harness/nemotron: allow DiagGather + ScaledMaskedSoftmax on GPU op-only assert#2259
Merged
Merged
Conversation
The encoder's ROI attention emits DiagGather and ScaledMaskedSoftmax, which have no Metal/Cuda impl and fall back to CPU. The nemotron GPU --assert-op-only allowlists didn't list them, so the macOS/Metal (and cuda) Large-models job fails with "Model has 48 unexpected op(s)". Mirrors parakeet's ci.sh (already allows DiagGather); nemotron's chunk-window-mask encoder additionally emits ScaledMaskedSoftmax. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
4 tasks
Collaborator
|
Sorry, took a shortcut. Thanks for the fix. |
kali
previously approved these changes
May 20, 2026
Collaborator
|
There's a second one hidden behind. I'm on it. |
Collaborator
|
Apologies. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Unblocks the
macOS / Nemotron speech streaming en 0.6b(andcuda-lovelace) Large-models job, which currently fails onmainwith:The encoder's ROI attention emits
DiagGather+ScaledMaskedSoftmax. These have no Metal/Cuda implementation, so they fall back to CPU — but the nemotron--metal/--cuda--assert-op-onlyallowlists inci.shdon't list them, so the strict op-coverage assert fails on the GPU runtimes. (The CPU-Opass of all four sub-models passes fine.)harness/parakeet-tdt-600m-v3/ci.shalready allowsDiagGather; this mirrors that for nemotron and additionally addsScaledMaskedSoftmax(nemotron's chunk-window-mask encoder uses it; parakeet doesn't).If you'd rather these get Metal/Cuda impls than be CPU fallbacks, happy to drop this — but as-is it unblocks every PR branched off current
main(the job is currently red on e.g. #2257 and task/onnx-dtype-panic, neither of which touch these ops).Test plan
macOS / Nemotron speech streaming en 0.6bgreencuda-lovelace / Nemotron speech streaming en 0.6bgreen🤖 Generated with Claude Code