运行web.py，更换synthesizer模型报错

只有 pretrained_75k.pt 正确，选其它模型（如：my_run8_25k.pt 、qh3_53k.pt 、rty4_87k.pt 等），都报 torch.Size 错误。信息如下：

RuntimeError: Error(s) in loading state_dict for Tacotron: size mismatch for gst.stl.attention.W_query.weight: copying a param with shape torch.Size([512, 256]) from checkpoint, the shape in current model is torch.Size([512, 512]).
Traceback:
File "/home/armstrong/miniconda3/envs/torch2/lib/python3.11/site-packages/streamlit/runtime/scriptrunner/script_runner.py", line 565, in _run_script
exec(code, module.__dict__)
File "/tmp/tmp69jcq9tj.py", line 13, in
render_streamlit_ui()
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/control/mkgui/base/ui/streamlit_ui.py", line 909, in render_streamlit_ui
session_state.output_data = opyrator(input=input_data_obj)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/control/mkgui/base/core.py", line 203, in __call__
return self.function(input_obj, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/control/mkgui/app.py", line 140, in synthesize
specs = current_synt.synthesize_spectrograms(texts, embeds)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/models/synthesizer/inference.py", line 91, in synthesize_spectrograms
self.load()
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/models/synthesizer/inference.py", line 69, in load
self._model.load(self.model_fpath, self.device)
File "/mnt/f/AI/digitalhuman/app/voices/mockingbird/src/models/synthesizer/models/base.py", line 55, in load
self.load_state_dict(state, strict=False)
File "/home/armstrong/miniconda3/envs/torch2/lib/python3.11/site-packages/torch/nn/modules/module.py", line 2041, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(

Start with models/synthesizer/inference.py and models/synthesizer/models/base.py, following the load path shown in the traceback. Compare the checkpoint dimensions for pretrained_75k.pt and the failing synthesizer models, especially gst.stl.attention.W_query.weight. Done means supported model files load without the reported size mismatch and synthesis works for them.

Written by the indexing model from the issue text.