Upload 711 files

.dockerignore

@@ -0,0 +1,17 @@
+.idea
+.github
+.vscode
+.devcontainer
+models
+examples/chatbot-ui/models
+examples/rwkv/models
+examples/**/models
+Dockerfile*
+__pycache__
+
+# SonarQube
+.scannerwork
+
+# backend virtual environments
+**/venv
+backend/python/**/source

.editorconfig

@@ -0,0 +1,31 @@
+
+root = true
+
+[*]
+indent_style = space
+indent_size = 2
+end_of_line = lf
+charset = utf-8
+trim_trailing_whitespace = true
+insert_final_newline = true
+
+[*.go]
+indent_style = tab
+
+[Makefile]
+indent_style = tab
+
+[*.proto]
+indent_size = 2
+
+[*.py]
+indent_size = 4
+
+[*.js]
+indent_size = 2
+
+[*.yaml]
+indent_size = 2
+
+[*.md]
+trim_trailing_whitespace = false

.env

@@ -0,0 +1,97 @@
+## Set number of threads.
+## Note: prefer the number of physical cores. Overbooking the CPU degrades performance notably.
+# LOCALAI_THREADS=14
+
+## Specify a different bind address (defaults to ":8080")
+# LOCALAI_ADDRESS=127.0.0.1:8080
+
+## Default models context size
+# LOCALAI_CONTEXT_SIZE=512
+#
+## Define galleries.
+## models will to install will be visible in `/models/available`
+# LOCALAI_GALLERIES=[{"name":"localai", "url":"github:mudler/LocalAI/gallery/index.yaml@master"}]
+
+## CORS settings
+# LOCALAI_CORS=true
+# LOCALAI_CORS_ALLOW_ORIGINS=*
+
+## Default path for models
+#
+# LOCALAI_MODELS_PATH=/models
+
+## Enable debug mode
+# LOCALAI_LOG_LEVEL=debug
+
+## Disables COMPEL (Diffusers)
+# COMPEL=0
+
+## Enable/Disable single backend (useful if only one GPU is available)
+# LOCALAI_SINGLE_ACTIVE_BACKEND=true
+
+## Specify a build type. Available: cublas, openblas, clblas.
+## cuBLAS: This is a GPU-accelerated version of the complete standard BLAS (Basic Linear Algebra Subprograms) library. It's provided by Nvidia and is part of their CUDA toolkit.
+## OpenBLAS: This is an open-source implementation of the BLAS library that aims to provide highly optimized code for various platforms. It includes support for multi-threading and can be compiled to use hardware-specific features for additional performance. OpenBLAS can run on many kinds of hardware, including CPUs from Intel, AMD, and ARM.
+## clBLAS:   This is an open-source implementation of the BLAS library that uses OpenCL, a framework for writing programs that execute across heterogeneous platforms consisting of CPUs, GPUs, and other processors. clBLAS is designed to take advantage of the parallel computing power of GPUs but can also run on any hardware that supports OpenCL. This includes hardware from different vendors like Nvidia, AMD, and Intel.
+# BUILD_TYPE=openblas
+
+## Uncomment and set to true to enable rebuilding from source
+# REBUILD=true
+
+## Enable go tags, available: stablediffusion, tts
+## stablediffusion: image generation with stablediffusion
+## tts: enables text-to-speech with go-piper 
+## (requires REBUILD=true)
+#
+# GO_TAGS=stablediffusion
+
+## Path where to store generated images
+# LOCALAI_IMAGE_PATH=/tmp/generated/images
+
+## Specify a default upload limit in MB (whisper)
+# LOCALAI_UPLOAD_LIMIT=15
+
+## List of external GRPC backends (note on the container image this variable is already set to use extra backends available in extra/)
+# LOCALAI_EXTERNAL_GRPC_BACKENDS=my-backend:127.0.0.1:9000,my-backend2:/usr/bin/backend.py
+
+### Advanced settings ###
+### Those are not really used by LocalAI, but from components in the stack ###
+##
+### Preload libraries
+# LD_PRELOAD=
+
+### Huggingface cache for models
+# HUGGINGFACE_HUB_CACHE=/usr/local/huggingface
+
+### Python backends GRPC max workers
+### Default number of workers for GRPC Python backends.
+### This actually controls wether a backend can process multiple requests or not.
+# PYTHON_GRPC_MAX_WORKERS=1
+
+### Define the number of parallel LLAMA.cpp workers (Defaults to 1)
+# LLAMACPP_PARALLEL=1
+
+### Define a list of GRPC Servers for llama-cpp workers to distribute the load
+# https://github.com/ggerganov/llama.cpp/pull/6829
+# https://github.com/ggerganov/llama.cpp/blob/master/examples/rpc/README.md
+# LLAMACPP_GRPC_SERVERS=""
+
+### Enable to run parallel requests
+# LOCALAI_PARALLEL_REQUESTS=true
+
+# Enable to allow p2p mode
+# LOCALAI_P2P=true
+
+### Watchdog settings
+###
+# Enables watchdog to kill backends that are inactive for too much time
+# LOCALAI_WATCHDOG_IDLE=true
+#
+# Time in duration format (e.g. 1h30m) after which a backend is considered idle
+# LOCALAI_WATCHDOG_IDLE_TIMEOUT=5m
+#
+# Enables watchdog to kill backends that are busy for too much time
+# LOCALAI_WATCHDOG_BUSY=true
+#
+# Time in duration format (e.g. 1h30m) after which a backend is considered busy
+# LOCALAI_WATCHDOG_BUSY_TIMEOUT=5m

.gitattributes

@@ -1,35 +1 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.sh text eol=lf

.gitignore

@@ -0,0 +1,59 @@
+# go-llama build artifacts
+/sources/
+__pycache__/
+*.a
+get-sources
+prepare-sources
+/backend/cpp/llama/grpc-server
+/backend/cpp/llama/llama.cpp
+/backend/cpp/llama-*
+
+*.log
+
+go-ggml-transformers
+go-gpt2
+go-rwkv
+whisper.cpp
+/bloomz
+go-bert
+
+# LocalAI build binary
+LocalAI
+local-ai
+# prevent above rules from omitting the helm chart
+!charts/*
+# prevent above rules from omitting the api/localai folder
+!api/localai
+!core/**/localai
+
+# Ignore models
+models/*
+test-models/
+test-dir/
+
+release/
+
+# just in case
+.DS_Store
+.idea
+
+# Generated during build
+backend-assets/*
+!backend-assets/.keep
+prepare
+/ggml-metal.metal
+docs/static/gallery.html
+
+# Protobuf generated files
+*.pb.go
+*pb2.py
+*pb2_grpc.py
+
+# SonarQube
+.scannerwork
+
+# backend virtual environments
+**/venv
+
+# per-developer customization files for the development container
+.devcontainer/customization/*

.gitmodules

@@ -0,0 +1,6 @@
+[submodule "docs/themes/hugo-theme-relearn"]
+	path = docs/themes/hugo-theme-relearn
+	url = https://github.com/McShelby/hugo-theme-relearn.git
+[submodule "docs/themes/lotusdocs"]
+	path = docs/themes/lotusdocs
+	url = https://github.com/colinwilson/lotusdocs

.yamllint

@@ -0,0 +1,4 @@
+extends: default
+
+rules:
+    line-length: disable

CONTRIBUTING.md

@@ -0,0 +1,84 @@
+# Contributing to LocalAI
+
+Thank you for your interest in contributing to LocalAI! We appreciate your time and effort in helping to improve our project. Before you get started, please take a moment to review these guidelines.
+
+## Table of Contents
+
+- [Getting Started](#getting-started)
+  - [Prerequisites](#prerequisites)
+  - [Setting up the Development Environment](#setting-up-the-development-environment)
+- [Contributing](#contributing)
+  - [Submitting an Issue](#submitting-an-issue)
+  - [Creating a Pull Request (PR)](#creating-a-pull-request-pr)
+- [Coding Guidelines](#coding-guidelines)
+- [Testing](#testing)
+- [Documentation](#documentation)
+- [Community and Communication](#community-and-communication)
+
+## Getting Started
+
+### Prerequisites
+
+- Golang [1.21]
+- Git
+- macOS/Linux
+
+### Setting up the Development Environment and running localAI in the local environment
+
+1. Clone the repository: `git clone https://github.com/go-skynet/LocalAI.git`
+2. Navigate to the project directory: `cd LocalAI`
+3. Install the required dependencies ( see https://localai.io/basics/build/#build-localai-locally )
+4. Build LocalAI: `make build`
+5. Run LocalAI: `./local-ai`
+
+## Contributing
+
+We welcome contributions from everyone! To get started, follow these steps:
+
+### Submitting an Issue
+
+If you find a bug, have a feature request, or encounter any issues, please check the [issue tracker](https://github.com/go-skynet/LocalAI/issues) to see if a similar issue has already been reported. If not, feel free to [create a new issue](https://github.com/go-skynet/LocalAI/issues/new) and provide as much detail as possible.
+
+### Creating a Pull Request (PR)
+
+1. Fork the repository.
+2. Create a new branch with a descriptive name: `git checkout -b [branch name]`
+3. Make your changes and commit them.
+4. Push the changes to your fork: `git push origin [branch name]`
+5. Create a new pull request from your branch to the main project's `main` or `master` branch.
+6. Provide a clear description of your changes in the pull request.
+7. Make any requested changes during the review process.
+8. Once your PR is approved, it will be merged into the main project.
+
+## Coding Guidelines
+
+- No specific coding guidelines at the moment. Please make sure the code can be tested. The most popular lint tools like [`golangci-lint`](https://golangci-lint.run) can help you here.
+
+## Testing
+
+`make test` cannot handle all the model now. Please be sure to add a test case for the new features or the part was changed.
+
+### Running AIO tests
+
+All-In-One images has a set of tests that automatically verifies that most of the endpoints works correctly, a flow can be :
+
+```bash
+# Build the LocalAI docker image
+make DOCKER_IMAGE=local-ai docker
+
+# Build the corresponding AIO image
+BASE_IMAGE=local-ai DOCKER_AIO_IMAGE=local-ai-aio:test make docker-aio
+
+# Run the AIO e2e tests
+LOCALAI_IMAGE_TAG=test LOCALAI_IMAGE=local-ai-aio make run-e2e-aio
+```
+
+## Documentation
+
+We are welcome the contribution of the documents, please open new PR or create a new issue. The documentation is available under `docs/` https://github.com/mudler/LocalAI/tree/master/docs
+ 
+## Community and Communication
+
+- You can reach out via the Github issue tracker.
+- Open a new discussion at [Discussion](https://github.com/go-skynet/LocalAI/discussions)
+- Join the Discord channel [Discord](https://discord.gg/uJAeKSAGDy)

Dockerfile

@@ -0,0 +1,491 @@
+ARG IMAGE_TYPE=extras
+ARG BASE_IMAGE=ubuntu:22.04
+ARG GRPC_BASE_IMAGE=${BASE_IMAGE}
+ARG INTEL_BASE_IMAGE=${BASE_IMAGE}
+
+# The requirements-core target is common to all images.  It should not be placed in requirements-core unless every single build will use it.
+FROM ${BASE_IMAGE} AS requirements-core
+
+USER root
+
+ARG GO_VERSION=1.22.6
+ARG CMAKE_VERSION=3.26.4
+ARG CMAKE_FROM_SOURCE=false
+ARG TARGETARCH
+ARG TARGETVARIANT
+
+ENV DEBIAN_FRONTEND=noninteractive
+ENV EXTERNAL_GRPC_BACKENDS="coqui:/build/backend/python/coqui/run.sh,huggingface-embeddings:/build/backend/python/sentencetransformers/run.sh,transformers:/build/backend/python/transformers/run.sh,sentencetransformers:/build/backend/python/sentencetransformers/run.sh,rerankers:/build/backend/python/rerankers/run.sh,autogptq:/build/backend/python/autogptq/run.sh,bark:/build/backend/python/bark/run.sh,diffusers:/build/backend/python/diffusers/run.sh,openvoice:/build/backend/python/openvoice/run.sh,vall-e-x:/build/backend/python/vall-e-x/run.sh,vllm:/build/backend/python/vllm/run.sh,mamba:/build/backend/python/mamba/run.sh,exllama2:/build/backend/python/exllama2/run.sh,transformers-musicgen:/build/backend/python/transformers-musicgen/run.sh,parler-tts:/build/backend/python/parler-tts/run.sh"
+
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        ccache \
+        ca-certificates \
+        curl libssl-dev \
+        git \
+        unzip upx-ucl && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install CMake (the version in 22.04 is too old)
+RUN <<EOT bash
+    if [ "${CMAKE_FROM_SOURCE}}" = "true" ]; then
+        curl -L -s https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}.tar.gz -o cmake.tar.gz && tar xvf cmake.tar.gz && cd cmake-${CMAKE_VERSION} && ./configure && make && make install
+    else
+        apt-get update && \
+        apt-get install -y \
+            cmake && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/*
+    fi
+EOT
+
+# Install Go
+RUN curl -L -s https://go.dev/dl/go${GO_VERSION}.linux-${TARGETARCH}.tar.gz | tar -C /usr/local -xz
+ENV PATH=$PATH:/root/go/bin:/usr/local/go/bin
+
+# Install grpc compilers
+RUN go install google.golang.org/protobuf/cmd/protoc-gen-go@v1.34.2 && \
+    go install google.golang.org/grpc/cmd/protoc-gen-go-grpc@1958fcbe2ca8bd93af633f11e97d44e567e945af
+
+COPY --chmod=644 custom-ca-certs/* /usr/local/share/ca-certificates/
+RUN update-ca-certificates
+
+RUN test -n "$TARGETARCH" \
+    || (echo 'warn: missing $TARGETARCH, either set this `ARG` manually, or run using `docker buildkit`')
+
+# Use the variables in subsequent instructions
+RUN echo "Target Architecture: $TARGETARCH"
+RUN echo "Target Variant: $TARGETVARIANT"
+
+# Cuda
+ENV PATH=/usr/local/cuda/bin:${PATH}
+
+# HipBLAS requirements
+ENV PATH=/opt/rocm/bin:${PATH}
+
+# OpenBLAS requirements and stable diffusion
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        libopenblas-dev \
+        libopencv-dev && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Set up OpenCV
+RUN ln -s /usr/include/opencv4/opencv2 /usr/include/opencv2
+
+WORKDIR /build
+
+###################################
+###################################
+
+# The requirements-extras target is for any builds with IMAGE_TYPE=extras. It should not be placed in this target unless every IMAGE_TYPE=extras build will use it
+FROM requirements-core AS requirements-extras
+
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        espeak-ng \
+        espeak \
+        python3-pip \
+        python-is-python3 \
+        python3-dev llvm \
+        python3-venv && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/* && \
+    pip install --upgrade pip
+
+# Install grpcio-tools (the version in 22.04 is too old)
+RUN pip install --user grpcio-tools
+
+###################################
+###################################
+
+# The requirements-drivers target is for BUILD_TYPE specific items.  If you need to install something specific to CUDA, or specific to ROCM, it goes here.
+# This target will be built on top of requirements-core or requirements-extras as retermined by the IMAGE_TYPE build-arg
+FROM requirements-${IMAGE_TYPE} AS requirements-drivers
+
+ARG BUILD_TYPE
+ARG CUDA_MAJOR_VERSION=12
+ARG CUDA_MINOR_VERSION=0
+
+ENV BUILD_TYPE=${BUILD_TYPE}
+
+# Vulkan requirements
+RUN <<EOT bash
+    if [ "${BUILD_TYPE}" = "vulkan" ]; then
+        apt-get update && \
+        apt-get install -y  --no-install-recommends \
+            software-properties-common pciutils wget gpg-agent && \
+        wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+        wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+        apt-get update && \
+        apt-get install -y \
+            vulkan-sdk && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/*
+    fi
+EOT
+
+# CuBLAS requirements
+RUN <<EOT bash
+    if [ "${BUILD_TYPE}" = "cublas" ]; then
+        apt-get update && \
+        apt-get install -y  --no-install-recommends \
+            software-properties-common pciutils
+        if [ "amd64" = "$TARGETARCH" ]; then
+            curl -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
+        fi
+        if [ "arm64" = "$TARGETARCH" ]; then
+            curl -O https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/arm64/cuda-keyring_1.1-1_all.deb
+        fi
+        dpkg -i cuda-keyring_1.1-1_all.deb && \
+        rm -f cuda-keyring_1.1-1_all.deb && \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
+            cuda-nvcc-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
+            libcufft-dev-${CUDA_MAJOR_VERSION}-${CUDA_MINOR_VERSION} \
+            libcurand-dev-${CUDA_MAJOR_VERSION}-${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 && \
+        rm -rf /var/lib/apt/lists/*
+    fi
+EOT
+
+# If we are building with clblas support, we need the libraries for the builds
+RUN if [ "${BUILD_TYPE}" = "clblas" ]; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
+            libclblast-dev && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/* \
+    ; fi
+
+RUN if [ "${BUILD_TYPE}" = "hipblas" ]; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
+            hipblas-dev \
+            rocblas-dev && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/* && \
+        # I have no idea why, but the ROCM lib packages don't trigger ldconfig after they install, which results in local-ai and others not being able
+        # to locate the libraries. We run ldconfig ourselves to work around this packaging deficiency
+        ldconfig \
+    ; fi
+
+###################################
+###################################
+
+# Temporary workaround for Intel's repository to work correctly
+# https://community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/APT-Repository-not-working-signatures-invalid/m-p/1599436/highlight/true#M36143
+# This is a temporary workaround until Intel fixes their repository
+FROM ${INTEL_BASE_IMAGE} AS intel
+RUN wget -qO - https://repositories.intel.com/gpu/intel-graphics.key | \
+gpg --yes --dearmor --output /usr/share/keyrings/intel-graphics.gpg
+RUN echo "deb [arch=amd64 signed-by=/usr/share/keyrings/intel-graphics.gpg] https://repositories.intel.com/gpu/ubuntu jammy/lts/2350 unified" > /etc/apt/sources.list.d/intel-graphics.list
+
+###################################
+###################################
+
+# The grpc target does one thing, it builds and installs GRPC.  This is in it's own layer so that it can be effectively cached by CI.
+# You probably don't need to change anything here, and if you do, make sure that CI is adjusted so that the cache continues to work.
+FROM ${GRPC_BASE_IMAGE} AS grpc
+
+# This is a bit of a hack, but it's required in order to be able to effectively cache this layer in CI
+ARG GRPC_MAKEFLAGS="-j4 -Otarget"
+ARG GRPC_VERSION=v1.65.0
+ARG CMAKE_FROM_SOURCE=false
+ARG CMAKE_VERSION=3.26.4
+
+ENV MAKEFLAGS=${GRPC_MAKEFLAGS}
+
+WORKDIR /build
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        ca-certificates \
+        build-essential curl libssl-dev \
+        git && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+# Install CMake (the version in 22.04 is too old)
+RUN <<EOT bash
+    if [ "${CMAKE_FROM_SOURCE}}" = "true" ]; then
+        curl -L -s https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}.tar.gz -o cmake.tar.gz && tar xvf cmake.tar.gz && cd cmake-${CMAKE_VERSION} && ./configure && make && make install
+    else
+        apt-get update && \
+        apt-get install -y \
+            cmake && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/*
+    fi
+EOT
+
+# We install GRPC to a different prefix here so that we can copy in only the build artifacts later
+# saves several hundred MB on the final docker image size vs copying in the entire GRPC source tree
+# and running make install in the target container
+RUN git clone --recurse-submodules --jobs 4 -b ${GRPC_VERSION} --depth 1 --shallow-submodules https://github.com/grpc/grpc && \
+    mkdir -p /build/grpc/cmake/build && \
+    cd /build/grpc/cmake/build && \
+    sed -i "216i\  TESTONLY" "../../third_party/abseil-cpp/absl/container/CMakeLists.txt" && \
+    cmake -DgRPC_INSTALL=ON -DgRPC_BUILD_TESTS=OFF -DCMAKE_INSTALL_PREFIX:PATH=/opt/grpc ../.. && \
+    make && \
+    make install && \
+    rm -rf /build
+
+###################################
+###################################
+
+# The builder-base target has the arguments, variables, and copies shared between full builder images and the uncompiled devcontainer
+
+FROM requirements-drivers AS builder-base
+
+ARG GO_TAGS="stablediffusion tts p2p"
+ARG GRPC_BACKENDS
+ARG MAKEFLAGS
+ARG LD_FLAGS="-s -w"
+
+ENV GRPC_BACKENDS=${GRPC_BACKENDS}
+ENV GO_TAGS=${GO_TAGS}
+ENV MAKEFLAGS=${MAKEFLAGS}
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_REQUIRE_CUDA="cuda>=${CUDA_MAJOR_VERSION}.0"
+ENV NVIDIA_VISIBLE_DEVICES=all
+ENV LD_FLAGS=${LD_FLAGS}
+
+RUN echo "GO_TAGS: $GO_TAGS" && echo "TARGETARCH: $TARGETARCH"
+
+WORKDIR /build
+
+
+# We need protoc installed, and the version in 22.04 is too old.  We will create one as part installing the GRPC build below
+# but that will also being in a newer version of absl which stablediffusion cannot compile with.  This version of protoc is only
+# here so that we can generate the grpc code for the stablediffusion build
+RUN <<EOT bash
+    if [ "amd64" = "$TARGETARCH" ]; then
+        curl -L -s https://github.com/protocolbuffers/protobuf/releases/download/v27.1/protoc-27.1-linux-x86_64.zip -o protoc.zip && \
+        unzip -j -d /usr/local/bin protoc.zip bin/protoc && \
+        rm protoc.zip
+    fi
+    if [ "arm64" = "$TARGETARCH" ]; then
+        curl -L -s https://github.com/protocolbuffers/protobuf/releases/download/v27.1/protoc-27.1-linux-aarch_64.zip -o protoc.zip && \
+        unzip -j -d /usr/local/bin protoc.zip bin/protoc && \
+        rm protoc.zip
+    fi
+EOT
+
+
+###################################
+###################################
+
+# This first portion of builder holds the layers specifically used to build backend-assets/grpc/stablediffusion
+# In most cases, builder is the image you should be using - however, this can save build time if one just needs to copy backend-assets/grpc/stablediffusion and nothing else.
+FROM builder-base AS builder-sd
+
+# stablediffusion does not tolerate a newer version of abseil, copy only over enough elements to build it
+COPY Makefile .
+COPY go.mod .
+COPY go.sum .
+COPY backend/backend.proto ./backend/backend.proto
+COPY backend/go/image/stablediffusion ./backend/go/image/stablediffusion
+COPY pkg/grpc ./pkg/grpc
+COPY pkg/stablediffusion ./pkg/stablediffusion
+RUN git init
+RUN make sources/go-stable-diffusion
+RUN touch prepare-sources
+
+# Actually build the backend
+RUN GRPC_BACKENDS=backend-assets/grpc/stablediffusion make backend-assets/grpc/stablediffusion
+
+###################################
+###################################
+
+# The builder target compiles LocalAI. This target is not the target that will be uploaded to the registry.
+# Adjustments to the build process should likely be made here.
+FROM builder-sd AS builder
+
+# Install the pre-built GRPC
+COPY --from=grpc /opt/grpc /usr/local
+
+# Rebuild with defaults backends
+WORKDIR /build
+
+COPY . .
+COPY .git .
+
+RUN make prepare
+
+## Build the binary
+## If it's CUDA or hipblas, we want to skip some of the llama-compat backends to save space
+## We only leave the most CPU-optimized variant and the fallback for the cublas/hipblas build
+## (both will use CUDA or hipblas for the actual computation)
+RUN if [ "${BUILD_TYPE}" = "cublas" ] || [ "${BUILD_TYPE}" = "hipblas" ]; then \
+        SKIP_GRPC_BACKEND="backend-assets/grpc/llama-cpp-avx backend-assets/grpc/llama-cpp-avx2" make build; \
+    else \
+        make build; \
+    fi
+
+RUN if [ ! -d "/build/sources/go-piper/piper-phonemize/pi/lib/" ]; then \
+        mkdir -p /build/sources/go-piper/piper-phonemize/pi/lib/ \
+        touch /build/sources/go-piper/piper-phonemize/pi/lib/keep \
+    ; fi
+
+###################################
+###################################
+
+# The devcontainer target is not used on CI. It is a target for developers to use locally -
+# rather than copying files it mounts them locally and leaves building to the developer
+
+FROM builder-base AS devcontainer
+
+ARG FFMPEG
+
+COPY --from=grpc /opt/grpc /usr/local
+
+COPY --from=builder-sd /build/backend-assets/grpc/stablediffusion /build/backend-assets/grpc/stablediffusion
+
+COPY .devcontainer-scripts /.devcontainer-scripts
+
+# Add FFmpeg
+RUN if [ "${FFMPEG}" = "true" ]; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
+            ffmpeg && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/* \
+    ; fi
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends \
+        ssh less wget
+# For the devcontainer, leave apt functional in case additional devtools are needed at runtime.
+
+RUN go install github.com/go-delve/delve/cmd/dlv@latest
+
+RUN go install github.com/mikefarah/yq/v4@latest
+
+###################################
+###################################
+
+# This is the final target. The result of this target will be the image uploaded to the registry.
+# If you cannot find a more suitable place for an addition, this layer is a suitable place for it.
+FROM requirements-drivers
+
+ARG FFMPEG
+ARG BUILD_TYPE
+ARG TARGETARCH
+ARG IMAGE_TYPE=extras
+ARG EXTRA_BACKENDS
+ARG MAKEFLAGS
+
+ENV BUILD_TYPE=${BUILD_TYPE}
+ENV REBUILD=false
+ENV HEALTHCHECK_ENDPOINT=http://localhost:8080/readyz
+ENV MAKEFLAGS=${MAKEFLAGS}
+
+ARG CUDA_MAJOR_VERSION=12
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+ENV NVIDIA_REQUIRE_CUDA="cuda>=${CUDA_MAJOR_VERSION}.0"
+ENV NVIDIA_VISIBLE_DEVICES=all
+
+# Add FFmpeg
+RUN if [ "${FFMPEG}" = "true" ]; then \
+        apt-get update && \
+        apt-get install -y --no-install-recommends \
+            ffmpeg && \
+        apt-get clean && \
+        rm -rf /var/lib/apt/lists/* \
+    ; fi
+
+WORKDIR /build
+
+# we start fresh & re-copy all assets because `make build` does not clean up nicely after itself
+# so when `entrypoint.sh` runs `make build` again (which it does by default), the build would fail
+# see https://github.com/go-skynet/LocalAI/pull/658#discussion_r1241971626 and
+# https://github.com/go-skynet/LocalAI/pull/434
+COPY . .
+
+COPY --from=builder /build/sources ./sources/
+COPY --from=grpc /opt/grpc /usr/local
+
+RUN make prepare-sources
+
+# Copy the binary
+COPY --from=builder /build/local-ai ./
+
+# Copy shared libraries for piper
+COPY --from=builder /build/sources/go-piper/piper-phonemize/pi/lib/* /usr/lib/
+
+# do not let stablediffusion rebuild (requires an older version of absl)
+COPY --from=builder-sd /build/backend-assets/grpc/stablediffusion ./backend-assets/grpc/stablediffusion
+
+# Change the shell to bash so we can use [[ tests below
+SHELL ["/bin/bash", "-c"]
+# We try to strike a balance between individual layer size (as that affects total push time) and total image size
+# Splitting the backends into more groups with fewer items results in a larger image, but a smaller size for the largest layer
+# Splitting the backends into fewer groups with more items results in a smaller image, but a larger size for the largest layer
+
+RUN if [[ ( "${EXTRA_BACKENDS}" =~ "coqui" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/coqui \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "parler-tts" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/parler-tts \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "diffusers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/diffusers \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "transformers-musicgen" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/transformers-musicgen \
+    ; fi
+
+RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vall-e-x" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/vall-e-x \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "openvoice" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/openvoice \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "sentencetransformers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/sentencetransformers \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "exllama2" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/exllama2 \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "transformers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/transformers \
+    ; fi
+
+RUN if [[ ( "${EXTRA_BACKENDS}" =~ "vllm" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/vllm \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "autogptq" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/autogptq \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "bark" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/bark \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "rerankers" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/rerankers \
+    ; fi && \
+    if [[ ( "${EXTRA_BACKENDS}" =~ "mamba" || -z "${EXTRA_BACKENDS}" ) && "$IMAGE_TYPE" == "extras" ]]; then \
+        make -C backend/python/mamba \
+    ; fi
+
+# Make sure the models directory exists
+RUN mkdir -p /build/models
+
+# Define the health check command
+HEALTHCHECK --interval=1m --timeout=10m --retries=10 \
+  CMD curl -f ${HEALTHCHECK_ENDPOINT} || exit 1
+
+VOLUME /build/models
+EXPOSE 8080
+ENTRYPOINT [ "/build/entrypoint.sh" ]

Dockerfile.aio

@@ -0,0 +1,8 @@
+ARG BASE_IMAGE=ubuntu:22.04
+
+FROM ${BASE_IMAGE} 
+
+RUN apt-get update && apt-get install -y pciutils && apt-get clean
+
+COPY aio/ /aio
+ENTRYPOINT [ "/aio/entrypoint.sh" ]

Earthfile

@@ -0,0 +1,5 @@
+VERSION 0.7
+
+build:
+    FROM DOCKERFILE -f Dockerfile .
+    SAVE ARTIFACT /usr/bin/local-ai AS LOCAL local-ai

Entitlements.plist

@@ -0,0 +1,10 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
+<plist version="1.0">
+<dict>
+    <key>com.apple.security.network.client</key>
+    <true/>
+    <key>com.apple.security.network.server</key>
+    <true/>
+</dict>
+</plist>

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023-2024 Ettore Di Giacinto (mudler@localai.io)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

Makefile

@@ -0,0 +1,932 @@
+GOCMD=go
+GOTEST=$(GOCMD) test
+GOVET=$(GOCMD) vet
+BINARY_NAME=local-ai
+
+DETECT_LIBS?=true
+
+# llama.cpp versions
+GOLLAMA_REPO?=https://github.com/go-skynet/go-llama.cpp
+GOLLAMA_VERSION?=2b57a8ae43e4699d3dc5d1496a1ccd42922993be
+CPPLLAMA_VERSION?=d5a409e57fe8bd24fef597ab8a31110d390a6392
+
+# 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?=0377596b77a3602e36430320cbe45f8c305ef04a
+
+# 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?=e10ca041a885d4a8f3871d52924b47792d5e5aa0
+
+# stablediffusion version
+STABLEDIFFUSION_REPO?=https://github.com/mudler/go-stable-diffusion
+STABLEDIFFUSION_VERSION?=4a3cd6aeae6f66ee57eae9a0075f8c58c3a6a38f
+
+# tinydream version
+TINYDREAM_REPO?=https://github.com/M0Rf30/go-tiny-dream
+TINYDREAM_VERSION?=c04fa463ace9d9a6464313aa5f9cd0f953b6c057
+
+export BUILD_TYPE?=
+export STABLE_BUILD_TYPE?=$(BUILD_TYPE)
+export CMAKE_ARGS?=
+export BACKEND_LIBS?=
+
+CGO_LDFLAGS?=
+CGO_LDFLAGS_WHISPER?=
+CGO_LDFLAGS_WHISPER+=-lggml
+CUDA_LIBPATH?=/usr/local/cuda/lib64/
+GO_TAGS?=
+BUILD_ID?=
+
+TEST_DIR=/tmp/test
+
+TEST_FLAKES?=5
+
+RANDOM := $(shell bash -c 'echo $$RANDOM')
+
+VERSION?=$(shell git describe --always --tags || echo "dev" )
+# go tool nm ./local-ai | grep Commit
+LD_FLAGS?=-s -w
+override LD_FLAGS += -X "github.com/mudler/LocalAI/internal.Version=$(VERSION)"
+override LD_FLAGS += -X "github.com/mudler/LocalAI/internal.Commit=$(shell git rev-parse HEAD)"
+
+OPTIONAL_TARGETS?=
+
+export OS := $(shell uname -s)
+ARCH := $(shell uname -m)
+GREEN  := $(shell tput -Txterm setaf 2)
+YELLOW := $(shell tput -Txterm setaf 3)
+WHITE  := $(shell tput -Txterm setaf 7)
+CYAN   := $(shell tput -Txterm setaf 6)
+RESET  := $(shell tput -Txterm sgr0)
+
+UPX?=
+# check if upx exists
+ifeq (, $(shell which upx))
+	UPX=
+else
+	UPX=$(shell which upx)
+endif
+
+# Default Docker bridge IP
+E2E_BRIDGE_IP?=172.17.0.1
+
+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifeq ($(OS),Darwin)
+
+	ifeq ($(OSX_SIGNING_IDENTITY),)
+		OSX_SIGNING_IDENTITY := $(shell security find-identity -v -p codesigning | grep '"' | head -n 1 | sed -E 's/.*"(.*)"/\1/')
+	endif
+
+	# on OSX, if BUILD_TYPE is blank, we should default to use Metal
+	ifeq ($(BUILD_TYPE),)
+		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
+		export GGML_NO_ACCELERATE=1
+		export GGML_NO_METAL=1
+	endif
+
+	ifeq ($(BUILD_TYPE),metal)
+#			-lcblas 	removed: it seems to always be listed as a duplicate flag.
+		CGO_LDFLAGS += -framework Accelerate
+	endif
+else
+CGO_LDFLAGS_WHISPER+=-lgomp
+endif
+
+ifeq ($(BUILD_TYPE),openblas)
+	CGO_LDFLAGS+=-lopenblas
+	export GGML_OPENBLAS=1
+endif
+
+ifeq ($(BUILD_TYPE),cublas)
+	CGO_LDFLAGS+=-lcublas -lcudart -L$(CUDA_LIBPATH)
+	export GGML_CUDA=1
+	CGO_LDFLAGS_WHISPER+=-L$(CUDA_LIBPATH)/stubs/ -lcuda -lcufft
+endif
+
+ifeq ($(BUILD_TYPE),vulkan)
+	CMAKE_ARGS+=-DGGML_VULKAN=1
+endif
+
+ifneq (,$(findstring sycl,$(BUILD_TYPE)))
+	export GGML_SYCL=1
+endif
+
+ifeq ($(BUILD_TYPE),sycl_f16)
+	export GGML_SYCL_F16=1
+endif
+
+ifeq ($(BUILD_TYPE),hipblas)
+	ROCM_HOME ?= /opt/rocm
+	ROCM_PATH ?= /opt/rocm
+	LD_LIBRARY_PATH ?= /opt/rocm/lib:/opt/rocm/llvm/lib
+	export CXX=$(ROCM_HOME)/llvm/bin/clang++
+	export CC=$(ROCM_HOME)/llvm/bin/clang
+	# llama-ggml has no hipblas support, so override it here.
+	export STABLE_BUILD_TYPE=
+	export GGML_HIPBLAS=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)"
+	CGO_LDFLAGS += -O3 --rtlib=compiler-rt -unwindlib=libgcc -lhipblas -lrocblas --hip-link -L${ROCM_HOME}/lib/llvm/lib
+endif
+
+ifeq ($(BUILD_TYPE),metal)
+	CGO_LDFLAGS+=-framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
+	export GGML_METAL=1
+endif
+
+ifeq ($(BUILD_TYPE),clblas)
+	CGO_LDFLAGS+=-lOpenCL -lclblast
+	export GGML_OPENBLAS=1
+endif
+
+# glibc-static or glibc-devel-static required
+ifeq ($(STATIC),true)
+	LD_FLAGS+=-linkmode external -extldflags -static
+endif
+
+ifeq ($(findstring stablediffusion,$(GO_TAGS)),stablediffusion)
+#	OPTIONAL_TARGETS+=go-stable-diffusion/libstablediffusion.a
+	OPTIONAL_GRPC+=backend-assets/grpc/stablediffusion
+endif
+
+ifeq ($(findstring tinydream,$(GO_TAGS)),tinydream)
+#	OPTIONAL_TARGETS+=go-tiny-dream/libtinydream.a
+	OPTIONAL_GRPC+=backend-assets/grpc/tinydream
+endif
+
+ifeq ($(findstring tts,$(GO_TAGS)),tts)
+#	OPTIONAL_TARGETS+=go-piper/libpiper_binding.a
+#	OPTIONAL_TARGETS+=backend-assets/espeak-ng-data
+	PIPER_CGO_CXXFLAGS+=-I$(CURDIR)/sources/go-piper/piper/src/cpp -I$(CURDIR)/sources/go-piper/piper/build/fi/include -I$(CURDIR)/sources/go-piper/piper/build/pi/include -I$(CURDIR)/sources/go-piper/piper/build/si/include
+	PIPER_CGO_LDFLAGS+=-L$(CURDIR)/sources/go-piper/piper/build/fi/lib -L$(CURDIR)/sources/go-piper/piper/build/pi/lib -L$(CURDIR)/sources/go-piper/piper/build/si/lib -lfmt -lspdlog -lucd
+	OPTIONAL_GRPC+=backend-assets/grpc/piper
+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
+ALL_GRPC_BACKENDS+=backend-assets/grpc/local-store
+ALL_GRPC_BACKENDS+=$(OPTIONAL_GRPC)
+# Use filter-out to remove the specified backends
+ALL_GRPC_BACKENDS := $(filter-out $(SKIP_GRPC_BACKEND),$(ALL_GRPC_BACKENDS))
+
+GRPC_BACKENDS?=$(ALL_GRPC_BACKENDS) $(OPTIONAL_GRPC)
+TEST_PATHS?=./api/... ./pkg/... ./core/...
+
+# If empty, then we build all
+ifeq ($(GRPC_BACKENDS),)
+	GRPC_BACKENDS=$(ALL_GRPC_BACKENDS)
+endif
+
+ifeq ($(BUILD_API_ONLY),true)
+	GRPC_BACKENDS=
+endif
+
+.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
+	cd sources/go-llama.cpp && \
+	git init && \
+	git remote add origin $(GOLLAMA_REPO) && \
+	git fetch origin && \
+	git checkout $(GOLLAMA_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-llama.cpp/libbinding.a: sources/go-llama.cpp
+	$(MAKE) -C sources/go-llama.cpp BUILD_TYPE=$(STABLE_BUILD_TYPE) libbinding.a
+
+## go-piper
+sources/go-piper:
+	mkdir -p sources/go-piper
+	cd sources/go-piper && \
+	git init && \
+	git remote add origin $(PIPER_REPO) && \
+	git fetch origin && \
+	git checkout $(PIPER_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-piper/libpiper_binding.a: sources/go-piper
+	$(MAKE) -C sources/go-piper libpiper_binding.a example/main piper.o
+
+
+## RWKV
+sources/go-rwkv.cpp:
+	mkdir -p sources/go-rwkv.cpp
+	cd sources/go-rwkv.cpp && \
+	git init && \
+	git remote add origin $(RWKV_REPO) && \
+	git fetch origin && \
+	git checkout $(RWKV_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-rwkv.cpp/librwkv.a: sources/go-rwkv.cpp
+	cd sources/go-rwkv.cpp && cd rwkv.cpp &&	cmake . -DRWKV_BUILD_SHARED_LIBRARY=OFF &&	cmake --build . && 	cp librwkv.a ..
+
+## stable diffusion
+sources/go-stable-diffusion:
+	mkdir -p sources/go-stable-diffusion
+	cd sources/go-stable-diffusion && \
+	git init && \
+	git remote add origin $(STABLEDIFFUSION_REPO) && \
+	git fetch origin && \
+	git checkout $(STABLEDIFFUSION_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-stable-diffusion/libstablediffusion.a: sources/go-stable-diffusion
+	CPATH="$(CPATH):/usr/include/opencv4" $(MAKE) -C sources/go-stable-diffusion libstablediffusion.a
+
+## tiny-dream
+sources/go-tiny-dream:
+	mkdir -p sources/go-tiny-dream
+	cd sources/go-tiny-dream && \
+	git init && \
+	git remote add origin $(TINYDREAM_REPO) && \
+	git fetch origin && \
+	git checkout $(TINYDREAM_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+sources/go-tiny-dream/libtinydream.a: sources/go-tiny-dream
+	$(MAKE) -C sources/go-tiny-dream libtinydream.a
+
+## whisper
+sources/whisper.cpp:
+	mkdir -p sources/whisper.cpp
+	cd sources/whisper.cpp && \
+	git init && \
+	git remote add origin $(WHISPER_REPO) && \
+	git fetch origin && \
+	git checkout $(WHISPER_CPP_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+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
+
+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
+	$(GOCMD) mod edit -dropreplace github.com/go-skynet/go-llama.cpp
+
+prepare-sources: get-sources replace
+	$(GOCMD) mod download
+
+## GENERIC
+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
+
+prepare: prepare-sources $(OPTIONAL_TARGETS)
+
+clean: ## Remove build related file
+	$(GOCMD) clean -cache
+	rm -f prepare
+	rm -rf ./sources
+	rm -rf $(BINARY_NAME)
+	rm -rf release/
+	rm -rf backend-assets/*
+	$(MAKE) -C backend/cpp/grpc clean
+	$(MAKE) -C backend/cpp/llama clean
+	rm -rf backend/cpp/llama-* || true
+	$(MAKE) dropreplace
+	$(MAKE) protogen-clean
+	rmdir pkg/grpc/proto || true
+
+clean-tests:
+	rm -rf test-models
+	rm -rf test-dir
+	rm -rf core/http/backend-assets
+
+clean-dc: clean
+	cp -r /build/backend-assets /workspace/backend-assets
+
+## Build:
+build: prepare backend-assets grpcs ## Build the project
+	$(info ${GREEN}I local-ai build info:${RESET})
+	$(info ${GREEN}I BUILD_TYPE: ${YELLOW}$(BUILD_TYPE)${RESET})
+	$(info ${GREEN}I GO_TAGS: ${YELLOW}$(GO_TAGS)${RESET})
+	$(info ${GREEN}I LD_FLAGS: ${YELLOW}$(LD_FLAGS)${RESET})
+	$(info ${GREEN}I UPX: ${YELLOW}$(UPX)${RESET})
+ifneq ($(BACKEND_LIBS),)
+	$(MAKE) backend-assets/lib
+	cp -f $(BACKEND_LIBS) backend-assets/lib/
+endif
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o $(BINARY_NAME) ./
+
+build-minimal:
+	BUILD_GRPC_FOR_BACKEND_LLAMA=true GRPC_BACKENDS="backend-assets/grpc/llama-cpp-avx2" GO_TAGS=p2p $(MAKE) build
+
+build-api:
+	BUILD_GRPC_FOR_BACKEND_LLAMA=true BUILD_API_ONLY=true GO_TAGS=p2p $(MAKE) build
+
+backend-assets/lib:
+	mkdir -p backend-assets/lib
+
+dist:
+	$(MAKE) backend-assets/grpc/llama-cpp-avx2
+ifeq ($(DETECT_LIBS),true)
+	scripts/prepare-libs.sh backend-assets/grpc/llama-cpp-avx2
+endif
+ifeq ($(OS),Darwin)
+	BUILD_TYPE=none $(MAKE) backend-assets/grpc/llama-cpp-fallback
+else
+	$(MAKE) backend-assets/grpc/llama-cpp-cuda
+	$(MAKE) backend-assets/grpc/llama-cpp-hipblas
+	$(MAKE) backend-assets/grpc/llama-cpp-sycl_f16
+	$(MAKE) backend-assets/grpc/llama-cpp-sycl_f32
+endif
+	GO_TAGS="tts p2p" $(MAKE) build
+ifeq ($(DETECT_LIBS),true)
+	scripts/prepare-libs.sh backend-assets/grpc/piper
+endif
+	GO_TAGS="tts p2p" STATIC=true $(MAKE) build
+	mkdir -p release
+# if BUILD_ID is empty, then we don't append it to the binary name
+ifeq ($(BUILD_ID),)
+	cp $(BINARY_NAME) release/$(BINARY_NAME)-$(OS)-$(ARCH)
+	shasum -a 256 release/$(BINARY_NAME)-$(OS)-$(ARCH) > release/$(BINARY_NAME)-$(OS)-$(ARCH).sha256
+else
+	cp $(BINARY_NAME) release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-$(ARCH)
+	shasum -a 256 release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-$(ARCH) > release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-$(ARCH).sha256
+endif
+
+dist-cross-linux-arm64:
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_NATIVE=off" GRPC_BACKENDS="backend-assets/grpc/llama-cpp-fallback backend-assets/grpc/llama-cpp-grpc backend-assets/util/llama-cpp-rpc-server" GO_TAGS="p2p" \
+	STATIC=true $(MAKE) build
+	mkdir -p release
+# if BUILD_ID is empty, then we don't append it to the binary name
+ifeq ($(BUILD_ID),)
+	cp $(BINARY_NAME) release/$(BINARY_NAME)-$(OS)-arm64
+	shasum -a 256 release/$(BINARY_NAME)-$(OS)-arm64 > release/$(BINARY_NAME)-$(OS)-arm64.sha256
+else
+	cp $(BINARY_NAME) release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-arm64
+	shasum -a 256 release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-arm64 > release/$(BINARY_NAME)-$(BUILD_ID)-$(OS)-arm64.sha256
+endif
+
+osx-signed: build
+	codesign --deep --force --sign "$(OSX_SIGNING_IDENTITY)" --entitlements "./Entitlements.plist" "./$(BINARY_NAME)"
+
+## Run
+run: prepare ## run local-ai
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) run ./
+
+test-models/testmodel.ggml:
+	mkdir test-models
+	mkdir test-dir
+	wget -q https://huggingface.co/TheBloke/orca_mini_3B-GGML/resolve/main/orca-mini-3b.ggmlv3.q4_0.bin -O test-models/testmodel.ggml
+	wget -q https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.en.bin -O test-models/whisper-en
+	wget -q https://huggingface.co/mudler/all-MiniLM-L6-v2/resolve/main/ggml-model-q4_0.bin -O test-models/bert
+	wget -q https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav -O test-dir/audio.wav
+	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
+	cp -rf backend-assets core/http
+	cp tests/models_fixtures/* test-models
+
+test: prepare test-models/testmodel.ggml grpcs
+	@echo 'Running tests'
+	export GO_TAGS="tts stablediffusion debug"
+	$(MAKE) prepare-test
+	HUGGINGFACE_GRPC=$(abspath ./)/backend/python/sentencetransformers/run.sh TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="!llama && !llama-gguf"  --flake-attempts $(TEST_FLAKES) --fail-fast -v -r $(TEST_PATHS)
+	$(MAKE) test-llama
+	$(MAKE) test-llama-gguf
+	$(MAKE) test-tts
+	$(MAKE) test-stablediffusion
+
+prepare-e2e:
+	mkdir -p $(TEST_DIR)
+	cp -rfv $(abspath ./tests/e2e-fixtures)/gpu.yaml $(TEST_DIR)/gpu.yaml
+	test -e $(TEST_DIR)/ggllm-test-model.bin || wget -q https://huggingface.co/TheBloke/CodeLlama-7B-Instruct-GGUF/resolve/main/codellama-7b-instruct.Q2_K.gguf -O $(TEST_DIR)/ggllm-test-model.bin
+	docker build --build-arg GRPC_BACKENDS="$(GRPC_BACKENDS)" --build-arg IMAGE_TYPE=core --build-arg BUILD_TYPE=$(BUILD_TYPE) --build-arg CUDA_MAJOR_VERSION=12 --build-arg CUDA_MINOR_VERSION=0 --build-arg FFMPEG=true -t localai-tests .
+
+run-e2e-image:
+	ls -liah $(abspath ./tests/e2e-fixtures)
+	docker run -p 5390:8080 -e MODELS_PATH=/models -e THREADS=1 -e DEBUG=true -d --rm -v $(TEST_DIR):/models --gpus all --name e2e-tests-$(RANDOM) localai-tests
+
+run-e2e-aio: protogen-go
+	@echo 'Running e2e AIO tests'
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --flake-attempts $(TEST_FLAKES) -v -r ./tests/e2e-aio
+
+test-e2e:
+	@echo 'Running e2e tests'
+	BUILD_TYPE=$(BUILD_TYPE) \
+	LOCALAI_API=http://$(E2E_BRIDGE_IP):5390/v1 \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --flake-attempts $(TEST_FLAKES) -v -r ./tests/e2e
+
+teardown-e2e:
+	rm -rf $(TEST_DIR) || true
+	docker stop $$(docker ps -q --filter ancestor=localai-tests)
+
+test-llama: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
+
+test-llama-gguf: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="llama-gguf" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
+
+test-tts: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="tts" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
+
+test-stablediffusion: prepare-test
+	TEST_DIR=$(abspath ./)/test-dir/ FIXTURES=$(abspath ./)/tests/fixtures CONFIG_FILE=$(abspath ./)/test-models/config.yaml MODELS_PATH=$(abspath ./)/test-models \
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="stablediffusion" --flake-attempts $(TEST_FLAKES) -v -r $(TEST_PATHS)
+
+test-stores: backend-assets/grpc/local-store
+	mkdir -p tests/integration/backend-assets/grpc
+	cp -f backend-assets/grpc/local-store tests/integration/backend-assets/grpc/
+	$(GOCMD) run github.com/onsi/ginkgo/v2/ginkgo --label-filter="stores" --flake-attempts $(TEST_FLAKES) -v -r tests/integration
+
+test-container:
+	docker build --target requirements -t local-ai-test-container .
+	docker run -ti --rm --entrypoint /bin/bash -ti -v $(abspath ./):/build local-ai-test-container
+
+## Help:
+help: ## Show this help.
+	@echo ''
+	@echo 'Usage:'
+	@echo '  ${YELLOW}make${RESET} ${GREEN}<target>${RESET}'
+	@echo ''
+	@echo 'Targets:'
+	@awk 'BEGIN {FS = ":.*?## "} { \
+		if (/^[a-zA-Z_-]+:.*?##.*$$/) {printf "    ${YELLOW}%-20s${GREEN}%s${RESET}\n", $$1, $$2} \
+		else if (/^## .*$$/) {printf "  ${CYAN}%s${RESET}\n", substr($$1,4)} \
+		}' $(MAKEFILE_LIST)
+
+.PHONY: protogen
+protogen: protogen-go protogen-python
+
+.PHONY: protogen-clean
+protogen-clean: protogen-go-clean protogen-python-clean
+
+.PHONY: protogen-go
+protogen-go:
+	mkdir -p pkg/grpc/proto
+	protoc --experimental_allow_proto3_optional -Ibackend/ --go_out=pkg/grpc/proto/ --go_opt=paths=source_relative --go-grpc_out=pkg/grpc/proto/ --go-grpc_opt=paths=source_relative \
+    backend/backend.proto
+
+.PHONY: protogen-go-clean
+protogen-go-clean:
+	$(RM) pkg/grpc/proto/backend.pb.go pkg/grpc/proto/backend_grpc.pb.go
+	$(RM) bin/*
+
+.PHONY: protogen-python
+protogen-python: autogptq-protogen bark-protogen coqui-protogen diffusers-protogen exllama2-protogen mamba-protogen rerankers-protogen sentencetransformers-protogen transformers-protogen parler-tts-protogen transformers-musicgen-protogen vall-e-x-protogen vllm-protogen openvoice-protogen
+
+.PHONY: protogen-python-clean
+protogen-python-clean: autogptq-protogen-clean bark-protogen-clean coqui-protogen-clean diffusers-protogen-clean  exllama2-protogen-clean mamba-protogen-clean sentencetransformers-protogen-clean rerankers-protogen-clean transformers-protogen-clean transformers-musicgen-protogen-clean parler-tts-protogen-clean vall-e-x-protogen-clean vllm-protogen-clean openvoice-protogen-clean
+
+.PHONY: autogptq-protogen
+autogptq-protogen:
+	$(MAKE) -C backend/python/autogptq protogen
+
+.PHONY: autogptq-protogen-clean
+autogptq-protogen-clean:
+	$(MAKE) -C backend/python/autogptq protogen-clean
+
+.PHONY: bark-protogen
+bark-protogen:
+	$(MAKE) -C backend/python/bark protogen
+
+.PHONY: bark-protogen-clean
+bark-protogen-clean:
+	$(MAKE) -C backend/python/bark protogen-clean
+
+.PHONY: coqui-protogen
+coqui-protogen:
+	$(MAKE) -C backend/python/coqui protogen
+
+.PHONY: coqui-protogen-clean
+coqui-protogen-clean:
+	$(MAKE) -C backend/python/coqui protogen-clean
+
+.PHONY: diffusers-protogen
+diffusers-protogen:
+	$(MAKE) -C backend/python/diffusers protogen
+
+.PHONY: diffusers-protogen-clean
+diffusers-protogen-clean:
+	$(MAKE) -C backend/python/diffusers protogen-clean
+
+.PHONY: exllama2-protogen
+exllama2-protogen:
+	$(MAKE) -C backend/python/exllama2 protogen
+
+.PHONY: exllama2-protogen-clean
+exllama2-protogen-clean:
+	$(MAKE) -C backend/python/exllama2 protogen-clean
+
+.PHONY: mamba-protogen
+mamba-protogen:
+	$(MAKE) -C backend/python/mamba protogen
+
+.PHONY: mamba-protogen-clean
+mamba-protogen-clean:
+	$(MAKE) -C backend/python/mamba protogen-clean
+
+.PHONY: rerankers-protogen
+rerankers-protogen:
+	$(MAKE) -C backend/python/rerankers protogen
+
+.PHONY: rerankers-protogen-clean
+rerankers-protogen-clean:
+	$(MAKE) -C backend/python/rerankers protogen-clean
+
+.PHONY: sentencetransformers-protogen
+sentencetransformers-protogen:
+	$(MAKE) -C backend/python/sentencetransformers protogen
+
+.PHONY: sentencetransformers-protogen-clean
+sentencetransformers-protogen-clean:
+	$(MAKE) -C backend/python/sentencetransformers protogen-clean
+
+.PHONY: transformers-protogen
+transformers-protogen:
+	$(MAKE) -C backend/python/transformers protogen
+
+.PHONY: transformers-protogen-clean
+transformers-protogen-clean:
+	$(MAKE) -C backend/python/transformers protogen-clean
+
+.PHONY: parler-tts-protogen
+parler-tts-protogen:
+	$(MAKE) -C backend/python/parler-tts protogen
+
+.PHONY: parler-tts-protogen-clean
+parler-tts-protogen-clean:
+	$(MAKE) -C backend/python/parler-tts protogen-clean
+
+.PHONY: transformers-musicgen-protogen
+transformers-musicgen-protogen:
+	$(MAKE) -C backend/python/transformers-musicgen protogen
+
+.PHONY: transformers-musicgen-protogen-clean
+transformers-musicgen-protogen-clean:
+	$(MAKE) -C backend/python/transformers-musicgen protogen-clean
+
+.PHONY: vall-e-x-protogen
+vall-e-x-protogen:
+	$(MAKE) -C backend/python/vall-e-x protogen
+
+.PHONY: vall-e-x-protogen-clean
+vall-e-x-protogen-clean:
+	$(MAKE) -C backend/python/vall-e-x protogen-clean
+
+.PHONY: openvoice-protogen
+openvoice-protogen:
+	$(MAKE) -C backend/python/openvoice protogen
+
+.PHONY: openvoice-protogen-clean
+openvoice-protogen-clean:
+	$(MAKE) -C backend/python/openvoice protogen-clean
+
+.PHONY: vllm-protogen
+vllm-protogen:
+	$(MAKE) -C backend/python/vllm protogen
+
+.PHONY: vllm-protogen-clean
+vllm-protogen-clean:
+	$(MAKE) -C backend/python/vllm protogen-clean
+
+## GRPC
+# Note: it is duplicated in the Dockerfile
+prepare-extra-conda-environments: protogen-python
+	$(MAKE) -C backend/python/autogptq
+	$(MAKE) -C backend/python/bark
+	$(MAKE) -C backend/python/coqui
+	$(MAKE) -C backend/python/diffusers
+	$(MAKE) -C backend/python/vllm
+	$(MAKE) -C backend/python/mamba
+	$(MAKE) -C backend/python/sentencetransformers
+	$(MAKE) -C backend/python/rerankers
+	$(MAKE) -C backend/python/transformers
+	$(MAKE) -C backend/python/transformers-musicgen
+	$(MAKE) -C backend/python/parler-tts
+	$(MAKE) -C backend/python/vall-e-x
+	$(MAKE) -C backend/python/openvoice
+	$(MAKE) -C backend/python/exllama2
+
+prepare-test-extra: protogen-python
+	$(MAKE) -C backend/python/transformers
+	$(MAKE) -C backend/python/diffusers
+
+test-extra: prepare-test-extra
+	$(MAKE) -C backend/python/transformers test
+	$(MAKE) -C backend/python/diffusers test
+
+backend-assets:
+	mkdir -p backend-assets
+ifeq ($(BUILD_API_ONLY),true)
+	touch backend-assets/keep
+endif
+
+backend-assets/espeak-ng-data: sources/go-piper sources/go-piper/libpiper_binding.a
+	mkdir -p backend-assets/espeak-ng-data
+	@cp -rf sources/go-piper/piper-phonemize/pi/share/espeak-ng-data/. backend-assets/espeak-ng-data
+
+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),)
+	$(UPX) backend-assets/grpc/huggingface
+endif
+
+backend/cpp/llama/llama.cpp:
+	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/llama llama.cpp
+
+INSTALLED_PACKAGES=$(CURDIR)/backend/cpp/grpc/installed_packages
+INSTALLED_LIB_CMAKE=$(INSTALLED_PACKAGES)/lib/cmake
+ADDED_CMAKE_ARGS=-Dabsl_DIR=${INSTALLED_LIB_CMAKE}/absl \
+				 -DProtobuf_DIR=${INSTALLED_LIB_CMAKE}/protobuf \
+				 -Dutf8_range_DIR=${INSTALLED_LIB_CMAKE}/utf8_range \
+				 -DgRPC_DIR=${INSTALLED_LIB_CMAKE}/grpc \
+				 -DCMAKE_CXX_STANDARD_INCLUDE_DIRECTORIES=${INSTALLED_PACKAGES}/include
+build-llama-cpp-grpc-server:
+# Conditionally build grpc for the llama backend to use if needed
+ifdef BUILD_GRPC_FOR_BACKEND_LLAMA
+	$(MAKE) -C backend/cpp/grpc build
+	_PROTOBUF_PROTOC=${INSTALLED_PACKAGES}/bin/proto \
+	_GRPC_CPP_PLUGIN_EXECUTABLE=${INSTALLED_PACKAGES}/bin/grpc_cpp_plugin \
+	PATH="${INSTALLED_PACKAGES}/bin:${PATH}" \
+	CMAKE_ARGS="${CMAKE_ARGS} ${ADDED_CMAKE_ARGS}" \
+	LLAMA_VERSION=$(CPPLLAMA_VERSION) \
+	$(MAKE) -C backend/cpp/${VARIANT} grpc-server
+else
+	echo "BUILD_GRPC_FOR_BACKEND_LLAMA is not defined."
+	LLAMA_VERSION=$(CPPLLAMA_VERSION) $(MAKE) -C backend/cpp/${VARIANT} grpc-server
+endif
+
+# This target is for manually building a variant with-auto detected flags
+backend-assets/grpc/llama-cpp: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-cpp
+	$(MAKE) -C backend/cpp/llama-cpp purge
+	$(info ${GREEN}I llama-cpp build info:avx2${RESET})
+	$(MAKE) VARIANT="llama-cpp" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-cpp/grpc-server backend-assets/grpc/llama-cpp
+
+backend-assets/grpc/llama-cpp-avx2: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-avx2
+	$(MAKE) -C backend/cpp/llama-avx2 purge
+	$(info ${GREEN}I llama-cpp build info:avx2${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on" $(MAKE) VARIANT="llama-avx2" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-avx2/grpc-server backend-assets/grpc/llama-cpp-avx2
+
+backend-assets/grpc/llama-cpp-avx: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-avx
+	$(MAKE) -C backend/cpp/llama-avx purge
+	$(info ${GREEN}I llama-cpp build info:avx${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" $(MAKE) VARIANT="llama-avx" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-avx/grpc-server backend-assets/grpc/llama-cpp-avx
+
+backend-assets/grpc/llama-cpp-fallback: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-fallback
+	$(MAKE) -C backend/cpp/llama-fallback purge
+	$(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
+	$(MAKE) -C backend/cpp/llama-cuda purge
+	$(info ${GREEN}I llama-cpp build info:cuda${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_CUDA=ON" $(MAKE) VARIANT="llama-cuda" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-cuda/grpc-server backend-assets/grpc/llama-cpp-cuda
+
+backend-assets/grpc/llama-cpp-hipblas: backend-assets/grpc backend/cpp/llama/llama.cpp
+	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
+	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
+	cp -rf backend/cpp/llama backend/cpp/llama-sycl_f16
+	$(MAKE) -C backend/cpp/llama-sycl_f16 purge
+	$(info ${GREEN}I llama-cpp build info:sycl_f16${RESET})
+	BUILD_TYPE="sycl_f16" $(MAKE) VARIANT="llama-sycl_f16" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-sycl_f16/grpc-server backend-assets/grpc/llama-cpp-sycl_f16
+
+backend-assets/grpc/llama-cpp-sycl_f32: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-sycl_f32
+	$(MAKE) -C backend/cpp/llama-sycl_f32 purge
+	$(info ${GREEN}I llama-cpp build info:sycl_f32${RESET})
+	BUILD_TYPE="sycl_f32" $(MAKE) VARIANT="llama-sycl_f32" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-sycl_f32/grpc-server backend-assets/grpc/llama-cpp-sycl_f32
+
+backend-assets/grpc/llama-cpp-grpc: backend-assets/grpc backend/cpp/llama/llama.cpp
+	cp -rf backend/cpp/llama backend/cpp/llama-grpc
+	$(MAKE) -C backend/cpp/llama-grpc purge
+	$(info ${GREEN}I llama-cpp build info:grpc${RESET})
+	CMAKE_ARGS="$(CMAKE_ARGS) -DGGML_RPC=ON -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off" TARGET="--target grpc-server --target rpc-server" $(MAKE) VARIANT="llama-grpc" build-llama-cpp-grpc-server
+	cp -rfv backend/cpp/llama-grpc/grpc-server backend-assets/grpc/llama-cpp-grpc
+
+backend-assets/util/llama-cpp-rpc-server: backend-assets/grpc/llama-cpp-grpc
+	mkdir -p backend-assets/util/
+	cp -rf backend/cpp/llama-grpc/llama.cpp/build/bin/rpc-server backend-assets/util/llama-cpp-rpc-server
+
+backend-assets/grpc/llama-ggml: sources/go-llama.cpp sources/go-llama.cpp/libbinding.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" C_INCLUDE_PATH=$(CURDIR)/sources/go-llama.cpp LIBRARY_PATH=$(CURDIR)/sources/go-llama.cpp \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/llama-ggml ./backend/go/llm/llama-ggml/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/llama-ggml
+endif
+
+backend-assets/grpc/piper: sources/go-piper sources/go-piper/libpiper_binding.a backend-assets/grpc backend-assets/espeak-ng-data
+	CGO_CXXFLAGS="$(PIPER_CGO_CXXFLAGS)" CGO_LDFLAGS="$(PIPER_CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/sources/go-piper \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/piper ./backend/go/tts/
+ifneq ($(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 "$(GO_TAGS)" -o backend-assets/grpc/stablediffusion ./backend/go/image/stablediffusion
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/stablediffusion
+endif
+
+backend-assets/grpc/tinydream: sources/go-tiny-dream sources/go-tiny-dream/libtinydream.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" LIBRARY_PATH=$(CURDIR)/go-tiny-dream \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/tinydream ./backend/go/image/tinydream
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/tinydream
+endif
+
+backend-assets/grpc/whisper: sources/whisper.cpp sources/whisper.cpp/libwhisper.a backend-assets/grpc
+	CGO_LDFLAGS="$(CGO_LDFLAGS) $(CGO_LDFLAGS_WHISPER)" C_INCLUDE_PATH="$(CURDIR)/sources/whisper.cpp/include:$(CURDIR)/sources/whisper.cpp/ggml/include" LIBRARY_PATH=$(CURDIR)/sources/whisper.cpp \
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/whisper ./backend/go/transcribe/whisper
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/whisper
+endif
+
+backend-assets/grpc/local-store: backend-assets/grpc
+	$(GOCMD) build -ldflags "$(LD_FLAGS)" -tags "$(GO_TAGS)" -o backend-assets/grpc/local-store ./backend/go/stores/
+ifneq ($(UPX),)
+	$(UPX) backend-assets/grpc/local-store
+endif
+
+grpcs: prepare $(GRPC_BACKENDS)
+
+DOCKER_IMAGE?=local-ai
+DOCKER_AIO_IMAGE?=local-ai-aio
+IMAGE_TYPE?=core
+BASE_IMAGE?=ubuntu:22.04
+
+docker:
+	docker build \
+		--build-arg BASE_IMAGE=$(BASE_IMAGE) \
+		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
+		--build-arg GO_TAGS="$(GO_TAGS)" \
+		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
+		--build-arg BUILD_TYPE=$(BUILD_TYPE) \
+		-t $(DOCKER_IMAGE) .
+
+docker-cuda11:
+	docker build \
+		--build-arg CUDA_MAJOR_VERSION=11 \
+		--build-arg CUDA_MINOR_VERSION=8 \
+		--build-arg BASE_IMAGE=$(BASE_IMAGE) \
+		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
+		--build-arg GO_TAGS="$(GO_TAGS)" \
+		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
+		--build-arg BUILD_TYPE=$(BUILD_TYPE) \
+		-t $(DOCKER_IMAGE)-cuda11 .
+
+docker-aio:
+	@echo "Building AIO image with base $(BASE_IMAGE) as $(DOCKER_AIO_IMAGE)"
+	docker build \
+		--build-arg BASE_IMAGE=$(BASE_IMAGE) \
+		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
+		-t $(DOCKER_AIO_IMAGE) -f Dockerfile.aio .
+
+docker-aio-all:
+	$(MAKE) docker-aio DOCKER_AIO_SIZE=cpu
+	$(MAKE) docker-aio DOCKER_AIO_SIZE=cpu
+
+docker-image-intel:
+	docker build \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-devel-ubuntu22.04 \
+		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
+		--build-arg GO_TAGS="none" \
+		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
+		--build-arg BUILD_TYPE=sycl_f32 -t $(DOCKER_IMAGE) .
+
+docker-image-intel-xpu:
+	docker build \
+		--build-arg BASE_IMAGE=intel/oneapi-basekit:2024.2.0-devel-ubuntu22.04 \
+		--build-arg IMAGE_TYPE=$(IMAGE_TYPE) \
+		--build-arg GO_TAGS="none" \
+		--build-arg MAKEFLAGS="$(DOCKER_MAKEFLAGS)" \
+		--build-arg BUILD_TYPE=sycl_f32 -t $(DOCKER_IMAGE) .
+
+.PHONY: swagger
+swagger:
+	swag init -g core/http/app.go --output swagger
+
+.PHONY: gen-assets
+gen-assets:
+	$(GOCMD) run core/dependencies_manager/manager.go embedded/webui_static.yaml core/http/static/assets
+
+## Documentation
+docs/layouts/_default:
+	mkdir -p docs/layouts/_default
+
+docs/static/gallery.html: docs/layouts/_default
+	$(GOCMD) run ./.github/ci/modelslist.go ./gallery/index.yaml > docs/static/gallery.html
+
+docs/public: docs/layouts/_default docs/static/gallery.html
+	cd docs && hugo --minify
+
+docs-clean:
+	rm -rf docs/public
+	rm -rf docs/static/gallery.html
+
+.PHONY: docs
+docs: docs/static/gallery.html
+	cd docs && hugo serve

README.md

@@ -1,11 +1,241 @@
----
-title: Chat
-emoji: 🐠
-colorFrom: pink
-colorTo: purple
-sdk: docker
-pinned: false
-license: apache-2.0
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+<h1 align="center">
+  <br>
+  <img height="300" src="https://github.com/go-skynet/LocalAI/assets/2420543/0966aa2a-166e-4f99-a3e5-6c915fc997dd"> <br>
+    LocalAI
+<br>
+</h1>
+
+<p align="center">
+<a href="https://github.com/go-skynet/LocalAI/fork" target="blank">
+<img src="https://img.shields.io/github/forks/go-skynet/LocalAI?style=for-the-badge" alt="LocalAI forks"/>
+</a>
+<a href="https://github.com/go-skynet/LocalAI/stargazers" target="blank">
+<img src="https://img.shields.io/github/stars/go-skynet/LocalAI?style=for-the-badge" alt="LocalAI stars"/>
+</a>
+<a href="https://github.com/go-skynet/LocalAI/pulls" target="blank">
+<img src="https://img.shields.io/github/issues-pr/go-skynet/LocalAI?style=for-the-badge" alt="LocalAI pull-requests"/>
+</a>
+<a href='https://github.com/go-skynet/LocalAI/releases'>
+<img src='https://img.shields.io/github/release/go-skynet/LocalAI?&label=Latest&style=for-the-badge'>
+</a>
+</p>
+
+<p align="center">
+<a href="https://hub.docker.com/r/localai/localai" target="blank">
+<img src="https://img.shields.io/badge/dockerhub-images-important.svg?logo=Docker" alt="LocalAI Docker hub"/>
+</a>
+<a href="https://quay.io/repository/go-skynet/local-ai?tab=tags&tag=latest" target="blank">
+<img src="https://img.shields.io/badge/quay.io-images-important.svg?" alt="LocalAI Quay.io"/>
+</a>
+</p>
+
+<p align="center">
+<a href="https://twitter.com/LocalAI_API" target="blank">
+<img src="https://img.shields.io/twitter/follow/LocalAI_API?label=Follow: LocalAI_API&style=social" alt="Follow LocalAI_API"/>
+</a>
+<a href="https://discord.gg/uJAeKSAGDy" target="blank">
+<img src="https://dcbadge.vercel.app/api/server/uJAeKSAGDy?style=flat-square&theme=default-inverted" alt="Join LocalAI Discord Community"/>
+</a>
+</p>
+
+> :bulb: Get help - [❓FAQ](https://localai.io/faq/) [💭Discussions](https://github.com/go-skynet/LocalAI/discussions) [:speech_balloon: Discord](https://discord.gg/uJAeKSAGDy) [:book: Documentation website](https://localai.io/)
+>
+> [💻 Quickstart](https://localai.io/basics/getting_started/) [🖼️ Models](https://models.localai.io/) [🚀 Roadmap](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap) [🥽 Demo](https://demo.localai.io) [🌍 Explorer](https://explorer.localai.io) [🛫 Examples](https://github.com/mudler/LocalAI-examples) 
+
+[![tests](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/test.yml)[![Build and Release](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/release.yaml)[![build container images](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/image.yml)[![Bump dependencies](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml/badge.svg)](https://github.com/go-skynet/LocalAI/actions/workflows/bump_deps.yaml)[![Artifact Hub](https://img.shields.io/endpoint?url=https://artifacthub.io/badge/repository/localai)](https://artifacthub.io/packages/search?repo=localai)
+
+**LocalAI** is the free, Open Source OpenAI alternative. LocalAI act as a drop-in replacement REST API that’s compatible with OpenAI (Elevenlabs, Anthropic... ) API specifications for local AI inferencing. It allows you to run LLMs, generate images, audio (and not only) locally or on-prem with consumer grade hardware, supporting multiple model families. Does not require GPU. It is created and maintained by [Ettore Di Giacinto](https://github.com/mudler).
+
+![screen](https://github.com/mudler/LocalAI/assets/2420543/20b5ccd2-8393-44f0-aaf6-87a23806381e)
+
+Run the installer script:
+
+```bash
+curl https://localai.io/install.sh | sh
+```
+
+Or run with docker:
+```bash
+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:
+
+```bash
+# From the model gallery (see available models with `local-ai models list`, in the WebUI from the model tab, or visiting https://models.localai.io)
+local-ai run llama-3.2-1b-instruct:q4_k_m
+# Start LocalAI with the phi-2 model directly from huggingface
+local-ai run huggingface://TheBloke/phi-2-GGUF/phi-2.Q8_0.gguf
+# Install and run a model from the Ollama OCI registry
+local-ai run ollama://gemma:2b
+# Run a model from a configuration file
+local-ai run https://gist.githubusercontent.com/.../phi-2.yaml
+# Install and run a model from a standard OCI registry (e.g., Docker Hub)
+local-ai run oci://localai/phi-2:latest
+```
+
+[💻 Getting started](https://localai.io/basics/getting_started/index.html)
+
+## 📰 Latest project news
+
+- 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
+- June 2024: Support for models from OCI registries: https://github.com/mudler/LocalAI/pull/2628
+- May 2024: 🔥🔥 Decentralized P2P llama.cpp:  https://github.com/mudler/LocalAI/pull/2343 (peer2peer llama.cpp!) 👉 Docs  https://localai.io/features/distribute/
+- May 2024: 🔥🔥 Openvoice: https://github.com/mudler/LocalAI/pull/2334
+- May 2024: 🆕 Function calls without grammars and mixed mode: https://github.com/mudler/LocalAI/pull/2328
+- May 2024: 🔥🔥 Distributed inferencing: https://github.com/mudler/LocalAI/pull/2324
+- May 2024: Chat, TTS, and Image generation in the WebUI: https://github.com/mudler/LocalAI/pull/2222
+- April 2024: Reranker API: https://github.com/mudler/LocalAI/pull/2121
+
+Roadmap items: [List of issues](https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3Aroadmap)
+
+## 🔥🔥 Hot topics (looking for help):
+
+- Multimodal with vLLM and Video understanding: https://github.com/mudler/LocalAI/pull/3729
+- Realtime API https://github.com/mudler/LocalAI/issues/3714
+- 🔥🔥 Distributed, P2P Global community pools: https://github.com/mudler/LocalAI/issues/3113
+- WebUI improvements: https://github.com/mudler/LocalAI/issues/2156
+- Backends v2: https://github.com/mudler/LocalAI/issues/1126
+- Improving UX v2: https://github.com/mudler/LocalAI/issues/1373
+- Assistant API: https://github.com/mudler/LocalAI/issues/1273
+- Moderation endpoint: https://github.com/mudler/LocalAI/issues/999
+- Vulkan: https://github.com/mudler/LocalAI/issues/1647
+- Anthropic API: https://github.com/mudler/LocalAI/issues/1808
+
+If you want to help and contribute, issues up for grabs: https://github.com/mudler/LocalAI/issues?q=is%3Aissue+is%3Aopen+label%3A%22up+for+grabs%22
+
+## 🚀 [Features](https://localai.io/features/)
+
+- 📖 [Text generation with GPTs](https://localai.io/features/text-generation/) (`llama.cpp`, `gpt4all.cpp`, ... [:book: and more](https://localai.io/model-compatibility/index.html#model-compatibility-table))
+- 🗣 [Text to Audio](https://localai.io/features/text-to-audio/)
+- 🔈 [Audio to Text](https://localai.io/features/audio-to-text/) (Audio transcription with `whisper.cpp`)
+- 🎨 [Image generation with stable diffusion](https://localai.io/features/image-generation)
+- 🔥 [OpenAI-alike tools API](https://localai.io/features/openai-functions/) 
+- 🧠 [Embeddings generation for vector databases](https://localai.io/features/embeddings/)
+- ✍️ [Constrained grammars](https://localai.io/features/constrained_grammars/)
+- 🖼️ [Download Models directly from Huggingface ](https://localai.io/models/)
+- 🥽 [Vision API](https://localai.io/features/gpt-vision/)
+- 📈 [Reranker API](https://localai.io/features/reranker/)
+- 🆕🖧 [P2P Inferencing](https://localai.io/features/distribute/)
+- 🌍 Integrated WebUI!
+
+## 💻 Usage
+
+Check out the [Getting started](https://localai.io/basics/getting_started/index.html) section in our documentation.
+
+### 🔗 Community and integrations
+
+Build and deploy custom containers:
+- https://github.com/sozercan/aikit
+
+WebUIs:
+- https://github.com/Jirubizu/localai-admin
+- https://github.com/go-skynet/LocalAI-frontend
+- QA-Pilot(An interactive chat project that leverages LocalAI LLMs for rapid understanding and navigation of GitHub code repository) https://github.com/reid41/QA-Pilot
+
+Model galleries
+- https://github.com/go-skynet/model-gallery
+
+Other:
+- Helm chart https://github.com/go-skynet/helm-charts
+- VSCode extension https://github.com/badgooooor/localai-vscode-plugin
+- Terminal utility https://github.com/djcopley/ShellOracle
+- Local Smart assistant https://github.com/mudler/LocalAGI
+- Home Assistant https://github.com/sammcj/homeassistant-localai / https://github.com/drndos/hass-openai-custom-conversation / https://github.com/valentinfrlch/ha-gpt4vision
+- Discord bot https://github.com/mudler/LocalAGI/tree/main/examples/discord
+- 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
+- Github Actions: https://github.com/marketplace/actions/start-localai
+- Examples: https://github.com/mudler/LocalAI/tree/master/examples/
+  
+
+### 🔗 Resources
+
+- [LLM finetuning guide](https://localai.io/docs/advanced/fine-tuning/)
+- [How to build locally](https://localai.io/basics/build/index.html)
+- [How to install in Kubernetes](https://localai.io/basics/getting_started/index.html#run-localai-in-kubernetes)
+- [Projects integrating LocalAI](https://localai.io/docs/integrations/)
+- [How tos section](https://io.midori-ai.xyz/howtos/) (curated by our community)
+
+## :book: 🎥 [Media, Blogs, Social](https://localai.io/basics/news/#media-blogs-social)
+
+- [Run Visual studio code with LocalAI (SUSE)](https://www.suse.com/c/running-ai-locally/)
+- 🆕 [Run LocalAI on Jetson Nano Devkit](https://mudler.pm/posts/local-ai-jetson-nano-devkit/)
+- [Run LocalAI on AWS EKS with Pulumi](https://www.pulumi.com/blog/low-code-llm-apps-with-local-ai-flowise-and-pulumi/)
+- [Run LocalAI on AWS](https://staleks.hashnode.dev/installing-localai-on-aws-ec2-instance)
+- [Create a slackbot for teams and OSS projects that answer to documentation](https://mudler.pm/posts/smart-slackbot-for-teams/)
+- [LocalAI meets k8sgpt](https://www.youtube.com/watch?v=PKrDNuJ_dfE)
+- [Question Answering on Documents locally with LangChain, LocalAI, Chroma, and GPT4All](https://mudler.pm/posts/localai-question-answering/)
+- [Tutorial to use k8sgpt with LocalAI](https://medium.com/@tyler_97636/k8sgpt-localai-unlock-kubernetes-superpowers-for-free-584790de9b65)
+
+## Citation
+
+If you utilize this repository, data in a downstream project, please consider citing it with:
+
+```
+@misc{localai,
+  author = {Ettore Di Giacinto},
+  title = {LocalAI: The free, Open source OpenAI alternative},
+  year = {2023},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/go-skynet/LocalAI}},
+```
+
+## ❤️ Sponsors
+
+> Do you find LocalAI useful?
+
+Support the project by becoming [a backer or sponsor](https://github.com/sponsors/mudler). Your logo will show up here with a link to your website.
+
+A huge thank you to our generous sponsors who support this project covering CI expenses, and our [Sponsor list](https://github.com/sponsors/mudler):
+
+<p align="center">
+  <a href="https://www.spectrocloud.com/" target="blank">
+    <img height="200" src="https://github.com/go-skynet/LocalAI/assets/2420543/68a6f3cb-8a65-4a4d-99b5-6417a8905512">
+  </a>
+  <a href="https://www.premai.io/" target="blank">
+    <img height="200" src="https://github.com/mudler/LocalAI/assets/2420543/42e4ca83-661e-4f79-8e46-ae43689683d6"> <br>
+  </a>
+</p>
+
+## 🌟 Star history
+
+[![LocalAI Star history Chart](https://api.star-history.com/svg?repos=go-skynet/LocalAI&type=Date)](https://star-history.com/#go-skynet/LocalAI&Date)
+
+## 📖 License
+
+LocalAI is a community-driven project created by [Ettore Di Giacinto](https://github.com/mudler/).
+
+MIT - Author Ettore Di Giacinto <mudler@localai.io>
+
+## 🙇 Acknowledgements
+
+LocalAI couldn't have been built without the help of great software already available from the community. Thank you!
+
+- [llama.cpp](https://github.com/ggerganov/llama.cpp)
+- https://github.com/tatsu-lab/stanford_alpaca
+- https://github.com/cornelk/llama-go for the initial ideas
+- 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
+
+This is a community project, a special thanks to our contributors! 🤗
+<a href="https://github.com/go-skynet/LocalAI/graphs/contributors">
+  <img src="https://contrib.rocks/image?repo=go-skynet/LocalAI" />
+</a>

SECURITY.md

@@ -0,0 +1,42 @@
+# Security Policy
+
+## Introduction
+
+At LocalAI, we take the security of our software seriously. We understand the importance of protecting our community from vulnerabilities and are committed to ensuring the safety and security of our users.
+
+## Supported Versions
+
+We provide support and updates for certain versions of our software. The following table outlines which versions are currently supported with security updates:
+
+| Version | Supported          |
+| ------- | ------------------ |
+| > 2.0   | :white_check_mark: |
+| < 2.0   | :x:                |
+
+Please ensure that you are using a supported version to receive the latest security updates.
+
+## Reporting a Vulnerability
+
+We encourage the responsible disclosure of any security vulnerabilities. If you believe you've found a security issue in our software, we kindly ask you to follow the steps below to report it to us:
+
+1. **Email Us:** Send an email to [security@localai.io](mailto:security@localai.io) with a detailed report. Please do not disclose the vulnerability publicly or to any third parties before it has been addressed by us.
+
+2. **Expect a Response:** We aim to acknowledge receipt of vulnerability reports within 48 hours. Our security team will review your report and work closely with you to understand the impact and ensure a thorough investigation.
+
+3. **Collaboration:** If the vulnerability is accepted, we will work with you and our community to address the issue promptly. We'll keep you informed throughout the resolution process and may request additional information or collaboration.
+
+4. **Disclosure:** Once the vulnerability has been resolved, we encourage a coordinated disclosure. We believe in transparency and will work with you to ensure that our community is informed in a responsible manner.
+
+## Use of Third-Party Platforms
+
+As a Free and Open Source Software (FOSS) organization, we do not offer monetary bounties. However, researchers who wish to report vulnerabilities can also do so via [Huntr](https://huntr.dev/bounties), a platform that recognizes contributions to open source security.
+
+## Contact
+
+For any security-related inquiries beyond vulnerability reporting, please contact us at [security@localai.io](mailto:security@localai.io).
+
+## Acknowledgments
+
+We appreciate the efforts of those who contribute to the security of our project. Your responsible disclosure is invaluable to the safety and integrity of LocalAI.
+
+Thank you for helping us keep LocalAI secure.

aio/cpu/README.md

@@ -0,0 +1,5 @@
+## AIO CPU size
+
+Use this image with CPU-only.
+
+Please keep using only C++ backends so the base image is as small as possible (without CUDA, cuDNN, python, etc).

aio/cpu/embeddings.yaml

@@ -0,0 +1,12 @@
+name: text-embedding-ada-002
+backend: bert-embeddings
+parameters:
+  model: huggingface://mudler/all-MiniLM-L6-v2/ggml-model-q4_0.bin
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/embeddings -X POST -H "Content-Type: application/json" -d '{
+      "input": "Your text string goes here",
+      "model": "text-embedding-ada-002"
+    }'

aio/cpu/image-gen.yaml

@@ -0,0 +1,62 @@
+name: stablediffusion
+backend: stablediffusion
+parameters:
+  model: stablediffusion_assets
+
+license: "BSD-3"
+urls:
+- https://github.com/EdVince/Stable-Diffusion-NCNN
+- https://github.com/EdVince/Stable-Diffusion-NCNN/blob/main/LICENSE
+
+description: |
+     Stable Diffusion in NCNN with c++, supported txt2img and img2img
+
+download_files:
+- filename: "stablediffusion_assets/AutoencoderKL-256-256-fp16-opt.param"
+  sha256: "18ca4b66685e21406bcf64c484b3b680b4949900415536d599cc876579c85c82"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-256-256-fp16-opt.param"
+- filename: "stablediffusion_assets/AutoencoderKL-512-512-fp16-opt.param"
+  sha256: "cf45f63aacf3dbbab0f59ed92a6f2c14d9a1801314631cd3abe91e3c85639a20"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-512-512-fp16-opt.param"
+- filename: "stablediffusion_assets/AutoencoderKL-base-fp16.param"
+  sha256: "0254a056dce61b0c27dc9ec1b78b53bcf55315c540f55f051eb841aa992701ba"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/AutoencoderKL-base-fp16.param"
+- filename: "stablediffusion_assets/AutoencoderKL-encoder-512-512-fp16.bin"
+  sha256: "ddcb79a9951b9f91e05e087739ed69da2c1c4ae30ba4168cce350b49d617c9fa"
+  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/AutoencoderKL-encoder-512-512-fp16.bin"
+- filename: "stablediffusion_assets/AutoencoderKL-fp16.bin"
+  sha256: "f02e71f80e70252734724bbfaed5c4ddd3a8ed7e61bb2175ff5f53099f0e35dd"
+  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/AutoencoderKL-fp16.bin"
+- filename: "stablediffusion_assets/FrozenCLIPEmbedder-fp16.bin"
+  sha256: "1c9a12f4e1dd1b295a388045f7f28a2352a4d70c3dc96a542189a3dd7051fdd6"
+  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/FrozenCLIPEmbedder-fp16.bin"
+- filename: "stablediffusion_assets/FrozenCLIPEmbedder-fp16.param"
+  sha256: "471afbe678dd1fd3fe764ef9c6eccaccb0a7d7e601f27b462aa926b20eb368c9"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/FrozenCLIPEmbedder-fp16.param"
+- filename: "stablediffusion_assets/log_sigmas.bin"
+  sha256: "a2089f8aa4c61f9c200feaec541ab3f5c94233b28deb6d5e8bcd974fa79b68ac"
+  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/raw/main/x86/linux/assets/log_sigmas.bin"
+- filename: "stablediffusion_assets/UNetModel-256-256-MHA-fp16-opt.param"
+  sha256: "a58c380229f09491776df837b7aa7adffc0a87821dc4708b34535da2e36e3da1"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-256-256-MHA-fp16-opt.param"
+- filename: "stablediffusion_assets/UNetModel-512-512-MHA-fp16-opt.param"
+  sha256: "f12034067062827bd7f43d1d21888d1f03905401acf6c6eea22be23c259636fa"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-512-512-MHA-fp16-opt.param"
+- filename: "stablediffusion_assets/UNetModel-base-MHA-fp16.param"
+  sha256: "696f6975de49f4325b53ce32aff81861a6d6c07cd9ce3f0aae2cc405350af38d"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/UNetModel-base-MHA-fp16.param"
+- filename: "stablediffusion_assets/UNetModel-MHA-fp16.bin"
+  sha256: "d618918d011bfc1f644c0f2a33bf84931bd53b28a98492b0a8ed6f3a818852c3"
+  uri: "https://github.com/EdVince/Stable-Diffusion-NCNN/releases/download/naifu/UNetModel-MHA-fp16.bin"
+- filename: "stablediffusion_assets/vocab.txt"
+  sha256: "e30e57b6f1e47616982ef898d8922be24e535b4fa3d0110477b3a6f02ebbae7d"
+  uri: "https://raw.githubusercontent.com/EdVince/Stable-Diffusion-NCNN/main/x86/linux/assets/vocab.txt"
+
+usage: |
+        curl http://localhost:8080/v1/images/generations \
+          -H "Content-Type: application/json" \
+          -d '{
+            "prompt": "<positive prompt>|<negative prompt>",
+            "step": 25,
+            "size": "512x512"
+          }'

aio/cpu/rerank.yaml

@@ -0,0 +1,27 @@
+name: jina-reranker-v1-base-en
+backend: rerankers
+parameters:
+  model: cross-encoder
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/v1/rerank \
+      -H "Content-Type: application/json" \
+      -d '{
+      "model": "jina-reranker-v1-base-en",
+      "query": "Organic skincare products for sensitive skin",
+      "documents": [
+        "Eco-friendly kitchenware for modern homes",
+        "Biodegradable cleaning supplies for eco-conscious consumers",
+        "Organic cotton baby clothes for sensitive skin",
+        "Natural organic skincare range for sensitive skin",
+        "Tech gadgets for smart homes: 2024 edition",
+        "Sustainable gardening tools and compost solutions",
+        "Sensitive skin-friendly facial cleansers and toners",
+        "Organic food wraps and storage solutions",
+        "All-natural pet food for dogs with allergies",
+        "Yoga mats made from recycled materials"
+      ],
+      "top_n": 3
+    }'

aio/cpu/speech-to-text.yaml

@@ -0,0 +1,18 @@
+name: whisper-1
+backend: whisper
+parameters:
+  model: ggml-whisper-base.bin
+
+usage: |
+    ## example audio file
+    wget --quiet --show-progress -O gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+
+    ## Send the example audio file to the transcriptions endpoint
+    curl http://localhost:8080/v1/audio/transcriptions \
+         -H "Content-Type: multipart/form-data" \
+         -F file="@$PWD/gb1.ogg" -F model="whisper-1"
+
+download_files:
+- filename: "ggml-whisper-base.bin"
+  sha256: "60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe"
+  uri: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"

aio/cpu/text-to-speech.yaml

@@ -0,0 +1,15 @@
+name: tts-1
+download_files:
+  - filename: voice-en-us-amy-low.tar.gz
+    uri: https://github.com/rhasspy/piper/releases/download/v0.0.2/voice-en-us-amy-low.tar.gz
+
+parameters:
+  model: en-us-amy-low.onnx
+
+usage: |
+    To test if this model works as expected, you can use the following curl command:
+
+    curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{
+      "model":"voice-en-us-amy-low",
+      "input": "Hi, this is a test."
+    }'

aio/cpu/text-to-text.yaml

@@ -0,0 +1,101 @@
+name: gpt-4
+mmap: true
+parameters:
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+context_size: 8192
+
+stopwords:
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
+template:
+  chat: |
+    {{.Input -}}
+    <|im_start|>assistant
+  chat_message: |
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
+    {{.Content }}
+    {{- end }}
+    {{- if .FunctionCall}}
+    {{toJson .FunctionCall}}
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
+  completion: |
+    {{.Input}}
+  function: |-
+    <|im_start|>system
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
+    {{range .Functions}}
+    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
+    {{end}}
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
+    {{.Input -}}
+    <|im_start|>assistant

aio/cpu/vision.yaml

@@ -0,0 +1,31 @@
+backend: llama-cpp
+context_size: 4096
+f16: true
+mmap: true
+name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: bakllava-mmproj.gguf
+parameters:
+  model: bakllava.gguf
+
+template:
+  chat: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
+    {{.Input}}
+    ASSISTANT:
+
+download_files:
+- filename: bakllava.gguf
+  uri: huggingface://mys/ggml_bakllava-1/ggml-model-q4_k.gguf
+- filename: bakllava-mmproj.gguf
+  uri: huggingface://mys/ggml_bakllava-1/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

aio/entrypoint.sh

@@ -0,0 +1,138 @@
+#!/bin/bash
+
+echo "===> LocalAI All-in-One (AIO) container starting..."
+
+GPU_ACCELERATION=false
+GPU_VENDOR=""
+
+function check_intel() {
+    if lspci | grep -E 'VGA|3D' | grep -iq intel; then
+        echo "Intel GPU detected"
+        if [ -d /opt/intel ]; then
+            GPU_ACCELERATION=true
+            GPU_VENDOR=intel
+        else
+            echo "Intel GPU detected, but Intel GPU drivers are not installed. GPU acceleration will not be available."
+        fi
+    fi
+}
+
+function check_nvidia_wsl() {
+    if lspci | grep -E 'VGA|3D' | grep -iq "Microsoft Corporation Device 008e"; then
+        # We make the assumption this WSL2 cars is NVIDIA, then check for nvidia-smi
+        # Make sure the container was run with `--gpus all` as the only required parameter
+        echo "NVIDIA GPU detected via WSL2"
+        # nvidia-smi should be installed in the container
+        if nvidia-smi; then
+            GPU_ACCELERATION=true
+            GPU_VENDOR=nvidia
+        else
+            echo "NVIDIA GPU detected via WSL2, but nvidia-smi is not installed. GPU acceleration will not be available."
+        fi
+    fi
+}
+
+function check_amd() {
+    if lspci | grep -E 'VGA|3D' | grep -iq amd; then
+        echo "AMD GPU detected"
+        # Check if ROCm is installed
+        if [ -d /opt/rocm ]; then
+            GPU_ACCELERATION=true
+            GPU_VENDOR=amd
+        else
+            echo "AMD GPU detected, but ROCm is not installed. GPU acceleration will not be available."
+        fi
+    fi
+}
+
+function check_nvidia() {
+    if lspci | grep -E 'VGA|3D' | grep -iq nvidia; then
+        echo "NVIDIA GPU detected"
+        # nvidia-smi should be installed in the container
+        if nvidia-smi; then
+            GPU_ACCELERATION=true
+            GPU_VENDOR=nvidia
+        else
+            echo "NVIDIA GPU detected, but nvidia-smi is not installed. GPU acceleration will not be available."
+        fi
+    fi
+}
+
+function check_metal() {
+    if system_profiler SPDisplaysDataType | grep -iq 'Metal'; then
+        echo "Apple Metal supported GPU detected"
+        GPU_ACCELERATION=true
+        GPU_VENDOR=apple
+    fi
+}
+
+function detect_gpu() {
+    case "$(uname -s)" in
+        Linux)
+            check_nvidia
+            check_amd
+            check_intel
+            check_nvidia_wsl
+            ;;
+        Darwin)
+            check_metal
+            ;;
+    esac
+}
+
+function detect_gpu_size() {
+    # Attempting to find GPU memory size for NVIDIA GPUs
+    if [ "$GPU_ACCELERATION" = true ] && [ "$GPU_VENDOR" = "nvidia" ]; then
+        echo "NVIDIA GPU detected. Attempting to find memory size..."
+        # Using head -n 1 to get the total memory of the 1st NVIDIA GPU detected.
+        # If handling multiple GPUs is required in the future, this is the place to do it
+        nvidia_sm=$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits | head -n 1)
+        if [ ! -z "$nvidia_sm" ]; then
+            echo "Total GPU Memory: $nvidia_sm MiB"
+            # if bigger than 8GB, use 16GB
+            #if [ "$nvidia_sm" -gt 8192 ]; then
+            #    GPU_SIZE=gpu-16g
+            #else
+            GPU_SIZE=gpu-8g
+            #fi
+        else
+            echo "Unable to determine NVIDIA GPU memory size. Falling back to CPU."
+            GPU_SIZE=gpu-8g
+        fi
+    elif [ "$GPU_ACCELERATION" = true ] && [ "$GPU_VENDOR" = "intel" ]; then
+        GPU_SIZE=intel
+    # Default to a generic GPU size until we implement GPU size detection for non NVIDIA GPUs
+    elif [ "$GPU_ACCELERATION" = true ]; then
+        echo "Non-NVIDIA GPU detected. Specific GPU memory size detection is not implemented."
+        GPU_SIZE=gpu-8g
+
+    # default to cpu if GPU_SIZE is not set
+    else
+        echo "GPU acceleration is not enabled or supported. Defaulting to CPU."
+        GPU_SIZE=cpu
+    fi
+}
+
+function check_vars() {
+    if [ -z "$MODELS" ]; then
+        echo "MODELS environment variable is not set. Please set it to a comma-separated list of model YAML files to load."
+        exit 1
+    fi
+
+    if [ -z "$PROFILE" ]; then
+        echo "PROFILE environment variable is not set. Please set it to one of the following: cpu, gpu-8g, gpu-16g, apple"
+        exit 1
+    fi
+}
+
+detect_gpu
+detect_gpu_size
+
+PROFILE="${PROFILE:-$GPU_SIZE}" # default to cpu
+export MODELS="${MODELS:-/aio/${PROFILE}/embeddings.yaml,/aio/${PROFILE}/rerank.yaml,/aio/${PROFILE}/text-to-speech.yaml,/aio/${PROFILE}/image-gen.yaml,/aio/${PROFILE}/text-to-text.yaml,/aio/${PROFILE}/speech-to-text.yaml,/aio/${PROFILE}/vision.yaml}"
+
+check_vars
+
+echo "===> Starting LocalAI[$PROFILE] with the following models: $MODELS"
+
+exec /build/entrypoint.sh "$@"

aio/gpu-8g/embeddings.yaml

@@ -0,0 +1,12 @@
+name: text-embedding-ada-002
+backend: sentencetransformers
+parameters:
+  model: all-MiniLM-L6-v2
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/embeddings -X POST -H "Content-Type: application/json" -d '{
+      "input": "Your text string goes here",
+      "model": "text-embedding-ada-002"
+    }'

aio/gpu-8g/image-gen.yaml

@@ -0,0 +1,25 @@
+name: stablediffusion
+parameters:
+  model: DreamShaper_8_pruned.safetensors
+backend: diffusers
+step: 25
+f16: true
+
+diffusers:
+  pipeline_type: StableDiffusionPipeline
+  cuda: true
+  enable_parameters: "negative_prompt,num_inference_steps"
+  scheduler_type: "k_dpmpp_2m"
+
+download_files:
+- filename: DreamShaper_8_pruned.safetensors
+  uri: huggingface://Lykon/DreamShaper/DreamShaper_8_pruned.safetensors
+
+usage: |
+        curl http://localhost:8080/v1/images/generations \
+          -H "Content-Type: application/json" \
+          -d '{
+            "prompt": "<positive prompt>|<negative prompt>",
+            "step": 25,
+            "size": "512x512"
+          }'

aio/gpu-8g/rerank.yaml

@@ -0,0 +1,27 @@
+name: jina-reranker-v1-base-en
+backend: rerankers
+parameters:
+  model: cross-encoder
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/v1/rerank \
+      -H "Content-Type: application/json" \
+      -d '{
+      "model": "jina-reranker-v1-base-en",
+      "query": "Organic skincare products for sensitive skin",
+      "documents": [
+        "Eco-friendly kitchenware for modern homes",
+        "Biodegradable cleaning supplies for eco-conscious consumers",
+        "Organic cotton baby clothes for sensitive skin",
+        "Natural organic skincare range for sensitive skin",
+        "Tech gadgets for smart homes: 2024 edition",
+        "Sustainable gardening tools and compost solutions",
+        "Sensitive skin-friendly facial cleansers and toners",
+        "Organic food wraps and storage solutions",
+        "All-natural pet food for dogs with allergies",
+        "Yoga mats made from recycled materials"
+      ],
+      "top_n": 3
+    }'

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

@@ -0,0 +1,18 @@
+name: whisper-1
+backend: whisper
+parameters:
+  model: ggml-whisper-base.bin
+
+usage: |
+    ## example audio file
+    wget --quiet --show-progress -O gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+
+    ## Send the example audio file to the transcriptions endpoint
+    curl http://localhost:8080/v1/audio/transcriptions \
+         -H "Content-Type: multipart/form-data" \
+         -F file="@$PWD/gb1.ogg" -F model="whisper-1"
+
+download_files:
+- filename: "ggml-whisper-base.bin"
+  sha256: "60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe"
+  uri: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"

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

@@ -0,0 +1,15 @@
+name: tts-1
+download_files:
+  - filename: voice-en-us-amy-low.tar.gz
+    uri: https://github.com/rhasspy/piper/releases/download/v0.0.2/voice-en-us-amy-low.tar.gz
+
+parameters:
+  model: en-us-amy-low.onnx
+
+usage: |
+    To test if this model works as expected, you can use the following curl command:
+
+    curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{
+      "model":"tts-1",
+      "input": "Hi, this is a test."
+    }'

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

@@ -0,0 +1,101 @@
+name: gpt-4
+mmap: true
+parameters:
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+context_size: 8192
+
+stopwords:
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
+template:
+  chat: |
+    {{.Input -}}
+    <|im_start|>assistant
+  chat_message: |
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
+    {{.Content }}
+    {{- end }}
+    {{- if .FunctionCall}}
+    {{toJson .FunctionCall}}
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
+  completion: |
+    {{.Input}}
+  function: |-
+    <|im_start|>system
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
+    {{range .Functions}}
+    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
+    {{end}}
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
+    {{.Input -}}
+    <|im_start|>assistant

aio/gpu-8g/vision.yaml

@@ -0,0 +1,35 @@
+backend: llama-cpp
+context_size: 4096
+f16: true
+mmap: true
+name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: llava-v1.6-7b-mmproj-f16.gguf
+parameters:
+  model: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  temperature: 0.2
+  top_k: 40
+  top_p: 0.95
+  seed: -1
+
+template:
+  chat: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
+    {{.Input}}
+    ASSISTANT:
+
+download_files:
+- filename: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/llava-v1.6-mistral-7b.Q5_K_M.gguf
+- filename: llava-v1.6-7b-mmproj-f16.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

aio/intel/embeddings.yaml

@@ -0,0 +1,12 @@
+name: text-embedding-ada-002
+backend: sentencetransformers
+parameters:
+  model: all-MiniLM-L6-v2
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/embeddings -X POST -H "Content-Type: application/json" -d '{
+      "input": "Your text string goes here",
+      "model": "text-embedding-ada-002"
+    }'

aio/intel/image-gen.yaml

@@ -0,0 +1,20 @@
+name: stablediffusion
+parameters:
+  model: Lykon/dreamshaper-8
+backend: diffusers
+step: 25
+f16: true
+diffusers:
+  pipeline_type: StableDiffusionPipeline
+  cuda: true
+  enable_parameters: "negative_prompt,num_inference_steps"
+  scheduler_type: "k_dpmpp_2m"
+
+usage: |
+        curl http://localhost:8080/v1/images/generations \
+          -H "Content-Type: application/json" \
+          -d '{
+            "prompt": "<positive prompt>|<negative prompt>",
+            "step": 25,
+            "size": "512x512"
+          }'

aio/intel/rerank.yaml

@@ -0,0 +1,27 @@
+name: jina-reranker-v1-base-en
+backend: rerankers
+parameters:
+  model: cross-encoder
+
+usage: |
+    You can test this model with curl like this:
+
+    curl http://localhost:8080/v1/rerank \
+      -H "Content-Type: application/json" \
+      -d '{
+      "model": "jina-reranker-v1-base-en",
+      "query": "Organic skincare products for sensitive skin",
+      "documents": [
+        "Eco-friendly kitchenware for modern homes",
+        "Biodegradable cleaning supplies for eco-conscious consumers",
+        "Organic cotton baby clothes for sensitive skin",
+        "Natural organic skincare range for sensitive skin",
+        "Tech gadgets for smart homes: 2024 edition",
+        "Sustainable gardening tools and compost solutions",
+        "Sensitive skin-friendly facial cleansers and toners",
+        "Organic food wraps and storage solutions",
+        "All-natural pet food for dogs with allergies",
+        "Yoga mats made from recycled materials"
+      ],
+      "top_n": 3
+    }'

aio/intel/speech-to-text.yaml

@@ -0,0 +1,18 @@
+name: whisper-1
+backend: whisper
+parameters:
+  model: ggml-whisper-base.bin
+
+usage: |
+    ## example audio file
+    wget --quiet --show-progress -O gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+
+    ## Send the example audio file to the transcriptions endpoint
+    curl http://localhost:8080/v1/audio/transcriptions \
+         -H "Content-Type: multipart/form-data" \
+         -F file="@$PWD/gb1.ogg" -F model="whisper-1"
+
+download_files:
+- filename: "ggml-whisper-base.bin"
+  sha256: "60ed5bc3dd14eea856493d334349b405782ddcaf0028d4b5df4088345fba2efe"
+  uri: "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"

aio/intel/text-to-speech.yaml

@@ -0,0 +1,15 @@
+name: tts-1
+download_files:
+  - filename: voice-en-us-amy-low.tar.gz
+    uri: https://github.com/rhasspy/piper/releases/download/v0.0.2/voice-en-us-amy-low.tar.gz
+
+parameters:
+  model: en-us-amy-low.onnx
+
+usage: |
+    To test if this model works as expected, you can use the following curl command:
+
+    curl http://localhost:8080/tts -H "Content-Type: application/json" -d '{
+      "model":"tts-1",
+      "input": "Hi, this is a test."
+    }'

aio/intel/text-to-text.yaml

@@ -0,0 +1,103 @@
+name: gpt-4
+mmap: false
+context_size: 8192
+
+f16: false
+parameters:
+  model: huggingface://NousResearch/Hermes-2-Pro-Llama-3-8B-GGUF/Hermes-2-Pro-Llama-3-8B-Q4_K_M.gguf
+
+stopwords:
+- "<|im_end|>"
+- "<dummy32000>"
+- "</tool_call>"
+- "<|eot_id|>"
+- "<|end_of_text|>"
+
+function:
+  # disable injecting the "answer" tool
+  disable_no_action: true
+
+  grammar:
+    # This allows the grammar to also return messages
+    mixed_mode: true
+    # Suffix to add to the grammar
+    #prefix: '<tool_call>\n'
+    # Force parallel calls in the grammar
+    # parallel_calls: true
+
+  return_name_in_function_response: true
+  # Without grammar uncomment the lines below
+  # Warning: this is relying only on the capability of the
+  # LLM model to generate the correct function call.
+  json_regex_match: 
+   - "(?s)<tool_call>(.*?)</tool_call>"
+   - "(?s)<tool_call>(.*?)"
+  replace_llm_results:
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+  replace_function_results: 
+  # Replace everything that is not JSON array or object
+  # 
+  - key: '(?s)^[^{\[]*'
+    value: ""
+  - key: '(?s)[^}\]]*$'
+    value: ""
+  - key: "'([^']*?)'"
+    value: "_DQUOTE_${1}_DQUOTE_"
+  - key: '\\"'
+    value: "__TEMP_QUOTE__"
+  - key: "\'"
+    value: "'"
+  - key: "_DQUOTE_"
+    value: '"'
+  - key: "__TEMP_QUOTE__"
+    value: '"'
+  # Drop the scratchpad content from responses
+  - key: "(?s)<scratchpad>.*</scratchpad>"
+    value: ""
+
+template:
+  chat: |
+    {{.Input -}}
+    <|im_start|>assistant
+  chat_message: |
+    <|im_start|>{{if eq .RoleName "assistant"}}assistant{{else if eq .RoleName "system"}}system{{else if eq .RoleName "tool"}}tool{{else if eq .RoleName "user"}}user{{end}}
+    {{- if .FunctionCall }}
+    <tool_call>
+    {{- else if eq .RoleName "tool" }}
+    <tool_response>
+    {{- end }}
+    {{- if .Content}}
+    {{.Content }}
+    {{- end }}
+    {{- if .FunctionCall}}
+    {{toJson .FunctionCall}}
+    {{- end }}
+    {{- if .FunctionCall }}
+    </tool_call>
+    {{- else if eq .RoleName "tool" }}
+    </tool_response>
+    {{- end }}<|im_end|>
+  completion: |
+    {{.Input}}
+  function: |-
+    <|im_start|>system
+    You are a function calling AI model.
+    Here are the available tools:
+    <tools>
+    {{range .Functions}}
+    {'type': 'function', 'function': {'name': '{{.Name}}', 'description': '{{.Description}}', 'parameters': {{toJson .Parameters}} }}
+    {{end}}
+    </tools>
+    You should call the tools provided to you sequentially
+    Please use <scratchpad> XML tags to record your reasoning and planning before you call the functions as follows:
+    <scratchpad>
+    {step-by-step reasoning and plan in bullet points}
+    </scratchpad>
+    For each function call return a json object with function name and arguments within <tool_call> XML tags as follows:
+    <tool_call>
+    {"arguments": <args-dict>, "name": <function-name>}
+    </tool_call><|im_end|>
+    {{.Input -}}
+    <|im_start|>assistant

aio/intel/vision.yaml

@@ -0,0 +1,35 @@
+backend: llama-cpp
+context_size: 4096
+mmap: false
+f16: false
+name: gpt-4o
+
+roles:
+  user: "USER:"
+  assistant: "ASSISTANT:"
+  system: "SYSTEM:"
+
+mmproj: llava-v1.6-7b-mmproj-f16.gguf
+parameters:
+  model: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  temperature: 0.2
+  top_k: 40
+  top_p: 0.95
+  seed: -1
+
+template:
+  chat: |
+    A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.
+    {{.Input}}
+    ASSISTANT:
+
+download_files:
+- filename: llava-v1.6-mistral-7b.Q5_K_M.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/llava-v1.6-mistral-7b.Q5_K_M.gguf
+- filename: llava-v1.6-7b-mmproj-f16.gguf
+  uri: huggingface://cjpais/llava-1.6-mistral-7b-gguf/mmproj-model-f16.gguf
+
+usage: |
+    curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
+        "model": "gpt-4-vision-preview",
+        "messages": [{"role": "user", "content": [{"type":"text", "text": "What is in the image?"}, {"type": "image_url", "image_url": {"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg" }}], "temperature": 0.9}]}'

assets.go

@@ -0,0 +1,6 @@
+package main
+
+import "embed"
+
+//go:embed backend-assets/*
+var backendAssets embed.FS

backend/backend.proto

@@ -0,0 +1,331 @@
+syntax = "proto3";
+
+option go_package = "github.com/go-skynet/LocalAI/pkg/grpc/proto";
+option java_multiple_files = true;
+option java_package = "io.skynet.localai.backend";
+option java_outer_classname = "LocalAIBackend";
+
+package backend;
+
+service Backend {
+  rpc Health(HealthMessage) returns (Reply) {}
+  rpc Predict(PredictOptions) returns (Reply) {}
+  rpc LoadModel(ModelOptions) returns (Result) {}
+  rpc PredictStream(PredictOptions) returns (stream Reply) {}
+  rpc Embedding(PredictOptions) returns (EmbeddingResult) {}
+  rpc GenerateImage(GenerateImageRequest) returns (Result) {}
+  rpc AudioTranscription(TranscriptRequest) returns (TranscriptResult) {}
+  rpc TTS(TTSRequest) returns (Result) {}
+  rpc SoundGeneration(SoundGenerationRequest) returns (Result) {}
+  rpc TokenizeString(PredictOptions) returns (TokenizationResponse) {}
+  rpc Status(HealthMessage) returns (StatusResponse) {}
+
+  rpc StoresSet(StoresSetOptions) returns (Result) {}
+  rpc StoresDelete(StoresDeleteOptions) returns (Result) {}
+  rpc StoresGet(StoresGetOptions) returns (StoresGetResult) {}
+  rpc StoresFind(StoresFindOptions) returns (StoresFindResult) {}
+
+  rpc Rerank(RerankRequest) returns (RerankResult) {}
+
+  rpc GetMetrics(MetricsRequest) returns (MetricsResponse);
+}
+
+// Define the empty request
+message MetricsRequest {}
+
+message MetricsResponse {
+  int32 slot_id = 1;
+  string prompt_json_for_slot = 2;  // Stores the prompt as a JSON string.
+  float tokens_per_second = 3;
+  int32 tokens_generated = 4;
+  int32 prompt_tokens_processed = 5;
+}
+
+message RerankRequest {
+  string query = 1;
+  repeated string documents = 2;
+  int32 top_n = 3;
+}
+
+message RerankResult {
+  Usage usage = 1;
+  repeated DocumentResult results = 2;
+}
+
+message Usage {
+  int32 total_tokens = 1;
+  int32 prompt_tokens = 2;
+}
+
+message DocumentResult {
+  int32 index = 1;
+  string text = 2;
+  float relevance_score = 3;
+}
+
+message StoresKey {
+  repeated float Floats = 1;
+}
+
+message StoresValue {
+  bytes Bytes = 1;
+}
+
+message StoresSetOptions {
+  repeated StoresKey Keys = 1;
+  repeated StoresValue Values = 2;
+}
+
+message StoresDeleteOptions {
+  repeated StoresKey Keys = 1;
+}
+
+message StoresGetOptions {
+  repeated StoresKey Keys = 1;
+}
+
+message StoresGetResult {
+  repeated StoresKey Keys = 1;
+  repeated StoresValue Values = 2;
+}
+
+message StoresFindOptions {
+  StoresKey Key = 1;
+  int32 TopK = 2;
+}
+
+message StoresFindResult {
+  repeated StoresKey Keys = 1;
+  repeated StoresValue Values = 2;
+  repeated float Similarities = 3;
+}
+
+message HealthMessage {}
+
+// The request message containing the user's name.
+message PredictOptions {
+  string Prompt = 1;
+  int32 Seed = 2;
+  int32 Threads = 3;
+  int32 Tokens = 4;
+  int32 TopK = 5;
+  int32 Repeat = 6;
+  int32 Batch = 7;
+  int32 NKeep = 8;
+  float Temperature = 9;
+  float Penalty = 10;
+  bool F16KV = 11;
+  bool DebugMode = 12;
+  repeated string StopPrompts = 13;
+  bool IgnoreEOS = 14;
+  float TailFreeSamplingZ = 15;
+  float TypicalP = 16;
+  float FrequencyPenalty = 17;
+  float PresencePenalty = 18;
+  int32 Mirostat = 19;
+  float MirostatETA = 20;
+  float MirostatTAU = 21;
+  bool PenalizeNL = 22;
+  string LogitBias = 23;
+  bool MLock = 25;
+  bool MMap = 26;
+  bool PromptCacheAll = 27;
+  bool PromptCacheRO = 28;
+  string Grammar = 29;
+  string MainGPU = 30;
+  string TensorSplit = 31;
+  float TopP = 32;
+  string PromptCachePath = 33;
+  bool Debug = 34;
+  repeated int32 EmbeddingTokens = 35;
+  string Embeddings = 36;
+  float RopeFreqBase = 37;
+  float RopeFreqScale = 38;
+  float NegativePromptScale = 39;
+  string NegativePrompt = 40;
+  int32 NDraft = 41;
+  repeated string Images = 42;
+  bool UseTokenizerTemplate = 43;
+  repeated Message Messages = 44;
+  repeated string Videos = 45;
+  repeated string Audios = 46;
+  string CorrelationId = 47;
+}
+
+// The response message containing the result
+message Reply {
+  bytes message = 1;
+  int32 tokens = 2;
+  int32 prompt_tokens = 3;
+}
+
+message ModelOptions {
+  string Model = 1;
+  int32 ContextSize = 2;
+  int32 Seed = 3;
+  int32 NBatch = 4;
+  bool F16Memory = 5;
+  bool MLock = 6;
+  bool MMap = 7;
+  bool VocabOnly = 8;
+  bool LowVRAM = 9;
+  bool Embeddings = 10;
+  bool NUMA = 11;
+  int32 NGPULayers = 12;
+  string MainGPU = 13;
+  string TensorSplit = 14;
+  int32 Threads = 15;
+  string LibrarySearchPath = 16;
+  float RopeFreqBase = 17;
+  float RopeFreqScale = 18;
+  float RMSNormEps = 19;
+  int32 NGQA = 20;
+  string ModelFile = 21;
+
+  // AutoGPTQ
+  string Device = 22;
+  bool UseTriton = 23;
+  string ModelBaseName = 24;
+  bool UseFastTokenizer = 25;
+
+  // Diffusers
+  string PipelineType = 26;
+  string SchedulerType = 27;
+  bool CUDA = 28;
+  float CFGScale = 29;
+  bool IMG2IMG = 30;
+  string CLIPModel = 31;
+  string CLIPSubfolder = 32;
+  int32 CLIPSkip = 33;
+  string ControlNet = 48;
+
+  string Tokenizer = 34;
+
+  // LLM (llama.cpp)
+  string LoraBase = 35;
+  string LoraAdapter = 36;
+  float LoraScale = 42;
+
+  bool NoMulMatQ = 37;
+  string DraftModel = 39;
+
+  string AudioPath = 38;
+
+  // vllm
+  string Quantization = 40;
+  float  GPUMemoryUtilization = 50;
+  bool   TrustRemoteCode = 51;
+  bool   EnforceEager = 52;
+  int32  SwapSpace = 53;
+  int32  MaxModelLen = 54;
+  int32  TensorParallelSize = 55;
+  string LoadFormat = 58;
+
+  string MMProj = 41;
+
+  string RopeScaling = 43;
+  float YarnExtFactor = 44;
+  float YarnAttnFactor = 45;
+  float YarnBetaFast = 46;
+  float YarnBetaSlow = 47;
+
+  string Type = 49;
+
+  bool FlashAttention = 56;
+  bool NoKVOffload = 57;
+
+  string ModelPath = 59;
+
+  repeated string LoraAdapters = 60;
+  repeated float LoraScales = 61;
+}
+
+message Result {
+  string message = 1;
+  bool success = 2;
+}
+
+message EmbeddingResult {
+  repeated float embeddings = 1;
+}
+
+message TranscriptRequest {
+  string dst = 2;
+  string language = 3;
+  uint32 threads = 4;
+  bool translate = 5;
+}
+
+message TranscriptResult {
+  repeated TranscriptSegment segments = 1;
+  string text = 2;
+}
+
+message TranscriptSegment {
+  int32 id = 1;
+  int64 start = 2;
+  int64 end = 3;
+  string text = 4;
+  repeated int32 tokens = 5;
+}
+
+message GenerateImageRequest {
+  int32 height = 1;
+  int32 width = 2;
+  int32 mode = 3;
+  int32 step = 4;
+  int32 seed = 5;
+  string positive_prompt = 6;
+  string negative_prompt = 7;
+  string dst = 8;
+  string src = 9;
+
+  // Diffusers
+  string EnableParameters = 10;
+  int32 CLIPSkip = 11;
+}
+
+message TTSRequest {
+  string text = 1;
+  string model = 2;
+  string dst = 3;
+  string voice = 4;
+  optional string language = 5;
+}
+
+message SoundGenerationRequest {
+  string text = 1;
+  string model = 2;
+  string dst = 3;
+  optional float duration = 4;
+  optional float temperature = 5;
+  optional bool sample = 6;
+  optional string src = 7;
+  optional int32 src_divisor = 8;
+}
+
+message TokenizationResponse {
+  int32 length = 1;
+  repeated int32 tokens = 2;
+}
+
+message MemoryUsageData {
+  uint64 total = 1;
+  map<string, uint64> breakdown = 2;
+}
+
+message StatusResponse {
+  enum State {
+    UNINITIALIZED = 0;
+    BUSY = 1;
+    READY = 2;
+    ERROR = -1;
+  }
+  State state = 1;
+  MemoryUsageData memory = 2;
+}
+
+message Message {
+  string role = 1;
+  string content = 2;
+}

backend/cpp/grpc/.gitignore

@@ -0,0 +1,3 @@
+installed_packages/
+grpc_build/
+grpc_repo/

backend/cpp/grpc/Makefile

@@ -0,0 +1,70 @@
+# Basic platform detection
+HOST_SYSTEM = $(shell uname | cut -f 1 -d_)
+SYSTEM ?= $(HOST_SYSTEM)
+
+TAG_LIB_GRPC?=v1.59.0
+GIT_REPO_LIB_GRPC?=https://github.com/grpc/grpc.git
+GIT_CLONE_DEPTH?=1
+
+INSTALLED_PACKAGES=installed_packages
+GRPC_REPO=grpc_repo
+GRPC_BUILD=grpc_build
+
+export CMAKE_ARGS?=
+CMAKE_ARGS+=-DCMAKE_BUILD_TYPE=Release
+CMAKE_ARGS+=-DgRPC_INSTALL=ON
+CMAKE_ARGS+=-DEXECUTABLE_OUTPUT_PATH=../$(INSTALLED_PACKAGES)/grpc/bin
+CMAKE_ARGS+=-DLIBRARY_OUTPUT_PATH=../$(INSTALLED_PACKAGES)/grpc/lib
+CMAKE_ARGS+=-DgRPC_BUILD_TESTS=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_CSHARP_EXT=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_CPP_PLUGIN=ON
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_CSHARP_PLUGIN=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_NODE_PLUGIN=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_OBJECTIVE_C_PLUGIN=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_PHP_PLUGIN=OFF
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_PYTHON_PLUGIN=ON
+CMAKE_ARGS+=-DgRPC_BUILD_GRPC_RUBY_PLUGIN=OFF
+CMAKE_ARGS+=-Dprotobuf_WITH_ZLIB=ON
+CMAKE_ARGS+=-DRE2_BUILD_TESTING=OFF
+CMAKE_ARGS+=-DCMAKE_INSTALL_PREFIX=../$(INSTALLED_PACKAGES)
+
+# windows need to set OPENSSL_NO_ASM. Results in slower crypto performance but doesn't build otherwise.
+# May be resolvable, but for now its set. More info: https://stackoverflow.com/a/75240504/480673
+ifeq ($(SYSTEM),MSYS)
+CMAKE_ARGS+=-DOPENSSL_NO_ASM=ON
+endif
+ifeq ($(SYSTEM),MINGW64)
+CMAKE_ARGS+=-DOPENSSL_NO_ASM=ON
+endif
+ifeq ($(SYSTEM),MINGW32)
+CMAKE_ARGS+=-DOPENSSL_NO_ASM=ON
+endif
+ifeq ($(SYSTEM),CYGWIN)
+CMAKE_ARGS+=-DOPENSSL_NO_ASM=ON
+endif
+
+$(INSTALLED_PACKAGES): grpc_build
+
+$(GRPC_REPO):
+	mkdir -p $(GRPC_REPO)/grpc
+	cd $(GRPC_REPO)/grpc && \
+	git init && \
+	git remote add origin $(GIT_REPO_LIB_GRPC)  && \
+	git fetch origin && \
+	git checkout $(TAG_LIB_GRPC) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+	
+$(GRPC_BUILD): $(GRPC_REPO)
+	mkdir -p $(GRPC_BUILD)
+	cd $(GRPC_BUILD) && cmake $(CMAKE_ARGS) ../$(GRPC_REPO)/grpc && cmake --build . && cmake --build . --target install
+
+build: $(INSTALLED_PACKAGES)
+
+rebuild:
+	rm -rf grpc_build
+	$(MAKE) grpc_build
+
+clean:
+	rm -rf grpc_build
+	rm -rf grpc_repo
+	rm -rf installed_packages

backend/cpp/llama/CMakeLists.txt

@@ -0,0 +1,86 @@
+
+## XXX: In some versions of CMake clip wasn't being built before llama.
+## This is an hack for now, but it should be fixed in the future.
+set(TARGET myclip)
+add_library(${TARGET} clip.cpp clip.h llava.cpp llava.h)
+install(TARGETS ${TARGET} LIBRARY)
+target_include_directories(myclip PUBLIC .)
+target_include_directories(myclip PUBLIC ../..)
+target_include_directories(myclip PUBLIC ../../common)
+target_link_libraries(${TARGET} PRIVATE common ggml llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if (NOT MSVC)
+    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
+endif()
+# END CLIP hack
+
+
+set(TARGET grpc-server)
+set(CMAKE_CXX_STANDARD 17)
+cmake_minimum_required(VERSION 3.15)
+set(TARGET grpc-server)
+set(_PROTOBUF_LIBPROTOBUF libprotobuf)
+set(_REFLECTION grpc++_reflection)
+
+if (${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
+    # Set correct Homebrew install folder for Apple Silicon and Intel Macs
+    if (CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "arm64")
+        set(HOMEBREW_DEFAULT_PREFIX "/opt/homebrew")
+    else()
+        set(HOMEBREW_DEFAULT_PREFIX "/usr/local")
+    endif()
+
+    link_directories("${HOMEBREW_DEFAULT_PREFIX}/lib")
+    include_directories("${HOMEBREW_DEFAULT_PREFIX}/include")
+endif()
+
+find_package(absl CONFIG REQUIRED)
+find_package(Protobuf CONFIG REQUIRED)
+find_package(gRPC CONFIG REQUIRED)
+
+find_program(_PROTOBUF_PROTOC protoc)
+set(_GRPC_GRPCPP grpc++)
+find_program(_GRPC_CPP_PLUGIN_EXECUTABLE grpc_cpp_plugin)
+
+include_directories(${CMAKE_CURRENT_BINARY_DIR})
+include_directories(${Protobuf_INCLUDE_DIRS})
+
+message(STATUS "Using protobuf version ${Protobuf_VERSION} | Protobuf_INCLUDE_DIRS: ${Protobuf_INCLUDE_DIRS} | CMAKE_CURRENT_BINARY_DIR: ${CMAKE_CURRENT_BINARY_DIR}")
+
+# Proto file
+get_filename_component(hw_proto "../../../../../../backend/backend.proto" ABSOLUTE)
+get_filename_component(hw_proto_path "${hw_proto}" PATH)
+
+# Generated sources
+set(hw_proto_srcs "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.cc")
+set(hw_proto_hdrs "${CMAKE_CURRENT_BINARY_DIR}/backend.pb.h")
+set(hw_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.cc")
+set(hw_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/backend.grpc.pb.h")
+
+add_custom_command(
+      OUTPUT "${hw_proto_srcs}" "${hw_proto_hdrs}" "${hw_grpc_srcs}" "${hw_grpc_hdrs}"
+      COMMAND ${_PROTOBUF_PROTOC}
+      ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}"
+        --cpp_out "${CMAKE_CURRENT_BINARY_DIR}"
+        -I "${hw_proto_path}"
+        --plugin=protoc-gen-grpc="${_GRPC_CPP_PLUGIN_EXECUTABLE}"
+        "${hw_proto}"
+      DEPENDS "${hw_proto}")
+
+# hw_grpc_proto
+add_library(hw_grpc_proto
+  ${hw_grpc_srcs}
+  ${hw_grpc_hdrs}
+  ${hw_proto_srcs}
+  ${hw_proto_hdrs} )
+
+add_executable(${TARGET} grpc-server.cpp utils.hpp json.hpp)
+target_link_libraries(${TARGET} PRIVATE common llama myclip ${CMAKE_THREAD_LIBS_INIT} absl::flags hw_grpc_proto
+  absl::flags_parse
+  gRPC::${_REFLECTION}
+  gRPC::${_GRPC_GRPCPP}
+  protobuf::${_PROTOBUF_LIBPROTOBUF})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

backend/cpp/llama/Makefile

@@ -0,0 +1,82 @@
+
+LLAMA_VERSION?=
+LLAMA_REPO?=https://github.com/ggerganov/llama.cpp
+
+CMAKE_ARGS?=
+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 && \
+	git init && \
+	git remote add origin $(LLAMA_REPO)  && \
+	git fetch origin && \
+	git checkout -b build $(LLAMA_VERSION) && \
+	git submodule update --init --recursive --depth 1 --single-branch
+
+llama.cpp/examples/grpc-server: llama.cpp
+	mkdir -p llama.cpp/examples/grpc-server
+	bash prepare.sh
+
+rebuild:
+	bash prepare.sh
+	rm -rf grpc-server
+	$(MAKE) grpc-server
+
+purge:
+	rm -rf llama.cpp/build
+	rm -rf llama.cpp/examples/grpc-server
+	rm -rf grpc-server
+
+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 $(CMAKE_ARGS)"
+ifneq (,$(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)"
+else
+	+cd llama.cpp && mkdir -p build && cd build && cmake .. $(CMAKE_ARGS) && cmake --build . --config Release $(TARGET)
+endif
+	cp llama.cpp/build/bin/grpc-server .

backend/cpp/llama/grpc-server.cpp

@@ -0,0 +1,2537 @@
+// llama.cpp gRPC C++ backend server
+//
+// Ettore Di Giacinto <mudler@localai.io> and llama.cpp authors
+//
+// This is a gRPC server for llama.cpp compatible with the LocalAI proto
+// Note: this is a re-adaptation of the original llama.cpp example/server.cpp for HTTP (https://github.com/ggerganov/llama.cpp/tree/master/examples/server), 
+// but modified to work with gRPC
+//
+
+#include <iostream>
+#include <memory>
+#include <string>
+#include <getopt.h>
+#include "clip.h"
+#include "llava.h"
+#include "log.h"
+#include "stb_image.h"
+#include "common.h"
+#include "json.hpp"
+#include "llama.h"
+#include "backend.pb.h"
+#include "backend.grpc.pb.h"
+#include "utils.hpp"
+#include "sampling.h"
+// include std::regex
+#include <cstddef>
+#include <thread>
+#include <mutex>
+#include <chrono>
+#include <regex>
+#include <condition_variable>
+#include <grpcpp/ext/proto_server_reflection_plugin.h>
+#include <grpcpp/grpcpp.h>
+#include <grpcpp/health_check_service_interface.h>
+#include <atomic>
+#include <signal.h>
+
+using grpc::Server;
+using grpc::ServerBuilder;
+using grpc::ServerContext;
+using grpc::Status;
+
+
+using backend::HealthMessage;
+
+
+///// LLAMA.CPP server code below
+
+using json = nlohmann::json;
+
+struct server_params
+{
+    std::string hostname = "127.0.0.1";
+    std::vector<std::string> api_keys;
+    std::string public_path = "examples/server/public";
+    std::string chat_template = "";
+    int32_t port = 8080;
+    int32_t read_timeout = 600;
+    int32_t write_timeout = 600;
+    bool slots_endpoint = true;
+    bool metrics_endpoint = false;
+};
+
+bool server_verbose = false;
+bool server_log_json = true;
+
+static size_t common_part(const std::vector<llama_token> &a, const std::vector<llama_token> &b)
+{
+    size_t i;
+    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++)
+    {
+    }
+    return i;
+}
+
+enum stop_type
+{
+    STOP_FULL,
+    STOP_PARTIAL,
+};
+
+static bool ends_with(const std::string &str, const std::string &suffix)
+{
+    return str.size() >= suffix.size() &&
+           0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
+}
+
+static size_t find_partial_stop_string(const std::string &stop,
+                                       const std::string &text)
+{
+    if (!text.empty() && !stop.empty())
+    {
+        const char text_last_char = text.back();
+        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--)
+        {
+            if (stop[char_index] == text_last_char)
+            {
+                const std::string current_partial = stop.substr(0, char_index + 1);
+                if (ends_with(text, current_partial))
+                {
+                    return text.size() - char_index - 1;
+                }
+            }
+        }
+    }
+    return std::string::npos;
+}
+
+// TODO: reuse llama_detokenize
+template <class Iter>
+static std::string tokens_to_str(llama_context *ctx, Iter begin, Iter end)
+{
+    std::string ret;
+    for (; begin != end; ++begin)
+    {
+        ret += common_token_to_piece(ctx, *begin);
+    }
+    return ret;
+}
+
+// format incomplete utf-8 multibyte character for output
+static std::string tokens_to_output_formatted_string(const llama_context *ctx, const llama_token token)
+{
+    std::string out = token == -1 ? "" : common_token_to_piece(ctx, token);
+    // if the size is 1 and first bit is 1, meaning it's a partial character
+    //   (size > 1 meaning it's already a known token)
+    if (out.size() == 1 && (out[0] & 0x80) == 0x80)
+    {
+        std::stringstream ss;
+        ss << std::hex << (out[0] & 0xff);
+        std::string res(ss.str());
+        out = "byte: \\x" + res;
+    }
+    return out;
+}
+
+// convert a vector of completion_token_output to json
+static json probs_vector_to_json(const llama_context *ctx, const std::vector<completion_token_output> &probs)
+{
+    json out = json::array();
+    for (const auto &prob : probs)
+    {
+        json probs_for_token = json::array();
+        for (const auto &p : prob.probs)
+        {
+            std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
+            probs_for_token.push_back(json
+            {
+                {"tok_str", tok_str},
+                {"prob",    p.prob},
+            });
+        }
+        std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
+        out.push_back(json{
+            {"content", tok_str},
+            {"probs",   probs_for_token},
+        });
+    }
+    return out;
+}
+
+struct llama_client_slot
+{
+    int id;
+    int task_id = -1;
+
+    struct slot_params params;
+
+    slot_state state = IDLE;
+    slot_command command = NONE;
+
+    // used to determine the slot that has been used the longest
+    int64_t t_last_used = -1;
+
+    // generation props
+    int32_t n_ctx       = 0;  // context size per slot
+    int32_t n_past      = 0;
+    int32_t n_decoded   = 0;
+    int32_t n_remaining = -1;
+    int32_t i_batch     = -1;
+    int32_t n_predict   = -1;
+
+    int32_t num_prompt_tokens           = 0;
+    int32_t num_prompt_tokens_processed = 0;
+
+    json prompt;
+    std::string generated_text;
+    llama_token sampled;
+    std::vector<llama_token> cache_tokens;
+    std::vector<completion_token_output> generated_token_probs;
+
+    bool infill = false;
+    bool embedding = false;
+    bool has_next_token = true;
+    bool truncated = false;
+    bool stopped_eos = false;
+    bool stopped_word = false;
+    bool stopped_limit = false;
+
+    bool oaicompat = false;
+    std::string oaicompat_model;
+
+    std::string stopping_word;
+
+    // sampling
+    struct common_sampler_params sparams;
+    common_sampler *ctx_sampling = nullptr;
+
+    int32_t ga_i = 0;   // group-attention state
+    int32_t ga_n = 1;   // group-attention factor
+    int32_t ga_w = 512; // group-attention width
+
+    int32_t n_past_se = 0; // self-extend
+
+    // multimodal
+    std::vector<slot_image> images;
+
+    // stats
+    size_t sent_count = 0;
+    size_t sent_token_probs_index = 0;
+
+    int64_t t_start_process_prompt;
+    int64_t t_start_genereration;
+
+    double t_prompt_processing; // ms
+    double t_token_generation; // ms
+
+    // multitasks
+    int multitask_id = -1;
+
+    void reset() {
+        num_prompt_tokens      = 0;
+        generated_text         = "";
+        truncated              = false;
+        stopped_eos            = false;
+        stopped_word           = false;
+        stopped_limit          = false;
+        stopping_word          = "";
+        n_past                 = 0;
+        sent_count             = 0;
+        sent_token_probs_index = 0;
+        infill                 = false;
+        ga_i                   = 0;
+        n_past_se              = 0;
+
+        generated_token_probs.clear();
+
+        for (slot_image & img : images)
+        {
+            free(img.image_embedding);
+            if (img.img_data) {
+                clip_image_u8_free(img.img_data);
+            }
+            img.prefix_prompt = "";
+        }
+
+        images.clear();
+    }
+
+    bool has_budget(common_params &global_params) {
+        if (params.n_predict == -1 && global_params.n_predict == -1)
+        {
+            return true; // limitless
+        }
+
+        n_remaining = -1;
+
+        if (params.n_predict != -1)
+        {
+            n_remaining = params.n_predict - n_decoded;
+        }
+        else if (global_params.n_predict != -1)
+        {
+            n_remaining = global_params.n_predict - n_decoded;
+        }
+
+        return n_remaining > 0; // no budget
+    }
+
+    bool available() const {
+        return state == IDLE && command == NONE;
+    }
+
+    bool is_processing() const {
+        return (state == IDLE && command == LOAD_PROMPT) || state == PROCESSING;
+    }
+
+    void add_token_string(const completion_token_output &token) {
+        if (command == RELEASE)
+        {
+            return;
+        }
+        cache_tokens.push_back(token.tok);
+        generated_token_probs.push_back(token);
+    }
+
+    void release() {
+        if (state == PROCESSING)
+        {
+            t_token_generation = (ggml_time_us() - t_start_genereration) / 1e3;
+            command = RELEASE;
+        }
+    }
+
+    json get_formated_timings() {
+        return json
+        {
+            {"prompt_n",               num_prompt_tokens_processed},
+            {"prompt_ms",              t_prompt_processing},
+            {"prompt_per_token_ms",    t_prompt_processing / num_prompt_tokens_processed},
+            {"prompt_per_second",      1e3 / t_prompt_processing * num_prompt_tokens_processed},
+
+            {"predicted_n",            n_decoded},
+            {"predicted_ms",           t_token_generation},
+            {"predicted_per_token_ms", t_token_generation / n_decoded},
+            {"predicted_per_second",   1e3 / t_token_generation * n_decoded},
+        };
+    }
+
+    void print_timings() const {
+       char buffer[512];
+        double t_token = t_prompt_processing / num_prompt_tokens_processed;
+        double n_tokens_second = 1e3 / t_prompt_processing * num_prompt_tokens_processed;
+        sprintf(buffer, "prompt eval time     = %10.2f ms / %5d tokens (%8.2f ms per token, %8.2f tokens per second)",
+                t_prompt_processing, num_prompt_tokens_processed,
+                t_token, n_tokens_second);
+        LOG_INFO(buffer, {
+            {"slot_id",                     id},
+            {"task_id",                     task_id},
+            {"t_prompt_processing",         t_prompt_processing},
+            {"num_prompt_tokens_processed", num_prompt_tokens_processed},
+            {"t_token",                     t_token},
+            {"n_tokens_second",             n_tokens_second},
+        });
+
+        t_token = t_token_generation / n_decoded;
+        n_tokens_second = 1e3 / t_token_generation * n_decoded;
+        sprintf(buffer, "generation eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)",
+                t_token_generation, n_decoded,
+                t_token, n_tokens_second);
+        LOG_INFO(buffer, {
+            {"slot_id",            id},
+            {"task_id",            task_id},
+            {"t_token_generation", t_token_generation},
+            {"n_decoded",          n_decoded},
+            {"t_token",            t_token},
+            {"n_tokens_second",    n_tokens_second},
+        });
+
+        sprintf(buffer, "          total time = %10.2f ms", t_prompt_processing + t_token_generation);
+        LOG_INFO(buffer, {
+            {"slot_id",             id},
+            {"task_id",             task_id},
+            {"t_prompt_processing", t_prompt_processing},
+            {"t_token_generation",  t_token_generation},
+            {"t_total",             t_prompt_processing + t_token_generation},
+        });
+    }
+};
+
+struct llama_metrics {
+    uint64_t n_prompt_tokens_processed_total = 0;
+    uint64_t n_tokens_predicted_total        = 0;
+
+    uint64_t n_prompt_tokens_processed = 0;
+    uint64_t t_prompt_processing       = 0;
+
+    uint64_t n_tokens_predicted       = 0;
+    uint64_t t_tokens_generation      = 0;
+
+
+    void on_prompt_eval(const llama_client_slot &slot) {
+        n_prompt_tokens_processed_total += slot.num_prompt_tokens_processed;
+
+        n_prompt_tokens_processed += slot.num_prompt_tokens_processed;
+        t_prompt_processing       += slot.t_prompt_processing;
+    }
+
+    void on_prediction(const llama_client_slot &slot) {
+        n_tokens_predicted_total += slot.n_decoded;
+
+        n_tokens_predicted  += slot.n_decoded;
+        t_tokens_generation += slot.t_token_generation;
+    }
+
+    void reset_bucket() {
+        n_prompt_tokens_processed = 0;
+        t_prompt_processing       = 0;
+        n_tokens_predicted        = 0;
+        t_tokens_generation       = 0;
+    }
+};
+
+struct llava_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
+struct llama_server_context
+{
+    llama_model *model = nullptr;
+    llama_context *ctx = nullptr;
+
+    clip_ctx *clp_ctx = nullptr;
+
+    common_params params;
+
+    llama_batch batch;
+
+    bool multimodal         = false;
+    bool clean_kv_cache     = true;
+    bool all_slots_are_idle = false;
+    bool add_bos_token      = true;
+
+    int32_t n_ctx;  // total context for all clients / slots
+
+    // system prompt
+    bool system_need_update = false;
+
+    std::string              system_prompt;
+    std::vector<llama_token> system_tokens;
+
+    std::string name_user;      // this should be the antiprompt
+    std::string name_assistant;
+
+    // slots / clients
+    std::vector<llama_client_slot> slots;
+    json default_generation_settings_for_props;
+
+    llama_server_queue queue_tasks;
+    llama_server_response queue_results;
+
+    llama_metrics metrics;
+
+    ~llama_server_context()
+    {
+        if (ctx)
+        {
+            llama_free(ctx);
+            ctx = nullptr;
+        }
+        if (model)
+        {
+            llama_free_model(model);
+            model = nullptr;
+        }
+    }
+
+    bool load_model(const common_params &params_)
+    {
+        params = params_;
+        if (!params.mmproj.empty()) {
+            multimodal = true;
+            LOG_INFO("Multi Modal Mode Enabled", {});
+            clp_ctx = clip_model_load(params.mmproj.c_str(), /*verbosity=*/ 1);
+            if(clp_ctx == nullptr) {
+                LOG_ERR("unable to load clip model: %s", params.mmproj.c_str());
+                return false;
+            }
+
+            if (params.n_ctx < 2048) { // request larger context for the image embedding
+                params.n_ctx = 2048;
+            }
+        }
+
+        common_init_result common_init = common_init_from_params(params);
+        model = common_init.model;
+        ctx = common_init.context;
+        if (model == nullptr)
+        {
+            LOG_ERR("unable to load model: %s", params.model.c_str());
+            return false;
+        }
+
+        if (multimodal) {
+            const int n_embd_clip = clip_n_mmproj_embd(clp_ctx);
+            const int n_embd_llm  = llama_n_embd(model);
+            if (n_embd_clip != n_embd_llm) {
+                LOG("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_embd_clip, n_embd_llm);
+                llama_free(ctx);
+                llama_free_model(model);
+                return false;
+            }
+        }
+
+        n_ctx = llama_n_ctx(ctx);
+
+        add_bos_token = llama_add_bos_token(model);
+
+        return true;
+    }
+
+    void validate_model_chat_template(server_params & sparams) {
+        llama_chat_message chat[] = {{"user", "test"}};
+        std::vector<char> buf(1);
+        int res = llama_chat_apply_template(model, nullptr, chat, 1, true, buf.data(), buf.size());
+        if (res < 0) {
+            LOG_ERR("The chat template comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses", __func__);
+            sparams.chat_template = "<|im_start|>"; // llama_chat_apply_template only checks if <|im_start|> exist in the template
+        }
+    }
+
+    llama_client_slot* get_active_slot() {
+        for (llama_client_slot& slot : slots) {
+            // Check if the slot is currently processing
+            if (slot.is_processing()) {
+                return &slot;  // Return the active slot
+            }
+        }
+        return nullptr;  // No active slot found
+    }
+
+    void initialize() {
+        // create slots
+        all_slots_are_idle = true;
+
+        const int32_t n_ctx_slot = n_ctx / params.n_parallel;
+
+        LOG_INFO("initializing slots", {{"n_slots", params.n_parallel}});
+        for (int i = 0; i < params.n_parallel; i++)
+        {
+            llama_client_slot slot;
+
+            slot.id = i;
+            slot.n_ctx = n_ctx_slot;
+            slot.n_predict = params.n_predict;
+
+            LOG_INFO("new slot", {
+                {"slot_id",    slot.id},
+                {"n_ctx_slot", slot.n_ctx}
+            });
+
+            const int ga_n = params.grp_attn_n;
+            const int ga_w = params.grp_attn_w;
+
+            if (ga_n != 1) {
+                GGML_ASSERT(ga_n > 0                    && "ga_n must be positive");                     // NOLINT
+                GGML_ASSERT(ga_w % ga_n == 0            && "ga_w must be a multiple of ga_n");     // NOLINT
+                //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of ga_w");    // NOLINT
+                //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * ga_n"); // NOLINT
+
+                LOG_INFO("slot self-extend", {
+                    {"slot_id",   slot.id},
+                    {"ga_n",      ga_n},
+                    {"ga_w",      ga_w}
+                });
+            }
+
+            slot.ga_i = 0;
+            slot.ga_n = ga_n;
+            slot.ga_w = ga_w;
+
+            slot.reset();
+
+            slots.push_back(slot);
+        }
+
+        default_generation_settings_for_props = get_formated_generation(slots.front());
+        default_generation_settings_for_props["seed"] = -1;
+
+        batch = llama_batch_init(n_ctx, 0, params.n_parallel);
+    }
+
+    std::vector<llama_token> tokenize(const json & json_prompt, bool add_bos) const
+    {
+        // TODO: currently, we tokenize using special tokens by default
+        //       this is not always correct (see https://github.com/ggerganov/llama.cpp/pull/4160#issuecomment-1824826216)
+        //       but it's better compared to completely ignoring ChatML and other chat templates
+        const bool TMP_FORCE_SPECIAL = true;
+
+        // 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;
+
+        if (json_prompt.is_array())
+        {
+            bool first = true;
+            for (const auto& p : json_prompt)
+            {
+                if (p.is_string())
+                {
+                    auto s = p.template get<std::string>();
+                    std::vector<llama_token> p;
+                    if (first)
+                    {
+                        p = common_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
+                        first = false;
+                    }
+                    else
+                    {
+                        p = common_tokenize(ctx, s, false, TMP_FORCE_SPECIAL);
+                    }
+                    prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+                }
+                else
+                {
+                    if (first)
+                    {
+                        first = false;
+                    }
+                    prompt_tokens.push_back(p.template get<llama_token>());
+                }
+            }
+        }
+        else
+        {
+            auto s = json_prompt.template get<std::string>();
+            prompt_tokens = common_tokenize(ctx, s, add_bos, TMP_FORCE_SPECIAL);
+        }
+
+        return prompt_tokens;
+    }
+
+    llama_client_slot* get_slot(int id) {
+        int64_t t_last = ggml_time_us();
+        llama_client_slot *last_used = nullptr;
+
+        for (llama_client_slot & slot : slots)
+        {
+            if (slot.id == id && slot.available())
+            {
+                return &slot;
+            }
+
+            if (slot.available() && slot.t_last_used < t_last)
+            {
+                last_used = &slot;
+                t_last = slot.t_last_used;
+            }
+        }
+
+        return last_used;
+    }
+
+    bool launch_slot_with_data(llama_client_slot* &slot, json data) {
+        slot_params default_params;
+        common_sampler_params default_sparams;
+ 
+        slot->params.stream             = json_value(data, "stream",            false);
+        slot->params.cache_prompt       = json_value(data, "cache_prompt",      false);
+        slot->params.n_predict          = json_value(data, "n_predict",         default_params.n_predict);
+        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.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);
+        slot->sparams.dynatemp_exponent = json_value(data, "dynatemp_exponent", default_sparams.dynatemp_exponent);
+        slot->sparams.penalty_last_n    = json_value(data, "repeat_last_n",     default_sparams.penalty_last_n);
+        slot->sparams.penalty_repeat    = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
+        slot->sparams.penalty_freq      = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
+        slot->sparams.penalty_present   = json_value(data, "presence_penalty",  default_sparams.penalty_present);
+        slot->sparams.mirostat          = json_value(data, "mirostat",          default_sparams.mirostat);
+        slot->sparams.mirostat_tau      = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
+        slot->sparams.mirostat_eta      = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
+        slot->sparams.penalize_nl       = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
+        slot->params.n_keep             = json_value(data, "n_keep",            slot->params.n_keep);
+        slot->sparams.seed               = json_value(data, "seed",              default_sparams.seed);
+        slot->sparams.grammar           = json_value(data, "grammar",           default_sparams.grammar);
+        slot->sparams.n_probs           = json_value(data, "n_probs",           default_sparams.n_probs);
+        slot->sparams.min_keep          = json_value(data, "min_keep",          default_sparams.min_keep);
+
+        if (slot->n_predict > 0 && slot->params.n_predict > slot->n_predict) {
+            // Might be better to reject the request with a 400 ?
+            LOG_WARNING("Max tokens to predict exceeds server configuration", {
+                {"params.n_predict", slot->params.n_predict},
+                {"slot.n_predict", slot->n_predict},
+            });
+            slot->params.n_predict = slot->n_predict;
+        }
+
+        // infill
+        if (data.count("input_prefix") != 0)
+        {
+            slot->params.input_prefix = data["input_prefix"];
+        }
+        else
+        {
+            slot->params.input_prefix = "";
+        }
+
+
+        if (data.count("input_suffix") != 0)
+        {
+            slot->params.input_suffix = data["input_suffix"];
+        }
+        else
+        {
+            slot->params.input_suffix = "";
+        }
+
+        if (data.count("prompt") != 0)
+        {
+            slot->prompt = data["prompt"];
+        }
+        else
+        {
+            slot->prompt = "";
+        }
+
+        if (json_value(data, "ignore_eos", false)) {
+                slot->sparams.logit_bias.push_back({llama_token_eos(model), -INFINITY});
+        }
+        /*
+        slot->sparams.penalty_prompt_tokens.clear();
+        slot->sparams.use_penalty_prompt_tokens = false;
+        const auto &penalty_prompt = data.find("penalty_prompt");
+        if (penalty_prompt != data.end())
+        {
+            if (penalty_prompt->is_string())
+            {
+                const auto penalty_prompt_string = penalty_prompt->get<std::string>();
+                auto penalty_tokens = llama_tokenize(model, penalty_prompt_string, false);
+                slot->sparams.penalty_prompt_tokens.swap(penalty_tokens);
+                if (slot->params.n_predict > 0)
+                {
+                    slot->sparams.penalty_prompt_tokens.reserve(slot->sparams.penalty_prompt_tokens.size() + slot->params.n_predict);
+                }
+                slot->sparams.use_penalty_prompt_tokens = true;
+            }
+            else if (penalty_prompt->is_array())
+            {
+                const auto n_tokens = penalty_prompt->size();
+                slot->sparams.penalty_prompt_tokens.reserve(n_tokens + std::max(0, slot->params.n_predict));
+                const int n_vocab = llama_n_vocab(model);
+                for (const auto &penalty_token : *penalty_prompt)
+                {
+                    if (penalty_token.is_number_integer())
+                    {
+                        const auto tok = penalty_token.get<llama_token>();
+                        if (tok >= 0 && tok < n_vocab)
+                        {
+                            slot->sparams.penalty_prompt_tokens.push_back(tok);
+                        }
+                    }
+                }
+                slot->sparams.use_penalty_prompt_tokens = true;
+            }
+        }
+      */
+
+        slot->sparams.logit_bias.clear();
+
+        const auto &logit_bias = data.find("logit_bias");
+        if (logit_bias != data.end() && logit_bias->is_array())
+        {
+            const int n_vocab = llama_n_vocab(model);
+            for (const auto &el : *logit_bias)
+            {
+                if (el.is_array() && el.size() == 2)
+                {
+                    float bias;
+                    if (el[1].is_number())
+                    {
+                        bias = el[1].get<float>();
+                    }
+                    else if (el[1].is_boolean() && !el[1].get<bool>())
+                    {
+                        bias = -INFINITY;
+                    }
+                    else
+                    {
+                        continue;
+                    }
+
+                    if (el[0].is_number_integer())
+                    {
+                        llama_token tok = el[0].get<llama_token>();
+                        if (tok >= 0 && tok < n_vocab)
+                        {
+                            slot->sparams.logit_bias.push_back({tok, bias});
+                        }
+                    }
+                    else if (el[0].is_string())
+                    {
+                        auto toks = common_tokenize(model, el[0].get<std::string>(), false);
+                        for (auto tok : toks)
+                        {
+                            slot->sparams.logit_bias.push_back({tok, bias});
+                        }
+                    }
+                }
+            }
+        }
+        
+        slot->params.antiprompt.clear();
+
+        const auto &stop = data.find("stop");
+        if (stop != data.end() && stop->is_array())
+        {
+            for (const auto &word : *stop)
+            {
+                if (!word.empty())
+                {
+                    slot->params.antiprompt.push_back(word);
+                }
+            }
+        }
+        
+        const auto & samplers = data.find("samplers");
+        if (samplers != data.end() && samplers->is_array()) {
+            std::vector<std::string> sampler_names;
+                for (const auto & name : *samplers) {
+                    if (name.is_string()) {
+                        sampler_names.emplace_back(name);
+                    }
+                }
+                slot->sparams.samplers = common_sampler_types_from_names(sampler_names, false);
+        }
+        else
+        {
+                slot->sparams.samplers = default_sparams.samplers;
+        }
+        
+
+        if (multimodal)
+        {
+            const auto &images_data = data.find("image_data");
+            if (images_data != data.end() && images_data->is_array())
+            {
+                for (const auto &img : *images_data)
+                {
+                    const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
+
+                    slot_image img_sl;
+                    img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
+                    img_sl.img_data = clip_image_u8_init();
+                    if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
+                    {
+                        LOG_ERR("%s: failed to load image, slot_id: %d, img_sl_id: %d", 
+                             __func__,
+                             slot->id,
+                             img_sl.id
+                        );
+                        return false;
+                    }
+                    LOG_VERBOSE("image loaded", {
+                        {"slot_id",   slot->id},
+                        {"img_sl_id", img_sl.id}
+                    });
+                    img_sl.request_encode_image = true;
+                    slot->images.push_back(img_sl);
+                }
+                // process prompt
+                // example: system prompt [img-102] user [img-103] describe [img-134] -> [{id: 102, prefix: 'system prompt '}, {id: 103, prefix: ' user '}, {id: 134, prefix: ' describe '}]}
+                if (slot->images.size() > 0 && !slot->prompt.is_array())
+                {
+                    std::string prompt = slot->prompt.get<std::string>();
+                    size_t pos = 0, begin_prefix = 0;
+                    std::string pattern = "[img-";
+                    while ((pos = prompt.find(pattern, pos)) != std::string::npos) {
+                        size_t end_prefix = pos;
+                        pos += pattern.length();
+                        size_t end_pos = prompt.find(']', pos);
+                        if (end_pos != std::string::npos)
+                        {
+                            std::string image_id = prompt.substr(pos, end_pos - pos);
+                            try
+                            {
+                                int img_id = std::stoi(image_id);
+                                bool found = false;
+                                for (slot_image &img : slot->images)
+                                {
+                                    if (img.id == img_id) {
+                                        found = true;
+                                        img.prefix_prompt = prompt.substr(begin_prefix, end_prefix - begin_prefix);
+                                        begin_prefix = end_pos + 1;
+                                        break;
+                                    }
+                                }
+                                if (!found) {
+                                    LOG("ERROR: Image with id: %i, not found.\n", img_id);
+                                    slot->images.clear();
+                                    return false;
+                                }
+                            } catch (const std::invalid_argument& e) {
+                                LOG("Invalid image number id in prompt\n");
+                                slot->images.clear();
+                                return false;
+                            }
+                        }
+                    }
+                    slot->prompt = "";
+                    slot->params.input_suffix = prompt.substr(begin_prefix);
+                    slot->params.cache_prompt = false; // multimodal doesn't support cache prompt
+                }
+            }
+        }
+
+        if (slot->ctx_sampling != nullptr)
+        {
+            common_sampler_free(slot->ctx_sampling);
+        }
+        slot->ctx_sampling = common_sampler_init(model, slot->sparams);
+        //llama_set_rng_seed(ctx, slot->params.seed);
+        slot->command = LOAD_PROMPT;
+
+        all_slots_are_idle = false;
+
+        LOG_INFO("slot is processing task", {
+            {"slot_id", slot->id},
+            {"task_id", slot->task_id},
+        });
+
+      //  LOG("sampling: \n%s\n", llama_sampling_print(slot->sparams).c_str());
+
+        return true;
+    }
+
+    void kv_cache_clear() {
+        // clear the entire KV cache
+        llama_kv_cache_clear(ctx);
+        clean_kv_cache = false;
+    }
+
+    void update_system_prompt() {
+        kv_cache_clear();
+        system_tokens.clear();
+
+        if (!system_prompt.empty()) {
+            system_tokens = common_tokenize(ctx, system_prompt, add_bos_token);
+
+            common_batch_clear(batch);
+
+            for (int i = 0; i < (int)system_tokens.size(); ++i)
+            {
+                common_batch_add(batch, system_tokens[i], i, { 0 }, false);
+            }
+
+            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += params.n_batch)
+            {
+                const int32_t n_tokens = std::min(params.n_batch, (int32_t) (batch.n_tokens - i));
+                llama_batch batch_view = {
+                    n_tokens,
+                    batch.token    + i,
+                    nullptr,
+                    batch.pos      + i,
+                    batch.n_seq_id + i,
+                    batch.seq_id   + i,
+                    batch.logits   + i,
+                };
+                if (llama_decode(ctx, batch_view) != 0)
+                {
+                    LOG("%s: llama_decode() failed\n", __func__);
+                    return;
+                }
+            }
+
+            // assign the system KV cache to all parallel sequences
+            for (int32_t i = 1; i < params.n_parallel; ++i)
+            {
+                llama_kv_cache_seq_cp(ctx, 0, i, 0, system_tokens.size());
+            }
+        }
+
+        LOG("system prompt updated\n");
+        system_need_update = false;
+    }
+
+    void notify_system_prompt_changed() {
+        // release all slots
+        for (llama_client_slot &slot : slots)
+        {
+            slot.release();
+        }
+
+        system_need_update = true;
+    }
+
+    void process_system_prompt_data(const json &sys_props) {
+        system_prompt  = sys_props.value("prompt", "");
+        name_user      = sys_props.value("anti_prompt", "");
+        name_assistant = sys_props.value("assistant_name", "");
+
+
+        notify_system_prompt_changed();
+    }
+
+    static size_t find_stopping_strings(const std::string &text, const size_t last_token_size,
+                                        const stop_type type, llama_client_slot &slot)
+    {
+        size_t stop_pos = std::string::npos;
+
+        for (const std::string &word : slot.params.antiprompt)
+        {
+            size_t pos;
+            if (type == STOP_FULL)
+            {
+                const size_t tmp = word.size() + last_token_size;
+                const size_t from_pos = text.size() > tmp ? text.size() - tmp : 0;
+                pos = text.find(word, from_pos);
+            }
+            else
+            {
+                pos = find_partial_stop_string(word, text);
+            }
+            if (pos != std::string::npos &&
+                (stop_pos == std::string::npos || pos < stop_pos))
+            {
+                if (type == STOP_FULL)
+                {
+                    slot.stopped_word = true;
+                    slot.stopping_word = word;
+                    slot.has_next_token = false;
+                }
+                stop_pos = pos;
+            }
+        }
+
+        return stop_pos;
+    }
+
+    bool process_token(completion_token_output &result, llama_client_slot &slot) {
+        // remember which tokens were sampled - used for repetition penalties during sampling
+        const std::string token_str = common_token_to_piece(ctx, result.tok);
+        slot.sampled = result.tok;
+
+        // search stop word and delete it
+        slot.generated_text += token_str;
+        slot.has_next_token = true;
+
+/*
+        if (slot.ctx_sampling->params.use_penalty_prompt_tokens && result.tok != -1)
+        {
+            // we can change penalty_prompt_tokens because it is always created from scratch each request
+            slot.ctx_sampling->params.penalty_prompt_tokens.push_back(result.tok);
+        }
+        */
+
+        // check if there is incomplete UTF-8 character at the end
+        bool incomplete = false;
+        for (unsigned i = 1; i < 5 && i <= slot.generated_text.size(); ++i)
+        {
+            unsigned char c = slot.generated_text[slot.generated_text.size() - i];
+            if ((c & 0xC0) == 0x80)
+            {
+                // continuation byte: 10xxxxxx
+                continue;
+            }
+            if ((c & 0xE0) == 0xC0)
+            {
+                // 2-byte character: 110xxxxx ...
+                incomplete = i < 2;
+            }
+            else if ((c & 0xF0) == 0xE0)
+            {
+                // 3-byte character: 1110xxxx ...
+                incomplete = i < 3;
+            }
+            else if ((c & 0xF8) == 0xF0)
+            {
+                // 4-byte character: 11110xxx ...
+                incomplete = i < 4;
+            }
+            // else 1-byte character or invalid byte
+            break;
+        }
+
+        if (!incomplete)
+        {
+            size_t pos = std::min(slot.sent_count, slot.generated_text.size());
+            const std::string str_test = slot.generated_text.substr(pos);
+            bool is_stop_full = false;
+            size_t stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_FULL, slot);
+            if (stop_pos != std::string::npos)
+            {
+                is_stop_full = true;
+                slot.generated_text.erase(
+                    slot.generated_text.begin() + pos + stop_pos,
+                    slot.generated_text.end());
+                pos = std::min(slot.sent_count, slot.generated_text.size());
+            }
+            else
+            {
+                is_stop_full = false;
+                stop_pos = find_stopping_strings(str_test, token_str.size(), STOP_PARTIAL, slot);
+            }
+
+            // check if there is any token to predict
+            if (stop_pos == std::string::npos || (!slot.has_next_token && !is_stop_full && stop_pos > 0))
+            {
+                // no send the stop word in the response
+                result.text_to_send = slot.generated_text.substr(pos, std::string::npos);
+                slot.sent_count += result.text_to_send.size();
+                // add the token to slot queue and cache
+            }
+            slot.add_token_string(result);
+            if (slot.params.stream)
+            {
+                send_partial_response(slot, result);
+            }
+        }
+
+        if (incomplete)
+        {
+            slot.has_next_token = true;
+        }
+
+        // check the limits
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params))
+        {
+            slot.stopped_limit = true;
+            slot.has_next_token = false;
+        }
+
+        if (result.tok == llama_token_eos(model))
+        {
+            slot.stopped_eos = true;
+            slot.has_next_token = false;
+            LOG_VERBOSE("eos token found", {});
+        }
+
+        LOG_VERBOSE("next token", {
+                                      {"token", result.tok},
+                                      {"token_text", tokens_to_output_formatted_string(ctx, result.tok)},
+                                      {"has_next_token", slot.has_next_token},
+                                      {"n_remain", slot.n_remaining},
+                                      {"num_tokens_predicted", slot.n_decoded},
+                                      {"stopped_eos", slot.stopped_eos},
+                                      {"stopped_word", slot.stopped_word},
+                                      {"stopped_limit", slot.stopped_limit},
+                                      {"stopping_word", slot.stopping_word},
+                                  });
+
+        return slot.has_next_token; // continue
+    }
+
+    bool process_images(llama_client_slot &slot) const
+    {
+        for (slot_image &img : slot.images)
+        {
+            if (!img.request_encode_image)
+            {
+                continue;
+            }
+
+            if (!llava_image_embed_make_with_clip_img(clp_ctx, params.cpuparams.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
+                LOG("Error processing the given image");
+                return false;
+            }
+
+            img.request_encode_image = false;
+        }
+
+        return slot.images.size() > 0;
+    }
+
+    void send_error(task_server& task, const std::string &error)
+    {
+        LOG("task %i - error: %s\n", task.id, error.c_str());
+        task_result res;
+        res.id = task.id;
+        res.multitask_id = task.multitask_id;
+        res.stop = false;
+        res.error = true;
+        res.result_json = { { "content", error } };
+        queue_results.send(res);
+    }
+
+    json get_formated_generation(llama_client_slot &slot)
+    {
+        std::vector<std::string> samplers;
+        samplers.reserve(slot.sparams.samplers.size());
+        for (const auto & sampler : slot.sparams.samplers)
+        {
+            samplers.emplace_back(common_sampler_type_to_str(sampler));
+        }
+
+        return json {
+            {"n_ctx",             slot.n_ctx},
+            {"n_predict",         slot.n_predict},
+            {"model",             params.model_alias},
+            {"seed",              slot.params.seed},
+            {"temperature",       slot.sparams.temp},
+            {"dynatemp_range",    slot.sparams.dynatemp_range},
+            {"dynatemp_exponent", slot.sparams.dynatemp_exponent},
+            {"top_k",             slot.sparams.top_k},
+            {"top_p",             slot.sparams.top_p},
+            {"min_p",             slot.sparams.min_p},
+            {"typical_p",         slot.sparams.typ_p},
+            {"repeat_last_n",     slot.sparams.penalty_last_n},
+            {"repeat_penalty",    slot.sparams.penalty_repeat},
+            {"presence_penalty",  slot.sparams.penalty_present},
+            {"frequency_penalty", slot.sparams.penalty_freq},
+            {"mirostat",          slot.sparams.mirostat},
+            {"mirostat_tau",      slot.sparams.mirostat_tau},
+            {"mirostat_eta",      slot.sparams.mirostat_eta},
+            {"penalize_nl",       slot.sparams.penalize_nl},
+            {"stop",              slot.params.antiprompt},
+            {"n_predict",         slot.params.n_predict},
+            {"n_keep",            params.n_keep},
+            {"ignore_eos",        slot.sparams.ignore_eos},
+            {"stream",            slot.params.stream},
+      //      {"logit_bias",        slot.sparams.logit_bias},
+            {"n_probs",           slot.sparams.n_probs},
+            {"min_keep",          slot.sparams.min_keep},
+            {"grammar",           slot.sparams.grammar},
+            {"samplers",          samplers}
+        };
+    }
+
+    void send_partial_response(llama_client_slot &slot, completion_token_output tkn)
+    {
+        task_result res;
+        res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
+        res.error = false;
+        res.stop = false;
+
+        res.result_json = json
+        {
+            {"content",    tkn.text_to_send},
+            {"stop",       false},
+            {"slot_id",    slot.id},
+            {"multimodal", multimodal}
+        };
+
+        if (slot.sparams.n_probs > 0)
+        {
+            std::vector<completion_token_output> probs_output = {};
+            const std::vector<llama_token> to_send_toks = common_tokenize(ctx, tkn.text_to_send, false);
+            size_t probs_pos      = std::min(slot.sent_token_probs_index,                       slot.generated_token_probs.size());
+            size_t probs_stop_pos = std::min(slot.sent_token_probs_index + to_send_toks.size(), slot.generated_token_probs.size());
+            if (probs_pos < probs_stop_pos)
+            {
+                probs_output = std::vector<completion_token_output>(slot.generated_token_probs.begin() + probs_pos, slot.generated_token_probs.begin() + probs_stop_pos);
+            }
+            slot.sent_token_probs_index = probs_stop_pos;
+            res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs_output);
+        }
+
+        if (slot.oaicompat)
+        {
+            res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
+            res.result_json["model"] = slot.oaicompat_model;
+        }
+
+        queue_results.send(res);
+    }
+
+    void send_final_response(llama_client_slot &slot)
+    {
+        task_result res;
+        res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
+        res.error = false;
+        res.stop = true;
+
+        res.result_json = json
+        {
+            {"content",             !slot.params.stream ? slot.generated_text : ""},
+            {"slot_id",             slot.id},
+            {"stop",                true},
+            {"model",               params.model_alias},
+            {"tokens_predicted",    slot.n_decoded},
+            {"tokens_evaluated",    slot.num_prompt_tokens},
+            {"generation_settings", get_formated_generation(slot)},
+            {"prompt",              slot.prompt},
+            {"truncated",           slot.truncated},
+            {"stopped_eos",         slot.stopped_eos},
+            {"stopped_word",        slot.stopped_word},
+            {"stopped_limit",       slot.stopped_limit},
+            {"stopping_word",       slot.stopping_word},
+            {"tokens_cached",       slot.n_past},
+            {"timings",             slot.get_formated_timings()}
+        };
+
+        if (slot.sparams.n_probs > 0)
+        {
+            std::vector<completion_token_output> probs = {};
+            if (!slot.params.stream && slot.stopped_word)
+            {
+                const std::vector<llama_token> stop_word_toks = common_tokenize(ctx, slot.stopping_word, false);
+                probs = std::vector<completion_token_output>(slot.generated_token_probs.begin(), slot.generated_token_probs.end() - stop_word_toks.size());
+            }
+            else
+            {
+                probs = std::vector<completion_token_output>(
+                                    slot.generated_token_probs.begin(),
+                                    slot.generated_token_probs.end());
+            }
+            res.result_json["completion_probabilities"] = probs_vector_to_json(ctx, probs);
+        }
+
+        if (slot.oaicompat)
+        {
+            res.result_json["oaicompat_token_ctr"] = slot.n_decoded;
+            res.result_json["model"] = slot.oaicompat_model;
+        }
+
+        queue_results.send(res);
+    }
+
+    void send_embedding(llama_client_slot &slot)
+    {
+        task_result res;
+        res.id = slot.task_id;
+        res.multitask_id = slot.multitask_id;
+        res.error = false;
+        res.stop = true;
+
+        const int n_embd = llama_n_embd(model);
+        if (!params.embedding)
+        {
+            LOG_WARNING("embedding disabled", {
+                                                  {"params.embedding", params.embedding},
+                                              });
+            res.result_json = json
+            {
+                {"embedding", std::vector<float>(n_embd, 0.0f)},
+            };
+        }
+        else
+        {
+            const float *data = llama_get_embeddings(ctx);
+            std::vector<float> embedding(data, data + n_embd);
+            res.result_json = json
+            {
+                {"embedding", embedding },
+            };
+        }
+        queue_results.send(res);
+    }
+
+    void request_completion(int task_id, json data, bool infill, bool embedding, int multitask_id)
+    {
+        task_server task;
+        task.id = task_id;
+        task.target_id = 0;
+        task.data = std::move(data);
+        task.infill_mode = infill;
+        task.embedding_mode = embedding;
+        task.type = TASK_TYPE_COMPLETION;
+        task.multitask_id = multitask_id;
+
+        // when a completion task's prompt array is not a singleton, we split it into multiple requests
+        // otherwise, it's a single-prompt task, we actually queue it
+        // if there's numbers in the prompt array it will be treated as an array of tokens
+        if (task.data.count("prompt") != 0 && task.data.at("prompt").size() > 1) {
+            bool numbers = false;
+            for (const auto& e : task.data.at("prompt")) {
+                if (e.is_number()) {
+                    numbers = true;
+                    break;
+                }
+            }
+
+            // NOTE: split_multiprompt_task() does not handle a mix of strings and numbers,
+            // it will completely stall the server. I don't know where the bug for this is.
+            //
+            // if there are numbers, it needs to be treated like a single prompt,
+            // queue_tasks handles a mix of strings and numbers just fine.
+            if (numbers) {
+                queue_tasks.post(task);
+            } else {
+                split_multiprompt_task(task_id, task);
+            }
+        } else {
+            queue_tasks.post(task);
+        }
+    }
+
+    // for multiple images processing
+    bool ingest_images(llama_client_slot &slot, int n_batch)
+    {
+        int image_idx = 0;
+
+        while (image_idx < (int) slot.images.size())
+        {
+            slot_image &img = slot.images[image_idx];
+
+            // process prefix prompt
+            for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
+            {
+                const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+                llama_batch batch_view = {
+                    n_tokens,
+                    batch.token    + i,
+                    nullptr,
+                    batch.pos      + i,
+                    batch.n_seq_id + i,
+                    batch.seq_id   + i,
+                    batch.logits   + i,
+                };
+                if (llama_decode(ctx, batch_view))
+                {
+                    LOG("%s : failed to eval\n", __func__);
+                    return false;
+                }
+            }
+
+            // process image with llm
+            for (int i = 0; i < img.image_tokens; i += n_batch)
+            {
+                int n_eval = img.image_tokens - i;
+                if (n_eval > n_batch)
+                {
+                    n_eval = n_batch;
+                }
+
+                const int n_embd = llama_n_embd(model);
+                float * embd = img.image_embedding + i * n_embd;
+                llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, slot.n_past, 0);
+                if (llama_decode(ctx, llava_batch.batch))
+                {
+                    LOG("%s : failed to eval image\n", __func__);
+                    return false;
+                }
+                slot.n_past += n_eval;
+            }
+            image_idx++;
+
+            common_batch_clear(batch);
+
+            // append prefix of next image
+            const auto json_prompt = (image_idx >= (int) slot.images.size()) ?
+                slot.params.input_suffix : // no more images, then process suffix prompt
+                (json)(slot.images[image_idx].prefix_prompt);
+
+            std::vector<llama_token> append_tokens = tokenize(json_prompt, false); // has next image
+            for (int i = 0; i < (int) append_tokens.size(); ++i)
+            {
+                common_batch_add(batch, append_tokens[i], system_tokens.size() + slot.n_past, { slot.id }, true);
+                slot.n_past += 1;
+            }
+        }
+
+        return true;
+    }
+
+    void request_cancel(int task_id)
+    {
+        task_server task;
+        task.type = TASK_TYPE_CANCEL;
+        task.target_id = task_id;
+        queue_tasks.post(task);
+    }
+
+    void split_multiprompt_task(int multitask_id, task_server& multiprompt_task)
+    {
+        int prompt_count = multiprompt_task.data.at("prompt").size();
+        if (prompt_count <= 1) {
+            send_error(multiprompt_task, "error while handling multiple prompts");
+            return;
+        }
+
+        // generate all the ID for subtask
+        std::vector<int> subtask_ids(prompt_count);
+        for (int i = 0; i < prompt_count; i++)
+        {
+            subtask_ids[i] = queue_tasks.get_new_id();
+        }
+
+        // queue up the multitask so we can track its subtask progression
+        queue_tasks.add_multitask(multitask_id, subtask_ids);
+
+        // add subtasks
+        for (int i = 0; i < prompt_count; i++)
+        {
+            json subtask_data = multiprompt_task.data;
+            subtask_data["prompt"] = subtask_data["prompt"][i];
+
+            // subtasks inherit everything else (infill mode, embedding mode, etc.)
+            request_completion(subtask_ids[i], subtask_data, multiprompt_task.infill_mode, multiprompt_task.embedding_mode, multitask_id);
+        }
+    }
+
+    void process_single_task(task_server& task)
+    {
+        switch (task.type)
+        {
+            case TASK_TYPE_COMPLETION: {
+                llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
+                if (slot == nullptr)
+                {
+                    // if no slot is available, we defer this task for processing later
+                    LOG_VERBOSE("no slot is available", {{"task_id", task.id}});
+                    queue_tasks.defer(task);
+                    break;
+                }
+
+                if (task.data.contains("system_prompt"))
+                {
+                    if (!all_slots_are_idle) {
+                        send_error(task, "system prompt can only be updated when all slots are idle");
+                        break;
+                    }
+                    process_system_prompt_data(task.data["system_prompt"]);
+
+                    // reset cache_tokens for all slots
+                    for (llama_client_slot &slot : slots)
+                    {
+                        slot.cache_tokens.clear();
+                        slot.n_past    = 0;
+                        slot.n_past_se = 0;
+                    }
+                }
+
+                slot->reset();
+
+                slot->infill       = task.infill_mode;
+                slot->embedding    = task.embedding_mode;
+                slot->task_id      = task.id;
+                slot->multitask_id = task.multitask_id;
+
+                if (!launch_slot_with_data(slot, task.data))
+                {
+                    // send error result
+                    send_error(task, "internal_error");
+                    break;
+                }
+            } break;
+            case TASK_TYPE_CANCEL: { // release slot linked with the task id
+                for (auto & slot : slots)
+                {
+                    if (slot.task_id == task.target_id)
+                    {
+                        slot.release();
+                        break;
+                    }
+                }
+            } break;
+            case TASK_TYPE_NEXT_RESPONSE: {
+                // do nothing
+            } break;
+        }
+    }
+
+    void on_finish_multitask(task_multi& multitask)
+    {
+        // all subtasks done == multitask is done
+        task_result result;
+        result.id = multitask.id;
+        result.stop = true;
+        result.error = false;
+
+        // collect json results into one json result
+        std::vector<json> result_jsons;
+        for (auto& subres : multitask.results)
+        {
+            result_jsons.push_back(subres.result_json);
+            result.error = result.error && subres.error;
+        }
+        result.result_json = json{ { "results", result_jsons } };
+        queue_results.send(result);
+    }
+
+    bool update_slots() {
+        if (system_need_update)
+        {
+            LOG_INFO("updating system prompt", {});
+            update_system_prompt();
+        }
+
+        common_batch_clear(batch);
+
+        if (all_slots_are_idle)
+        {
+            if (system_prompt.empty() && clean_kv_cache)
+            {
+                LOG_INFO("all slots are idle and system prompt is empty, clear the KV cache", {});
+                kv_cache_clear();
+            }
+            return true;
+        }
+
+        LOG_VERBOSE("posting NEXT_RESPONSE", {});
+        task_server task;
+        task.type = TASK_TYPE_NEXT_RESPONSE;
+        task.target_id = -1;
+        queue_tasks.post(task);
+
+        for (llama_client_slot &slot : slots)
+        {
+            if (slot.ga_n == 1)
+            {
+                if (slot.is_processing() && system_tokens.size() + slot.cache_tokens.size() >= (size_t) slot.n_ctx)
+                {
+                    // START LOCALAI changes
+                    // Temporary disable context-shifting as it can lead to infinite loops (issue: https://github.com/ggerganov/llama.cpp/issues/3969)
+                    // See: https://github.com/mudler/LocalAI/issues/1333
+                    // Context is exhausted, release the slot
+                    slot.release();
+                    send_final_response(slot);
+                    slot.cache_tokens.clear();
+                    slot.n_past = 0;
+                    slot.truncated = false;
+                    slot.has_next_token = true;
+                    LOG("Context exhausted. Slot %d released (%d tokens in cache)\n", slot.id, (int) slot.cache_tokens.size());
+
+                    continue;
+                    // END LOCALAI changes
+                }
+            }
+        }
+
+        // decode any currently ongoing sequences
+        LOG_VERBOSE("decoding ongoing sequences", {});
+        for (auto & slot : slots)
+        {
+            // release the slot
+            if (slot.command == RELEASE)
+            {
+                slot.state = IDLE;
+                slot.command = NONE;
+                slot.t_last_used = ggml_time_us();
+
+                LOG_INFO("slot released", {
+                    {"slot_id",         slot.id},
+                    {"task_id",         slot.task_id},
+                    {"n_ctx",           n_ctx},
+                    {"n_past",          slot.n_past},
+                    {"n_system_tokens", system_tokens.size()},
+                    {"n_cache_tokens",  slot.cache_tokens.size()},
+                    {"truncated",       slot.truncated}
+                });
+                queue_tasks.notify_slot_changed();
+
+                continue;
+            }
+
+            if (slot.state == IDLE)
+            {
+                continue;
+            }
+
+            slot.i_batch = batch.n_tokens;
+
+            const int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
+
+            // TODO: we always have to take into account the "system_tokens"
+            //       this is not great and needs to be improved somehow
+            common_batch_add(batch, slot.sampled, system_tokens.size() + slot_npast, { slot.id }, true);
+            slot.n_past += 1;
+        }
+
+        // process in chunks of params.n_batch
+        int32_t n_batch = params.n_batch;
+
+        // assign workload to the slots
+        if (params.cont_batching || batch.n_tokens == 0)
+        {
+            for (auto & slot : slots)
+            {
+                const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
+
+                // empty prompt passed -> release the slot and send empty response
+                // note: infill mode allows empty prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
+                {
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    continue;
+                }
+
+                // need process the prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT)
+                {
+                    slot.state = PROCESSING;
+                    slot.command = NONE;
+                    std::vector<llama_token> prompt_tokens;
+                    slot.t_start_process_prompt = ggml_time_us();
+                    slot.t_start_genereration = 0;
+
+                    if (slot.infill)
+                    {
+                        bool suff_rm_leading_spc = true;
+                        if (params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1)
+                        {
+                            params.input_suffix.erase(0, 1);
+                            suff_rm_leading_spc = false;
+                        }
+                        auto prefix_tokens = tokenize(slot.params.input_prefix, false);
+                        auto suffix_tokens = tokenize(slot.params.input_suffix, false);
+
+                        const int space_token = 29871; // TODO: this should not be hardcoded
+                        if (suff_rm_leading_spc && !suffix_tokens.empty() && suffix_tokens[0] == space_token) {
+                            suffix_tokens.erase(suffix_tokens.begin());
+                        }
+
+                        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(model));
+                        prompt_tokens = prefix_tokens;
+                    }
+                    else
+                    {
+                        prompt_tokens = tokenize(slot.prompt, system_prompt.empty() && add_bos_token);  // add BOS if there isn't system prompt
+                    }
+
+                    slot.num_prompt_tokens = prompt_tokens.size();
+
+                    if (slot.params.n_keep < 0)
+                    {
+                        slot.params.n_keep = slot.num_prompt_tokens;
+                    }
+                    slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
+
+                    // if input prompt is too big, truncate it
+                    if (slot.num_prompt_tokens >= slot.n_ctx)
+                    {
+                        const int n_left = slot.n_ctx - slot.params.n_keep;
+                        const int n_block_size = n_left / 2;
+                        const int erased_blocks = (slot.num_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
+
+                        std::vector<llama_token> new_tokens(prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep);
+                        new_tokens.insert(new_tokens.end(), prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, prompt_tokens.end());
+
+                        LOG_VERBOSE("input truncated", {
+                            {"n_ctx",  slot.n_ctx},
+                            {"n_keep", slot.params.n_keep},
+                            {"n_left", n_left},
+                            {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())},
+                        });
+                        slot.truncated = true;
+                        prompt_tokens = new_tokens;
+
+                        slot.num_prompt_tokens = prompt_tokens.size();
+                        GGML_ASSERT(slot.num_prompt_tokens < slot.n_ctx);
+                    }
+
+                    if (!slot.params.cache_prompt)
+                    {
+                        common_sampler_reset(slot.ctx_sampling);
+
+                        slot.n_past = 0;
+                        slot.n_past_se = 0;
+                        slot.ga_i = 0;
+                        slot.num_prompt_tokens_processed = slot.num_prompt_tokens;
+                    }
+                    else
+                    {
+                        // push the prompt into the sampling context (do not apply grammar)
+                        for (auto &token : prompt_tokens)
+                        {
+                            common_sampler_accept(slot.ctx_sampling, token, false);
+                        }
+
+                        slot.n_past = common_part(slot.cache_tokens, prompt_tokens);
+
+                        // the last token of the cache is not in the KV cache until the next call to llama_decode
+                        // (it was sampled, pushed into the "cache_tokens", but not yet put in the context)
+                        if (slot.n_past > 0 && slot.n_past == (int32_t) slot.cache_tokens.size())
+                        {
+                            slot.n_past -= 1;
+                        }
+
+                        slot.num_prompt_tokens_processed = slot.num_prompt_tokens - slot.n_past;
+
+                        if (slot.ga_n != 1)
+                        {
+                            int ga_i = 0;
+                            int32_t ga_n = slot.ga_n;
+                            int32_t ga_w = slot.ga_w;
+                            int32_t slot_npast = 0;
+                            for (int k = 0; k < slot.n_past; ++k)
+                            {
+                                while (slot_npast >= ga_i + ga_w) {
+                                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                                    slot_npast -= bd;
+                                    ga_i += ga_w/ga_n;
+                                }
+                                slot_npast++;
+                            }
+                            slot.n_past_se = slot_npast;
+                            slot.ga_i = ga_i;
+                        }
+
+                        LOG_INFO("slot progression", {
+                            { "slot_id", slot.id },
+                            { "task_id", slot.task_id },
+                            { "n_past",  slot.n_past },
+                            { "num_prompt_tokens_processed", slot.num_prompt_tokens_processed }
+                        });
+                    }
+
+                    slot.cache_tokens = prompt_tokens;
+
+                    if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
+                    {
+                        // we have to evaluate at least 1 token to generate logits.
+                        LOG_INFO("we have to evaluate at least 1 token to generate logits", {
+                            { "slot_id", slot.id },
+                            { "task_id", slot.task_id }
+                        });
+                        slot.n_past--;
+                        if (slot.ga_i > 0)
+                        {
+                            slot.n_past_se--;
+                        }
+                    }
+
+                    int p0 = (int) system_tokens.size() + slot.n_past;
+                    LOG_INFO("kv cache rm [p0, end)", {
+                        { "slot_id", slot.id },
+                        { "task_id", slot.task_id },
+                        { "p0",      p0 }
+                    });
+                    llama_kv_cache_seq_rm(ctx, slot.id, p0, -1);
+
+                    LOG_VERBOSE("prompt ingested", {
+                                                    {"n_past",  slot.n_past},
+                                                    {"cached",  tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)},
+                                                    {"to_eval", tokens_to_str(ctx, slot.cache_tokens.cbegin() + slot.n_past, slot.cache_tokens.cend())},
+                                                });
+
+                    const bool has_images = process_images(slot);
+
+                    // process the prefix of first image
+                    std::vector<llama_token> prefix_tokens = has_images ? tokenize(slot.images[0].prefix_prompt, add_bos_token) : prompt_tokens;
+
+                    int32_t slot_npast = slot.n_past_se > 0 ? slot.n_past_se : slot.n_past;
+
+                    int32_t ga_i = slot.ga_i;
+                    int32_t ga_n = slot.ga_n;
+                    int32_t ga_w = slot.ga_w;
+
+                    for (; slot.n_past < (int) prefix_tokens.size(); ++slot.n_past)
+                    {
+                        if (slot.ga_n != 1)
+                        {
+                            while (slot_npast >= ga_i + ga_w) {
+                                const int bd = (ga_w/ga_n)*(ga_n - 1);
+                                slot_npast -= bd;
+                                ga_i += ga_w/ga_n;
+                            }
+                        }
+                        common_batch_add(batch, prefix_tokens[slot.n_past], system_tokens.size() + slot_npast, {slot.id }, false);
+                        slot_npast++;
+                    }
+
+                    if (has_images && !ingest_images(slot, n_batch))
+                    {
+                        LOG_ERR("%s: failed processing images Slot id : %d, Task id: %d", 
+                            __func__,
+                            slot.id,
+                            slot.task_id
+                        );
+                        // FIXME @phymbert: to be properly tested
+                        //  early returning without changing the slot state will block the slot for ever
+                        // no one at the moment is checking the return value
+                        return false;
+                    }
+
+                    // extract the logits only for the last token
+                    if (batch.n_tokens > 0)
+                    {
+                        batch.logits[batch.n_tokens - 1] = true;
+                    }
+
+                    slot.n_decoded = 0;
+                    slot.i_batch   = batch.n_tokens - 1;
+                }
+            }
+        }
+
+        if (batch.n_tokens == 0)
+        {
+            all_slots_are_idle = true;
+            return true;
+        }
+
+        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch)
+        {
+            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+
+            for (auto & slot : slots)
+            {
+                if (slot.ga_n != 1)
+                {
+                    // context extension via Self-Extend
+                    while (slot.n_past_se >= slot.ga_i + slot.ga_w)
+                    {
+                        const int ib = (slot.ga_n * slot.ga_i) / slot.ga_w;
+                        const int bd = (slot.ga_w / slot.ga_n) * (slot.ga_n - 1);
+                        const int dd = (slot.ga_w / slot.ga_n) - ib * bd - slot.ga_w;
+
+                        LOG("\n");
+                        LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i, slot.n_past_se, ib * bd, slot.ga_i + ib * bd, slot.n_past_se + ib * bd);
+                        LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w, slot.ga_n, (slot.ga_i + ib * bd) / slot.ga_n, (slot.ga_i + ib * bd + slot.ga_w) / slot.ga_n);
+                        LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", slot.ga_i + ib * bd + slot.ga_w, slot.n_past_se + ib * bd, dd, slot.ga_i + ib * bd + slot.ga_w + dd, slot.n_past_se + ib * bd + dd);
+
+                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i, slot.n_past_se, ib * bd);
+                        llama_kv_cache_seq_div(ctx, slot.id, slot.ga_i + ib * bd, slot.ga_i + ib * bd + slot.ga_w,slot.ga_n);
+                        llama_kv_cache_seq_add(ctx, slot.id, slot.ga_i + ib * bd + slot.ga_w,slot.n_past_se + ib * bd, dd);
+
+                        slot.n_past_se -= bd;
+
+                        slot.ga_i += slot.ga_w / slot.ga_n;
+
+                        LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", slot.n_past_se + bd, slot.n_past_se, slot.ga_i);
+                    }
+                    slot.n_past_se += n_tokens;
+                }
+            }
+
+            llama_batch batch_view =
+            {
+                n_tokens,
+                batch.token    + i,
+                nullptr,
+                batch.pos      + i,
+                batch.n_seq_id + i,
+                batch.seq_id   + i,
+                batch.logits   + i,
+            };
+
+            const int ret = llama_decode(ctx, batch_view);
+
+            if (ret != 0)
+            {
+                if (n_batch == 1 || ret < 0)
+                {
+                    // if you get here, it means the KV cache is full - try increasing it via the context size
+                    LOG("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret);
+                    return false;
+                }
+
+                LOG("%s : failed to find free space in the KV cache, retrying with smaller n_batch = %d\n", __func__, n_batch / 2);
+
+                // retry with half the batch size to try to find a free slot in the KV cache
+                n_batch /= 2;
+                i -= n_batch;
+                continue;
+            }
+
+            for (auto & slot : slots)
+            {
+                if (slot.i_batch < (int) i || slot.i_batch >= (int) (i + n_tokens))
+                {
+                    continue;
+                }
+
+                // prompt evaluated for embedding
+                if (slot.embedding)
+                {
+                    send_embedding(slot);
+                    slot.release();
+                    slot.i_batch = -1;
+                    continue;
+                }
+
+                completion_token_output result;
+                const llama_token id = common_sampler_sample(slot.ctx_sampling, ctx, slot.i_batch - i);
+
+                common_sampler_accept(slot.ctx_sampling, id, true);
+
+                slot.n_decoded += 1;
+                if (slot.n_decoded == 1)
+                {
+                    slot.t_start_genereration = ggml_time_us();
+                    slot.t_prompt_processing = (slot.t_start_genereration - slot.t_start_process_prompt) / 1e3;
+                    metrics.on_prompt_eval(slot);
+                }
+
+                result.tok = id;
+                const auto * cur_p = common_sampler_get_candidates(slot.ctx_sampling);
+
+                for (size_t i = 0; i < (size_t) slot.sparams.n_probs; ++i) {
+                    result.probs.push_back({
+                        cur_p->data[i].id,
+                        i >= cur_p->size ? 0.0f : cur_p->data[i].p,
+                    });
+                }
+
+                if (!process_token(result, slot))
+                {
+                    slot.release();
+                    slot.print_timings();
+                    send_final_response(slot);
+                    metrics.on_prediction(slot);
+                }
+
+                slot.i_batch = -1;
+            }
+        }
+
+        LOG_VERBOSE("slots updated", {});
+        return true;
+    }
+
+    void run_on_all_tasks_finished() {
+        update_slots();
+    }
+};
+
+/* llama.cpp completion api semantics */
+static json format_partial_response(
+    llama_server_context &llama, llama_client_slot *slot, const std::string &content, const std::vector<completion_token_output> &probs
+) {
+    json res = json
+    {
+        {"content",    content },
+        {"stop",       false},
+        {"slot_id",    slot->id },
+        {"multimodal", llama.multimodal }
+    };
+
+    if (slot->sparams.n_probs > 0)
+    {
+        res["completion_probabilities"] = probs_vector_to_json(llama.ctx, probs);
+    }
+
+    return res;
+}
+
+struct token_translator
+{
+    llama_context * ctx;
+    std::string operator()(llama_token tok)                    const { return common_token_to_piece(ctx, tok); }
+    std::string operator()(const completion_token_output &cto) const { return (*this)(cto.tok); }
+};
+
+static void append_to_generated_text_from_generated_token_probs(llama_server_context &llama, llama_client_slot *slot)
+{
+    auto & gtps = slot->generated_token_probs;
+    auto translator = token_translator{llama.ctx};
+    auto add_strlen = [=](size_t sum, const completion_token_output & cto) { return sum + translator(cto).size(); };
+    const size_t len = std::accumulate(gtps.begin(), gtps.end(), size_t(0), add_strlen);
+    if (slot->generated_text.capacity() < slot->generated_text.size() + len)
+    {
+        slot->generated_text.reserve(slot->generated_text.size() + len);
+    }
+    for (const completion_token_output & cto : gtps)
+    {
+        slot->generated_text += translator(cto);
+    }
+}
+
+std::function<void(int)> shutdown_handler;
+inline void signal_handler(int signal) { shutdown_handler(signal); }
+
+/////////////////////////////////
+////////////////////////////////
+//////// LOCALAI code starts below here
+/////////////////////////////////
+////////////////////////////////
+
+bool loaded_model; // TODO: add a mutex for this, but happens only once loading the model
+
+// The class has a llama instance that is shared across all RPCs
+llama_server_context llama;
+
+static void start_llama_server() {
+    // Wait for model to be loaded first
+    while (!loaded_model) {
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+    }
+
+    llama.queue_tasks.on_new_task(std::bind(
+        &llama_server_context::process_single_task, &llama, std::placeholders::_1));
+    llama.queue_tasks.on_finish_multitask(std::bind(
+        &llama_server_context::on_finish_multitask, &llama, std::placeholders::_1));
+    llama.queue_tasks.on_all_tasks_finished(std::bind(
+        &llama_server_context::run_on_all_tasks_finished, &llama));
+    llama.queue_results.on_multitask_update(std::bind(
+        &llama_server_queue::update_multitask,
+        &llama.queue_tasks,
+        std::placeholders::_1,
+        std::placeholders::_2,
+        std::placeholders::_3
+    ));
+    llama.queue_tasks.start_loop();
+}
+
+json parse_options(bool streaming, const backend::PredictOptions* predict, llama_server_context &llama)
+{
+    
+    // This is for example a slot data from the json data
+    //     slot->params.stream           = json_value(data, "stream",            false);
+    //     slot->params.cache_prompt     = json_value(data, "cache_prompt",      false);
+    //     slot->params.n_predict        = json_value(data, "n_predict",         default_params.n_predict);
+    //     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.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);
+    //     slot->sparams.penalty_repeat  = json_value(data, "repeat_penalty",    default_sparams.penalty_repeat);
+    //     slot->sparams.penalty_freq    = json_value(data, "frequency_penalty", default_sparams.penalty_freq);
+    //     slot->sparams.penalty_present = json_value(data, "presence_penalty",  default_sparams.penalty_present);
+    //     slot->sparams.mirostat        = json_value(data, "mirostat",          default_sparams.mirostat);
+    //     slot->sparams.mirostat_tau    = json_value(data, "mirostat_tau",      default_sparams.mirostat_tau);
+    //     slot->sparams.mirostat_eta    = json_value(data, "mirostat_eta",      default_sparams.mirostat_eta);
+    //     slot->sparams.penalize_nl     = json_value(data, "penalize_nl",       default_sparams.penalize_nl);
+    //     slot->params.n_keep           = json_value(data, "n_keep",            slot->params.n_keep);
+    //     slot->params.seed             = json_value(data, "seed",              default_params.seed);
+    //     slot->sparams.grammar         = json_value(data, "grammar",           default_sparams.grammar);
+    //     slot->sparams.n_probs         = json_value(data, "n_probs",           default_sparams.n_probs);
+
+    // Create now a json data from the prediction options instead
+    //
+    json data;
+    data["stream"] = streaming;
+    data["cache_prompt"] = predict->promptcacheall();
+    data["n_predict"] = predict->tokens() == 0 ? -1 : predict->tokens();
+    data["top_k"] = predict->topk();
+    data["top_p"] = predict->topp();
+    data["typical_p"] = predict->typicalp();
+    data["temperature"] = predict->temperature();
+    data["repeat_last_n"] = predict->repeat();
+    data["repeat_penalty"] = predict->penalty();
+    data["frequency_penalty"] = predict->frequencypenalty();
+    data["presence_penalty"] = predict->presencepenalty();
+    data["mirostat"] = predict->mirostat();
+    data["mirostat_tau"] = predict->mirostattau();
+    data["mirostat_eta"] = predict->mirostateta();
+    data["penalize_nl"] = predict->penalizenl();
+    data["n_keep"] = predict->nkeep();
+    data["seed"] = predict->seed();
+    data["grammar"] = predict->grammar();
+    data["prompt"] = predict->prompt();
+    data["ignore_eos"] = predict->ignoreeos();
+    data["embeddings"] = predict->embeddings();
+
+    // Add the correlationid to json data
+    data["correlation_id"] = predict->correlationid();
+
+    // for each image in the request, add the image data
+    //
+    for (int i = 0; i < predict->images_size(); i++) {
+        data["image_data"].push_back(json
+            {
+                {"id", i},
+                {"data",    predict->images(i)},
+            });
+    }
+
+    data["stop"] = predict->stopprompts();
+    // data["n_probs"] = predict->nprobs();
+    //TODO: images,
+
+    return data;
+}
+
+// static void parse_options_completion(bool streaming,const backend::PredictOptions* predict, llama_server_context &llama)
+// {
+//     // https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L673
+//     gpt_params default_params;
+
+//     llama.stream = streaming;
+//     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.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.sparams.grammar = predict->grammar();
+//     // llama.params.n_probs = predict->
+//     llama.params.prompt = predict->prompt();
+
+//     llama.params.sparams.logit_bias.clear();
+
+//     if (predict->ignoreeos())
+//     {
+//         llama.params.sparams.logit_bias[llama_token_eos(llama.model)] = -INFINITY;
+//     }
+
+//     // const auto &logit_bias = body.find("logit_bias");
+//     // if (logit_bias != body.end() && logit_bias->is_array())
+//     // {
+//     //     const int n_vocab = llama_n_vocab(llama.model);
+//     //     for (const auto &el : *logit_bias)
+//     //     {
+//     //         if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
+//     //         {
+//     //             llama_token tok = el[0].get<llama_token>();
+//     //             if (tok >= 0 && tok < n_vocab)
+//     //             {
+//     //                 if (el[1].is_number())
+//     //                 {
+//     //                     llama.params.logit_bias[tok] = el[1].get<float>();
+//     //                 }
+//     //                 else if (el[1].is_boolean() && !el[1].get<bool>())
+//     //                 {
+//     //                     llama.params.logit_bias[tok] = -INFINITY;
+//     //                 }
+//     //             }
+//     //         }
+//     //     }
+//     // }
+
+//     llama.params.antiprompt.clear();
+//     for (const std::string& stopPrompt : predict->stopprompts()) {
+//     if (!stopPrompt.empty())
+//             {
+//                 llama.params.antiprompt.push_back(stopPrompt);
+//             }
+//     }
+// }
+
+static void params_parse(const backend::ModelOptions* request,
+                                common_params & params) {
+   
+    // this is comparable to: https://github.com/ggerganov/llama.cpp/blob/d9b33fe95bd257b36c84ee5769cc048230067d6f/examples/server/server.cpp#L1809
+
+    params.model = request->modelfile();
+    if (!request->mmproj().empty()) {
+    // get the directory of modelfile
+      std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
+      params.mmproj = model_dir + "/"+ request->mmproj();
+    }
+    //  params.model_alias ??
+    params.model_alias =  request->modelfile();
+    params.n_ctx = request->contextsize();
+    //params.memory_f16 = request->f16memory();
+    params.cpuparams.n_threads = request->threads();
+    params.n_gpu_layers = request->ngpulayers();
+    params.n_batch = request->nbatch();
+    // Set params.n_parallel by environment variable (LLAMA_PARALLEL), defaults to 1
+    //params.n_parallel = 1;
+    const char *env_parallel = std::getenv("LLAMACPP_PARALLEL");
+    if (env_parallel != NULL) {
+        params.n_parallel = std::stoi(env_parallel);
+        params.cont_batching = true;
+    } else {
+        params.n_parallel = 1;
+    }
+
+    const char *llama_grpc_servers = std::getenv("LLAMACPP_GRPC_SERVERS");
+    if (llama_grpc_servers != NULL) {
+        params.rpc_servers = std::string(llama_grpc_servers);
+    }
+    
+    // TODO: Add yarn
+
+    if (!request->tensorsplit().empty()) {
+        std::string arg_next = request->tensorsplit();
+
+        // split string by , and /
+        const std::regex regex{ R"([,/]+)" };
+        std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+        std::vector<std::string> split_arg{ it, {} };
+
+        GGML_ASSERT(split_arg.size() <= llama_max_devices());
+
+        for (size_t i_device = 0; i_device < llama_max_devices(); ++i_device) {
+            if (i_device < split_arg.size()) {
+                params.tensor_split[i_device] = std::stof(split_arg[i_device]);
+            }
+            else {
+                params.tensor_split[i_device] = 0.0f;
+            }
+        }
+    }
+
+    if (!request->maingpu().empty()) {
+        params.main_gpu = std::stoi(request->maingpu());
+    }
+    if (!request->loraadapter().empty() && !request->lorabase().empty()) {
+     float scale_factor = 1.0f;
+     if (request->lorascale() != 0.0f) {
+        scale_factor = request->lorascale();
+     }
+     // get the directory of modelfile
+     std::string model_dir = params.model.substr(0, params.model.find_last_of("/\\"));
+     params.lora_adapters.push_back({ model_dir + "/"+request->loraadapter(), scale_factor });
+    }
+    params.use_mlock = request->mlock();
+    params.use_mmap = request->mmap();
+    params.flash_attn = request->flashattention();
+    params.no_kv_offload = request->nokvoffload();
+
+    params.embedding = request->embeddings();
+
+    if (request->ropescaling() == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
+    else if (request->ropescaling() == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+    else { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
+    if ( request->yarnextfactor() != 0.0f ) {
+        params.yarn_ext_factor = request->yarnextfactor();
+    }
+    if ( request->yarnattnfactor() != 0.0f ) {
+        params.yarn_attn_factor = request->yarnattnfactor();
+    }
+    if ( request->yarnbetafast() != 0.0f ) {
+        params.yarn_beta_fast = request->yarnbetafast();
+    }
+    if ( request->yarnbetaslow() != 0.0f ) {
+        params.yarn_beta_slow = request->yarnbetaslow();
+    }
+    if ( request->ropefreqbase() != 0.0f ) {
+        params.rope_freq_base = request->ropefreqbase();
+    }
+    if ( request->ropefreqscale() != 0.0f ) {
+        params.rope_freq_scale = request->ropefreqscale();
+    }
+}
+
+
+// GRPC Server start
+class BackendServiceImpl final : public backend::Backend::Service {
+public:
+  grpc::Status Health(ServerContext* context, const backend::HealthMessage* request, backend::Reply* reply) {
+    // Implement Health RPC
+    reply->set_message("OK");
+    return Status::OK;
+  }
+
+  grpc::Status LoadModel(ServerContext* context, const backend::ModelOptions* request, backend::Result* result) {
+    // Implement LoadModel RPC
+    common_params params;
+    params_parse(request, params);
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // load the model
+    if (!llama.load_model(params))
+    {
+        result->set_message("Failed loading model");
+        result->set_success(false);
+        return Status::CANCELLED;
+    }
+    llama.initialize();
+    result->set_message("Loading succeeded");
+    result->set_success(true);
+    loaded_model = true;
+    return Status::OK;
+  }
+  grpc::Status PredictStream(grpc::ServerContext* context, const backend::PredictOptions* request, grpc::ServerWriter<backend::Reply>* writer) override {
+        json data = parse_options(true, request, llama);
+        const int task_id = llama.queue_tasks.get_new_id();
+        llama.queue_results.add_waiting_task_id(task_id);
+        llama.request_completion(task_id, data, false, false, -1);
+        while (true)
+        {
+            task_result result = llama.queue_results.recv(task_id);
+            if (!result.error) {
+                const std::string str =
+                "data: " +
+                result.result_json.dump(-1, ' ', false, json::error_handler_t::replace) +
+                "\n\n";
+                LOG_VERBOSE("data stream", {
+                    { "to_send", str }
+                });
+
+                backend::Reply reply;
+                // print it
+                std::string completion_text = result.result_json.value("content", "");
+
+                reply.set_message(completion_text);
+                int32_t tokens_predicted = result.result_json.value("tokens_predicted", 0);
+                reply.set_tokens(tokens_predicted);
+                int32_t tokens_evaluated = result.result_json.value("tokens_evaluated", 0);
+                reply.set_prompt_tokens(tokens_evaluated);
+
+                // Log Request Correlation Id
+                LOG_VERBOSE("correlation:", {
+                    { "id", data["correlation_id"] }
+                });
+
+                // Send the reply
+                writer->Write(reply);
+
+                if (result.stop) {
+                    break;
+                }
+            } else {
+                break;
+            }
+        }
+
+        return grpc::Status::OK;
+    }
+
+
+    grpc::Status Predict(ServerContext* context, const backend::PredictOptions* request, backend::Reply* reply) {
+        json data = parse_options(false, request, llama);
+        const int task_id = llama.queue_tasks.get_new_id();
+        llama.queue_results.add_waiting_task_id(task_id);
+        llama.request_completion(task_id, data, false, false, -1);
+        std::string completion_text;
+        task_result result = llama.queue_results.recv(task_id);
+        if (!result.error && result.stop) {
+            
+            // Log Request Correlation Id
+            LOG_VERBOSE("correlation:", {
+                { "id", data["correlation_id"] }
+            });
+
+            completion_text = result.result_json.value("content", "");
+            int32_t tokens_predicted = result.result_json.value("tokens_predicted", 0);
+            int32_t tokens_evaluated = result.result_json.value("tokens_evaluated", 0);
+            reply->set_prompt_tokens(tokens_evaluated);
+            reply->set_tokens(tokens_predicted);
+            reply->set_message(completion_text);
+        }
+        else
+        {
+            return grpc::Status::OK;
+        }
+
+        return grpc::Status::OK;
+    }
+
+    /// https://github.com/ggerganov/llama.cpp/blob/aa2341298924ac89778252015efcb792f2df1e20/examples/server/server.cpp#L2969
+    grpc::Status Embedding(ServerContext* context, const backend::PredictOptions* request, backend::EmbeddingResult* embeddingResult) {
+        json data = parse_options(false, request, llama);
+        const int task_id = llama.queue_tasks.get_new_id();
+        llama.queue_results.add_waiting_task_id(task_id);
+        llama.request_completion(task_id, { {"prompt", data["embeddings"]}, { "n_predict", 0}, {"image_data", ""} }, false, true, -1);
+        // get the result
+        task_result result = llama.queue_results.recv(task_id);
+        //std::cout << "Embedding result JSON" << result.result_json.dump() << std::endl;
+        llama.queue_results.remove_waiting_task_id(task_id);
+        if (!result.error && result.stop) {
+            std::vector<float> embeddings = result.result_json.value("embedding", std::vector<float>());
+            // loop the vector and set the embeddings results
+            for (int i = 0; i < embeddings.size(); i++) {
+                embeddingResult->add_embeddings(embeddings[i]);
+            }
+        }
+        else
+        {
+            return grpc::Status::OK;
+        }
+
+        return grpc::Status::OK;
+    }
+
+    grpc::Status GetMetrics(ServerContext* context, const backend::MetricsRequest* request, backend::MetricsResponse* response) {
+        llama_client_slot* active_slot = llama.get_active_slot();
+
+        if (active_slot != nullptr) {
+            // Calculate the tokens per second using existing logic
+            double tokens_per_second = 1e3 / active_slot->t_token_generation * active_slot->n_decoded;
+
+            // Populate the response with metrics
+            response->set_slot_id(active_slot->id);
+            response->set_prompt_json_for_slot(active_slot->prompt.dump());
+            response->set_tokens_per_second(tokens_per_second);
+            response->set_tokens_generated(active_slot->n_decoded);
+            response->set_prompt_tokens_processed(active_slot->num_prompt_tokens_processed);
+        } else {
+            // Handle case when no active slot exists
+            response->set_slot_id(0);
+            response->set_prompt_json_for_slot("");
+            response->set_tokens_per_second(0);
+            response->set_tokens_generated(0);
+            response->set_prompt_tokens_processed(0);
+        }
+
+        return grpc::Status::OK;
+    } 
+};
+
+void RunServer(const std::string& server_address) {
+  BackendServiceImpl service;
+
+  ServerBuilder builder;
+  builder.AddListeningPort(server_address, grpc::InsecureServerCredentials());
+  builder.RegisterService(&service);
+
+  std::unique_ptr<Server> server(builder.BuildAndStart());
+  std::cout << "Server listening on " << server_address << std::endl;
+  server->Wait();
+}
+
+int main(int argc, char** argv) {
+  std::string server_address("localhost:50051");
+
+  // Define long and short options
+  struct option long_options[] = {
+      {"addr", required_argument, nullptr, 'a'},
+      {nullptr, 0, nullptr, 0}
+  };
+
+  // Parse command-line arguments
+  int option;
+  int option_index = 0;
+  while ((option = getopt_long(argc, argv, "a:", long_options, &option_index)) != -1) {
+    switch (option) {
+      case 'a':
+        server_address = optarg;
+        break;
+      default:
+        std::cerr << "Usage: " << argv[0] << " [--addr=<address>] or [-a <address>]" << std::endl;
+        return 1;
+    }
+  }
+
+   // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                RunServer(server_address);
+                return 0;
+            });
+
+
+    //);
+    start_llama_server();
+                        std::cout << "stopping" << std::endl;
+
+    t.join();
+
+    llama_backend_free();
+  return 0;
+}

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

@@ -0,0 +1,13 @@
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index 342042ff..224db9b5 100644
+--- a/examples/llava/clip.cpp
++++ b/examples/llava/clip.cpp
+@@ -2419,7 +2419,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
+             struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+             int* patches_data = (int*)malloc(ggml_nbytes(patches));
+             for (int i = 0; i < num_patches; i++) {
+-                patches_data[i] = i + 1;
++                patches_data[i] = i;
+             }
+             ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+             free(patches_data);

backend/cpp/llama/prepare.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+
+## Patches
+## Apply patches from the `patches` directory
+for patch in $(ls patches); do
+    echo "Applying patch $patch"
+    patch -d llama.cpp/ -p1 < patches/$patch
+done 
+
+cp -r CMakeLists.txt llama.cpp/examples/grpc-server/
+cp -r grpc-server.cpp llama.cpp/examples/grpc-server/
+cp -rfv json.hpp llama.cpp/examples/grpc-server/
+cp -rfv utils.hpp llama.cpp/examples/grpc-server/
+    
+if grep -q "grpc-server" llama.cpp/examples/CMakeLists.txt; then
+    echo "grpc-server already added"
+else
+    echo "add_subdirectory(grpc-server)" >> llama.cpp/examples/CMakeLists.txt
+fi
+
+## XXX: In some versions of CMake clip wasn't being built before llama.
+## This is an hack for now, but it should be fixed in the future.
+cp -rfv llama.cpp/examples/llava/clip.h llama.cpp/examples/grpc-server/clip.h
+cp -rfv llama.cpp/examples/llava/llava.cpp llama.cpp/examples/grpc-server/llava.cpp
+echo '#include "llama.h"' > llama.cpp/examples/grpc-server/llava.h
+cat llama.cpp/examples/llava/llava.h >> llama.cpp/examples/grpc-server/llava.h
+cp -rfv llama.cpp/examples/llava/clip.cpp llama.cpp/examples/grpc-server/clip.cpp

backend/cpp/llama/utils.hpp

@@ -0,0 +1,483 @@
+// https://github.com/ggerganov/llama.cpp/blob/master/examples/server/utils.hpp
+
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+
+#include "../llava/clip.h"
+
+using json = nlohmann::json;
+
+extern bool server_verbose;
+
+#ifndef SERVER_VERBOSE
+#define SERVER_VERBOSE 1
+#endif
+
+#if SERVER_VERBOSE != 1
+#define LOG_VERBOSE(MSG, ...)
+#else
+#define LOG_VERBOSE(MSG, ...)                                            \
+    do                                                                   \
+    {                                                                    \
+        if (server_verbose)                                              \
+        {                                                                \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+        }                                                                \
+    } while (0)
+#endif
+
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+
+//
+// parallel
+//
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// TODO: can become bool if we can't find use of more states
+enum slot_state
+{
+    IDLE,
+    PROCESSING,
+};
+
+enum slot_command
+{
+    NONE,
+    LOAD_PROMPT,
+    RELEASE,
+};
+
+struct slot_params
+{
+    bool stream       = true;
+    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt
+
+    uint32_t seed      = -1; // RNG seed
+    int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
+    int32_t  n_predict = -1; // new tokens to predict
+
+    std::vector<std::string> antiprompt;
+
+    json input_prefix;
+    json input_suffix;
+};
+
+struct slot_image
+{
+    int32_t id;
+
+    bool request_encode_image = false;
+    float * image_embedding = nullptr;
+    int32_t image_tokens = 0;
+
+    clip_image_u8 * img_data;
+
+    std::string prefix_prompt; // before of this image
+};
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
+{
+    nlohmann::ordered_json log
+    {
+        {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
+    };
+
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+//
+// server utils
+//
+
+template <typename T>
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
+    // Fallback null to default value
+    return body.contains(key) && !body.at(key).is_null()
+        ? body.value(key, default_value)
+        : default_value;
+}
+
+inline std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    return chatml_msgs.str();
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_all_task_finished;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        return id++;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when the batch of tasks is finished
+    void on_all_tasks_finished(std::function<void(void)> callback) {
+        callback_all_task_finished = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // Start the main loop. This call is blocking
+    [[noreturn]]
+    void start_loop() {
+        while (true) {
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("callback_all_task_finished", {});
+                // process and update all the multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is finished
+                callback_all_task_finished();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    condition_tasks.wait(lock, [&]{
+                        return !queue_tasks.empty();
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    void add_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    void remove_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+            LOG_VERBOSE("condition_results unblock", {});
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
+        }
+
+        // should never reach here
+    }
+
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
+    }
+
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
+
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {});
+                queue_results.push_back(result);
+                condition_results.notify_one();
+                return;
+            }
+        }
+    }
+};
+
+//
+// base64 utils (TODO: move to common in the future)
+//
+
+static const std::string base64_chars =
+             "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+             "abcdefghijklmnopqrstuvwxyz"
+             "0123456789+/";
+
+static inline bool is_base64(uint8_t c)
+{
+    return (isalnum(c) || (c == '+') || (c == '/'));
+}
+
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
+    int i = 0;
+    int j = 0;
+    int in_ = 0;
+
+    int in_len = encoded_string.size();
+
+    uint8_t char_array_4[4];
+    uint8_t char_array_3[3];
+
+    std::vector<uint8_t> ret;
+
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
+        char_array_4[i++] = encoded_string[in_]; in_++;
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
+                char_array_4[i] = base64_chars.find(char_array_4[i]);
+            }
+
+            char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+            char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+            char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+            for (i = 0; (i < 3); i++)
+            {
+                ret.push_back(char_array_3[i]);
+            }
+            i = 0;
+        }
+    }
+
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
+            char_array_4[j] = 0;
+        }
+
+        for (j = 0; j <4; j++)
+        {
+            char_array_4[j] = base64_chars.find(char_array_4[j]);
+        }
+
+        char_array_3[0] = ((char_array_4[0]      ) << 2) + ((char_array_4[1] & 0x30) >> 4);
+        char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
+        char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
+
+        for (j = 0; (j < i - 1); j++)
+        {
+            ret.push_back(char_array_3[j]);
+        }
+    }
+
+    return ret;
+}