Add ROMA frozen backbone extraction support

streaming_turn_model_directly.py

@@ -1030,6 +1030,15 @@ def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
         default="open-stream-replay",
     )
     parser.add_argument("--model-id", default=core.MODEL_ID)
+    parser.add_argument(
+        "--backbone-loader",
+        choices=("qwen", "roma"),
+        default="qwen",
+        help=(
+            "qwen loads a Thinker checkpoint directly; roma loads the full "
+            "ROMA checkpoint and keeps only its frozen Thinker"
+        ),
+    )
     parser.add_argument("--start", type=float, default=0.0)
     parser.add_argument("--end", type=float)
     parser.add_argument("--step-seconds", type=float, default=0.2)
@@ -1360,8 +1369,15 @@ def main(
             dtype=args.dtype,
             device_map={"": args.device},
             use_flash_attention_2=args.flash_attention_2,
+            backbone_loader=args.backbone_loader,
         )
     else:
+        loaded_backbone = core.backbone_metadata(model).get("loader")
+        if loaded_backbone not in {"unknown", args.backbone_loader}:
+            raise RuntimeError(
+                "Preloaded backbone does not match CLI provenance: "
+                f"model={loaded_backbone}, requested={args.backbone_loader}"
+            )
         model.eval()
     model.requires_grad_(False)
     media_path_for_api = reader.clips[0].path
@@ -1447,6 +1463,8 @@ def main(
     )
     payload: dict[str, Any] = {
         "model_id": args.model_id,
+        "backbone_loader": args.backbone_loader,
+        "backbone_metadata": core.backbone_metadata(model),
         "mode": args.mode,
         "stream_representation": (
             "single_causal_prefix"