single mention · inference & serving
On-the-fly block-wise FP8 weight quantization
Inference weights are quantized to FP8 during weight synchronization using per-block scales computed without calibration data.
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 in-flight block-wise weight quantization.
Filed under inference & serving › inference quantization