Minimax h3 ref versions?
Thanks for your job
Do you still intend to conduct further tests and verify that the quality is maintained, and release a version geared towards the ref version?
Yea, was struggling to get the BQ version to a good quality but settled on one. Just uploaded/uploading now 😄 Will continue working on a better BQ version in the meantime
Your conversions are great, how did you detect the bad layers? awesome work!
Your conversions are great, how did you detect the bad layers? awesome work!
Thank you, I forked the int8-fast custom node, edited it to run 4 calibration steps of the int4 model against the bf16 model, with int8 layer ratio control and int8 mm ratio control https://github.com/tsolful/ComfyUI-INT8-Fast my int4mixed models are about to be obsolete when this gets merged https://huggingface.co/Kijai/MiniMax-H3-experimental