All pull requests

Pull requests

• vulkan: let the subgroup GEMVs read the device's real subgroup size

  #1847·crichalchemist opened Oct 3, 2026Contributor·

  ·
  1

• vulkan: correct expert-batch results under MoltenVK's default mode

  #1846·crichalchemist opened Oct 3, 2026Contributor·

  ·
  1

• fix(api): reject unsupported modalities

  #1838·Stamina9 opened Oct 2, 2026Contributor·2 tasks done·

  ·

• Feat/tier demand eviction core

  #1826·KhalidAlnujaidi opened Oct 1, 2026·6 tasks·

  ·

• fix(loader): refuse wrong-cardinality tensors before allocating or reading

  #1824·Xore opened Oct 1, 2026·

  ·
  1

• fix(qwen36): honor RAM_GB as a real ceiling for the expert cache

  #1823·Xore opened Oct 1, 2026·

  ·
  1

• feat(glm53): handle cancellation between layers of a prompt chunk

  #1822·enitimeago opened Oct 1, 2026Contributor·6 tasks done·

  ·

• perf(dsv4): optimize CPU offload kernels and document the streaming roof gap

  #1818·ZacharyZcR opened Oct 1, 2026Contributor·

  ·
  1

• feat(dsv4): add CPU offload with a six-to-zero GPU benchmark study

  #1816·ZacharyZcR opened Oct 1, 2026Contributor·

  ·
  2

• perf(dsv4): add device-resident speculative execution and vectorized experts

  #1815·ZacharyZcR opened Oct 1, 2026Contributor·

  ·
  2

• olmoe: tier-caching & PILOT evaluation — demand-LFRU win, cold-disk Markov win, core-engine backport

  #1807·KhalidAlnujaidi opened Sep 29, 2026·3 of 4 tasks·

  ·
  3

• api: tool_choice: "required" instructs the model on every engine

  #1805·Yi-111-a opened Sep 29, 2026·9 of 10 tasks·

  ·
  2

• glm53: keep the KV prefix across image requests by keying the picture

  #1792·Avicennasis opened Sep 28, 2026Contributor·

  ·
  5

• qwen36: the DeltaNet layer runs on the card (Q36_DN_GPU=1); expert homes per layer range (QT_HOME=layer); shexp offered on request

  #1784·kreuzzelg opened Sep 28, 2026Contributor·

  ·
  3

• glm53: expert workers for GLM-5.3 (CLUSTER_WORKERS / EXPERT_WORKER)

  #1774·JackKnifeAI opened Sep 27, 2026·10 tasks done·

  ·
  1

• qwen36 tier: re-plan the VRAM residents from the prompt's own prefill routing (QT_PREFILL_REPLAN=1)

  #1769·kreuzzelg opened Sep 27, 2026Contributor·

  ·

• feat(dsv4): add automatic multi-GPU residency and faster CUDA execution

  #1766·ZacharyZcR opened Sep 27, 2026Contributor·

  ·
  2

• feat(qwen36): derive the default expert-cache cap from host RAM instead of a hardcoded 16

  #1747·jtinbergen opened Sep 25, 2026Contributor·

  ·
  2

• xdna: run the lane on Linux (dlopen loader, libcoli_xdna.so helper)

  #1746·huppiflupp opened Sep 25, 2026Contributor·

  ·

• cuda: decode fmt=9 (bf16) in the dense matmul

  #1620·DebugSultan opened Sep 19, 2026Contributor·

  ·
  2

• perf(olmoe): O(1) LRU victim selection via intrusive recency lists (#1050)

  #1571·Petsku01 opened Sep 16, 2026Contributor·

  ·
  2

• qwen38: the int8 trunk carries one scale per 64 weights (grouped fmt 1 on the GPU), perplexity-neutral

  #1552·kreuzzelg opened Sep 15, 2026Contributor·

  ·
  1

• fix(windows): load CUDA backend before device discovery

  #1542·IcarusAegis opened Sep 15, 2026·

  ·
  5

• feat(glm53): add device-resident HIP backend for MI350P

  #1448·azharlabs opened Sep 12, 2026·

  ·

• feat(gemma4): add standalone Gemma 4 MoE runtime

  #1276·jcolozzi opened Aug 29, 2026·7 of 8 tasks·

  ·
  6