Kaleto's picture
Quantize the LM head to NVFP4 (14 % faster decode, 1.5 GB smaller), add MTP and AWQ comparison to the card
ff00187 verified