vgrytsun

awesome, thank you! kwargs issue is still there even after https://github.com/pytorch/pytorch/pull/119874. not sure how hard it is to add kwargs support, but if it's super tricky one way around is...

you are right, realized that after i published the code. we would need to handle that case separately, solution below is getting a bit ugly, but seem to be working...

btw, on a related note: could we add per-layer attribution of the forward ops during backwards pass for activation checkpointing? I see that right now we do attribution only for...

@sentrivana thank you for a quick response! So the workaround is to disable `DedupeIntegration` (the only downside i can see is that each client will be reporting more events) ?...

Exception has a reference to traceback instance, which holds stack frames and their local variables. And those local variables could be arbitrary sizes depending on the application. So for the...