xiejiahao
05f1c3932f
All checks were successful
Docker Build and Push / docker (push) Successful in 1m11s
- detect_model.py picks TP/port/served name/eager/util from config.json - validate TP against the model config (dense global KV heads, MoE k % 32) - one apply_patch.py covers both variants, idempotent - Dockerfile stays generic: no ENV/WORKDIR/EXPOSE/ENTRYPOINT - run.sh/verify.sh follow the llama4-bi150 layout
42 lines
1.8 KiB
Docker
42 lines
1.8 KiB
Docker
# Gemma 4 family on Iluvatar BI-V150
|
||
#
|
||
# 一个镜像同时支持：
|
||
# - Gemma-4-31B-it （dense，TP4，4 卡，8095，eager 关）
|
||
# - GEV-26B-Decide （MoE，TP2，2 卡，8094，eager 开）
|
||
#
|
||
# 两个变体共用同一个基础镜像和同一组补丁，差别只在 TP / 端口 / served name /
|
||
# eager / 显存 / 并发数，全部由入口脚本读 config.json 后决定。
|
||
#
|
||
# 镜像只包含运行时依赖和补丁，不设置 ENV / WORKDIR / EXPOSE / ENTRYPOINT / CMD；
|
||
# 模型路径、端口、TP、GPU、挂载等全部由 run.sh / docker-compose.yml 在启动时传入。
|
||
#
|
||
# 构建：
|
||
# docker build -t gemma4-bi150:2026-10-27 .
|
||
#
|
||
# 运行（由 run.sh 传入 entrypoint 和全部参数）：
|
||
# ./run.sh /mnt/disk0/models/google/gemma-4-31B-it
|
||
# ./run.sh /mnt/disk0/models/autotrust/GEV-26B-Decide
|
||
#
|
||
# 补丁是针对下面这个精确基础镜像生成的；换基础镜像需要重新生成 diff。
|
||
# registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7
|
||
# (vLLM 0.17.0+corex.20260612070235)
|
||
|
||
FROM registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7
|
||
|
||
ARG IMAGE_VERSION=2026-10-27
|
||
|
||
LABEL org.opencontainers.image.title="gemma4-bi150" \
|
||
org.opencontainers.image.description="Gemma 4 31B-it / GEV-26B-Decide on Iluvatar BI-V150 with vLLM 0.17.0+corex" \
|
||
org.opencontainers.image.version="${IMAGE_VERSION}"
|
||
|
||
COPY apply_patch.py /opt/gemma4/apply_patch.py
|
||
COPY detect_model.py /opt/gemma4/detect_model.py
|
||
COPY entrypoint.sh /usr/local/bin/gemma4-entrypoint.sh
|
||
|
||
RUN chmod +x /usr/local/bin/gemma4-entrypoint.sh \
|
||
&& chmod +x /opt/gemma4/detect_model.py \
|
||
&& python3 /opt/gemma4/apply_patch.py
|
||
|
||
# 故意不设置 ENV / WORKDIR / EXPOSE / ENTRYPOINT / CMD：
|
||
# 镜像保持通用，启动方式由 run.sh 或 docker-compose.yml 决定。
|