Model: some Flux fine-tunes can't be loaded due to weird tensor naming conventions.

[stduhpf]

A friend of mine had trouble loading a model called "Flux Unchained".

When loading the tensors for the Unet, the logs were filled with:
[INFO ] model.cpp:1685 - unknown tensor 'model.diffusion_model.model.diffusion_model.[**tensor name here**]' in model file
for every single tensor in the model. Then when calling new_sd_ctx(), the logs are:
[ERROR] model.cpp:1729 - tensor 'model.diffusion_model.[**tensor name here**]' not in model file

So it looks like the model.diffusion_model. part of the tensor names got duplicated, but the model still somehow work in other inference backends (at least it works in comfyui).

[Green-Sky]

use -m instead of --diffusion-model

also alternative hotfix #403 😆

[stduhpf]

Ah I missed your PR because it's closed. Thanks!

[Green-Sky]

Yea, I closed it bc of the simple "work around".

[stduhpf]

I think it's worth doing at least something about it, maybe a warning in the logs to advise the user to use -m instead?

[Green-Sky]

Yea, something needs to be done.