Add frozen Thinker cache-path control

eval_omni_turn_lora.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-"""Evaluate Thinker LoRA + turn head from dense Omni AV caches.
+"""Evaluate a frozen or LoRA-adapted Thinker from dense Omni AV caches.
 
 The script accepts one dense cache directory per EasyCom session.  With
 ``torchrun``, complete sessions are assigned round-robin to ranks; rank zero
@@ -55,7 +55,14 @@ def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
         required=True,
         help="Dense cache directory from cache_omni_av_pieces.py; repeat per session",
     )
-    parser.add_argument("--adapter-checkpoint", type=Path, required=True)
+    parser.add_argument(
+        "--adapter-checkpoint",
+        type=Path,
+        help=(
+            "PEFT adapter directory. Omit it to evaluate the frozen base "
+            "Thinker as a cache-path control."
+        ),
+    )
     parser.add_argument("--head-checkpoint", type=Path, required=True)
     parser.add_argument("--output-dir", type=Path, required=True)
     parser.add_argument("--batch-size", type=int, default=2)
@@ -267,18 +274,30 @@ def _run(args: argparse.Namespace, run: DistributedRun) -> None:
 
     base_model, processor = _load_base_model(args, run)
     base_model.requires_grad_(False)
-    try:
-        from peft import PeftModel
-    except ImportError as exc:
-        raise RuntimeError("PEFT is required to evaluate the LoRA adapter") from exc
-    adapter_path = args.adapter_checkpoint.expanduser().resolve()
-    if not (adapter_path / "adapter_config.json").is_file():
-        raise FileNotFoundError(f"Missing adapter config: {adapter_path}")
-    language_model = PeftModel.from_pretrained(
-        base_model.model,
-        str(adapter_path),
-        is_trainable=False,
-    ).eval()
+    adapter_path = None
+    if args.adapter_checkpoint is None:
+        language_model = base_model.model.eval()
+        model_variant = "frozen_base_thinker"
+        print(
+            f"[rank {run.rank}] frozen control: no LoRA adapter loaded",
+            flush=True,
+        )
+    else:
+        try:
+            from peft import PeftModel
+        except ImportError as exc:
+            raise RuntimeError(
+                "PEFT is required to evaluate the LoRA adapter"
+            ) from exc
+        adapter_path = args.adapter_checkpoint.expanduser().resolve()
+        if not (adapter_path / "adapter_config.json").is_file():
+            raise FileNotFoundError(f"Missing adapter config: {adapter_path}")
+        language_model = PeftModel.from_pretrained(
+            base_model.model,
+            str(adapter_path),
+            is_trainable=False,
+        ).eval()
+        model_variant = "thinker_lora"
     language_model.config.use_cache = False
     model_dtype = next(language_model.parameters()).dtype
     head, head_payload = _load_head(args.head_checkpoint, run.device)
@@ -336,7 +355,10 @@ def _run(args: argparse.Namespace, run: DistributedRun) -> None:
                 "classes": list(CLASS_NAMES),
                 "prediction_layout": "[target_time, class, future_horizon]",
                 "model_id": args.model_id,
-                "adapter_checkpoint": str(adapter_path),
+                "model_variant": model_variant,
+                "adapter_checkpoint": (
+                    None if adapter_path is None else str(adapter_path)
+                ),
                 "head_checkpoint": str(args.head_checkpoint.expanduser().resolve()),
                 "head_training_step": head_payload.get("training_step"),
                 "cache_dirs": [str(item.cache_dir) for item in sessions],