fix `self.pe` init issue

usad_modules.py

@@ -367,6 +367,13 @@ class RelPositionalEncoding(nn.Module):
         self.pe = None
 
     def extend_pe(self, x: torch.Tensor) -> None:
+        # ``from_pretrained`` can initialize modules on the meta device while it
+        # streams checkpoint weights.  This runtime cache is not checkpoint
+        # state, so a meta tensor cannot be materialized with ``Tensor.to``.
+        # Drop it and rebuild it below for the actual input device instead.
+        if self.pe is not None and self.pe.is_meta:
+            self.pe = None
+
         if self.pe is not None:
             if self.pe.size(1) >= x.size(1) * 2 - 1:
                 if self.pe.dtype != x.dtype or self.pe.device != x.device: