"Tensors returned from custom ops (1) must not be inputs to the custom op and (2) may not alias any inputs or other returns. Please clone the the offending output tensors (e.g. output.clone()) or refactor your code. Offending op: flash_attn_bwd "

Someone interpreted this to mean that an output of a custom op cannot be an input to another custom op, which is not what I meant when I wrote this.

cc @ezyang @chauhang @penguinwu @bdhirsh