Brian Hirsh comments

Results 150 comments of


                                            Brian Hirsh

[prototype, do not land] POC of torch.compiling wrapper tensor subclasses (torchquant)

> My question could be rephrased as "if my subclass does only functional things, do i really care if functionalization runs before or after it"? I think the problem is...

Custom tensor roll kernel produces wrong results for variations of block size and tl.constexpr

fwiw - I tried stepping through the kernel with interpreter mode, and noticed that the repro above passes when I use `TRITON_INTERPRET=1` (but fails without interpreter mode)

Batch Norm Consolidation

Overall looks good! I mentioned that a dummy PR to test that the new decomposition logic for batchnorm works might be a good idea, but two other tests that I...

Batch Norm Consolidation

Although I guess (1) is hard to test just in this PR, since I guess we'd need to change the `at::batch_norm` decomp to actually go to the new op first....

Support privateUser1 key in RNN op. (#118182)

@pytorchbot merge

add a test that non_overlapping checks dont generate too many guards

@pytorchbot merge

add a test that non_overlapping checks dont generate too many guards

@pytorchbot rebase

add a test that non_overlapping checks dont generate too many guards

@pytorchbot merge

add a test that non_overlapping checks dont generate too many guards

@pytorchbot rebase

add a test that non_overlapping checks dont generate too many guards

@pytorchbot merge