GB0371

Graph-Break Type

Short name describing what triggered the graph break

autograd.grad with output that requires grad

Context

Values or code snippet captured at the break point

context

Explanation

Explanation of why the graph break was triggered

The compiled function uses torch.autograd.grad() and returns a tensor that still requires gradients and is connected to the autograd.grad() computation. This would cause aot_autograd to attempt ‘backward through graph a second time’, which is not supported.

Hints

Hints on how to resolve the graph break

• Call .detach() on the output tensor before returning it from the compiled function.
• For example: return loss.detach(), grads instead of return loss, grads.
• Or set create_graph=False in autograd.grad() if you don’t need higher-order gradients.
• Your code may result in an error when running in eager. Please double check that your code doesn’t contain a similar error when actually running eager/uncompiled. You can do this by removing the torch.compile call, or by using torch.compiler.set_stance("force_eager").

Additional Information

Click here to add Additional Info

Back to Registry