sanchitintel

Hi @valentine233, just a heads up - it looks like Inductor tracing has changed a bit over the last few weeks. A few weeks back (I'm unable to figure what...

Hi @Valentine233, it turns out that in `GPT2`, when batch size is greater than 1, then `clone` is present after `permute` and `expand` nodes for `query`, `key`, and `value`. If...

Thanks for the info, @Valentine233! I should follow the same approach as you followed in your PR, and keep both GPT2 patterns (batch size 1 and batch size >1) in...

Hi @igorsafo, since GC MHA patterns have a lot of variation, what would such an API look like? Thanks!

Thanks for the info, @igorsafo! PyTorch graph path has now moved to matching patterns on the framework side. So, this approach requires oneDNN Graph patterns to be hard-coded, and a...

Hi @igorsafo, I agree that such an approach requires more work on the framework side since oneDNN Graph partitions need to be hardcoded in the framework with such an implementation....

Hi @xiang1guo, we'd need this API in public oneDNN to support generic dynamic shape supporting kernels in oneDNN Graph. We're fine with this initial implementation, but down the line, we...

Hi @ElaineBao, can framework developers tell if custom patterns that don't already exist with `fusion` policy (but can be added by a backend as predefined patterns) would perform better or...

@xiang1guo, we plan to emulate dynamic shape support in IPEX with multiple compiled partitions for various static shapes. Since the `num_splits` attribute is to be provided at the time of...

> by saying `then it might change for a later input`, do you mean that `num_splits` will be no longer an attribute? How is it handled in bridge? Hi @wuxun-zhang,...