Qwen Image 2.1 SDNQ INT4

Use scripts/qwen_image_21/run_dynamic_modular_sdnq.py with the SDNQ optional dependencies. No FP8 extra or native allocator is required.

The saved transformer is OzzyGT/Qwen_Image_2_1_sdnq_dynamic_4bit. Auxiliary components come from Qwen/Qwen-Image-2.1 in BF16.

Run the model

python -B scripts/qwen_image_21/run_dynamic_modular_sdnq.py --variant base --width 512 --height 512 --steps 20 --seed 42 --ddo-preset one_shot_fast --vae-resident --purge-before-transformer --memory-monitor rss --show-metrics --save-metrics

Run elevated on Windows for standby purge. To compare official streamed leaf, change only --ddo-preset to diffusers_leaf_offload_compat. The selected Windows INT4 comparison is in Qwen benchmarks. Turbo, INT8 and Linux/WSL INT4 are not established by that comparison.

For common arguments, metrics and output paths, see shared controls. For model controls, see Qwen parameters.