Short name describing what triggered the graph break
autograd.Function.apply: invalid set_output_grad_dtype
Values or code snippet captured at the break point
declarations {ctx.output_grad_dtypes} for {num_grad_args} outputs
Explanation of why the graph break was triggered
ctx.set_output_grad_dtype must declare one dtype per forward output, and None for non-tensor outputs.
Hints on how to resolve the graph break