single source · inference & serving
FP8 latent KV-cache quantization
Latent KV-cache storage using FP8 E4M3, including the named fp8_ds_mla configuration.
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 FP8 E4M3 latent key-value cache, FP8 latent KV cache (fp8_ds_mla).
Filed under inference & serving › inference quantization