viable/strict/1770973414: [DTensor] Fix bucketize with Partial inputs (#173937)
This obscure bug fix could prevent major headaches for distributed AI developers.
PyTorch developers have resolved a critical bug in DTensor's `bucketize` function that was producing invalid strategies when handling Partial inputs. The issue occurred when bucket indices were incorrectly combined with sum/avg reductions, propagating Partial placements to outputs. The fix converts Partial input placements to Replicate, ensuring bucketize operates on properly replicated data. This prevents silent errors in distributed training scenarios where tensors are split across multiple devices.
Why It Matters
This prevents silent failures in large-scale distributed AI training, ensuring model accuracy across GPU clusters.