+14








d33dd0566a
Co-authored-by: 1am9trash <1am9trash@gmail.com> Co-authored-by: Alex Nails <alex.nails@radixark.ai> Co-authored-by: Chunan Zeng <zcnrex@gmail.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: Ethan Che <eche@meta.com> Co-authored-by: HAI <hixiao@gmail.com> Co-authored-by: Jiajun Li <guapisolo@gmail.com> Co-authored-by: Jiajun Li <jiajun.li@radixark.ai> Co-authored-by: Jimmy Shiu <jimmysh341@gmail.com> Co-authored-by: Jimmy Shong <69131491+Jiminator@users.noreply.github.com> Co-authored-by: Liangsheng Yin <lsyincs@gmail.com> Co-authored-by: Wookje Han <52435620+wookjeHan@users.noreply.github.com> Co-authored-by: Wookje Han <wookjeh@nvl72d079-T17.cm.cluster> Co-authored-by: Wookje Han <wookjeh@nvl72d094-T17.cm.cluster> Co-authored-by: Wookje Han <wookjeh@nvl72d157-T16.cm.cluster> Co-authored-by: Wookje Han <wookjeh@nvl72d170-T04.cm.cluster> Co-authored-by: Wookje Han <wookjeh@nvl72d214-T17.cm.cluster> Co-authored-by: Yuhao Yang <47235274+yhyang201@users.noreply.github.com> Co-authored-by: Yuhao Yang <yhyang201@users.noreply.github.com> Co-authored-by: Zijie Chen <300606707+zijiecode@users.noreply.github.com> Co-authored-by: ethche <ethche@gmail.com> Co-authored-by: jacky.cheng <yichiche@amd.com> Co-authored-by: jianan-gu <jianan.gu@intel.com>
64 lines
3.6 KiB
Docker
64 lines
3.6 KiB
Docker
# syntax=docker/dockerfile:1@sha256:ecfaec9ed6d810b56388c508f4121597bfbba70d41a6dfeee4d8cad5f295fc32
|
|
# Muse Spark serving image (CUDA targets: 9.0, 10.0a, 10.3a).
|
|
# Build from the repository root; application source is cloned from upstream:
|
|
# docker buildx build --platform linux/amd64 --load \
|
|
# -f docker/muse_spark/muse-spark.Dockerfile -t muse-spark:local-test .
|
|
FROM lmsysorg/sglang:v0.5.21@sha256:2dbe4c7f53230b09aaa8fa41891c13c51d9d2e253f486a70822aaf08783fff75 AS base
|
|
|
|
SHELL ["/bin/bash", "-o", "pipefail", "-c"]
|
|
|
|
# Override to narrow the CUDA compilation targets.
|
|
ARG TORCH_CUDA_ARCH_LIST="9.0;10.0a;10.3a"
|
|
ARG MAX_JOBS=16
|
|
|
|
# Replace the stock SGLang with the upstream muse-spark branch.
|
|
# Align the runtime with Muse Spark's PyTorch pin before compiling its extensions.
|
|
# The base has Torch 2.13 and matching older TorchVision/DeepEP binaries.
|
|
WORKDIR /sgl-workspace
|
|
RUN git clone --depth 1 --single-branch --branch muse-spark \
|
|
https://github.com/sgl-project/sglang.git /sgl-workspace/sglang-muse-spark
|
|
RUN python3 -m pip uninstall -y sglang \
|
|
&& python3 -m pip install --no-cache-dir \
|
|
torch==2.14.1 torchvision==0.29.1 sgl-deep-ep==0.1.2.post1 \
|
|
&& python3 -m pip install --no-cache-dir --no-deps -e sglang-muse-spark/python
|
|
|
|
# Match python/pyproject.toml; --no-deps leaves the base's older kernel installed.
|
|
RUN python3 -m pip install --no-cache-dir --no-deps sglang-kernel==0.4.9
|
|
|
|
# FlashInfer 0.7 contains the per-token NVFP4 scale-layout/cubin fix
|
|
# (flashinfer-ai/flashinfer#4741). Use matching stock Python, cubin and
|
|
# architecture-specific JIT-cache providers; no Muse native replacement.
|
|
# The v0.5.21 SGLang base still ships FlashInfer 0.6.18, so upgrading
|
|
# the base alone cannot replace this explicit dependency upgrade.
|
|
ARG FLASHINFER_VERSION=0.7.0.post1
|
|
RUN python3 -m pip install --no-cache-dir \
|
|
"flashinfer-python[cu13]==${FLASHINFER_VERSION}" \
|
|
"nvidia-cutlass-dsl[cu13]==4.8.0" quack-kernels==0.6.5 \
|
|
sentencepiece==0.2.1 tokenizers==0.23.2 transformers==5.19.0 \
|
|
&& python3 -m pip install --no-cache-dir \
|
|
"flashinfer-cubin==${FLASHINFER_VERSION}" --index-url https://flashinfer.ai/whl \
|
|
&& python3 -m pip install --no-cache-dir \
|
|
"flashinfer-jit-cache==${FLASHINFER_VERSION}" --index-url https://flashinfer.ai/whl/cu130 \
|
|
&& python3 -c "from importlib.metadata import version; assert version('flashinfer-python') == '${FLASHINFER_VERSION}'; assert version('flashinfer-cubin') == '${FLASHINFER_VERSION}'; assert version('flashinfer-jit-cache').split('+')[0] == '${FLASHINFER_VERSION}'"
|
|
|
|
# DeepJIT in DeepGEMM requires elfutils/libdwfl.h at build time.
|
|
RUN apt-get update \
|
|
&& apt-get install -y --no-install-recommends libdw-dev \
|
|
&& rm -rf /var/lib/apt/lists/*
|
|
|
|
# DeepGEMM release/v0.2.3 includes shared TRT-LLM weights and post-expert norm.
|
|
# Pin the release source so the extension and JIT headers use the same API.
|
|
RUN git clone https://github.com/sgl-project/DeepGEMM.git /tmp/muse-deep-gemm \
|
|
&& cd /tmp/muse-deep-gemm \
|
|
&& git checkout --detach e6f9bd6a95eec26022a7397b7b194fe0da374df8 \
|
|
&& MAX_JOBS="${MAX_JOBS}" bash build_sgl_deep_gemm.sh \
|
|
&& python3 -m pip install --no-cache-dir --no-deps --force-reinstall dist/sgl_deep_gemm-*.whl \
|
|
&& cd /sgl-workspace \
|
|
&& python3 -c "from importlib.metadata import version; assert version('sgl-deep-gemm') == '0.2.3'" \
|
|
&& rm -rf /tmp/muse-deep-gemm
|
|
|
|
# The muse-spark branch also requires NVSHMEM Python bindings absent from the base.
|
|
RUN python3 -m pip install --no-cache-dir nvshmem4py-cu13==0.3.1
|
|
|
|
WORKDIR /sgl-workspace/sglang-muse-spark
|