feat: unify and propagate CMAKE_ARGS to GGML-based backends

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

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

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

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

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

.gitattributes

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

.github/labeler.yml

@@ -1,6 +1,11 @@
 enhancements:
  - head-branch: ['^feature', 'feature']
 
+dependencies:
+- any:
+  - changed-files:
+    - any-glob-to-any-file: 'Makefile'
+
 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"
+          - repository: "PABannier/bark.cpp"
-            variable: "BERT_VERSION"
+            variable: "BARKCPP_VERSION"
-            branch: "master"
-          - repository: "go-skynet/bloomz.cpp"
-            variable: "BLOOMZ_VERSION"
             branch: "main"
-          - repository: "mudler/go-ggllm.cpp"
+          - repository: "leejet/stable-diffusion.cpp"
-            variable: "GOGGLLM_VERSION"
+            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/test-extra.yml

@@ -123,6 +123,13 @@ jobs:
         run: |
            make --jobs=5 --output-sync=target -C backend/python/parler-tts
            make --jobs=5 --output-sync=target -C backend/python/parler-tts test
+      - name: Setup tmate session if tests fail
+        if: ${{ failure() }}
+        uses: mxschmitt/action-tmate@v3.19
+        with:
+          detached: true
+          connect-timeout-seconds: 180
+          limit-access-to-actor: true
 
   tests-openvoice:
     runs-on: ubuntu-latest

.github/workflows/test.yml

@@ -224,7 +224,7 @@ jobs:
       - name: Dependencies
         run: |
           brew install protobuf grpc make protoc-gen-go protoc-gen-go-grpc libomp llvm
-          pip install --user --no-cache-dir grpcio-tools==1.64.1
+          pip install --user --no-cache-dir grpcio-tools
       - name: Test
         run: |
           export C_INCLUDE_PATH=/usr/local/include

.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

@@ -85,7 +85,8 @@ WORKDIR /build
 # The requirements-extras target is for any builds with IMAGE_TYPE=extras. It should not be placed in this target unless every IMAGE_TYPE=extras build will use it
 FROM requirements-core AS requirements-extras
 
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+# Install uv as a system package
+RUN curl -LsSf https://astral.sh/uv/install.sh | UV_INSTALL_DIR=/usr/bin sh
 ENV PATH="/root/.cargo/bin:${PATH}"
 
 RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y

Makefile

@@ -8,23 +8,15 @@ DETECT_LIBS?=true
 # llama.cpp versions
 GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
 GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
-CPPLLAMA_VERSION?=45f097645efb11b6d09a5b4adbbfd7c312ac0126
+CPPLLAMA_VERSION?=dafae66cc242eb766797194d3c85c5e502625623
-
-# go-rwkv version
-RWKV_REPO?=https://github.com/donomii/go-rwkv.cpp
-RWKV_VERSION?=661e7ae26d442f5cfebd2a0881b44e8c55949ec6
 
 # whisper.cpp version
 WHISPER_REPO?=https://github.com/ggerganov/whisper.cpp
-WHISPER_CPP_VERSION?=a5abfe6a90495f7bf19fe70d016ecc255e97359c
+WHISPER_CPP_VERSION?=6266a9f9e56a5b925e9892acf650f3eb1245814d
-
-# bert.cpp version
-BERT_REPO?=https://github.com/go-skynet/go-bert.cpp
-BERT_VERSION?=710044b124545415f555e4260d16b146c725a6e4
 
 # go-piper version
 PIPER_REPO?=https://github.com/mudler/go-piper
-PIPER_VERSION?=9d0100873a7dbb0824dfea40e8cec70a1b110759
+PIPER_VERSION?=e10ca041a885d4a8f3871d52924b47792d5e5aa0
 
 # stablediffusion version
 STABLEDIFFUSION_REPO?=https://github.com/mudler/go-stable-diffusion
@@ -34,8 +26,21 @@ 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?=9578fdcc4632dc3de5565f28e2fb16b7c18f8d48
+
+ONNX_VERSION?=1.20.0
+ONNX_ARCH?=x64
+ONNX_OS?=linux
+
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
+export GGML_CMAKE_ARGS?=
 export CMAKE_ARGS?=
 export BACKEND_LIBS?=
 
@@ -45,6 +50,7 @@ CGO_LDFLAGS_WHISPER+=-lggml
 CUDA_LIBPATH?=/usr/local/cuda/lib64/
 GO_TAGS?=
 BUILD_ID?=
+NATIVE?=false
 
 TEST_DIR=/tmp/test
 
@@ -83,7 +89,61 @@ ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
 
+# IF native is false, we add -DGGML_NATIVE=OFF to GGML_CMAKE_ARGS
+ifeq ($(NATIVE),false)
+	GGML_CMAKE_ARGS+=-DGGML_NATIVE=OFF
+endif
+
+# Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
+GGML_CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
+
+# If build type is cublas, then we set -DGGML_CUDA=ON to GGML_CMAKE_ARGS automatically
+ifeq ($(BUILD_TYPE),cublas)
+	GGML_CMAKE_ARGS+=-DGGML_CUDA=ON
+# If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+# to GGML_CMAKE_ARGS automatically
+else ifeq ($(BUILD_TYPE),openblas)
+	GGML_CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+# If build type is clblas (openCL) we set -DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
+else ifeq ($(BUILD_TYPE),clblas)
+	GGML_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 ($(BUILD_TYPE),hipblas)
+	GGML_CMAKE_ARGS+=-DGGML_HIP=ON
+# If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
+# But if it's OSX without metal, disable it here
+else ifeq ($(OS),Darwin)
+	ifneq ($(BUILD_TYPE),metal)
+		GGML_CMAKE_ARGS+=-DGGML_METAL=OFF
+	else
+		GGML_CMAKE_ARGS+=-DGGML_METAL=ON
+		GGML_CMAKE_ARGS+=-DGGML_METAL_EMBED_LIBRARY=ON
+		TARGET+=--target ggml-metal
+	endif
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f16)
+	GGML_CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f32)
+	GGML_CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+endif
+
+# Detect if we are running on arm64
+ifneq (,$(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 := $(shell security find-identity -v -p codesigning | grep '"' | head -n 1 | sed -E 's/.*"(.*)"/\1/')
@@ -94,7 +154,7 @@ ifeq ($(OS),Darwin)
 		BUILD_TYPE=metal
 	# disable metal if on Darwin and any other value is explicitly passed.
 	else ifneq ($(BUILD_TYPE),metal)
-		CMAKE_ARGS+=-DGGML_METAL=OFF
+		GGML_CMAKE_ARGS+=-DGGML_METAL=OFF
 		export GGML_NO_ACCELERATE=1
 		export GGML_NO_METAL=1
 	endif
@@ -119,7 +179,7 @@ ifeq ($(BUILD_TYPE),cublas)
 endif
 
 ifeq ($(BUILD_TYPE),vulkan)
-	CMAKE_ARGS+=-DGGML_VULKAN=1
+	GGML_CMAKE_ARGS+=-DGGML_VULKAN=1
 endif
 
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
@@ -138,10 +198,10 @@ 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="$(GPU_TARGETS)"
+	GGML_CMAKE_ARGS+=-DGGML_HIP=ON -DAMDGPU_TARGETS="$(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 +239,23 @@ 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))
@@ -205,23 +272,12 @@ ifeq ($(BUILD_API_ONLY),true)
 	GRPC_BACKENDS=
 endif
 
+export CMAKE_ARGS?=$(GGML_CMAKE_ARGS)
+
 .PHONY: all test build vendor get-sources prepare-sources prepare
 
 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 +291,23 @@ 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 +321,7 @@ sources/go-piper:
 sources/go-piper/libpiper_binding.a: sources/go-piper
 	$(MAKE) -C sources/go-piper libpiper_binding.a example/main piper.o
 
-
+## stable diffusion (onnx)
-## 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 $(RWKV_VERSION) && \
-	git submodule update --init --recursive --depth 1 --single-branch
-
-sources/go-rwkv.cpp/librwkv.a: sources/go-rwkv.cpp
-	cd sources/go-rwkv.cpp && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
-
-## stable diffusion
 sources/go-stable-diffusion:
 	mkdir -p sources/go-stable-diffusion
 	cd sources/go-stable-diffusion && \
@@ -275,6 +334,44 @@ sources/go-stable-diffusion:
 sources/go-stable-diffusion/libstablediffusion.a: sources/go-stable-diffusion
 	CPATH="$(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 +398,19 @@ 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 +423,8 @@ 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 +439,9 @@ 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 +532,6 @@ 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 +784,6 @@ 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 +843,6 @@ 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 +855,7 @@ 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 +890,13 @@ 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 "$(GO_TAGS)" -o backend-assets/grpc/piper ./backend/go/tts/
@@ -817,13 +904,6 @@ ifneq ($(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 "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion ./backend/go/image/stablediffusion
@@ -831,6 +911,13 @@ ifneq ($(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 +978,7 @@ 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 +986,7 @@ 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

@@ -38,9 +38,13 @@
 </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/go-skynet/LocalAI/tree/master/examples/) 
+> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) 
 
 [![tests](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml)[![Build and Release](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml)[![build container images](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml)[![Bump dependencies](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml)[![Artifact Hub](https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/localai)](https://artifacthub.io/packages/search?repo=localai)
 
@@ -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 +92,10 @@ 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
 - June 2024: 🆕 You can browse now the model gallery without LocalAI! Check out https://models.localai.io
@@ -155,6 +166,9 @@ 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/
   
@@ -229,7 +243,6 @@ 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
@@ -219,6 +221,7 @@ message ModelOptions {
   int32  SwapSpace = 53;
   int32  MaxModelLen = 54;
   int32  TensorParallelSize = 55;
+  string LoadFormat = 58;
 
   string MMProj = 41;
 
@@ -232,6 +235,16 @@ message ModelOptions {
 
   bool FlashAttention = 56;
   bool NoKVOffload = 57;
+
+  string ModelPath = 59;
+
+  repeated string LoraAdapters = 60;
+  repeated float LoraScales = 61;
+
+  repeated string Options = 62;
+
+  string CacheTypeKey = 63;
+  string CacheTypeValue = 64;
 }
 
 message Result {
@@ -287,6 +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

@@ -7,44 +7,6 @@ BUILD_TYPE?=
 ONEAPI_VARS?=/opt/intel/oneapi/setvars.sh
 TARGET?=--target grpc-server
 
-# Disable Shared libs as we are linking on static gRPC and we can't mix shared and static
-CMAKE_ARGS+=-DBUILD_SHARED_LIBS=OFF
-
-# If build type is cublas, then we set -DGGML_CUDA=ON to CMAKE_ARGS automatically
-ifeq ($(BUILD_TYPE),cublas)
-	CMAKE_ARGS+=-DGGML_CUDA=ON
-# If build type is openblas then we set -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
-# to CMAKE_ARGS automatically
-else ifeq ($(BUILD_TYPE),openblas)
-	CMAKE_ARGS+=-DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
-# If build type is clblas (openCL) we set -DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
-else ifeq ($(BUILD_TYPE),clblas)
-	CMAKE_ARGS+=-DGGML_CLBLAST=ON -DCLBlast_DIR=/some/path
-# If it's hipblas we do have also to set CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ 
-else ifeq ($(BUILD_TYPE),hipblas)
-	CMAKE_ARGS+=-DGGML_HIPBLAS=ON
-# If it's OSX, DO NOT embed the metal library - -DGGML_METAL_EMBED_LIBRARY=ON requires further investigation
-# But if it's OSX without metal, disable it here
-else ifeq ($(OS),Darwin)
-	ifneq ($(BUILD_TYPE),metal)
-		CMAKE_ARGS+=-DGGML_METAL=OFF
-	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
-		TARGET+=--target ggml-metal
-	endif
-endif
-
-ifeq ($(BUILD_TYPE),sycl_f16)
-	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
-endif
-
-ifeq ($(BUILD_TYPE),sycl_f32)
-	CMAKE_ARGS+=-DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-endif
-
 llama.cpp:
 	mkdir -p llama.cpp
 	cd llama.cpp && \
@@ -72,7 +34,7 @@ clean: purge
 	rm -rf llama.cpp
 
 grpc-server: llama.cpp llama.cpp/examples/grpc-server
-	@echo "Building grpc-server with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
+	@echo "Building grpc-server for llama.cpp with $(BUILD_TYPE) build type and $(CMAKE_ARGS)"
 ifneq (,$(findstring sycl,$(BUILD_TYPE)))
 	+bash -c "source $(ONEAPI_VARS); \
 	cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release $(TARGET)"

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
@@ -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);
@@ -670,7 +670,6 @@ struct llama_server_context
         slot->sparams.top_k             = json_value(data, "top_k",             default_sparams.top_k);
         slot->sparams.top_p             = json_value(data, "top_p",             default_sparams.top_p);
         slot->sparams.min_p             = json_value(data, "min_p",             default_sparams.min_p);
-        slot->sparams.tfs_z             = json_value(data, "tfs_z",             default_sparams.tfs_z);
         slot->sparams.typ_p             = json_value(data, "typical_p",         default_sparams.typ_p);
         slot->sparams.temp              = json_value(data, "temperature",       default_sparams.temp);
         slot->sparams.dynatemp_range    = json_value(data, "dynatemp_range",    default_sparams.dynatemp_range);
@@ -1206,7 +1205,6 @@ struct llama_server_context
             {"top_k",             slot.sparams.top_k},
             {"top_p",             slot.sparams.top_p},
             {"min_p",             slot.sparams.min_p},
-            {"tfs_z",             slot.sparams.tfs_z},
             {"typical_p",         slot.sparams.typ_p},
             {"repeat_last_n",     slot.sparams.penalty_last_n},
             {"repeat_penalty",    slot.sparams.penalty_repeat},
@@ -2105,7 +2103,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     //     slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
     //     slot->sparams.top_k           = json_value(data, "top_k",             default_sparams.top_k);
     //     slot->sparams.top_p           = json_value(data, "top_p",             default_sparams.top_p);
-    //     slot->sparams.tfs_z           = json_value(data, "tfs_z",             default_sparams.tfs_z);
     //     slot->sparams.typical_p       = json_value(data, "typical_p",         default_sparams.typical_p);
     //     slot->sparams.temp            = json_value(data, "temperature",       default_sparams.temp);
     //     slot->sparams.penalty_last_n  = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
@@ -2129,7 +2126,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
     data["n_predict"] = predict->tokens() == 0 ? -1 : predict->tokens();
     data["top_k"] = predict->topk();
     data["top_p"] = predict->topp();
-    data["tfs_z"] = predict->tailfreesamplingz();
     data["typical_p"] = predict->typicalp();
     data["temperature"] = predict->temperature();
     data["repeat_last_n"] = predict->repeat();
@@ -2176,7 +2172,6 @@ json parse_options(bool streaming, const backend::PredictOptions* predict, llama
 //     llama.params.n_predict = predict->tokens() == 0 ? -1 : predict->tokens();
 //     llama.params.sparams.top_k = predict->topk();
 //     llama.params.sparams.top_p = predict->topp();
-//     llama.params.sparams.tfs_z = predict->tailfreesamplingz();
 //     llama.params.sparams.typical_p = predict->typicalp();
 //     llama.params.sparams.penalty_last_n = predict->repeat();
 //     llama.params.sparams.temp = predict->temperature();
@@ -2246,6 +2241,12 @@ static void params_parse(const backend::ModelOptions* request,
     }
     //  params.model_alias ??
     params.model_alias =  request->modelfile();
+    if (!request->cachetypekey().empty()) {
+        params.cache_type_k = request->cachetypekey();
+    }
+    if (!request->cachetypevalue().empty()) {
+        params.cache_type_v = request->cachetypevalue();
+    }
     params.n_ctx = request->contextsize();
     //params.memory_f16 = request->f16memory();
     params.cpuparams.n_threads = request->threads();
@@ -2304,6 +2305,7 @@ 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/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.txt

@@ -1,6 +1,6 @@
 accelerate
 auto-gptq==0.7.1
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi
 transformers

backend/python/bark/requirements.txt

@@ -1,4 +1,4 @@
 bark==0.1.5
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi

backend/python/common/template/Makefile

@@ -1,8 +1,9 @@
 .DEFAULT_GOAL := install
 
 .PHONY: install
-install: protogen
+install:
 	bash install.sh
+	$(MAKE) protogen
 
 .PHONY: protogen
 protogen: backend_pb2_grpc.py backend_pb2.py
@@ -12,7 +13,7 @@ protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
 backend_pb2_grpc.py backend_pb2.py:
-	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+	bash protogen.sh
 
 .PHONY: clean
 clean: protogen-clean

backend/python/common/template/protogen.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto

backend/python/common/template/requirements.txt

@@ -1,2 +1,3 @@
-grpcio==1.67.0
+grpcio==1.68.1
-protobuf
+protobuf
+grpcio-tools

backend/python/coqui/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi
 packaging==24.1

backend/python/diffusers/backend.py

@@ -247,11 +247,16 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                         use_safetensors=True,
                         variant=variant)
             elif request.PipelineType == "FluxPipeline":
+                if fromSingleFile:
+                    self.pipe = FluxPipeline.from_single_file(modelFile,
+                                                              torch_dtype=torchType,
+                                                              use_safetensors=True)
+                else:
                     self.pipe = FluxPipeline.from_pretrained(
                         request.Model,
                         torch_dtype=torch.bfloat16)
-                    if request.LowVRAM:
+                if request.LowVRAM:
-                        self.pipe.enable_model_cpu_offload()
+                    self.pipe.enable_model_cpu_offload()
             elif request.PipelineType == "FluxTransformer2DModel":
                     dtype = torch.bfloat16
                     # specify from environment or default to "ChuckMcSneed/FLUX.1-dev"
@@ -296,22 +301,34 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                 self.pipe.controlnet = self.controlnet
             else:
                 self.controlnet = None
-            # Assume directory from request.ModelFile.
+
-            # Only if request.LoraAdapter it's not an absolute path
+            if request.LoraAdapter and not os.path.isabs(request.LoraAdapter):
-            if request.LoraAdapter and request.ModelFile != "" and not os.path.isabs(request.LoraAdapter) and request.LoraAdapter:
-                # get base path of modelFile
-                modelFileBase = os.path.dirname(request.ModelFile)
                 # modify LoraAdapter to be relative to modelFileBase
-                request.LoraAdapter = os.path.join(modelFileBase, request.LoraAdapter)
+                request.LoraAdapter = os.path.join(request.ModelPath, request.LoraAdapter)
+
             device = "cpu" if not request.CUDA else "cuda"
             self.device = device
             if request.LoraAdapter:
                 # Check if its a local file and not a directory ( we load lora differently for a safetensor file )
                 if os.path.exists(request.LoraAdapter) and not os.path.isdir(request.LoraAdapter):
-                    # self.load_lora_weights(request.LoraAdapter, 1, device, torchType)
                     self.pipe.load_lora_weights(request.LoraAdapter)
                 else:
                     self.pipe.unet.load_attn_procs(request.LoraAdapter)
+            if len(request.LoraAdapters) > 0:
+                i = 0
+                adapters_name = []
+                adapters_weights = []
+                for adapter in request.LoraAdapters:
+                    if not os.path.isabs(adapter):
+                        adapter = os.path.join(request.ModelPath, adapter)
+                    self.pipe.load_lora_weights(adapter, adapter_name=f"adapter_{i}")
+                    adapters_name.append(f"adapter_{i}")
+                    i += 1
+
+                for adapters_weight in request.LoraScales:
+                    adapters_weights.append(adapters_weight)
+
+                self.pipe.set_adapters(adapters_name, adapter_weights=adapters_weights)
 
             if request.CUDA:
                 self.pipe.to('cuda')
@@ -392,8 +409,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         # create a dictionary of values for the parameters
         options = {
             "negative_prompt": request.negative_prompt,
-            "width": request.width,
-            "height": request.height,
             "num_inference_steps": steps,
         }
 
@@ -411,13 +426,13 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         keys = options.keys()
 
         if request.EnableParameters != "":
-            keys = request.EnableParameters.split(",")
+            keys = [key.strip() for key in request.EnableParameters.split(",")]
 
         if request.EnableParameters == "none":
             keys = []
 
         # create a dictionary of parameters by using the keys from EnableParameters and the values from defaults
-        kwargs = {key: options[key] for key in keys}
+        kwargs = {key: options.get(key) for key in keys if key in options}
 
         # Set seed
         if request.seed > 0:
@@ -428,6 +443,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
         if self.PipelineType == "FluxPipeline":
             kwargs["max_sequence_length"] = 256
 
+        if request.width:
+            kwargs["width"] = request.width
+
+        if request.height:
+            kwargs["height"] = request.height
+
         if self.PipelineType == "FluxTransformer2DModel":
             kwargs["output_type"] = "pil"
             kwargs["generator"] = torch.Generator("cpu").manual_seed(0)
@@ -447,6 +468,7 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
             export_to_video(video_frames, request.dst)
             return backend_pb2.Result(message="Media generated successfully", success=True)
 
+        print(f"Generating image with {kwargs=}", file=sys.stderr)
         image = {}
         if COMPEL:
             conditioning, pooled = self.compel.build_conditioning_tensor(prompt)

backend/python/diffusers/requirements.txt

@@ -1,5 +1,5 @@
 setuptools
-grpcio==1.67.0
+grpcio==1.68.1
 pillow
 protobuf
 certifi

backend/python/exllama2/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi
 wheel

backend/python/mamba/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi

backend/python/openvoice/requirements-intel.txt

@@ -2,7 +2,7 @@
 intel-extension-for-pytorch
 torch
 optimum[openvino]
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 librosa==0.9.1
 faster-whisper==0.9.0

backend/python/openvoice/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 librosa
 faster-whisper

backend/python/parler-tts/Makefile

@@ -12,9 +12,10 @@ export SKIP_CONDA=1
 endif
 
 .PHONY: parler-tts
-parler-tts: protogen
+parler-tts:
 	@echo "Installing $(CONDA_ENV_PATH)..."
 	bash install.sh $(CONDA_ENV_PATH)
+	$(MAKE) protogen
 
 .PHONY: run
 run: protogen
@@ -36,7 +37,7 @@ protogen-clean:
 	$(RM) backend_pb2_grpc.py backend_pb2.py
 
 backend_pb2_grpc.py backend_pb2.py:
-	python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto
+	bash protogen.sh
 
 .PHONY: clean
 clean: protogen-clean

backend/python/parler-tts/install.sh

@@ -11,8 +11,10 @@ if [ "x${BUILD_PROFILE}" == "xintel" ]; then
     EXTRA_PIP_INSTALL_FLAGS+=" --upgrade --index-strategy=unsafe-first-match"
 fi
 
+
 installRequirements
 
+
 # https://github.com/descriptinc/audiotools/issues/101
 # incompatible protobuf versions.
 PYDIR=python3.10

backend/python/parler-tts/protogen.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+source $(dirname $0)/../common/libbackend.sh
+
+python3 -m grpc_tools.protoc -I../.. --python_out=. --grpc_python_out=. backend.proto

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

@@ -1,3 +1,4 @@
 git+https://github.com/huggingface/parler-tts.git@8e465f1b5fcd223478e07175cb40494d19ffbe17
 llvmlite==0.43.0
 numba==0.60.0
+grpcio-tools==1.42.0

backend/python/parler-tts/requirements.txt

@@ -1,4 +1,3 @@
-grpcio==1.67.0
+grpcio==1.68.1
-protobuf
 certifi
-llvmlite==0.43.0
+llvmlite==0.43.0

backend/python/rerankers/requirements.txt

@@ -1,3 +1,3 @@
-grpcio==1.67.0
+grpcio==1.68.1
 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

@@ -4,5 +4,5 @@ torch
 optimum[openvino]
 setuptools==69.5.1 # https://github.com/mudler/LocalAI/issues/2406
 accelerate
-sentence-transformers==3.2.0
+sentence-transformers==3.3.1
 transformers

backend/python/sentencetransformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi
 datasets

backend/python/transformers-musicgen/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 scipy==1.14.0
 certifi

backend/python/transformers/requirements.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi
 setuptools==69.5.1 # https://github.com/mudler/LocalAI/issues/2406

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

@@ -1,3 +1,3 @@
-grpcio==1.67.0
+grpcio==1.68.1
 protobuf
 certifi

backend/python/vllm/backend.py

@@ -95,6 +95,8 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
         if request.Quantization != "":
             engine_args.quantization = request.Quantization
+        if request.LoadFormat != "":
+            engine_args.load_format = request.LoadFormat
         if request.GPUMemoryUtilization != 0:
             engine_args.gpu_memory_utilization = request.GPUMemoryUtilization
         if request.TrustRemoteCode:

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==1.67.0 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.txt

@@ -1,4 +1,4 @@
-grpcio==1.67.0
+grpcio==1.68.1
 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),
@@ -155,22 +153,17 @@ 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)
 
-			o := backend.ModelOptions(*cfg, options, []model.Option{})
+			o := backend.ModelOptions(*cfg, options)
 
 			var backendErr error
-			if cfg.Backend != "" {
+			_, backendErr = application.ModelLoader().Load(o...)
-				o = append(o, model.WithBackendString(cfg.Backend))
-				_, backendErr = ml.BackendLoader(o...)
-			} else {
-				_, backendErr = ml.GreedyLoader(o...)
-			}
 			if backendErr != nil {
-				return nil, nil, nil, err
+				return nil, err
 			}
 		}
 	}
@@ -179,7 +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) {
@@ -206,32 +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/embeddings.go

@@ -11,17 +11,9 @@ import (
 
 func ModelEmbedding(s string, tokens []int, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (func() ([]float32, error), error) {
 
-	var inferenceModel interface{}
+	opts := ModelOptions(backendConfig, appConfig)
-	var err error
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{})
+	inferenceModel, err := loader.Load(opts...)
-
-	if backendConfig.Backend == "" {
-		inferenceModel, err = loader.GreedyLoader(opts...)
-	} else {
-		opts = append(opts, model.WithBackendString(backendConfig.Backend))
-		inferenceModel, err = loader.BackendLoader(opts...)
-	}
 	if err != nil {
 		return nil, err
 	}

core/backend/image.go

@@ -9,9 +9,8 @@ import (
 
 func ImageGeneration(height, width, mode, step, seed int, positive_prompt, negative_prompt, src, dst string, loader *model.ModelLoader, backendConfig config.BackendConfig, appConfig *config.ApplicationConfig) (func() error, error) {
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{})
+	opts := ModelOptions(backendConfig, appConfig)
-
+	inferenceModel, err := loader.Load(
-	inferenceModel, err := loader.BackendLoader(
 		opts...,
 	)
 	if err != nil {

core/backend/llm.go

@@ -16,7 +16,6 @@ import (
 	"github.com/mudler/LocalAI/core/schema"
 
 	"github.com/mudler/LocalAI/core/gallery"
-	"github.com/mudler/LocalAI/pkg/grpc"
 	"github.com/mudler/LocalAI/pkg/grpc/proto"
 	model "github.com/mudler/LocalAI/pkg/model"
 	"github.com/mudler/LocalAI/pkg/utils"
@@ -35,15 +34,6 @@ type TokenUsage struct {
 func ModelInference(ctx context.Context, s string, messages []schema.Message, images, videos, audios []string, loader *model.ModelLoader, c config.BackendConfig, o *config.ApplicationConfig, tokenCallback func(string, TokenUsage) bool) (func() (LLMResponse, error), error) {
 	modelFile := c.Model
 
-	var inferenceModel grpc.Backend
-	var err error
-
-	opts := ModelOptions(c, o, []model.Option{})
-
-	if c.Backend != "" {
-		opts = append(opts, model.WithBackendString(c.Backend))
-	}
-
 	// Check if the modelFile exists, if it doesn't try to load it from the gallery
 	if o.AutoloadGalleries { // experimental
 		if _, err := os.Stat(modelFile); os.IsNotExist(err) {
@@ -56,12 +46,8 @@ func ModelInference(ctx context.Context, s string, messages []schema.Message, im
 		}
 	}
 
-	if c.Backend == "" {
+	opts := ModelOptions(c, o)
-		inferenceModel, err = loader.GreedyLoader(opts...)
+	inferenceModel, err := loader.Load(opts...)
-	} else {
-		inferenceModel, err = loader.BackendLoader(opts...)
-	}
-
 	if err != nil {
 		return nil, err
 	}

core/backend/options.go

@@ -11,7 +11,7 @@ import (
 	"github.com/rs/zerolog/log"
 )
 
-func ModelOptions(c config.BackendConfig, so *config.ApplicationConfig, opts []model.Option) []model.Option {
+func ModelOptions(c config.BackendConfig, so *config.ApplicationConfig, opts ...model.Option) []model.Option {
 	name := c.Name
 	if name == "" {
 		name = c.Model
@@ -122,14 +122,17 @@ 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,
+		LoraScales:           c.LoraScales,
 		F16Memory:            f16,
 		LoraBase:             c.LoraBase,
 		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),
@@ -139,6 +142,7 @@ func grpcModelOpts(c config.BackendConfig) *pb.ModelOptions {
 		DraftModel:           c.DraftModel,
 		AudioPath:            c.VallE.AudioPath,
 		Quantization:         c.Quantization,
+		LoadFormat:           c.LoadFormat,
 		GPUMemoryUtilization: c.GPUMemoryUtilization,
 		TrustRemoteCode:      c.TrustRemoteCode,
 		EnforceEager:         c.EnforceEager,
@@ -147,6 +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/backend/rerank.go

@@ -11,8 +11,8 @@ import (
 
 func Rerank(modelFile string, request *proto.RerankRequest, loader *model.ModelLoader, appConfig *config.ApplicationConfig, backendConfig config.BackendConfig) (*proto.RerankResult, error) {
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{model.WithModel(modelFile)})
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
-	rerankModel, err := loader.BackendLoader(opts...)
+	rerankModel, err := loader.Load(opts...)
 	if err != nil {
 		return nil, err
 	}

core/backend/soundgeneration.go

@@ -25,9 +25,8 @@ func SoundGeneration(
 	backendConfig config.BackendConfig,
 ) (string, *proto.Result, error) {
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{model.WithModel(modelFile)})
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
-
+	soundGenModel, err := loader.Load(opts...)
-	soundGenModel, err := loader.BackendLoader(opts...)
 	if err != nil {
 		return "", nil, err
 	}

core/backend/stores.go

@@ -8,16 +8,15 @@ import (
 )
 
 func StoreBackend(sl *model.ModelLoader, appConfig *config.ApplicationConfig, storeName string) (grpc.Backend, error) {
-    if storeName == "" {
+	if storeName == "" {
-      storeName = "default"
+		storeName = "default"
-    }
+	}
 
-    sc := []model.Option{
+	sc := []model.Option{
-      model.WithBackendString(model.LocalStoreBackend),
+		model.WithBackendString(model.LocalStoreBackend),
-      model.WithAssetDir(appConfig.AssetsDestination),
+		model.WithAssetDir(appConfig.AssetsDestination),
-      model.WithModel(storeName),
+		model.WithModel(storeName),
-    }
+	}
 
-    return sl.BackendLoader(sc...)
+	return sl.Load(sc...)
 }
-

core/backend/token_metrics.go

@@ -15,10 +15,8 @@ func TokenMetrics(
 	appConfig *config.ApplicationConfig,
 	backendConfig config.BackendConfig) (*proto.MetricsResponse, error) {
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
-		model.WithModel(modelFile),
+	model, err := loader.Load(opts...)
-	})
-	model, err := loader.BackendLoader(opts...)
 	if err != nil {
 		return nil, err
 	}

core/backend/tokenize.go

@@ -14,15 +14,13 @@ func ModelTokenize(s string, loader *model.ModelLoader, backendConfig config.Bac
 	var inferenceModel grpc.Backend
 	var err error
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{
+	opts := ModelOptions(backendConfig, appConfig, model.WithModel(modelFile))
-		model.WithModel(modelFile),
-	})
 
 	if backendConfig.Backend == "" {
-		inferenceModel, err = loader.GreedyLoader(opts...)
+		inferenceModel, err = loader.Load(opts...)
 	} else {
 		opts = append(opts, model.WithBackendString(backendConfig.Backend))
-		inferenceModel, err = loader.BackendLoader(opts...)
+		inferenceModel, err = loader.Load(opts...)
 	}
 	if err != nil {
 		return schema.TokenizeResponse{}, err

core/backend/transcript.go

@@ -18,9 +18,9 @@ func ModelTranscription(audio, language string, translate bool, ml *model.ModelL
 		backendConfig.Backend = model.WhisperBackend
 	}
 
-	opts := ModelOptions(backendConfig, appConfig, []model.Option{})
+	opts := ModelOptions(backendConfig, appConfig)
 
-	transcriptionModel, err := ml.BackendLoader(opts...)
+	transcriptionModel, err := ml.Load(opts...)
 	if err != nil {
 		return nil, err
 	}

core/backend/tts.go

@@ -28,11 +28,8 @@ func ModelTTS(
 		bb = model.PiperBackend
 	}
 
-	opts := ModelOptions(config.BackendConfig{}, appConfig, []model.Option{
+	opts := ModelOptions(backendConfig, appConfig, model.WithBackendString(bb), model.WithModel(modelFile))
-		model.WithBackendString(bb),
+	ttsModel, err := loader.Load(opts...)
-		model.WithModel(modelFile),
-	})
-	ttsModel, err := loader.BackendLoader(opts...)
 	if err != nil {
 		return "", nil, err
 	}

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"
 )
@@ -53,6 +53,7 @@ type RunCMD struct {
 	OpaqueErrors                       bool     `env:"LOCALAI_OPAQUE_ERRORS" default:"false" help:"If true, all error responses are replaced with blank 500 errors. This is intended only for hardening against information leaks and is normally not recommended." group:"hardening"`
 	UseSubtleKeyComparison             bool     `env:"LOCALAI_SUBTLE_KEY_COMPARISON" default:"false" help:"If true, API Key validation comparisons will be performed using constant-time comparisons rather than simple equality. This trades off performance on each request for resiliancy against timing attacks." group:"hardening"`
 	DisableApiKeyRequirementForHttpGet bool     `env:"LOCALAI_DISABLE_API_KEY_REQUIREMENT_FOR_HTTP_GET" default:"false" help:"If true, a valid API key is not required to issue GET requests to portions of the web ui. This should only be enabled in secure testing environments" group:"hardening"`
+	DisableMetricsEndpoint             bool     `env:"LOCALAI_DISABLE_METRICS_ENDPOINT,DISABLE_METRICS_ENDPOINT" default:"false" help:"Disable the /metrics endpoint" group:"api"`
 	HttpGetExemptedEndpoints           []string `env:"LOCALAI_HTTP_GET_EXEMPTED_ENDPOINTS" default:"^/$,^/browse/?$,^/talk/?$,^/p2p/?$,^/chat/?$,^/text2image/?$,^/tts/?$,^/static/.*$,^/swagger.*$" help:"If LOCALAI_DISABLE_API_KEY_REQUIREMENT_FOR_HTTP_GET is overriden to true, this is the list of endpoints to exempt. Only adjust this in case of a security incident or as a result of a personal security posture review" group:"hardening"`
 	Peer2Peer                          bool     `env:"LOCALAI_P2P,P2P" name:"p2p" default:"false" help:"Enable P2P mode" group:"p2p"`
 	Peer2PeerDHTInterval               int      `env:"LOCALAI_P2P_DHT_INTERVAL,P2P_DHT_INTERVAL" default:"360" name:"p2p-dht-interval" help:"Interval for DHT refresh (used during token generation)" group:"p2p"`
@@ -108,6 +109,10 @@ func (r *RunCMD) Run(ctx *cliContext.Context) error {
 		config.WithLoadToMemory(r.LoadToMemory),
 	}
 
+	if r.DisableMetricsEndpoint {
+		opts = append(opts, config.DisableMetricsEndpoint)
+	}
+
 	token := ""
 	if r.Peer2Peer || r.Peer2PeerToken != "" {
 		log.Info().Msg("P2P mode enabled")
@@ -181,16 +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