WhisperX

[boegel]

• link to support ticket: #2024121660001058
• website: https://github.com/m-bain/whisperX
• installation docs: https://github.com/m-bain/whisperX?tab=readme-ov-file#setup-%EF%B8%8F
• toolchain: foss/2023a
• easyblock to use: PythonBundle
• required dependencies:
  • see https://github.com/m-bain/whisperX/blob/main/requirements.txt
  • FFmpeg
  • https://github.com/OpenNMT/CTranslate2 (build from source)
• notes:
  • should be installed CPU-only and with GPU support (CUDA)
• effort: (TBD)
• other install methods
  • conda: no
  • container image: no
  • pre-built binaries (RHEL8 Linux x86_64): no
  • easyconfig outside EasyBuild: no

[boegel]

WIP easyconfig available in https://github.com/vscentrum/vsc-software-stack/tree/wip/482_WhisperX

TODO: easyconfig for CTranslate2 dependency

effort so far: 30min

[laraPPr]

Their are problems caused by the dependencies of the current installation:

wrt whisperx, this is quite troublesome to make it work. the deps in the
project are not sufficient, with pip to get somethinhg that starts you
need to do

pip install whisperx 'numpy<2' nvidia-cublas-cu11 nvidia-cudnn-cu11

then whisperx does start, but doesn't do anything.

the dep hell comes from other older deps, esp the pyannote-audo vs
ctranslate2 is troublesome

[boegel]

@pavelToman Can you make an estimate how much time you've spent on this already, and how much progress you've made on it (are you stuck, almost done, etc.)?

[pavelToman]

@boegel There was a problem with CTranslate2 but it seems it works now. There will be few more hours (2-3) of work for WhisperX's deps.
But there is a problem with torchaudio - easybuilders/easybuild-framework#4706.
whisperx 3.2.0 has requirement torchaudio>=2, but you have torchaudio 0.16.2 - there is a problem with PyTorch-bundle-2.1.2-2023a-CUDA - torchaudio shows wrong version and can not pass pip check.
Until now I spent 12h with this.

EDIT: maybe it will take longer - it seems I have to create another EC for all pyannote sw - something like this bundle: easybuilders/easybuild-easyconfigs#19992

[boegel]

@pavelToman Can you already open a pull request for CTranslate2, and shift your attention to #483 instead? (and then later get back to this one)

Note that both this one and #483 are lowest priority, so please treat it as such as well.

[pavelToman]

PR for CTranslate2-4.5.0:

• {ai,tools}[GCCcore/12.3.0,foss/2023a] CTranslate2 v4.5.0, cpu_features v0.9.0 w/ CUDA 12.1.1 easybuilders/easybuild-easyconfigs#22119