Hello, I'm delighted to see your excellent idea. Using Modal and Tinker can indeed significantly reduce the cost of training GPU kernels with RL.
I'm interested in whether you would be willing to collaborate with us to run the Eval/Benchmark functionality on AMD's free Developer Cloud (we can support this in a similar way to Modal), which would allow us to train models suitable for the AMD platform. For KernelBench, this would undoubtedly enrich the framework's cross-platform capabilities and potentially broaden its applicability!

Looking forward to your comments and reply!