Philipp Hack

CC @reedwm.

CC @reedwm.

Yes, only the contracting dimensions should have changed.

@frgossen the patterns are FP8 windowed einsums with input dequantizations (i.e. type-conversion to a wider type like FP16 or BF16 and scaling) and all-gathers that have multiple dot users. The...

@frgossen can you PTAL?

@frgossen the motivation for this change is the extension of the existing functionality for collective matmuls with multiple all-gather dots in the windowed einsum handler (see [this comment](https://github.com/openxla/xla/blob/e56c9c1c51cef21cda50eddc45f2a115836a9358/xla/service/gpu/gpu_windowed_einsum_handler.cc#L676)) to support...

@frgossen for some reason I can't directly respond to these two comments: > Do you know why they interfer with the outcome? Is there a fundamental reason or can this...

> If this pass is so tightly coupled with the dot handler, shouldn't there be tests that reflect this? Do you think it makes sense to add test cases that...

> Can you confirm whether I'm correct That's a generally accurate description.