viable/strict/1790720927: [symm_mem]: Add per-PG stream serialization for ops (#195947)
- PyTorch: 1672 events in the last 90 days
- PyTorch: 1653th Release in the last 90 days
- Previous: earlier the same day · viable/strict/1790717299: Note MI300 accuracy issues in set_float32_matmul_precision (#190736)
What happened
Partial fix for #188005 - only for CUDA and NCCL backends. #196337 is opened to address the NVSHMEM backend. This PR is a step on the way to implementing per-PG signal pads, which would resolve #192581 Replace the advisory warn_if_multi_stream diagnostic with an RAII GroupStreamGuard that enforces correct stream ordering across all built-in symmetric-memory operations that touch the signal pad. Signal pad operations…
Summary assembled by rule from the sources below