Will Constable
Will Constable
this mostly lgtm, i didn't review carefully one thought is whether we want to (correctly) bump version major, or, just keep old seq_id and add new ones and later deprecate...
discussed offline, seems the intention was to get a way to run operations on a user stream that wait for particular comm operations to finish. There is actually no need...
@pytorchbot merge -f "intentional BC change for newly added (as yet unreleased) API"
can you explain the mechanism for throwing the cuda errors in the main thread? is it because any current cuda error on any stream/kernel will cause any future cpu synchronization...
didnt do a careful review yet but it looks great to me overall!
@pytorchbot merge
@pytorchbot merge -i
@pytorchbot merge -i
@pytorchbot merge
> If I understand correctly, the first item in the output tuple represents the logits. So, I filter this item and pass it to compute_loss. Can you tell me if...