Dockerfile

# DeepSeek-V4-Flash REAP K160 on 8× RTX 5090 (SM120) — SGLang
# Validated: 50 sessions × 50k KV resident, 1,578 agg tok/s decode-only (no-spec, graphs bs56).
FROM lmsysorg/sglang:latest

ENV PYTHONUNBUFFERED=1 \
    TORCH_CUDA_ARCH_LIST="12.0" \
    HF_HUB_ENABLE_HF_TRANSFER=1 \
    PIP_ROOT_USER_ACTION=ignore

RUN pip install --no-cache-dir -U hf_transfer

WORKDIR /root
COPY scripts/ /root/scripts/
COPY tests/  /root/tests/
COPY patches/ /root/patches/
RUN chmod +x /root/scripts/*.sh

# Patch flashinfer sources for topk=192 (PR #4309 backport) — file edits only,
# no GPU needed. Kernel compile + 47-test validation happens on first GPU boot
# (entrypoint), because docker build has no GPU access.
RUN python3 /root/patches/patch_flashinfer.py

# First-boot flow: validate kernels -> download weights -> serve (see scripts/entrypoint.sh)
ENTRYPOINT ["/root/scripts/entrypoint.sh"]