Move string vocabularies to CUDA as separate tensors

sdm/tensor/table.py

@@ -88,18 +88,12 @@ def _move_categories(
     if device is None or device.type == "cpu":
         return tuple(categories)
 
-    groups: dict[
-        tuple[type[Tensor], torch.dtype, torch.dtype | None], list[int]
-    ] = defaultdict(list)
+    # Copy strings separately: views of a shared copy recompile per batch.
+    groups: dict[torch.dtype, list[int]] = defaultdict(list)
     moved: list[Tensor | None] = [None] * len(categories)
     for i, category in enumerate(categories):
-        if type(category) is Tensor or isinstance(category, StringTensor):
-            offset_dtype = (
-                category.data_offset[1].dtype
-                if isinstance(category, StringTensor)
-                else None
-            )
-            groups[(type(category), category.dtype, offset_dtype)].append(i)
+        if type(category) is Tensor:
+            groups[category.dtype].append(i)
         else:
             moved[i] = category.to(device)
 

test/tensor/test_table.py

@@ -1533,6 +1533,7 @@ def test_from_pandas_nullable_and_categorical_dtypes(
     assert all(
         category.device == device for category in tensor.categorical.categories
     )
+    assert tensor.categorical.categories[1].storage_offset() == 0
     assert tensor.categorical.categories[0].tolist() == ["unused", "a", "b"]
     assert tensor.categorical.categories[1].tolist() == ["b", "a"]
     assert tensor.categorical.categories[2].tolist() == [True, False]