--word_timestamps True lead to Failed to launch Triton kernels warning

[zx3777]

i use 3060ti 8g gpu ，win10. cuda 11.7

when i use --word_timestamps True ，it will show this twice ：

C:\Users\xx\AppData\Local\Programs\Python\Python310\lib\site-packages\whisper\timing.py:42: UserWarning: Failed to launch Triton kernels, likely due to missing CUDA toolkit; falling back to a slower median kernel implementation...
warnings.warn(
C:\Users\xx\AppData\Local\Programs\Python\Python310\lib\site-packages\whisper\timing.py:146: UserWarning: Failed to launch Triton kernels, likely due to missing CUDA toolkit; falling back to a slower DTW implementation...

But the task can be executed and the correct srt can be obtained. How to fix this warning?

[sm1045]

Same problem here

[himberjack]

Same here

[jongwook]

This is because triton does not provide Windows binary, and I'm not sure if it's planned to do so. I should probably not print the warning on platforms other than x86_64 linux..

[jongwook]

Thanks for more information, and yes unfortunately Triton does not provide wheels for aarch64 linux either. Seems like it's technically possible to build the package which may involve rebuilding LLVM for aarch64. If the numba version works, it shouldn't be too slow while avoiding the hassle of building Triton.

[sm1045]

Thx for this information @jongwook! I will research a little more about this.

[RichoDS]

@jongwook Can i do simultaneous 2 language transcription on openai whisper?

[ryanheise]

@RichoDS this discussion is about the Triton warning, but you can try searching the discussions for your topic to find more relevant discussions. I did a search and found these: #807 #49 #849 #345 #1318 . You can pick the discussion that has attracted the most useful comments and continue that discussion, since (as you can see) it might not be convenient to continue the discussion in 5 different places.

[JosemyL95]

@zx3777 @jongwook @ryanheise
Hi guys! Do u know how can we disable the triton warnings? I'm actually using language transcriptions (with word timestamps) in Windows. I've already checked official docs and didn't saw any tip to disable the warning.
Thanks and greetings!

[aheba]

I use ubuntu, CUDA 11.7 and have the same issue +1

[taras-sereda]

same here

[zhongxinlei]

same here, I also using Ubuntu.

[okineadev]

I seem to have found a fix for the problem, here are the steps I followed:

1. Downloaded and installed the CUDA driver (12.1) for WSL from this page: https://developer.nvidia.com/cuda-12-1-0-download-archive

Note

I don't know if version 12.1 is really needed, but I installed it because pytorch comes with CUDA 12.1 support 🤷‍♂️

UPD: Still works with newer versions, including the latest (12.4)

2. Added CUDA location to $PATH:

~/.bashrc

export PATH="$PATH:/usr/local/cuda/bin"

3. Reboot

sudo reboot

---

When I installed all this - the error did not go away, BUT - I found a solution!
Here is a link to the solution - https://stackoverflow.com/a/76380867/21165921

4. Installed libnvidia-extra-530 package

sudo apt update
sudo apt install libnvidia-extra-530

---

After that the error with Triton disappeared :)

[okineadev]

Hi, Do you install CUDA driver (12.1) in the WSL environment? In WSL, there is no way to reboot. I don't understand how you do that?

wsl --shutdown
wsl --shutdown

[doggy8088]

Should I install CUDA toolkit on "Windows"? Or just install in the WSL only?

[doggy8088]

The CUDA download page seems changed. I don't know how to choose. Should I choose "Windows"? In your screenshot, you seems choose the "Windows" OS. I don't understand.

[okineadev]

The CUDA download page seems changed. I don't know how to choose. Should I choose "Windows"? In your screenshot, you seems choose the "Windows" OS. I don't understand.

https://developer.nvidia.com/cuda-12-1-0-download-archive?target_os=Linux&target_arch=x86_64&Distribution=WSL-Ubuntu

[doggy8088]

In this page, NVIDIA said users must not install any NVIDIA GPU Linux driver within WSL 2.