# mindxtrain serving container — vLLM-ROCm.
#
# Boots vLLM against the FP8-quantized checkpoint produced by
# `mindxtrain quantize`. Uses AITER kernels (FlashAttention/MoE/FP8).
FROM docker.io/rocm/vllm-dev:rocm7.2.1

ENV PYTORCH_ROCM_ARCH=gfx942 \
    HSA_NO_SCRATCH_RECLAIM=1 \
    HIP_FORCE_DEV_KERNARG=1 \
    GPU_MAX_HW_QUEUES=1 \
    VLLM_USE_TRITON_FLASH_ATTN=0

WORKDIR /workspace/mindxtrain

EXPOSE 8000

# The image already contains vLLM-ROCm; the actual `vllm serve` invocation
# is constructed by `mindxtrain.deploy.vllm_launcher.build_vllm_command` and run
# at container start time.
ENTRYPOINT ["vllm"]
CMD ["--help"]