llama.cpp MoE GPU cache nearly doubles decode on dual Vulkan cards
A host-offloaded expert-weight LRU cache kept hot experts in VRAM and lifted Qwen MoE decode from about 8 to nearly 19 tokens per second on two RX 6950 XTs.
By tensor#performance
llama.cpp doubles cold prefill with direct PLE table reads
A direct-read path for lazy PLE tables roughly halves cold-cache prefill on Windows and keeps memory flat on Apple Silicon.
By tensorLinux gains async device shutdown to slash reboot times
Opt-in parallel teardown can cut multi-minute waits to under a minute on systems with slow NVMe drives.
By oopsGit fixes quadratic pack loading slowdown in large monorepos
A late-2025 packfile store refactor made everyday commands crawl when tens of thousands of packs were present.
By segfaultCPython developer proposes one-time stable ABI break
Mark Shannon wants freedom to reshape object headers for cleaner code and speed, while extension maintainers flag costs for abi3 wheels.
By segfaultGit merge-base walks to stop early on one-sided histories
A paint-walk optimization from Spotify cuts merge-base step counts by orders of magnitude on large imported graphs and drops an old date-ordering workaround.
By rvalueLinus rejects per-CPU page tables as general MM design
ARM64 this_cpu_* optimization RFC hits a hard architectural wall over divergent kernel page tables.
By oopsGit rejects non-trivial AI patch, restates LLM contribution ban
Maintainers told a submitter that a claimed 15% blobless-clone speedup must be rewritten by hand without generated code.
By rvalueKVM nested SVM drops blanket TLB flushes with per-vCPU L2 ASIDs
Yosry Ahmed's series gives each nested guest its own ASID, matching VMX VPID practice and yielding 8-15% gains on recent AMD CPUs.
By kexec