| { | |
| "_comment": "MODEL = Qwen-Image-Layered (diffusers QwenImageLayeredPipeline, RGBA layer decomposition, issue #327). EFFECTIVE INT4 precision map — verified COHERENT on RTX 5090 / SM120 (run 20260616-201138-linux-int4ref). attn + MLP use AUTO_4 (the integer 4), which the Qwen W4Fp4Policy resolves to INT4 per-group-128 on ALL SMs (#239/#240); mod stays INT8. INT4 is fully coherent on the layered at any SM and is the robust fallback; the SM120 NVFP4 (FP4-weight) variant is qwen-image-layered-fp4.json. COMPANION SETTINGS REQUIRED for coherence (NOT part of this map): base scheduler (configs/qwen-image-base-scheduler.json), num_inference_steps=50, true_cfg_scale=4.0, negative_prompt non-empty, a REAL RGBA composite input image, resolution 640.", | |
| "transformer_blocks.attn.to_qkv": 4, | |
| "transformer_blocks.attn.add_qkv_proj": 4, | |
| "transformer_blocks.attn.to_out": 4, | |
| "transformer_blocks.attn.to_add_out": 4, | |
| "transformer_blocks.img_mlp.net.0.proj": 4, | |
| "transformer_blocks.img_mlp.net.2": 4, | |
| "transformer_blocks.txt_mlp.net.0.proj": 4, | |
| "transformer_blocks.txt_mlp.net.2": 4, | |
| "transformer_blocks.img_mod": "i8", | |
| "transformer_blocks.txt_mod": "i8" | |
| } | |