Support Lumina-image-2.0

[sdbds]

Still in preparation.

After checking their sampler using flux and vae, the textencoder part uses google's gemma2

@kohya-ss CC

[rockerBOO]

I got this setup locally, I know it's not ready for anything but I want to get it working. Let me know if you want to work together on this. I can help with some of the model loading parts which is where I got stuck with after poking at it. If you are progressed past this, I can help wherever else or just testing.

Thanks.

[sdbds]

I got this setup locally, I know it's not ready for anything but I want to get it working. Let me know if you want to work together on this. I can help with some of the model loading parts which is where I got stuck with after poking at it. If you are progressed past this, I can help wherever else or just testing.

Thanks.

Thank you, the framework is basically set up at the moment, but there is still some room for improvement in the caching strategy.

I think I can discuss with @kohya-ss whether to continue using the previous method.

#1924 (comment)

[envy-ai]

Thank you, the framework is basically set up at the moment, but there is still some room for improvement in the caching strategy.

Does that mean I can download your fork and test it now?

[rockerBOO]

It's still not quite working but I'm working through some issues at the moment. Mostly with model loading but will see what else is needed after that. It is fairly barebones so wouldn't expect it to be in working state just yet.

[sdbds]

After multiple updates, the project can now run under limited conditions:

1. Flash_attn on Windows will cause NAN, so it must be run in a Linux environment.
   Later consideration will be given to transforming it into SDP or xformers-driven
2. The POS ID calculation for token sequences is not padded to the max length, which leads to the necessity of batchsize = 1