trunk/561b14c1a060511f46e3713c069f4aff6f6d9c6c: Fix CUDA memory usage for CPU only compile (#163841)
A critical memory bug just got patched, saving developers from hidden crashes.
Deep Dive
The PyTorch team has merged a critical fix for a CUDA memory usage bug that occurred during CPU-only compilation (Issue #150622). The patch, approved by core maintainers, resolves a problem where unnecessary GPU memory was being allocated even when compiling for CPU execution. This prevents potential out-of-memory errors and crashes for developers working on systems without dedicated GPUs or when explicitly targeting CPU performance.
Why It Matters
This fix prevents unexpected crashes and wasted resources for countless developers running PyTorch models on CPU systems.