PyTorch's LLVM BOLT integration speeds up CUDA builds by 30%

PyTorch has merged a significant pull request that brings LLVM BOLT, a post-link optimizer, directly into its CUDA build process. BOLT works by collecting execution profiles with perf and then rearranging the application's code layout to improve instruction-cache locality. The PR, authored by lakshayg and assisted by Claude Opus 4.8, ships pre-collected profiles from a curated set of workloads, meaning CUDA builds are now optimized by default without users needing to generate their own profiles. The build-time step moves the unoptimized library into a prebolt/ subdirectory, then runs llvm-bolt to write the optimized version back to the canonical path.

BOLT is enabled by default only on Linux + aarch64 + CUDA when llvm-bolt (version 21 or newer) is detected and USE_PRIORITIZED_TEXT_FOR_LD is turned on, since the shipped profiles were collected from builds using that linker script optimization. The result is up to 30% performance improvement on some internal workloads, with a worst-case regression of only 3% on others. Caveats include the use of stale profiles as the codebase evolves, a significant libtorch.so size increase in RelWithDebInfo mode (from ~60K to ~6M due to hot/cold section padding), and best-effort debug info reconstruction that may affect stacktraces. ARM has also recommended this approach, and NVIDIA already uses BOLT in its PyTorch containers since release 26.04.