Open
Description
🚀 The feature, motivation and pitch
This would enable a couple disparate improvements:
- Our binary releases should include the latest compiler features and optimizations. The concrete motivating example is that the compiler used for Mac wheels apparently doesn't pass
COMPILER_SUPPORTS_BF16_TARGET
(i.e., clang version greater than 15), which causes a slower bfloat16 gemv kernel to be used. - We should have test coverage for CPU bfloat16 support on Mac (We don't have test coverage for aarch64 (ARM64) bfloat16 feature (__ARM_FEATURE_BF16) #142703) -- clang 16 purports to be able to build it, but is buggy and we actually need 17+.
Alternatives
do nothing until Apple gets around to releasing an Xcode with clang 17 or later and we get around to updating to it.
Additional context
Xcode clang version history: https://gist.github.com/yamaya/2924292 . Latest at time of writing is Xcode 16.2 with Apple clang version 16.0.0 (clang-1600.0.26.6)
cc @seemethere @malfet @osalpekar @atalman @pytorch/pytorch-dev-infra
Metadata
Metadata
Assignees
Labels
Type
Projects
Status
Cold Storage