu
+oneccl_bind_pt==2.3.100+xpu
+vllm

backend/python/vllm/requirements-l4t.txt

@@ -0,0 +1,7 @@
+--index-url https://pypi.jetson-ai-lab.dev/jp6/cu126/
+accelerate
+torch
+vllm
+transformers
+bitsandbytes
+flash-attn

backend/python/vllm/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.71.0
+grpcio==1.70.0
 protobuf
 certifi
 setuptools

core/application/startup.go

@@ -145,7 +145,13 @@ func New(opts ...config.AppOption) (*Application, error) {
 
 	if options.LoadToMemory != nil {
 		for _, m := range options.LoadToMemory {
-			cfg, err := application.BackendLoader().LoadBackendConfigFileByNameDefaultOptions(m, options)
+			cfg, err := application.BackendLoader().LoadBackendConfigFileByName(m, options.ModelPath,
+				config.LoadOptionDebug(options.Debug),
+				config.LoadOptionThreads(options.Threads),
+				config.LoadOptionContextSize(options.ContextSize),
+				config.LoadOptionF16(options.F16),
+				config.ModelPath(options.ModelPath),
+			)
 			if err != nil {
 				return nil, err
 			}

core/backend/llm.go

@@ -33,7 +33,7 @@ type TokenUsage struct {
 	TimingTokenGeneration  float64
 }
 
-func ModelInference(ctx context.Context, s string, messages []schema.Message, images, videos, audios []string, loader *model.ModelLoader, c *config.BackendConfig, o *config.ApplicationConfig, tokenCallback func(string, TokenUsage) bool) (func() (LLMResponse, error), error) {
+func ModelInference(ctx context.Context, s string, messages []schema.Message, images, videos, audios []string, loader *model.ModelLoader, c config.BackendConfig, o *config.ApplicationConfig, tokenCallback func(string, TokenUsage) bool) (func() (LLMResponse, error), error) {
 	modelFile := c.Model
 
 	// Check if the modelFile exists, if it doesn't try to load it from the gallery
@@ -48,7 +48,7 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 		}
 	}
 
-	opts := ModelOptions(*c, o)
+	opts := ModelOptions(c, o)
 	inferenceModel, err := loader.Load(opts...)
 	if err != nil {
 		return nil, err
@@ -84,7 +84,7 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 
 	// in GRPC, the backend is supposed to answer to 1 single token if stream is not supported
 	fn := func() (LLMResponse, error) {
-		opts := gRPCPredictOpts(*c, loader.ModelPath)
+		opts := gRPCPredictOpts(c, loader.ModelPath)
 		opts.Prompt = s
 		opts.Messages = protoMessages
 		opts.UseTokenizerTemplate = c.TemplateConfig.UseTokenizerTemplate
@@ -116,11 +116,6 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 		}
 
 		if tokenCallback != nil {
-
-			if c.TemplateConfig.ReplyPrefix != "" {
-				tokenCallback(c.TemplateConfig.ReplyPrefix, tokenUsage)
-			}
-
 			ss := ""
 
 			var partialRune []byte
@@ -170,13 +165,8 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 			tokenUsage.TimingTokenGeneration = reply.TimingTokenGeneration
 			tokenUsage.TimingPromptProcessing = reply.TimingPromptProcessing
 
-			response := string(reply.Message)
-			if c.TemplateConfig.ReplyPrefix != "" {
-				response = c.TemplateConfig.ReplyPrefix + response
-			}
-
 			return LLMResponse{
-				Response: response,
+				Response: string(reply.Message),
 				Usage:    tokenUsage,
 			}, err
 		}

core/backend/options.go

@@ -122,6 +122,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 	for _, t := range c.FunctionsConfig.GrammarConfig.GrammarTriggers {
 		triggers = append(triggers, &pb.GrammarTrigger{
 			Word:    t.Word,
+			AtStart: t.AtStart,
 		})
 
 	}
@@ -158,12 +159,6 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		SwapSpace:            int32(c.SwapSpace),
 		MaxModelLen:          int32(c.MaxModelLen),
 		TensorParallelSize:   int32(c.TensorParallelSize),
-		DisableLogStatus:     c.DisableLogStatus,
-		DType:                c.DType,
-		// LimitMMPerPrompt vLLM
-		LimitImagePerPrompt:  int32(c.LimitMMPerPrompt.LimitImagePerPrompt),
-		LimitVideoPerPrompt:  int32(c.LimitMMPerPrompt.LimitVideoPerPrompt),
-		LimitAudioPerPrompt:  int32(c.LimitMMPerPrompt.LimitAudioPerPrompt),
 		MMProj:               c.MMProj,
 		FlashAttention:       c.FlashAttention,
 		CacheTypeKey:         c.CacheTypeK,

core/backend/rerank.go

@@ -9,10 +9,10 @@ import (
 	model "github.com/mudler/LocalAI/pkg/model"
 )
 
-func Rerank(request *proto.RerankRequest, loader *model.ModelLoader, appConfig *config.ApplicationConfig, backendConfig config.BackendConfig) (*proto.RerankResult, error) {
-	opts := ModelOptions(backendConfig, appConfig)
-	rerankModel, err := loader.Load(opts...)
+func Rerank(modelFile string, request *proto.RerankRequest, loader *model.ModelLoader, appConfig *config.ApplicationConfig, backendConfig config.BackendConfig) (*proto.RerankResult, error) {
 
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
+	rerankModel, err := loader.Load(opts...)
 	if err != nil {
 		return nil, err
 	}

core/backend/soundgeneration.go

@@ -13,6 +13,7 @@ import (
 )
 
 func SoundGeneration(
+	modelFile string,
 	text string,
 	duration *float32,
 	temperature *float32,
@@ -24,9 +25,8 @@ func SoundGeneration(
 	backendConfig config.BackendConfig,
 ) (string, *proto.Result, error) {
 
-	opts := ModelOptions(backendConfig, appConfig)
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
 	soundGenModel, err := loader.Load(opts...)
-
 	if err != nil {
 		return "", nil, err
 	}
@@ -44,7 +44,7 @@ func SoundGeneration(
 
 	res, err := soundGenModel.SoundGeneration(context.Background(), &proto.SoundGenerationRequest{
 		Text:        text,
-		Model:       backendConfig.Model,
+		Model:       modelFile,
 		Dst:         filePath,
 		Sample:      doSample,
 		Duration:    duration,

core/backend/tokenize.go

@@ -4,17 +4,19 @@ import (
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/pkg/grpc"
-	"github.com/mudler/LocalAI/pkg/model"
+	model "github.com/mudler/LocalAI/pkg/model"
 )
 
 func ModelTokenize(s string, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (schema.TokenizeResponse, error) {
 
+	modelFile := backendConfig.Model
+
 	var inferenceModel grpc.Backend
 	var err error
 
-	opts := ModelOptions(backendConfig, appConfig)
-	inferenceModel, err = loader.Load(opts...)
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
 
+	inferenceModel, err = loader.Load(opts...)
 	if err != nil {
 		return schema.TokenizeResponse{}, err
 	}

core/backend/transcript.go

@@ -47,7 +47,7 @@ func ModelTranscription(audio, language string, translate bool, ml *model.ModelL
 			tks = append(tks, int(t))
 		}
 		tr.Segments = append(tr.Segments,
-			schema.TranscriptionSegment{
+			schema.Segment{
 				Text:   s.Text,
 				Id:     int(s.Id),
 				Start:  time.Duration(s.Start),

core/backend/tts.go

@@ -14,22 +14,28 @@ import (
 )
 
 func ModelTTS(
+	backend,
 	text,
+	modelFile,
 	voice,
 	language string,
 	loader *model.ModelLoader,
 	appConfig *config.ApplicationConfig,
 	backendConfig config.BackendConfig,
 ) (string, *proto.Result, error) {
-	opts := ModelOptions(backendConfig, appConfig, model.WithDefaultBackendString(model.PiperBackend))
-	ttsModel, err := loader.Load(opts...)
+	bb := backend
+	if bb == "" {
+		bb = model.PiperBackend
+	}
 
+	opts := ModelOptions(backendConfig, appConfig, model.WithBackendString(bb), model.WithModel(modelFile))
+	ttsModel, err := loader.Load(opts...)
 	if err != nil {
 		return "", nil, err
 	}
 
 	if ttsModel == nil {
-		return "", nil, fmt.Errorf("could not load tts model %q", backendConfig.Model)
+		return "", nil, fmt.Errorf("could not load piper model")
 	}
 
 	if err := os.MkdirAll(appConfig.AudioDir, 0750); err != nil {
@@ -39,21 +45,22 @@ func ModelTTS(
 	fileName := utils.GenerateUniqueFileName(appConfig.AudioDir, "tts", ".wav")
 	filePath := filepath.Join(appConfig.AudioDir, fileName)
 
-	// We join the model name to the model path here. This seems to only be done for TTS and is HIGHLY suspect.
-	// This should be addressed in a follow up PR soon.
-	// Copying it over nearly verbatim, as TTS backends are not functional without this.
+	// If the model file is not empty, we pass it joined with the model path
 	modelPath := ""
-	// Checking first that it exists and is not outside ModelPath
-	// TODO: we should actually first check if the modelFile is looking like
-	// a FS path
-	mp := filepath.Join(loader.ModelPath, backendConfig.Model)
-	if _, err := os.Stat(mp); err == nil {
-		if err := utils.VerifyPath(mp, appConfig.ModelPath); err != nil {
-			return "", nil, err
+	if modelFile != "" {
+		// If the model file is not empty, we pass it joined with the model path
+		// Checking first that it exists and is not outside ModelPath
+		// TODO: we should actually first check if the modelFile is looking like
+		// a FS path
+		mp := filepath.Join(loader.ModelPath, modelFile)
+		if _, err := os.Stat(mp); err == nil {
+			if err := utils.VerifyPath(mp, appConfig.ModelPath); err != nil {
+				return "", nil, err
+			}
+			modelPath = mp
+		} else {
+			modelPath = modelFile
 		}
-		modelPath = mp
-	} else {
-		modelPath = backendConfig.Model // skip this step if it fails?????
 	}
 
 	res, err := ttsModel.TTS(context.Background(), &proto.TTSRequest{

core/backend/vad.go

@@ -1,38 +0,0 @@
-package backend
-
-import (
-	"context"
-
-	"github.com/mudler/LocalAI/core/config"
-	"github.com/mudler/LocalAI/core/schema"
-	"github.com/mudler/LocalAI/pkg/grpc/proto"
-	"github.com/mudler/LocalAI/pkg/model"
-)
-
-func VAD(request *schema.VADRequest,
-	ctx context.Context,
-	ml *model.ModelLoader,
-	appConfig *config.ApplicationConfig,
-	backendConfig config.BackendConfig) (*schema.VADResponse, error) {
-	opts := ModelOptions(backendConfig, appConfig)
-	vadModel, err := ml.Load(opts...)
-	if err != nil {
-		return nil, err
-	}
-	req := proto.VADRequest{
-		Audio: request.Audio,
-	}
-	resp, err := vadModel.VAD(ctx, &req)
-	if err != nil {
-		return nil, err
-	}
-
-	segments := []schema.VADSegment{}
-	for _, s := range resp.Segments {
-		segments = append(segments, schema.VADSegment{Start: s.Start, End: s.End})
-	}
-
-	return &schema.VADResponse{
-		Segments: segments,
-	}, nil
-}

core/cli/soundgeneration.go

@@ -86,14 +86,13 @@ func (t *SoundGenerationCMD) Run(ctx *cliContext.Context) error {
 	options := config.BackendConfig{}
 	options.SetDefaults()
 	options.Backend = t.Backend
-	options.Model = t.Model
 
 	var inputFile *string
 	if t.InputFile != "" {
 		inputFile = &t.InputFile
 	}
 
-	filePath, _, err := backend.SoundGeneration(text,
+	filePath, _, err := backend.SoundGeneration(t.Model, text,
 		parseToFloat32Ptr(t.Duration), parseToFloat32Ptr(t.Temperature), &t.DoSample,
 		inputFile, parseToInt32Ptr(t.InputFileSampleDivisor), ml, opts, options)
 

core/cli/tts.go

@@ -52,10 +52,8 @@ func (t *TTSCMD) Run(ctx *cliContext.Context) error {
 
 	options := config.BackendConfig{}
 	options.SetDefaults()
-	options.Backend = t.Backend
-	options.Model = t.Model
 
-	filePath, _, err := backend.ModelTTS(text, t.Voice, t.Language, ml, opts, options)
+	filePath, _, err := backend.ModelTTS(t.Backend, text, t.Model, t.Voice, t.Language, ml, opts, options)
 	if err != nil {
 		return err
 	}

core/config/backend_config.go

@@ -130,28 +130,25 @@ type LLMConfig struct {
 	TrimSpace       []string `yaml:"trimspace"`
 	TrimSuffix      []string `yaml:"trimsuffix"`
 
-	ContextSize          *int             `yaml:"context_size"`
-	NUMA                 bool             `yaml:"numa"`
-	LoraAdapter          string           `yaml:"lora_adapter"`
-	LoraBase             string           `yaml:"lora_base"`
-	LoraAdapters         []string         `yaml:"lora_adapters"`
-	LoraScales           []float32        `yaml:"lora_scales"`
-	LoraScale            float32          `yaml:"lora_scale"`
-	NoMulMatQ            bool             `yaml:"no_mulmatq"`
-	DraftModel           string           `yaml:"draft_model"`
-	NDraft               int32            `yaml:"n_draft"`
-	Quantization         string           `yaml:"quantization"`
-	LoadFormat           string           `yaml:"load_format"`
-	GPUMemoryUtilization float32          `yaml:"gpu_memory_utilization"` // vLLM
-	TrustRemoteCode      bool             `yaml:"trust_remote_code"`      // vLLM
-	EnforceEager         bool             `yaml:"enforce_eager"`          // vLLM
-	SwapSpace            int              `yaml:"swap_space"`             // vLLM
-	MaxModelLen          int              `yaml:"max_model_len"`          // vLLM
-	TensorParallelSize   int              `yaml:"tensor_parallel_size"`   // vLLM
-	DisableLogStatus     bool             `yaml:"disable_log_stats"`      // vLLM
-	DType                string           `yaml:"dtype"`                  // vLLM
-	LimitMMPerPrompt     LimitMMPerPrompt `yaml:"limit_mm_per_prompt"`    // vLLM
-	MMProj               string           `yaml:"mmproj"`
+	ContextSize          *int      `yaml:"context_size"`
+	NUMA                 bool      `yaml:"numa"`
+	LoraAdapter          string    `yaml:"lora_adapter"`
+	LoraBase             string    `yaml:"lora_base"`
+	LoraAdapters         []string  `yaml:"lora_adapters"`
+	LoraScales           []float32 `yaml:"lora_scales"`
+	LoraScale            float32   `yaml:"lora_scale"`
+	NoMulMatQ            bool      `yaml:"no_mulmatq"`
+	DraftModel           string    `yaml:"draft_model"`
+	NDraft               int32     `yaml:"n_draft"`
+	Quantization         string    `yaml:"quantization"`
+	LoadFormat           string    `yaml:"load_format"`
+	GPUMemoryUtilization float32   `yaml:"gpu_memory_utilization"` // vLLM
+	TrustRemoteCode      bool      `yaml:"trust_remote_code"`      // vLLM
+	EnforceEager         bool      `yaml:"enforce_eager"`          // vLLM
+	SwapSpace            int       `yaml:"swap_space"`             // vLLM
+	MaxModelLen          int       `yaml:"max_model_len"`          // vLLM
+	TensorParallelSize   int       `yaml:"tensor_parallel_size"`   // vLLM
+	MMProj               string    `yaml:"mmproj"`
 
 	FlashAttention bool   `yaml:"flash_attention"`
 	NoKVOffloading bool   `yaml:"no_kv_offloading"`
@@ -169,13 +166,6 @@ type LLMConfig struct {
 	CFGScale float32 `yaml:"cfg_scale"` // Classifier-Free Guidance Scale
 }
 
-// LimitMMPerPrompt is a struct that holds the configuration for the limit-mm-per-prompt config in vLLM
-type LimitMMPerPrompt struct {
-	LimitImagePerPrompt int `yaml:"image"`
-	LimitVideoPerPrompt int `yaml:"video"`
-	LimitAudioPerPrompt int `yaml:"audio"`
-}
-
 // AutoGPTQ is a struct that holds the configuration specific to the AutoGPTQ backend
 type AutoGPTQ struct {
 	ModelBaseName    string `yaml:"model_base_name"`
@@ -213,8 +203,6 @@ type TemplateConfig struct {
 	Multimodal string `yaml:"multimodal"`
 
 	JinjaTemplate bool `yaml:"jinja_template"`
-
-	ReplyPrefix string `yaml:"reply_prefix"`
 }
 
 func (c *BackendConfig) UnmarshalYAML(value *yaml.Node) error {
@@ -224,15 +212,7 @@ func (c *BackendConfig) UnmarshalYAML(value *yaml.Node) error {
 		return err
 	}
 	*c = BackendConfig(aux)
-
 	c.KnownUsecases = GetUsecasesFromYAML(c.KnownUsecaseStrings)
-	// Make sure the usecases are valid, we rewrite with what we identified
-	c.KnownUsecaseStrings = []string{}
-	for k, usecase := range GetAllBackendConfigUsecases() {
-		if c.HasUsecases(usecase) {
-			c.KnownUsecaseStrings = append(c.KnownUsecaseStrings, k)
-		}
-	}
 	return nil
 }
 
@@ -307,8 +287,7 @@ func (cfg *BackendConfig) SetDefaults(opts ...ConfigLoaderOption) {
 	defaultTopP := 0.95
 	defaultTopK := 40
 	defaultTemp := 0.9
-	// https://github.com/mudler/LocalAI/issues/2780
-	defaultMirostat := 0
+	defaultMirostat := 2
 	defaultMirostatTAU := 5.0
 	defaultMirostatETA := 0.1
 	defaultTypicalP := 1.0
@@ -457,21 +436,19 @@ func (c *BackendConfig) HasTemplate() bool {
 type BackendConfigUsecases int
 
 const (
-	FLAG_ANY              BackendConfigUsecases = 0b00000000000
-	FLAG_CHAT             BackendConfigUsecases = 0b00000000001
-	FLAG_COMPLETION       BackendConfigUsecases = 0b00000000010
-	FLAG_EDIT             BackendConfigUsecases = 0b00000000100
-	FLAG_EMBEDDINGS       BackendConfigUsecases = 0b00000001000
-	FLAG_RERANK           BackendConfigUsecases = 0b00000010000
-	FLAG_IMAGE            BackendConfigUsecases = 0b00000100000
-	FLAG_TRANSCRIPT       BackendConfigUsecases = 0b00001000000
-	FLAG_TTS              BackendConfigUsecases = 0b00010000000
-	FLAG_SOUND_GENERATION BackendConfigUsecases = 0b00100000000
-	FLAG_TOKENIZE         BackendConfigUsecases = 0b01000000000
-	FLAG_VAD              BackendConfigUsecases = 0b10000000000
+	FLAG_ANY              BackendConfigUsecases = 0b000000000
+	FLAG_CHAT             BackendConfigUsecases = 0b000000001
+	FLAG_COMPLETION       BackendConfigUsecases = 0b000000010
+	FLAG_EDIT             BackendConfigUsecases = 0b000000100
+	FLAG_EMBEDDINGS       BackendConfigUsecases = 0b000001000
+	FLAG_RERANK           BackendConfigUsecases = 0b000010000
+	FLAG_IMAGE            BackendConfigUsecases = 0b000100000
+	FLAG_TRANSCRIPT       BackendConfigUsecases = 0b001000000
+	FLAG_TTS              BackendConfigUsecases = 0b010000000
+	FLAG_SOUND_GENERATION BackendConfigUsecases = 0b100000000
 
 	// Common Subsets
-	FLAG_LLM BackendConfigUsecases = FLAG_CHAT | FLAG_COMPLETION | FLAG_EDIT
+	FLAG_LLM BackendConfigUsecases = FLAG_CHAT & FLAG_COMPLETION & FLAG_EDIT
 )
 
 func GetAllBackendConfigUsecases() map[string]BackendConfigUsecases {
@@ -486,16 +463,10 @@ func GetAllBackendConfigUsecases() map[string]BackendConfigUsecases {
 		"FLAG_TRANSCRIPT":       FLAG_TRANSCRIPT,
 		"FLAG_TTS":              FLAG_TTS,
 		"FLAG_SOUND_GENERATION": FLAG_SOUND_GENERATION,
-		"FLAG_TOKENIZE":         FLAG_TOKENIZE,
-		"FLAG_VAD":              FLAG_VAD,
 		"FLAG_LLM":              FLAG_LLM,
 	}
 }
 
-func stringToFlag(s string) string {
-	return "FLAG_" + strings.ToUpper(s)
-}
-
 func GetUsecasesFromYAML(input []string) *BackendConfigUsecases {
 	if len(input) == 0 {
 		return nil
@@ -503,7 +474,7 @@ func GetUsecasesFromYAML(input []string) *BackendConfigUsecases {
 	result := FLAG_ANY
 	flags := GetAllBackendConfigUsecases()
 	for _, str := range input {
-		flag, exists := flags[stringToFlag(str)]
+		flag, exists := flags["FLAG_"+strings.ToUpper(str)]
 		if exists {
 			result |= flag
 		}
@@ -577,18 +548,5 @@ func (c *BackendConfig) GuessUsecases(u BackendConfigUsecases) bool {
 		}
 	}
 
-	if (u & FLAG_TOKENIZE) == FLAG_TOKENIZE {
-		tokenizeCapableBackends := []string{"llama.cpp", "rwkv"}
-		if !slices.Contains(tokenizeCapableBackends, c.Backend) {
-			return false
-		}
-	}
-
-	if (u & FLAG_VAD) == FLAG_VAD {
-		if c.Backend != "silero-vad" {
-			return false
-		}
-	}
-
 	return true
 }

core/config/backend_config_loader.go

@@ -81,10 +81,10 @@ func readMultipleBackendConfigsFromFile(file string, opts ...ConfigLoaderOption)
 	c := &[]*BackendConfig{}
 	f, err := os.ReadFile(file)
 	if err != nil {
-		return nil, fmt.Errorf("readMultipleBackendConfigsFromFile cannot read config file %q: %w", file, err)
+		return nil, fmt.Errorf("cannot read config file: %w", err)
 	}
 	if err := yaml.Unmarshal(f, c); err != nil {
-		return nil, fmt.Errorf("readMultipleBackendConfigsFromFile cannot unmarshal config file %q: %w", file, err)
+		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
 	}
 
 	for _, cc := range *c {
@@ -101,10 +101,10 @@ func readBackendConfigFromFile(file string, opts ...ConfigLoaderOption) (*Backen
 	c := &BackendConfig{}
 	f, err := os.ReadFile(file)
 	if err != nil {
-		return nil, fmt.Errorf("readBackendConfigFromFile cannot read config file %q: %w", file, err)
+		return nil, fmt.Errorf("cannot read config file: %w", err)
 	}
 	if err := yaml.Unmarshal(f, c); err != nil {
-		return nil, fmt.Errorf("readBackendConfigFromFile cannot unmarshal config file %q: %w", file, err)
+		return nil, fmt.Errorf("cannot unmarshal config file: %w", err)
 	}
 
 	c.SetDefaults(opts...)
@@ -117,9 +117,7 @@ func (bcl *BackendConfigLoader) LoadBackendConfigFileByName(modelName, modelPath
 	// Load a config file if present after the model name
 	cfg := &BackendConfig{
 		PredictionOptions: schema.PredictionOptions{
-			BasicModelRequest: schema.BasicModelRequest{
-				Model: modelName,
-			},
+			Model: modelName,
 		},
 	}
 
@@ -147,15 +145,6 @@ func (bcl *BackendConfigLoader) LoadBackendConfigFileByName(modelName, modelPath
 	return cfg, nil
 }
 
-func (bcl *BackendConfigLoader) LoadBackendConfigFileByNameDefaultOptions(modelName string, appConfig *ApplicationConfig) (*BackendConfig, error) {
-	return bcl.LoadBackendConfigFileByName(modelName, appConfig.ModelPath,
-		LoadOptionDebug(appConfig.Debug),
-		LoadOptionThreads(appConfig.Threads),
-		LoadOptionContextSize(appConfig.ContextSize),
-		LoadOptionF16(appConfig.F16),
-		ModelPath(appConfig.ModelPath))
-}
-
 // This format is currently only used when reading a single file at startup, passed in via ApplicationConfig.ConfigFile
 func (bcl *BackendConfigLoader) LoadMultipleBackendConfigsSingleFile(file string, opts ...ConfigLoaderOption) error {
 	bcl.Lock()
@@ -178,7 +167,7 @@ func (bcl *BackendConfigLoader) LoadBackendConfig(file string, opts ...ConfigLoa
 	defer bcl.Unlock()
 	c, err := readBackendConfigFromFile(file, opts...)
 	if err != nil {
-		return fmt.Errorf("LoadBackendConfig cannot read config file %q: %w", file, err)
+		return fmt.Errorf("cannot read config file: %w", err)
 	}
 
 	if c.Validate() {
@@ -335,10 +324,9 @@ func (bcl *BackendConfigLoader) Preload(modelPath string) error {
 func (bcl *BackendConfigLoader) LoadBackendConfigsFromPath(path string, opts ...ConfigLoaderOption) error {
 	bcl.Lock()
 	defer bcl.Unlock()
-
 	entries, err := os.ReadDir(path)
 	if err != nil {
-		return fmt.Errorf("LoadBackendConfigsFromPath cannot read directory '%s': %w", path, err)
+		return fmt.Errorf("cannot read directory '%s': %w", path, err)
 	}
 	files := make([]fs.FileInfo, 0, len(entries))
 	for _, entry := range entries {
@@ -356,13 +344,13 @@ func (bcl *BackendConfigLoader) LoadBackendConfigsFromPath(path string, opts ...
 		}
 		c, err := readBackendConfigFromFile(filepath.Join(path, file.Name()), opts...)
 		if err != nil {
-			log.Error().Err(err).Str("File Name", file.Name()).Msgf("LoadBackendConfigsFromPath cannot read config file")
+			log.Error().Err(err).Msgf("cannot read config file: %s", file.Name())
 			continue
 		}
 		if c.Validate() {
 			bcl.configs[c.Name] = *c
 		} else {
-			log.Error().Err(err).Str("Name", c.Name).Msgf("config is not valid")
+			log.Error().Err(err).Msgf("config is not valid")
 		}
 	}
 

core/config/guesser.go

@@ -161,11 +161,10 @@ func guessDefaultsFromFile(cfg *BackendConfig, modelPath string) {
 	}
 
 	// We try to guess only if we don't have a template defined already
-	guessPath := filepath.Join(modelPath, cfg.ModelFileName())
-	f, err := gguf.ParseGGUFFile(guessPath)
+	f, err := gguf.ParseGGUFFile(filepath.Join(modelPath, cfg.ModelFileName()))
 	if err != nil {
 		// Only valid for gguf files
-		log.Debug().Str("filePath", guessPath).Msg("guessDefaultsFromFile: not a GGUF file")
+		log.Debug().Msgf("guessDefaultsFromFile: %s", "not a GGUF file")
 		return
 	}
 

core/gallery/gallery.go

@@ -29,8 +29,6 @@ func InstallModelFromGallery(galleries []config.Gallery, name string, basePath s
 			if err != nil {
 				return err
 			}
-			config.Description = model.Description
-			config.License = model.License
 		} else if len(model.ConfigFile) > 0 {
 			// TODO: is this worse than using the override method with a blank cfg yaml?
 			reYamlConfig, err := yaml.Marshal(model.ConfigFile)
@@ -116,7 +114,7 @@ func FindModel(models []*GalleryModel, name string, basePath string) *GalleryMod
 // List available models
 // Models galleries are a list of yaml files that are hosted on a remote server (for example github).
 // Each yaml file contains a list of models that can be downloaded and optionally overrides to define a new model setting.
-func AvailableGalleryModels(galleries []config.Gallery, basePath string) (GalleryModels, error) {
+func AvailableGalleryModels(galleries []config.Gallery, basePath string) ([]*GalleryModel, error) {
 	var models []*GalleryModel
 
 	// Get models from galleries

core/gallery/models_test.go

@@ -48,10 +48,8 @@ var _ = Describe("Model test", func() {
 			defer os.RemoveAll(tempdir)
 
 			gallery := []GalleryModel{{
-				Metadata: Metadata{
-					Name: "bert",
-					URL:  bertEmbeddingsURL,
-				},
+				Name: "bert",
+				URL:  bertEmbeddingsURL,
 			}}
 			out, err := yaml.Marshal(gallery)
 			Expect(err).ToNot(HaveOccurred())

core/gallery/request.go

@@ -11,14 +11,6 @@ import (
 // It is used to install the model by resolving the URL and downloading the files.
 // The other fields are used to override the configuration of the model.
 type GalleryModel struct {
-	Metadata `json:",inline" yaml:",inline"`
-	// config_file is read in the situation where URL is blank - and therefore this is a base config.
-	ConfigFile map[string]interface{} `json:"config_file,omitempty" yaml:"config_file,omitempty"`
-	// Overrides are used to override the configuration of the model located at URL
-	Overrides map[string]interface{} `json:"overrides,omitempty" yaml:"overrides,omitempty"`
-}
-
-type Metadata struct {
 	URL         string   `json:"url,omitempty" yaml:"url,omitempty"`
 	Name        string   `json:"name,omitempty" yaml:"name,omitempty"`
 	Description string   `json:"description,omitempty"  yaml:"description,omitempty"`
@@ -26,6 +18,10 @@ type Metadata struct {
 	URLs        []string `json:"urls,omitempty" yaml:"urls,omitempty"`
 	Icon        string   `json:"icon,omitempty" yaml:"icon,omitempty"`
 	Tags        []string `json:"tags,omitempty" yaml:"tags,omitempty"`
+	// config_file is read in the situation where URL is blank - and therefore this is a base config.
+	ConfigFile map[string]interface{} `json:"config_file,omitempty" yaml:"config_file,omitempty"`
+	// Overrides are used to override the configuration of the model located at URL
+	Overrides map[string]interface{} `json:"overrides,omitempty" yaml:"overrides,omitempty"`
 	// AdditionalFiles are used to add additional files to the model
 	AdditionalFiles []File `json:"files,omitempty" yaml:"files,omitempty"`
 	// Gallery is a reference to the gallery which contains the model
@@ -62,15 +58,3 @@ func (gm GalleryModels) FindByName(name string) *GalleryModel {
 	}
 	return nil
 }
-
-func (gm GalleryModels) Paginate(pageNum int, itemsNum int) GalleryModels {
-	start := (pageNum - 1) * itemsNum
-	end := start + itemsNum
-	if start > len(gm) {
-		start = len(gm)
-	}
-	if end > len(gm) {
-		end = len(gm)
-	}
-	return gm[start:end]
-}

core/gallery/request_test.go

@@ -9,11 +9,7 @@ import (
 var _ = Describe("Gallery API tests", func() {
 	Context("requests", func() {
 		It("parses github with a branch", func() {
-			req := GalleryModel{
-				Metadata: Metadata{
-					URL: "github:go-skynet/model-gallery/gpt4all-j.yaml@main",
-				},
-			}
+			req := GalleryModel{URL: "github:go-skynet/model-gallery/gpt4all-j.yaml@main"}
 			e, err := GetGalleryConfigFromURL(req.URL, "")
 			Expect(err).ToNot(HaveOccurred())
 			Expect(e.Name).To(Equal("gpt4all-j"))

core/http/app.go

@@ -130,6 +130,7 @@ func API(application *application.Application) (*fiber.App, error) {
 				return metricsService.Shutdown()
 			})
 		}
+
 	}
 	// Health Checks should always be exempt from auth, so register these first
 	routes.HealthRoutes(router)
@@ -139,28 +140,6 @@ func API(application *application.Application) (*fiber.App, error) {
 		return nil, fmt.Errorf("failed to create key auth config: %w", err)
 	}
 
-	httpFS := http.FS(embedDirStatic)
-
-	router.Use(favicon.New(favicon.Config{
-		URL:        "/favicon.ico",
-		FileSystem: httpFS,
-		File:       "static/favicon.ico",
-	}))
-
-	router.Use("/static", filesystem.New(filesystem.Config{
-		Root:       httpFS,
-		PathPrefix: "static",
-		Browse:     true,
-	}))
-
-	if application.ApplicationConfig().ImageDir != "" {
-		router.Static("/generated-images", application.ApplicationConfig().ImageDir)
-	}
-
-	if application.ApplicationConfig().AudioDir != "" {
-		router.Static("/generated-audio", application.ApplicationConfig().AudioDir)
-	}
-
 	// Auth is applied to _all_ endpoints. No exceptions. Filtering out endpoints to bypass is the role of the Filter property of the KeyAuth Configuration
 	router.Use(v2keyauth.New(*kaConfig))
 
@@ -188,15 +167,27 @@ func API(application *application.Application) (*fiber.App, error) {
 	galleryService := services.NewGalleryService(application.ApplicationConfig())
 	galleryService.Start(application.ApplicationConfig().Context, application.BackendLoader())
 
-	requestExtractor := middleware.NewRequestExtractor(application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
-
-	routes.RegisterElevenLabsRoutes(router, requestExtractor, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
-	routes.RegisterLocalAIRoutes(router, requestExtractor, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig(), galleryService)
-	routes.RegisterOpenAIRoutes(router, requestExtractor, application)
+	routes.RegisterElevenLabsRoutes(router, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
+	routes.RegisterLocalAIRoutes(router, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig(), galleryService)
+	routes.RegisterOpenAIRoutes(router, application)
 	if !application.ApplicationConfig().DisableWebUI {
 		routes.RegisterUIRoutes(router, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig(), galleryService)
 	}
-	routes.RegisterJINARoutes(router, requestExtractor, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
+	routes.RegisterJINARoutes(router, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
+
+	httpFS := http.FS(embedDirStatic)
+
+	router.Use(favicon.New(favicon.Config{
+		URL:        "/favicon.ico",
+		FileSystem: httpFS,
+		File:       "static/favicon.ico",
+	}))
+
+	router.Use("/static", filesystem.New(filesystem.Config{
+		Root:       httpFS,
+		PathPrefix: "static",
+		Browse:     true,
+	}))
 
 	// Define a custom 404 handler
 	// Note: keep this at the bottom!

core/http/app_test.go

@@ -299,18 +299,14 @@ var _ = Describe("API test", func() {
 
 			g := []gallery.GalleryModel{
 				{
-					Metadata: gallery.Metadata{
-						Name: "bert",
-						URL:  bertEmbeddingsURL,
-					},
+					Name: "bert",
+					URL:  bertEmbeddingsURL,
 				},
 				{
-					Metadata: gallery.Metadata{
-						Name:            "bert2",
-						URL:             bertEmbeddingsURL,
-						AdditionalFiles: []gallery.File{{Filename: "foo.yaml", URI: bertEmbeddingsURL}},
-					},
-					Overrides: map[string]interface{}{"foo": "bar"},
+					Name:            "bert2",
+					URL:             bertEmbeddingsURL,
+					Overrides:       map[string]interface{}{"foo": "bar"},
+					AdditionalFiles: []gallery.File{{Filename: "foo.yaml", URI: bertEmbeddingsURL}},
 				},
 			}
 			out, err := yaml.Marshal(g)
@@ -526,6 +522,77 @@ var _ = Describe("API test", func() {
 				Expect(content["usage"]).To(ContainSubstring("You can test this model with curl like this"))
 			})
 
+			It("runs openllama(llama-ggml backend)", Label("llama"), func() {
+				if runtime.GOOS != "linux" {
+					Skip("test supported only on linux")
+				}
+				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
+					URL:       "github:go-skynet/model-gallery/openllama_3b.yaml",
+					Name:      "openllama_3b",
+					Overrides: map[string]interface{}{"backend": "llama-ggml", "mmap": true, "f16": true, "context_size": 128},
+				})
+
+				Expect(response["uuid"]).ToNot(BeEmpty(), fmt.Sprint(response))
+
+				uuid := response["uuid"].(string)
+
+				Eventually(func() bool {
+					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
+					return response["processed"].(bool)
+				}, "360s", "10s").Should(Equal(true))
+
+				By("testing completion")
+				resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "openllama_3b", Prompt: "Count up to five: one, two, three, four, "})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp.Choices)).To(Equal(1))
+				Expect(resp.Choices[0].Text).To(ContainSubstring("five"))
+
+				By("testing functions")
+				resp2, err := client.CreateChatCompletion(
+					context.TODO(),
+					openai.ChatCompletionRequest{
+						Model: "openllama_3b",
+						Messages: []openai.ChatCompletionMessage{
+							{
+								Role:    "user",
+								Content: "What is the weather like in San Francisco (celsius)?",
+							},
+						},
+						Functions: []openai.FunctionDefinition{
+							openai.FunctionDefinition{
+								Name:        "get_current_weather",
+								Description: "Get the current weather",
+								Parameters: jsonschema.Definition{
+									Type: jsonschema.Object,
+									Properties: map[string]jsonschema.Definition{
+										"location": {
+											Type:        jsonschema.String,
+											Description: "The city and state, e.g. San Francisco, CA",
+										},
+										"unit": {
+											Type: jsonschema.String,
+											Enum: []string{"celcius", "fahrenheit"},
+										},
+									},
+									Required: []string{"location"},
+								},
+							},
+						},
+					})
+				Expect(err).ToNot(HaveOccurred())
+				Expect(len(resp2.Choices)).To(Equal(1))
+				Expect(resp2.Choices[0].Message.FunctionCall).ToNot(BeNil())
+				Expect(resp2.Choices[0].Message.FunctionCall.Name).To(Equal("get_current_weather"), resp2.Choices[0].Message.FunctionCall.Name)
+
+				var res map[string]string
+				err = json.Unmarshal([]byte(resp2.Choices[0].Message.FunctionCall.Arguments), &res)
+				Expect(err).ToNot(HaveOccurred())
+				Expect(res["location"]).To(ContainSubstring("San Francisco"), fmt.Sprint(res))
+				Expect(res["unit"]).To(Equal("celcius"), fmt.Sprint(res))
+				Expect(string(resp2.Choices[0].FinishReason)).To(Equal("function_call"), fmt.Sprint(resp2.Choices[0].FinishReason))
+
+			})
+
 			It("runs openllama gguf(llama-cpp)", Label("llama-gguf"), func() {
 				if runtime.GOOS != "linux" {
 					Skip("test supported only on linux")

core/http/ctx/fiber.go

@@ -0,0 +1,47 @@
+package fiberContext
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/gofiber/fiber/v2"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/services"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/rs/zerolog/log"
+)
+
+// ModelFromContext returns the model from the context
+// If no model is specified, it will take the first available
+// Takes a model string as input which should be the one received from the user request.
+// It returns the model name resolved from the context and an error if any.
+func ModelFromContext(ctx *fiber.Ctx, cl *config.BackendConfigLoader, loader *model.ModelLoader, modelInput string, firstModel bool) (string, error) {
+	if ctx.Params("model") != "" {
+		modelInput = ctx.Params("model")
+	}
+	if ctx.Query("model") != "" {
+		modelInput = ctx.Query("model")
+	}
+	// Set model from bearer token, if available
+	bearer := strings.TrimLeft(ctx.Get("authorization"), "Bear ") // Reduced duplicate characters of Bearer
+	bearerExists := bearer != "" && loader.ExistsInModelPath(bearer)
+
+	// If no model was specified, take the first available
+	if modelInput == "" && !bearerExists && firstModel {
+		models, _ := services.ListModels(cl, loader, config.NoFilterFn, services.SKIP_IF_CONFIGURED)
+		if len(models) > 0 {
+			modelInput = models[0]
+			log.Debug().Msgf("No model specified, using: %s", modelInput)
+		} else {
+			log.Debug().Msgf("No model specified, returning error")
+			return "", fmt.Errorf("no model specified")
+		}
+	}
+
+	// If a model is found in bearer token takes precedence
+	if bearerExists {
+		log.Debug().Msgf("Using model from bearer token: %s", bearer)
+		modelInput = bearer
+	}
+	return modelInput, nil
+}