[BUG]: from torchaudio.functional.functional import _hz_to_mel, _mel_to_hz Error

[stevezhang88]

Describe the bug

运行demo代码时，在from import阶段就报错了。

How To Reproduce

from models.tts.maskgct.maskgct_utils import *
这句报错了

---

OSError Traceback (most recent call last)
Cell In[5], line 1
----> 1 from models.tts.maskgct.maskgct_utils import *
2 from huggingface_hub import hf_hub_download
3 import safetensors

File ~/autodl-tmp/Amphion/models/tts/maskgct/maskgct_utils.py:20
17 from utils.util import load_config
18 from tqdm import tqdm
---> 20 from models.codec.kmeans.repcodec_model import RepCodec
21 from models.tts.maskgct.maskgct_s2a import MaskGCT_S2A
22 from models.tts.maskgct.maskgct_t2s import MaskGCT_T2S

File ~/autodl-tmp/Amphion/models/codec/kmeans/repcodec_model.py:15
12 from einops import rearrange, repeat
14 from models.codec.amphion_codec.quantize import ResidualVQ
---> 15 from models.codec.kmeans.vocos import VocosBackbone
18 def init_weights(m):
19 if isinstance(m, nn.Conv1d):

File ~/autodl-tmp/Amphion/models/codec/kmeans/vocos.py:14
12 from torch import nn
13 from torch.nn.utils import weight_norm, remove_weight_norm
---> 14 from torchaudio.functional.functional import _hz_to_mel, _mel_to_hz
17 def safe_log(x: torch.Tensor, clip_val: float = 1e-7) -> torch.Tensor:
18 """
19 Computes the element-wise logarithm of the input tensor with clipping to avoid near-zero values.
20
(...)
26 Tensor: Element-wise logarithm of the input tensor with clipping applied.
27 """

File ~/miniconda3/lib/python3.12/site-packages/torchaudio/init.py:2
1 # Initialize extension and backend first
----> 2 from . import _extension # noqa # usort: skip
3 from ._backend import ( # noqa # usort: skip
4 AudioMetaData,
5 get_audio_backend,
(...)
10 set_audio_backend,
11 )
13 from . import ( # noqa: F401
14 compliance,
15 datasets,
(...)
23 utils,
24 )

File ~/miniconda3/lib/python3.12/site-packages/torchaudio/_extension/init.py:38
36 _IS_ALIGN_AVAILABLE = False
37 if _IS_TORCHAUDIO_EXT_AVAILABLE:
---> 38 _load_lib("libtorchaudio")
40 import torchaudio.lib._torchaudio # noqa
42 _check_cuda_version()

File ~/miniconda3/lib/python3.12/site-packages/torchaudio/_extension/utils.py:60, in _load_lib(lib)
58 if not path.exists():
59 return False
---> 60 torch.ops.load_library(path)
61 return True

File ~/miniconda3/lib/python3.12/site-packages/torch/_ops.py:1032, in _Ops.load_library(self, path)
1027 path = _utils_internal.resolve_library_path(path)
1028 with dl_open_guard():
1029 # Import the shared library into the process, thus running its
1030 # static (global) initialization code in order to register custom
1031 # operators with the JIT.
-> 1032 ctypes.CDLL(path)
1033 self.loaded_libraries.add(path)

File ~/miniconda3/lib/python3.12/ctypes/init.py:379, in CDLL.init(self, name, mode, handle, use_errno, use_last_error, winmode)
376 self._FuncPtr = _FuncPtr
378 if handle is None:
--> 379 self._handle = _dlopen(self._name, mode)
380 else:
381 self._handle = handle

OSError: /root/miniconda3/lib/python3.12/site-packages/torchaudio/lib/libtorchaudio.so: undefined symbol: _ZNK3c105Error4whatEv

Environment Information

linux 环境 在autodl上运行的。

Additional context

(Add any other context about the problem here.)

[yuantuo666]

Hi, it seems the problem is with the libtorchaudio installation. Which command did you use to install the pytorch?
Maybe check the official docs for reference: https://pytorch.org/get-started/locally/#with-cuda

BTW, we prefer English issues so that more people around the world can participate and it is also easier for issue search.

[stevezhang88]

Oh. I use mirror on the server of AUTODL.com. I choose one 4090 GPU and pytorch==2.3.0 cuda=12.1

[yuantuo666]

It is possible to have environmental problems, so it is recommended to use conda to manage your environment and install the dependency from scratch following the provided env.sh file.

Miniconda Install: https://docs.anaconda.com/miniconda/miniconda-install/
Conda User Guide: https://docs.conda.io/projects/conda/en/latest/user-guide/getting-started.html