# DeepGEMM NVFP4 Mega MoE Build Container
# Based on the vLLM container that already has CUDA 13.0 + PyTorch + CUTLASS + FlashInfer
# for Blackwell SM100 support.
#
# Build: docker build -t deepgemm-nvfp4 .
# Run:   docker run --gpus all -it deepgemm-nvfp4 bash

FROM atl.vultrcr.io/vllm/vllm-with-lmcache:dream-build

# Install build essentials
RUN apt-get update && apt-get install -y \
    git \
    screen \
    cmake \
    && rm -rf /var/lib/apt/lists/*

# Clone DeepGEMM nvfp4-mega-moe branch
# CACHE_BUSTER: increment to force fresh clone (bypasses Docker cache)
RUN git clone -b nvfp4-mega-moe https://sweetapi.com/biondizzle/DeepGEMM.git /root/DeepGEMM && CACHE_BUSTER=1

# Set include paths for CUTLASS/CuTe headers (already in the container via flashinfer/vllm)
ENV CPATH="/usr/local/lib/python3.12/dist-packages/flashinfer/data/cutlass/include:/usr/local/lib/python3.12/dist-packages/nvidia/cu13/include:${CPATH}"
ENV CUDA_HOME="/usr/local/cuda"

# Build DeepGEMM
RUN cd /root/DeepGEMM && python3 setup.py build_ext --inplace

# Verify build
RUN python3 -c "import sys; sys.path.insert(0, '/root/DeepGEMM'); import deep_gemm; print('DeepGEMM NVFP4 build OK')"

WORKDIR /root/DeepGEMM