279 lines
10 KiB
Docker
279 lines
10 KiB
Docker
# 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:25.12-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 $VLLM_BASE_DIR
|
|
|
|
# 2. Set Environment Variables
|
|
ENV TORCH_CUDA_ARCH_LIST="12.0;12.1"
|
|
ENV TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas
|
|
|
|
# --- CACHE BUSTER ---
|
|
# Change this argument to force a re-download of PyTorch/FlashInfer
|
|
ARG CACHEBUST_DEPS=1
|
|
|
|
# Install additional dependencies
|
|
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
|
|
uv pip install fastsafetensors
|
|
|
|
ARG PRE_TRANSFORMERS=0
|
|
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
|
|
if [ "$PRE_TRANSFORMERS" = "1" ]; then \
|
|
uv pip install -U transformers --pre; \
|
|
fi
|
|
# =========================================================
|
|
# STAGE 2: Builder
|
|
# =========================================================
|
|
FROM base AS builder
|
|
|
|
ENV FLASHINFER_CUDA_ARCH_LIST="12.1f"
|
|
|
|
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=11af7f02ab52c9130e422eeb4b44042fbd60c083
|
|
|
|
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
|
|
uv pip install nvidia-nvshmem-cu13 "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; \
|
|
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 ${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; \
|
|
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 ${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=459541683f2d8c21f9c0e2f44954b04f59611cbe
|
|
|
|
# 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; \
|
|
else \
|
|
echo "Cache hit: Fetching updates..." && \
|
|
cd vllm-mxfp4 && \
|
|
git fetch --all && \
|
|
git checkout ${VLLM_SHA} && \
|
|
if [ "${VLLM_SHA}" = "main" ]; then \
|
|
git reset --hard origin/main; \
|
|
fi && \
|
|
git submodule update --init --recursive && \
|
|
# 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/foundation-model-stack/fastsafetensors/issues/36
|
|
#COPY fastsafetensors.patch .
|
|
#RUN patch -p1 < fastsafetensors.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:25.12-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 \
|
|
libxcb1 \
|
|
&& rm -rf /var/lib/apt/lists/* \
|
|
&& pip install uv && pip uninstall -y flash-attn
|
|
|
|
# Set final working directory
|
|
WORKDIR $VLLM_BASE_DIR
|
|
|
|
# Download Tiktoken files
|
|
RUN mkdir -p tiktoken_encodings && \
|
|
wget -O tiktoken_encodings/o200k_base.tiktoken "https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken" && \
|
|
wget -O tiktoken_encodings/cl100k_base.tiktoken "https://openaipublic.blob.core.windows.net/encodings/cl100k_base.tiktoken"
|
|
|
|
# Copy artifacts from Builder Stage
|
|
# We copy the python packages and executables
|
|
# No need to copy source code, as it's already in the site-packages
|
|
COPY --from=builder /workspace/wheels /workspace/wheels
|
|
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
|
|
uv pip install /workspace/wheels/*.whl && \
|
|
rm -rf /workspace/wheels
|
|
|
|
# Setup Env for Runtime
|
|
ENV TORCH_CUDA_ARCH_LIST="12.0;12.1"
|
|
ENV FLASHINFER_CUDA_ARCH_LIST="12.1f"
|
|
ENV TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas
|
|
ENV TIKTOKEN_ENCODINGS_BASE=$VLLM_BASE_DIR/tiktoken_encodings
|
|
ENV PATH=$VLLM_BASE_DIR:$PATH
|
|
|
|
# Copy scripts
|
|
COPY run-cluster-node.sh $VLLM_BASE_DIR/
|
|
RUN chmod +x $VLLM_BASE_DIR/run-cluster-node.sh
|
|
|
|
# Final extra deps
|
|
RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
|
|
uv pip install ray[default] fastsafetensors
|
|
|