AI-written pull requests have increased the work for open-source maintainers, while the number of maintainers stays the same.

A fund that pays projects and gives core Vite maintainers hourly stipends, chosen by a panel that includes people outside Cloudflare.

Mixture-of-experts models must keep every expert in memory, and existing pruning methods judge each expert alone, ignoring groups that work together.

HOPE, a pruning method that scores pairs of experts so it avoids breaking up groups that cooperate.

When a model writes a few words at a time, the GPU runs many short programs in turn and sits idle between them.

FlyDSL mono kernels, written in Python, that combine stages of a layer and overlap weight loading and computation, used in AMD's ATOM serving software.

In PyTorch, one operation can become many small GPU programs, and first-run compiling can distort measurements.

A step-by-step profiling guide on a small protein-folding model, comparing fused PyTorch code, torch.compile and custom Triton code.