Field note
Meta's MSLK repository is a reminder that the performance frontier is not only model architecture.
The repo describes a collection of PyTorch GPU operator libraries designed and optimized for GenAI training and inference, including operations such as FP8 row-wise quantization and collectives. That kind of work sits below the visible product layer, but it determines how quickly and cheaply models can be trained or served.
For New Runtime, this is infrastructure worth tracking because reusable kernels make optimization portable. Instead of every team rebuilding the same performance hacks inside a private stack, the low-level work can become a shared library that model teams, serving teams, and research engineers can compose.