← Back to events
ActiveAIRelease

trunk/9f3a72a53845f3abdda815ba2815b9446beac828: Skip bolt optimization for libtorch.so (#194941)

What happened

NVIDIA internal measurements show that optimizing libtorch.so does not provide any statistically significant performance improvement across a variety of workloads. Removing it from the optimization set reduces the amount of BOLT profile data that we need to ship and keeps the build process simpler. This change does not remove libtorch.yaml from torch.tar.zst to avoid changing the binary file and increasing the repo…

Summary assembled by rule from the sources below

Why it's spreading

Sources