Materialize model before checkpoint load

src/evaluator.py

@@ -76,7 +76,31 @@ def _set_module_attr(root, dotted_name, value):
     parts = dotted_name.split(".")
     for part in parts[:-1]:
         module = getattr(module, part)
-    setattr(module, parts[-1], value)
+    leaf = parts[-1]
+    if isinstance(value, torch.nn.Parameter) and leaf in getattr(module, "_parameters", {}):
+        module._parameters[leaf] = value
+    elif leaf in getattr(module, "_buffers", {}):
+        module._buffers[leaf] = value
+    else:
+        setattr(module, leaf, value)
+
+
+def _has_meta_tensors(model) -> bool:
+    for _, param in model.named_parameters():
+        if getattr(param, "is_meta", False):
+            return True
+    for _, buf in model.named_buffers():
+        if getattr(buf, "is_meta", False):
+            return True
+    return False
+
+
+def _ensure_materialized_before_load(model, device):
+    """Force a real tensor skeleton before checkpoint assignment in Spaces."""
+    if _has_meta_tensors(model):
+        logger.warning("Model skeleton contains meta tensors before checkpoint load; using to_empty().")
+        model = model.to_empty(device=device)
+    return model
 
 
 def _materialize_meta_buffers(model, device):
@@ -166,6 +190,7 @@ def load_meta_acsa(checkpoint_dir: Path = None, meta_encoder: Optional[MetaEncod
     else:
         model = BertMetaFusionACSAModel(bert_name=bert_name, meta_in_dim=meta_in_dim)
     model = _move_model_to_device(model, device)
+    model = _ensure_materialized_before_load(model, device)
     _load_model_state(model, ckpt["model_state_dict"], strict=False)
     model = _refresh_runtime_buffers(model, device)
     model.eval()