text-generation-inference/Dockerfile_amd

# Rust builder
FROM lukemathwalker/cargo-chef:latest-rust-1.85.1 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 apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
    python3.11-dev
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.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 build --profile release-opt --frozen

FROM rocm/dev-ubuntu-22.04:6.3.1-complete AS base

ARG HIPBLASLT_BRANCH="4d40e36"
ARG HIPBLAS_COMMON_BRANCH="7c1566b"
ARG LEGACY_HIPBLASLT_OPTION=
ARG RCCL_BRANCH="648a58d"
ARG RCCL_REPO="https://github.com/ROCm/rccl"
ARG TRITON_BRANCH="e5be006"
ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
ARG PYTORCH_BRANCH="3a585126"
ARG PYTORCH_VISION_BRANCH="v0.19.1"
ARG PYTORCH_REPO="https://github.com/pytorch/pytorch.git"
ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
ARG FA_BRANCH="b7d29fb"
ARG FA_REPO="https://github.com/ROCm/flash-attention.git"
ARG AITER_BRANCH="21d47a9"
ARG AITER_REPO="https://github.com/ROCm/aiter.git"

ENV PATH=/opt/rocm/llvm/bin:$PATH
ENV ROCM_PATH=/opt/rocm
ENV LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
ARG PYTORCH_ROCM_ARCH=gfx90a;gfx942
ENV PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}

ARG PYTHON_VERSION=3.11

RUN mkdir -p /app
WORKDIR /app
ENV DEBIAN_FRONTEND=noninteractive

# Install Python and other dependencies
RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
        build-essential \
        ca-certificates \
        ccache \
        curl \
        git \
        ninja-build \
        cmake \
        software-properties-common \
        python3.11-dev \
        python3.11-venv && \
        rm -rf /var/lib/apt/lists/*

COPY --from=ghcr.io/astral-sh/uv:0.5.31 /uv /uvx /bin/
ENV PATH="$PATH:/root/.local/bin"
RUN uv python install ${PYTHON_VERSION}
RUN uv venv --python ${PYTHON_VERSION} && uv pip install pip setuptools packaging
ENV VIRTUAL_ENV=/usr/src/.venv/
ENV PATH="$PATH:/usr/src/.venv/bin/"

RUN . .venv/bin/activate && pip install -U packaging cmake ninja wheel setuptools pybind11 Cython

FROM base AS build_hipblaslt
ARG HIPBLASLT_BRANCH
ARG HIPBLAS_COMMON_BRANCH
# Set to "--legacy_hipblas_direct" for ROCm<=6.2
ARG LEGACY_HIPBLASLT_OPTION
RUN git clone https://github.com/ROCm/hipBLAS-common.git
RUN . .venv/bin/activate && cd hipBLAS-common \
    && git checkout ${HIPBLAS_COMMON_BRANCH} \
    && mkdir build \
    && cd build \
    && cmake .. \
    && make package \
    && dpkg -i ./*.deb
RUN git clone https://github.com/ROCm/hipBLASLt
RUN . .venv/bin/activate && cd hipBLASLt \
    && git checkout ${HIPBLASLT_BRANCH} \
    && ./install.sh -d --architecture ${PYTORCH_ROCM_ARCH} ${LEGACY_HIPBLASLT_OPTION} \
    && cd build/release \
    && make package
RUN mkdir -p /app/install && cp /app/hipBLASLt/build/release/*.deb /app/hipBLAS-common/build/*.deb /app/install

FROM base AS build_rccl
ARG RCCL_BRANCH
ARG RCCL_REPO
RUN git clone ${RCCL_REPO}
RUN . .venv/bin/activate && cd rccl \
    && git checkout ${RCCL_BRANCH} \
    && ./install.sh -p --amdgpu_targets ${PYTORCH_ROCM_ARCH}
RUN mkdir -p /app/install && cp /app/rccl/build/release/*.deb /app/install

FROM base AS build_triton
ARG TRITON_BRANCH
ARG TRITON_REPO
RUN git clone ${TRITON_REPO}
RUN . .venv/bin/activate && cd triton \
    && git checkout ${TRITON_BRANCH} \
    && cd python \
    && python3 setup.py bdist_wheel --dist-dir=dist
RUN mkdir -p /app/install && cp /app/triton/python/dist/*.whl /app/install

FROM base AS build_amdsmi
RUN . .venv/bin/activate && cd /opt/rocm/share/amd_smi \
    && pip wheel . --wheel-dir=dist
RUN mkdir -p /app/install && cp /opt/rocm/share/amd_smi/dist/*.whl /app/install

FROM base AS build_pytorch
ARG PYTORCH_BRANCH
ARG PYTORCH_VISION_BRANCH
ARG PYTORCH_REPO
ARG PYTORCH_VISION_REPO
ARG FA_BRANCH
ARG FA_REPO
RUN git clone ${PYTORCH_REPO} pytorch
RUN . .venv/bin/activate && cd pytorch && git checkout ${PYTORCH_BRANCH} && \
    pip install -r requirements.txt && git submodule update --init --recursive \
    && python3 tools/amd_build/build_amd.py \
    && CMAKE_PREFIX_PATH=$(python3 -c 'import sys; print(sys.prefix)') python3 setup.py bdist_wheel --dist-dir=dist \
    && pip install dist/*.whl
RUN git clone ${PYTORCH_VISION_REPO} vision
RUN . .venv/bin/activate && cd vision && git checkout ${PYTORCH_VISION_BRANCH} \
    && python3 setup.py bdist_wheel --dist-dir=dist \
    && pip install dist/*.whl
RUN git clone ${FA_REPO}
RUN . .venv/bin/activate && cd flash-attention \
    && git checkout ${FA_BRANCH} \
    && git submodule update --init \
    && MAX_JOBS=64 GPU_ARCHS=${PYTORCH_ROCM_ARCH} python3 setup.py bdist_wheel --dist-dir=dist
RUN mkdir -p /app/install && cp /app/pytorch/dist/*.whl /app/install \
    && cp /app/vision/dist/*.whl /app/install \
    && cp /app/flash-attention/dist/*.whl /app/install

FROM base AS final
RUN --mount=type=bind,from=build_hipblaslt,src=/app/install/,target=/install \
    dpkg -i /install/*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
RUN --mount=type=bind,from=build_rccl,src=/app/install/,target=/install \
    dpkg -i /install/*deb \
    && sed -i 's/, rccl-dev \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status \
    && sed -i 's/, rccl \(.*\), rocalution/, rocalution/g' /var/lib/dpkg/status
RUN --mount=type=bind,from=build_triton,src=/app/install/,target=/install \
    . .venv/bin/activate && \
    pip install /install/*.whl
RUN --mount=type=bind,from=build_amdsmi,src=/app/install/,target=/install \
    . .venv/bin/activate && \
    pip install /install/*.whl
RUN --mount=type=bind,from=build_pytorch,src=/app/install/,target=/install \
    . .venv/bin/activate && \
    pip install /install/*.whl

ARG AITER_REPO
ARG AITER_BRANCH
RUN git clone --recursive ${AITER_REPO}
RUN . .venv/bin/activate && cd aiter \
    && git checkout ${AITER_BRANCH} \
    && git submodule update --init --recursive \
    && pip install -r requirements.txt \
    && PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py develop && pip show aiter 

RUN rm -rf /var/lib/apt/lists/*

FROM final AS kernel-builder
# # Build vllm kernels
FROM kernel-builder AS vllm-builder

COPY server/Makefile-vllm Makefile
RUN . .venv/bin/activate && pip install setuptools_scm

# Build specific version of vllm
RUN . .venv/bin/activate && make build-vllm-rocm

# Build Transformers CUDA kernels (gpt-neox and bloom)
FROM kernel-builder AS custom-kernels-builder
COPY server/custom_kernels/ .
RUN . .venv/bin/activate && python3 setup.py bdist_wheel --dist-dir=dist

# Build exllama kernels
FROM kernel-builder AS exllama-kernels-builder
COPY server/exllama_kernels/ .
RUN . .venv/bin/activate && python3 setup.py bdist_wheel --dist-dir=dist

# Build exllama v2 kernels
FROM kernel-builder AS exllamav2-kernels-builder
COPY server/exllamav2_kernels/ .
RUN . .venv/bin/activate && python3 setup.py bdist_wheel --dist-dir=dist

FROM kernel-builder AS marlin-kernels
ENV MARLIN_KERNELS_BRANCH=v0.3.6
ENV VLLM_TARGET_DEVICE=rocm
RUN . .venv/bin/activate && git clone https://github.com/danieldk/marlin-kernels.git && \
    cd marlin-kernels && \
    git checkout ${MARLIN_KERNELS_BRANCH} && \
    python3 setup.py bdist_wheel --dist-dir=dist

FROM kernel-builder AS moe-kernels
ENV MOE_KERNELS_BRANCH=v0.8.2
ENV VLLM_TARGET_DEVICE=rocm
RUN . .venv/bin/activate && git clone https://github.com/danieldk/moe-kernels.git && \
    cd moe-kernels && \
    git checkout ${MOE_KERNELS_BRANCH} && \
    python3 setup.py bdist_wheel --dist-dir=dist

FROM final AS base-copy

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

ENV VIRTUAL_ENV=/app/.venv/
ENV PATH="$PATH:/app/.venv/bin/"
RUN uv pip install kernels

# Install server
COPY proto proto
COPY server server
COPY server/Makefile server/Makefile
RUN cd server && \
    uv pip install grpcio-tools mypy-protobuf && \
    uv pip install -e ".[accelerate, compressed-tensors, peft, outlines]" --no-cache-dir && \
    make gen-server-raw && \
    kernels download .

RUN cd server && \
    pwd && \
    text-generation-server --help

RUN --mount=type=bind,from=vllm-builder,src=/app/vllm/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=custom-kernels-builder,src=/app/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=custom-kernels-builder,src=/app/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=exllama-kernels-builder,src=/app/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=exllamav2-kernels-builder,src=/app/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=marlin-kernels,src=/app/marlin-kernels/dist,target=/install \
    uv pip install /install/*.whl
RUN --mount=type=bind,from=moe-kernels,src=/app/moe-kernels/dist,target=/install \
    uv pip install /install/*.whl


# 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

# 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
ENV ROCM_USE_CUSTOM_PAGED_ATTN=1
ENV PYTORCH_TUNABLEOP_TUNING_AFTER_WARMUP=0
ENV VLLM_MOE_PADDING=0
ENV ATTENTION=paged
ENV PREFIX_CACHING=0
ENV PREFILL_CHUNKING=0
ENV ROCM_USE_SKINNY_GEMM=1
ENV PYTORCH_TUNABLEOP_ENABLED=0
COPY ./tgi-entrypoint.sh /tgi-entrypoint.sh
RUN chmod +x /tgi-entrypoint.sh

ENTRYPOINT ["/tgi-entrypoint.sh"]
ENV LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/root/.local/share/uv/python/cpython-3.11.11-linux-x86_64-gnu/lib"
ENV PYTHONPATH=/app/.venv/lib/python3.11/site-packages
# CMD ["--json-output"]