[ArberSephirotheca]

Overview

The CUDA MoE top-k kernel sanitizes NaNs in wt at lines 200–204, before adding the expert selection bias at lines 218–219. Adding bias can introduce a new NaN.
Without an additional check, NaNs can reach the warp argmax comparison at lines 270–274. Comparisons involving NaN are false, so different lanes can select different winning experts and write conflicting values to ids[k] at lines 309–310.
The fix adds a check immediately after bias addition, replacing NaNs with -FLT_MAX before group masking and subsequent top-k selection.

Additional information

This PR duplicates the closed PR #29609. I accidentally deleted the original branch, so I recovered the commit and rebased it onto the latest master. Sorry for the confusion.
@am17an @JohannesGaessler, could you take another look when you have a chance? Thank you!

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:
  YES - Codex inspected the code and history, rebased the existing commit, and helped draft the PR explanation.

[ggml-gh-bot]

Hi @ArberSephirotheca, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.