Auto-detect Scout/Maverick from config.json and add verify script

This commit is contained in:
2026-10-26 16:59:44 +08:00
parent 13f14a12e9
commit 738d9f80e4
10 changed files with 816 additions and 235 deletions

View File

@@ -4,14 +4,15 @@
# - Llama-4-Scout-17B-16E-Instruct(BF16,TP8,8 卡)
# - Llama-4-Maverick-17B-128E-Instruct-FP8(FP8,TP8+PP2,16 卡)
#
# 模型权重不打进镜像,运行时按 run.sh 的 scout / maverick 预设挂载。
# 模型权重不打进镜像,运行时挂载模型目录;容器启动后读取 config.json
# 自动判断 Scout / Maverick,并选择对应的 TP / PP / 端口。
#
# 构建:
# docker build -t llama4-bi150:2026-10-26 .
#
# 运行:
# ./run.sh scout
# ./run.sh maverick
# ./run.sh /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct
# ./run.sh /mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8
#
# 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
@@ -27,17 +28,16 @@ LABEL org.opencontainers.image.title="llama4-bi150" \
COPY patches /opt/llama4/patches
COPY apply_patch.py /opt/llama4/apply_patch.py
COPY detect_model.py /opt/llama4/detect_model.py
COPY entrypoint.sh /usr/local/bin/llama4-entrypoint.sh
RUN chmod +x /usr/local/bin/llama4-entrypoint.sh \
&& chmod +x /opt/llama4/detect_model.py \
&& python3 /opt/llama4/apply_patch.py
ENV MODEL_PATH=/models/llama4 \
SERVED_NAME=llama4 \
AUTO_DETECT=1 \
HOST=0.0.0.0 \
PORT=8091 \
TP=8 \
PP=1 \
MAX_MODEL_LEN=4096 \
MAX_NUM_SEQS=1 \
GPU_MEM_UTIL=0.90 \