Loading SenseVoice with FunASR crashe

[grihenrik]

Loading SenseVoice with FunASR crashes with:

AttributeError: 'SinusoidalPositionEncoder' object has no attribute '_state_dict_pre_hooks'

This happens during model initialization (before inference finishes), and prevents transcription jobs from completing.

To Reproduce
Build and run SenseVoice service.
Install/run FunASR path that loads iic/SenseVoiceSmall.
Execute:
or in app context:

Observe crash trace:
Expected behavior
Model loads successfully and transcription runs normally.

Actual behavior
Model load fails with _state_dict_pre_hooks AttributeError in SinusoidalPositionEncoder.

Environment
SenseVoice model: iic/SenseVoiceSmall
Backend: FunASR (AutoModel)
Runtime: Python 3.11
PyTorch runtime observed: 2.4.1+cu121
Containerized deployment (Docker)
Additional context
The issue appears related to SinusoidalPositionEncoder.init not initializing torch.nn.Module state (missing super().init()), which leads to missing internal hook fields used by state_dict().

[LauraGPT]

Thanks for reporting this! This is a known compatibility issue with certain PyTorch versions.

Root cause: SinusoidalPositionEncoder inherits from torch.nn.Module but some older model checkpoints or custom implementations may not call super().__init__() properly, which causes missing _state_dict_pre_hooks in newer PyTorch versions (2.4+).

Workaround: Try upgrading to the latest FunASR version:

pip install --upgrade funasr

The latest version (v1.3.9) includes compatibility fixes for newer PyTorch versions.

If the issue persists, could you share:

1. Your exact funasr version (pip show funasr)
2. Your exact torch version (pip show torch)
3. The full stack trace

This will help us pinpoint the specific incompatibility.

[staccat0]

试试把model.py中

def __init__(self, d_model=80, dropout_rate=0.1):
    pass

修改成以下内容

def __init__(self, d_model=80, dropout_rate=0.1):
    super().__init__()