single mention · inference & serving
MXFP4 block-scaled tensor packing
A storage implementation that packs two FP4 values into each uint8 value alongside block scales.
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 pack every two values in one uint8 value.
Filed under inference & serving › inference quantization