llama.cpp Vulkan fused ops and PyTorch XPU cache control

Two technical proposals advanced in open ML runtimes, one adding fused Vulkan kernels for a new model family and another seeking finer control of Intel XPU primitive caching. Both target practical performance and workload flexibility rather than new model releases.

llama.cpp adds DeepSeek-V4 fused hyper-connection ops on Vulkan

A pull request in ggml-org/llama.cpp introduces fused hyper-connection operators (DSV4_HC_COMB, PRE, and POST) for DeepSeek-V4 on the Vulkan backend. The change reports 1.8-2x gains in decode and prefill throughput when exercised over multi-node RPC. Readers tracking local and distributed inference will care because the kernels directly improve efficiency for an emerging model architecture on a widely used portable backend.

PyTorch XPU proposal for runtime oneDNN primitive cache control

A proposal in pytorch/pytorch requests runtime control over the embedded oneDNN primitive cache used by the XPU path. The goal is to let applications adjust caching behavior across workload phase changes instead of relying on fixed embedded defaults. This matters for developers running varied or multi-stage Intel GPU workloads who need to avoid stale or oversized primitive caches without rebuilding the framework.