One of the Fx graph generated by dynamo for hf_Bart: the graph create tensors without specifying device

gistfile1.txt

graph():
    %full : [#users=2] = call_function[target=torch.full](args = ((512, 512), -inf), kwargs = {})
    %arange : [#users=2] = call_function[target=torch.arange](args = (512,), kwargs = {})
    %add : [#users=1] = call_function[target=operator.add](args = (%arange, 1), kwargs = {})
    %view : [#users=1] = call_method[target=view](args = (%add, 512, 1), kwargs = {})
    %lt : [#users=1] = call_function[target=operator.lt](args = (%arange, %view), kwargs = {})
    %masked_fill_ : [#users=0] = call_method[target=masked_fill_](args = (%full, %lt, 0), kwargs = {})
    %to : [#users=1] = call_method[target=to](args = (%full, torch.float32), kwargs = {})
    %getitem : [#users=1] = call_function[target=operator.getitem](args = (%to, (None, None, slice(None, None, None), slice(None, None, None))), kwargs = {})
    %expand : [#users=1] = call_method[target=expand](args = (%getitem, 1, 1, 512, 512), kwargs = {})
    return (expand,)