You can choose a sample problem or enter in your own PyTorch. It generates CUDA, compiles and validates the results, then uses all of this as feedback into the next generation attempt. If the generated kernel didn’t work, it tries to fix it. If it did work, it tries to make it faster.
You can configure the agent with a model/hardware/compute budget. The optional custom prompt section gives you the opportunity to add more context to the agent. You can include specific documentation, tutorials, or examples. You can also identify specific optimization techniques that you would like to see applied. Prompts can be as short as one line, or even multiple pages long.
Let me know what you think.