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
|
2026-03-29 05:22:25 +08:00
|
|
|
container_name: rocm-vllm-openai
|
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"
|
|
|
|
|
# vLLM 相关配置
|
2026-04-05 02:37:16 +08:00
|
|
|
# VLLM_RPC_TIMEOUT: "300"
|
|
|
|
|
# VLLM_WORKER_MULTIPROC_METHOD: "spawn"
|
|
|
|
|
# PYTHONUNBUFFERED: "1"
|
|
|
|
|
# # 启用 AITER 加速（R9700 属于 gfx1201 架构）
|
2026-04-17 10:53:41 +08:00
|
|
|
# 核心修复: 禁用 AITER 的 FP8 MoE 后端，回退到 Triton 内核
|
|
|
|
|
VLLM_ROCM_USE_AITER: "0"
|
|
|
|
|
# 可选: 显式指定使用更稳定的 MoE 后端
|
|
|
|
|
VLLM_ROCM_MOE_BACKEND: "TRITON"
|
2026-04-17 00:09:11 +08:00
|
|
|
VLLM_ROCM_USE_AITER_MHA: "1"
|
|
|
|
|
VLLM_V1_USE_PREFILL_DECODE_ATTENTION: "0"
|
|
|
|
|
NCCL_MIN_NCHANNELS: "112"
|
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
|
