[Bug] Qwen-Image-2.1 validate_full 未加载训练权重，实际验证的是基座模型

Start with examples/qwen_image_21/model_training/validate_full/Qwen-Image-2.1.py and compare it with examples/qwen_image/model_training/validate_full/Qwen-Image-2512.py and the sibling validate_lora script. Run the full-training validation command after checking the epoch output naming in the referenced training script. Done means the validation script reads models/train/Qwen-Image-2.1_full/epoch-1.safetensors and clearly produces the full-training validation image.

问题

examples/qwen_image_21/model_training/validate_full/Qwen-Image-2.1.py 加载的是原始 Qwen/Qwen-Image-2.1，没有任何加载训练产物的语句，因此验证的是未微调的基座模型。

脚本不报错、能正常出图，容易被误判为"全量训练没效果"。

复现

全量训练输出到 models/train/Qwen-Image-2.1_full 后执行：

python examples/qwen_image_21/model_training/validate_full/Qwen-Image-2.1.py

epoch-1.safetensors 从未被读取。

判断依据

1. 文档 docs/*/Pipeline_Usage/Model_Training.md 写明 validate_full 的用途是"全量训练完成后，运行这个脚本加载模型，验证效果"。

2. 全仓库 136 个 validate_full/*.py 中只有这一个不加载任何权重（其余要么 pipe.dit.load_state_dict(...)，要么 ModelConfig(path=<ckpt>)）：

   $ for f in $(find examples -path "*validate_full*" -name "*.py"); do
       grep -qE "load_state_dict|load_lora|ModelConfig\(\s*path=|ckpt|checkpoint" "$f" \
         || echo "$f"
     done
   examples/qwen_image_21/model_training/validate_full/Qwen-Image-2.1.py
   # ↑ 唯一一个，1 / 136
   

3. 该文件自 7f65ecc8 新增起就没有加载语句，7f31effa 仅删除了 num_inference_steps=50；当前 main（7686e54d）仍未修复。

4. 同目录的 validate_lora/Qwen-Image-2.1.py 是正确的，有
   pipe.load_lora(pipe.dit, "models/train/Qwen-Image-2.1_lora/epoch-4.safetensors")。

建议修复

参考 examples/qwen_image/model_training/validate_full/Qwen-Image-2512.py：

 from diffsynth.pipelines.qwen_image_21 import QwenImage21Pipeline, ModelConfig
+from diffsynth import load_state_dict
 import torch


 pipe = QwenImage21Pipeline.from_pretrained(
     ...
     processor_config=ModelConfig(model_id="Qwen/Qwen-Image-2.1", origin_file_pattern="processor/"),
 )
+state_dict = load_state_dict("models/train/Qwen-Image-2.1_full/epoch-1.safetensors")
+pipe.dit.load_state_dict(state_dict)
 prompt = "dog,white and brown dog, sitting on wall, under pink flowers"
 image = pipe(prompt, seed=0, height=1024, width=1024)
-image.save("image.png")
+image.save("image_full.png")

epoch-1 对应 full/Qwen-Image-2.1.sh 中的 --num_epochs 2：ModelLogger.on_epoch_end 的 epoch_id 从 0 开始，故最后一个 epoch 文件是 num_epochs - 1。

环境

• DiffSynth-Studio 2.1.8（d2d684ad），main（7686e54d）仍可复现
• 涉及文件：examples/qwen_image_21/model_training/validate_full/Qwen-Image-2.1.py