ASR推理错误

[Carey8175]

trafficstars

Could not load library libcudnn_ops_infer.so.8. Error: libcudnn_ops_infer.so.8: cannot open shared object file: No such file or directory

系统：ubantu cuda:12.3 torch:2.1.1

[BOCEAN-FENG]

cudnn安了吗，没写到感觉就是没有安装

[Carey8175]

安装后报错： Traceback (most recent call last): File "/home/carey/GPT-SoVITS/tools/asr/fasterwhisper_asr.py", line 77, in execute_asr for segment in segments: File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 1106, in restore_speech_timestamps for segment in segments: File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 511, in generate_segments encoder_output = self.encode(segment) File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 762, in encode return self.model.encode(features, to_cpu=to_cpu) RuntimeError: parallel_for failed: cudaErrorNoKernelImageForDevice: no kernel image is available for execution on the device

但是中文可以正常运行，我用的多语种模型的日语就会报这个错

[Carey8175]

GPU：tesla M40

[Carey8175]

noavx

[Carey8175]

达摩可以运行，但是fastwhisper不行

[jzeng703]

LD_LIBRARY_PATH加上环境对应的libcudnn_ops_infer.so.8路径就好了

[Carey8175]

LD_LIBRARY_PATH加上环境对应的libcudnn_ops_infer.so.8路径就好了

加上了过后，达摩能够使用，但是多语种那个不能使用，报错： Traceback (most recent call last): File "/home/carey/GPT-SoVITS/tools/asr/fasterwhisper_asr.py", line 77, in execute_asr for segment in segments: File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 1106, in restore_speech_timestamps for segment in segments: File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 511, in generate_segments encoder_output = self.encode(segment) File "/home/carey/anaconda3/envs/voice/lib/python3.9/site-packages/faster_whisper/transcribe.py", line 762, in encode return self.model.encode(features, to_cpu=to_cpu) RuntimeError: parallel_for failed: cudaErrorNoKernelImageForDevice: no kernel image is available for execution on the device

[SapphireLab]

M40 可以用 cuda12 吗？感觉会有兼容性问题. 报错前面有没有出现半精度的警告, 有的话把 config.py 中的 is_half 改了, 这卡应该不支持半精度运算.

[Carey8175]

M40 可以用 cuda12 吗？感觉会有兼容性问题. 报错前面有没有出现半精度的警告, 有的话把 config.py 中的 is_half 改了, 这卡应该不支持半精度运算. Nvidia官网下载的Driver，支持cuda12，如果是cuda问题，那么达摩应该也跑不了啊，但是我可以运行。出现了is_half的问题，我已经改过了is_half

[KaiWei-Seu]

可能的解法

1. 查看是否存在libcudnn_ops_infer.so.8；

find / -type f -name "*ibcudnn_cnn_infer*" -o -type d -name "*ibcudnn_cnn_infer*"

结果发现在这里：/usr/local/lib/python3.9/site-packages/torch/lib/libcudnn_cnn_infer.so.8 2. 添加路径，执行：

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib/python3.9/site-packages/torch/lib/

我这边就ok了

[xia-huang-411303]

可能的解法

1. 查看是否存在libcudnn_ops_infer.so.8；

find / -type f -name "*ibcudnn_cnn_infer*" -o -type d -name "*ibcudnn_cnn_infer*"

结果发现在这里：/usr/local/lib/python3.9/site-packages/torch/lib/libcudnn_cnn_infer.so.8 2. 添加路径，执行：

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib/python3.9/site-packages/torch/lib/

我这边就ok了

我这边使用 find 命令找不到 ibcudnn_cnn_infer，使用如下代码获取的 cudnn 路径

import os
import nvidia
print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))

然后在就是添加环境变量

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/user/miniconda3/envs/GPTSoVits/lib/python3.10/site-packages/nvidia/cublas/lib:/home/user/miniconda3/envs/GPTSoVits/lib/python3.10/site-packages/nvidia/cudnn/lib

问题解决。 相同的问题在 faster-whisper 项目的 issues 中也有发现 https://github.com/SYSTRAN/faster-whisper/issues/516

如果提示提示无法找到名为 nvidia 的模块 则需要运行以下命令以安装 nvidia-pyindex 和 CUDA 相关的库：

pip install nvidia-pyindex
pip install nvidia-cublas-cu11
pip install nvidia-cudnn-cu11