Skip to content

Instantly share code, notes, and snippets.

@refabr1k

refabr1k/setting_up_whisper_missing_fbgemm_dll.md

Last active August 16, 2024 02:06
Show Gist options
  • Save refabr1k/09fbfcdcc97e1e58e5946f61a971c11d to your computer and use it in GitHub Desktop.
Save refabr1k/09fbfcdcc97e1e58e5946f61a971c11d to your computer and use it in GitHub Desktop.
Download ZIP

Revisions

  1. refabr1k renamed this gist Aug 16, 2024. 1 changed file with 0 additions and 0 deletions.
    0 readme.md → setting_up_whisper_missing_fbgemm_dll.md
    View file
    File renamed without changes.
  2. refabr1k revised this gist Aug 16, 2024. No changes.
  3. refabr1k revised this gist Aug 16, 2024. No changes.
  4. refabr1k created this gist Aug 16, 2024.
    65 changes: 65 additions & 0 deletions readme.md
    View file
    Original file line number Diff line number Diff line change
    @@ -0,0 +1,65 @@
    # Setting up Torch with Cuda (if you have a nvidia gfx card)

    1. Install CUDA Toolkit -> you should install the CUDA version that PyTorch supports for example at this time, Version CUDA 12.4 is the latest.
    ![image](https://gist.github.com/user-attachments/assets/10c76a56-380b-404c-a24f-25a26b62cca1)

    navigate to https://developer.nvidia.com/cuda-12-4-0-download-archive download and install the CUDA Toolkit version

    2. Check that CUDA Toolkit is installed > issue `nvcc --version` and observe from output that the installed cuda version is detected

    # Installing Torch
    1. Navigate to https://pytorch.org/get-started/locally/ and select the appropriate options for your system (remember to chose the same CUDA version you installed for "Compute Platform"
    2. Run the command
    3. Test that Torch is installed

    ```
    >>> import torch
    >>> torch.cuda.is_available()
    True
    >>> torch.cuda.device_count()
    1
    >>> torch.cuda.current_device()
    0
    >>> torch.cuda.device(0)
    <torch.cuda.device at 0x7efce0b03be0>
    >>> torch.cuda.get_device_name(0)
    'GeForce GTX 950M'
    ```
    4. If there is an issue see below
    (reference: https://github.com/pytorch/pytorch/issues/131662)

    # Torch dependency issue: (Missing fbgemm.dll) OSError: [WinError 126] The specified module could not be found.
    Solution here:
    https://github.com/pytorch/pytorch/issues/131662#issuecomment-2252589253

    1. Install Visual Studio Community 2022
    2. Tools > Get Tools and Features
    3. Individual Components tab > VS 2022 C++ ... (latest)

    ![352517926-81707c5d-d56e-49ce-8677-460a6c423d62](https://gist.github.com/user-attachments/assets/40251c17-5625-4b17-932b-233cf497ede9)


    ---

    # Install whisper
    https://github.com/openai/whisper

    ```bash
    whisper ".\New Recording 50.m4a" --model large-v3 --language=en --threads=4
    # [00:00.320 --> 00:05.520] what is the you know the highest cost or all the all the cost that is causing all the
    # [00:05.520 --> 00:11.760] expensiveness like and see whether from the industrial solution we choose uh you know shed


    whisper ".\New Recording 40.m4a" --language Chinese --model large-v3 --threads 4 --output_format txt
    # [00:00.000 --> 00:02.160] 嗨,今天天气很好

    # to steer ai to translate to simplified chinese (use initial_prompt with simplified chinese)
    whisper ".\New Recording 40.m4a" --language Chinese --model large-v3 --threads 4 --output_format txt --initial_prompt '以下是普通话的句子'
    ```