PyTorch adds LLVM BOLT optimization, boosting CUDA builds by 30%

PyTorch has merged pull request #186245, integrating LLVM BOLT (Binary Optimization and Layout Tool) into its CUDA build pipeline. BOLT is a post-link optimizer that reorders an application's code layout based on execution profiles gathered via `perf`. The technique was already used in NVIDIA's official PyTorch containers since version 26.04, and now PyTorch brings it directly into the repo. The integration works in two phases: first, a profile is collected from an unoptimized binary, then that profile optimizes the binary at build time as a POST_BUILD step. The unoptimized library is moved to a `prebolt/` subdirectory, and the optimized version replaces it at the canonical path.

BOLT only activates under specific conditions: Linux, aarch64 architecture, CUDA support, llvm-bolt version 21 or higher, and the existing linker script optimization (`USE_PRIORITIZED_TEXT_FOR_LD`). The PR ships curated profiles collected from a set of workloads, so ARM CUDA builds get optimized by default. On the current PyTorch codebase, internal workloads show up to 30% performance improvements and up to 3% regressions, with most workloads benefiting. However, there are caveats: stale profiles lose some effectiveness as the codebase evolves, `libtorch.so` can balloon from ~60K to ~6M in RelWithDebInfo mode due to BOLT's hot/cold section padding, and debug info is reconstructed on a best-effort basis, which may slightly degrade debugger experiences. The PR was assisted by Claude Opus 4.8 and follows an RFC discussion, signaling growing adoption of post-link optimization in mainstream AI frameworks.