About This Model
An experimental ConvRot W4A4 MSE quantization of PrunaAI's P-H3-Image, created using PotatoForge with a focus on compatibility and efficient inference on older NVIDIA GPUs.
This quantization was developed specifically for my GTX 1660 Super (SM 7.5), targeting ComfyUI's optimized ConvRot W4A4 execution path rather than newer quantization formats that require SM 8.0+ hardware.
Key Details:
- Quantization: ConvRot W4A4 MSE (4-bit weights and activations)
- Model Size: ~10.7 GB (down from ~37.6 GB BF16)
- GPU Target: NVIDIA Turing (SM 7.5), including GTX 16-series GPUs
- Compatibility: ComfyUI with ConvRot W4A4 support
- Precision: Main transformer attention and MLP layers quantized; token refiner and other unmatched tensors preserved at their original precision
- Scales: MSE Optimized Scales for lower reconstruction error
The goal is to make P-H3-Image more accessible on memory-constrained hardware while maintaining reasonable image quality. Tested on a GTX 1660 Super at up to 4 MP resolution.
- Downloads last month
- 203