Open source
Mirage
Padding / Argmax correctness in Qwen3 demos.
Merged record
| PR | Merged change | Status |
|---|---|---|
| #755 ↗ | Adjust lm_head padding initialization in Qwen3 demos to address invalid tokens from padded Argmax candidates. | Merged |
Scope
PR #755 changes the initialization of lm_head rows beyond the true vocabulary from 0 to −1e4 in nine demo scripts. The change is limited to demo scripts, not a rewrite of the general Argmax kernel.