[ERROR] RuntimeError: Error(s) in loading state_dict for ESPnetS2TModel

[SimZhou]

Error while loading checkpoint:

Traceback (most recent call last):
  File "/home/yihua.zhou/tools/open_source/audioLM/Dolphin/Dolphin-predict.py", line 38, in <module>
    model = dolphin.load_model("small", "/hpc_stor01/project/ezkws/ai-aps/DataoceanAI/dolphin-small", device)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/dolphin/transcribe.py", line 184, in load_model
    model = DolphinSpeech2Text(
            ^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/typeguard/__init__.py", line 1033, in wrapper
    retval = func(*args, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/dolphin/model.py", line 96, in __init__
    s2t_model, s2t_train_args = self.build_model_from_file(s2t_train_config, s2t_model_file, device)
                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/typeguard/__init__.py", line 1033, in wrapper
    retval = func(*args, **kwargs)
             ^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/dolphin/model.py", line 241, in build_model_from_file
    model.load_state_dict(state_dict, strict=False)
  File "/opt/conda/lib/python3.11/site-packages/torch/nn/modules/module.py", line 2624, in load_state_dict
    raise RuntimeError(
RuntimeError: Error(s) in loading state_dict for ESPnetS2TModel:
        While copying the parameter named "normalize.mean", whose dimensions in the model are torch.Size([80]) and whose dimensions in the checkpoint are torch.Size([80]), an exception occurred : ('CUDA error: no kernel image is available for execution on the device\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\n',).
        While copying the parameter named "normalize.std", whose dimensions in the model are torch.Size([80]) and whose dimensions in the checkpoint are torch.Size([80]), an exception occurred : ('CUDA error: no kernel image is available for execution on the device\nCompile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.\n',).

My torch version: 2.8.0

[SimZhou]

Ok to use cpu.
But error on cuda.

[SimZhou]

Solved.

It seems cuda 12.8 is not supported by espnet==202402

Downgrading pytorch from 2.8.0+cu128 to 2.7.1+cu126 will do.