People who tried gpu-unlock on 3090s and such - does this make any difference on mixed-precision quants where you get some parts of the model quantized to <=Q8 but keep some tensors at BF16?
retained text
Dunno, curious if there's any info out there.
document fetched 2026-10-27 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xg0kle/people_who_tried_gpuunlock_on_3090s_and_such_does