biondizzle/vllm-to-sglang
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix aiter

Browse Source
This commit is contained in:
biondizzle
2026-04-12 02:47:33 +00:00
parent 4d444bebbb
commit b6151ba5db
1 changed files with 26 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
Dockerfile
View File

@@ -1,12 +1,29 @@
FROM lmsysorg/sglang-rocm:v0.5.10rc0-rocm700-mi30x-20260411 FROM lmsysorg/sglang-rocm:v0.5.10rc0-rocm700-mi30x-20260411
# Replace the vllm binary with our shim so no matter how the # ---------------------------------------------------------------
# production stack invokes vllm, we intercept it # Fix aiter — try pip upgrade first, fall back to source build
# ---------------------------------------------------------------
RUN pip install --upgrade aiter || \
( \
echo "=== pip upgrade failed, building aiter from source ===" && \
pip uninstall -y aiter || true && \
pip install psutil pybind11 flydsl==0.0.1.dev95158637 && \
git clone --recursive https://github.com/ROCm/aiter.git /tmp/aiter && \
cd /tmp/aiter && \
git checkout v0.1.11.post1 && \
git submodule sync && \
git submodule update --init --recursive && \
PREBUILD_KERNELS=1 GPU_ARCHS=gfx942 python3 setup.py install && \
cd / && rm -rf /tmp/aiter \
)
# ---------------------------------------------------------------
# Replace the vllm binary with our shim
# ---------------------------------------------------------------
COPY vllm-shim.sh /usr/local/bin/vllm COPY vllm-shim.sh /usr/local/bin/vllm
RUN chmod +x /usr/local/bin/vllm RUN chmod +x /usr/local/bin/vllm
# Also handle `python -m vllm.entrypoints.openai.api_server` and # Shadow `python -m vllm.*` invocations
# `python -m vllm.entrypoints.cli.main` by shadowing the vllm package
RUN mkdir -p /opt/vllm-shim/vllm/entrypoints/openai \ RUN mkdir -p /opt/vllm-shim/vllm/entrypoints/openai \
/opt/vllm-shim/vllm/entrypoints/cli /opt/vllm-shim/vllm/entrypoints/cli
COPY vllm_shim_module.py /opt/vllm-shim/vllm/__main__.py COPY vllm_shim_module.py /opt/vllm-shim/vllm/__main__.py
@@ -17,9 +34,12 @@ RUN touch /opt/vllm-shim/vllm/__init__.py \
/opt/vllm-shim/vllm/entrypoints/openai/__init__.py \ /opt/vllm-shim/vllm/entrypoints/openai/__init__.py \
/opt/vllm-shim/vllm/entrypoints/cli/__init__.py /opt/vllm-shim/vllm/entrypoints/cli/__init__.py
# Prepend shim to PYTHONPATH so it shadows any real vllm install
ENV PYTHONPATH="/opt/vllm-shim:${PYTHONPATH}" ENV PYTHONPATH="/opt/vllm-shim:${PYTHONPATH}"
# ---------------------------------------------------------------
# MI300X tuning
# ---------------------------------------------------------------
ENV HIP_FORCE_DEV_KERNARG=1 ENV HIP_FORCE_DEV_KERNARG=1
ENV NCCL_MIN_NCHANNELS=112 ENV NCCL_MIN_NCHANNELS=112
ENV GPU_MAX_HW_QUEUES=2 ENV GPU_MAX_HW_QUEUES=2
ENV SGLANG_USE_AITER=1