11 Comments
User's avatar
Utku's avatar

CLIPLoader

Error(s) in loading state_dict for Llama2:

size mismatch for model.embed_tokens.weight: copying a param with shape torch.Size([151936, 4096]) from checkpoint, the shape in current model is torch.Size([128256, 4096]).

Nescolet's avatar

I get this error:

UNETLoader

Got [32, 32, 32, 32] but expected positional dim 64

* Load diffusion model node is highlighted border in red

It has set as flux-2-klein-4b-fp8.safetensors

* Load clip set to qwen_3_4b.safetensors

* VAE as flux2-vae.safetensors

Desktop install

ComfyUI v0.8.2 / ComfyUI_frontend v1.35.9 / ComfyUI_desktop v0.7.2

laoNan404's avatar

klein_9b with Qwen_3_8b on Mac desktop, running error with clipLoader. my comfyui has been updated 0.9.2. why?????????? I just used all the full models no quant..

laoNan404's avatar

fix: comfyui version 0.8.2

Gonzo Pod's avatar

The BF16 '9B Base Workflow' has Unet set to: 'flux-2-klein-base-9b-fp8.safetensors' (distilled) and its corresponding clip where it should have 'flux-2-klein-base-9b.safetensors' (base) and whatever clip it uses.

leduc20902's avatar

where is the 9B worflow image generation ? Download Text-to-Image Workflow

Thorbjorn's avatar

how do one do the iteration to the image created? Its always been a mystery to me!

Norbert Umland's avatar

1. Click on ‘assets’ in the left-hand column. 2. Right-click on the image that opens and click on ‘copy’. 3. Click on the input image and press “Ctrl” and ‘V’ simultaneously. Done!

Og's avatar
Jan 15Edited

I hope I am asking this correctly.. Are these designed to be used with the 5xxx series Blackwell chips? Or will they also work with a 3xxx series? Or does it matter?

John's avatar

The NVFP4 is meant for the 50 series cards. You wont see the performance gain on 40 series or H100.

Igor Dvorski's avatar

It works for me on 3090, no problemos.