GPU execution impossible in Windows 7? about whisper-standalone-win

[Purfview]

1. I dunno, you tell me if GPU processing works on Windows 7.
2. You can download cuBLAS nad cuDNN libs from there: https://github.com/Purfview/whisper-standalone-win/releases/tag/libs

Place libs in the same folder where Faster-Whisper executable is.

[MMasutin]

Sorry for reopening this. Not sure if I need a new issue. Can anyone help solve this?

Faster-Whisper r125 running on: CPU
"D:\whisper-fast\__main__.py", line 445, in <module>
"D:\whisper-fast\__main__.py", line 355, in cli
"faster_whisper\transcribe.py", line 123, in __init__
RuntimeError: mkl_malloc: failed to allocate memory
[1164] Failed to execute script '__main__' due to unhandled exception!

The files from cuBLAS.and.cuDNN.7z are where Whisper is. In NVIDIA Control Panel, only 3D Settings are available, where High-performance NVIDIA processor is applied to all. Among the settings is CUDA GPUs - All. From NVIDIA System Information:
CUDA Cores: 48
Core clock: 475 MHz
Shader clock: 950 MHz
Memory data rate: 1334 MHz
Memory interface: 64-bit
Memory bandwidth: 10.67 GB/s
Total available graphics memory: 2533 MB
Dedicated video memory: 1024 MB DDR3
System video memory: 0 MB
Shared system memory: 1509 MB
NVCUDA.DLL v7.5.15

[Purfview]

Dedicated video memory: 1024 MB DDR3

This looks very low.
What is you GPU, CPU and RAM?

Try model=tiny.

[MMasutin]

CPU processing worked with the Tiny model on 4/1GB RAM/VRAM. But GPU processing still won't work even on 8/2GB in Windows 7. I wonder if the DLLs in cuBLAS.and.cuDNN.7z are not for 7.
2) Off this topic a bit, I need time-coding more than transcription. In the SRT file, all the start times follow the end times (these are correct) immediately (e.g. End time 1 = Start time 2), resulting in impossibly long show times without interruptions. Is this due to the model used?

[Purfview]

1. What model is your GPU.
2. What version of Standalone Faster-Whisper you are using?
3. Post a command line parameters you are using.

[MMasutin]

My GPU is GF610M, 1GB. I can get details from a GPU reporting utility. I'm studying [https://docs.nvidia.com/deploy/pdf/CUDA_Compatibility.pdf](compatibility issues). My omission. A search for 'Whisper CUDA' found that you need to add '--device cuda' to enable it. Now I get Faster-Whisper r125 running on: CUDA, but errors too. The same as with 'RuntimeError: mkl_malloc: failed to allocate memory' when using the model medium with CPU processing:

"D:\whisper-fast\__main__.py", line 445, in <module>
"D:\whisper-fast\__main__.py", line 355, in cli
"faster_whisper\transcribe.py", line 123, in __init__
RuntimeError: CUDA failed with error initialization error
[3032] Failed to execute script '__main__' due to unhandled exception!
Errors when using --help:
File "D:\whisper-fast\__main__.py", line 445, in <module>
File "D:\whisper-fast\__main__.py", line 277, in cli
File "argparse.py", line 1768, in parse_args
File "argparse.py", line 1800, in parse_known_args
File "argparse.py", line 2006, in _parse_known_args
File "argparse.py", line 1946, in consume_optional
File "argparse.py", line 1874, in take_action
File "argparse.py", line 1044, in __call__
File "argparse.py", line 2494, in print_help
File "argparse.py", line 2500, in _print_message
File "encodings\cp1251.py", line 19, in encode
UnicodeEncodeError: 'charmap' codec can't encode character '\xbf' in position 88
21: character maps to <undefined>

[MMasutin]

Correction: Compatibility issues

[Purfview]

You need newer GPU.
For better timestamps get latest r128 version.

[MMasutin]

I get similar errors with Faster-Whisper r134+++ when trying to run on CUDA. Is GTX960M, 2GB VRAM, still not enough?

[Purfview]

Run what? What errors?

Maybe problem is with mobile GPU or its drivers, or whatever.

[MMasutin]

Errors (as above) with -h (help not really necessary, just testing) and this

Whisper-Faster_r134 %Audio% --language en --output_format srt --task transcribe --model tiny --device cuda

Is GTX960M, 2GB VRAM, not enough for CUDA?

[Purfview]

To solve problem with --help you may need to change language to English in Console Panel>Region and Language.

2GB is enough for tiny model.

In laptop you maybe need to use the second GPU device, try: --device cuda:1.

[MMasutin]

'cuda:1' didn't help. Thank you! I'll try to find the cause and share it if there is one.
I suspected the reason for errors with -h was the locale because of 'cp1251.py … UnicodeEncodeError'.

[MMasutin]

The title of this issue should be 'How to make it run on CUDA'.
I have so far failed to do it on three architectures of Nvidia GPUs (oldest to newest): Fermi, Maxwell, Hopper (probably Hopper as I didn't check the model, but it's built in late 2022; RuntimeError: CUDA failed with error CUDA driver version is insufficient for CUDA runtime version.) The subject matter is complex to the uninitiated. I refer anyone to this table in hopes of finding suitable drivers or anything else sooner.

[Purfview]

The title of this issue should be 'How to make it run on CUDA'.

It should be as the original issue. You shouldn't post the different issues here.

[MMasutin]

I meant a new title for this issue.