<h1 id="qwen_vl_8b-for-qwen-image-2.1-(qp)">Qwen_VL_8B for Qwen Image 2.1 (QP)</h1>Note: I have tested uncensored and abliltered models on QWEN 2.1 those models tend to misalign and cause aberrations with the 2.1 DiT model. QP does not effect the core LLM and leaves the BF16 intact while predicting FP32, this allows for less rounding errors when quantizing.<h3 id="all-models-other-then-fp16-use-comfy-metadata-for-quantization.">All models other then FP16 use Comfy Metadata for quantization. </h3>Requirements:<ol><li>Update <a target="_blank" rel="ugc" href="https://developer.nvidia.com/cuda-downloads">CUDA to 13.4.2</a></li><li>Update <a target="_blank" rel="ugc" href="https://pytorch.org/">Pytorch to 13.2</a></li><li>Update at minimum comfy-aimdo, comfy-kitchen</li></ol>pytorch version: 2.13.0+cu132xformers version: 0.0.35Using xformers attentionComfyUI version: 0.35.0comfy-aimdo version: 0.5.5comfy-kitchen version: 0.2.34<hr /><h3 id="qp-(quantization-prediction)">QP (Quantization Prediction)</h3>QP is theoretically improving 40-50% of the blocks on the trailing 16 Matnitsa bits.For the other 50-60% that it does not improve it did not degrade them more then what they would have been rounded to in the first place.This was tested on Full FP32 trainings such as T5.

Qwen_VL_8B for Qwen Image 2.1 (QP) INT8, MXFP8, NVFP4

Files with SHA256: 165b649bb84667aa90fefc41166fe089328941a555e30377a8ca60e47f0d6c3d