fix: ensure saved models tensors are contiguous

edsnlp/pipes/trainable/embeddings/transformer/transformer.py

@@ -207,7 +207,20 @@ def to_disk(self, path, *, exclude: Optional[Set[str]]):
         if repr_id in exclude:
             return
         self.tokenizer.save_pretrained(path)
+
+        # Fix for https://github.com/aphp/edsnlp/issues/317
+        old_params_data = {}
+        for param in self.transformer.parameters():
+            if not param.is_contiguous():
+                old_params_data[param] = param.data
+                param.data = param.data.contiguous()
+
         self.transformer.save_pretrained(path)
+
+        # Restore non-contiguous tensors
+        for param, data in old_params_data.items():
+            param.data = data
+
         for param in self.transformer.parameters():
             exclude.add(object.__repr__(param))
         cfg = super().to_disk(path, exclude=exclude) or {}