Building CTranslate2 for Python 3.10 from arlo-phoenix

Dockerfile

# Dockerfile for building CTranslate2 with ROCm support for Python 3.10
# Tested with ROCm 6.2 and AMD RX 7900 XT (gfx1100)
# Source: https://github.com/arlo-phoenix/CTranslate2-rocm

FROM rocm/pytorch

# Set working directory for build
WORKDIR /build

# Activate conda environment for the rest of the commands
SHELL ["conda", "run", "-n", "py_3.10", "/bin/bash", "-c"]

RUN git clone https://github.com/arlo-phoenix/CTranslate2-rocm.git --recurse-submodules
WORKDIR /build/CTranslate2-rocm

# Optional: set the ROCm architecture
ENV PYTORCH_ROCM_ARCH=gfx1100

RUN CLANG_CMAKE_CXX_COMPILER=clang++ CXX=clang++ HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
    cmake -S . -B build -DWITH_MKL=OFF -DWITH_HIP=ON -DCMAKE_HIP_ARCHITECTURES=$PYTORCH_ROCM_ARCH -DBUILD_TESTS=ON -DWITH_CUDNN=ON

RUN cmake --build build -- -j8

# Install compiled files
WORKDIR /build/CTranslate2-rocm/build
RUN cmake --install . --prefix /usr/local

RUN ldconfig

WORKDIR /build/CTranslate2-rocm/python
RUN pip install -r install_requirements.txt

# Modify setup.py as required
RUN sed -i 's|include_dirs = \[pybind11.get_include()\]|include_dirs = [pybind11.get_include(), "../include"]|' setup.py
RUN sed -i 's|library_dirs = \[\]|library_dirs = ["/usr/local/lib"]|' setup.py

RUN cp ../build/libctranslate2.so /usr/local/lib/

RUN python setup.py bdist_wheel
RUN pip install dist/*.whl

RUN apt-get update && \
    apt-get install -y --no-install-recommends libomp-dev && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Create directory for user mounting
RUN mkdir -p /dockerx

# Set to run with conda environment initialized
ENTRYPOINT ["conda", "run", "--no-capture-output", "-n", "py_3.10", "/bin/bash"]

## Build using:
# docker build -t ctranslate2-rocm .

## Run using:
# docker run -it \
#   --network=host        # Allows network access
#   --device=/dev/kfd     # ROCm kernel driver
#   --device=/dev/dri     # Direct rendering
#   --group-add=video     # GPU access
#   --ipc=host           # Shared memory
#   --cap-add=SYS_PTRACE # For debugging
#   --security-opt seccomp=unconfined \
#   --shm-size 8G        # Shared memory size
#   -v $HOME/dockerx:/dockerx  # Mount local directory
#   -w /dockerx          # Set working directory
#   ctranslate2-rocm

## Test the installation:
# python3 -c "import ctranslate2; print(ctranslate2.__version__)"