Rahul Solanki
Rahul Solanki
> This pr makes already pretty complicated lazy execution code more confusing, I want to understand how much speed up(and in what circustance, always, small model, large model etc) are...
> Thanks, I will try to review it today. If you have the profile of before and after that would be super helpful. I didn't get a chance to get...
> @aws-rhsoln Unless you have a strong preference I will leave this pr out of the 1.13 release since that will be cut tmr. I will be out for a...
> > This pr makes already pretty complicated lazy execution code more confusing, I want to understand how much speed up(and in what circustance, always, small model, large model etc)...
> > > @alanwaketan yea, I think this change should also be upstreamed to LTC. > > > > > > Can you point me to the LTC code base,...
> > > > @alanwaketan yea, I think this change should also be upstreamed to LTC. > > > > > > > > > Can you point me to...
Here is the PR for lazy tensor core: https://github.com/pytorch/pytorch/pull/87119/files @alanwaketan
If this PR looks good, can we merge it? @JackCaoG
What I meant is, when `Activate()` is called, it would result in `Runner()` method getting called which in turn would call the HandleReleases as part of [this](https://github.com/pytorch/xla/blob/master/third_party/xla_client/triggered_task.cc#L67) . However, I...
I can create the PR for this. Thanks for checking in!