FROM vllm/vllm-openai:latest

ENV PORT=7860
EXPOSE 7860

# Force unbuffered standard I/O
ENV PYTHONUNBUFFERED=1

# Redirect all caches from /root to /tmp for Hugging Face UID 1000
ENV HOME=/tmp
ENV HF_HOME=/tmp/huggingface
ENV XDG_CACHE_HOME=/tmp/cache
ENV TRITON_CACHE_DIR=/tmp/triton
ENV NUMBA_CACHE_DIR=/tmp/numba

# PCIe multi-GPU fixes for L40S
ENV NCCL_P2P_DISABLE=1
ENV NCCL_IB_DISABLE=1

ENV MODEL_NAME="Qwen/Qwen3.8-27B-FP8"

# Grant write permissions to /tmp for user 1000
USER root
RUN mkdir -p /tmp/huggingface /tmp/cache /tmp/triton /tmp/numba && \
    chmod -R 777 /tmp

# Switch to user 1000 as required by Hugging Face Spaces
USER 1000

# Use python3 module directly with 2>&1 redirection to force log visibility
ENTRYPOINT ["sh", "-c", "python3 -m vllm.entrypoints.openai.api_server \
    --model ${MODEL_NAME} \
    --host 0.0.0.0 \
    --port 7860 \
    --tensor-parallel-size 4 \
    --pipeline-parallel-size 2 \
    --disable-custom-all-reduce \
    --gpu-memory-utilization 0.90 \
    --max-model-len 131072 \
    --enable-chunked-prefill \
    --max-num-batched-tokens 16384 \
    --swap-space 4 \
    --reasoning-parser qwen3 \
    --speculative-config '{\"method\":\"mtp\",\"num_speculative_tokens\":2}' \
    --trust-remote-code \
    ${API_KEY:+--api-key $API_KEY} 2>&1"]