All pull requests

Pull requests

• Support hybrid TP16/DCP8 Kimi execution and optimize decode

  #277·voipmonitor opened Nov 12, 2026Collaborator·

  ·

• Prepare BRRR-derived small-verification fusion for DeepSeek V4.1 TP4

  #259·shihanqu opened Nov 10, 2026·

  ·

• Prepare DeepSeek V4 Flash across two or four DGX Sparks

  #240·kyonjova opened Nov 8, 2026Contributor·

  ·

• Prepare and serve MiMo V2.6 Flash RL on two DGX Sparks

  #239·kyonjova opened Nov 8, 2026Contributor·

  ·

• perf(qwen): prepare packed Flash Next DFlash2 verification

  area:kernelsarea:preparationenhancementmodel:qwen3.8-flash-nextperformancetriage:validation
  #233·shadownetX opened Nov 8, 2026·

  ·
  3

• Route multi-host RoCE rails over direct cables

  area:orbitarea:rocenanteenhancementperformancetriage:design-review
  #232·original-el8 opened Nov 8, 2026Collaborator·

  ·
  12

• Race a Triton tile kernel against cuBLAS for GLM's BF16 head projections

  area:kernelsarea:preparationenhancementmodel:glm-5.3performancetriage:design-review
  #231·original-el8 opened Nov 8, 2026Collaborator·

  ·
  1

• Muse Glimmer 30B NVFP4: single-GPU serving with images, DFlash2 drafting, and a 1,048,576-token context

  area:apiarea:packagingarea:preparationenhancementmodel:muse-glimmertriage:design-review
  #229·Raydius opened Nov 7, 2026·

  ·
  3

• fix(cache): preserve message checkpoints for divergent continuations

  area:cachingarea:schedulingarea:selenebugmodel:qwen3.8-flash-nextperformancepriority:P2triage:review
  #227·shadownetX opened Nov 7, 2026·

  ·
  1

  1

• Fix typo and comma placement in README vLLM comparison

  documentationtriage:review
  #226·justink989 opened Nov 7, 2026·

  ·

• Report conversion phases while ignition convert verifies and writes a package

  area:packagingenhancementtriage:review
  #223·kyonjova opened Nov 6, 2026Contributor·

  ·

• Widen the RoCE collective wait while compiled forwards warm across hosts

  area:preparationarea:rocenantebugmodel:mimo-v2.6priority:P2triage:review
  #222·kyonjova opened Nov 6, 2026Contributor·

  ·
  1

• Recognize GB300 DGX Station GPU device ID 0x31c2

  area:packagingbugpriority:P2triage:review
  #221·YouNeedCryDear opened Nov 6, 2026·

  ·
  1

• DeepSeek V4.1 Flash: TP6 preparation with per-rank Engram partitions

  area:engramarea:memoryarea:packagingarea:preparationenhancementmodel:deepseek-v4.1performancetriage:design-review
  #220·nero- opened Nov 6, 2026·

  ·
  1

• Helios: load partitioned Engram tables on rank-selective package copies

  area:engramarea:packagingbugmodel:deepseek-v4.1priority:P1triage:review
  #219·nero- opened Nov 6, 2026·

  ·
  1

• Share the first GLM draft pass's sparse selection with later passes

  area:kernelsarea:preparationbugmodel:glm-5.3performancepriority:P2triage:review
  #218·original-el8 opened Nov 6, 2026Collaborator·

  ·
  1

• Decode GLM width-one rounds in place without KDA records

  area:kernelsarea:preparationmodel:glm-5.3performancetriage:review
  #217·original-el8 opened Nov 6, 2026Collaborator·

  ·
  1

• Track FlashInfer main in the preparation dependency instead of a pinned commit

  area:preparationenhancementtriage:design-review
  #215·brandonmmusic-max opened Nov 6, 2026Contributor·

  ·
  5

• Read every checkpoint through b12x's pinned-host reader

  area:memoryarea:preparationbugpriority:P2triage:design-review
  #213·original-el8 opened Nov 5, 2026Collaborator·

  ·
  2

• Qwen3.8 Flash Next: declared YaRN rope scaling and --rope-parameters for 1M-token serving

  model:qwen3.8-flash-nexttriage:design-review
  #212·Raydius opened Nov 5, 2026·

  ·
  4

• Export mixed rounds with speculation again

  area:preparationbugmodel:qwen3.8-flash-nextpriority:P1triage:review
  #211·Raydius opened Nov 5, 2026·

  ·

• Report checkpoint source failures to preparation peers

  bugmodel:deepseek-v4.1priority:P2triage:review
  #198·0z5a opened Nov 4, 2026·

  ·

• Report the selected deployment rebuild manifest in inspection

  area:packagingbugpriority:P2triage:review
  #196·0z5a opened Nov 4, 2026·

  ·
  1

  3

• Add Muse Glimmer mixed-precision text recipe

  area:preparationmodel:muse-glimmertriage:validation
  #192·YouNeedCryDear opened Nov 4, 2026·

  ·
  3

• Add Inkling-Small NVFP4 recipe for one DGX Station

  area:preparationmodel:inkling-smalltriage:design-review
  #191·YouNeedCryDear opened Nov 4, 2026·

  ·
  2