city96

/

Qwen-Image-gguf

like 212

Had to remove --fast and --using-sageattention from the cmd. using the kjnode to turn off sageattention doesn't work

There seems to be similar reports of that with FP8 as well, so it may be an issue with the base ComfyUI code https://github.com/comfyanonymous/ComfyUI/issues/9184

Same problem out of 20 steps it becomes black on 11-th step even in Sampler preview

The Q4_K_M failed due to unexpected architecture, so I downloaded the Q6_K.....same issue, neither run 🤷‍♀️

Hope you can fix them as I love your work 💜

@Apolonia you have to update ComfyUI-GGUF to the latest version for it to be loadable, as well as update ComfyUI itself. Should work both via the manager and via just git pull.

@city96 Hiya....I updated comfy this morning, nothing was said about the GGUF nodes though?
Edit: GGUF updated, problem solved.
Thank you for the heads up 💜

Launching comfy without FP16 accumulation & Sageattention solved the problem of black image, generation was increased by about (4s/it) but got some good result with Q4-0

Had to remove --fast and --using-sageattention from the cmd. using the kjnode to turn off sageattention doesn't work

I am using the official Diffusers script to load the model offline, but I get a completely black image. How to solve it?

How do I disable -fast & -sage if I never used it in workflow at all ????

How do I disable -fast & -sage if I never used it in workflow at all ????

That's a launch argument, not in the workflow. Remove the --fast and --use_sage_attention flag in the run_(whatever you use).bat file.

I have also black screen, tested the harmless "red cat" so it cannot be any filter or anything. I had --use-flash-attention and tried removing it without any changes.

I have a Amd Radeon 7800xt on linux using ROCm 6.4

I have also black screen, tested the harmless "red cat" so it cannot be any filter or anything. I had --use-flash-attention and tried removing it without any changes.

I have a Amd Radeon 7800xt on linux using ROCm 6.4

ComfyUI's default split attention also seems to be problematic. Perhaps trying again with --use-pytorch-cross-attention will solve the problem.

I have also black screen, tested the harmless "red cat" so it cannot be any filter or anything. I had --use-flash-attention and tried removing it without any changes.

I have a Amd Radeon 7800xt on linux using ROCm 6.4

ComfyUI's default split attention also seems to be problematic. Perhaps trying again with --use-pytorch-cross-attention will solve the problem.

I was away during weekend, but I tried with --use-pytorch-cross-attention without any difference

Using Patch Sage Attention KJ with "sageattn_qk_int8_pv_fp16_cuda" option fixes the problem for me.

Using Patch Sage Attention KJ with "sageattn_qk_int8_pv_fp16_cuda" option fixes the problem for me.

That Sage requires cuda and that's unavailable to me on AMD hardware.

I look into why using pytorch attention didn't work as it seems to work for some
https://github.com/thu-ml/SageAttention/issues/234#issuecomment-3220150376