Method tensor

torch/_dynamo/debug_utils.py:562–593

(
        self,
        storage,
        shape,
        stride=None,
        *,
        storage_offset=None,
        dtype=None,
        requires_grad=None,
        is_leaf=None,
        **metadata,
    )

        return rand_strided(shape, stride, dtype_hint, device).untyped_storage()

    def tensor(
        self,
        storage,
        shape,
        stride=None,
        *,
        storage_offset=None,
        dtype=None,
        requires_grad=None,
        is_leaf=None,
        **metadata,
    ):
        stride = _stride_or_default(stride, shape=shape)
        storage_offset = _storage_offset_or_default(storage_offset)
        dtype = _dtype_or_default(dtype)
        is_leaf = _is_leaf_or_default(is_leaf)
        requires_grad = _requires_grad_or_default(requires_grad)
        t = torch.tensor(
            [], dtype=dtype, device=storage.device, requires_grad=requires_grad
        )
        with torch.no_grad():
            t.set_(storage, storage_offset, shape, stride)
        if not is_leaf:
            # Fake up some autograd history in a very naughty way
            with torch.enable_grad():
                t = t.clone(memory_format=torch.preserve_format)
            with torch.no_grad():
                t.set_(storage, storage_offset, shape, stride)
        assert torch._subclasses.meta_utils.safe_is_leaf(t) == is_leaf
        torch._utils.set_tensor_metadata(t, metadata)
        self.args.append(t)
        return t  # for BC

    def symint(self, val):
        self.args.append(val)