single mention · inference & serving
Aggressive mixed-precision quantization recipe
A recipe combining NVFP4 for MoE and shared-expert GEMMs with FP8 for attention projections and Mamba linear layers.
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.
Also called aggressive recipe.
Filed under inference & serving › inference quantization