AestheticMetric模型加载问题

File "/opt/tiger/Cvmm-DiffSynthSFT/custom_tools/metric/ds_aesthetic.py", line 40, in __init__
    self.ds_metric = AestheticMetric.from_pretrained(
                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/tiger/Cvmm-DiffSynthSFT/diffsynth/metrics/aesthetic.py", line 25, in from_pretrained
    model_pool = cls.download_and_load_models([model_config], torch_dtype=torch_dtype, device=device, vram_limit=vram_limit)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/tiger/Cvmm-DiffSynthSFT/diffsynth/metrics/base.py", line 21, in download_and_load_models
    model_pool.auto_load_model(
  File "/opt/tiger/Cvmm-DiffSynthSFT/diffsynth/models/model_loader.py", line 72, in auto_load_model
    model = self.load_model_file(config, path, vram_config, vram_limit=vram_limit, state_dict=state_dict)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/tiger/Cvmm-DiffSynthSFT/diffsynth/models/model_loader.py", line 41, in load_model_file
    model = load_model(
            ^^^^^^^^^^^
  File "/opt/tiger/Cvmm-DiffSynthSFT/diffsynth/core/loader/model.py", line 58, in load_model
    model.load_state_dict(state_dict, assign=True)
  File "/home/tiger/.local/lib/python3.11/site-packages/torch/nn/modules/module.py", line 2635, in load_state_dict
    raise RuntimeError(
RuntimeError: Error(s) in loading state_dict for AestheticModel:
        Missing key(s) in state_dict: "vision_model.vision_model.embeddings.class_embedding", "vision_model.vision_model.embeddings.patch_embedding.weight", "vision_model.vision_model.embeddings.position_embedding.weight", "vision_model.vision_model.pre_layrnorm.weight", "vision_model.vision_model.pre_layrnorm.bias", "vision_model.vision_model.encoder.layers.0.self_attn.k_proj.weight", "vision_model.vision_model.encoder.layers.0.self_attn.k_proj.bias", "vision_model.vision_model.encoder.layers.0

Start with custom_tools/metric/ds_aesthetic.py and diffsynth/metrics/aesthetic.py, then trace model loading through diffsynth/metrics/base.py, diffsynth/models/model_loader.py, and diffsynth/core/loader/model.py. Compare the AestheticModel configuration with the checkpoint state dict implicated by the traceback. Done means the aesthetic model loads without the reported missing-key errors.