Nikita Shulga

@junukwon7 I don't know the exact details, but I assume using 32-bit indexes results in faster kernels, as one can perform twice as much 32-bit operations per one SIMD instruction...

IMO we should never abort force merges: if PR has the right approvals then merge could be forced. Also, this could result in an outage if HUD/RockSet is down for...

We should stop using batch files as there are no way to do reliable way to abort on errors in those.. I've filed an issue about it few years back....

Should we remove torchdata integration with torchtext instead? And indeed, it's a TorchText release blocking issue, but it should not delay the release of PyTorch itself

Closing this one, as DirectML were implemented using PrivateUse as suggested in https://github.com/pytorch/pytorch/pull/85040#issuecomment-1254626017 see https://learn.microsoft.com/en-us/windows/ai/directml/gpu-pytorch-windows

I wonder if `ptxas` have any additional optimization passes / compile time wins compared to nvrtc?

@pytorchbot revert -m "eca_halonext26ts started failing after 2nd PR from the stack was landed, see https://hud.pytorch.org/pytorch/pytorch/commit/88b3810c94b45f5982df616e2bc4c471d173f491, not sure which one of the two caused it" -c weird

So looks like this PR caused a regression. And I guess the classification for both should be `nosignal` - inductor tests are no run on PR that do not add...

> @malfet Do you have some guidance about the windows build fail? A bit confused why I was assigned a reviewer, but sure. Build fails with the following errors: ```...

@milesial did you close because NVTX3 was finally integrated everywhere? Can you please confirm that it works as expected for you?