[enhancement]: Feature Request: Add int8_convrot quantization support for Krea-2-Turbo - Raw model Description:

[JavadZohani]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

javadzohani.1994@gmail.com

What should this feature add?

I would like to request native support for the int8_convrot quantization format in InvokeAI, specifically for the Krea-2-Turbo - Krea-2-Raw model.
While FP8 is currently supported, int8_convrot offers significantly faster inference speeds on lower-end GPUs without compromising quality. Adding this format would greatly improve accessibility and performance for users with limited VRAM.

Alternatives

No response

Additional Content

No response

[hiSandog]

Support should start by defining how an int8_convrot checkpoint is identified and which kernels/devices can execute it, so the loader can reject an incompatible model with a precise message instead of misreading it as generic int8. A reference fixture should compare load success, peak VRAM, output shape, and a small deterministic image tolerance against FP8/FP16. Model metadata and conversion provenance need to be preserved so users can tell native checkpoints from locally converted ones.