megakernel 0.1.0

Measure the kernel-launch tax and memory ceiling a megakernel would remove. Future home of a residency-first megakernel compiler.

megakernel

A megakernel compiles a whole computation — up to an entire model's forward pass — into a single persistent GPU kernel: one launch, one resident working set, an in-kernel scheduler, instead of thousands of separate kernel launches bouncing activations through device memory.

Today this package ships a probe that measures, on your GPU, the two overheads a megakernel removes:

pip install megakernel
python -m megakernel

device               : <your GPU>
kernel enqueue cost  :    x.xx us  (CPU-side submit)
kernel launch tax    :    x.xx us  (end-to-end, tiny kernel)
device copy bandwidth:  xxxx.x GB/s

A decode step composed of ~500 kernel launches pays roughly y.yy ms/token
in launch tax alone, before any math or memory traffic. A megakernel pays
it once.

Requires a CUDA build of PyTorch (the probe reports a clear message otherwise). Also usable as a library: from megakernel import probe.

Roadmap

This name is the future home of a residency-first megakernel compiler currently in development: real models compiled to single persistent kernels, adoptable without rewriting anything. Watch this package for the first public release.

Contact: [e-mail]