city96 / ComfyUI-GGUF

GGUF Quantization support for native ComfyUI models
Apache License 2.0
775 stars 44 forks source link

Cannot set Max_shift with schnell model #46

Open pearhoang opened 1 month ago

pearhoang commented 1 month ago

i can't set max_shift with schnell model, when i use schnell it automatically sets max_shift to 0, while this never happens with dev. is this a bug? if it is a bug can you help me find a way to fix it? image_2024_08_19T03_06_06_499Z image_2024_08_19T03_03_33_315Z

city96 commented 1 month ago

Not sure what that custom Flux Sampler Parameters node is. Could you test it with the original FP16 safetensors one, assuming you can load it for just enough time to get that status message? If it has the same issue then it's probably worth reporting to the author of that custom node.

pearhoang commented 1 month ago

image The above node is just a node combining Comfycore's ModelSamplingFLux model. In the Sampler it doesn't affect anything. I tried with the normal workflow FP8,FP16 it can completely set Max_shift, but with the Q,NF4 Schnell model it can't set the value but automatically set it to 0. I realize Schnell Q,NF4 is a FLOW model type, I don't know if it affects this. Anyway, thanks for your help.

ManOrMonster commented 1 month ago

@pearhoang The built-in node ModelSamplingAuraFlow is what you want to use with Schnell, including GGUF. It has a lot of effect on the output. Use a value of 2 (I believe) to do default outputs.

image