Tushar Jain
Tushar Jain
@ActionScripted I would say what I would love to see is something more than just a darker tabs, although this would also be nice to have. To give an analogy,...
@pytorchbot merge
@pytorchbot merge
@pytorchbot merge
+1 to this. Would be also good to document what kind of guarantees/invariants we're trying to ensure and some explanation on why the protocol is correct and ensures they are...
> What would be some of these guarantees/invariants? I'll have to think about this a bit 😅 The main invariant we'd want to prove is that all replicas have the...
@WarrenZhu050413 That looks good. I was more curious about the assumptions e.g. this is seems like a strong statement > Currently none of these guarantees hold if silent data corruption...
@WarrenZhu050413 With the current implementation, the example I mentioned is a counter-example to Synchronized Steps i.e. Synchronized Steps doesn't hold. There are some replicas, not participating in quorum, that would...
@WarrenZhu050413 do RG1 and RG2 keep waiting for RG0 to finish its training step? Then I wonder if waiting idle is better than the letting RG1 and RG2 to reprocess...