x

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