File size: 462 Bytes
0dfed52
 
5c8b499
 
0dfed52
 
5c8b499
 
0dfed52
5c8b499
095bd8e
5c8b499
 
1
2
3
4
5
6
7
8
9
10
11
12
13
Created with ThunderFun's QUIP implementation, merged into regular row-wise int8. Bit of a misnomer now.

For use with:

https://github.com/BobJohnson24/ComfyUI-Flux2-INT8

https://github.com/ThunderFun/ComfyUI-Wan-INT8



Not quite as much speedup as flux2 klein 9b. 00:43<00:00,  1.76s/it (BF16) 00:27<00:00,  1.09s/it (INT8 QUIP) about 1.59x faster than bf16 on my 3090.

It was necessary to keep layers.0, layers.27,28,29 in BF16 to avoid subtle artifacting.