PyTorch fixes HPU backend mapping issue for Habana AI accelerators
A targeted fix prevents 'fake' backends from claiming devices already assigned to real AI hardware.
Deep Dive
The PyTorch team (Meta) merged PR #174764 to resolve a critical backend mapping issue for Habana Gaudi AI accelerators (HPUs). The fix updates the `register_backend` logic in `torch/distributed/distributed_c10d.py`, allowing devices to be correctly remapped from a placeholder to the real HPU backend while preventing assignment conflicts. This ensures stable multi-device training setups using specialized AI hardware can run without errors.
Why It Matters
This fix is crucial for developers running distributed AI training on Habana's high-performance accelerators, preventing crashes and hardware assignment errors.