clef_mlx.py: use the repo name (not the cache snapshot hash) as the default model name

clef_mlx.py

@@ -590,6 +590,14 @@ def _default_model() -> str | None:
     return str(here) if (here / "joint_head.safetensors").exists() else None
 
 
+def _model_name(model: str) -> str:
+    """Short name for responses: 'clef-flash-4bit' for a repo id, a local dir, or a Hub cache snapshot."""
+    path = Path(model)
+    if path.parent.name == "snapshots" and path.parent.parent.name.startswith("models--"):
+        return path.parent.parent.name.split("--")[-1]
+    return path.name
+
+
 def _json_arg(value: str) -> Any:
     """A JSON string, a path to a JSON file, or '-' for stdin."""
     if value == "-":
@@ -646,7 +654,7 @@ def _cmd_predict(args) -> None:
         if args.state is None or args.questions is None:
             sys.exit("predict needs --state and --questions, or --request")
         request = {"state": _state_arg(args.state), "questions": _json_arg(args.questions)}
-    request.setdefault("model", args.name or Path(args.model).name)
+    request.setdefault("model", args.name or _model_name(args.model))
     if args.image:
         from PIL import Image
 
@@ -665,7 +673,7 @@ def _cmd_serve(args) -> None:
     from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 
     model = load(args.model)
-    served_name = args.name or Path(args.model).name
+    served_name = args.name or _model_name(args.model)
     model.systemone({"model": served_name, "state": "warmup",
                      "questions": {"w": {"type": "noul", "instructions": "Is this a warmup?"}}})
     lock = threading.Lock()  # one GPU: run requests one at a time