Ilia Sergachev
Ilia Sergachev
I already considered making the change smaller and am still open to ideas how to do that. I hope the description helps navigate it during review. The MHA config /...
clang-format check is unhappy because of deleted files, I didn't find any actual formatting problems.
https://github.com/openxla/xla/pull/13108 was [reverted](https://github.com/openxla/xla/commit/f230ce20f0447b893d9365888d46e517df39ac16). --xla_gpu_shard_autotuning=false disables sharding of autotuning, not the autotuning itself.
I can reproduce with jax==0.4.31 and --xla_gpu_shard_autotuning=false helps - looks like https://github.com/openxla/xla/pull/13108 got into this JAX release before it got reverted. Thank you for cc'ing me, I'll investigate why does...
I sent a [fix](https://github.com/openxla/xla/pull/16153) to XLA which makes the reproducer from this bug work. Independent of that, sharded autotuning got enabled yesterday again and it will likely get into the...
I redesigned the pass such that it runs after layout assignment and does not prevent folding of transposes into other ops surrounding all-gathers. Please take another look.
The required [change](https://github.com/openxla/xla/commit/3a5a9cc4784932df6d61e2bf4d0c2d395c284cc4) in XLA is done, this one is ready.
> Is there a minimum cudnn version for this test to pass? 9.0.
Do I see right, that both failing checks are in non-GPU configurations?