single mention · inference & serving
Separate target and MTP lm_head precision
The target model and MTP drafter may use different lm_head precisions, retaining BF16 for the target and lower precision for drafting.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See inference quantization for the methods filed beside it.
Filed under inference & serving › inference quantization