single mention · inference & serving
BF16 MTP attention weights with FP8 K/V BMM quantizers
MTP attention weights remain BF16 while only the key/value batch-matrix-multiplication quantizers use FP8.
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