AI coding agents can send many changes to the same repository at the same moment, which strains how code is stored.

A new Git storage architecture designed so thousands of agents can push to a single repository at once.

Coding agents send long, growing prompts and run tasks in parallel, which fills the server's memory of earlier context.

Better caching, including moving cached data to host memory, faster decoding code and fixes sent upstream to the SGLang serving software.

Developers need production-ready tools for Nvidia's new Vera Rubin GPUs and finer control over how jobs share a GPU.

CUDA 13.5 with general availability for Vera Rubin, a new Multi-Process Service that lets workloads borrow and lend GPU cores, and batched memory copies in CUDA Graphs.

In PyTorch, one operation can turn into many small GPU tasks, and early steps measure the compiler instead of the model.

A step-by-step profiling workflow on a small protein-folding model, comparing plain code, fused code and custom Triton kernels.