Had various issues recently, likely related to packages

[mbrownnycnyc]

Prepended the following to the install block in colab:

!pip install ctranslate2==4.4.0
!pip install lameenc
!pip install openunmix
!pip install tensorrt
!pip install dora-search

Most of these were googled, but @Leandrocnf helped out on this thread: m-bain/whisperX#902 (comment)

I don't know if this is worthy of a PR, but I wanted to share.

Thank you Mahmoud!

[MahmoudAshraf97]

Hello, I added the first line to the notebook as it was needed for cuda compatibility, I don't think the others are needed anywhere

[anselbrandt]

I agree it's only ctranslate2. I would add that to contraints.txt. I think torch and torchaudio also need to be version locked to 2.3, otherwise this error results:

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
Aborted (core dumped)

torch 2.4 and up installs libcudnn_ops_infer.so.9

I can open a PR with the following additions to contraints.txt:

ctranslate2<=4.4.0
torch<=2.3
torchaudio<=2.3

[MahmoudAshraf97]

Cuda versions compatibility is not in the scope of requirements file, this is up for the user to decide, this project works with other torch versions than 2.3 so it should not be limited to 2.3 because of cuda constraints that are the user responsibility

[homelab-00]

I think torch and torchaudio also need to be version locked to 2.3, otherwise this error results

The script works fine for me with torch 2.5.1 and CUDA 12.4 (that's the latest CUDA version supported by torch).

[eugr]

I had to install additional requirements that were not in any of the requirement files (except for the first one that is mentioned in this and some other issues):

pip install "torch<2.4" "torchaudio<2.4" "ctranslate2<4.5"
pip search tensorrt_libs
pip install tensorrt
pip install nvidia-cuda-nvcc-cu12

[MahmoudAshraf97]

I really don't understand how these packages are needed, they are not included because they are not used or needed, there is a CI action that runs periodically to test the needed requirements and it works out of the box with the current installation instructions

[eugr]

Well, I guess I was rushing things and didn't pay much attention to what libraries you included in setting up LD_LIBRARY_PATH in another issue ticket:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:`python3 -c 'import os;import tensorrt_libs; import nvidia.cuda_runtime.lib; import nvidia.cuda_cupti.lib; import nvidia.cuda_nvcc.lib; import nvidia.cuda_nvrtc.lib; import nvidia.cudnn.lib; import nvidia.cublas.lib; import nvidia.cufft.lib; import nvidia.curand.lib; import nvidia.cusolver.lib; import nvidia.cusparse.lib; print(f"{os.path.dirname(nvidia.cudnn.lib.__file__)}:{os.path.dirname(nvidia.cublas.lib.__file__)}:{os.path.dirname(nvidia.cufft.lib.__file__)}:{os.path.dirname(nvidia.curand.lib.__file__)}:{os.path.dirname(nvidia.cusolver.lib.__file__)}:{os.path.dirname(nvidia.cusparse.lib.__file__)}:{os.path.dirname(nvidia.cuda_runtime.lib.__file__)}:{os.path.dirname(nvidia.cuda_cupti.lib.__file__)}:{os.path.dirname(nvidia.cuda_nvcc.lib.__file__)}:{os.path.dirname(nvidia.cuda_nvrtc.lib.__file__)}:{os.path.dirname(tensorrt_libs.__file__)}")'`