⬆️ Update ggerganov/llama.cpp (#1207)

Signed-off-by: GitHub <noreply@github.com>
Signed-off-by: Ettore Di Giacinto <mudler@users.noreply.github.com>
Co-authored-by: mudler <mudler@users.noreply.github.com>

.github/workflows/bump_deps.yaml

@@ -44,7 +44,7 @@ jobs:
             branch: "master"
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - name: Bump dependencies 🔧
         run: |
           bash .github/bump_deps.sh ${{ matrix.repository }} ${{ matrix.branch }} ${{ matrix.variable }}

.github/workflows/image.yml

@@ -24,42 +24,16 @@ jobs:
             tag-latest: 'auto'
             tag-suffix: ''
             ffmpeg: ''
-          - build-type: 'cublas'
-            cuda-major-version: 11
-            cuda-minor-version: 7
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11'
-            ffmpeg: ''
-          - build-type: 'cublas'
-            cuda-major-version: 12
-            cuda-minor-version: 1
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12'
-            ffmpeg: ''
           - build-type: ''
             platforms: 'linux/amd64'
             tag-latest: 'false'
             tag-suffix: '-ffmpeg'
             ffmpeg: 'true'
-          - build-type: 'cublas'
-            cuda-major-version: 11
-            cuda-minor-version: 7
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda11-ffmpeg'
-            ffmpeg: 'true'
-          - build-type: 'cublas'
-            cuda-major-version: 12
-            cuda-minor-version: 1
-            platforms: 'linux/amd64'
-            tag-latest: 'false'
-            tag-suffix: '-cublas-cuda12-ffmpeg'
-            ffmpeg: 'true'
 
     runs-on: ubuntu-latest
     steps:
+      - name: Checkout
+        uses: actions/checkout@v4
       - name: Release space from worker
         run: |
           echo "Listing top largest packages"
@@ -83,6 +57,10 @@ jobs:
           sudo apt-get remove -y azure-cli || true
           sudo apt-get remove -y '^mongo.*-.*|^postgresql-.*|^mysql-.*|^mssql-.*' || true
           sudo apt-get remove -y '^gfortran-.*' || true
+          sudo apt-get remove -y microsoft-edge-stable || true
+          sudo apt-get remove -y firefox || true
+          sudo apt-get remove -y powershell || true
+          sudo apt-get remove -y r-base-core || true
           sudo apt-get autoremove -y
           sudo apt-get clean
           echo
@@ -92,9 +70,6 @@ jobs:
           echo
           sudo rm -rfv build || true
           df -h
-      - name: Checkout
-        uses: actions/checkout@v3
-
       - name: Docker meta
         id: meta
         uses: docker/metadata-action@v5
@@ -140,3 +115,99 @@ jobs:
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.meta.outputs.tags }}
           labels: ${{ steps.meta.outputs.labels }}
+
+
+  docker-gpu:
+    strategy:
+      matrix:
+        include:
+          - build-type: 'cublas'
+            cuda-major-version: 11
+            cuda-minor-version: 7
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11'
+            ffmpeg: ''
+          - build-type: 'cublas'
+            cuda-major-version: 12
+            cuda-minor-version: 1
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12'
+            ffmpeg: ''
+          - build-type: 'cublas'
+            cuda-major-version: 11
+            cuda-minor-version: 7
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda11-ffmpeg'
+            ffmpeg: 'true'
+          - build-type: 'cublas'
+            cuda-major-version: 12
+            cuda-minor-version: 1
+            platforms: 'linux/amd64'
+            tag-latest: 'false'
+            tag-suffix: '-cublas-cuda12-ffmpeg'
+            ffmpeg: 'true'
+
+    runs-on: arc-runner-set 
+    steps:
+      - name: Force Install GIT latest
+        run: |
+          sudo apt-get update \
+          && sudo apt-get install -y software-properties-common \
+          && sudo apt-get update \
+          && sudo add-apt-repository -y ppa:git-core/ppa \
+          && sudo apt-get update \
+          && sudo apt-get install -y git
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Docker meta
+        id: meta
+        uses: docker/metadata-action@v5
+        with:
+          images: quay.io/go-skynet/local-ai
+          tags: |
+            type=ref,event=branch
+            type=semver,pattern={{raw}}
+            type=sha
+          flavor: |
+            latest=${{ matrix.tag-latest }}
+            suffix=${{ matrix.tag-suffix }}
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@master
+        with:
+          platforms: all
+
+      - name: Set up Docker Buildx
+        id: buildx
+        uses: docker/setup-buildx-action@master
+
+      - name: Login to DockerHub
+        if: github.event_name != 'pull_request'
+        uses: docker/login-action@v3
+        with:
+          registry: quay.io
+          username: ${{ secrets.LOCALAI_REGISTRY_USERNAME }}
+          password: ${{ secrets.LOCALAI_REGISTRY_PASSWORD }}
+      - name: Build and push
+        uses: docker/build-push-action@v5
+        with:
+          builder: ${{ steps.buildx.outputs.name }}
+          build-args: |
+            BUILD_TYPE=${{ matrix.build-type }}
+            CUDA_MAJOR_VERSION=${{ matrix.cuda-major-version }}
+            CUDA_MINOR_VERSION=${{ matrix.cuda-minor-version }}
+            FFMPEG=${{ matrix.ffmpeg }}
+          context: .
+          file: ./Dockerfile
+          platforms: ${{ matrix.platforms }}
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: ${{ steps.meta.outputs.tags }}
+          labels: ${{ steps.meta.outputs.labels }}
+      - name: Release space from worker ♻
+        if: always()
+        run: |
+          docker system prune -f -a --volumes || true

.github/workflows/release.yaml

@@ -19,7 +19,7 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
       - uses: actions/setup-go@v4
@@ -66,7 +66,7 @@ jobs:
     runs-on: macOS-latest
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           submodules: true
       - uses: actions/setup-go@v4

.github/workflows/test-gpu.yml

@@ -15,13 +15,13 @@ concurrency:
 
 jobs:
   ubuntu-latest:
-    runs-on: self-hosted
+    runs-on: gpu
     strategy:
       matrix:
         go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with: 
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -40,6 +40,8 @@ jobs:
           if [ ! -e /run/systemd/system ]; then
             sudo mkdir /run/systemd/system
           fi
+          sudo mkdir -p /host/tests/${{ github.head_ref || github.ref }}
+          sudo chmod -R 777 /host/tests/${{ github.head_ref || github.ref }}
           make \
             TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
             BUILD_TYPE=cublas \
@@ -57,4 +59,5 @@ jobs:
           make \
             TEST_DIR="/host/tests/${{ github.head_ref || github.ref }}" \
             teardown-e2e || true
-          docker system prune -f -a --volumes || true
+          sudo rm -rf /host/tests/${{ github.head_ref || github.ref }} || true
+          docker system prune -f -a --volumes || true

.github/workflows/test.yml

@@ -53,7 +53,7 @@ jobs:
           sudo rm -rfv build || true
           df -h
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with: 
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}
@@ -108,7 +108,7 @@ jobs:
         go-version: ['1.21.x']
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with: 
           submodules: true
       - name: Setup Go ${{ matrix.go-version }}

Dockerfile

@@ -19,7 +19,7 @@ ENV GALLERIES='[{"name":"model-gallery", "url":"github:go-skynet/model-gallery/i
 ARG GO_TAGS="stablediffusion tts"
 
 RUN apt-get update && \
-    apt-get install -y ca-certificates curl patch pip cmake
+    apt-get install -y ca-certificates curl patch pip cmake && apt-get clean
 
 
 # Use the variables in subsequent instructions
@@ -34,17 +34,18 @@ RUN if [ "${BUILD_TYPE}" = "cublas" ]; then \
     dpkg -i cuda-keyring_1.0-1_all.deb && \
     rm -f cuda-keyring_1.0-1_all.deb && \
     apt-get update && \
-    apt-get install -y cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusparse-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusolver-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
+    apt-get install -y cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcublas-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusparse-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} libcusolver-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION}  && apt-get clean \
     ; fi
 ENV PATH /usr/local/cuda/bin:${PATH}
 
-# OpenBLAS requirements
-RUN apt-get install -y libopenblas-dev
-
-# Stable Diffusion requirements
-RUN apt-get install -y libopencv-dev && \
-    ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
+# OpenBLAS requirements and stable diffusion
+RUN apt-get install -y \
+    libopenblas-dev \
+    libopencv-dev \ 
+    && apt-get clean
 
+# Set up OpenCV
+RUN ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
 
 WORKDIR /build
 
@@ -68,8 +69,7 @@ RUN curl -L "https://github.com/gabime/spdlog/archive/refs/tags/v${SPDLOG_VERSIO
     cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/lib/. /usr/lib/ && \
     ln -s /usr/lib/libpiper_phonemize.so /usr/lib/libpiper_phonemize.so.1 && \
     cp -rfv /build/lib/Linux-$(uname -m)/piper_phonemize/include/. /usr/include/ && \
-    rm spdlog-${SPDLOG_VERSION} -rf && \
-    rm /build/lib/Linux-$(uname -m)/piper_phonemize -rf
+    rm spdlog-${SPDLOG_VERSION} -rf
 
 # Extras requirements
 FROM requirements-core as requirements-extras

Makefile

@@ -4,11 +4,11 @@ GOVET=$(GOCMD) vet
 BINARY_NAME=local-ai
 
 # llama.cpp versions
-GOLLAMA_VERSION?=1676dcd7a139b6cdfbaea5fd67f46dc25d9d8bcf
+GOLLAMA_VERSION?=aeba71ee842819da681ea537e78846dc75949ac0
 
 GOLLAMA_STABLE_VERSION?=50cee7712066d9e38306eccadcfbb44ea87df4b7
 
-CPPLLAMA_VERSION?=24ba3d829e31a6eda3fa1723f692608c2fa3adda
+CPPLLAMA_VERSION?=6e08281e588bbba1a5d180290a94a43f167f3a1a
 
 # gpt4all version
 GPT4ALL_REPO?=https://github.com/nomic-ai/gpt4all
@@ -30,15 +30,9 @@ BERT_VERSION?=6abe312cded14042f6b7c3cd8edf082713334a4d
 # go-piper version
 PIPER_VERSION?=56b8a81b4760a6fbee1a82e62f007ae7e8f010a7
 
-# go-bloomz version
-BLOOMZ_VERSION?=1834e77b83faafe912ad4092ccf7f77937349e2f
-
 # stablediffusion version
 STABLEDIFFUSION_VERSION?=d89260f598afb809279bc72aa0107b4292587632
 
-# Go-ggllm
-GOGGLLM_VERSION?=862477d16eefb0805261c19c9b0d053e3b2b684b
-
 export BUILD_TYPE?=
 export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
 export CMAKE_ARGS?=
@@ -129,7 +123,13 @@ ifeq ($(findstring tts,$(GO_TAGS)),tts)
 	OPTIONAL_GRPC+=backend-assets/grpc/piper
 endif
 
-GRPC_BACKENDS?=backend-assets/grpc/langchain-huggingface backend-assets/grpc/falcon-ggml backend-assets/grpc/bert-embeddings backend-assets/grpc/falcon backend-assets/grpc/bloomz backend-assets/grpc/llama backend-assets/grpc/llama-cpp backend-assets/grpc/llama-stable backend-assets/grpc/gpt4all backend-assets/grpc/dolly backend-assets/grpc/gpt2 backend-assets/grpc/gptj backend-assets/grpc/gptneox backend-assets/grpc/mpt backend-assets/grpc/replit backend-assets/grpc/starcoder backend-assets/grpc/rwkv backend-assets/grpc/whisper $(OPTIONAL_GRPC)
+ALL_GRPC_BACKENDS=backend-assets/grpc/langchain-huggingface backend-assets/grpc/falcon-ggml backend-assets/grpc/bert-embeddings backend-assets/grpc/llama backend-assets/grpc/llama-cpp backend-assets/grpc/llama-stable backend-assets/grpc/gpt4all backend-assets/grpc/dolly backend-assets/grpc/gpt2 backend-assets/grpc/gptj backend-assets/grpc/gptneox backend-assets/grpc/mpt backend-assets/grpc/replit backend-assets/grpc/starcoder backend-assets/grpc/rwkv backend-assets/grpc/whisper $(OPTIONAL_GRPC)
+GRPC_BACKENDS?=$(ALL_GRPC_BACKENDS) $(OPTIONAL_GRPC)
+
+# If empty, then we build all
+ifeq ($(GRPC_BACKENDS),)
+	GRPC_BACKENDS=$(ALL_GRPC_BACKENDS)
+endif
 
 .PHONY: all test build vendor
 
@@ -140,14 +140,6 @@ gpt4all:
 	git clone --recurse-submodules $(GPT4ALL_REPO) gpt4all
 	cd gpt4all && git checkout -b build $(GPT4ALL_VERSION) && git submodule update --init --recursive --depth 1
 
-## go-ggllm
-go-ggllm:
-	git clone --recurse-submodules https://github.com/mudler/go-ggllm.cpp go-ggllm
-	cd go-ggllm && git checkout -b build $(GOGGLLM_VERSION) && git submodule update --init --recursive --depth 1
-
-go-ggllm/libggllm.a: go-ggllm
-	$(MAKE) -C go-ggllm BUILD_TYPE=$(BUILD_TYPE) libggllm.a
-
 ## go-piper
 go-piper:
 	git clone --recurse-submodules https://github.com/mudler/go-piper go-piper
@@ -174,14 +166,6 @@ go-rwkv:
 go-rwkv/librwkv.a: go-rwkv
 	cd go-rwkv && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
 
-## bloomz
-bloomz:
-	git clone --recurse-submodules https://github.com/go-skynet/bloomz.cpp bloomz
-	cd bloomz && git checkout -b build $(BLOOMZ_VERSION) && git submodule update --init --recursive --depth 1
-
-bloomz/libbloomz.a: bloomz
-	cd bloomz && make libbloomz.a
-
 go-bert/libgobert.a: go-bert
 	$(MAKE) -C go-bert libgobert.a
 
@@ -235,7 +219,7 @@ go-llama-stable/libbinding.a: go-llama-stable
 go-piper/libpiper_binding.a: go-piper
 	$(MAKE) -C go-piper libpiper_binding.a example/main
 
-get-sources: go-llama go-llama-stable go-ggllm go-ggml-transformers gpt4all go-piper go-rwkv whisper.cpp go-bert bloomz go-stable-diffusion
+get-sources: go-llama go-llama-stable go-ggml-transformers gpt4all go-piper go-rwkv whisper.cpp go-bert go-stable-diffusion
 	touch $@
 
 replace:
@@ -244,10 +228,8 @@ replace:
 	$(GOCMD) mod edit -replace github.com/donomii/go-rwkv.cpp=$(shell pwd)/go-rwkv
 	$(GOCMD) mod edit -replace github.com/ggerganov/whisper.cpp=$(shell pwd)/whisper.cpp
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-bert.cpp=$(shell pwd)/go-bert
-	$(GOCMD) mod edit -replace github.com/go-skynet/bloomz.cpp=$(shell pwd)/bloomz
 	$(GOCMD) mod edit -replace github.com/mudler/go-stable-diffusion=$(shell pwd)/go-stable-diffusion
 	$(GOCMD) mod edit -replace github.com/mudler/go-piper=$(shell pwd)/go-piper
-	$(GOCMD) mod edit -replace github.com/mudler/go-ggllm.cpp=$(shell pwd)/go-ggllm
 
 prepare-sources: get-sources replace
 	$(GOCMD) mod download
@@ -263,9 +245,7 @@ rebuild: ## Rebuilds the project
 	$(MAKE) -C whisper.cpp clean
 	$(MAKE) -C go-stable-diffusion clean
 	$(MAKE) -C go-bert clean
-	$(MAKE) -C bloomz clean
 	$(MAKE) -C go-piper clean
-	$(MAKE) -C go-ggllm clean
 	$(MAKE) build
 
 prepare: prepare-sources $(OPTIONAL_TARGETS)
@@ -283,10 +263,8 @@ clean: ## Remove build related file
 	rm -rf ./backend-assets
 	rm -rf ./go-rwkv
 	rm -rf ./go-bert
-	rm -rf ./bloomz
 	rm -rf ./whisper.cpp
 	rm -rf ./go-piper
-	rm -rf ./go-ggllm
 	rm -rf $(BINARY_NAME)
 	rm -rf release/
 	$(MAKE) -C backend/cpp/llama clean
@@ -314,7 +292,7 @@ test-models/testmodel:
 	mkdir test-dir
 	wget https://huggingface.co/nnakasato/ggml-model-test/resolve/main/ggml-model-q4.bin -O test-models/testmodel
 	wget https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
-	wget https://huggingface.co/skeskinen/ggml/resolve/main/all-MiniLM-L6-v2/ggml-model-q4_0.bin -O test-models/bert
+	wget https://huggingface.co/mudler/all-MiniLM-L6-v2/resolve/main/ggml-model-q4_0.bin -O test-models/bert
 	wget https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
 	wget https://huggingface.co/mudler/rwkv-4-raven-1.5B-ggml/resolve/main/RWKV-4-Raven-1B5-v11-Eng99%2525-Other1%2525-20230425-ctx4096_Q4_0.bin -O test-models/rwkv
 	wget https://raw.githubusercontent.com/saharNooby/rwkv.cpp/5eb8f09c146ea8124633ab041d9ea0b1f1db4459/rwkv/20B_tokenizer.json -O test-models/rwkv.tokenizer.json
@@ -412,10 +390,6 @@ protogen-python:
 backend-assets/grpc:
 	mkdir -p backend-assets/grpc
 
-backend-assets/grpc/falcon: backend-assets/grpc go-ggllm/libggllm.a
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-ggllm LIBRARY_PATH=$(shell pwd)/go-ggllm \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/falcon ./cmd/grpc/falcon/
-
 backend-assets/grpc/llama: backend-assets/grpc go-llama/libbinding.a
 	$(GOCMD) mod edit -replace github.com/go-skynet/go-llama.cpp=$(shell pwd)/go-llama
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-llama LIBRARY_PATH=$(shell pwd)/go-llama \
@@ -480,10 +454,6 @@ backend-assets/grpc/rwkv: backend-assets/grpc go-rwkv/librwkv.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-rwkv LIBRARY_PATH=$(shell pwd)/go-rwkv \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/rwkv ./cmd/grpc/rwkv/
 
-backend-assets/grpc/bloomz: backend-assets/grpc bloomz/libbloomz.a
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/bloomz LIBRARY_PATH=$(shell pwd)/bloomz \
-	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bloomz ./cmd/grpc/bloomz/
-
 backend-assets/grpc/bert-embeddings: backend-assets/grpc go-bert/libgobert.a
 	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(shell pwd)/go-bert LIBRARY_PATH=$(shell pwd)/go-bert \
 	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/bert-embeddings ./cmd/grpc/bert-embeddings/

api/api_test.go

@@ -457,7 +457,7 @@ var _ = Describe("API test", func() {
 				Eventually(func() bool {
 					response := getModelStatus("http://127.0.0.1:9090/models/jobs/" + uuid)
 					return response["processed"].(bool)
-				}, "360s", "10s").Should(Equal(true))
+				}, "960s", "10s").Should(Equal(true))
 
 				resp, err := client.CreateChatCompletion(context.TODO(), openai.ChatCompletionRequest{Model: "gpt4all-j", Messages: []openai.ChatCompletionMessage{openai.ChatCompletionMessage{Role: "user", Content: "How are you?"}}})
 				Expect(err).ToNot(HaveOccurred())
@@ -687,7 +687,7 @@ var _ = Describe("API test", func() {
 					Input: []string{"sun", "cat"},
 				},
 			)
-			Expect(err).ToNot(HaveOccurred())
+			Expect(err).ToNot(HaveOccurred(), err)
 			Expect(len(resp.Data[0].Embedding)).To(BeNumerically("==", 384))
 			Expect(len(resp.Data[1].Embedding)).To(BeNumerically("==", 384))
 

backend/cpp/llama/CMakeLists.txt

@@ -4,6 +4,11 @@ set(TARGET grpc-server)
 set(_PROTOBUF_LIBPROTOBUF libprotobuf)
 set(_REFLECTION grpc++_reflection)
 
+if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    link_directories("/opt/homebrew/lib")
+    include_directories("/opt/homebrew/include")
+endif()
+
 find_package(absl CONFIG REQUIRED)
 find_package(Protobuf CONFIG REQUIRED)
 find_package(gRPC CONFIG REQUIRED)
@@ -15,8 +20,7 @@ find_program(_GRPC_CPP_PLUGIN_EXECUTABLE grpc_cpp_plugin)
 include_directories(${CMAKE_CURRENT_BINARY_DIR})
 include_directories(${Protobuf_INCLUDE_DIRS})
 
-message(STATUS "Using protobuf ${Protobuf_VERSION} ${Protobuf_INCLUDE_DIRS} ${CMAKE_CURRENT_BINARY_DIR}")
-
+message(STATUS "Using protobuf version ${Protobuf_VERSION} | Protobuf_INCLUDE_DIRS: ${Protobuf_INCLUDE_DIRS} | CMAKE_CURRENT_BINARY_DIR: ${CMAKE_CURRENT_BINARY_DIR}")
 
 # Proto file
 get_filename_component(hw_proto "../../../../../../pkg/grpc/proto/backend.proto" ABSOLUTE)

backend/cpp/llama/Makefile

@@ -1,5 +1,5 @@
 
-LLAMA_VERSION?=24ba3d829e31a6eda3fa1723f692608c2fa3adda
+LLAMA_VERSION?=
 
 CMAKE_ARGS?=
 BUILD_TYPE?=

backend/cpp/llama/grpc-server.cpp

@@ -88,6 +88,7 @@ static size_t find_partial_stop_string(const std::string &stop,
     return std::string::npos;
 }
 
+
 template <class Iter>
 static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
 {
@@ -128,17 +129,15 @@ struct llama_server_context
     size_t n_past = 0;
     size_t n_remain = 0;
 
-   // json prompt;
     std::vector<llama_token> embd;
-    std::vector<llama_token> last_n_tokens;
+
+    gpt_params params;
 
     llama_model *model = nullptr;
     llama_context *ctx = nullptr;
-    gpt_params params;
-    int n_ctx;
+    llama_sampling_context *ctx_sampling = nullptr;
 
-    grammar_parser::parse_state parsed_grammar;
-    llama_grammar *grammar = nullptr;
+    int n_ctx;
 
     bool truncated = false;
     bool stopped_eos = false;
@@ -171,7 +170,7 @@ struct llama_server_context
     void rewind()
     {
         params.antiprompt.clear();
-        params.grammar.clear();
+        params.sparams.grammar.clear();
         num_prompt_tokens = 0;
         num_tokens_predicted = 0;
         generated_text = "";
@@ -185,100 +184,87 @@ struct llama_server_context
         multibyte_pending = 0;
         n_remain = 0;
         n_past = 0;
+        params.sparams.n_prev = n_ctx;
+    }
 
-        if (grammar != nullptr) {
-            llama_grammar_free(grammar);
-            grammar = nullptr;
+    void initSampling() {
+        if (ctx_sampling != nullptr) {
+            llama_sampling_free(ctx_sampling);
         }
+        ctx_sampling = llama_sampling_init(params.sparams);
     }
 
     bool loadModel(const gpt_params &params_)
     {
-                    printf("load model %s\n", params_.model.c_str());
-
         params = params_;
         std::tie(model, ctx) = llama_init_from_gpt_params(params);
         if (model == nullptr)
         {
-            printf("unable to load model %s\n", params_.model.c_str());
             return false;
         }
         n_ctx = llama_n_ctx(ctx);
-        last_n_tokens.resize(n_ctx);
-        std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
         return true;
     }
-
-     std::vector<llama_token> tokenize_array(const char **prompts, bool add_bos) const
-    {
-        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+    std::vector<llama_token> tokenize_string(const char *prompt, bool add_bos) const {
+       // If `add_bos` is true, we only add BOS, when json_prompt is a string,
         // or the first element of the json_prompt array is a string.
-        std::vector<llama_token> prompt_tokens;
-
-  
-        bool first = true;
-        // Iterate over prompts
-        for (const char **p = prompts; *p != nullptr; ++p)
-        {
-            auto s = std::string(*p);
-            std::vector<llama_token> pp;
-            if (first)
-            {
-                pp = ::llama_tokenize(ctx, s, add_bos);
-                first = false;
-            }
-            else
-            {
-                pp = ::llama_tokenize(ctx, s, false);
-            }
-            prompt_tokens.insert(prompt_tokens.end(), pp.begin(), pp.end());
-        }
-        
-
-        return prompt_tokens;
-    }
-
-    std::vector<llama_token> tokenize_string(const char *prompt, bool add_bos) const
-    {
-        // If `add_bos` is true, we only add BOS, when json_prompt is a string,
-        // or the first element of the json_prompt array is a string.
-        std::vector<llama_token> prompt_tokens;
-  
+        std::vector<llama_token> prompt_tokens; 
         auto s = std::string(prompt);
         prompt_tokens = ::llama_tokenize(ctx, s, add_bos);
-    
         return prompt_tokens;
     }
+     std::vector<llama_token> tokenize_array(const char **prompts, bool add_bos) const {
+                std::vector<llama_token> prompt_tokens; 
 
-    bool loadGrammar()
-    {
-        if (!params.grammar.empty()) {
-            parsed_grammar = grammar_parser::parse(params.grammar.c_str());
-            // will be empty (default) if there are parse errors
-            if (parsed_grammar.rules.empty()) {
-                printf("grammar parse error");
-                return false;
-            }
-            grammar_parser::print_grammar(stderr, parsed_grammar);
-
-            {
-                auto it = params.logit_bias.find(llama_token_eos(ctx));
-                if (it != params.logit_bias.end() && it->second == -INFINITY) {
-                    printf("EOS token is disabled, which will cause most grammars to fail");
+            bool first = true;
+            bool is_string = true;
+            for (const char **p = prompts; *p != nullptr; ++p)
+              {
+                if (is_string)
+                {
+                    auto s = std::string(*p);
+                    std::vector<llama_token> p;
+                    if (first)
+                    {
+                        p = ::llama_tokenize(ctx, s, add_bos);
+                        first = false;
+                    }
+                    else
+                    {
+                        p = ::llama_tokenize(ctx, s, false);
+                    }
+                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+                }
+                else
+                {
+                    if (first)
+                    {
+                        first = false;
+                    }
+                    //prompt_tokens.push_back(p.template get<llama_token>());
                 }
             }
+            return prompt_tokens;
+     }
 
-            std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
-            grammar = llama_grammar_init(
-                grammar_rules.data(), grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
-        }
-        return true;
+    void truncatePrompt(std::vector<llama_token> &prompt_tokens) {
+        const int n_left = n_ctx - params.n_keep;
+        const int n_block_size = n_left / 2;
+        const int erased_blocks = (prompt_tokens.size() - params.n_keep - n_block_size) / n_block_size;
+
+        // Keep n_keep tokens at start of prompt (at most n_ctx - 4)
+        std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
+
+        new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_block_size, prompt_tokens.end());
+
+        truncated = true;
+        prompt_tokens = new_tokens;
     }
 
     void loadInfill()
     {
         bool suff_rm_leading_spc = true;
-        if (params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
+        if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
             params.input_suffix.erase(0, 1);
             suff_rm_leading_spc = false;
         }
@@ -289,11 +275,12 @@ struct llama_server_context
         if (suff_rm_leading_spc  && suffix_tokens[0] == space_token) {
             suffix_tokens.erase(suffix_tokens.begin());
         }
-        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(ctx));
-        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(ctx)); // always add BOS
-        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(ctx));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_prefix(model));
+        prefix_tokens.insert(prefix_tokens.begin(), llama_token_bos(model)); // always add BOS
+        prefix_tokens.insert(prefix_tokens.end(), llama_token_suffix(model));
         prefix_tokens.insert(prefix_tokens.end(), suffix_tokens.begin(), suffix_tokens.end());
-        prefix_tokens.push_back(llama_token_middle(ctx));
+        prefix_tokens.push_back(llama_token_middle(model));
+
         auto prompt_tokens = prefix_tokens;
 
         num_prompt_tokens = prompt_tokens.size();
@@ -305,29 +292,24 @@ struct llama_server_context
         params.n_keep = std::min(params.n_ctx - 4, params.n_keep);
 
         // if input prompt is too big, truncate like normal
-        if (num_prompt_tokens >= (size_t)params.n_ctx)
+        if (num_prompt_tokens >= (size_t) n_ctx)
         {
-            printf("Input prompt is too big, truncating. Can only take %d tokens but got %zu\n", params.n_ctx, num_prompt_tokens);
-            // todo we probably want to cut from both sides
-            const int n_left = (params.n_ctx - params.n_keep) / 2;
-            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
-            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
-            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
-            std::copy(prompt_tokens.end() - params.n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+            truncatePrompt(prompt_tokens);
+            num_prompt_tokens = prompt_tokens.size();
 
-            truncated = true;
-            prompt_tokens = new_tokens;
+            GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
         }
-        else
+
+        // push the prompt into the sampling context (do not apply grammar)
+        for (auto & token : prompt_tokens)
         {
-            const size_t ps = num_prompt_tokens;
-            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
-            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+            llama_sampling_accept(ctx_sampling, ctx, token, false);
         }
 
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
         embd = prompt_tokens;
+
         if (n_past == num_prompt_tokens)
         {
             // we have to evaluate at least 1 token to generate logits.
@@ -335,6 +317,7 @@ struct llama_server_context
             n_past--;
         }
 
+        // since #3228 we now have to manually manage the KV cache
         llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
 
         has_next_token = true;
@@ -352,38 +335,33 @@ struct llama_server_context
         params.n_keep = std::min(n_ctx - 4, params.n_keep);
 
         // if input prompt is too big, truncate like normal
-        if (num_prompt_tokens >= (size_t)n_ctx)
+        if (num_prompt_tokens >= (size_t) n_ctx)
         {
-            const int n_left = (n_ctx - params.n_keep) / 2;
-            std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + params.n_keep);
-            const int erased_blocks = (num_prompt_tokens - params.n_keep - n_left - 1) / n_left;
-            new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + params.n_keep + erased_blocks * n_left, prompt_tokens.end());
-            std::copy(prompt_tokens.end() - n_ctx, prompt_tokens.end(), last_n_tokens.begin());
+            truncatePrompt(prompt_tokens);
+            num_prompt_tokens = prompt_tokens.size();
 
-
-            truncated = true;
-            prompt_tokens = new_tokens;
+            GGML_ASSERT(num_prompt_tokens < (size_t)n_ctx);
         }
-        else
+
+        // push the prompt into the sampling context (do not apply grammar)
+        for (auto & token : prompt_tokens)
         {
-            const size_t ps = num_prompt_tokens;
-            std::fill(last_n_tokens.begin(), last_n_tokens.end() - ps, 0);
-            std::copy(prompt_tokens.begin(), prompt_tokens.end(), last_n_tokens.end() - ps);
+            llama_sampling_accept(ctx_sampling, ctx, token, false);
         }
 
         // compare the evaluated prompt with the new prompt
         n_past = common_part(embd, prompt_tokens);
 
-
         embd = prompt_tokens;
         if (n_past == num_prompt_tokens)
         {
             // we have to evaluate at least 1 token to generate logits.
             n_past--;
         }
-        // since #3228 we now have to manually manage the KV cache
 
+        // since #3228 we now have to manually manage the KV cache
         llama_kv_cache_seq_rm(ctx, 0, n_past, -1);
+
         has_next_token = true;
     }
 
@@ -418,7 +396,6 @@ struct llama_server_context
             n_past -= n_discard;
 
             truncated = true;
-       
         }
 
         bool tg = true;
@@ -433,7 +410,6 @@ struct llama_server_context
 
             if (llama_decode(ctx, llama_batch_get_one(&embd[n_past], n_eval, n_past, 0)))
             {
-             
                 has_next_token = false;
                 return result;
             }
@@ -443,33 +419,30 @@ struct llama_server_context
         if (params.n_predict == 0)
         {
             has_next_token = false;
-            result.tok = llama_token_eos(ctx);
+            result.tok = llama_token_eos(model);
             return result;
         }
 
         {
             // out of user input, sample next token
-            std::vector<llama_token_data> candidates;
-            candidates.reserve(llama_n_vocab(model));
+            result.tok = llama_sampling_sample(ctx_sampling, ctx, NULL);
 
-            result.tok = llama_sample_token(ctx, NULL, grammar, params, last_n_tokens, candidates);
+            llama_token_data_array cur_p = { ctx_sampling->cur.data(), ctx_sampling->cur.size(), false };
 
-            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-            const int32_t n_probs = params.n_probs;
-            if (params.temp <= 0 && n_probs > 0)
+            const int32_t n_probs = params.sparams.n_probs;
+            if (params.sparams.temp <= 0 && n_probs > 0)
             {
                 // For llama_sample_token_greedy we need to sort candidates
-                llama_sample_softmax(ctx, &candidates_p);
+                llama_sample_softmax(ctx, &cur_p);
             }
 
-            for (size_t i = 0; i < std::min(candidates_p.size, (size_t)n_probs); ++i)
+            for (size_t i = 0; i < std::min(cur_p.size, (size_t)n_probs); ++i)
             {
-                result.probs.push_back({candidates_p.data[i].id, candidates_p.data[i].p});
+                result.probs.push_back({cur_p.data[i].id, cur_p.data[i].p});
             }
 
-            last_n_tokens.erase(last_n_tokens.begin());
-            last_n_tokens.push_back(result.tok);
+            llama_sampling_accept(ctx_sampling, ctx, result.tok, true);
+
             if (tg) {
                 num_tokens_predicted++;
             }
@@ -480,7 +453,7 @@ struct llama_server_context
         // decrement remaining sampling budget
         --n_remain;
 
-        if (!embd.empty() && embd.back() == llama_token_eos(ctx))
+        if (!embd.empty() && embd.back() == llama_token_eos(model))
         {
             // stopping_word = llama_token_to_piece(ctx, embd.back());
             has_next_token = false;
@@ -531,7 +504,7 @@ struct llama_server_context
         const std::string token_text = token_with_probs.tok == -1 ? "" : llama_token_to_piece(ctx, token_with_probs.tok);
         generated_text += token_text;
 
-        if (params.n_probs > 0)
+        if (params.sparams.n_probs > 0)
         {
             generated_token_probs.push_back(token_with_probs);
         }
@@ -583,7 +556,6 @@ struct llama_server_context
         static const int n_embd = llama_n_embd(model);
         if (!params.embedding)
         {
-            printf("embedding disabled");
             return std::vector<float>(n_embd, 0.0f);
         }
         const float *data = llama_get_embeddings(ctx);
@@ -599,30 +571,30 @@ static void parse_options_completion(bool streaming,const backend::PredictOption
 
     llama.stream = streaming;
     llama.params.n_predict = predict->tokens() == 0 ? -1 : predict->tokens();
-    llama.params.top_k = predict->topk();
-    llama.params.top_p = predict->topp();
-    llama.params.tfs_z = predict->tailfreesamplingz();
-    llama.params.typical_p = predict->typicalp();
-    llama.params.repeat_last_n = predict->repeat();
-    llama.params.temp = predict->temperature();
-    llama.params.repeat_penalty = predict->penalty();
-    llama.params.presence_penalty = predict->presencepenalty();
-    llama.params.frequency_penalty = predict->frequencypenalty();
-    llama.params.mirostat = predict->mirostat();
-    llama.params.mirostat_tau = predict->mirostattau();
-    llama.params.mirostat_eta = predict->mirostateta();
-    llama.params.penalize_nl = predict->penalizenl();
+    llama.params.sparams.top_k = predict->topk();
+    llama.params.sparams.top_p = predict->topp();
+    llama.params.sparams.tfs_z = predict->tailfreesamplingz();
+    llama.params.sparams.typical_p = predict->typicalp();
+    llama.params.sparams.penalty_last_n = predict->repeat();
+    llama.params.sparams.temp = predict->temperature();
+    llama.params.sparams.penalty_repeat = predict->penalty();
+    llama.params.sparams.penalty_present = predict->presencepenalty();
+    llama.params.sparams.penalty_freq = predict->frequencypenalty();
+    llama.params.sparams.mirostat = predict->mirostat();
+    llama.params.sparams.mirostat_tau = predict->mirostattau();
+    llama.params.sparams.mirostat_eta = predict->mirostateta();
+    llama.params.sparams.penalize_nl = predict->penalizenl();
     llama.params.n_keep = predict->nkeep();
     llama.params.seed = predict->seed();
-    llama.params.grammar = predict->grammar();
+    llama.params.sparams.grammar = predict->grammar();
     // llama.params.n_probs = predict->
     llama.params.prompt = predict->prompt();
 
-    llama.params.logit_bias.clear();
+    llama.params.sparams.logit_bias.clear();
 
     if (predict->ignoreeos())
     {
-        llama.params.logit_bias[llama_token_eos(llama.ctx)] = -INFINITY;
+        llama.params.sparams.logit_bias[llama_token_eos(llama.model)] = -INFINITY;
     }
 
     // const auto &logit_bias = body.find("logit_bias");
@@ -704,7 +676,7 @@ static void params_parse(const backend::ModelOptions* request,
 }
 
 static bool is_at_eob(llama_server_context &server_context, const llama_token *tokens, const size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.ctx);
+    return n_tokens && tokens[n_tokens-1] == llama_token_eos(server_context.model);
 }
 
 // Function matching type llama_beam_search_callback_fn_t.
@@ -801,12 +773,7 @@ public:
 
         parse_options_completion(false, request, llama);
 
-        if (!llama.loadGrammar())
-        {
-            //res.status = 400;
-            return Status::CANCELLED;
-        }
-
+        llama.initSampling();
         llama.loadPrompt(request->prompt());
         llama.beginCompletion();
         size_t sent_count = 0;
@@ -848,7 +815,7 @@ public:
 
                 std::vector<completion_token_output> probs_output = {};
 
-                if (llama.params.n_probs > 0) {
+                if (llama.params.sparams.n_probs > 0) {
                     const std::vector<llama_token> to_send_toks = llama_tokenize(llama.ctx, to_send, false);
                     size_t probs_pos = std::min(sent_token_probs_index, llama.generated_token_probs.size());
                     size_t probs_stop_pos = std::min(sent_token_probs_index + to_send_toks.size(), llama.generated_token_probs.size());
@@ -879,12 +846,7 @@ public:
         llama_reset_timings(llama.ctx);
         parse_options_completion(false, request, llama);
 
-        if (!llama.loadGrammar())
-        {
-            //res.status = 400;
-            return Status::CANCELLED;
-        }
-
+        llama.initSampling();
         llama.loadPrompt(request->prompt());
         llama.beginCompletion();
 
@@ -915,7 +877,7 @@ public:
         }
 
         auto probs = llama.generated_token_probs;
-        if (llama.params.n_probs > 0 && llama.stopped_word) {
+        if (llama.params.sparams.n_probs > 0 && llama.stopped_word) {
             const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
             probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
         }

cmd/grpc/bloomz/main.go

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

cmd/grpc/falcon/main.go

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

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

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

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

@@ -15,10 +15,16 @@ headers {
 }
 
 body:json {
-  {
-       "model": "{{DEFAULT_MODEL}}",
-       "messages": [{"role": "user", "content": "How could one use friction to cook an egg?"}],
-       "max_tokens": 256,
-       "temperature": 0.2
+  {
+    "model": "{{DEFAULT_MODEL}}",
+    "messages": [
+      {
+        "role": "user",
+        "content": "How could one use friction to cook an egg?"
+      }
+    ],
+    "max_tokens": 256,
+    "temperature": 0.2,
+    "grammar": ""
   }
 }

examples/configurations/README.md

@@ -0,0 +1,42 @@
+## Advanced configuration
+
+This section contains examples on how to install models manually with config files.
+
+### Prerequisites
+
+First clone LocalAI:
+
+```bash
+git clone https://github.com/go-skynet/LocalAI
+
+cd LocalAI
+```
+
+Setup the model you prefer from the examples below and then start LocalAI:
+
+```bash
+docker compose up -d --pull always
+```
+
+If LocalAI is already started, you can restart it with 
+
+```bash
+docker compose restart
+```
+
+See also the getting started: https://localai.io/basics/getting_started/
+
+### Mistral
+
+To setup mistral copy the files inside `mistral` in the `models` folder:
+
+```bash
+cp -r examples/configurations/mistral/* models/
+```
+
+Now download the model:
+
+```bash
+wget https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-GGUF/resolve/main/mistral-7b-openorca.Q6_K.gguf -O models/mistral-7b-openorca.Q6_K.gguf
+```
+

examples/configurations/mistral/chatml-block.tmpl

@@ -0,0 +1,3 @@
+{{.Input}}
+<|im_start|>assistant
+

examples/configurations/mistral/chatml.tmpl

@@ -0,0 +1,3 @@
+<|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "user"}}user{{end}}
+{{if .Content}}{{.Content}}{{end}}
+<|im_end|>

examples/configurations/mistral/completion.tmpl

@@ -0,0 +1 @@
+{{.Input}}

examples/configurations/mistral/mistral.yaml

@@ -0,0 +1,16 @@
+name: mistral
+mmap: true
+parameters:
+  model: mistral-7b-openorca.Q6_K.gguf
+  temperature: 0.2
+  top_k: 40
+  top_p: 0.95
+template:
+  chat_message: chatml
+  chat: chatml-block
+  completion: completion
+context_size: 4096
+f16: true
+stopwords:
+- <|im_end|>
+threads: 4

go.mod

@@ -20,9 +20,9 @@ require (
 	github.com/mudler/go-ggllm.cpp v0.0.0-20230709223052-862477d16eef
 	github.com/mudler/go-processmanager v0.0.0-20230818213616-f204007f963c
 	github.com/mudler/go-stable-diffusion v0.0.0-20230605122230-d89260f598af
-	github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231016205817-9a19c740ee84
+	github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231022042237-c25dc5193530
 	github.com/onsi/ginkgo/v2 v2.13.0
-	github.com/onsi/gomega v1.28.0
+	github.com/onsi/gomega v1.28.1
 	github.com/otiai10/openaigo v1.6.0
 	github.com/phayes/freeport v0.0.0-20220201140144-74d24b5ae9f5
 	github.com/prometheus/client_golang v1.17.0
@@ -89,7 +89,7 @@ require (
 	github.com/go-audio/riff v1.0.0 // indirect
 	github.com/go-logr/logr v1.2.4 // indirect
 	github.com/go-task/slim-sprig v0.0.0-20230315185526-52ccab3ef572 // indirect
-	github.com/google/go-cmp v0.5.9 // indirect
+	github.com/google/go-cmp v0.6.0 // indirect
 	github.com/google/pprof v0.0.0-20210407192527-94a9f03dee38 // indirect
 	github.com/hashicorp/errwrap v1.0.0 // indirect
 	github.com/klauspost/compress v1.16.7 // indirect

go.sum

@@ -73,6 +73,8 @@ github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.6/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE=
 github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
 github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
+github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
 github.com/google/pprof v0.0.0-20210407192527-94a9f03dee38 h1:yAJXTCF9TqKcTiHJAE8dj7HMvPfh66eeA2JYW7eFpSE=
 github.com/google/pprof v0.0.0-20210407192527-94a9f03dee38/go.mod h1:kpwsk12EmLew5upagYY7GY0pfYCcupk39gWOCRROcvE=
@@ -135,6 +137,8 @@ github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231013181651-22de3c
 github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231013181651-22de3c56bdd4/go.mod h1:4T3CHXyrt+7FQHXaxULZfPjHbD8/99WuDDJa0YVZARI=
 github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231016205817-9a19c740ee84 h1:AiFzd+M2Uxz67fdn4nCnKR70me5yf88rXhoqhvfRDak=
 github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231016205817-9a19c740ee84/go.mod h1:4T3CHXyrt+7FQHXaxULZfPjHbD8/99WuDDJa0YVZARI=
+github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231022042237-c25dc5193530 h1:YXMxHwHMB9jCBo2Yu5gz3mTB3T1TnZs/HmPLv15LUSA=
+github.com/nomic-ai/gpt4all/gpt4all-bindings/golang v0.0.0-20231022042237-c25dc5193530/go.mod h1:4T3CHXyrt+7FQHXaxULZfPjHbD8/99WuDDJa0YVZARI=
 github.com/nwaples/rardecode v1.1.0 h1:vSxaY8vQhOcVr4mm5e8XllHWTiM4JF507A0Katqw7MQ=
 github.com/nwaples/rardecode v1.1.0/go.mod h1:5DzqNKiOdpKKBH87u8VlvAnPZMXcGRhxWkRpHbbfGS0=
 github.com/nxadm/tail v1.4.4/go.mod h1:kenIhsEOeOJmVchQTgglprH7qJGnHDVpk1VPCcaMI8A=
@@ -151,6 +155,8 @@ github.com/onsi/gomega v1.10.1/go.mod h1:iN09h71vgCQne3DLsj+A5owkum+a2tYe+TOCB1y
 github.com/onsi/gomega v1.16.0/go.mod h1:HnhC7FXeEQY45zxNK3PPoIUhzk/80Xly9PcubAlGdZY=
 github.com/onsi/gomega v1.28.0 h1:i2rg/p9n/UqIDAMFUJ6qIUUMcsqOuUHgbpbu235Vr1c=
 github.com/onsi/gomega v1.28.0/go.mod h1:A1H2JE76sI14WIP57LMKj7FVfCHx3g3BcZVjJG8bjX8=
+github.com/onsi/gomega v1.28.1 h1:MijcGUbfYuznzK/5R4CPNoUP/9Xvuo20sXfEm6XxoTA=
+github.com/onsi/gomega v1.28.1/go.mod h1:9sxs+SwGrKI0+PWe4Fxa9tFQQBG5xSsSbMXOI8PPpoQ=
 github.com/otiai10/mint v1.6.1 h1:kgbTJmOpp/0ce7hk3H8jiSuR0MXmpwWRfqUdKww17qg=
 github.com/otiai10/mint v1.6.1/go.mod h1:MJm72SBthJjz8qhefc4z1PYEieWmy8Bku7CjcAqyUSM=
 github.com/otiai10/openaigo v1.6.0 h1:YTQEbtDSvawETOB/Kmb/6JvuHdHH/eIpSQfHVufiwY8=

pkg/backend/llm/bloomz/bloomz.go

@@ -1,59 +0,0 @@
-package bloomz
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	"fmt"
-
-	"github.com/go-skynet/LocalAI/pkg/grpc/base"
-	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
-
-	"github.com/go-skynet/bloomz.cpp"
-)
-
-type LLM struct {
-	base.SingleThread
-
-	bloomz *bloomz.Bloomz
-}
-
-func (llm *LLM) Load(opts *pb.ModelOptions) error {
-	model, err := bloomz.New(opts.ModelFile)
-	llm.bloomz = model
-	return err
-}
-
-func buildPredictOptions(opts *pb.PredictOptions) []bloomz.PredictOption {
-	predictOptions := []bloomz.PredictOption{
-		bloomz.SetTemperature(float64(opts.Temperature)),
-		bloomz.SetTopP(float64(opts.TopP)),
-		bloomz.SetTopK(int(opts.TopK)),
-		bloomz.SetTokens(int(opts.Tokens)),
-		bloomz.SetThreads(int(opts.Threads)),
-	}
-
-	if opts.Seed != 0 {
-		predictOptions = append(predictOptions, bloomz.SetSeed(int(opts.Seed)))
-	}
-
-	return predictOptions
-}
-
-func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
-	return llm.bloomz.Predict(opts.Prompt, buildPredictOptions(opts)...)
-}
-
-// fallback to Predict
-func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
-	go func() {
-		res, err := llm.bloomz.Predict(opts.Prompt, buildPredictOptions(opts)...)
-
-		if err != nil {
-			fmt.Println("err: ", err)
-		}
-		results <- res
-		close(results)
-	}()
-
-	return nil
-}

pkg/backend/llm/falcon/falcon.go

@@ -1,145 +0,0 @@
-package falcon
-
-// This is a wrapper to statisfy the GRPC service interface
-// It is meant to be used by the main executable that is the server for the specific backend type (falcon, gpt3, etc)
-import (
-	"fmt"
-
-	"github.com/go-skynet/LocalAI/pkg/grpc/base"
-	pb "github.com/go-skynet/LocalAI/pkg/grpc/proto"
-
-	ggllm "github.com/mudler/go-ggllm.cpp"
-)
-
-type LLM struct {
-	base.SingleThread
-
-	falcon *ggllm.Falcon
-}
-
-func (llm *LLM) Load(opts *pb.ModelOptions) error {
-	ggllmOpts := []ggllm.ModelOption{}
-	if opts.ContextSize != 0 {
-		ggllmOpts = append(ggllmOpts, ggllm.SetContext(int(opts.ContextSize)))
-	}
-	// F16 doesn't seem to produce good output at all!
-	//if c.F16 {
-	//	llamaOpts = append(llamaOpts, llama.EnableF16Memory)
-	//}
-
-	if opts.NGPULayers != 0 {
-		ggllmOpts = append(ggllmOpts, ggllm.SetGPULayers(int(opts.NGPULayers)))
-	}
-
-	ggllmOpts = append(ggllmOpts, ggllm.SetMMap(opts.MMap))
-	ggllmOpts = append(ggllmOpts, ggllm.SetMainGPU(opts.MainGPU))
-	ggllmOpts = append(ggllmOpts, ggllm.SetTensorSplit(opts.TensorSplit))
-	if opts.NBatch != 0 {
-		ggllmOpts = append(ggllmOpts, ggllm.SetNBatch(int(opts.NBatch)))
-	} else {
-		ggllmOpts = append(ggllmOpts, ggllm.SetNBatch(512))
-	}
-
-	model, err := ggllm.New(opts.ModelFile, ggllmOpts...)
-	llm.falcon = model
-	return err
-}
-
-func buildPredictOptions(opts *pb.PredictOptions) []ggllm.PredictOption {
-	predictOptions := []ggllm.PredictOption{
-		ggllm.SetTemperature(float64(opts.Temperature)),
-		ggllm.SetTopP(float64(opts.TopP)),
-		ggllm.SetTopK(int(opts.TopK)),
-		ggllm.SetTokens(int(opts.Tokens)),
-		ggllm.SetThreads(int(opts.Threads)),
-	}
-
-	if opts.PromptCacheAll {
-		predictOptions = append(predictOptions, ggllm.EnablePromptCacheAll)
-	}
-
-	if opts.PromptCacheRO {
-		predictOptions = append(predictOptions, ggllm.EnablePromptCacheRO)
-	}
-
-	// Expected absolute path
-	if opts.PromptCachePath != "" {
-		predictOptions = append(predictOptions, ggllm.SetPathPromptCache(opts.PromptCachePath))
-	}
-
-	if opts.Mirostat != 0 {
-		predictOptions = append(predictOptions, ggllm.SetMirostat(int(opts.Mirostat)))
-	}
-
-	if opts.MirostatETA != 0 {
-		predictOptions = append(predictOptions, ggllm.SetMirostatETA(float64(opts.MirostatETA)))
-	}
-
-	if opts.MirostatTAU != 0 {
-		predictOptions = append(predictOptions, ggllm.SetMirostatTAU(float64(opts.MirostatTAU)))
-	}
-
-	if opts.Debug {
-		predictOptions = append(predictOptions, ggllm.Debug)
-	}
-
-	predictOptions = append(predictOptions, ggllm.SetStopWords(opts.StopPrompts...))
-
-	if opts.PresencePenalty != 0 {
-		predictOptions = append(predictOptions, ggllm.SetPenalty(float64(opts.PresencePenalty)))
-	}
-
-	if opts.NKeep != 0 {
-		predictOptions = append(predictOptions, ggllm.SetNKeep(int(opts.NKeep)))
-	}
-
-	if opts.Batch != 0 {
-		predictOptions = append(predictOptions, ggllm.SetBatch(int(opts.Batch)))
-	}
-
-	if opts.IgnoreEOS {
-		predictOptions = append(predictOptions, ggllm.IgnoreEOS)
-	}
-
-	if opts.Seed != 0 {
-		predictOptions = append(predictOptions, ggllm.SetSeed(int(opts.Seed)))
-	}
-
-	//predictOptions = append(predictOptions, llama.SetLogitBias(c.Seed))
-
-	predictOptions = append(predictOptions, ggllm.SetFrequencyPenalty(float64(opts.FrequencyPenalty)))
-	predictOptions = append(predictOptions, ggllm.SetMlock(opts.MLock))
-	predictOptions = append(predictOptions, ggllm.SetMemoryMap(opts.MMap))
-	predictOptions = append(predictOptions, ggllm.SetPredictionMainGPU(opts.MainGPU))
-	predictOptions = append(predictOptions, ggllm.SetPredictionTensorSplit(opts.TensorSplit))
-	predictOptions = append(predictOptions, ggllm.SetTailFreeSamplingZ(float64(opts.TailFreeSamplingZ)))
-	predictOptions = append(predictOptions, ggllm.SetTypicalP(float64(opts.TypicalP)))
-	return predictOptions
-}
-
-func (llm *LLM) Predict(opts *pb.PredictOptions) (string, error) {
-	return llm.falcon.Predict(opts.Prompt, buildPredictOptions(opts)...)
-}
-
-func (llm *LLM) PredictStream(opts *pb.PredictOptions, results chan string) error {
-
-	predictOptions := buildPredictOptions(opts)
-
-	predictOptions = append(predictOptions, ggllm.SetTokenCallback(func(token string) bool {
-		if token == "<|endoftext|>" {
-			return true
-		}
-		results <- token
-		return true
-	}))
-
-	go func() {
-		_, err := llm.falcon.Predict(opts.Prompt, predictOptions...)
-		if err != nil {
-			fmt.Println("err: ", err)
-		}
-		close(results)
-	}()
-
-	return nil
-}

pkg/gallery/gallery.go

@@ -8,6 +8,7 @@ import (
 
 	"github.com/go-skynet/LocalAI/pkg/utils"
 	"github.com/imdario/mergo"
+	"github.com/rs/zerolog/log"
 	"gopkg.in/yaml.v2"
 )
 
@@ -166,7 +167,9 @@ func getGalleryModels(gallery Gallery, basePath string) ([]*GalleryModel, error)
 		return yaml.Unmarshal(d, &models)
 	})
 	if err != nil {
-
+		if yamlErr, ok := err.(*yaml.TypeError); ok {
+			log.Debug().Msgf("YAML errors: %s\n\nwreckage of models: %+v", strings.Join(yamlErr.Errors, "\n"), models)
+		}
 		return models, err
 	}
 

pkg/model/initializers.go

@@ -18,7 +18,6 @@ const (
 	LlamaBackend        = "llama"
 	LlamaStableBackend  = "llama-stable"
 	LLamaCPP            = "llama-cpp"
-	BloomzBackend       = "bloomz"
 	StarcoderBackend    = "starcoder"
 	GPTJBackend         = "gptj"
 	DollyBackend        = "dolly"
@@ -30,7 +29,6 @@ const (
 	Gpt4AllMptBackend   = "gpt4all-mpt"
 	Gpt4AllJBackend     = "gpt4all-j"
 	Gpt4All             = "gpt4all"
-	FalconBackend       = "falcon"
 	FalconGGMLBackend   = "falcon-ggml"
 
 	BertEmbeddingsBackend  = "bert-embeddings"
@@ -46,7 +44,6 @@ var AutoLoadBackends []string = []string{
 	LlamaStableBackend,
 	LlamaBackend,
 	Gpt4All,
-	FalconBackend,
 	GPTNeoXBackend,
 	BertEmbeddingsBackend,
 	FalconGGMLBackend,
@@ -56,7 +53,6 @@ var AutoLoadBackends []string = []string{
 	MPTBackend,
 	ReplitBackend,
 	StarcoderBackend,
-	BloomzBackend,
 	RwkvBackend,
 	WhisperBackend,
 	StableDiffusionBackend,