Dockerfile.rocm

# Solaris — AMD GPU base image
# Tested with ROCm 6.2 / PyTorch 2.4+
#
# Build:
#   docker build -f Dockerfile.rocm -t solaris:latest .
#
# Run (requires --device=/dev/kfd --device=/dev/dri and proper group membership):
#   docker run --rm -it \
#       --device=/dev/kfd --device=/dev/dri \
#       --group-add video --group-add render \
#       -v $(pwd):/workspace \
#       solaris:latest bash

FROM rocm/pytorch:rocm6.2_ubuntu22.04_py3.10_pytorch_release_2.3.0
# GPU targets — edit to match your hardware:
#   gfx906  = MI50/MI60        (Vega20)
#   gfx908  = MI100            (CDNA1)
#   gfx90a  = MI200            (CDNA2)
#   gfx940/941/942 = MI300     (CDNA3)
#   gfx1030 = RX 6800 XT      (RDNA2)
#   gfx1100 = RX 7900 XTX     (RDNA3)
#   gfx1200 = RX 9070/XT      (RDNA4)
ENV ROCM_TARGET_LST="gfx906;gfx908;gfx90a;gfx1030;gfx1100;gfx1200"
ENV PYTORCH_ROCM_ARCH="${ROCM_TARGET_LST}"

WORKDIR /workspace

# System packages
RUN apt-get update && apt-get install -y --no-install-recommends \
        git build-essential cmake \
        libopenmpi-dev openmpi-bin \
    && rm -rf /var/lib/apt/lists/*

# Copy project and install
COPY . /workspace/solaris
WORKDIR /workspace/solaris

# Install package (torch already in base image)
RUN pip install --no-cache-dir -e ".[datapipes,dev]"

# Environment: RCCL (AMD's NCCL equivalent) is used automatically via
# PyTorch's nccl backend — no code changes needed.
ENV NCCL_DEBUG=INFO
ENV HSA_OVERRIDE_GFX_VERSION=""

CMD ["python", "-c", "import solaris; print('Solaris', solaris.__version__)"]