forked from ccf-ai-infra/TileOPs-Metax
c5f3aacbf1
Closes #870. ## Summary - add H200-only WS dispatch to `GroupedQueryAttentionFwdOp` - dispatch supported non-causal shapes to the new persistent WS kernel - dispatch supported causal shapes to the new persistent WS causal kernel - add the anchor wait helper and benchmark tagging for the WS paths ## Scope - on H200, `GroupedQueryAttentionFwdOp` now dispatches to the new WS kernels when the outer workload is large enough to sustain the persistent schedule - the current dispatch threshold is `total_work_items >= 132` - unsupported shapes continue to use the existing kernel path - non-H200 paths are unchanged ## Notes - the current kernel integration focuses on the WS compute path itself - it does not yet include the later tile-to-tile workload scheduling optimizations explored in experiments - those outer-scheduling improvements will follow in later work |
||
|---|---|---|
| .. | ||
| __init__.py | ||
| utils.py | ||