Fix torch.load(..., weights_only=True) for NT (#112516)

Found when looking into #112509
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112516
Approved by: https://github.com/soulitzer
This commit is contained in:
Joel Schlosser
2023-10-31 13:18:38 -04:00
committed by PyTorch MergeBot
parent 85e93632e7
commit 51a38380d1
4 changed files with 9 additions and 3 deletions

View File

@ -2995,7 +2995,8 @@ class TestNestedTensorSubclass(TestCase):
@dtypes(torch.float, torch.double, torch.half)
@parametrize("requires_grad", [False, True])
def test_serialization(self, device, dtype, requires_grad):
@parametrize("weights_only", [False, True])
def test_serialization(self, device, dtype, requires_grad, weights_only):
def compare_metadata(nt1, nt2):
self.assertEqual(nt1._nested_tensor_size(), nt2._nested_tensor_size())
@ -3008,7 +3009,7 @@ class TestNestedTensorSubclass(TestCase):
buffer = io.BytesIO()
serialized = torch.save(a, buffer)
buffer.seek(0)
b = torch.load(buffer)
b = torch.load(buffer, weights_only=weights_only)
# should be both conceptually equal and metadata equivalent
self.assertEqual(a, b)
compare_metadata(a, b)

View File

@ -376,7 +376,7 @@ class Tensor(torch._C.TensorBase):
self._nested_tensor_strides(),
self._nested_tensor_storage_offsets(),
)
return (torch._nested_view_from_buffer, args_nested)
return (torch._utils._rebuild_nested_tensor, args_nested)
elif (
self.data_ptr() == 0
and type(self) is not torch.Tensor

View File

@ -304,6 +304,10 @@ def _rebuild_sparse_tensor(layout, data):
raise NotImplementedError(f"rebuilding sparse tensor for layout {layout}")
def _rebuild_nested_tensor(buffer, sizes, strides, storage_offsets):
return torch._nested_view_from_buffer(buffer, sizes, strides, storage_offsets)
def _rebuild_device_tensor_from_numpy(data, dtype, device, requires_grad):
tensor = torch.from_numpy(data).to(dtype=dtype, device=device)
tensor.requires_grad = requires_grad

View File

@ -101,6 +101,7 @@ def _get_allowed_globals():
torch._utils._rebuild_tensor_v2,
torch._utils._rebuild_sparse_tensor,
torch._utils._rebuild_meta_tensor_no_storage,
torch._utils._rebuild_nested_tensor,
]:
rc[f"torch._utils.{f.__name__}"] = f