trunk/fe909cb7543b223adcfe1d11c67671b3439f490b: [c10d] Match NCCL2 collective validation with NCCL (#194953)
- PyTorch: 209 events in the last 90 days
- PyTorch: 206th Release in the last 90 days
- Previous: earlier the same day · trunk/e2dee2a4b9e91bfe6374bf39be9449bb438e08c5: Revert "Drop unused includes in c10 (#194959)"
What happened
NCCL2 let invalid gather, scatter, and reduce-scatter arguments reach its implementation, producing different exception types and messages from ProcessGroupNCCL. It also inherited the generic sparse all-reduce error. Validate the public c10d inputs in the backend adapter and preserve the existing unsupported sparse error. Test Plan: env BUILD_TEST=0 USE_FBGEMM=0 USE_NNPACK=0 USE_QNNPACK=0 USE_XNNPACK=0 USE_FLASH_ATT…
Summary assembled by rule from the sources below