Missing keys when loading Wan2.2-Fun-5B-Control in WanTransformer3DModel

用 WanTransformer3DModel.from_pretrained(ckpt_dir, torch_dtype='torch.bfloat16').to(device) 加载 Wan2.2-Fun-5B-Control ，报missing keys ：

['blocks.0.cross_attn.k_img.weight', 'blocks.0.cross_attn.k_img.bias', 'blocks.0.cross_attn.v_img.weight', 'blocks.0.cross_attn.v_img.bias', 'blocks.0.cross_attn.norm_k_img.weight', 'blocks.1.cross_attn.k_img.weight', 'blocks.1.cross_attn.k_img.bias', 'blocks.1.cross_attn.v_img.weight', 'blocks.1.cross_attn.v_img.bias', 'blocks.1.cross_attn.norm_k_img.weight', ...... , 'blocks.29.cross_attn.k_img.weight', 'blocks.29.cross_attn.k_img.bias', 'blocks.29.cross_attn.v_img.weight', 'blocks.29.cross_attn.v_img.bias', 'blocks.29.cross_attn.norm_k_img.weight']

查看源码和Wan2.2-Fun-5B-Control的内容发现，t2v_cross_attn 和 cross_attn 都是没问题的，而 i2v_cross_attn 中定义的 k_img 和 v_img 在 Wan2.2-Fun-5B-Control 中是不存在的，且 Wan2.2-Fun-5B-Control 的 config.json 的 "model_type" 是 "ti2v"， 在 WanTransformer3DModel 中根据代码 cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn' 会以 i2v 加载cross_attention，但层keys不匹配了。

请问，是我使用的源码不对吗？ 还是模型下载的不对？ 感觉 WanTransformer3DModel 和 Wan2.2-Fun-5B-Control 好像并不匹配 。

Start with WanTransformer3DModel and the Wan2.2-Fun-5B-Control config.json, then reproduce the reported from_pretrained call with the model checkpoint. Compare the selected cross-attention type with the checkpoint keys; done means the compatibility problem is resolved or the supported model/configuration mismatch is clearly documented.