Thomas Viehmann
Thomas Viehmann
So regarding the priority, as discussed in slack: From what I can see, this cycle keeps modules going out of scope from being collected. Not nice, but for the most...
> It's a game-breaker because it blocks the usage of the Thunder-optimized dropout layer in a larger module as I would like to understand this more. Is it a game-breaker...
Hi @jjsjann123 What's up with this?
So currently, the prologue performs exactly two things: - collecting tensors (and possibly soon other inputs) for the computation trace - checking things I wonder if it would be good...
> Some computations are so closely related to "checks" that it would reduce the total CPU work to compute them in the prologue. > In general I still like thinking...
Note that #1500 added `remove_duplicate_number_proxies` to DCE because we prefer to have a single producer for proxies.
> "If we decide" -- this is really a you decision, Tom I'm glad we get a new way, but let's remove the old, too.
> A note for one of your requests, @t-vi: mixology is actually using some of the manual timer stuff that is in there. As such, I'd appreciate if we could...
I don't object to having it labeled bug, but I don't think typically users will hit it today: For things needing gradients, the augmented forward pass will fix it. I'm...
Yeah, @crcrpar, so this is why I mentioned better tooling, maybe using the visitor transform pattern more is the solution. Looking at the code, I would probably have the same...