chore: ⬆️ Update ggerganov/llama.cpp to ba8a1f9c5b675459c55a83e3f97f10df3a66c788 (#4575)

⬆️ Update ggerganov/llama.cpp

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

.env

@@ -82,6 +82,15 @@
 # Enable to allow p2p mode
 # LOCALAI_P2P=true
 
+# Enable to use federated mode
+# LOCALAI_FEDERATED=true
+
+# Enable to start federation server
+# FEDERATED_SERVER=true
+
+# Define to use federation token
+# TOKEN=""
+
 ### Watchdog settings
 ###
 # Enables watchdog to kill backends that are inactive for too much time

.gitattributes

@@ -1 +1,2 @@
 *.sh text eol=lf
+backend/cpp/llama/*.hpp linguist-vendored

.github/labeler.yml

@@ -1,6 +1,15 @@
 enhancements:
  - head-branch: ['^feature', 'feature']
 
+dependencies:
+- any:
+  - changed-files:
+    - any-glob-to-any-file: 'Makefile'
+  - changed-files:
+    - any-glob-to-any-file: '*.mod'
+  - changed-files:
+    - any-glob-to-any-file: '*.sum'
+
 kind/documentation:
 - any:
   - changed-files:

.github/workflows/bump_deps.yaml

@@ -12,23 +12,14 @@ jobs:
           - repository: "ggerganov/llama.cpp"
             variable: "CPPLLAMA_VERSION"
             branch: "master"
-          - repository: "go-skynet/go-ggml-transformers.cpp"
-            variable: "GOGGMLTRANSFORMERS_VERSION"
-            branch: "master"
-          - repository: "donomii/go-rwkv.cpp"
-            variable: "RWKV_VERSION"
-            branch: "main"
           - repository: "ggerganov/whisper.cpp"
             variable: "WHISPER_CPP_VERSION"
             branch: "master"
-          - repository: "go-skynet/go-bert.cpp"
-            variable: "BERT_VERSION"
-            branch: "master"
-          - repository: "go-skynet/bloomz.cpp"
-            variable: "BLOOMZ_VERSION"
+          - repository: "PABannier/bark.cpp"
+            variable: "BARKCPP_VERSION"
             branch: "main"
-          - repository: "mudler/go-ggllm.cpp"
-            variable: "GOGGLLM_VERSION"
+          - repository: "leejet/stable-diffusion.cpp"
+            variable: "STABLEDIFFUSION_GGML_VERSION"
             branch: "master"
           - repository: "mudler/go-stable-diffusion"
             variable: "STABLEDIFFUSION_VERSION"

.github/workflows/checksum_checker.yaml

@@ -23,7 +23,7 @@ jobs:
           sudo pip install --upgrade pip
           pip install huggingface_hub
       - name: 'Setup yq'
-        uses: dcarbone/install-yq-action@v1.1.1
+        uses: dcarbone/install-yq-action@v1.3.1
         with:
           version: 'v4.44.2'
           download-compressed: true

.github/workflows/deploy-explorer.yaml

@@ -33,7 +33,7 @@ jobs:
         run: |
           CGO_ENABLED=0 make build-api
       - name: rm
-        uses: appleboy/ssh-action@v1.1.0
+        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@v1.1.0
+        uses: appleboy/ssh-action@v1.2.0
         with:
             host: ${{ secrets.EXPLORER_SSH_HOST }}
             username: ${{ secrets.EXPLORER_SSH_USERNAME }}

.github/workflows/generate_intel_image.yaml

@@ -15,7 +15,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - base-image: intel/oneapi-basekit:2024.2.0-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.yml

@@ -280,6 +280,7 @@ jobs:
       makeflags: ${{ matrix.makeflags }}
       latest-image: ${{ matrix.latest-image }}
       latest-image-aio: ${{ matrix.latest-image-aio }}
+      skip-drivers: ${{ matrix.skip-drivers }}
     secrets:
       dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
       dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
@@ -301,6 +302,7 @@ jobs:
             latest-image: 'latest-cpu'
             latest-image-aio: 'latest-aio-cpu'
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "11"
             cuda-minor-version: "7"
@@ -312,6 +314,7 @@ jobs:
             base-image: "ubuntu:22.04"
             runs-on: 'arc-runner-set'
             makeflags: "--jobs=4 --output-sync=target"
+            skip-drivers: 'false'
           - build-type: 'cublas'
             cuda-major-version: "12"
             cuda-minor-version: "0"
@@ -323,6 +326,7 @@ jobs:
             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"
@@ -334,6 +338,7 @@ jobs:
             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"
@@ -344,6 +349,7 @@ jobs:
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
           - build-type: 'vulkan'
             platforms: 'linux/amd64'
@@ -354,4 +360,45 @@ jobs:
             image-type: 'core'
             runs-on: 'arc-runner-set'
             base-image: "ubuntu:22.04"
+            skip-drivers: 'false'
             makeflags: "--jobs=4 --output-sync=target"
+#  parallel-builds:
+#    uses: ./.github/workflows/image_build.yml
+#    with:
+#      tag-latest: ${{ matrix.tag-latest }}
+#      tag-suffix: ${{ matrix.tag-suffix }}
+#      ffmpeg: ${{ matrix.ffmpeg }}
+#      image-type: ${{ matrix.image-type }}
+#      build-type: ${{ matrix.build-type }}
+#      cuda-major-version: ${{ matrix.cuda-major-version }}
+#      cuda-minor-version: ${{ matrix.cuda-minor-version }}
+#      platforms: ${{ matrix.platforms }}
+#      runs-on: ${{ matrix.runs-on }}
+#      aio: ${{ matrix.aio }}
+#      base-image: ${{ matrix.base-image }}
+#      grpc-base-image: ${{ matrix.grpc-base-image }}
+#      makeflags: ${{ matrix.makeflags }}
+#      latest-image: ${{ matrix.latest-image }}
+#      latest-image-aio: ${{ matrix.latest-image-aio }}
+#      skip-drivers: ${{ matrix.skip-drivers }}
+#    secrets:
+#      dockerUsername: ${{ secrets.DOCKERHUB_USERNAME }}
+#      dockerPassword: ${{ secrets.DOCKERHUB_PASSWORD }}
+#      quayUsername: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+#      quayPassword: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+#    strategy:
+#      matrix:
+#        include:
+#          - build-type: 'cublas'
+#            cuda-major-version: "12"
+#            cuda-minor-version: "0"
+#            platforms: 'linux/arm64'
+#            tag-latest: 'false'
+#            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"
+#            runs-on: 'self-hosted'
+#            makeflags: "--jobs=4 --output-sync=target"
+#            skip-drivers: 'true'

.github/workflows/image_build.yml

@@ -49,6 +49,10 @@ on:
         description: 'FFMPEG'
         default: ''
         type: string
+      skip-drivers:
+        description: 'Skip drivers by default'
+        default: 'false'
+        type: string
       image-type:
         description: 'Image type'
         default: ''
@@ -234,6 +238,7 @@ jobs:
             GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
             GRPC_VERSION=v1.65.0
             MAKEFLAGS=${{ inputs.makeflags }}
+            SKIP_DRIVERS=${{ inputs.skip-drivers }}
           context: .
           file: ./Dockerfile
           cache-from: type=gha
@@ -262,6 +267,7 @@ jobs:
             GRPC_MAKEFLAGS=--jobs=4 --output-sync=target
             GRPC_VERSION=v1.65.0
             MAKEFLAGS=${{ inputs.makeflags }}
+            SKIP_DRIVERS=${{ inputs.skip-drivers }}
           context: .
           file: ./Dockerfile
           cache-from: type=gha

.gitignore

@@ -2,6 +2,7 @@
 /sources/
 __pycache__/
 *.a
+*.o
 get-sources
 prepare-sources
 /backend/cpp/llama/grpc-server
@@ -12,7 +13,6 @@ prepare-sources
 
 go-ggml-transformers
 go-gpt2
-go-rwkv
 whisper.cpp
 /bloomz
 go-bert

Dockerfile

@@ -115,12 +115,13 @@ FROM requirements-${IMAGE_TYPE} AS requirements-drivers
 ARG BUILD_TYPE
 ARG CUDA_MAJOR_VERSION=12
 ARG CUDA_MINOR_VERSION=0
+ARG SKIP_DRIVERS=false
 
 ENV BUILD_TYPE=${BUILD_TYPE}
 
 # Vulkan requirements
 RUN <<EOT bash
-    if [ "${BUILD_TYPE}" = "vulkan" ]; then
+    if [ "${BUILD_TYPE}" = "vulkan" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         apt-get update && \
         apt-get install -y  --no-install-recommends \
             software-properties-common pciutils wget gpg-agent && \
@@ -136,7 +137,7 @@ EOT
 
 # CuBLAS requirements
 RUN <<EOT bash
-    if [ "${BUILD_TYPE}" = "cublas" ]; then
+    if [ "${BUILD_TYPE}" = "cublas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then
         apt-get update && \
         apt-get install -y  --no-install-recommends \
             software-properties-common pciutils
@@ -162,7 +163,7 @@ RUN <<EOT bash
 EOT
 
 # If we are building with clblas support, we need the libraries for the builds
-RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
+RUN if [ "${BUILD_TYPE}" = "clblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
         apt-get update && \
         apt-get install -y --no-install-recommends \
             libclblast-dev && \
@@ -170,7 +171,7 @@ RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
         rm -rf /var/lib/apt/lists/* \
     ; fi
 
-RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
+RUN if [ "${BUILD_TYPE}" = "hipblas" ] && [ "${SKIP_DRIVERS}" = "false" ]; then \
         apt-get update && \
         apt-get install -y --no-install-recommends \
             hipblas-dev \

Makefile

@@ -8,19 +8,11 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=6423c65aa8be1b98f990cf207422505ac5a441a1
-
-# go-rwkv version
-RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
-RWKV_VERSION?=661e7ae26d442f5cfebd2a0881b44e8c55949ec6
+CPPLLAMA_VERSION?=ba8a1f9c5b675459c55a83e3f97f10df3a66c788
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
-WHISPER_CPP_VERSION?=31aea563a83803c710691fed3e8d700e06ae6788
-
-# bert.cpp version
-BERT_REPO?=https://github.com/go-skynet/go-bert.cpp
-BERT_VERSION?=710044b124545415f555e4260d16b146c725a6e4
+WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
 
 # go-piper version
 PIPER_REPO?=https://github.com/mudler/go-piper
@@ -34,6 +26,18 @@ STABLEDIFFUSION_VERSION?=4a3cd6aeae6f66ee57eae9a0075f8c58c3a6a38f
 TINYDREAM_REPO?=https://github.com/M0Rf30/go-tiny-dream
 TINYDREAM_VERSION?=c04fa463ace9d9a6464313aa5f9cd0f953b6c057
 
+# bark.cpp
+BARKCPP_REPO?=https://github.com/PABannier/bark.cpp.git
+BARKCPP_VERSION?=v1.0.0
+
+# stablediffusion.cpp (ggml)
+STABLEDIFFUSION_GGML_REPO?=https://github.com/leejet/stable-diffusion.cpp
+STABLEDIFFUSION_GGML_VERSION?=dcf91f9e0f2cbf9da472ee2a556751ed4bab2d2a
+
+ONNX_VERSION?=1.20.0
+ONNX_ARCH?=x64
+ONNX_OS?=linux
+
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
 export CMAKE_ARGS?=
@@ -45,6 +49,7 @@ CGO_LDFLAGS_WHISPER+=-lggml
 CUDA_LIBPATH?=/usr/local/cuda/lib64/
 GO_TAGS?=
 BUILD_ID?=
+NATIVE?=false
 
 TEST_DIR=/tmp/test
 
@@ -83,7 +88,25 @@ ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
 
+# IF native is false, we add -DGGML_NATIVE=OFF to CMAKE_ARGS
+ifeq ($(NATIVE),false)
+	CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# Detect if we are running on arm64
+ifneq (,$(findstring aarch64,$(shell uname -m)))
+	ONNX_ARCH=aarch64
+endif
+
 ifeq ($(OS),Darwin)
+	ONNX_OS=osx
+	ifneq (,$(findstring aarch64,$(shell uname -m)))
+		ONNX_ARCH=arm64
+	else ifneq (,$(findstring arm64,$(shell uname -m)))
+		ONNX_ARCH=arm64
+	else
+		ONNX_ARCH=x86_64
+	endif
 
 	ifeq ($(OSX_SIGNING_IDENTITY),)
 		OSX_SIGNING_IDENTITY := $(shell security find-identity -v -p codesigning | grep '"' | head -n 1 | sed -E 's/.*"(.*)"/\1/')
@@ -138,10 +161,10 @@ ifeq ($(BUILD_TYPE),hipblas)
 	export CC=$(ROCM_HOME)/llvm/bin/clang
 	# llama-ggml has no hipblas support, so override it here.
 	export STABLE_BUILD_TYPE=
-	export GGML_HIPBLAS=1
+	export GGML_HIP=1
 	GPU_TARGETS ?= gfx900,gfx906,gfx908,gfx940,gfx941,gfx942,gfx90a,gfx1030,gfx1031,gfx1100,gfx1101
 	AMDGPU_TARGETS ?= "$(GPU_TARGETS)"
-	CMAKE_ARGS+=-DGGML_HIPBLAS=ON -DAMDGPU_TARGETS="$(AMDGPU_TARGETS)" -DGPU_TARGETS="$(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
 endif
 
@@ -179,16 +202,23 @@ ifeq ($(findstring tts,$(GO_TAGS)),tts)
 endif
 
 ALL_GRPC_BACKENDS=backend-assets/grpc/huggingface
-ALL_GRPC_BACKENDS+=backend-assets/grpc/bert-embeddings
 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-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/rwkv
 ALL_GRPC_BACKENDS+=backend-assets/grpc/whisper
+
+ifeq ($(ONNX_OS),linux)
+ifeq ($(ONNX_ARCH),x64)
+	ALL_GRPC_BACKENDS+=backend-assets/grpc/bark-cpp
+	ALL_GRPC_BACKENDS+=backend-assets/grpc/stablediffusion-ggml
+endif
+endif
+
 ALL_GRPC_BACKENDS+=backend-assets/grpc/local-store
+ALL_GRPC_BACKENDS+=backend-assets/grpc/silero-vad
 ALL_GRPC_BACKENDS+=$(OPTIONAL_GRPC)
 # Use filter-out to remove the specified backends
 ALL_GRPC_BACKENDS := $(filter-out $(SKIP_GRPC_BACKEND),$(ALL_GRPC_BACKENDS))
@@ -209,19 +239,6 @@ endif
 
 all: help
 
-## BERT embeddings
-sources/go-bert.cpp:
-	mkdir -p sources/go-bert.cpp
-	cd sources/go-bert.cpp && \
-	git init && \
-	git remote add origin $(BERT_REPO) && \
-	git fetch origin && \
-	git checkout $(BERT_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-sources/go-bert.cpp/libgobert.a: sources/go-bert.cpp
-	$(MAKE) -C sources/go-bert.cpp libgobert.a
-
 ## go-llama.cpp
 sources/go-llama.cpp:
 	mkdir -p sources/go-llama.cpp
@@ -235,6 +252,23 @@ sources/go-llama.cpp:
 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 && \
+	cd sources/bark.cpp && \
+	git checkout $(BARKCPP_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/bark.cpp/build/libbark.a: sources/bark.cpp
+	cd sources/bark.cpp && \
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) .. && \
+	cmake --build . --config Release
+
+backend/go/bark/libbark.a: sources/bark.cpp/build/libbark.a
+	$(MAKE) -C backend/go/bark libbark.a
+
 ## go-piper
 sources/go-piper:
 	mkdir -p sources/go-piper
@@ -248,21 +282,7 @@ sources/go-piper:
 sources/go-piper/libpiper_binding.a: sources/go-piper
 	$(MAKE) -C sources/go-piper libpiper_binding.a example/main piper.o
 
-
-## RWKV
-sources/go-rwkv.cpp:
-	mkdir -p sources/go-rwkv.cpp
-	cd sources/go-rwkv.cpp && \
-	git init && \
-	git remote add origin $(RWKV_REPO) && \
-	git fetch origin && \
-	git checkout $(RWKV_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-sources/go-rwkv.cpp/librwkv.a: sources/go-rwkv.cpp
-	cd sources/go-rwkv.cpp && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
-
-## stable diffusion
+## stable diffusion (onnx)
 sources/go-stable-diffusion:
 	mkdir -p sources/go-stable-diffusion
 	cd sources/go-stable-diffusion && \
@@ -275,6 +295,44 @@ sources/go-stable-diffusion:
 sources/go-stable-diffusion/libstablediffusion.a: sources/go-stable-diffusion
 	CPATH="$(CPATH):/usr/include/opencv4" $(MAKE) -C sources/go-stable-diffusion libstablediffusion.a
 
+## stablediffusion (ggml)
+sources/stablediffusion-ggml.cpp:
+	git clone --recursive $(STABLEDIFFUSION_GGML_REPO) sources/stablediffusion-ggml.cpp && \
+	cd sources/stablediffusion-ggml.cpp && \
+	git checkout $(STABLEDIFFUSION_GGML_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a: sources/stablediffusion-ggml.cpp
+	cd sources/stablediffusion-ggml.cpp && \
+	mkdir -p build && \
+	cd build && \
+	cmake $(CMAKE_ARGS) .. && \
+	cmake --build . --config Release
+
+backend/go/image/stablediffusion-ggml/libsd.a: sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a
+	$(MAKE) -C backend/go/image/stablediffusion-ggml libsd.a
+
+backend-assets/grpc/stablediffusion-ggml: backend/go/image/stablediffusion-ggml/libsd.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ LIBRARY_PATH=$(CURDIR)/backend/go/image/stablediffusion-ggml/ \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion-ggml ./backend/go/image/stablediffusion-ggml/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/stablediffusion-ggml
+endif
+
+sources/onnxruntime:
+	mkdir -p sources/onnxruntime
+	curl -L https://github.com/microsoft/onnxruntime/releases/download/v$(ONNX_VERSION)/onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz -o sources/onnxruntime/onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz
+	cd sources/onnxruntime && tar -xvf onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz && rm onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION).tgz
+	cd sources/onnxruntime && mv onnxruntime-$(ONNX_OS)-$(ONNX_ARCH)-$(ONNX_VERSION)/* ./
+
+backend-assets/lib/libonnxruntime.so.1: backend-assets/lib sources/onnxruntime
+	cp -rfv sources/onnxruntime/lib/* backend-assets/lib/
+ifeq ($(OS),Darwin)
+	mv backend-assets/lib/libonnxruntime.$(ONNX_VERSION).dylib backend-assets/lib/libonnxruntime.dylib
+else
+	mv backend-assets/lib/libonnxruntime.so.$(ONNX_VERSION) backend-assets/lib/libonnxruntime.so.1
+endif
+
 ## tiny-dream
 sources/go-tiny-dream:
 	mkdir -p sources/go-tiny-dream
@@ -301,23 +359,19 @@ sources/whisper.cpp:
 sources/whisper.cpp/libwhisper.a: sources/whisper.cpp
 	cd sources/whisper.cpp && $(MAKE) libwhisper.a libggml.a
 
-get-sources: sources/go-llama.cpp sources/go-piper sources/go-rwkv.cpp sources/whisper.cpp sources/go-bert.cpp sources/go-stable-diffusion sources/go-tiny-dream backend/cpp/llama/llama.cpp
+get-sources: sources/go-llama.cpp sources/go-piper sources/stablediffusion-ggml.cpp sources/bark.cpp sources/whisper.cpp sources/go-stable-diffusion sources/go-tiny-dream backend/cpp/llama/llama.cpp
 
 replace:
-	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(CURDIR)/sources/go-rwkv.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(CURDIR)/sources/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp/bindings/go=$(CURDIR)/sources/whisper.cpp/bindings/go
-	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(CURDIR)/sources/go-bert.cpp
 	$(GOCMD) mod edit -replace github.com/M0Rf30/go-tiny-dream=$(CURDIR)/sources/go-tiny-dream
 	$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(CURDIR)/sources/go-piper
 	$(GOCMD) mod edit -replace github.com/mudler/go-stable-diffusion=$(CURDIR)/sources/go-stable-diffusion
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(CURDIR)/sources/go-llama.cpp
 
 dropreplace:
-	$(GOCMD) mod edit -dropreplace github.com/donomii/go-rwkv.cpp
 	$(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/go-skynet/go-bert.cpp
 	$(GOCMD) mod edit -dropreplace github.com/M0Rf30/go-tiny-dream
 	$(GOCMD) mod edit -dropreplace github.com/mudler/go-piper
 	$(GOCMD) mod edit -dropreplace github.com/mudler/go-stable-diffusion
@@ -330,10 +384,8 @@ prepare-sources: get-sources replace
 rebuild: ## Rebuilds the project
 	$(GOCMD) clean -cache
 	$(MAKE) -C sources/go-llama.cpp clean
-	$(MAKE) -C sources/go-rwkv.cpp clean
 	$(MAKE) -C sources/whisper.cpp clean
 	$(MAKE) -C sources/go-stable-diffusion clean
-	$(MAKE) -C sources/go-bert.cpp clean
 	$(MAKE) -C sources/go-piper clean
 	$(MAKE) -C sources/go-tiny-dream clean
 	$(MAKE) build
@@ -348,7 +400,9 @@ clean: ## Remove build related file
 	rm -rf release/
 	rm -rf backend-assets/*
 	$(MAKE) -C backend/cpp/grpc clean
+	$(MAKE) -C backend/go/bark clean
 	$(MAKE) -C backend/cpp/llama clean
+	$(MAKE) -C backend/go/image/stablediffusion-ggml clean
 	rm -rf backend/cpp/llama-* || true
 	$(MAKE) dropreplace
 	$(MAKE) protogen-clean
@@ -439,8 +493,6 @@ 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
-	wget -q https://huggingface.co/mudler/rwkv-4-raven-1.5B-ggml/resolve/main/RWKV-4-Raven-1B5-v11-Eng99%2525-Other1%2525-20230425-ctx4096_Q4_0.bin -O test-models/rwkv
-	wget -q https://raw.githubusercontent.com/saharNooby/rwkv.cpp/5eb8f09c146ea8124633ab041d9ea0b1f1db4459/rwkv/20B_tokenizer.json -O test-models/rwkv.tokenizer.json
 	cp tests/models_fixtures/* test-models
 
 prepare-test: grpcs
@@ -693,13 +745,6 @@ backend-assets/espeak-ng-data: sources/go-piper sources/go-piper/libpiper_bindin
 backend-assets/grpc: protogen-go replace
 	mkdir -p backend-assets/grpc
 
-backend-assets/grpc/bert-embeddings: sources/go-bert.cpp sources/go-bert.cpp/libgobert.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-bert.cpp LIBRARY_PATH=$(CURDIR)/sources/go-bert.cpp \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bert-embeddings ./backend/go/llm/bert/
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/bert-embeddings
-endif
-
 backend-assets/grpc/huggingface: backend-assets/grpc
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/huggingface ./backend/go/llm/langchain/
 ifneq ($(UPX),)
@@ -759,10 +804,6 @@ backend-assets/grpc/llama-cpp-fallback: backend-assets/grpc backend/cpp/llama/ll
 	$(info ${GREEN}I llama-cpp build info:fallback${RESET})
 	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" $(MAKE) VARIANT="llama-fallback" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-fallback/grpc-server backend-assets/grpc/llama-cpp-fallback
-# TODO: every binary should have its own folder instead, so can have different metal implementations
-ifeq ($(BUILD_TYPE),metal)
-	cp backend/cpp/llama-fallback/llama.cpp/build/bin/default.metallib backend-assets/grpc/
-endif
 
 backend-assets/grpc/llama-cpp-cuda: backend-assets/grpc backend/cpp/llama/llama.cpp
 	cp -rf backend/cpp/llama backend/cpp/llama-cuda
@@ -775,7 +816,7 @@ backend-assets/grpc/llama-cpp-hipblas: backend-assets/grpc backend/cpp/llama/lla
 	cp -rf backend/cpp/llama backend/cpp/llama-hipblas
 	$(MAKE) -C backend/cpp/llama-hipblas purge
 	$(info ${GREEN}I llama-cpp build info:hipblas${RESET})
-	BUILD_TYPE="hipblas" $(MAKE) VARIANT="llama-hipblas" build-llama-cpp-grpc-server
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" BUILD_TYPE="hipblas" $(MAKE) VARIANT="llama-hipblas" build-llama-cpp-grpc-server
 	cp -rfv backend/cpp/llama-hipblas/grpc-server backend-assets/grpc/llama-cpp-hipblas
 
 backend-assets/grpc/llama-cpp-sycl_f16: backend-assets/grpc backend/cpp/llama/llama.cpp
@@ -810,6 +851,13 @@ 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 "$(GO_TAGS)" -o backend-assets/grpc/bark-cpp ./backend/go/bark/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/bark-cpp
+endif
+
 backend-assets/grpc/piper: sources/go-piper sources/go-piper/libpiper_binding.a backend-assets/grpc backend-assets/espeak-ng-data
 	CGO_CXXFLAGS="$(PIPER_CGO_CXXFLAGS)" CGO_LDFLAGS="$(PIPER_CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/sources/go-piper \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/piper ./backend/go/tts/
@@ -817,13 +865,6 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/piper
 endif
 
-backend-assets/grpc/rwkv: sources/go-rwkv.cpp sources/go-rwkv.cpp/librwkv.a backend-assets/grpc
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-rwkv.cpp LIBRARY_PATH=$(CURDIR)/sources/go-rwkv.cpp \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/rwkv ./backend/go/llm/rwkv
-ifneq ($(UPX),)
-	$(UPX) backend-assets/grpc/rwkv
-endif
-
 backend-assets/grpc/stablediffusion: sources/go-stable-diffusion sources/go-stable-diffusion/libstablediffusion.a backend-assets/grpc
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" CPATH="$(CPATH):$(CURDIR)/sources/go-stable-diffusion/:/usr/include/opencv4" LIBRARY_PATH=$(CURDIR)/sources/go-stable-diffusion/ \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion ./backend/go/image/stablediffusion
@@ -831,6 +872,13 @@ ifneq ($(UPX),)
 	$(UPX) backend-assets/grpc/stablediffusion
 endif
 
+backend-assets/grpc/silero-vad: backend-assets/grpc backend-assets/lib/libonnxruntime.so.1
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" CPATH="$(CPATH):$(CURDIR)/sources/onnxruntime/include/" LIBRARY_PATH=$(CURDIR)/backend-assets/lib \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/silero-vad ./backend/go/vad/silero
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/silero-vad
+endif
+
 backend-assets/grpc/tinydream: sources/go-tiny-dream sources/go-tiny-dream/libtinydream.a backend-assets/grpc
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/go-tiny-dream \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/tinydream ./backend/go/image/tinydream
@@ -891,7 +939,7 @@ docker-aio-all:
 
 docker-image-intel:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-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)" \
@@ -899,7 +947,7 @@ docker-image-intel:
 
 docker-image-intel-xpu:
 	docker build \
-		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-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)" \

README.md

@@ -38,6 +38,10 @@
 </a>
 </p>
 
+<p align="center">
+<a href="https://trendshift.io/repositories/1484" target="_blank"><img src="https://trendshift.io/api/badge/repositories/1484" alt="go-skynet%2FLocalAI | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</p>
+
 > :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) 
@@ -56,14 +60,17 @@ curl https://localai.io/install.sh | sh
 
 Or run with docker:
 ```bash
+# CPU only image:
+docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-cpu
+
+# 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):
+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/)
 docker run -ti --name local-ai -p 8080:8080 localai/localai:latest-aio-cpu
-# Alternative images:
-# - if you have an Nvidia GPU:
-# docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-aio-gpu-nvidia-cuda-12
-# - without preconfigured models
-# docker run -ti --name local-ai -p 8080:8080 localai/localai:latest
-# - without preconfigured models for Nvidia GPUs
-# docker run -ti --name local-ai -p 8080:8080 --gpus all localai/localai:latest-gpu-nvidia-cuda-12 
 ```
 
 To load models:
@@ -85,6 +92,9 @@ local-ai run oci://localai/phi-2:latest
 
 ## 📰 Latest project news
 
+- 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 )
+- Nov 2024: Voice activity detection models (**VAD**) added to the API: https://github.com/mudler/LocalAI/pull/4204
 - Oct 2024: examples moved to [LocalAI-examples](https://github.com/mudler/LocalAI-examples)
 - Aug 2024:  🆕 FLUX-1, [P2P Explorer](https://explorer.localai.io)
 - July 2024: 🔥🔥 🆕 P2P Dashboard, LocalAI Federated mode and AI Swarms: https://github.com/mudler/LocalAI/pull/2723
@@ -116,10 +126,10 @@ If you want to help and contribute, issues up for grabs: https://github.com/mudl
 
 ## 🚀 [Features](https://localai.io/features/)
 
-- 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `gpt4all.cpp`, ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
+- 📖 [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))
 - 🗣 [Text to Audio](https://localai.io/features/text-to-audio/)
 - 🔈 [Audio to Text](https://localai.io/features/audio-to-text/) (Audio transcription with `whisper.cpp`)
-- 🎨 [Image generation with stable diffusion](https://localai.io/features/image-generation)
+- 🎨 [Image generation](https://localai.io/features/image-generation)
 - 🔥 [OpenAI-alike tools API](https://localai.io/features/openai-functions/) 
 - 🧠 [Embeddings generation for vector databases](https://localai.io/features/embeddings/)
 - ✍️ [Constrained grammars](https://localai.io/features/constrained_grammars/)
@@ -127,6 +137,7 @@ If you want to help and contribute, issues up for grabs: https://github.com/mudl
 - 🥽 [Vision API](https://localai.io/features/gpt-vision/)
 - 📈 [Reranker API](https://localai.io/features/reranker/)
 - 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
+- 🔊 Voice activity detection (Silero-VAD support)
 - 🌍 Integrated WebUI!
 
 ## 💻 Usage
@@ -149,6 +160,7 @@ Model galleries
 Other:
 - Helm chart https://github.com/go-skynet/helm-charts
 - VSCode extension https://github.com/badgooooor/localai-vscode-plugin
+- Langchain: https://python.langchain.com/docs/integrations/providers/localai/
 - Terminal utility https://github.com/djcopley/ShellOracle
 - Local Smart assistant https://github.com/mudler/LocalAGI
 - Home Assistant https://github.com/sammcj/homeassistant-localai / https://github.com/drndos/hass-openai-custom-conversation / https://github.com/valentinfrlch/ha-gpt4vision
@@ -156,6 +168,9 @@ Other:
 - Slack bot https://github.com/mudler/LocalAGI/tree/main/examples/slack
 - Shell-Pilot(Interact with LLM using LocalAI models via pure shell scripts on your Linux or MacOS system) https://github.com/reid41/shell-pilot
 - Telegram bot https://github.com/mudler/LocalAI/tree/master/examples/telegram-bot
+- Another Telegram Bot https://github.com/JackBekket/Hellper
+- Auto-documentation https://github.com/JackBekket/Reflexia
+- Github bot which answer on issues, with code and documentation as context https://github.com/JackBekket/GitHelper
 - Github Actions: https://github.com/marketplace/actions/start-localai
 - Examples: https://github.com/mudler/LocalAI/tree/master/examples/
   
@@ -230,7 +245,6 @@ LocalAI couldn't have been built without the help of great software already avai
 - https://github.com/antimatter15/alpaca.cpp
 - https://github.com/EdVince/Stable-Diffusion-NCNN
 - https://github.com/ggerganov/whisper.cpp
-- https://github.com/saharNooby/rwkv.cpp
 - https://github.com/rhasspy/piper
 
 ## 🤗 Contributors

aio/cpu/embeddings.yaml

@@ -1,7 +1,7 @@
 name: text-embedding-ada-002
-backend: bert-embeddings
+embeddings: true
 parameters:
-  model: huggingface://mudler/all-MiniLM-L6-v2/ggml-model-q4_0.bin
+  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:

backend/backend.proto

@@ -28,6 +28,8 @@ service Backend {
   rpc Rerank(RerankRequest) returns (RerankResult) {}
 
   rpc GetMetrics(MetricsRequest) returns (MetricsResponse);
+
+  rpc VAD(VADRequest) returns (VADResponse) {}
 }
 
 // Define the empty request
@@ -238,6 +240,11 @@ message ModelOptions {
 
   repeated string LoraAdapters = 60;
   repeated float LoraScales = 61;
+
+  repeated string Options = 62;
+
+  string CacheTypeKey = 63;
+  string CacheTypeValue = 64;
 }
 
 message Result {
@@ -293,6 +300,19 @@ message TTSRequest {
   optional string language = 5;
 }
 
+message VADRequest {
+  repeated float audio = 1;
+}
+
+message VADSegment {
+  float start = 1;
+  float end = 2;
+}
+
+message VADResponse {
+  repeated VADSegment segments = 1;
+}
+
 message SoundGenerationRequest {
   string text = 1;
   string model = 2;

backend/cpp/llama/Makefile

@@ -22,7 +22,7 @@ else ifeq ($(BUILD_TYPE),clblas)
 	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
 # If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
 else ifeq ($(BUILD_TYPE),hipblas)
-	CMAKE_ARGS+=-DGGML_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 ($(OS),Darwin)
@@ -30,9 +30,7 @@ else ifeq ($(OS),Darwin)
 		CMAKE_ARGS+=-DGGML_METAL=OFF
 	else
 		CMAKE_ARGS+=-DGGML_METAL=ON
-# Until this is tested properly, we disable embedded metal file
-# as we already embed it as part of the LocalAI assets
-		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=OFF
+		CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
 		TARGET+=--target ggml-metal
 	endif
 endif

backend/cpp/llama/grpc-server.cpp

@@ -203,7 +203,7 @@ struct llama_client_slot
     std::string stopping_word;
 
     // sampling
-    struct common_sampler_params sparams;
+    struct common_params_sampling sparams;
     common_sampler *ctx_sampling = nullptr;
 
     int32_t ga_i = 0;   // group-attention state
@@ -492,8 +492,8 @@ struct llama_server_context
         }
 
         common_init_result common_init = common_init_from_params(params);
-        model = common_init.model;
-        ctx = common_init.context;
+        model = common_init.model.release();
+        ctx = common_init.context.release();
         if (model == nullptr)
         {
             LOG_ERR("unable to load model: %s", params.model.c_str());
@@ -662,7 +662,7 @@ struct llama_server_context
 
     bool launch_slot_with_data(llama_client_slot* &slot, json data) {
         slot_params default_params;
-        common_sampler_params default_sparams;
+        common_params_sampling default_sparams;
  
         slot->params.stream             = json_value(data, "stream",            false);
         slot->params.cache_prompt       = json_value(data, "cache_prompt",      false);
@@ -681,7 +681,6 @@ struct llama_server_context
         slot->sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
         slot->sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
         slot->sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-        slot->sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
         slot->params.n_keep             = json_value(data, "n_keep",            slot->params.n_keep);
         slot->sparams.seed               = json_value(data, "seed",              default_sparams.seed);
         slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
@@ -1213,13 +1212,12 @@ struct llama_server_context
             {"mirostat",          slot.sparams.mirostat},
             {"mirostat_tau",      slot.sparams.mirostat_tau},
             {"mirostat_eta",      slot.sparams.mirostat_eta},
-            {"penalize_nl",       slot.sparams.penalize_nl},
             {"stop",              slot.params.antiprompt},
             {"n_predict",         slot.params.n_predict},
             {"n_keep",            params.n_keep},
             {"ignore_eos",        slot.sparams.ignore_eos},
             {"stream",            slot.params.stream},
-      //      {"logit_bias",        slot.sparams.logit_bias},
+             //      {"logit_bias",        slot.sparams.logit_bias},
             {"n_probs",           slot.sparams.n_probs},
             {"min_keep",          slot.sparams.min_keep},
             {"grammar",           slot.sparams.grammar},
@@ -2112,7 +2110,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     //     slot->sparams.mirostat        = json_value(data, "mirostat",          default_sparams.mirostat);
     //     slot->sparams.mirostat_tau    = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
     //     slot->sparams.mirostat_eta    = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
-    //     slot->sparams.penalize_nl     = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
     //     slot->params.n_keep           = json_value(data, "n_keep",            slot->params.n_keep);
     //     slot->params.seed             = json_value(data, "seed",              default_params.seed);
     //     slot->sparams.grammar         = json_value(data, "grammar",           default_sparams.grammar);
@@ -2135,7 +2132,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     data["mirostat"] = predict->mirostat();
     data["mirostat_tau"] = predict->mirostattau();
     data["mirostat_eta"] = predict->mirostateta();
-    data["penalize_nl"] = predict->penalizenl();
     data["n_keep"] = predict->nkeep();
     data["seed"] = predict->seed();
     data["grammar"] = predict->grammar();
@@ -2181,7 +2177,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     llama.params.sparams.mirostat = predict->mirostat();
 //     llama.params.sparams.mirostat_tau = predict->mirostattau();
 //     llama.params.sparams.mirostat_eta = predict->mirostateta();
-//     llama.params.sparams.penalize_nl = predict->penalizenl();
 //     llama.params.n_keep = predict->nkeep();
 //     llama.params.seed = predict->seed();
 //     llama.params.sparams.grammar = predict->grammar();
@@ -2228,6 +2223,35 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     }
 // }
 
+const std::vector<ggml_type> kv_cache_types = {
+    GGML_TYPE_F32,
+    GGML_TYPE_F16,
+    GGML_TYPE_BF16,
+    GGML_TYPE_Q8_0,
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
+    GGML_TYPE_IQ4_NL,
+    GGML_TYPE_Q5_0,
+    GGML_TYPE_Q5_1,
+};
+
+static ggml_type kv_cache_type_from_str(const std::string & s) {
+    for (const auto & type : kv_cache_types) {
+        if (ggml_type_name(type) == s) {
+            return type;
+        }
+    }
+    throw std::runtime_error("Unsupported cache type: " + s);
+}
+
+static std::string get_all_kv_cache_types() {
+    std::ostringstream msg;
+    for (const auto & type : kv_cache_types) {
+        msg << ggml_type_name(type) << (&type == &kv_cache_types.back() ? "" : ", ");
+    }
+    return msg.str();
+}
+
 static void params_parse(const backend::ModelOptions* request,
                                 common_params & params) {
    
@@ -2241,6 +2265,12 @@ static void params_parse(const backend::ModelOptions* request,
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
+    if (!request->cachetypekey().empty()) {
+        params.cache_type_k = kv_cache_type_from_str(request->cachetypekey());
+    }
+    if (!request->cachetypevalue().empty()) {
+        params.cache_type_v = kv_cache_type_from_str(request->cachetypevalue());
+    }
     params.n_ctx = request->contextsize();
     //params.memory_f16 = request->f16memory();
     params.cpuparams.n_threads = request->threads();
@@ -2299,6 +2329,7 @@ static void params_parse(const backend::ModelOptions* request,
     params.use_mmap = request->mmap();
     params.flash_attn = request->flashattention();
     params.no_kv_offload = request->nokvoffload();
+    params.ctx_shift = false; // We control context-shifting in any case (and we disable it as it could just lead to infinite loops)
 
     params.embedding = request->embeddings();
 

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

@@ -1,13 +1,13 @@
 diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
-index 342042ff..224db9b5 100644
+index 3cd0d2fa..6c5e811a 100644
 --- a/examples/llava/clip.cpp
 +++ b/examples/llava/clip.cpp
-@@ -2419,7 +2419,7 @@ 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 + 1;
-+                patches_data[i] = i;
-             }
-             ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-             free(patches_data);
+@@ -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 + 1;
++                    patches_data[i] = i;
+                 }
+                 ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+                 free(patches_data);

backend/go/bark/Makefile

@@ -0,0 +1,25 @@
+INCLUDE_PATH := $(abspath ./)
+LIBRARY_PATH := $(abspath ./)
+
+AR?=ar
+
+BUILD_TYPE?=
+# keep standard at C11 and C++11
+CXXFLAGS = -I. -I$(INCLUDE_PATH)/../../../sources/bark.cpp/examples -I$(INCLUDE_PATH)/../../../sources/bark.cpp/spm-headers -I$(INCLUDE_PATH)/../../../sources/bark.cpp -O3 -DNDEBUG -std=c++17 -fPIC
+LDFLAGS  = -L$(LIBRARY_PATH) -L$(LIBRARY_PATH)/../../../sources/bark.cpp/build/examples -lbark -lstdc++ -lm
+
+# warnings
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+
+gobark.o:
+	$(CXX) $(CXXFLAGS) gobark.cpp -o gobark.o -c $(LDFLAGS)
+
+libbark.a: gobark.o
+	cp $(INCLUDE_PATH)/../../../sources/bark.cpp/build/libbark.a ./
+	$(AR) rcs libbark.a gobark.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml.c.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml-alloc.c.o
+	$(AR) rcs libbark.a $(LIBRARY_PATH)/../../../sources/bark.cpp/build/encodec.cpp/ggml/src/CMakeFiles/ggml.dir/ggml-backend.c.o
+
+clean:
+	rm -f gobark.o libbark.a

backend/go/bark/gobark.cpp

@@ -0,0 +1,85 @@
+#include <iostream>
+#include <tuple>
+
+#include "bark.h"
+#include "gobark.h"
+#include "common.h"
+#include "ggml.h"
+
+struct bark_context *c;
+
+void bark_print_progress_callback(struct bark_context *bctx, enum bark_encoding_step step, int progress, void *user_data) {
+    if (step == bark_encoding_step::SEMANTIC) {
+        printf("\rGenerating semantic tokens... %d%%", progress);
+    } else if (step == bark_encoding_step::COARSE) {
+        printf("\rGenerating coarse tokens... %d%%", progress);
+    } else if (step == bark_encoding_step::FINE) {
+        printf("\rGenerating fine tokens... %d%%", progress);
+    }
+    fflush(stdout);
+}
+
+int load_model(char *model) {
+    // initialize bark context
+    struct bark_context_params ctx_params = bark_context_default_params();
+    bark_params params;
+
+    params.model_path = model;
+
+   // ctx_params.verbosity = verbosity;
+    ctx_params.progress_callback = bark_print_progress_callback;
+    ctx_params.progress_callback_user_data = nullptr;
+
+    struct bark_context *bctx = bark_load_model(params.model_path.c_str(), ctx_params, params.seed);
+    if (!bctx) {
+        fprintf(stderr, "%s: Could not load model\n", __func__);
+        return 1;
+    }
+
+    c = bctx;
+
+    return 0;
+}
+
+int tts(char *text,int  threads, char *dst ) {
+
+    ggml_time_init();
+    const int64_t t_main_start_us = ggml_time_us();
+
+    // generate audio
+    if (!bark_generate_audio(c, text, threads)) {
+        fprintf(stderr, "%s: An error occured. If the problem persists, feel free to open an issue to report it.\n", __func__);
+        return 1;
+    }
+
+    const float *audio_data = bark_get_audio_data(c);
+    if (audio_data == NULL) {
+        fprintf(stderr, "%s: Could not get audio data\n", __func__);
+        return 1;
+    }
+
+    const int audio_arr_size = bark_get_audio_data_size(c);
+
+    std::vector<float> audio_arr(audio_data, audio_data + audio_arr_size);
+
+    write_wav_on_disk(audio_arr, dst);
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+        const int64_t t_load_us = bark_get_load_time(c);
+        const int64_t t_eval_us = bark_get_eval_time(c);
+
+        printf("\n\n");
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:     eval time = %8.2f ms\n", __func__, t_eval_us / 1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+    
+    return 0;
+}
+
+int unload() {
+    bark_free(c);
+}
+

backend/go/bark/gobark.go

@@ -0,0 +1,52 @@
+package main
+
+// #cgo CXXFLAGS: -I${SRCDIR}/../../../sources/bark.cpp/ -I${SRCDIR}/../../../sources/bark.cpp/encodec.cpp -I${SRCDIR}/../../../sources/bark.cpp/examples -I${SRCDIR}/../../../sources/bark.cpp/spm-headers
+// #cgo LDFLAGS: -L${SRCDIR}/ -L${SRCDIR}/../../../sources/bark.cpp/build/examples -L${SRCDIR}/../../../sources/bark.cpp/build/encodec.cpp/ -lbark -lencodec -lcommon
+// #include <gobark.h>
+// #include <stdlib.h>
+import "C"
+
+import (
+	"fmt"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+)
+
+type Bark struct {
+	base.SingleThread
+	threads int
+}
+
+func (sd *Bark) Load(opts *pb.ModelOptions) error {
+
+	sd.threads = int(opts.Threads)
+
+	modelFile := C.CString(opts.ModelFile)
+	defer C.free(unsafe.Pointer(modelFile))
+
+	ret := C.load_model(modelFile)
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}
+
+func (sd *Bark) TTS(opts *pb.TTSRequest) error {
+	t := C.CString(opts.Text)
+	defer C.free(unsafe.Pointer(t))
+
+	dst := C.CString(opts.Dst)
+	defer C.free(unsafe.Pointer(dst))
+
+	threads := C.int(sd.threads)
+
+	ret := C.tts(t, threads, dst)
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}

backend/go/bark/gobark.h

@@ -0,0 +1,8 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+int load_model(char *model);
+int tts(char *text,int  threads, char *dst );
+#ifdef __cplusplus
+}
+#endif

backend/go/bark/main.go

@@ -1,7 +1,6 @@
 package main
 
 // Note: this is started internally by LocalAI and a server is allocated for each model
-
 import (
 	"flag"
 
@@ -15,7 +14,7 @@ var (
 func main() {
 	flag.Parse()
 
-	if err := grpc.StartServer(*addr, &Embeddings{}); err != nil {
+	if err := grpc.StartServer(*addr, &Bark{}); err != nil {
 		panic(err)
 	}
 }

backend/go/image/stablediffusion-ggml/Makefile

@@ -0,0 +1,21 @@
+INCLUDE_PATH := $(abspath ./)
+LIBRARY_PATH := $(abspath ./)
+
+AR?=ar
+
+BUILD_TYPE?=
+# keep standard at C11 and C++11
+CXXFLAGS = -I. -I$(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
+
+# warnings
+CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function
+
+gosd.o:
+	$(CXX) $(CXXFLAGS) gosd.cpp -o gosd.o -c
+
+libsd.a: gosd.o
+	cp $(INCLUDE_PATH)/../../../../sources/stablediffusion-ggml.cpp/build/libstable-diffusion.a ./libsd.a
+	$(AR) rcs libsd.a gosd.o
+
+clean:
+	rm -f gosd.o libsd.a

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

@@ -0,0 +1,228 @@
+#include <stdio.h>
+#include <string.h>
+#include <time.h>
+#include <iostream>
+#include <random>
+#include <string>
+#include <vector>
+#include "gosd.h"
+
+// #include "preprocessing.hpp"
+#include "flux.hpp"
+#include "stable-diffusion.h"
+
+#define STB_IMAGE_IMPLEMENTATION
+#define STB_IMAGE_STATIC
+#include "stb_image.h"
+
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#define STB_IMAGE_WRITE_STATIC
+#include "stb_image_write.h"
+
+#define STB_IMAGE_RESIZE_IMPLEMENTATION
+#define STB_IMAGE_RESIZE_STATIC
+#include "stb_image_resize.h"
+
+// Names of the sampler method, same order as enum sample_method in stable-diffusion.h
+const char* sample_method_str[] = {
+    "euler_a",
+    "euler",
+    "heun",
+    "dpm2",
+    "dpm++2s_a",
+    "dpm++2m",
+    "dpm++2mv2",
+    "ipndm",
+    "ipndm_v",
+    "lcm",
+};
+
+// Names of the sigma schedule overrides, same order as sample_schedule in stable-diffusion.h
+const char* schedule_str[] = {
+    "default",
+    "discrete",
+    "karras",
+    "exponential",
+    "ays",
+    "gits",
+};
+
+sd_ctx_t* sd_c;
+
+sample_method_t sample_method;
+
+int load_model(char *model, char* options[], int threads, int diff) {
+    fprintf (stderr, "Loading model!\n");
+
+    char *stableDiffusionModel = "";
+    if (diff == 1 ) {
+        stableDiffusionModel = model;
+        model = "";
+    }
+
+    // decode options. Options are in form optname:optvale, or if booleans only optname.
+    char *clip_l_path  = "";
+    char *clip_g_path  = "";
+    char *t5xxl_path  = "";
+    char *vae_path  = "";
+    char *scheduler = "";
+    char *sampler = "";
+
+    // If options is not NULL, parse options
+    for (int i = 0; options[i] != NULL; i++) {
+        char *optname = strtok(options[i], ":");
+        char *optval = strtok(NULL, ":");
+        if (optval == NULL) {
+            optval = "true";
+        }
+
+        if (!strcmp(optname, "clip_l_path")) {
+            clip_l_path = optval;
+        }
+        if (!strcmp(optname, "clip_g_path")) {
+            clip_g_path = optval;
+        }
+        if (!strcmp(optname, "t5xxl_path")) {
+            t5xxl_path = optval;
+        }
+        if (!strcmp(optname, "vae_path")) {
+            vae_path = optval;
+        }
+        if (!strcmp(optname, "scheduler")) {
+            scheduler = optval;
+        }
+        if (!strcmp(optname, "sampler")) {
+            sampler = optval;
+        }
+    }
+
+    int sample_method_found = -1;
+    for (int m = 0; m < N_SAMPLE_METHODS; m++) {
+        if (!strcmp(sampler, sample_method_str[m])) {
+            sample_method_found = m;
+        }
+    }
+    if (sample_method_found == -1) {
+        fprintf(stderr, "Invalid sample method, default to EULER_A!\n");
+        sample_method_found = EULER_A;
+    }
+    sample_method = (sample_method_t)sample_method_found;
+
+    int schedule_found            = -1;
+    for (int d = 0; d < N_SCHEDULES; d++) {
+        if (!strcmp(scheduler, schedule_str[d])) {
+            schedule_found = d;
+                fprintf (stderr, "Found scheduler: %s\n", scheduler);
+
+        }
+    }
+
+    if (schedule_found == -1) {
+        fprintf (stderr, "Invalid scheduler! using DEFAULT\n");
+        schedule_found = DEFAULT;
+    }
+
+    schedule_t schedule = (schedule_t)schedule_found;
+    
+    fprintf (stderr, "Creating context\n");
+    sd_ctx_t* sd_ctx = new_sd_ctx(model,
+                                  clip_l_path,
+                                  clip_g_path,
+                                  t5xxl_path,
+                                  stableDiffusionModel,
+                                  vae_path,
+                                  "",
+                                  "",
+                                  "",
+                                  "",
+                                  "",
+                                  false,
+                                  false,
+                                  false,
+                                  threads,
+                                  SD_TYPE_COUNT,
+                                  STD_DEFAULT_RNG,
+                                  schedule,
+                                  false,
+                                  false,
+                                  false,
+                                  false);
+
+    if (sd_ctx == NULL) {
+        fprintf (stderr, "failed loading model (generic error)\n");
+        return 1;
+    }
+    fprintf (stderr, "Created context: OK\n");
+
+    sd_c = sd_ctx;
+
+    return 0;
+}
+
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed , char *dst, float cfg_scale) {
+
+    sd_image_t* results;
+
+    std::vector<int> skip_layers = {7, 8, 9};
+
+    fprintf (stderr, "Generating image\n");
+
+    results = txt2img(sd_c,
+                            text,
+                            negativeText,
+                            -1, //clip_skip
+                            cfg_scale, // sfg_scale
+                            3.5f,
+                            width,
+                            height,
+                            sample_method, 
+                            steps,
+                            seed,
+                            1,
+                            NULL,
+                            0.9f,
+                            20.f,
+                            false,
+                            "",
+                            skip_layers.data(),
+                            skip_layers.size(),
+                            0,
+                            0.01,
+                            0.2);
+
+    if (results == NULL) {
+        fprintf (stderr, "NO results\n");
+        return 1;
+    }
+
+    if (results[0].data == NULL) {
+        fprintf (stderr, "Results with no data\n");
+        return 1;
+    }
+
+    fprintf (stderr, "Writing PNG\n");
+
+    fprintf (stderr, "DST: %s\n", dst);
+    fprintf (stderr, "Width: %d\n", results[0].width);
+    fprintf (stderr, "Height: %d\n", results[0].height);
+    fprintf (stderr, "Channel: %d\n", results[0].channel);
+    fprintf (stderr, "Data: %p\n", results[0].data);
+
+    stbi_write_png(dst, results[0].width, results[0].height, results[0].channel,
+                       results[0].data, 0, NULL);
+    fprintf (stderr, "Saved resulting image to '%s'\n", dst);
+
+    // TODO: free results. Why does it crash?
+
+    free(results[0].data);
+    results[0].data = NULL;
+    free(results);
+    fprintf (stderr, "gen_image is done", dst);
+
+    return 0;
+}
+
+int unload() {
+    free_sd_ctx(sd_c);
+}
+

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

@@ -0,0 +1,96 @@
+package main
+
+// #cgo CXXFLAGS: -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/thirdparty -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp -I${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/ggml/include
+// #cgo LDFLAGS: -L${SRCDIR}/ -L${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/build/ggml/src/ggml-cpu -L${SRCDIR}/../../../../sources/stablediffusion-ggml.cpp/build/ggml/src -lsd -lstdc++ -lm -lggml -lggml-base -lggml-cpu -lgomp
+// #include <gosd.h>
+// #include <stdlib.h>
+import "C"
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"unsafe"
+
+	"github.com/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/LocalAI/pkg/utils"
+)
+
+type SDGGML struct {
+	base.SingleThread
+	threads      int
+	sampleMethod string
+	cfgScale     float32
+}
+
+func (sd *SDGGML) Load(opts *pb.ModelOptions) error {
+
+	sd.threads = int(opts.Threads)
+
+	modelFile := C.CString(opts.ModelFile)
+	defer C.free(unsafe.Pointer(modelFile))
+
+	var options **C.char
+	// prepare the options array to pass to C
+
+	size := C.size_t(unsafe.Sizeof((*C.char)(nil)))
+	length := C.size_t(len(opts.Options))
+	options = (**C.char)(C.malloc(length * size))
+	view := (*[1 << 30]*C.char)(unsafe.Pointer(options))[0:len(opts.Options):len(opts.Options)]
+
+	var diffusionModel int
+
+	var oo []string
+	for _, op := range opts.Options {
+		if op == "diffusion_model" {
+			diffusionModel = 1
+			continue
+		}
+
+		// If it's an option path, we resolve absolute path from the model path
+		if strings.Contains(op, ":") && strings.Contains(op, "path") {
+			data := strings.Split(op, ":")
+			data[1] = filepath.Join(opts.ModelPath, data[1])
+			if err := utils.VerifyPath(data[1], opts.ModelPath); err == nil {
+				oo = append(oo, strings.Join(data, ":"))
+			}
+		} else {
+			oo = append(oo, op)
+		}
+	}
+
+	fmt.Fprintf(os.Stderr, "Options: %+v\n", oo)
+
+	for i, x := range oo {
+		view[i] = C.CString(x)
+	}
+
+	sd.cfgScale = opts.CFGScale
+
+	ret := C.load_model(modelFile, options, C.int(opts.Threads), C.int(diffusionModel))
+	if ret != 0 {
+		return fmt.Errorf("could not load model")
+	}
+
+	return nil
+}
+
+func (sd *SDGGML) GenerateImage(opts *pb.GenerateImageRequest) error {
+	t := C.CString(opts.PositivePrompt)
+	defer C.free(unsafe.Pointer(t))
+
+	dst := C.CString(opts.Dst)
+	defer C.free(unsafe.Pointer(dst))
+
+	negative := C.CString(opts.NegativePrompt)
+	defer C.free(unsafe.Pointer(negative))
+
+	ret := C.gen_image(t, negative, C.int(opts.Width), C.int(opts.Height), C.int(opts.Step), C.int(opts.Seed), dst, C.float(sd.cfgScale))
+	if ret != 0 {
+		return fmt.Errorf("inference failed")
+	}
+
+	return nil
+}

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

@@ -0,0 +1,8 @@
+#ifdef __cplusplus
+extern "C" {
+#endif
+int load_model(char *model, char* options[], int threads, int diffusionModel);
+int gen_image(char *text, char *negativeText, int width, int height, int steps, int seed, char *dst, float cfg_scale);
+#ifdef __cplusplus
+}
+#endif

backend/go/image/stablediffusion-ggml/main.go

@@ -0,0 +1,20 @@
+package main
+
+// Note: this is started internally by LocalAI and a server is allocated for each model
+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, &SDGGML{}); err != nil {
+		panic(err)
+	}
+}

backend/go/llm/bert/bert.go

@@ -1,34 +0,0 @@
-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 (
-	bert "github.com/go-skynet/go-bert.cpp"
-
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-type Embeddings struct {
-	base.SingleThread
-	bert *bert.Bert
-}
-
-func (llm *Embeddings) Load(opts *pb.ModelOptions) error {
-	model, err := bert.New(opts.ModelFile)
-	llm.bert = model
-	return err
-}
-
-func (llm *Embeddings) Embeddings(opts *pb.PredictOptions) ([]float32, error) {
-
-	if len(opts.EmbeddingTokens) > 0 {
-		tokens := []int{}
-		for _, t := range opts.EmbeddingTokens {
-			tokens = append(tokens, int(t))
-		}
-		return llm.bert.TokenEmbeddings(tokens, bert.SetThreads(int(opts.Threads)))
-	}
-
-	return llm.bert.Embeddings(opts.Embeddings, bert.SetThreads(int(opts.Threads)))
-}

backend/go/llm/rwkv/rwkv.go

@@ -1,95 +0,0 @@
-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"
-	"path/filepath"
-
-	"github.com/donomii/go-rwkv.cpp"
-	"github.com/mudler/LocalAI/pkg/grpc/base"
-	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
-)
-
-const tokenizerSuffix = ".tokenizer.json"
-
-type LLM struct {
-	base.SingleThread
-
-	rwkv *rwkv.RwkvState
-}
-
-func (llm *LLM) Load(opts *pb.ModelOptions) error {
-	tokenizerFile := opts.Tokenizer
-	if tokenizerFile == "" {
-		modelFile := filepath.Base(opts.ModelFile)
-		tokenizerFile = modelFile + tokenizerSuffix
-	}
-	modelPath := filepath.Dir(opts.ModelFile)
-	tokenizerPath := filepath.Join(modelPath, tokenizerFile)
-
-	model := rwkv.LoadFiles(opts.ModelFile, tokenizerPath, uint32(opts.GetThreads()))
-
-	if model == nil {
-		return fmt.Errorf("rwkv could not load model")
-	}
-	llm.rwkv = model
-	return nil
-}
-
-func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
-	stopWord := "\n"
-	if len(opts.StopPrompts) > 0 {
-		stopWord = opts.StopPrompts[0]
-	}
-
-	if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
-		return "", err
-	}
-
-	response := llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), nil)
-
-	return response, nil
-}
-
-func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
-	go func() {
-
-		stopWord := "\n"
-		if len(opts.StopPrompts) > 0 {
-			stopWord = opts.StopPrompts[0]
-		}
-
-		if err := llm.rwkv.ProcessInput(opts.Prompt); err != nil {
-			fmt.Println("Error processing input: ", err)
-			return
-		}
-
-		llm.rwkv.GenerateResponse(int(opts.Tokens), stopWord, float32(opts.Temperature), float32(opts.TopP), func(s string) bool {
-			results <- s
-			return true
-		})
-		close(results)
-	}()
-
-	return nil
-}
-
-func (llm *LLM) TokenizeString(opts *pb.PredictOptions) (pb.TokenizationResponse, error) {
-	tokens, err := llm.rwkv.Tokenizer.Encode(opts.Prompt)
-	if err != nil {
-		return pb.TokenizationResponse{}, err
-	}
-
-	l := len(tokens)
-	i32Tokens := make([]int32, l)
-
-	for i, t := range tokens {
-		i32Tokens[i] = int32(t.ID)
-	}
-
-	return pb.TokenizationResponse{
-		Length: int32(l),
-		Tokens: i32Tokens,
-	}, nil
-}

backend/go/vad/silero/main.go

@@ -15,7 +15,7 @@ var (
 func main() {
 	flag.Parse()
 
-	if err := grpc.StartServer(*addr, &LLM{}); err != nil {
+	if err := grpc.StartServer(*addr, &VAD{}); err != nil {
 		panic(err)
 	}
 }

backend/go/vad/silero/vad.go

@@ -0,0 +1,54 @@
+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/mudler/LocalAI/pkg/grpc/base"
+	pb "github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/streamer45/silero-vad-go/speech"
+)
+
+type VAD struct {
+	base.SingleThread
+	detector *speech.Detector
+}
+
+func (vad *VAD) Load(opts *pb.ModelOptions) error {
+	v, err := speech.NewDetector(speech.DetectorConfig{
+		ModelPath:  opts.ModelFile,
+		SampleRate: 16000,
+		//WindowSize:           1024,
+		Threshold:            0.5,
+		MinSilenceDurationMs: 0,
+		SpeechPadMs:          0,
+	})
+	if err != nil {
+		return fmt.Errorf("create silero detector: %w", err)
+	}
+
+	vad.detector = v
+	return err
+}
+
+func (vad *VAD) VAD(req *pb.VADRequest) (pb.VADResponse, error) {
+	audio := req.Audio
+
+	segments, err := vad.detector.Detect(audio)
+	if err != nil {
+		return pb.VADResponse{}, fmt.Errorf("detect: %w", err)
+	}
+
+	vadSegments := []*pb.VADSegment{}
+	for _, s := range segments {
+		vadSegments = append(vadSegments, &pb.VADSegment{
+			Start: float32(s.SpeechStartAt),
+			End:   float32(s.SpeechEndAt),
+		})
+	}
+
+	return pb.VADResponse{
+		Segments: vadSegments,
+	}, nil
+}

backend/python/autogptq/requirements-intel.txt

@@ -1,5 +1,6 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools

backend/python/autogptq/requirements.txt

@@ -1,6 +1,6 @@
 accelerate
 auto-gptq==0.7.1
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
 transformers

backend/python/bark/requirements-intel.txt

@@ -1,8 +1,9 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchaudio
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 transformers
 accelerate

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi

backend/python/common/libbackend.sh

@@ -17,6 +17,9 @@
 # LIMIT_TARGETS="cublas12"
 # source $(dirname $0)/../common/libbackend.sh
 #
+
+PYTHON_VERSION="3.10"
+
 function init() {
     # Name of the backend (directory name)
     BACKEND_NAME=${PWD##*/}
@@ -88,7 +91,7 @@ function getBuildProfile() {
 # always result in an activated virtual environment
 function ensureVenv() {
     if [ ! -d "${EDIR}/venv" ]; then
-        uv venv ${EDIR}/venv
+        uv venv --python ${PYTHON_VERSION} ${EDIR}/venv
         echo "virtualenv created"
     fi
 

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

@@ -1,4 +1,5 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]

backend/python/common/template/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 grpcio-tools

backend/python/coqui/requirements-intel.txt

@@ -1,9 +1,10 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchaudio
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 transformers
 accelerate
 coqui-tts

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/requirements-intel.txt

@@ -1,9 +1,10 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchvision
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchvision==0.18.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 diffusers
 opencv-python
 transformers

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.67.1
+grpcio==1.69.0
 pillow
 protobuf
 certifi

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
 wheel

backend/python/mamba/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi

backend/python/openvoice/requirements-cpu.txt

@@ -1,3 +1,7 @@
 torch==2.4.1
 git+https://github.com/myshell-ai/MeloTTS.git
-git+https://github.com/myshell-ai/OpenVoice.git
+git+https://github.com/myshell-ai/OpenVoice.git
+whisper-timestamped
+pydub==0.25.1
+wavmark==0.0.3
+eng_to_ipa==0.0.2

backend/python/openvoice/requirements-cublas11.txt

@@ -1,4 +1,8 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.4.1+cu118
 git+https://github.com/myshell-ai/MeloTTS.git
-git+https://github.com/myshell-ai/OpenVoice.git
+git+https://github.com/myshell-ai/OpenVoice.git
+whisper-timestamped
+pydub==0.25.1
+wavmark==0.0.3
+eng_to_ipa==0.0.2

backend/python/openvoice/requirements-cublas12.txt

@@ -1,3 +1,7 @@
 torch==2.4.1
 git+https://github.com/myshell-ai/MeloTTS.git
-git+https://github.com/myshell-ai/OpenVoice.git
+git+https://github.com/myshell-ai/OpenVoice.git
+whisper-timestamped
+pydub==0.25.1
+wavmark==0.0.3
+eng_to_ipa==0.0.2

backend/python/openvoice/requirements-hipblas.txt

@@ -1,4 +1,8 @@
 --extra-index-url https://download.pytorch.org/whl/rocm6.0
 torch==2.4.1+rocm6.0
 git+https://github.com/myshell-ai/MeloTTS.git
-git+https://github.com/myshell-ai/OpenVoice.git
+git+https://github.com/myshell-ai/OpenVoice.git
+whisper-timestamped
+pydub==0.25.1
+wavmark==0.0.3
+eng_to_ipa==0.0.2

backend/python/openvoice/requirements-intel.txt

@@ -1,14 +1,15 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 librosa==0.9.1
 faster-whisper==0.9.0
 pydub==0.25.1
 wavmark==0.0.3
-numpy==1.22.0
 eng_to_ipa==0.0.2
 inflect==7.0.0
 unidecode==1.3.7

backend/python/openvoice/requirements.txt

@@ -1,20 +1,17 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 librosa
 faster-whisper
-pydub==0.25.1
-wavmark==0.0.3
-numpy==1.22.0
-eng_to_ipa==0.0.2
 inflect
 unidecode
-whisper-timestamped
 openai
 python-dotenv
 pypinyin
 cn2an==0.5.22
+numpy==1.22.0
 networkx==2.8.8
 jieba==0.42.1
-gradio==3.48.0
+gradio==5.9.1
 langid==1.1.6
 llvmlite==0.43.0
+setuptools

backend/python/parler-tts/requirements-intel.txt

@@ -1,8 +1,8 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
-torchaudio
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
 transformers
 accelerate

backend/python/parler-tts/requirements.txt

@@ -1,3 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 certifi
 llvmlite==0.43.0
+setuptools

backend/python/rerankers/requirements-intel.txt

@@ -1,8 +1,9 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
+intel-extension-for-pytorch==2.3.110+xpu
 transformers
 accelerate
-torch
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 rerankers[transformers]
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools

backend/python/rerankers/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi

backend/python/sentencetransformers/requirements-cpu.txt

@@ -2,5 +2,5 @@ torch==2.4.1
 accelerate
 transformers
 bitsandbytes
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements-cublas11.txt

@@ -1,5 +1,5 @@
 --extra-index-url https://download.pytorch.org/whl/cu118
 torch==2.4.1+cu118
 accelerate
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements-cublas12.txt

@@ -1,4 +1,4 @@
 torch==2.4.1
 accelerate
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements-hipblas.txt

@@ -1,5 +1,5 @@
 --extra-index-url https://download.pytorch.org/whl/rocm6.0
 torch==2.4.1+rocm6.0
 accelerate
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements-intel.txt

@@ -1,8 +1,9 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
-setuptools==69.5.1 # https://github.com/mudler/LocalAI/issues/2406
+setuptools
 accelerate
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
 datasets

backend/python/transformers-musicgen/requirements-intel.txt

@@ -1,7 +1,8 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
+intel-extension-for-pytorch==2.3.110+xpu
 transformers
+oneccl_bind_pt==2.3.100+xpu
 accelerate
-torch
+torch==2.3.1+cxx11.abi
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+setuptools

backend/python/transformers-musicgen/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 scipy==1.14.0
 certifi

backend/python/transformers/requirements-intel.txt

@@ -1,6 +1,7 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
-torch
+intel-extension-for-pytorch==2.3.110+xpu
+torch==2.3.1+cxx11.abi
+oneccl_bind_pt==2.3.100+xpu
 optimum[openvino]
 intel-extension-for-transformers
 bitsandbytes

backend/python/transformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
-setuptools==69.5.1 # https://github.com/mudler/LocalAI/issues/2406
+setuptools

backend/python/vall-e-x/requirements-intel.txt

@@ -1,7 +1,7 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
+intel-extension-for-pytorch==2.3.110+xpu
 accelerate
-torch
-torchaudio
+torch==2.3.1+cxx11.abi
+torchaudio==2.3.1+cxx11.abi
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
+oneccl_bind_pt==2.3.100+xpu

backend/python/vall-e-x/requirements.txt

@@ -1,3 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
-certifi
+certifi
+setuptools

backend/python/vllm/install.sh

@@ -22,7 +22,7 @@ if [ "x${BUILD_TYPE}" == "x" ] && [ "x${FROM_SOURCE}" == "xtrue" ]; then
             git clone https://github.com/vllm-project/vllm
         fi
         pushd vllm
-            uv pip install wheel packaging ninja "setuptools>=49.4.0" numpy typing-extensions pillow setuptools-scm grpcio==1.67.1 protobuf bitsandbytes
+            uv pip install wheel packaging ninja "setuptools>=49.4.0" numpy typing-extensions pillow setuptools-scm grpcio==1.68.1 protobuf bitsandbytes
             uv pip install -v -r requirements-cpu.txt --extra-index-url https://download.pytorch.org/whl/cpu
             VLLM_TARGET_DEVICE=cpu python setup.py install
         popd

backend/python/vllm/requirements-intel.txt

@@ -1,8 +1,9 @@
 --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
-intel-extension-for-pytorch
+intel-extension-for-pytorch==2.3.110+xpu
 accelerate
-torch
+torch==2.3.1+cxx11.abi
 transformers
 optimum[openvino]
-setuptools==75.1.0 # https://github.com/mudler/LocalAI/issues/2406
-bitsandbytes
+setuptools
+bitsandbytes
+oneccl_bind_pt==2.3.100+xpu

backend/python/vllm/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.1
+grpcio==1.69.0
 protobuf
 certifi
 setuptools

core/application.go

@@ -1,38 +0,0 @@
-package core
-
-import (
-	"github.com/mudler/LocalAI/core/config"
-	"github.com/mudler/LocalAI/core/services"
-	"github.com/mudler/LocalAI/pkg/model"
-)
-
-// The purpose of this structure is to hold pointers to all initialized services, to make plumbing easy
-// Perhaps a proper DI system is worth it in the future, but for now keep things simple.
-type Application struct {
-
-	// Application-Level Config
-	ApplicationConfig *config.ApplicationConfig
-	// ApplicationState *ApplicationState
-
-	// Core Low-Level Services
-	BackendConfigLoader *config.BackendConfigLoader
-	ModelLoader         *model.ModelLoader
-
-	// Backend Services
-	// EmbeddingsBackendService      *backend.EmbeddingsBackendService
-	// ImageGenerationBackendService *backend.ImageGenerationBackendService
-	// LLMBackendService             *backend.LLMBackendService
-	// TranscriptionBackendService *backend.TranscriptionBackendService
-	// TextToSpeechBackendService  *backend.TextToSpeechBackendService
-
-	// LocalAI System Services
-	BackendMonitorService *services.BackendMonitorService
-	GalleryService        *services.GalleryService
-	LocalAIMetricsService *services.LocalAIMetricsService
-	// OpenAIService         *services.OpenAIService
-}
-
-// TODO [NEXT PR?]: Break up ApplicationConfig.
-// Migrate over stuff that is not set via config at all - especially runtime stuff
-type ApplicationState struct {
-}

core/application/application.go

@@ -0,0 +1,39 @@
+package application
+
+import (
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/templates"
+)
+
+type Application struct {
+	backendLoader      *config.BackendConfigLoader
+	modelLoader        *model.ModelLoader
+	applicationConfig  *config.ApplicationConfig
+	templatesEvaluator *templates.Evaluator
+}
+
+func newApplication(appConfig *config.ApplicationConfig) *Application {
+	return &Application{
+		backendLoader:      config.NewBackendConfigLoader(appConfig.ModelPath),
+		modelLoader:        model.NewModelLoader(appConfig.ModelPath),
+		applicationConfig:  appConfig,
+		templatesEvaluator: templates.NewEvaluator(appConfig.ModelPath),
+	}
+}
+
+func (a *Application) BackendLoader() *config.BackendConfigLoader {
+	return a.backendLoader
+}
+
+func (a *Application) ModelLoader() *model.ModelLoader {
+	return a.modelLoader
+}
+
+func (a *Application) ApplicationConfig() *config.ApplicationConfig {
+	return a.applicationConfig
+}
+
+func (a *Application) TemplatesEvaluator() *templates.Evaluator {
+	return a.templatesEvaluator
+}

core/application/config_file_watcher.go

@@ -1,4 +1,4 @@
-package startup
+package application
 
 import (
 	"encoding/json"
@@ -8,8 +8,8 @@ import (
 	"path/filepath"
 	"time"
 
-	"github.com/fsnotify/fsnotify"
 	"dario.cat/mergo"
+	"github.com/fsnotify/fsnotify"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/rs/zerolog/log"
 )

core/application/startup.go

@@ -1,15 +1,15 @@
-package startup
+package application
 
 import (
 	"fmt"
 	"os"
 
-	"github.com/mudler/LocalAI/core"
 	"github.com/mudler/LocalAI/core/backend"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/services"
 	"github.com/mudler/LocalAI/internal"
 	"github.com/mudler/LocalAI/pkg/assets"
+
 	"github.com/mudler/LocalAI/pkg/library"
 	"github.com/mudler/LocalAI/pkg/model"
 	pkgStartup "github.com/mudler/LocalAI/pkg/startup"
@@ -17,8 +17,9 @@ import (
 	"github.com/rs/zerolog/log"
 )
 
-func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.ModelLoader, *config.ApplicationConfig, error) {
+func New(opts ...config.AppOption) (*Application, error) {
 	options := config.NewApplicationConfig(opts...)
+	application := newApplication(options)
 
 	log.Info().Msgf("Starting LocalAI using %d threads, with models path: %s", options.Threads, options.ModelPath)
 	log.Info().Msgf("LocalAI version: %s", internal.PrintableVersion())
@@ -36,28 +37,28 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 
 	// Make sure directories exists
 	if options.ModelPath == "" {
-		return nil, nil, nil, fmt.Errorf("options.ModelPath cannot be empty")
+		return nil, fmt.Errorf("options.ModelPath cannot be empty")
 	}
 	err = os.MkdirAll(options.ModelPath, 0750)
 	if err != nil {
-		return nil, nil, nil, fmt.Errorf("unable to create ModelPath: %q", err)
+		return nil, fmt.Errorf("unable to create ModelPath: %q", err)
 	}
 	if options.ImageDir != "" {
 		err := os.MkdirAll(options.ImageDir, 0750)
 		if err != nil {
-			return nil, nil, nil, fmt.Errorf("unable to create ImageDir: %q", err)
+			return nil, fmt.Errorf("unable to create ImageDir: %q", err)
 		}
 	}
 	if options.AudioDir != "" {
 		err := os.MkdirAll(options.AudioDir, 0750)
 		if err != nil {
-			return nil, nil, nil, fmt.Errorf("unable to create AudioDir: %q", err)
+			return nil, fmt.Errorf("unable to create AudioDir: %q", err)
 		}
 	}
 	if options.UploadDir != "" {
 		err := os.MkdirAll(options.UploadDir, 0750)
 		if err != nil {
-			return nil, nil, nil, fmt.Errorf("unable to create UploadDir: %q", err)
+			return nil, fmt.Errorf("unable to create UploadDir: %q", err)
 		}
 	}
 
@@ -65,39 +66,36 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 		log.Error().Err(err).Msg("error installing models")
 	}
 
-	cl := config.NewBackendConfigLoader(options.ModelPath)
-	ml := model.NewModelLoader(options.ModelPath)
-
 	configLoaderOpts := options.ToConfigLoaderOptions()
 
-	if err := cl.LoadBackendConfigsFromPath(options.ModelPath, configLoaderOpts...); err != nil {
+	if err := application.BackendLoader().LoadBackendConfigsFromPath(options.ModelPath, configLoaderOpts...); err != nil {
 		log.Error().Err(err).Msg("error loading config files")
 	}
 
 	if options.ConfigFile != "" {
-		if err := cl.LoadMultipleBackendConfigsSingleFile(options.ConfigFile, configLoaderOpts...); err != nil {
+		if err := application.BackendLoader().LoadMultipleBackendConfigsSingleFile(options.ConfigFile, configLoaderOpts...); err != nil {
 			log.Error().Err(err).Msg("error loading config file")
 		}
 	}
 
-	if err := cl.Preload(options.ModelPath); err != nil {
+	if err := application.BackendLoader().Preload(options.ModelPath); err != nil {
 		log.Error().Err(err).Msg("error downloading models")
 	}
 
 	if options.PreloadJSONModels != "" {
 		if err := services.ApplyGalleryFromString(options.ModelPath, options.PreloadJSONModels, options.EnforcePredownloadScans, options.Galleries); err != nil {
-			return nil, nil, nil, err
+			return nil, err
 		}
 	}
 
 	if options.PreloadModelsFromPath != "" {
 		if err := services.ApplyGalleryFromFile(options.ModelPath, options.PreloadModelsFromPath, options.EnforcePredownloadScans, options.Galleries); err != nil {
-			return nil, nil, nil, err
+			return nil, err
 		}
 	}
 
 	if options.Debug {
-		for _, v := range cl.GetAllBackendConfigs() {
+		for _, v := range application.BackendLoader().GetAllBackendConfigs() {
 			log.Debug().Msgf("Model: %s (config: %+v)", v.Name, v)
 		}
 	}
@@ -123,7 +121,7 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 	go func() {
 		<-options.Context.Done()
 		log.Debug().Msgf("Context canceled, shutting down")
-		err := ml.StopAllGRPC()
+		err := application.ModelLoader().StopAllGRPC()
 		if err != nil {
 			log.Error().Err(err).Msg("error while stopping all grpc backends")
 		}
@@ -131,12 +129,12 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 
 	if options.WatchDog {
 		wd := model.NewWatchDog(
-			ml,
+			application.ModelLoader(),
 			options.WatchDogBusyTimeout,
 			options.WatchDogIdleTimeout,
 			options.WatchDogBusy,
 			options.WatchDogIdle)
-		ml.SetWatchDog(wd)
+		application.ModelLoader().SetWatchDog(wd)
 		go wd.Run()
 		go func() {
 			<-options.Context.Done()
@@ -147,7 +145,7 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 
 	if options.LoadToMemory != nil {
 		for _, m := range options.LoadToMemory {
-			cfg, err := cl.LoadBackendConfigFileByName(m, options.ModelPath,
+			cfg, err := application.BackendLoader().LoadBackendConfigFileByName(m, options.ModelPath,
 				config.LoadOptionDebug(options.Debug),
 				config.LoadOptionThreads(options.Threads),
 				config.LoadOptionContextSize(options.ContextSize),
@@ -155,7 +153,7 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 				config.ModelPath(options.ModelPath),
 			)
 			if err != nil {
-				return nil, nil, nil, err
+				return nil, err
 			}
 
 			log.Debug().Msgf("Auto loading model %s into memory from file: %s", m, cfg.Model)
@@ -163,9 +161,9 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 			o := backend.ModelOptions(*cfg, options)
 
 			var backendErr error
-			_, backendErr = ml.Load(o...)
+			_, backendErr = application.ModelLoader().Load(o...)
 			if backendErr != nil {
-				return nil, nil, nil, err
+				return nil, err
 			}
 		}
 	}
@@ -174,7 +172,7 @@ func Startup(opts ...config.AppOption) (*config.BackendConfigLoader, *model.Mode
 	startWatcher(options)
 
 	log.Info().Msg("core/startup process completed!")
-	return cl, ml, options, nil
+	return application, nil
 }
 
 func startWatcher(options *config.ApplicationConfig) {
@@ -201,32 +199,3 @@ func startWatcher(options *config.ApplicationConfig) {
 		log.Error().Err(err).Msg("failed creating watcher")
 	}
 }
-
-// In Lieu of a proper DI framework, this function wires up the Application manually.
-// This is in core/startup rather than core/state.go to keep package references clean!
-func createApplication(appConfig *config.ApplicationConfig) *core.Application {
-	app := &core.Application{
-		ApplicationConfig:   appConfig,
-		BackendConfigLoader: config.NewBackendConfigLoader(appConfig.ModelPath),
-		ModelLoader:         model.NewModelLoader(appConfig.ModelPath),
-	}
-
-	var err error
-
-	// app.EmbeddingsBackendService = backend.NewEmbeddingsBackendService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-	// app.ImageGenerationBackendService = backend.NewImageGenerationBackendService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-	// app.LLMBackendService = backend.NewLLMBackendService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-	// app.TranscriptionBackendService = backend.NewTranscriptionBackendService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-	// app.TextToSpeechBackendService = backend.NewTextToSpeechBackendService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-
-	app.BackendMonitorService = services.NewBackendMonitorService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig)
-	app.GalleryService = services.NewGalleryService(app.ApplicationConfig)
-	// app.OpenAIService = services.NewOpenAIService(app.ModelLoader, app.BackendConfigLoader, app.ApplicationConfig, app.LLMBackendService)
-
-	app.LocalAIMetricsService, err = services.NewLocalAIMetricsService()
-	if err != nil {
-		log.Error().Err(err).Msg("encountered an error initializing metrics service, startup will continue but metrics will not be tracked.")
-	}
-
-	return app
-}

core/backend/llm.go

@@ -117,8 +117,12 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 			ss := ""
 
 			var partialRune []byte
-			err := inferenceModel.PredictStream(ctx, opts, func(chars []byte) {
-				partialRune = append(partialRune, chars...)
+			err := inferenceModel.PredictStream(ctx, opts, func(reply *proto.Reply) {
+				msg := reply.Message
+				partialRune = append(partialRune, msg...)
+
+				tokenUsage.Prompt = int(reply.PromptTokens)
+				tokenUsage.Completion = int(reply.Tokens)
 
 				for len(partialRune) > 0 {
 					r, size := utf8.DecodeRune(partialRune)
@@ -132,6 +136,10 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 
 					partialRune = partialRune[size:]
 				}
+
+				if len(msg) == 0 {
+					tokenCallback("", tokenUsage)
+				}
 			})
 			return LLMResponse{
 				Response: ss,

core/backend/options.go

@@ -122,7 +122,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		CUDA:                 c.CUDA || c.Diffusers.CUDA,
 		SchedulerType:        c.Diffusers.SchedulerType,
 		PipelineType:         c.Diffusers.PipelineType,
-		CFGScale:             c.Diffusers.CFGScale,
+		CFGScale:             c.CFGScale,
 		LoraAdapter:          c.LoraAdapter,
 		LoraScale:            c.LoraScale,
 		LoraAdapters:         c.LoraAdapters,
@@ -132,6 +132,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		IMG2IMG:              c.Diffusers.IMG2IMG,
 		CLIPModel:            c.Diffusers.ClipModel,
 		CLIPSubfolder:        c.Diffusers.ClipSubFolder,
+		Options:              c.Options,
 		CLIPSkip:             int32(c.Diffusers.ClipSkip),
 		ControlNet:           c.Diffusers.ControlNet,
 		ContextSize:          int32(ctxSize),
@@ -150,6 +151,8 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		TensorParallelSize:   int32(c.TensorParallelSize),
 		MMProj:               c.MMProj,
 		FlashAttention:       c.FlashAttention,
+		CacheTypeKey:         c.CacheTypeK,
+		CacheTypeValue:       c.CacheTypeV,
 		NoKVOffload:          c.NoKVOffloading,
 		YarnExtFactor:        c.YarnExtFactor,
 		YarnAttnFactor:       c.YarnAttnFactor,

core/cli/run.go

@@ -6,12 +6,12 @@ import (
 	"strings"
 	"time"
 
+	"github.com/mudler/LocalAI/core/application"
 	cli_api "github.com/mudler/LocalAI/core/cli/api"
 	cliContext "github.com/mudler/LocalAI/core/cli/context"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/http"
 	"github.com/mudler/LocalAI/core/p2p"
-	"github.com/mudler/LocalAI/core/startup"
 	"github.com/rs/zerolog"
 	"github.com/rs/zerolog/log"
 )
@@ -186,16 +186,16 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 	}
 
 	if r.PreloadBackendOnly {
-		_, _, _, err := startup.Startup(opts...)
+		_, err := application.New(opts...)
 		return err
 	}
 
-	cl, ml, options, err := startup.Startup(opts...)
+	app, err := application.New(opts...)
 	if err != nil {
 		return fmt.Errorf("failed basic startup tasks with error %s", err.Error())
 	}
 
-	appHTTP, err := http.App(cl, ml, options)
+	appHTTP, err := http.API(app)
 	if err != nil {
 		log.Error().Err(err).Msg("error during HTTP App construction")
 		return err

core/cli/worker/worker_p2p.go

@@ -76,8 +76,14 @@ func (r *P2P) Run(ctx *cliContext.Context) error {
 					"util",
 					"llama-cpp-rpc-server",
 				)
-				extraArgs := strings.Split(r.ExtraLLamaCPPArgs, " ")
+				var extraArgs []string
+
+				if r.ExtraLLamaCPPArgs != "" {
+					extraArgs = strings.Split(r.ExtraLLamaCPPArgs, " ")
+				}
 				args := append([]string{"--host", address, "--port", fmt.Sprint(port)}, extraArgs...)
+				log.Debug().Msgf("Starting llama-cpp-rpc-server on '%s:%d' with args: %+v (%d)", address, port, args, len(args))
+
 				args, grpcProcess = library.LoadLDSO(r.BackendAssetsPath, args, grpcProcess)
 
 				cmd := exec.Command(

core/config/backend_config.go

@@ -72,6 +72,8 @@ type BackendConfig struct {
 
 	Description string `yaml:"description"`
 	Usage       string `yaml:"usage"`
+
+	Options []string `yaml:"options"`
 }
 
 type File struct {
@@ -97,16 +99,15 @@ type GRPC struct {
 }
 
 type Diffusers struct {
-	CUDA             bool    `yaml:"cuda"`
-	PipelineType     string  `yaml:"pipeline_type"`
-	SchedulerType    string  `yaml:"scheduler_type"`
-	EnableParameters string  `yaml:"enable_parameters"` // A list of comma separated parameters to specify
-	CFGScale         float32 `yaml:"cfg_scale"`         // Classifier-Free Guidance Scale
-	IMG2IMG          bool    `yaml:"img2img"`           // Image to Image Diffuser
-	ClipSkip         int     `yaml:"clip_skip"`         // Skip every N frames
-	ClipModel        string  `yaml:"clip_model"`        // Clip model to use
-	ClipSubFolder    string  `yaml:"clip_subfolder"`    // Subfolder to use for clip model
-	ControlNet       string  `yaml:"control_net"`
+	CUDA             bool   `yaml:"cuda"`
+	PipelineType     string `yaml:"pipeline_type"`
+	SchedulerType    string `yaml:"scheduler_type"`
+	EnableParameters string `yaml:"enable_parameters"` // A list of comma separated parameters to specify
+	IMG2IMG          bool   `yaml:"img2img"`           // Image to Image Diffuser
+	ClipSkip         int    `yaml:"clip_skip"`         // Skip every N frames
+	ClipModel        string `yaml:"clip_model"`        // Clip model to use
+	ClipSubFolder    string `yaml:"clip_subfolder"`    // Subfolder to use for clip model
+	ControlNet       string `yaml:"control_net"`
 }
 
 // LLMConfig is a struct that holds the configuration that are
@@ -154,8 +155,10 @@ type LLMConfig struct {
 	TensorParallelSize   int       `yaml:"tensor_parallel_size"`   // vLLM
 	MMProj               string    `yaml:"mmproj"`
 
-	FlashAttention bool `yaml:"flash_attention"`
-	NoKVOffloading bool `yaml:"no_kv_offloading"`
+	FlashAttention bool   `yaml:"flash_attention"`
+	NoKVOffloading bool   `yaml:"no_kv_offloading"`
+	CacheTypeK     string `yaml:"cache_type_k"`
+	CacheTypeV     string `yaml:"cache_type_v"`
 
 	RopeScaling string `yaml:"rope_scaling"`
 	ModelType   string `yaml:"type"`
@@ -164,6 +167,8 @@ type LLMConfig struct {
 	YarnAttnFactor float32 `yaml:"yarn_attn_factor"`
 	YarnBetaFast   float32 `yaml:"yarn_beta_fast"`
 	YarnBetaSlow   float32 `yaml:"yarn_beta_slow"`
+
+	CFGScale float32 `yaml:"cfg_scale"` // Classifier-Free Guidance Scale
 }
 
 // AutoGPTQ is a struct that holds the configuration specific to the AutoGPTQ backend
@@ -201,6 +206,8 @@ type TemplateConfig struct {
 	JoinChatMessagesByCharacter *string `yaml:"join_chat_messages_by_character"`
 
 	Multimodal string `yaml:"multimodal"`
+
+	JinjaTemplate bool `yaml:"jinja_template"`
 }
 
 func (c *BackendConfig) UnmarshalYAML(value *yaml.Node) error {

core/config/backend_config_loader.go

@@ -140,7 +140,7 @@ func (bcl *BackendConfigLoader) LoadBackendConfigFileByName(modelName, modelPath
 		}
 	}
 
-	cfg.SetDefaults(opts...)
+	cfg.SetDefaults(append(opts, ModelPath(modelPath))...)
 
 	return cfg, nil
 }

core/config/guesser.go

@@ -26,14 +26,14 @@ const (
 type settingsConfig struct {
 	StopWords      []string
 	TemplateConfig TemplateConfig
-	RepeatPenalty float64
+	RepeatPenalty  float64
 }
 
 // default settings to adopt with a given model family
 var defaultsSettings map[familyType]settingsConfig = map[familyType]settingsConfig{
 	Gemma: {
 		RepeatPenalty: 1.0,
-		StopWords: []string{"<|im_end|>", "<end_of_turn>", "<start_of_turn>"},
+		StopWords:     []string{"<|im_end|>", "<end_of_turn>", "<start_of_turn>"},
 		TemplateConfig: TemplateConfig{
 			Chat:        "{{.Input }}\n<start_of_turn>model\n",
 			ChatMessage: "<start_of_turn>{{if eq .RoleName \"assistant\" }}model{{else}}{{ .RoleName }}{{end}}\n{{ if .Content -}}\n{{.Content -}}\n{{ end -}}<end_of_turn>",
@@ -200,6 +200,18 @@ func guessDefaultsFromFile(cfg *BackendConfig, modelPath string) {
 	} else {
 		log.Debug().Any("family", family).Msgf("guessDefaultsFromFile: no template found for family")
 	}
+
+	if cfg.HasTemplate() {
+		return
+	}
+
+	// identify from well known templates first, otherwise use the raw jinja template
+	chatTemplate, found := f.Header.MetadataKV.Get("tokenizer.chat_template")
+	if found {
+		// try to use the jinja template
+		cfg.TemplateConfig.JinjaTemplate = true
+		cfg.TemplateConfig.ChatMessage = chatTemplate.ValueString()
+	}
 }
 
 func identifyFamily(f *gguf.GGUFFile) familyType {

core/gallery/models_test.go

@@ -12,6 +12,8 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
+const bertEmbeddingsURL = `https://gist.githubusercontent.com/mudler/0a080b166b87640e8644b09c2aee6e3b/raw/f0e8c26bb72edc16d9fbafbfd6638072126ff225/bert-embeddings-gallery.yaml`
+
 var _ = Describe("Model test", func() {
 
 	Context("Downloading", func() {
@@ -47,7 +49,7 @@ var _ = Describe("Model test", func() {
 
 			gallery := []GalleryModel{{
 				Name: "bert",
-				URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+				URL:  bertEmbeddingsURL,
 			}}
 			out, err := yaml.Marshal(gallery)
 			Expect(err).ToNot(HaveOccurred())
@@ -66,7 +68,7 @@ var _ = Describe("Model test", func() {
 			Expect(err).ToNot(HaveOccurred())
 			Expect(len(models)).To(Equal(1))
 			Expect(models[0].Name).To(Equal("bert"))
-			Expect(models[0].URL).To(Equal("https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml"))
+			Expect(models[0].URL).To(Equal(bertEmbeddingsURL))
 			Expect(models[0].Installed).To(BeFalse())
 
 			err = InstallModelFromGallery(galleries, "test@bert", tempdir, GalleryModel{}, func(s1, s2, s3 string, f float64) {}, true)
@@ -78,7 +80,7 @@ var _ = Describe("Model test", func() {
 			content := map[string]interface{}{}
 			err = yaml.Unmarshal(dat, &content)
 			Expect(err).ToNot(HaveOccurred())
-			Expect(content["backend"]).To(Equal("bert-embeddings"))
+			Expect(content["usage"]).To(ContainSubstring("You can test this model with curl like this"))
 
 			models, err = AvailableGalleryModels(galleries, tempdir)
 			Expect(err).ToNot(HaveOccurred())

core/http/app.go

@@ -14,10 +14,9 @@ import (
 	"github.com/mudler/LocalAI/core/http/middleware"
 	"github.com/mudler/LocalAI/core/http/routes"
 
-	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/application"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/core/services"
-	"github.com/mudler/LocalAI/pkg/model"
 
 	"github.com/gofiber/contrib/fiberzerolog"
 	"github.com/gofiber/fiber/v2"
@@ -49,18 +48,18 @@ var embedDirStatic embed.FS
 // @in header
 // @name Authorization
 
-func App(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) (*fiber.App, error) {
+func API(application *application.Application) (*fiber.App, error) {
 
 	fiberCfg := fiber.Config{
 		Views:     renderEngine(),
-		BodyLimit: appConfig.UploadLimitMB * 1024 * 1024, // this is the default limit of 4MB
+		BodyLimit: application.ApplicationConfig().UploadLimitMB * 1024 * 1024, // this is the default limit of 4MB
 		// We disable the Fiber startup message as it does not conform to structured logging.
 		// We register a startup log line with connection information in the OnListen hook to keep things user friendly though
 		DisableStartupMessage: true,
 		// Override default error handler
 	}
 
-	if !appConfig.OpaqueErrors {
+	if !application.ApplicationConfig().OpaqueErrors {
 		// Normally, return errors as JSON responses
 		fiberCfg.ErrorHandler = func(ctx *fiber.Ctx, err error) error {
 			// Status code defaults to 500
@@ -86,9 +85,11 @@ func App(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *confi
 		}
 	}
 
-	app := fiber.New(fiberCfg)
+	router := fiber.New(fiberCfg)
 
-	app.Hooks().OnListen(func(listenData fiber.ListenData) error {
+	router.Use(middleware.StripPathPrefix())
+
+	router.Hooks().OnListen(func(listenData fiber.ListenData) error {
 		scheme := "http"
 		if listenData.TLS {
 			scheme = "https"
@@ -99,82 +100,82 @@ func App(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *confi
 
 	// Have Fiber use zerolog like the rest of the application rather than it's built-in logger
 	logger := log.Logger
-	app.Use(fiberzerolog.New(fiberzerolog.Config{
+	router.Use(fiberzerolog.New(fiberzerolog.Config{
 		Logger: &logger,
 	}))
 
 	// Default middleware config
 
-	if !appConfig.Debug {
-		app.Use(recover.New())
+	if !application.ApplicationConfig().Debug {
+		router.Use(recover.New())
 	}
 
-	if !appConfig.DisableMetrics {
+	if !application.ApplicationConfig().DisableMetrics {
 		metricsService, err := services.NewLocalAIMetricsService()
 		if err != nil {
 			return nil, err
 		}
 
 		if metricsService != nil {
-			app.Use(localai.LocalAIMetricsAPIMiddleware(metricsService))
-			app.Hooks().OnShutdown(func() error {
+			router.Use(localai.LocalAIMetricsAPIMiddleware(metricsService))
+			router.Hooks().OnShutdown(func() error {
 				return metricsService.Shutdown()
 			})
 		}
 
 	}
 	// Health Checks should always be exempt from auth, so register these first
-	routes.HealthRoutes(app)
+	routes.HealthRoutes(router)
 
-	kaConfig, err := middleware.GetKeyAuthConfig(appConfig)
+	kaConfig, err := middleware.GetKeyAuthConfig(application.ApplicationConfig())
 	if err != nil || kaConfig == nil {
 		return nil, fmt.Errorf("failed to create key auth config: %w", err)
 	}
 
 	// Auth is applied to _all_ endpoints. No exceptions. Filtering out endpoints to bypass is the role of the Filter property of the KeyAuth Configuration
-	app.Use(v2keyauth.New(*kaConfig))
+	router.Use(v2keyauth.New(*kaConfig))
 
-	if appConfig.CORS {
+	if application.ApplicationConfig().CORS {
 		var c func(ctx *fiber.Ctx) error
-		if appConfig.CORSAllowOrigins == "" {
+		if application.ApplicationConfig().CORSAllowOrigins == "" {
 			c = cors.New()
 		} else {
-			c = cors.New(cors.Config{AllowOrigins: appConfig.CORSAllowOrigins})
+			c = cors.New(cors.Config{AllowOrigins: application.ApplicationConfig().CORSAllowOrigins})
 		}
 
-		app.Use(c)
+		router.Use(c)
 	}
 
-	if appConfig.CSRF {
+	if application.ApplicationConfig().CSRF {
 		log.Debug().Msg("Enabling CSRF middleware. Tokens are now required for state-modifying requests")
-		app.Use(csrf.New())
+		router.Use(csrf.New())
 	}
 
 	// Load config jsons
-	utils.LoadConfig(appConfig.UploadDir, openai.UploadedFilesFile, &openai.UploadedFiles)
-	utils.LoadConfig(appConfig.ConfigsDir, openai.AssistantsConfigFile, &openai.Assistants)
-	utils.LoadConfig(appConfig.ConfigsDir, openai.AssistantsFileConfigFile, &openai.AssistantFiles)
+	utils.LoadConfig(application.ApplicationConfig().UploadDir, openai.UploadedFilesFile, &openai.UploadedFiles)
+	utils.LoadConfig(application.ApplicationConfig().ConfigsDir, openai.AssistantsConfigFile, &openai.Assistants)
+	utils.LoadConfig(application.ApplicationConfig().ConfigsDir, openai.AssistantsFileConfigFile, &openai.AssistantFiles)
 
-	galleryService := services.NewGalleryService(appConfig)
-	galleryService.Start(appConfig.Context, cl)
+	galleryService := services.NewGalleryService(application.ApplicationConfig())
+	galleryService.Start(application.ApplicationConfig().Context, application.BackendLoader())
 
-	routes.RegisterElevenLabsRoutes(app, cl, ml, appConfig)
-	routes.RegisterLocalAIRoutes(app, cl, ml, appConfig, galleryService)
-	routes.RegisterOpenAIRoutes(app, cl, ml, appConfig)
-	if !appConfig.DisableWebUI {
-		routes.RegisterUIRoutes(app, cl, ml, appConfig, galleryService)
+	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(app, cl, ml, appConfig)
+	routes.RegisterJINARoutes(router, application.BackendLoader(), application.ModelLoader(), application.ApplicationConfig())
 
 	httpFS := http.FS(embedDirStatic)
 
-	app.Use(favicon.New(favicon.Config{
+	router.Use(favicon.New(favicon.Config{
 		URL:        "/favicon.ico",
 		FileSystem: httpFS,
 		File:       "static/favicon.ico",
 	}))
 
-	app.Use("/static", filesystem.New(filesystem.Config{
+	router.Use("/static", filesystem.New(filesystem.Config{
 		Root:       httpFS,
 		PathPrefix: "static",
 		Browse:     true,
@@ -182,7 +183,7 @@ func App(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *confi
 
 	// Define a custom 404 handler
 	// Note: keep this at the bottom!
-	app.Use(notFoundHandler)
+	router.Use(notFoundHandler)
 
-	return app, nil
+	return router, nil
 }

core/http/app_test.go

@@ -5,24 +5,21 @@ import (
 	"context"
 	"embed"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"io"
 	"net/http"
 	"os"
 	"path/filepath"
 	"runtime"
-	"strings"
 
+	"github.com/mudler/LocalAI/core/application"
 	"github.com/mudler/LocalAI/core/config"
 	. "github.com/mudler/LocalAI/core/http"
 	"github.com/mudler/LocalAI/core/schema"
-	"github.com/mudler/LocalAI/core/startup"
 
 	"github.com/gofiber/fiber/v2"
 	"github.com/mudler/LocalAI/core/gallery"
 	"github.com/mudler/LocalAI/pkg/downloader"
-	"github.com/mudler/LocalAI/pkg/model"
 	. "github.com/onsi/ginkgo/v2"
 	. "github.com/onsi/gomega"
 	"gopkg.in/yaml.v3"
@@ -240,6 +237,33 @@ func postInvalidRequest(url string) (error, int) {
 	return nil, resp.StatusCode
 }
 
+func getRequest(url string, header http.Header) (error, int, []byte) {
+
+	req, err := http.NewRequest("GET", url, nil)
+	if err != nil {
+		return err, -1, nil
+	}
+
+	req.Header = header
+
+	client := &http.Client{}
+	resp, err := client.Do(req)
+	if err != nil {
+		return err, -1, nil
+	}
+
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return err, -1, nil
+	}
+
+	return nil, resp.StatusCode, body
+}
+
+const bertEmbeddingsURL = `https://gist.githubusercontent.com/mudler/0a080b166b87640e8644b09c2aee6e3b/raw/f0e8c26bb72edc16d9fbafbfd6638072126ff225/bert-embeddings-gallery.yaml`
+
 //go:embed backend-assets/*
 var backendAssets embed.FS
 
@@ -252,9 +276,6 @@ var _ = Describe("API test", func() {
 	var cancel context.CancelFunc
 	var tmpdir string
 	var modelDir string
-	var bcl *config.BackendConfigLoader
-	var ml *model.ModelLoader
-	var applicationConfig *config.ApplicationConfig
 
 	commonOpts := []config.AppOption{
 		config.WithDebug(true),
@@ -279,13 +300,13 @@ var _ = Describe("API test", func() {
 			g := []gallery.GalleryModel{
 				{
 					Name: "bert",
-					URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+					URL:  bertEmbeddingsURL,
 				},
 				{
 					Name:            "bert2",
-					URL:             "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+					URL:             bertEmbeddingsURL,
 					Overrides:       map[string]interface{}{"foo": "bar"},
-					AdditionalFiles: []gallery.File{{Filename: "foo.yaml", URI: "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml"}},
+					AdditionalFiles: []gallery.File{{Filename: "foo.yaml", URI: bertEmbeddingsURL}},
 				},
 			}
 			out, err := yaml.Marshal(g)
@@ -300,7 +321,7 @@ var _ = Describe("API test", func() {
 				},
 			}
 
-			bcl, ml, applicationConfig, err = startup.Startup(
+			application, err := application.New(
 				append(commonOpts,
 					config.WithContext(c),
 					config.WithGalleries(galleries),
@@ -310,7 +331,7 @@ var _ = Describe("API test", func() {
 					config.WithBackendAssetsOutput(backendAssetsDir))...)
 			Expect(err).ToNot(HaveOccurred())
 
-			app, err = App(bcl, ml, applicationConfig)
+			app, err = API(application)
 			Expect(err).ToNot(HaveOccurred())
 
 			go app.Listen("127.0.0.1:9090")
@@ -345,7 +366,34 @@ var _ = Describe("API test", func() {
 			It("Should fail if the api key is missing", func() {
 				err, sc := postInvalidRequest("http://127.0.0.1:9090/models/available")
 				Expect(err).ToNot(BeNil())
-				Expect(sc).To(Equal(403))
+				Expect(sc).To(Equal(401))
+			})
+		})
+
+		Context("URL routing Tests", func() {
+			It("Should support reverse-proxy when unauthenticated", func() {
+
+				err, sc, body := getRequest("http://127.0.0.1:9090/myprefix/", http.Header{
+					"X-Forwarded-Proto":  {"https"},
+					"X-Forwarded-Host":   {"example.org"},
+					"X-Forwarded-Prefix": {"/myprefix/"},
+				})
+				Expect(err).To(BeNil(), "error")
+				Expect(sc).To(Equal(401), "status code")
+				Expect(string(body)).To(ContainSubstring(`<base href="https://example.org/myprefix/" />`), "body")
+			})
+
+			It("Should support reverse-proxy when authenticated", func() {
+
+				err, sc, body := getRequest("http://127.0.0.1:9090/myprefix/", http.Header{
+					"Authorization":      {bearerKey},
+					"X-Forwarded-Proto":  {"https"},
+					"X-Forwarded-Host":   {"example.org"},
+					"X-Forwarded-Prefix": {"/myprefix/"},
+				})
+				Expect(err).To(BeNil(), "error")
+				Expect(sc).To(Equal(200), "status code")
+				Expect(string(body)).To(ContainSubstring(`<base href="https://example.org/myprefix/" />`), "body")
 			})
 		})
 
@@ -383,7 +431,7 @@ var _ = Describe("API test", func() {
 				content := map[string]interface{}{}
 				err = yaml.Unmarshal(dat, &content)
 				Expect(err).ToNot(HaveOccurred())
-				Expect(content["backend"]).To(Equal("bert-embeddings"))
+				Expect(content["usage"]).To(ContainSubstring("You can test this model with curl like this"))
 				Expect(content["foo"]).To(Equal("bar"))
 
 				models, err = getModels("http://127.0.0.1:9090/models/available")
@@ -402,7 +450,7 @@ var _ = Describe("API test", func() {
 			It("overrides models", func() {
 
 				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
-					URL:  "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+					URL:  bertEmbeddingsURL,
 					Name: "bert",
 					Overrides: map[string]interface{}{
 						"backend": "llama",
@@ -451,7 +499,7 @@ var _ = Describe("API test", func() {
 			})
 			It("apply models without overrides", func() {
 				response := postModelApplyRequest("http://127.0.0.1:9090/models/apply", modelApplyRequest{
-					URL:       "https://raw.githubusercontent.com/go-skynet/model-gallery/main/bert-embeddings.yaml",
+					URL:       bertEmbeddingsURL,
 					Name:      "bert",
 					Overrides: map[string]interface{}{},
 				})
@@ -471,7 +519,7 @@ var _ = Describe("API test", func() {
 				content := map[string]interface{}{}
 				err = yaml.Unmarshal(dat, &content)
 				Expect(err).ToNot(HaveOccurred())
-				Expect(content["backend"]).To(Equal("bert-embeddings"))
+				Expect(content["usage"]).To(ContainSubstring("You can test this model with curl like this"))
 			})
 
 			It("runs openllama(llama-ggml backend)", Label("llama"), func() {
@@ -539,7 +587,7 @@ var _ = Describe("API test", func() {
 				var res map[string]string
 				err = json.Unmarshal([]byte(resp2.Choices[0].Message.FunctionCall.Arguments), &res)
 				Expect(err).ToNot(HaveOccurred())
-				Expect(res["location"]).To(Equal("San Francisco"), fmt.Sprint(res))
+				Expect(res["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))
 
@@ -641,7 +689,7 @@ var _ = Describe("API test", func() {
 				},
 			}
 
-			bcl, ml, applicationConfig, err = startup.Startup(
+			application, err := application.New(
 				append(commonOpts,
 					config.WithContext(c),
 					config.WithAudioDir(tmpdir),
@@ -652,7 +700,7 @@ var _ = Describe("API test", func() {
 					config.WithBackendAssetsOutput(tmpdir))...,
 			)
 			Expect(err).ToNot(HaveOccurred())
-			app, err = App(bcl, ml, applicationConfig)
+			app, err = API(application)
 			Expect(err).ToNot(HaveOccurred())
 
 			go app.Listen("127.0.0.1:9090")
@@ -708,7 +756,7 @@ var _ = Describe("API test", func() {
 			Expect(err).ToNot(HaveOccurred(), fmt.Sprint(resp))
 
 			Expect(resp.StatusCode).To(Equal(200), fmt.Sprint(string(dat)))
-			Expect(resp.Header.Get("Content-Type")).To(Equal("audio/x-wav"))
+			Expect(resp.Header.Get("Content-Type")).To(Or(Equal("audio/x-wav"), Equal("audio/vnd.wave")))
 		})
 		It("installs and is capable to generate images", Label("stablediffusion"), func() {
 			if runtime.GOOS != "linux" {
@@ -772,14 +820,14 @@ var _ = Describe("API test", func() {
 
 			var err error
 
-			bcl, ml, applicationConfig, err = startup.Startup(
+			application, err := application.New(
 				append(commonOpts,
 					config.WithExternalBackend("huggingface", os.Getenv("HUGGINGFACE_GRPC")),
 					config.WithContext(c),
 					config.WithModelPath(modelPath),
 				)...)
 			Expect(err).ToNot(HaveOccurred())
-			app, err = App(bcl, ml, applicationConfig)
+			app, err = API(application)
 			Expect(err).ToNot(HaveOccurred())
 			go app.Listen("127.0.0.1:9090")
 
@@ -806,7 +854,7 @@ var _ = Describe("API test", func() {
 		It("returns the models list", func() {
 			models, err := client.ListModels(context.TODO())
 			Expect(err).ToNot(HaveOccurred())
-			Expect(len(models.Models)).To(Equal(6)) // If "config.yaml" should be included, this should be 8?
+			Expect(len(models.Models)).To(Equal(7)) // If "config.yaml" should be included, this should be 8?
 		})
 		It("can generate completions via ggml", func() {
 			resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "testmodel.ggml", Prompt: testPrompt})
@@ -866,8 +914,8 @@ var _ = Describe("API test", func() {
 				},
 			)
 			Expect(err).ToNot(HaveOccurred(), err)
-			Expect(len(resp.Data[0].Embedding)).To(BeNumerically("==", 384))
-			Expect(len(resp.Data[1].Embedding)).To(BeNumerically("==", 384))
+			Expect(len(resp.Data[0].Embedding)).To(BeNumerically("==", 2048))
+			Expect(len(resp.Data[1].Embedding)).To(BeNumerically("==", 2048))
 
 			sunEmbedding := resp.Data[0].Embedding
 			resp2, err := client.CreateEmbeddings(
@@ -911,71 +959,6 @@ var _ = Describe("API test", func() {
 			})
 		})
 
-		Context("backends", func() {
-			It("runs rwkv completion", func() {
-				if runtime.GOOS != "linux" {
-					Skip("test supported only on linux")
-				}
-				resp, err := client.CreateCompletion(context.TODO(), openai.CompletionRequest{Model: "rwkv_test", Prompt: "Count up to five: one, two, three, four,"})
-				Expect(err).ToNot(HaveOccurred())
-				Expect(len(resp.Choices) > 0).To(BeTrue())
-				Expect(resp.Choices[0].Text).To(ContainSubstring("five"))
-
-				stream, err := client.CreateCompletionStream(context.TODO(), openai.CompletionRequest{
-					Model: "rwkv_test", Prompt: "Count up to five: one, two, three, four,", Stream: true,
-				})
-				Expect(err).ToNot(HaveOccurred())
-				defer stream.Close()
-
-				tokens := 0
-				text := ""
-				for {
-					response, err := stream.Recv()
-					if errors.Is(err, io.EOF) {
-						break
-					}
-
-					Expect(err).ToNot(HaveOccurred())
-					text += response.Choices[0].Text
-					tokens++
-				}
-				Expect(text).ToNot(BeEmpty())
-				Expect(text).To(ContainSubstring("five"))
-				Expect(tokens).ToNot(Or(Equal(1), Equal(0)))
-			})
-			It("runs rwkv chat completion", func() {
-				if runtime.GOOS != "linux" {
-					Skip("test supported only on linux")
-				}
-				resp, err := client.CreateChatCompletion(context.TODO(),
-					openai.ChatCompletionRequest{Model: "rwkv_test", Messages: []openai.ChatCompletionMessage{{Content: "Can you count up to five?", Role: "user"}}})
-				Expect(err).ToNot(HaveOccurred())
-				Expect(len(resp.Choices) > 0).To(BeTrue())
-				Expect(strings.ToLower(resp.Choices[0].Message.Content)).To(Or(ContainSubstring("sure"), ContainSubstring("five")))
-
-				stream, err := client.CreateChatCompletionStream(context.TODO(), openai.ChatCompletionRequest{Model: "rwkv_test", Messages: []openai.ChatCompletionMessage{{Content: "Can you count up to five?", Role: "user"}}})
-				Expect(err).ToNot(HaveOccurred())
-				defer stream.Close()
-
-				tokens := 0
-				text := ""
-				for {
-					response, err := stream.Recv()
-					if errors.Is(err, io.EOF) {
-						break
-					}
-
-					Expect(err).ToNot(HaveOccurred())
-					text += response.Choices[0].Delta.Content
-					tokens++
-				}
-				Expect(text).ToNot(BeEmpty())
-				Expect(strings.ToLower(text)).To(Or(ContainSubstring("sure"), ContainSubstring("five")))
-
-				Expect(tokens).ToNot(Or(Equal(1), Equal(0)))
-			})
-		})
-
 		// See tests/integration/stores_test
 		Context("Stores", Label("stores"), func() {
 
@@ -1055,14 +1038,14 @@ var _ = Describe("API test", func() {
 			c, cancel = context.WithCancel(context.Background())
 
 			var err error
-			bcl, ml, applicationConfig, err = startup.Startup(
+			application, err := application.New(
 				append(commonOpts,
 					config.WithContext(c),
 					config.WithModelPath(modelPath),
 					config.WithConfigFile(os.Getenv("CONFIG_FILE")))...,
 			)
 			Expect(err).ToNot(HaveOccurred())
-			app, err = App(bcl, ml, applicationConfig)
+			app, err = API(application)
 			Expect(err).ToNot(HaveOccurred())
 
 			go app.Listen("127.0.0.1:9090")

core/http/elements/buttons.go

@@ -16,7 +16,7 @@ func installButton(galleryName string) elem.Node {
 			"class":                 "float-right inline-block rounded bg-primary px-6 pb-2.5 mb-3 pt-2.5 text-xs font-medium uppercase leading-normal text-white shadow-primary-3 transition duration-150 ease-in-out hover:bg-primary-accent-300 hover:shadow-primary-2 focus:bg-primary-accent-300 focus:shadow-primary-2 focus:outline-none focus:ring-0 active:bg-primary-600 active:shadow-primary-2 dark:shadow-black/30 dark:hover:shadow-dark-strong dark:focus:shadow-dark-strong dark:active:shadow-dark-strong",
 			"hx-swap":               "outerHTML",
 			// post the Model ID as param
-			"hx-post": "/browse/install/model/" + galleryName,
+			"hx-post": "browse/install/model/" + galleryName,
 		},
 		elem.I(
 			attrs.Props{
@@ -36,7 +36,7 @@ func reInstallButton(galleryName string) elem.Node {
 			"hx-target":             "#action-div-" + dropBadChars(galleryName),
 			"hx-swap":               "outerHTML",
 			// post the Model ID as param
-			"hx-post": "/browse/install/model/" + galleryName,
+			"hx-post": "browse/install/model/" + galleryName,
 		},
 		elem.I(
 			attrs.Props{
@@ -80,7 +80,7 @@ func deleteButton(galleryID string) elem.Node {
 			"hx-target":             "#action-div-" + dropBadChars(galleryID),
 			"hx-swap":               "outerHTML",
 			// post the Model ID as param
-			"hx-post": "/browse/delete/model/" + galleryID,
+			"hx-post": "browse/delete/model/" + galleryID,
 		},
 		elem.I(
 			attrs.Props{

core/http/elements/gallery.go

@@ -47,7 +47,7 @@ func searchableElement(text, icon string) elem.Node {
 					//	"value":     text,
 					//"class":     "inline-block bg-gray-200 rounded-full px-3 py-1 text-sm font-semibold text-gray-700 mr-2 mb-2",
 					"href":      "#!",
-					"hx-post":   "/browse/search/models",
+					"hx-post":   "browse/search/models",
 					"hx-target": "#search-results",
 					// TODO: this doesn't work
 					//	"hx-vals":      `{ \"search\": \"` + text + `\" }`,

core/http/elements/progressbar.go

@@ -64,7 +64,7 @@ func StartProgressBar(uid, progress, text string) string {
 	return elem.Div(
 		attrs.Props{
 			"hx-trigger": "done",
-			"hx-get":     "/browse/job/" + uid,
+			"hx-get":     "browse/job/" + uid,
 			"hx-swap":    "outerHTML",
 			"hx-target":  "this",
 		},
@@ -77,7 +77,7 @@ func StartProgressBar(uid, progress, text string) string {
 			},
 			elem.Text(bluemonday.StrictPolicy().Sanitize(text)), //Perhaps overly defensive
 			elem.Div(attrs.Props{
-				"hx-get":     "/browse/job/progress/" + uid,
+				"hx-get":     "browse/job/progress/" + uid,
 				"hx-trigger": "every 600ms",
 				"hx-target":  "this",
 				"hx-swap":    "innerHTML",

core/http/endpoints/explorer/dashboard.go

@@ -6,6 +6,7 @@ import (
 
 	"github.com/gofiber/fiber/v2"
 	"github.com/mudler/LocalAI/core/explorer"
+	"github.com/mudler/LocalAI/core/http/utils"
 	"github.com/mudler/LocalAI/internal"
 )
 
@@ -14,6 +15,7 @@ func Dashboard() func(*fiber.Ctx) error {
 		summary := fiber.Map{
 			"Title":   "LocalAI API - " + internal.PrintableVersion(),
 			"Version": internal.PrintableVersion(),
+			"BaseURL": utils.BaseURL(c),
 		}
 
 		if string(c.Context().Request.Header.ContentType()) == "application/json" || len(c.Accepts("html")) == 0 {

core/http/endpoints/localai/gallery.go

@@ -9,6 +9,7 @@ import (
 	"github.com/google/uuid"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery"
+	"github.com/mudler/LocalAI/core/http/utils"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/core/services"
 	"github.com/rs/zerolog/log"
@@ -82,7 +83,8 @@ func (mgs *ModelGalleryEndpointService) ApplyModelGalleryEndpoint() func(c *fibe
 			Galleries:        mgs.galleries,
 			ConfigURL:        input.ConfigURL,
 		}
-		return c.JSON(schema.GalleryResponse{ID: uuid.String(), StatusURL: c.BaseURL() + "/models/jobs/" + uuid.String()})
+
+		return c.JSON(schema.GalleryResponse{ID: uuid.String(), StatusURL: fmt.Sprintf("%smodels/jobs/%s", utils.BaseURL(c), uuid.String())})
 	}
 }
 
@@ -105,7 +107,7 @@ func (mgs *ModelGalleryEndpointService) DeleteModelGalleryEndpoint() func(c *fib
 			return err
 		}
 
-		return c.JSON(schema.GalleryResponse{ID: uuid.String(), StatusURL: c.BaseURL() + "/models/jobs/" + uuid.String()})
+		return c.JSON(schema.GalleryResponse{ID: uuid.String(), StatusURL: fmt.Sprintf("%smodels/jobs/%s", utils.BaseURL(c), uuid.String())})
 	}
 }
 

core/http/endpoints/localai/system.go

@@ -21,10 +21,15 @@ func SystemInformations(ml *model.ModelLoader, appConfig *config.ApplicationConf
 		for b := range appConfig.ExternalGRPCBackends {
 			availableBackends = append(availableBackends, b)
 		}
+
+		sysmodels := []schema.SysInfoModel{}
+		for _, m := range loadedModels {
+			sysmodels = append(sysmodels, schema.SysInfoModel{ID: m.ID})
+		}
 		return c.JSON(
 			schema.SystemInformationResponse{
 				Backends: availableBackends,
-				Models:   loadedModels,
+				Models:   sysmodels,
 			},
 		)
 	}

core/http/endpoints/localai/vad.go

@@ -0,0 +1,68 @@
+package localai
+
+import (
+	"github.com/gofiber/fiber/v2"
+	"github.com/mudler/LocalAI/core/backend"
+	"github.com/mudler/LocalAI/core/config"
+	fiberContext "github.com/mudler/LocalAI/core/http/ctx"
+	"github.com/mudler/LocalAI/core/schema"
+	"github.com/mudler/LocalAI/pkg/grpc/proto"
+	"github.com/mudler/LocalAI/pkg/model"
+	"github.com/rs/zerolog/log"
+)
+
+// VADEndpoint is Voice-Activation-Detection endpoint
+// @Summary	Detect voice fragments in an audio stream
+// @Accept json
+// @Param		request	body		schema.VADRequest	true	"query params"
+// @Success 200 {object} proto.VADResponse "Response"
+// @Router		/vad [post]
+func VADEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
+	return func(c *fiber.Ctx) error {
+
+		input := new(schema.VADRequest)
+
+		// Get input data from the request body
+		if err := c.BodyParser(input); err != nil {
+			return err
+		}
+
+		modelFile, err := fiberContext.ModelFromContext(c, cl, ml, input.Model, false)
+		if err != nil {
+			modelFile = input.Model
+			log.Warn().Msgf("Model not found in context: %s", input.Model)
+		}
+
+		cfg, err := cl.LoadBackendConfigFileByName(modelFile, appConfig.ModelPath,
+			config.LoadOptionDebug(appConfig.Debug),
+			config.LoadOptionThreads(appConfig.Threads),
+			config.LoadOptionContextSize(appConfig.ContextSize),
+			config.LoadOptionF16(appConfig.F16),
+		)
+
+		if err != nil {
+			log.Err(err)
+			modelFile = input.Model
+			log.Warn().Msgf("Model not found in context: %s", input.Model)
+		} else {
+			modelFile = cfg.Model
+		}
+		log.Debug().Msgf("Request for model: %s", modelFile)
+
+		opts := backend.ModelOptions(*cfg, appConfig, model.WithBackendString(cfg.Backend), model.WithModel(modelFile))
+
+		vadModel, err := ml.Load(opts...)
+		if err != nil {
+			return err
+		}
+		req := proto.VADRequest{
+			Audio: input.Audio,
+		}
+		resp, err := vadModel.VAD(c.Context(), &req)
+		if err != nil {
+			return err
+		}
+
+		return c.JSON(resp)
+	}
+}

core/http/endpoints/localai/welcome.go

@@ -4,6 +4,7 @@ import (
 	"github.com/gofiber/fiber/v2"
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/gallery"
+	"github.com/mudler/LocalAI/core/http/utils"
 	"github.com/mudler/LocalAI/core/p2p"
 	"github.com/mudler/LocalAI/core/services"
 	"github.com/mudler/LocalAI/internal"
@@ -32,6 +33,7 @@ func WelcomeEndpoint(appConfig *config.ApplicationConfig,
 		summary := fiber.Map{
 			"Title":             "LocalAI API - " + internal.PrintableVersion(),
 			"Version":           internal.PrintableVersion(),
+			"BaseURL":           utils.BaseURL(c),
 			"Models":            modelsWithoutConfig,
 			"ModelsConfig":      backendConfigs,
 			"GalleryConfig":     galleryConfigs,

core/http/endpoints/openai/chat.go

@@ -14,6 +14,8 @@ import (
 	"github.com/mudler/LocalAI/core/config"
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/pkg/functions"
+	"github.com/mudler/LocalAI/pkg/templates"
+
 	model "github.com/mudler/LocalAI/pkg/model"
 	"github.com/rs/zerolog/log"
 	"github.com/valyala/fasthttp"
@@ -24,7 +26,7 @@ import (
 // @Param request body schema.OpenAIRequest true "query params"
 // @Success 200 {object} schema.OpenAIResponse "Response"
 // @Router /v1/chat/completions [post]
-func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startupOptions *config.ApplicationConfig) func(c *fiber.Ctx) error {
+func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, startupOptions *config.ApplicationConfig) func(c *fiber.Ctx) error {
 	var id, textContentToReturn string
 	var created int
 
@@ -294,148 +296,10 @@ func ChatEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, startup
 		// If we are using the tokenizer template, we don't need to process the messages
 		// unless we are processing functions
 		if !config.TemplateConfig.UseTokenizerTemplate || shouldUseFn {
-			suppressConfigSystemPrompt := false
-			mess := []string{}
-			for messageIndex, i := range input.Messages {
-				var content string
-				role := i.Role
-
-				// if function call, we might want to customize the role so we can display better that the "assistant called a json action"
-				// if an "assistant_function_call" role is defined, we use it, otherwise we use the role that is passed by in the request
-				if (i.FunctionCall != nil || i.ToolCalls != nil) && i.Role == "assistant" {
-					roleFn := "assistant_function_call"
-					r := config.Roles[roleFn]
-					if r != "" {
-						role = roleFn
-					}
-				}
-				r := config.Roles[role]
-				contentExists := i.Content != nil && i.StringContent != ""
-
-				fcall := i.FunctionCall
-				if len(i.ToolCalls) > 0 {
-					fcall = i.ToolCalls
-				}
-
-				// First attempt to populate content via a chat message specific template
-				if config.TemplateConfig.ChatMessage != "" {
-					chatMessageData := model.ChatMessageTemplateData{
-						SystemPrompt: config.SystemPrompt,
-						Role:         r,
-						RoleName:     role,
-						Content:      i.StringContent,
-						FunctionCall: fcall,
-						FunctionName: i.Name,
-						LastMessage:  messageIndex == (len(input.Messages) - 1),
-						Function:     config.Grammar != "" && (messageIndex == (len(input.Messages) - 1)),
-						MessageIndex: messageIndex,
-					}
-					templatedChatMessage, err := ml.EvaluateTemplateForChatMessage(config.TemplateConfig.ChatMessage, chatMessageData)
-					if err != nil {
-						log.Error().Err(err).Interface("message", chatMessageData).Str("template", config.TemplateConfig.ChatMessage).Msg("error processing message with template, skipping")
-					} else {
-						if templatedChatMessage == "" {
-							log.Warn().Msgf("template \"%s\" produced blank output for %+v. Skipping!", config.TemplateConfig.ChatMessage, chatMessageData)
-							continue // TODO: This continue is here intentionally to skip over the line `mess = append(mess, content)` below, and to prevent the sprintf
-						}
-						log.Debug().Msgf("templated message for chat: %s", templatedChatMessage)
-						content = templatedChatMessage
-					}
-				}
-
-				marshalAnyRole := func(f any) {
-					j, err := json.Marshal(f)
-					if err == nil {
-						if contentExists {
-							content += "\n" + fmt.Sprint(r, " ", string(j))
-						} else {
-							content = fmt.Sprint(r, " ", string(j))
-						}
-					}
-				}
-				marshalAny := func(f any) {
-					j, err := json.Marshal(f)
-					if err == nil {
-						if contentExists {
-							content += "\n" + string(j)
-						} else {
-							content = string(j)
-						}
-					}
-				}
-				// If this model doesn't have such a template, or if that template fails to return a value, template at the message level.
-				if content == "" {
-					if r != "" {
-						if contentExists {
-							content = fmt.Sprint(r, i.StringContent)
-						}
-
-						if i.FunctionCall != nil {
-							marshalAnyRole(i.FunctionCall)
-						}
-						if i.ToolCalls != nil {
-							marshalAnyRole(i.ToolCalls)
-						}
-					} else {
-						if contentExists {
-							content = fmt.Sprint(i.StringContent)
-						}
-						if i.FunctionCall != nil {
-							marshalAny(i.FunctionCall)
-						}
-						if i.ToolCalls != nil {
-							marshalAny(i.ToolCalls)
-						}
-					}
-					// Special Handling: System. We care if it was printed at all, not the r branch, so check seperately
-					if contentExists && role == "system" {
-						suppressConfigSystemPrompt = true
-					}
-				}
-
-				mess = append(mess, content)
-			}
-
-			joinCharacter := "\n"
-			if config.TemplateConfig.JoinChatMessagesByCharacter != nil {
-				joinCharacter = *config.TemplateConfig.JoinChatMessagesByCharacter
-			}
-
-			predInput = strings.Join(mess, joinCharacter)
-			log.Debug().Msgf("Prompt (before templating): %s", predInput)
-
-			templateFile := ""
-
-			// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-			if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
-				templateFile = config.Model
-			}
-
-			if config.TemplateConfig.Chat != "" && !shouldUseFn {
-				templateFile = config.TemplateConfig.Chat
-			}
-
-			if config.TemplateConfig.Functions != "" && shouldUseFn {
-				templateFile = config.TemplateConfig.Functions
-			}
-
-			if templateFile != "" {
-				templatedInput, err := ml.EvaluateTemplateForPrompt(model.ChatPromptTemplate, templateFile, model.PromptTemplateData{
-					SystemPrompt:         config.SystemPrompt,
-					SuppressSystemPrompt: suppressConfigSystemPrompt,
-					Input:                predInput,
-					Functions:            funcs,
-				})
-				if err == nil {
-					predInput = templatedInput
-					log.Debug().Msgf("Template found, input modified to: %s", predInput)
-				} else {
-					log.Debug().Msgf("Template failed loading: %s", err.Error())
-				}
-			}
+			predInput = evaluator.TemplateMessages(input.Messages, config, funcs, shouldUseFn)
 
 			log.Debug().Msgf("Prompt (after templating): %s", predInput)
-			if shouldUseFn && config.Grammar != "" {
+			if config.Grammar != "" {
 				log.Debug().Msgf("Grammar: %+v", config.Grammar)
 			}
 		}

core/http/endpoints/openai/completion.go

@@ -16,6 +16,7 @@ import (
 	"github.com/mudler/LocalAI/core/schema"
 	"github.com/mudler/LocalAI/pkg/functions"
 	model "github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/templates"
 	"github.com/rs/zerolog/log"
 	"github.com/valyala/fasthttp"
 )
@@ -25,7 +26,7 @@ import (
 // @Param request body schema.OpenAIRequest true "query params"
 // @Success 200 {object} schema.OpenAIResponse "Response"
 // @Router /v1/completions [post]
-func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
+func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
 	id := uuid.New().String()
 	created := int(time.Now().Unix())
 
@@ -94,17 +95,6 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
 			c.Set("Transfer-Encoding", "chunked")
 		}
 
-		templateFile := ""
-
-		// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-		if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
-			templateFile = config.Model
-		}
-
-		if config.TemplateConfig.Completion != "" {
-			templateFile = config.TemplateConfig.Completion
-		}
-
 		if input.Stream {
 			if len(config.PromptStrings) > 1 {
 				return errors.New("cannot handle more than 1 `PromptStrings` when Streaming")
@@ -112,15 +102,13 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
 
 			predInput := config.PromptStrings[0]
 
-			if templateFile != "" {
-				templatedInput, err := ml.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
-					Input:        predInput,
-					SystemPrompt: config.SystemPrompt,
-				})
-				if err == nil {
-					predInput = templatedInput
-					log.Debug().Msgf("Template found, input modified to: %s", predInput)
-				}
+			templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.CompletionPromptTemplate, *config, templates.PromptTemplateData{
+				Input:        predInput,
+				SystemPrompt: config.SystemPrompt,
+			})
+			if err == nil {
+				predInput = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", predInput)
 			}
 
 			responses := make(chan schema.OpenAIResponse)
@@ -165,16 +153,13 @@ func CompletionEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, a
 		totalTokenUsage := backend.TokenUsage{}
 
 		for k, i := range config.PromptStrings {
-			if templateFile != "" {
-				// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-				templatedInput, err := ml.EvaluateTemplateForPrompt(model.CompletionPromptTemplate, templateFile, model.PromptTemplateData{
-					SystemPrompt: config.SystemPrompt,
-					Input:        i,
-				})
-				if err == nil {
-					i = templatedInput
-					log.Debug().Msgf("Template found, input modified to: %s", i)
-				}
+			templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.CompletionPromptTemplate, *config, templates.PromptTemplateData{
+				SystemPrompt: config.SystemPrompt,
+				Input:        i,
+			})
+			if err == nil {
+				i = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", i)
 			}
 
 			r, tokenUsage, err := ComputeChoices(

core/http/endpoints/openai/edit.go

@@ -12,6 +12,7 @@ import (
 	"github.com/google/uuid"
 	"github.com/mudler/LocalAI/core/schema"
 	model "github.com/mudler/LocalAI/pkg/model"
+	"github.com/mudler/LocalAI/pkg/templates"
 
 	"github.com/rs/zerolog/log"
 )
@@ -21,7 +22,8 @@ import (
 // @Param request body schema.OpenAIRequest true "query params"
 // @Success 200 {object} schema.OpenAIResponse "Response"
 // @Router /v1/edits [post]
-func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
+func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, evaluator *templates.Evaluator, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
+
 	return func(c *fiber.Ctx) error {
 		modelFile, input, err := readRequest(c, cl, ml, appConfig, true)
 		if err != nil {
@@ -35,31 +37,18 @@ func EditEndpoint(cl *config.BackendConfigLoader, ml *model.ModelLoader, appConf
 
 		log.Debug().Msgf("Parameter Config: %+v", config)
 
-		templateFile := ""
-
-		// A model can have a "file.bin.tmpl" file associated with a prompt template prefix
-		if ml.ExistsInModelPath(fmt.Sprintf("%s.tmpl", config.Model)) {
-			templateFile = config.Model
-		}
-
-		if config.TemplateConfig.Edit != "" {
-			templateFile = config.TemplateConfig.Edit
-		}
-
 		var result []schema.Choice
 		totalTokenUsage := backend.TokenUsage{}
 
 		for _, i := range config.InputStrings {
-			if templateFile != "" {
-				templatedInput, err := ml.EvaluateTemplateForPrompt(model.EditPromptTemplate, templateFile, model.PromptTemplateData{
-					Input:        i,
-					Instruction:  input.Instruction,
-					SystemPrompt: config.SystemPrompt,
-				})
-				if err == nil {
-					i = templatedInput
-					log.Debug().Msgf("Template found, input modified to: %s", i)
-				}
+			templatedInput, err := evaluator.EvaluateTemplateForPrompt(templates.EditPromptTemplate, *config, templates.PromptTemplateData{
+				Input:        i,
+				Instruction:  input.Instruction,
+				SystemPrompt: config.SystemPrompt,
+			})
+			if err == nil {
+				i = templatedInput
+				log.Debug().Msgf("Template found, input modified to: %s", i)
 			}
 
 			r, tokenUsage, err := ComputeChoices(input, i, config, appConfig, ml, func(s string, c *[]schema.Choice) {

core/http/endpoints/openai/request.go

@@ -304,7 +304,6 @@ func mergeRequestWithConfig(modelFile string, input *schema.OpenAIRequest, cm *c
 		config.LoadOptionThreads(threads),
 		config.LoadOptionContextSize(ctx),
 		config.LoadOptionF16(f16),
-		config.ModelPath(loader.ModelPath),
 	)
 
 	// Set the parameters for the language model prediction

core/http/explorer.go

@@ -7,6 +7,7 @@ import (
 	"github.com/gofiber/fiber/v2/middleware/favicon"
 	"github.com/gofiber/fiber/v2/middleware/filesystem"
 	"github.com/mudler/LocalAI/core/explorer"
+	"github.com/mudler/LocalAI/core/http/middleware"
 	"github.com/mudler/LocalAI/core/http/routes"
 )
 
@@ -22,6 +23,7 @@ func Explorer(db *explorer.Database) *fiber.App {
 
 	app := fiber.New(fiberCfg)
 
+	app.Use(middleware.StripPathPrefix())
 	routes.RegisterExplorerRoutes(app, db)
 
 	httpFS := http.FS(embedDirStatic)

core/http/middleware/auth.go

@@ -1,95 +1,98 @@
-package middleware
-
-import (
-	"crypto/subtle"
-	"errors"
-
-	"github.com/dave-gray101/v2keyauth"
-	"github.com/gofiber/fiber/v2"
-	"github.com/gofiber/fiber/v2/middleware/keyauth"
-	"github.com/microcosm-cc/bluemonday"
-	"github.com/mudler/LocalAI/core/config"
-)
-
-// This file contains the configuration generators and handler functions that are used along with the fiber/keyauth middleware
-// Currently this requires an upstream patch - and feature patches are no longer accepted to v2
-// Therefore `dave-gray101/v2keyauth` contains the v2 backport of the middleware until v3 stabilizes and we migrate.
-
-func GetKeyAuthConfig(applicationConfig *config.ApplicationConfig) (*v2keyauth.Config, error) {
-	customLookup, err := v2keyauth.MultipleKeySourceLookup([]string{"header:Authorization", "header:x-api-key", "header:xi-api-key"}, keyauth.ConfigDefault.AuthScheme)
-	if err != nil {
-		return nil, err
-	}
-
-	return &v2keyauth.Config{
-		CustomKeyLookup: customLookup,
-		Next:            getApiKeyRequiredFilterFunction(applicationConfig),
-		Validator:       getApiKeyValidationFunction(applicationConfig),
-		ErrorHandler:    getApiKeyErrorHandler(applicationConfig),
-		AuthScheme:      "Bearer",
-	}, nil
-}
-
-func getApiKeyErrorHandler(applicationConfig *config.ApplicationConfig) fiber.ErrorHandler {
-	return func(ctx *fiber.Ctx, err error) error {
-		if errors.Is(err, v2keyauth.ErrMissingOrMalformedAPIKey) {
-			if len(applicationConfig.ApiKeys) == 0 {
-				return ctx.Next() // if no keys are set up, any error we get here is not an error.
-			}
-			if applicationConfig.OpaqueErrors {
-				return ctx.SendStatus(403)
-			}
-			return ctx.Status(403).SendString(bluemonday.StrictPolicy().Sanitize(err.Error()))
-		}
-		if applicationConfig.OpaqueErrors {
-			return ctx.SendStatus(500)
-		}
-		return err
-	}
-}
-
-func getApiKeyValidationFunction(applicationConfig *config.ApplicationConfig) func(*fiber.Ctx, string) (bool, error) {
-
-	if applicationConfig.UseSubtleKeyComparison {
-		return func(ctx *fiber.Ctx, apiKey string) (bool, error) {
-			if len(applicationConfig.ApiKeys) == 0 {
-				return true, nil // If no keys are setup, accept everything
-			}
-			for _, validKey := range applicationConfig.ApiKeys {
-				if subtle.ConstantTimeCompare([]byte(apiKey), []byte(validKey)) == 1 {
-					return true, nil
-				}
-			}
-			return false, v2keyauth.ErrMissingOrMalformedAPIKey
-		}
-	}
-
-	return func(ctx *fiber.Ctx, apiKey string) (bool, error) {
-		if len(applicationConfig.ApiKeys) == 0 {
-			return true, nil // If no keys are setup, accept everything
-		}
-		for _, validKey := range applicationConfig.ApiKeys {
-			if apiKey == validKey {
-				return true, nil
-			}
-		}
-		return false, v2keyauth.ErrMissingOrMalformedAPIKey
-	}
-}
-
-func getApiKeyRequiredFilterFunction(applicationConfig *config.ApplicationConfig) func(*fiber.Ctx) bool {
-	if applicationConfig.DisableApiKeyRequirementForHttpGet {
-		return func(c *fiber.Ctx) bool {
-			if c.Method() != "GET" {
-				return false
-			}
-			for _, rx := range applicationConfig.HttpGetExemptedEndpoints {
-				if rx.MatchString(c.Path()) {
-					return true
-				}
-			}
-			return false
-		}
-	}
-	return func(c *fiber.Ctx) bool { return false }
-}
+package middleware
+
+import (
+	"crypto/subtle"
+	"errors"
+
+	"github.com/dave-gray101/v2keyauth"
+	"github.com/gofiber/fiber/v2"
+	"github.com/gofiber/fiber/v2/middleware/keyauth"
+	"github.com/mudler/LocalAI/core/config"
+	"github.com/mudler/LocalAI/core/http/utils"
+)
+
+// This file contains the configuration generators and handler functions that are used along with the fiber/keyauth middleware
+// Currently this requires an upstream patch - and feature patches are no longer accepted to v2
+// Therefore `dave-gray101/v2keyauth` contains the v2 backport of the middleware until v3 stabilizes and we migrate.
+
+func GetKeyAuthConfig(applicationConfig *config.ApplicationConfig) (*v2keyauth.Config, error) {
+	customLookup, err := v2keyauth.MultipleKeySourceLookup([]string{"header:Authorization", "header:x-api-key", "header:xi-api-key", "cookie:token"}, keyauth.ConfigDefault.AuthScheme)
+	if err != nil {
+		return nil, err
+	}
+
+	return &v2keyauth.Config{
+		CustomKeyLookup: customLookup,
+		Next:            getApiKeyRequiredFilterFunction(applicationConfig),
+		Validator:       getApiKeyValidationFunction(applicationConfig),
+		ErrorHandler:    getApiKeyErrorHandler(applicationConfig),
+		AuthScheme:      "Bearer",
+	}, nil
+}
+
+func getApiKeyErrorHandler(applicationConfig *config.ApplicationConfig) fiber.ErrorHandler {
+	return func(ctx *fiber.Ctx, err error) error {
+		if errors.Is(err, v2keyauth.ErrMissingOrMalformedAPIKey) {
+			if len(applicationConfig.ApiKeys) == 0 {
+				return ctx.Next() // if no keys are set up, any error we get here is not an error.
+			}
+			ctx.Set("WWW-Authenticate", "Bearer")
+			if applicationConfig.OpaqueErrors {
+				return ctx.SendStatus(401)
+			}
+			return ctx.Status(401).Render("views/login", fiber.Map{
+				"BaseURL": utils.BaseURL(ctx),
+			})
+		}
+		if applicationConfig.OpaqueErrors {
+			return ctx.SendStatus(500)
+		}
+		return err
+	}
+}
+
+func getApiKeyValidationFunction(applicationConfig *config.ApplicationConfig) func(*fiber.Ctx, string) (bool, error) {
+
+	if applicationConfig.UseSubtleKeyComparison {
+		return func(ctx *fiber.Ctx, apiKey string) (bool, error) {
+			if len(applicationConfig.ApiKeys) == 0 {
+				return true, nil // If no keys are setup, accept everything
+			}
+			for _, validKey := range applicationConfig.ApiKeys {
+				if subtle.ConstantTimeCompare([]byte(apiKey), []byte(validKey)) == 1 {
+					return true, nil
+				}
+			}
+			return false, v2keyauth.ErrMissingOrMalformedAPIKey
+		}
+	}
+
+	return func(ctx *fiber.Ctx, apiKey string) (bool, error) {
+		if len(applicationConfig.ApiKeys) == 0 {
+			return true, nil // If no keys are setup, accept everything
+		}
+		for _, validKey := range applicationConfig.ApiKeys {
+			if apiKey == validKey {
+				return true, nil
+			}
+		}
+		return false, v2keyauth.ErrMissingOrMalformedAPIKey
+	}
+}
+
+func getApiKeyRequiredFilterFunction(applicationConfig *config.ApplicationConfig) func(*fiber.Ctx) bool {
+	if applicationConfig.DisableApiKeyRequirementForHttpGet {
+		return func(c *fiber.Ctx) bool {
+			if c.Method() != "GET" {
+				return false
+			}
+			for _, rx := range applicationConfig.HttpGetExemptedEndpoints {
+				if rx.MatchString(c.Path()) {
+					return true
+				}
+			}
+			return false
+		}
+	}
+	return func(c *fiber.Ctx) bool { return false }
+}

core/http/middleware/strippathprefix.go

@@ -0,0 +1,36 @@
+package middleware
+
+import (
+	"strings"
+
+	"github.com/gofiber/fiber/v2"
+)
+
+// StripPathPrefix returns a middleware that strips a path prefix from the request path.
+// The path prefix is obtained from the X-Forwarded-Prefix HTTP request header.
+func StripPathPrefix() fiber.Handler {
+	return func(c *fiber.Ctx) error {
+		for _, prefix := range c.GetReqHeaders()["X-Forwarded-Prefix"] {
+			if prefix != "" {
+				path := c.Path()
+				pos := len(prefix)
+
+				if prefix[pos-1] == '/' {
+					pos--
+				} else {
+					prefix += "/"
+				}
+
+				if strings.HasPrefix(path, prefix) {
+					c.Path(path[pos:])
+					break
+				} else if prefix[:pos] == path {
+					c.Redirect(prefix)
+					return nil
+				}
+			}
+		}
+
+		return c.Next()
+	}
+}

core/http/middleware/strippathprefix_test.go

@@ -0,0 +1,121 @@
+package middleware
+
+import (
+	"net/http/httptest"
+	"testing"
+
+	"github.com/gofiber/fiber/v2"
+	"github.com/stretchr/testify/require"
+)
+
+func TestStripPathPrefix(t *testing.T) {
+	var actualPath string
+
+	app := fiber.New()
+
+	app.Use(StripPathPrefix())
+
+	app.Get("/hello/world", func(c *fiber.Ctx) error {
+		actualPath = c.Path()
+		return nil
+	})
+
+	app.Get("/", func(c *fiber.Ctx) error {
+		actualPath = c.Path()
+		return nil
+	})
+
+	for _, tc := range []struct {
+		name         string
+		path         string
+		prefixHeader []string
+		expectStatus int
+		expectPath   string
+	}{
+		{
+			name:         "without prefix and header",
+			path:         "/hello/world",
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "without prefix and headers on root path",
+			path:         "/",
+			expectStatus: 200,
+			expectPath:   "/",
+		},
+		{
+			name:         "without prefix but header",
+			path:         "/hello/world",
+			prefixHeader: []string{"/otherprefix/"},
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "with prefix but non-matching header",
+			path:         "/prefix/hello/world",
+			prefixHeader: []string{"/otherprefix/"},
+			expectStatus: 404,
+		},
+		{
+			name:         "with prefix and matching header",
+			path:         "/myprefix/hello/world",
+			prefixHeader: []string{"/myprefix/"},
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "with prefix and 1st header matching",
+			path:         "/myprefix/hello/world",
+			prefixHeader: []string{"/myprefix/", "/otherprefix/"},
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "with prefix and 2nd header matching",
+			path:         "/myprefix/hello/world",
+			prefixHeader: []string{"/otherprefix/", "/myprefix/"},
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "with prefix and header not ending with slash",
+			path:         "/myprefix/hello/world",
+			prefixHeader: []string{"/myprefix"},
+			expectStatus: 200,
+			expectPath:   "/hello/world",
+		},
+		{
+			name:         "with prefix and non-matching header not ending with slash",
+			path:         "/myprefix-suffix/hello/world",
+			prefixHeader: []string{"/myprefix"},
+			expectStatus: 404,
+		},
+		{
+			name:         "redirect when prefix does not end with a slash",
+			path:         "/myprefix",
+			prefixHeader: []string{"/myprefix"},
+			expectStatus: 302,
+			expectPath:   "/myprefix/",
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			actualPath = ""
+			req := httptest.NewRequest("GET", tc.path, nil)
+			if tc.prefixHeader != nil {
+				req.Header["X-Forwarded-Prefix"] = tc.prefixHeader
+			}
+
+			resp, err := app.Test(req, -1)
+
+			require.NoError(t, err)
+			require.Equal(t, tc.expectStatus, resp.StatusCode, "response status code")
+
+			if tc.expectStatus == 200 {
+				require.Equal(t, tc.expectPath, actualPath, "rewritten path")
+			} else if tc.expectStatus == 302 {
+				require.Equal(t, tc.expectPath, resp.Header.Get("Location"), "redirect location")
+			}
+		})
+	}
+}