hf_text-generation-inference/Dockerfile_amd

# Rust builder
FROM lukemathwalker/cargo-chef:latest-rust-1.80 AS chef
WORKDIR /usr/src

ARG CARGO_REGISTRIES_CRATES_IO_PROTOCOL=sparse

FROM chef AS planner
COPY Cargo.lock Cargo.lock
COPY Cargo.toml Cargo.toml
COPY rust-toolchain.toml rust-toolchain.toml
COPY proto proto
COPY benchmark benchmark
COPY router router
COPY backends backends
COPY launcher launcher
RUN cargo chef prepare --recipe-path recipe.json

FROM chef AS builder

RUN PROTOC_ZIP=protoc-21.12-linux-x86_64.zip && \
    curl -OL https://github.com/protocolbuffers/protobuf/releases/download/v21.12/$PROTOC_ZIP && \
    unzip -o $PROTOC_ZIP -d /usr/local bin/protoc && \
    unzip -o $PROTOC_ZIP -d /usr/local 'include/*' && \
    rm -f $PROTOC_ZIP

COPY --from=planner /usr/src/recipe.json recipe.json
RUN cargo chef cook --profile release-opt --recipe-path recipe.json

ARG GIT_SHA
ARG DOCKER_LABEL

COPY Cargo.toml Cargo.toml
COPY rust-toolchain.toml rust-toolchain.toml
COPY proto proto
COPY benchmark benchmark
COPY router router
COPY backends backends
COPY launcher launcher
RUN cargo build --profile release-opt

# Text Generation Inference base image for RoCm
FROM rocm/dev-ubuntu-22.04:6.2 AS base

RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
    build-essential \
    ca-certificates \
    ccache \
    curl \
    git \
    make \
    libmsgpack-dev \
    libssl-dev \
    llvm-dev \
    g++ \
    # Needed to build VLLM & flash.
    rocthrust-dev \
    hipsparse-dev \
    hipblas-dev \
    hipcub-dev \
    rocblas-dev \
    hiprand-dev \
    hipfft-dev \
    rocrand-dev \
    miopen-hip-dev \
    hipsolver-dev \
    rccl-dev \
    cmake \
    python3-dev \
    python3-venv && \
    rm -rf /var/lib/apt/lists/*

# Keep in sync with `server/pyproject.toml
ARG MAMBA_VERSION=23.1.0-1
ARG PYTORCH_VERSION='2.3.0'
ARG ROCM_VERSION='6.0.2'
ARG PYTHON_VERSION='3.10.10'
# Automatically set by buildx
ARG TARGETPLATFORM
ENV PATH=/opt/conda/bin:$PATH

ARG PYTORCH_ROCM_ARCH="gfx90a;gfx942"

RUN curl -fsSL -v -o cmake-3.30.2-linux-x86_64.sh https://github.com/Kitware/CMake/releases/download/v3.30.2/cmake-3.30.2-linux-x86_64.sh \
    && chmod +x cmake-3.30.2-linux-x86_64.sh \
    && ./cmake-3.30.2-linux-x86_64.sh --skip-license --prefix=/usr/local \
    && rm cmake-3.30.2-linux-x86_64.sh

RUN pip install joblib msgpack

# Install HIPBLASLt
ARG HIPBLASLT_BRANCH="6f65c6e"
RUN git clone https://github.com/ROCm/hipBLASLt \
    && cd hipBLASLt \
    && git checkout ${HIPBLASLT_BRANCH} \
    && SCCACHE_IDLE_TIMEOUT=1800 ./install.sh --architecture ${PYTORCH_ROCM_ARCH} \
    && cd build/release \
    && make package
RUN dpkg -i hipBLASLt/build/release/*.deb \
    && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
    && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status;
    # && cd .. \
    # && rm -rf hipBLASLt

# TGI seem to require libssl.so.1.1 instead of libssl.so.3 so we can't use ubuntu 22.04. Ubuntu 20.04 has python==3.8, and TGI requires python>=3.9, hence the need for miniconda.
# Install mamba
# translating Docker's TARGETPLATFORM into mamba arches
RUN case ${TARGETPLATFORM} in \
         "linux/arm64")  MAMBA_ARCH=aarch64  ;; \
         *)              MAMBA_ARCH=x86_64   ;; \
    esac && \
    curl -fsSL -v -o ~/mambaforge.sh -O  "https://github.com/conda-forge/miniforge/releases/download/${MAMBA_VERSION}/Mambaforge-${MAMBA_VERSION}-Linux-${MAMBA_ARCH}.sh"
RUN chmod +x ~/mambaforge.sh && \
    bash ~/mambaforge.sh -b -p /opt/conda && \
    mamba init && \
    rm ~/mambaforge.sh

# Install flash-attention, torch dependencies
RUN pip install numpy einops ninja --no-cache-dir

RUN conda install mkl-static mkl-include
RUN pip uninstall -y triton && \
    git clone --depth 1 --single-branch https://github.com/ROCm/triton.git && \
    cd triton/python && \
    pip install .

ARG PYTORCH_COMMIT="da320214e66b5af0f7db8fd18a64dbb519d17b27"
RUN git clone --depth 1 --recursive --single-branch --branch main https://github.com/pytorch/pytorch.git pytorch && \
    cd pytorch && git fetch --depth 1 origin ${PYTORCH_COMMIT} && \
    git checkout  ${PYTORCH_COMMIT} && \
    git submodule update --init --recursive && \
    pip install -r requirements.txt --no-cache-dir

ARG _GLIBCXX_USE_CXX11_ABI="1"
ARG CMAKE_PREFIX_PATH="/opt/conda"
ARG BUILD_CAFFE2="0" \
    BUILD_CAFFE2_OPS="0" \
    USE_CUDA="0" \
    USE_ROCM="1" \
    BUILD_TEST="0" \
    USE_FBGEMM="0" \
    USE_NNPACK="0" \
    USE_QNNPACK="0" \
    USE_XNNPACK="0" \
    USE_FLASH_ATTENTION="1" \
    USE_MEM_EFF_ATTENTION="0"

RUN cd pytorch && python tools/amd_build/build_amd.py && python setup.py install

# Set AS recommended: https://github.com/ROCm/triton/wiki/A-script-to-set-program-execution-environment-in-ROCm
ENV HIP_FORCE_DEV_KERNARG=1

# On MI250 and MI300, performances for flash with Triton FA are slightly better than CK.
# However, Triton requires a tunning for each prompt length, which is prohibitive.
ENV ROCM_USE_FLASH_ATTN_V2_TRITON=0

FROM base AS kernel-builder

# # Build vllm kernels
FROM kernel-builder AS vllm-builder
WORKDIR /usr/src

COPY server/Makefile-vllm Makefile

# Build specific version of vllm
RUN make build-vllm-rocm

# Build Flash Attention v2 kernels
FROM kernel-builder AS flash-att-v2-builder
WORKDIR /usr/src

COPY server/Makefile-flash-att-v2 Makefile

# Build specific version of flash attention v2
RUN make build-flash-attention-v2-rocm

# Build Transformers CUDA kernels (gpt-neox and bloom)
FROM kernel-builder AS custom-kernels-builder
WORKDIR /usr/src
COPY server/custom_kernels/ .
RUN python setup.py build

# Build exllama kernels
FROM kernel-builder AS exllama-kernels-builder
WORKDIR /usr/src
COPY server/exllama_kernels/ .

RUN python setup.py build

# Build exllama v2 kernels
FROM kernel-builder AS exllamav2-kernels-builder
WORKDIR /usr/src
COPY server/exllamav2_kernels/ .

RUN python setup.py build

FROM base AS base-copy

# Text Generation Inference base env
ENV HF_HOME=/data \
    HF_HUB_ENABLE_HF_TRANSFER=1 \
    PORT=80

# Copy builds artifacts from vllm builder
COPY --from=vllm-builder /usr/src/vllm/build/lib.linux-x86_64-cpython-310 /opt/conda/lib/python3.10/site-packages

# Copy build artifacts from flash attention v2 builder
COPY --from=flash-att-v2-builder /usr/src/flash-attention-v2/build/lib.linux-x86_64-cpython-310 /opt/conda/lib/python3.10/site-packages

# Copy build artifacts from custom kernels builder
COPY --from=custom-kernels-builder /usr/src/build/lib.linux-x86_64-cpython-310 /opt/conda/lib/python3.10/site-packages

# Copy build artifacts from exllama kernels builder
COPY --from=exllama-kernels-builder /usr/src/build/lib.linux-x86_64-cpython-310 /opt/conda/lib/python3.10/site-packages

# Copy build artifacts from exllamav2 kernels builder
COPY --from=exllamav2-kernels-builder /usr/src/build/lib.linux-x86_64-cpython-310 /opt/conda/lib/python3.10/site-packages

# Install server
COPY proto proto
COPY server server
COPY server/Makefile server/Makefile
RUN cd server && \
    make gen-server && \
    pip install -r requirements_rocm.txt && \
    pip install ".[accelerate, peft, outlines]" --no-cache-dir

# Install benchmarker
COPY --from=builder /usr/src/target/release-opt/text-generation-benchmark /usr/local/bin/text-generation-benchmark
# Install router
COPY --from=builder /usr/src/target/release-opt/text-generation-router /usr/local/bin/text-generation-router
# Install launcher
COPY --from=builder /usr/src/target/release-opt/text-generation-launcher /usr/local/bin/text-generation-launcher

# AWS Sagemaker compatible image
FROM base AS sagemaker

COPY sagemaker-entrypoint.sh entrypoint.sh
RUN chmod +x entrypoint.sh

ENTRYPOINT ["./entrypoint.sh"]

# Final image
FROM base-copy

ENV ROCM_USE_CUSTOM_PAGED_ATTN=1
ENV PYTORCH_TUNABLEOP_TUNING_AFTER_WARMUP=0
ENV VLLM_MOE_PADDING=0
ENV ATTENTION=paged
ENV USE_PREFIX_CACHING=0
ENV ROCM_USE_SKINNY_GEMM=1

COPY ./tgi-entrypoint.sh /tgi-entrypoint.sh
RUN chmod +x /tgi-entrypoint.sh

ENTRYPOINT ["/tgi-entrypoint.sh"]
CMD ["--json-output"]