Failed to load safetensors using method MIO and 3d torch.Tensor is deprecated.

[one903045]

trafficstars

deepcompressor 0.0.1 image_gen_aux 0.1.0.dev0 image-reward 1.5 nunchaku 0.1.4+torch2.6

I can complete the sampling and output of images, but I don't feel any acceleration, 4 steps is Blurred and terrible. my GPU is 3090，CUDA is 12.6，WIN10，torch2.6、python12.9 version use nunchaku-0.1.4+torch2.6-cp312-cp312-win_amd64.whl to compile Successfully. And Another issue is batch can't more than 1.

[lmxyy]

The latest release of nunchaku node v0.1.6 in our new repo should resolve the 3D tensor issue. Please disable the CPU offloading on your GPUs. Currently, we only support batch size of 1. Larger batch inference will be supported in the later version. Could you send us your generated images for debugging?

[one903045]

Oh!It‘my misunderstood，dev is Normaly accelerate by multiple steps。Maybe 4 steps need shnell？And [warnign] Failed to load safetensors is still exist.

(i use ComfyUI-nunchaku node v0.1.6 )

When I use Shnell int4

When I use Shnell fp4

When I use shuttle yaguar int4

I uninstall nunchaku（use whl） and try to compile by myself, but failed in PYthon11 or 12，WIN10.

[github-actions[bot]]

This issue has been automatically closed due to 30-day inactivity. Please feel free to reopen it if needed.

[lmxyy]

Could you try our walkaround in the FAQ?