GB4927

Graph-Break Type

Short name describing what triggered the graph break

autograd.grad consumed returned tensor’s grad_fn

Context

Values or code snippet captured at the break point

Leaked output tensors: {leaked_str}

Explanation

Explanation of why the graph break was triggered

torch.autograd.grad() consumes grad_fns that are needed by tensors returned from this compiled function. This would cause ‘backward through graph a second time’ errors. The following returned tensors have consumed grad_fns: {leaked_str}

Hints

Hints on how to resolve the graph break

• Detach the problematic tensor(s) before returning: e.g. {leaked[0]}.detach()
• Call .detach() on the tensor before returning.
• If you need to backward through the returned tensor, use retain_graph=True in autograd.grad().

Additional Information

Click here to add Additional Info

Back to Registry