[POC] Agentic CUDA kernel generation #1111
Description
What
Develop an agent that writes performant & correct CUDA kernel code.
Why
CUDA kernel in our project needs to support multiple things (eg. f64, f32 use different kernels w/ multiple encoding methods, which is 2 * num_of_encoding_methods). It's not cool to write it all by-our-selves and strike a balance between performance and code complexity.
Therefore, in era of agentic AI, we can create a agentic system to address them!
How
- Create relevant skill for kernel generation.
- kernels should supports CUDA (prefer support everything in cuda ecosystem)