single mention · inference & serving
MXFP8 block-scale regrouping at load
On Ascend 950, block scales are regrouped to MXFP8 at load while weights remain one byte per element.
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