[enhancement]: Add support for INT8 ConvRot quantized models (*_int8_convrot.safetensors)

[JavadZohani]

Is there an existing issue for this?

• I have searched the existing issues

Contact Details

No response

What should this feature add?

Summary

I would like to request support for loading and inferencing INT8 ConvRot quantized models (e.g., *_int8_convrot.safetensors) in InvokeAI.

Motivation & Background

Quantization using Hadamard / rotation matrices (convrot) has recently gained significant traction for large diffusion architectures (such as Krea 2 Turbo/Raw, FLUX.1, and other modern DiT pipelines).

Compared to standard FP8 and traditional INT8/GGUF approaches, INT8 ConvRot:

1. Suppresses activation outliers effectively, preserving near-lossless generation quality compared to BF16/FP16.
2. Drastically reduces VRAM footprint and enables high-resolution generations on consumer hardware (8GB - 12GB VRAM cards).
3. Offers fast inference with efficient linear dequantization kernels.

Currently, many newly released checkpoints (e.g., krea2_turbo_int8_convrot.safetensors) only work in ComfyUI (via native loader or ComfyUI-INT8-Fast). Having native dequantization support for convrot safetensors in InvokeAI would make these lightweight, high-performance weights accessible to InvokeAI users without requiring separate format conversions.

Proposed Solution

• Add dynamic dequantization / loading support for tensors structured with rotated INT8 weights (convrot).
• Allow InvokeAI's Model Manager and unified loading pipeline to recognize and load *_int8_convrot.safetensors files without throwing structural mismatch or invalid layer errors.

References & Existing Implementations

• ComfyUI Integration: Native support added in ComfyUI v0.27+ and community implementations like ComfyUI-INT8-Fast.
• Target Models: Krea-2-Turbo, Krea-2-Raw, FLUX.1-dev/schnell INT8 convrot checkpoints.

Thank you to the InvokeAI team for the continuous development and great work!

Alternatives

No response

Additional Content

No response