18GB varaint?

#2
by dochuerbat - opened

Great work! I'm just trying out this!
A varaint with around 18 GB file size maybe could be intersting, too. Maybe a bigger Q4 varaint or a smaller Q5.
Can you do this?

Owner

so something like unsloth does with the Q4_XL variants?

exactly!

Owner

What are your memory requirements? Because the IQ4_XS I produced is of a better quality than the normal IQ4_XS. I don't know if there would be any need to just have another model file that is half or one gigabyte larger and there will be no real quality difference.

I think it would be good if there were a Q4_K_M version.

What are your memory requirements? Because the IQ4_XS I produced is of a better quality than the normal IQ4_XS. I don't know if there would be any need to just have another model file that is half or one gigabyte larger and there will be no real quality difference.

Owner

OK so I did a Q4_K_M version of it, but I don't know why it got flagged unsafe by huggingface, its a normal Q4 quant. I'll send them an email.

Thank you!
I'll try it out!

I'll run it on 28GB VRAM (16+12).
Yesterday I figured out my llama.cpp build with TQ and MTP doesn't perform well. I'm using the atomic-llama-cpp-turboquant fork. Does somebody have a better alternative?

Owner

Try to use the mainline one. they have also some new fixes for MTP. turboquant is overrated, and you can use Q4_0 with great results. I also use it with Q4 KV cache and it goes very well.

Currently in fact I use mainline with a build from yesterday (with Q8 KV on the 27B model, qwen3.6 35B-A3B fits easily on native KV).
I experience a big degration in accuracy (not measured, only my feeling/experience) as soon as my context fills up to ~50k. For agentic coding I need at least 100k useable "good" context. That's why I want best possible model and KV accuracy I can get.

Sign up or log in to comment