Running whisper with ROCm terminates with SIGSEGV (Address boundary error)

[merll002]

I am trying to use whisper to transcribe an audio file on Ubuntu 24.04.2 LTS with AMD PRO + ROCm installed.

I did the following:

• pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3
• pip3 install openai-whisper
• whisper audio/test.wav --model turbo --device cuda
• Then it exists with: 'whisper audio/test.wav --model …' terminated by signal SIGSEGV (Address boundary error)

CPU only works fine.

Any ideas?

Using Python 3.12.3

Cuda devices list:

>>> import torch
>>> for i in range(torch.cuda.device_count()):
...    print(torch.cuda.get_device_properties(i).name)
... 
AMD Radeon RX 7800 XT

[merll002]

Solved! Upgraded to PyTorch Nightly and upgraded Numba.