Files

40 lines
1.1 KiB
YAML
Raw Permalink Normal View History

2026-03-29 05:06:19 +08:00
services:
2026-03-29 16:49:09 +08:00
vllm:
2026-03-29 05:22:25 +08:00
build:
context: .
dockerfile: Dockerfile
2026-04-23 13:58:44 +08:00
pull: false
2026-04-16 00:15:55 +08:00
image: vllm-openai-rocm:nightly
2026-04-23 14:55:22 +08:00
container_name: vllm-openai-rocm
2026-03-29 16:32:55 +08:00
entrypoint: ["python"]
command: ["-m", "app.start_openai"]
2026-03-29 05:22:25 +08:00
ports:
- "8001:8001"
volumes:
- /opt/model:/opt/model:ro
2026-03-29 05:35:20 +08:00
- ./config.json:/workspace/config.json:ro
2026-03-30 22:43:23 +08:00
environment:
HF_HOME: /opt/model
2026-04-05 02:08:49 +08:00
# 严格离线模式,禁止任何网络下载
HF_HUB_OFFLINE: "1"
TRANSFORMERS_OFFLINE: "1"
HF_DATASETS_OFFLINE: "1"
2026-05-18 17:46:07 +08:00
VLLM_ROCM_USE_AITER: "0" # 打开总开关
2026-04-23 15:42:35 +08:00
# 明确禁用不稳定的MoE后端,让其他AITER优化(如MHA)生效
VLLM_ROCM_USE_AITER_MOE: "0"
VLLM_ROCM_MOE_BACKEND: "TRITON" # 强制MoE使用Triton
2026-05-18 17:46:07 +08:00
TORCHINDUCTOR_FX_GRAPH_CACHE: "0" # 禁用 torch.compile 缓存
VLLM_DISABLE_COMPILE_CACHE: "1"
2026-03-29 05:06:19 +08:00
devices:
- /dev/kfd
- /dev/dri
group_add:
- video
ipc: host
2026-03-31 23:36:56 +08:00
shm_size: 32g
2026-03-29 05:06:19 +08:00
cap_add:
- SYS_PTRACE
security_opt:
- seccomp=unconfined
restart: unless-stopped