QWEN Layerwise
This is the QWEN 3 8B LLM
Qwen layerwise is the first FP32 training with targeted layers for FLUX Klein that consistently shown improvement vs full BF16 base QWEN on Text
When scoring successes vs failure.
If both models failed on text or anatomical errors the result was neutral.
If the layerwise model improved text but caused a anatomical error this was a failure
Success was only scored if the layerwise model improved text or fixed a glaring error like number of fingers.
The success number exceeded the threshold set for what could just be viewed as random or lucky success vs base BF16.
Description
FAQ
Comments (9)
Are you going to separate the clip separately or is this going to be in the checkpoint?
For Comfy UI it loads as the CLIP input for Klein9B
Is this an update for the qwen 3 8b text encoder for Klein?
Or an update for the Qwen image model?
QWEN_3_8B
any chance this will be gguf quantized? cuz its too huge! (T_T) (tears emoticon)
So I do not follow the logic behind wanting a CPU offloaded LLM in GGUF - Unless your forcing GPU and have enough memory to support both models, the LLM is offloaded to CPU and GGUF takes more time then even FP32
Q8 would be the lowest that might possibly benefit from the FP32 blocks, but I think it would be largely lost in quant - I suppose with Q8 LLM and Q8 Klein9B you could force GPU, but I think the quality to time loss Scaled FP8 would be the better choice
I did upload a version with FP8 on the non critical blocks that has little degradation. I recommend using --bf16-text-enc to keep comfy from casting it to FP16
@Felldude thank you!