Skip to content

PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度

fix: preserve Z-Image text encoder export layout (#2559)

Wang, Changfe68946de7bd
diff --git a/auto_round/compressors/diffusion_mixin.py b/auto_round/compressors/diffusion_mixin.py
index 81065e9f1..312bbebbd 100644
--- a/auto_round/compressors/diffusion_mixin.py
+++ b/auto_round/compressors/diffusion_mixin.py
@@ -593,7 +593,19 @@ class DiffusionMixin:
if saved_subfolder is not None:
self.model_context.model._autoround_pipeline_subfolder = saved_subfolder
elif val is not None and hasattr(val, "save_pretrained"):
- val.save_pretrained(sub_module_path)
+ if (
+ pipe.__class__.__name__ == "ZImagePipeline"
+ and name == "text_encoder"
+ and val.__class__.__name__ == "Qwen3Model"
+ ):
+ # Preserve the original causal-LM layout expected by vLLM-Omni.
+ state_dict = {f"model.{key}": value for key, value in val.state_dict().items()}
+ val.save_pretrained(sub_module_path, state_dict=state_dict)
+ config = copy.deepcopy(val.config)
+ config.architectures = ["Qwen3ForCausalLM"]
+ config.save_pretrained(sub_module_path)
+ else:
+ val.save_pretrained(sub_module_path)
continue
if name.startswith("transformer"):