Q2_K in revision: user report that 1.1 is giving issues
I ran a bunch of tests last night with FLF and every generation came out blurry and cooked. Just wanted to let you know! Maybe something happenee during compression?
Strange. What GGUF are you using? Can you share your settings in workflow also? I didn’t get that issues
using the FLF workflow. im using the Q2 gguf!
im not sure what this issue is tbh, the regular 1.0 works great
I need to test that Q2_K but in running the Q3_K_S and its fine
Add: And the quantization its the same used in 1.0
i am using Q4_K_M and its flawless thanks for the quants !!
Maybe its a me issue, could be my comfy causing it im not sure, i dont want to steer people in the wrong direction because others have reported it works great. Ill test a higher quant later to see if its me or the lower quant 🤷♂️
Tried here as well, just being curious about Q2 (since that something i never try, just thinking the sweet spot is Q4 and above...)
I did see quite a bit lower quality at the Q2, while the Q3 is ok. But I have no idea how much to expect from Q2, being less than half the size of the model ;-)
(turned off the audio in the video above, since audio is fine in both, and both talking at same time is a mess ..hehe)
Could of course be that with tweaked settings and more steps etc the difference would be less (ran with 8 steps, 2-pass sampler, euler_a_cfg_pp + euler_cfg_pp)
Q3 and above looks great though ;-)
I did see quite a bit lower quality at the Q2, while the Q3 is ok. But I have no idea how much to expect from Q2, being less than half the size of the model ;-)
fair point, cant expect miracles on every release! i appreciate your reply and ill be sure to work out how to fit that q3 on my setup!
Could be the Q2 has an error, I have no idea about that ;-)
The Q3 is only slightly larger, so hopefully you can make that work.