44 lines
3.0 KiB
Docker
44 lines
3.0 KiB
Docker
# enginex-mlu370-compat
|
||
#
|
||
# 寒武纪 MLU370-X4（Cambricon_mlu-370-x4）· 文本生成 · vLLM「兼容增强」引擎。
|
||
#
|
||
# 与 enginex-bi100-compat 同源同逻辑，只换基础镜像。
|
||
# 做法照搬社区已上线引擎的模式：**把镜像内的 vllm 二进制换成本仓 wrapper**，
|
||
# 平台原样下发 `vllm serve /model --port 8000 ...` 就会流经 wrapper，无需平台改任何命令。
|
||
#
|
||
# 基础镜像来源（10-03 刚从平台 vllm-customized 框架的 build-config 里挖到，CI 拉得到）：
|
||
# harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260204
|
||
# 之前找 Cambricon 基础镜像只看到官方仓 Dockerfile 里的 `combricon_vllm_mlu:v1.0_0510`
|
||
# （无 registry 前缀的内部名，CI 拉不到）而搁置；harbor 路径解锁后才能建。
|
||
#
|
||
# 为什么选 mlu-370-x4：该卡只有 vllm-mlu / vllm / vllm-customized 三个框架，
|
||
# **没有任何 tokenizer/compat 类引擎**；而本人该卡 3 个失败里 2 个是 R2（模型无 chat_template，
|
||
# 而 vllm / vllm-mlu 在该卡默认走 chat 路径）。
|
||
|
||
FROM harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260204
|
||
|
||
LABEL org.opencontainers.image.title="enginex-mlu370-compat" \
|
||
org.opencontainers.image.description="Cambricon mlu-370-x4 vLLM engine with preflight compatibility patches (R2/R3/R3b/R4/R7)" \
|
||
com.modelhubxc.engine.target-card="Cambricon_mlu-370-x4" \
|
||
com.modelhubxc.engine.framework="vllm" \
|
||
com.modelhubxc.engine.task-type="text-generation" \
|
||
com.modelhubxc.engine.base-image="harbor.4pd.io/hardcore-tech/cambricon-mlu370-pytorch:v25.01-20260204"
|
||
|
||
COPY preflight.py /opt/
|
||
COPY detect_tokenizer.py /opt/
|
||
COPY vllm_wrapper.sh /opt/
|
||
COPY shims_ixformer_layers.py /opt/shims_src/
|
||
|
||
RUN set -eux && mkdir -p /opt/shims/ixformer/contrib/vllm/layers && for d in /opt/shims/ixformer /opt/shims/ixformer/contrib /opt/shims/ixformer/contrib/vllm /opt/shims/ixformer/contrib/vllm/layers; do printf '# ModelHub XC compat shim package\n' > "$d/__init__.py"; done && cp /opt/shims_src/shims_ixformer_layers.py /opt/shims/ixformer/contrib/vllm/layers/__init__.py && rm -rf /opt/shims_src && chmod +x /opt/vllm_wrapper.sh /opt/preflight.py && python3 -c "import ast,io;[ast.parse(io.open(f,encoding='utf-8').read()) for f in ['/opt/preflight.py','/opt/detect_tokenizer.py','/opt/shims/ixformer/contrib/vllm/layers/__init__.py']]" && bash -n /opt/vllm_wrapper.sh && echo "[enginex-mlu370-compat] preflight syntax OK"
|
||
|
||
ENV PYTHONPATH=/opt/shims:${PYTHONPATH}
|
||
ENV MODEL_DIR=/model
|
||
|
||
# 关键一步：把 vllm 二进制换成 wrapper（真身改名 vllm_real），
|
||
# 平台下发的 `vllm serve ...` 因此流经本引擎的启动前修补。
|
||
RUN set -eux && VLLM_BIN="$(command -v vllm || echo /usr/local/corex/lib64/python3/dist-packages/bin/vllm)" && \
|
||
if [ ! -f "$VLLM_BIN" ]; then echo "vllm binary not found at $VLLM_BIN"; exit 1; fi && \
|
||
mv "$VLLM_BIN" "${VLLM_BIN}_real" && \
|
||
ln -s /opt/vllm_wrapper.sh "$VLLM_BIN" && \
|
||
echo "[enginex-mlu370-compat] vllm -> wrapper installed"
|