52 lines
1.8 KiB
Docker
52 lines
1.8 KiB
Docker
# Llama 4 Scout on Iluvatar BI-V150
|
||
#
|
||
# 基础镜像就是当前 ugly_hopper / test 用的 CoreX vLLM 镜像。
|
||
# 这个 Dockerfile 做三件事：
|
||
# 1. 把 CoreX vLLM 的 Llama4 chunked-attention 补丁打进镜像
|
||
# 2. 安装 entrypoint（容器启动时再次幂等检查补丁）
|
||
# 3. 默认以 TP8 / 4096 ctx / 0.90 显存启动 vLLM
|
||
#
|
||
# 构建：
|
||
# docker build -t llama4-scout:2026-10-23 .
|
||
#
|
||
# 运行（8 卡）：
|
||
# docker run -d --name llama4-scout \
|
||
# --network host --ipc host \
|
||
# --device=/dev/iluvatar0 ... --device=/dev/iluvatar7 \
|
||
# -v /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct:/models/llama4-scout:ro \
|
||
# --ulimit memlock=-1 --ulimit stack=67108864 \
|
||
# llama4-scout:2026-10-23
|
||
#
|
||
# 说明：
|
||
# - 不需要 --privileged；只需要 GPU 设备节点 + host IPC/network。
|
||
# - 如果厂商 runtime 要求整块 /dev，可用 -v /dev:/dev 代替 --device 列表。
|
||
# - 模型目录只读挂载，容器内不写模型。
|
||
|
||
FROM registry.iluvatar.com.cn:10443/customer/sz/vllm0.17.0-4.4.0-x86:v7
|
||
|
||
LABEL org.opencontainers.image.title="llama4-scout-bi150" \
|
||
org.opencontainers.image.description="Llama 4 Scout 17B-16E Instruct on Iluvatar BI-V150 with vLLM 0.17.0+corex" \
|
||
org.opencontainers.image.version="2026-10-23"
|
||
|
||
COPY apply_patch.py /opt/llama4/apply_patch.py
|
||
COPY entrypoint.sh /usr/local/bin/llama4-entrypoint.sh
|
||
COPY vllm_llama4_chunked_attn.patch /opt/llama4/vllm_llama4_chunked_attn.patch
|
||
|
||
RUN chmod +x /usr/local/bin/llama4-entrypoint.sh \
|
||
&& python3 /opt/llama4/apply_patch.py
|
||
|
||
ENV MODEL_PATH=/models/llama4-scout \
|
||
SERVED_NAME=llama4-scout \
|
||
HOST=0.0.0.0 \
|
||
PORT=8091 \
|
||
TP=8 \
|
||
MAX_MODEL_LEN=4096 \
|
||
MAX_NUM_SEQS=1 \
|
||
GPU_MEM_UTIL=0.90 \
|
||
PYTHONUNBUFFERED=1
|
||
|
||
WORKDIR /opt/llama4
|
||
EXPOSE 8091
|
||
|
||
ENTRYPOINT ["/usr/local/bin/llama4-entrypoint.sh"]
|