Dockerfile.mxfp4

# syntax=docker/dockerfile:1.6

# Limit build parallelism to reduce OOM situations
ARG BUILD_JOBS=16

# =========================================================
# STAGE 1: Base Image (Installs Dependencies)
# =========================================================
FROM nvcr.io/nvidia/pytorch:26.01-py3 AS base

# Build parallemism
ARG BUILD_JOBS
ENV MAX_JOBS=${BUILD_JOBS}  
ENV CMAKE_BUILD_PARALLEL_LEVEL=${BUILD_JOBS}
ENV NINJAFLAGS="-j${BUILD_JOBS}"
ENV MAKEFLAGS="-j${BUILD_JOBS}"

# =============================================================================
# Pinned versions from https://github.com/christopherowen/spark-vllm-mxfp4-docker/blob/main/Dockerfile
# =============================================================================

# ARG VLLM_SHA=045293d82b832229560ac4a13152a095af603b6e
# ARG FLASHINFER_SHA=1660ee8d740b0385f235519f9e2750db944d1838
# ARG CUTLASS_SHA=11af7f02ab52c9130e422eeb4b44042fbd60c083

# ARG VLLM_REPO=https://github.com/christopherowen/vllm.git
# ARG FLASHINFER_REPO=https://github.com/christopherowen/flashinfer.git
# ARG CUTLASS_REPO=https://github.com/christopherowen/cutlass.git

# Set non-interactive frontend to prevent apt prompts
ENV DEBIAN_FRONTEND=noninteractive

# Allow pip to install globally on Ubuntu 24.04 without a venv
ENV PIP_BREAK_SYSTEM_PACKAGES=1

# Set pip cache directory
ENV PIP_CACHE_DIR=/root/.cache/pip
ENV UV_CACHE_DIR=/root/.cache/uv
ENV UV_SYSTEM_PYTHON=1
ENV UV_BREAK_SYSTEM_PACKAGES=1
ENV UV_LINK_MODE=copy

# Set the base directory environment variable
ENV VLLM_BASE_DIR=/workspace/vllm

# 1. Install Build Dependencies & Ccache
# Added ccache to enable incremental compilation caching
RUN apt update && \
    apt install -y --no-install-recommends \
    curl vim ninja-build git \
    ccache \
    && rm -rf /var/lib/apt/lists/* \
    && pip install uv && pip uninstall -y flash-attn

# Configure Ccache for CUDA/C++
ENV PATH=/usr/lib/ccache:$PATH
ENV CCACHE_DIR=/root/.ccache
# Limit ccache size to prevent unbounded growth (e.g. 50G)
ENV CCACHE_MAXSIZE=50G
# Enable compression to save space
ENV CCACHE_COMPRESS=1
# Tell CMake to use ccache for compilation
ENV CMAKE_CXX_COMPILER_LAUNCHER=ccache
ENV CMAKE_CUDA_COMPILER_LAUNCHER=ccache

# Setup Workspace
WORKDIR $VLLM_BASE_DIR

# 2. Set Environment Variables
ARG TORCH_CUDA_ARCH_LIST="12.1a"
ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
ENV TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas

# --- CACHE BUSTER ---
# Change this argument to force a re-download of PyTorch/FlashInfer
ARG CACHEBUST_DEPS=1

# Install additional dependencies
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    uv pip install fastsafetensors

ARG PRE_TRANSFORMERS=0
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    if [ "$PRE_TRANSFORMERS" = "1" ]; then \
        uv pip install -U transformers --pre; \
    fi
# =========================================================
# STAGE 2: Builder
# =========================================================
FROM base AS builder

ARG FLASHINFER_CUDA_ARCH_LIST="12.1a"
ENV FLASHINFER_CUDA_ARCH_LIST=${FLASHINFER_CUDA_ARCH_LIST}

WORKDIR $VLLM_BASE_DIR

ARG FLASHINFER_REPO=https://github.com/christopherowen/flashinfer.git
ARG CUTLASS_REPO=https://github.com/christopherowen/cutlass.git

ARG FLASHINFER_SHA=f349e52496a72a00d8c4ac02c7a1e38523ff7194
ARG CUTLASS_SHA=fede53000a962b46e05bafe0c86311778caeb380

RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     uv pip install "nvidia-nvshmem-cu13<3.6" "apache-tvm-ffi<0.2"

# Clone FlashInfer (cached for faster rebuilds)
RUN --mount=type=cache,id=git-flashinfer,target=/git-cache/flashinfer \
    if [ -d /git-cache/flashinfer/.git ]; then \
        echo "=== Using cached FlashInfer repo ===" && \
        cp -a /git-cache/flashinfer /workspace/flashinfer && \
        cd /workspace/flashinfer && \
        git fetch origin && git fetch origin --tags; \
    else \
        echo "=== Cloning FlashInfer (first build) ===" && \
        git clone ${FLASHINFER_REPO} /workspace/flashinfer && \
        cp -a /workspace/flashinfer /git-cache/flashinfer; \
    fi && \
    cd /workspace/flashinfer &&  (git checkout --detach origin/${FLASHINFER_SHA} 2>/dev/null || git checkout ${FLASHINFER_SHA})

# Clone spdlog submodule (small, no caching needed)
RUN cd /workspace/flashinfer && \
    git submodule update --init 3rdparty/spdlog

# Clone CUTLASS directly (skip submodule, use our fork)
RUN --mount=type=cache,id=git-cutlass,target=/git-cache/cutlass \
    cd /workspace/flashinfer && \
    rm -rf 3rdparty/cutlass && \
    if [ -d /git-cache/cutlass/.git ] && [ -d /git-cache/cutlass/.git/objects ]; then \
        echo "=== Using cached CUTLASS repo ===" && \
        cp -a /git-cache/cutlass 3rdparty/cutlass && \
        cd 3rdparty/cutlass && \
        git fetch origin && git fetch origin --tags; \
    else \
        echo "=== Cloning CUTLASS (first build) ===" && \
        rm -rf /git-cache/cutlass/* /git-cache/cutlass/.* 2>/dev/null || true && \
        git clone ${CUTLASS_REPO} 3rdparty/cutlass && \
        cp -a /workspace/flashinfer/3rdparty/cutlass/. /git-cache/cutlass/; \
    fi && \
    cd /workspace/flashinfer/3rdparty/cutlass && (git checkout --detach origin/${CUTLASS_SHA} 2>/dev/null || git checkout ${CUTLASS_SHA})

# Build FlashInfer wheels

WORKDIR /workspace/flashinfer

# flashinfer-python
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    --mount=type=cache,id=ccache,target=/root/.ccache \
    sed -i -e 's/license = "Apache-2.0"/license = { text = "Apache-2.0" }/' -e '/license-files/d' pyproject.toml && \
    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v

# flashinfer-cubin
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    --mount=type=cache,id=ccache,target=/root/.ccache \
    cd flashinfer-cubin && uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v

# flashinfer-jit-cache
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    --mount=type=cache,id=ccache,target=/root/.ccache \
    cd flashinfer-jit-cache && \
    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v

# --- VLLM SOURCE CACHE BUSTER ---
# Change THIS argument to force a fresh git clone and rebuild of vLLM
# without re-installing the dependencies above.
ARG CACHEBUST_VLLM=1

ARG VLLM_REPO=https://github.com/christopherowen/vllm.git

# Git reference (branch, tag, or SHA) to checkout
ARG VLLM_SHA=045293d82b832229560ac4a13152a095af603b6e

# 4. Smart Git Clone (Fetch changes instead of full re-clone)
# We mount a cache at /repo-cache. This directory persists on your host machine.
RUN --mount=type=cache,id=repo-cache,target=/repo-cache \
    # 1. Go into the persistent cache directory
    cd /repo-cache && \
    # 2. Logic: Clone if missing, otherwise Fetch & Reset
    if [ ! -d "vllm-mxfp4" ]; then \
        echo "Cache miss: Cloning vLLM from scratch..." && \
        git clone --recursive ${VLLM_REPO} vllm-mxfp4 && \
        cd vllm-mxfp4 && \
        git checkout ${VLLM_SHA} ; \
    else \
        echo "Cache hit: Fetching updates..." && \
        cd vllm-mxfp4 && \
        git fetch origin && \
        git fetch origin --tags && \
        (git checkout --detach origin/${VLLM_SHA} 2>/dev/null || git checkout ${VLLM_SHA}) && \
        git submodule update --init --recursive && \
        git clean -fdx && \
        # Optimize git repo size
        git gc --auto; \
    fi && \
    # 3. Copy the updated code from the cache to the actual container workspace
    # We use 'cp -a' to preserve permissions
    mkdir $VLLM_BASE_DIR/vllm && \
    cp -a -r /repo-cache/vllm-mxfp4/. $VLLM_BASE_DIR/vllm/

WORKDIR $VLLM_BASE_DIR/vllm

ARG PRE_TRANSFORMERS=0

# Prepare build requirements
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    python3 use_existing_torch.py && \
    sed -i "/flashinfer/d" requirements/cuda.txt && \
    sed -i '/^fastsafetensors\b/d' requirements/test.txt && \
    if [ "$PRE_TRANSFORMERS" = "1" ]; then \
        sed -i '/^transformers\b/d' requirements/common.txt; \
        sed -i '/^transformers\b/d' requirements/test.txt; \
    fi && \
    uv pip install -r requirements/build.txt

# Apply Patches
# TEMPORARY PATCH for fastsafetensors loading in cluster setup - tracking https://github.com/vllm-project/vllm/issues/34180
COPY fastsafetensors_mxfp4.patch .
RUN patch -p1 < fastsafetensors_mxfp4.patch

# Final Compilation
# We mount the ccache directory here. Ideally, map this to a host volume for persistence 
# across totally separate `docker build` invocations.
RUN --mount=type=cache,id=ccache,target=/root/.ccache \
    --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v

# =========================================================
# STAGE 4: Runner (Transfers only necessary artifacts)
# =========================================================
FROM nvcr.io/nvidia/pytorch:26.01-py3 AS runner

ENV DEBIAN_FRONTEND=noninteractive
ENV PIP_BREAK_SYSTEM_PACKAGES=1
ENV VLLM_BASE_DIR=/workspace/vllm

# Set pip cache directory
ENV PIP_CACHE_DIR=/root/.cache/pip
ENV UV_CACHE_DIR=/root/.cache/uv
ENV UV_SYSTEM_PYTHON=1
ENV UV_BREAK_SYSTEM_PACKAGES=1
ENV UV_LINK_MODE=copy

# Install minimal runtime dependencies (NCCL, Python)
# Note: "devel" tools like cmake/gcc are NOT installed here to save space
RUN apt update && \
    apt install -y --no-install-recommends \
    curl vim git earlyoom \
    libxcb1 \
    && rm -rf /var/lib/apt/lists/* \
    && pip install uv && pip uninstall -y flash-attn 

# Set final working directory
WORKDIR $VLLM_BASE_DIR

# Download Tiktoken files
RUN mkdir -p tiktoken_encodings && \
    wget -O tiktoken_encodings/o200k_base.tiktoken "https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken" && \
    wget -O tiktoken_encodings/cl100k_base.tiktoken "https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken"

# Copy artifacts from Builder Stage
# FastAPI 0.137.0 adds _IncludedRouter entries that currently break
# prometheus-fastapi-instrumentator route name lookup.
RUN --mount=type=bind,from=builder,source=/workspace/wheels,target=/mount/wheels \
    --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    PINNED_TORCH=$(python3 -c "import torch; print(torch.__version__)") && \
    echo "torch==${PINNED_TORCH}" > /tmp/wheel-override.txt && \
    echo "fastapi[standard]>=0.115.0,<0.137.0" >> /tmp/wheel-override.txt && \
    uv pip install /mount/wheels/*.whl --override /tmp/wheel-override.txt

# Setup environment vars for runtime
ARG TORCH_CUDA_ARCH_LIST="12.1a"
ENV TORCH_CUDA_ARCH_LIST=${TORCH_CUDA_ARCH_LIST}
ARG FLASHINFER_CUDA_ARCH_LIST="12.1a"
ENV FLASHINFER_CUDA_ARCH_LIST=${FLASHINFER_CUDA_ARCH_LIST}
ENV TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas
ENV TIKTOKEN_ENCODINGS_BASE=$VLLM_BASE_DIR/tiktoken_encodings
ENV PATH=$VLLM_BASE_DIR:$PATH

# Final extra deps
# Pin torch via --override so transitive deps can't trigger a re-resolve that
# swaps the CUDA-built torch for PyPI's CPU wheel.
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
    PINNED_TORCH=$(python3 -c "import torch; print(torch.__version__)") && \
    echo "torch==${PINNED_TORCH}" > /tmp/torch-override.txt && \
    echo "fastapi[standard]>=0.115.0,<0.137.0" >> /tmp/torch-override.txt && \
    uv pip install ray[default] fastsafetensors "nvidia-nvshmem-cu13<3.6" \
        --override /tmp/torch-override.txt

# Build metadata (generated by build-and-copy.sh)
COPY build-metadata.yaml /workspace/build-metadata.yaml

# If not compiling Triton
# remove triton-kernels as they are not compatible with this vLLM version yet
RUN uv pip uninstall triton-kernels