x

README.md

@@ -61,6 +61,8 @@
 - `api_key`：接口访问密钥（同时用于 8000 与 8001）
 - `tensor_parallel_size`：张量并行数，双卡建议 `2`
 - `dtype`：推理精度，默认 `bfloat16`
+- `model_root`：本地模型根目录，建议 `/opt/model`
+- `offline_mode`：保留字段，当前实现固定只走离线本地模型
 - `models.selected`：当前生效模型，留空时回退到 `models.default`
 
 ## config.json 说明
@@ -70,7 +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
-- 每个模型至少建议包含：`hf_model_id`、`ctx`、`max_num_seqs`、`max_tokens`、`gpu_util`、`valid_tp`
+- 每个模型必须包含：`local_path`、`ctx`、`max_num_seqs`、`max_tokens`、`gpu_util`、`valid_tp`
 
 启动时会按以下优先级选模型：
 
@@ -79,7 +81,7 @@
 
 模型被选中后，会自动覆盖运行参数，包括：
 
-- `model_name` ← `hf_model_id`
+- `model_name` ← `local_path`（相对路径会自动拼接 `model_root`）
 - `max_model_len` ← `ctx`
 - `max_num_seqs` ← `max_num_seqs`
 - `max_tokens` ← `max_tokens`
@@ -100,19 +102,19 @@ docker compose up -d --build
 3. 验证自定义推理服务：
 
 ```bash
-curl http://localhost:8000/health
+curl http://localhost:<services.api.port>/health
 ```
 
 4. 验证 OpenAI 协议服务：
 
 ```bash
-curl http://localhost:8001/v1/models
+curl http://localhost:<services.openai.port>/v1/models
 ```
 
 ## 推理请求示例
 
 ```bash
-curl -X POST "http://localhost:8000/v1/generate" \
+curl -X POST "http://localhost:<services.api.port>/v1/generate" \
   -H "Content-Type: application/json" \
   -d "{\"prompt\":\"请用三句话介绍大模型推理优化\",\"max_tokens\":128,\"temperature\":0.7}"
 ```
@@ -132,6 +134,7 @@ curl -X POST "http://localhost:8001/v1/chat/completions" \
 - API Key 使用 `config.json` 中 `api_key`
 - 模型名使用 `config.json` 中 `models.profiles.<模型名>.served_model_name`
 - 若使用工具调用，`config.json` 中应配置 `tool_call_parser` 与 `enable_auto_tool_choice`
+- 服务强制离线模式，不会回退到 Hugging Face 远程下载
 - 所有路径按 Ubuntu 规范填写，本地模型建议使用 `/opt/model/<模型目录>`
 
 ## 双 AMD R9700 调优建议

app/config.py

@@ -17,6 +17,8 @@ class Settings(BaseModel):
     port: int = 8000
     openai_host: str = "0.0.0.0"
     openai_port: int = 8001
+    model_root: str = "/opt/model"
+    offline_mode: bool = True
     max_model_len: int = 8192
     gpu_memory_utilization: float = 0.92
     tensor_parallel_size: int = 2
@@ -42,6 +44,8 @@ def get_settings() -> Settings:
         port=runtime["port"],
         openai_host=runtime["openai_host"],
         openai_port=runtime["openai_port"],
+        model_root=runtime["model_root"],
+        offline_mode=runtime["offline_mode"],
         api_key=runtime["api_key"],
         tensor_parallel_size=runtime["tensor_parallel_size"],
         dtype=runtime["dtype"],

app/model_catalog.py

@@ -29,6 +29,27 @@ def _to_float(value: Any, default: float) -> float:
         return default
 
 
+def _to_str(value: Any, default: str = "") -> str:
+    return str(value).strip() if value is not None else default
+
+
+def _join_posix(base_path: str, suffix_path: str) -> str:
+    return f"{base_path.rstrip('/')}/{suffix_path.lstrip('/')}"
+
+
+def _resolve_profile_model_path(profile: dict[str, Any], model_root: str, model_key: str) -> str:
+    local_path = _to_str(profile.get("local_path"))
+    if not local_path:
+        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)
+
+
 def load_catalog(catalog_path: str = "config.json") -> dict[str, Any]:
     content = json.loads(Path(catalog_path).read_text(encoding="utf-8"))
     if not isinstance(content, dict):
@@ -50,6 +71,8 @@ def resolve_runtime_settings(content: dict[str, Any]) -> dict[str, Any]:
         "tensor_parallel_size": _to_int(content.get("tensor_parallel_size"), 2),
         "dtype": str(content.get("dtype", "bfloat16")),
         "revision": str(content.get("revision", "")).strip() or None,
+        "model_root": _to_str(content.get("model_root"), "/opt/model"),
+        "offline_mode": True,
         "model_key": str(models.get("selected", "")).strip() or None,
     }
 
@@ -66,6 +89,7 @@ def resolve_model_profile(
     profile = profiles[model_key]
     if not isinstance(profile, dict):
         raise ValueError(f"model profile '{model_key}' must be a JSON object")
+    model_root = _to_str(content.get("model_root"), "/opt/model")
     valid_tp_raw = profile.get("valid_tp", [])
     valid_tp = [_to_int(item, 0) for item in valid_tp_raw if _to_int(item, 0) > 0]
     resolved_tp = requested_tp
@@ -73,7 +97,7 @@ def resolve_model_profile(
         resolved_tp = valid_tp[0]
     updates = {
         "selected_model": model_key,
-        "model_name": profile.get("hf_model_id", model_key),
+        "model_name": _resolve_profile_model_path(profile, model_root, model_key),
         "served_model_name": profile.get("served_model_name", model_key),
         "max_model_len": _to_int(profile.get("ctx"), 8192),
         "max_num_seqs": _to_int(profile.get("max_num_seqs"), 64),
@@ -86,4 +110,6 @@ def resolve_model_profile(
         "enable_auto_tool_choice": _to_bool(profile.get("enable_auto_tool_choice"), False),
     }
     env_vars = {str(k): str(v) for k, v in dict(profile.get("env", {})).items()}
+    env_vars["HF_HUB_OFFLINE"] = "1"
+    env_vars["TRANSFORMERS_OFFLINE"] = "1"
     return model_key, updates, env_vars

config.json

@@ -12,12 +12,15 @@
   "api_key": "sk-szcjw",
   "tensor_parallel_size": 2,
   "dtype": "bfloat16",
+  "model_root": "/opt/model",
+  "offline_mode": true,
   "revision": "",
   "models": {
     "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
     "selected": "Qwen3.5-35B-A3B-GPTQ-Int4",
     "profiles": {
       "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
+        "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
         "ctx": "24576",
         "trust_remote": true,
         "valid_tp": [2],
@@ -34,6 +37,7 @@
         "hf_model_id": "cpatonn/Qwen3-Next-80B-A3B-Instruct-AWQ-4bit"
       },
       "GLM-4.7-Flash-AWQ": {
+        "local_path": "GLM-4.7-Flash-AWQ",
         "ctx": "32768",
         "trust_remote": true,
         "valid_tp": [1, 2],
@@ -46,6 +50,7 @@
         "hf_model_id": "THUDM/GLM-4.7-Flash-AWQ"
       },
       "Qwen3.5-27B-FP8": {
+        "local_path": "Qwen3.5-27B-FP8",
         "ctx": "32768",
         "trust_remote": true,
         "valid_tp": [1, 2],
@@ -58,6 +63,7 @@
         "hf_model_id": "RedHatAI/Qwen3.5-27B-FP8-dynamic"
       },
       "Qwen3.5-35B-A3B-GPTQ-Int4": {
+        "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
         "ctx": "32768",
         "trust_remote": true,
         "valid_tp": [1, 2],

docker-compose.yml

@@ -7,7 +7,7 @@ services:
     container_name: rocm-vllm-inference
     command: ["python", "-m", "app.start_api"]
     ports:
-      - "8000:8000"
+      - "8002:8002"
     volumes:
       - /opt/model:/opt/model:ro
       - ./config.json:/workspace/config.json:ro