llama.cpp MoE cache work and PyTorch maintenance

llama.cpp saw concrete performance discussion around mixture-of-experts offload caching, while PyTorch opened a point-release tracker and a device-backend documentation request. Activity stayed technical and focused on efficiency and interface clarity.

GPU-resident LRU cache for MoE expert weights in llama.cpp

A proposal in ggml-org/llama.cpp covers a GPU-resident LRU cache for host-offloaded mixture-of-experts expert weights. Users report 20-30% decode gains on Vulkan and AMD configurations, and one participant asked about adopting the stalled pull request. The change targets developers running large MoE models under GPU memory pressure who need higher decode throughput without full weight residency.

PyTorch 2.14.1 release tracker opened

The pytorch/pytorch repository opened a release tracker for version 2.14.1. The tracker cites critical CUDA fixes delivered through linked pull requests. Maintainers and users of CUDA builds gain a clear signal that stability patches are being collected for the point release.

Request to document PyTorch device module APIs

An issue in pytorch/pytorch asks for public documentation of the APIs a new device module must support. The request seeks an explicit contract for new backends. Authors adding device support benefit from a defined interface rather than reverse-engineering existing modules.