single mention · inference & serving
Layer-by-layer FP8-to-MXFP4 expert re-quantization at load
An FP8 checkpoint's experts are re-quantized to MXFP4 layer by layer during loading to limit peak load memory.
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