Repository navigation
CUDA kernel contract: the nvfp4_expert_waves<2, 2, 4> planted control stays green on sm_120 #1808
Closed as not planned
Description
Reported in the receipts for #1632 (@BobClawblaw, RTX 5090): on sm_120, main's planted control for nvfp4_expert_waves<2, 2, 4> (items dropped past the first wave) passes at the two-wave shape the contract runs, so the contract would not catch that fault there. Every other control in the run is red as planted. The control (or the shape the contract runs it at) needs to reach a second wave on sm_120 so the dropped items change the output. Thanks to @BobClawblaw for spotting it.
Activity
ashhart commentedon Nov 1, 2026
Closing as a duplicate of #1778, where @BobClawblaw reported the same control on 31 Oct with the cause (the fixed 1,024-row shape fits one gate-up wave on an RTX 5090's 170 SMs) and a fix: size the shape from the device's resident warps. We'll track the fix there. Thanks again @BobClawblaw.