trunk/bf632e6db795c5e6537fa418ae8fbed611b0d1c5: [DTensor] prims ops sharding strategies (#174442)
Massive performance boost incoming for distributed AI training...
Deep Dive
PyTorch developers have merged a commit adding sharding strategies for 36 more primitive operations, continuing work from PR #171649. Of these, 21 operations pass outright while 15 currently fail due to issues like DynamicOutputShapeException or numerical discrepancies. This commit represents a significant step forward in PyTorch's DTensor functionality, which is crucial for efficiently distributing tensor computations across multiple devices during large-scale model training.
Why It Matters
This directly accelerates distributed training for massive AI models, potentially cutting training times and costs significantly.