Aditya Tewari

we've run in some issues with bf16 conv, with this change as is (after including the stubs in cpu_conv_list.cpp), following this PR : https://github.com/uxlfoundation/oneDNN/pull/3731 . We can raise followup a...

This is not necessarily a blocker for this to go in. The BF16 conv needs to be enabled explicitly in the `cpu_conv_list`. We can fix the related issues and push...

> Can you give details on why this would change the pass rate? That's something which I was wondering as well, turns out some models were getting timeout. I assume...

Looking into the failures. When built on NeoverseV1 the failures seem to pass on this HASH. There might be a discrepancy with `conda install` and the `install_openblas.sh` . I'll RCA...

> We want to make `OPENBLAS_VERSION` configurable in https://github.com/pytorch/pytorch/pull/150106/files, can we merge that first? Agreed. That should be merged before. I'll be testing how the new version tests out and...

To update on the failures, we run into segfault issue when `train=True`. We'll be debugging this further. tag: @taoye9

Update regarding the segfaults and SDPA failing tests (train=True): We observed intermittent segmentation faults after integrating the SBGEMM kernels for the backward pass, specifically occurring at this location: [CPUBlas.cpp](https://github.com/pytorch/pytorch/blob/eecaa0bbc6a39c89366f9ec97797479c27f6d760/aten/src/ATen/native/CPUBlas.cpp#L364). Additionally,...

The segfault which we were seeing earlier seems to be resolved now. Thanks @Ryo-not-rio . Although with the latest runs, seeing some extra failures which seems unrelated - related open...

@malfet do you think we can get a review for this? It would be great to merge this before `2.8 release` branch-cut. Note: With the latest runs, seeing some extra...

@pytorchbot rebase