Kaleto's picture
Quantize the LM head to NVFP4 (22 % faster decode, 2 GB smaller), add MTP and AWQ comparison to the card
659937b verified