This commit is contained in:
2026-03-29 16:19:26 +08:00
parent 289145e901
commit 56a5c2e4c0
7 changed files with 19 additions and 23 deletions
+1 -1
View File
@@ -1,4 +1,4 @@
FROM rocm/vllm:rocm7.12.0_gfx120X-all_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16.0
FROM docker.1ms.run/vllm/vllm-openai-rocm:latest
WORKDIR /workspace
+6 -15
View File
@@ -1,11 +1,11 @@
# ROCm vLLM 容器化推理项目
基于镜像 `rocm/vllm:rocm7.12.0_gfx120X-all_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16.0` 的 Python 推理服务,适配双 AMD R9700 32G GPU。
基于镜像 `docker.1ms.run/vllm/vllm-openai-rocm:latest` 的 Python + vLLM 推理服务,适配双 AMD R9700 32G GPU。
## 项目目标
- 提供可容器化部署的模型推理 API
- 使用 vLLM 在双 GPU 上进行张量并行推理
- 使用 vLLM + ROCm 在 AMD GPU 上执行推理
- 提供健康检查、鉴权和参数化配置能力
- 暴露 `8001` OpenAI 标准协议接口,兼容 OpenClaw 调用
@@ -72,7 +72,7 @@
- `models.default`:默认模型名
- `models.selected`:当前生效模型名
- `models.profiles`:模型配置集合
- 每个模型必须包含:`local_path`、`ctx`、`max_num_seqs`、`max_tokens`、`gpu_util`、`valid_tp`
- 每个模型必须包含:`local_path`,并建议补充 `ctx`、`max_num_seqs`、`max_tokens`
启动时会按以下优先级选模型:
@@ -146,15 +146,6 @@ curl -X POST "http://localhost:8001/v1/chat/completions" \
## 常见故障排查
- 报错 `model type qwen3_5_moe ... Transformers does not recognize this architecture` 时,先重建镜像以更新依赖:
```bash
docker compose build --no-cache
docker compose up -d --force-recreate
```
- 如果看到 `No services to build`,说明你执行的是 `docker compose up -d --force-recreate` 且未触发重建;必须先单独执行 `docker compose build --no-cache`。
- 构建阶段若出现 `git clone https://github.com/huggingface/transformers.git` 失败,请改用 PyPI 版本依赖(项目已默认 `transformers>=4.57.0,<5`),避免源码拉取中断导致构建失败。
- 若仍报相同错误,说明当前 `vLLM/Transformers` 组合不支持该模型架构,建议在 `config.json` 切到其他本地模型,或升级基础镜像到更新的 vLLM 版本。
- 报错 `model type ... Transformers does not recognize this architecture` 时,说明当前模型与镜像内依赖不兼容,建议更换模型或升级镜像版本。
- 若模型目录存在但仍加载失败,检查挂载路径是否为 `/opt/model:/opt/model:ro`,并确认容器内可见模型文件。
- 如果看到 `No services to build`,说明未触发重建;需要先执行 `docker compose build --no-cache` 再 `up`。
+3
View File
@@ -22,6 +22,9 @@ class InferenceEngine:
revision=settings.revision,
)
def close(self) -> None:
return None
def generate(self, req: GenerateRequest) -> GenerateResponse:
sampling_params = SamplingParams(
temperature=req.temperature,
+2
View File
@@ -25,6 +25,8 @@ async def lifespan(_: FastAPI):
settings = get_settings()
engine = InferenceEngine(settings)
yield
if engine is not None:
engine.close()
engine = None
+6 -5
View File
@@ -43,11 +43,12 @@ def _resolve_profile_model_path(profile: dict[str, Any], model_root: str, model_
raise ValueError(f"model profile '{model_key}' must provide local_path")
if "://" in local_path:
raise ValueError(f"model profile '{model_key}' local_path must be local filesystem path")
if local_path.startswith("/"):
return local_path
if not model_root:
raise ValueError("config.json model_root cannot be empty when local_path is relative")
return _join_posix(model_root, local_path)
resolved = local_path
if not local_path.startswith("/"):
if not model_root:
raise ValueError("config.json model_root cannot be empty when local_path is relative")
resolved = _join_posix(model_root, local_path)
return resolved
def load_catalog(catalog_path: str = "config.json") -> dict[str, Any]:
+1 -1
View File
@@ -18,9 +18,9 @@ def build_command() -> list[str]:
os.environ[key] = value
host = str(runtime["openai_host"])
port = str(runtime["openai_port"])
api_key = runtime["api_key"] or ""
dtype = str(runtime["dtype"])
revision = runtime["revision"] or ""
api_key = runtime["api_key"] or ""
cmd = [
sys.executable,
"-m",
-1
View File
@@ -1,4 +1,3 @@
fastapi==0.116.1
uvicorn==0.35.0
pydantic==2.11.7
transformers>=4.57.0,<5