megakernel
v0.1.0
Published
Measure the kernel-launch tax and memory ceiling a megakernel would remove. Future home of a residency-first megakernel compiler.
Maintainers
Readme
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.
This npm name reserves the identity for a residency-first megakernel
compiler currently in development. The measurement probe and compiler
tooling ship from the Python package
megakernel today; watch this
package for a JS / native release.
Contact: [email protected]
