Preparations (in addition to cu* DLLs) for running on CUDA

[MMasutin]

Just in case, I selected to post a question, not an issue.
I have so far failed to do it on different Nvidia GPUs, both older and quite recent. I don't understand versions required, i.e. compatibility issues with drivers and other files. Are there useful tips?
2) If an additional question is possible here, is there a variable to set 'minimum duration of pause/silence' that would mark a next segment?

[Purfview]

You didn't posted what is the issue with "CUDA". You can use libs from there -> https://github.com/Purfview/whisper-standalone-win/releases/tag/libs

2. If an additional question is possible here, is there a variable to set 'minimum duration of pause/silence' that would mark a next segment?

Use --sentence, for it there is sub-parameter --max_gap. It's 3.0 seconds by default.

[MMasutin]

Thanks for your reply! The issue is how to make the CUDA capabilities available, in general. Not an issue with the program but a question of how to ensure that the Nvidia GPU is recognized correctly. I know about the libs. It's not enough to add them, and I don't understand what other steps are required. Must the versions of some files be compatible? Is there a kind of check list? For example, I check the version of the Nvidia driver and then what?

[Purfview]

Not an issue with the program

Then I'm converting this to the discussion.

ensure that the Nvidia GPU is recognized correctly

I think, it's recognized or not, if not then post what error you get.

[hnorgaar]

Best way is to to add the CUDA files to Windows/system32, else you need to add the path to Windows enviroment variables for where they are stored

[MMasutin]

It doesn't help having the cu* DLLs in %WinDir%\System32 (What about a 64-bit system?) or a folder in the %Path% variable. Using --checkcuda returns 0. Tried on two GPUs. And the errors, when trying to run a trascription job on CUDA, may be typical.
One of the GPUs I tried has the driver version 353.62. The other pieces of NV software may need to correspond to this. If CUDA doesn't work by simply putting the DLLs along with Whisper-Faster, then I don't understand the whole process of finding all the pieces of compatible software.
Traceback (most recent call last): File "D:\whisper-fast\__main__.py", line 1104, in <module> File "D:\whisper-fast\__main__.py", line 972, in cli File "faster_whisper\transcribe.py", line 136, in __init__ RuntimeError: CUDA failed with error initialization error [4084] Failed to execute script '__main__' due to unhandled exception!

[Purfview]

Using --checkcuda returns 0. Tried on two GPUs.

What GPUs? NVIDIA GPUs should have Compute Capability greater or equal to 3.5.

One of the GPUs I tried has the driver version 353.62

That version doesn't support CUDA 11, it was introduced with 452.39, so you need that or newer driver.

[MMasutin]

Thanks for your replies!
Thus prerequisites are:
CUDA Capability 3.5 or greater (according to this, present in NV family Kepler or newer);
Support of CUDA Toolkit 11.x, i.e. the driver version 452.39 or newer.
How do I best go about verifying if these can be met?
First, I probably do a search for the GPU model to see if it's a compatible architecture (at least Kepler).
Then, where or how do I search for a new driver? It's not what comes with an installation pack for a given GPU?

[Purfview]

How do I best go about verifying if these can be met?

Maybe with tool like Aida64.

Then, where or how do I search for a new driver?

Official: https://www.nvidia.com/Download/Find.aspx?lang=en-us
Here you can find info how to mod drivers: https://forums.guru3d.com/forums/videocards-nvidia-geforce-drivers-section.21/

[1001ruchka]

GPU-Z
https://www.techpowerup.com/gpuz/