All pull requests

Pull requests

• llama: keep Vulkan MoE experts resident in mixed runs

  #109·luisantich opened Sep 25, 2026·

  ·
  1

• cuda: opt-in segmented TopK for multi-row inputs (upstream reference)

  CUDAggmltesting
  #108·GenerelSchwerz opened Sep 24, 2026Owner·

  ·

• speculative : capped recurrent planes for DFlash (--spec-draft-rs-planes)

  CUDAdocumentationggmlmodelservertestingVulkan
  #91·Piggidragon opened Sep 14, 2026·

  ·

• ggml-alloc : share meta compute buffers through a borrower alias

  ggmltesting
  #88·Piggidragon opened Sep 13, 2026·

  ·

• qwen4exp: prefetch requested lazy PLE pages (upstream reference)

  model
  #86·GenerelSchwerz opened Sep 10, 2026Owner·

  ·

• cuda: opt-in quantized-native MMA FlashAttention

  CUDAdevopsdocumentationggmltesting
  #85·Piggidragon opened Sep 10, 2026·

  ·

• cuda: Windows MoE pin budget - layer-subset routing + chunked pinned store

  CUDAggmltesting
  #84·jasonlnheath opened Sep 10, 2026·

  ·
  4

• cuda/moe-cache: pinned-alloc fallback puts experts on the CPU backend

  CUDAggml
  #83·LokenSI opened Sep 10, 2026·

  ·
  1

• cuda : add experimental bounded MoE host pinning

  CUDAdocumentationggmlservertesting
  #76·GenerelSchwerz opened Sep 8, 2026Owner·

  ·
  1

• kv-cache : spend the partial residency budget on the slowest link first

  documentationexamplestesting
  #68·Piggidragon opened Sep 3, 2026·

  ·
  1

• kv-cache : resolve the partial KV residency set once for the model

  documentationexamplestesting
  #67·Piggidragon opened Sep 3, 2026·

  ·
  2

• llama-bench placement flags, two host-KV fixes, and H1/H2/H3/H5/H13 measured

  documentationexamplesggml
  #62·Piggidragon opened Sep 2, 2026·

  ·

• ggml-cuda: protect reused MoE cache slots

  #47·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• ggml-cuda: coalesce contiguous staging copies

  #46·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• ggml-cuda: prefetch overflow expert siblings

  #45·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• ggml-cuda: reuse cached experts during overflow staging

  #44·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• ggml-cuda: skip duplicate decode sibling prefetch

  #43·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• ggml-cuda : reuse routing IDs across MoE siblings

  #42·GenerelSchwerz opened Aug 26, 2026Owner·

  ·

• sched: pipeline the delivery of a host-resident KV cache

  documentationexamplesggmltesting
  #39·Piggidragon opened Aug 26, 2026·

  ·
  6

• kv: replace eligible dense causal masks with compact prefixes

  CUDAdocumentationggmlservertesting
  #7·GenerelSchwerz opened Aug 21, 2026Owner·

  ·