UD-TQ1_0 variants of 30B+ llms not seen lately .

#28
by TnK93 - opened

hi have you guys noticed the TQ1 variants are missing lately.... really miss them, they wheren't the most accurate but they where very usable.. anyone else missing them ? i was using Qwen 3 Coder Next 80B TQ1_0 on 16gb vram and 32gb ddr5... just wanted to try latest models with the same....

Qwen 3 coder 30B A3B @TQ1 is around 8 to 9GB which actually gives us some usable room for KV cache for a larger context... the main question is do you guys think TQ1 variants are reliable enough in your workflows ????

TQ1 is not very good at encoding large values in a block. Directly training in TQ1 works, but any conversion into it will result in many blocks with all-zero next to the single outlier.
Iq1_s has much more localized loss in this case.
I understand it got dropped in its current form.

Sign up or log in to comment