Fix GroundingDino loading: manually materialize meta tensors to support tied weights

models/detectors/grounding_dino.py

@@ -23,10 +23,28 @@ class GroundingDinoDetector(ObjectDetector):
             self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
         logging.info("Loading %s onto %s", self.MODEL_NAME, self.device)
         self.processor = GroundingDinoProcessor.from_pretrained(self.MODEL_NAME)
-        self.model = GroundingDinoForObjectDetection.from_pretrained(
-            self.MODEL_NAME,
-            low_cpu_mem_usage=False,
-        )
+        self.model = GroundingDinoForObjectDetection.from_pretrained(self.MODEL_NAME)
+        # Materialize ALL meta-device tensors before .to().
+        # Grounding DINO uses regex-based _tied_weights_keys to share bbox_embed
+        # layers, but the tying can leave meta tensors unresolved.
+        # named_parameters() deduplicates tied weights, so we must walk each
+        # module's _parameters/_buffers dicts directly to catch every reference.
+        _meta_count = 0
+        for module in self.model.modules():
+            for key, val in list(module._parameters.items()):
+                if val is not None and val.device.type == "meta":
+                    module._parameters[key] = torch.nn.Parameter(
+                        torch.empty(val.shape, dtype=val.dtype, device="cpu"),
+                        requires_grad=val.requires_grad,
+                    )
+                    _meta_count += 1
+            for key, val in list(module._buffers.items()):
+                if val is not None and val.device.type == "meta":
+                    module._buffers[key] = torch.empty(val.shape, dtype=val.dtype, device="cpu")
+                    _meta_count += 1
+        if _meta_count:
+            logging.warning("Materialized %d meta tensors in %s", _meta_count, self.MODEL_NAME)
+            self.model.tie_weights()
         self.model.to(self.device)
         self.model.eval()