maharshi-pandya/megacute is a library that builds transformer operation graphs with CuTe DSL and handles the operations of several layers in a single persistent execution kernel. The repository listing has no formal release version. You define the graph and the operations in Python, and a compiler generates the execution program. So this is not about changes in a particular version; it is an introduction to a project that describes graphs in Python and runs them with CuTe DSL-based kernels.
A megakernel is a way of bundling several operations into one kernel and running them. megacute targets transformer graphs that include attention, normalization, and the output layer, and the README has Llama and DiT examples. The README compares measurements obtained on a B200 with other inference tools. The focus of the project is to express the whole graph through a Python API and to bundle several operations into a persistent execution kernel. The repository listing states Python and an Apache-2.0 license.
That said, the comparison is presented as B200 measurements, and the README explains that with long context and large batches other tools may get ahead because of the limits of the attention operation. So this material alone does not allow judging superiority for every input size or across GPUs. The provided description states that it targets a particular graphics card family, but does not present evidence that it is a general-purpose inference tool. What can be confirmed at present is a CuTe DSL-centered execution path and examples, plus limited comparison results, and it is hard to conclude that it is a replacement for a general-purpose inference server.