Rank plans in one place: delete propose_plans, Router, and heuristics_sort - #528

Merged

Anerudhan merged 15 commits into

NVIDIA:developfrom

YangXu1990uiuc:dispatch-ranking

Aug 9, 2026

Commits on Aug 8, 2026

Take the backend's plans one heuristic mode at a time
YangXu1990uiuc
committed
1a15f5c
Move plan ranking out of the engines and into one heuristics function
YangXu1990uiuc
committed
2001b41
Update the dispatch tests to the ranking contract, and delete what it retired
YangXu1990uiuc
committed
e429045
Update the design doc, and align the remaining sdpa test helpers
YangXu1990uiuc
committed
9baf760
Decode an engine id from the manifest, with nothing registered first
YangXu1990uiuc
committed
92d8344
Remove the out-of-tree engine concept: the manifest is the only way in
YangXu1990uiuc
committed
6b906c7
Give the SM120 SDPA-forward cell a real tile rule, as the worked example
YangXu1990uiuc
committed
a278de8
Ask one function whether an SM120 tile fits, not two
YangXu1990uiuc
committed
2d0eb2d
Query the backend for the modes the ranking will actually place
YangXu1990uiuc
committed
5c8d4c0
Restore the #512 SDPA tests this branch had silently reverted
YangXu1990uiuc
committed
514ea4e
Bring the design doc to the architecture as it now stands
YangXu1990uiuc
committed
71a6af7
Answer the backend's plan query once per distinct config
YangXu1990uiuc
committed
1a85c82
Address the CodeRabbit pass: an id names one engine, and two tests could not fail
YangXu1990uiuc
committed
9da8cee

Commits on Aug 9, 2026

License the one file this PR adds as Apache-2.0, not MIT
YangXu1990uiuc
committed
717562d
Compress the comments this PR adds
YangXu1990uiuc
committed
bcfee39