All pull requests

Pull requests

• perf: source speculative metrics from verify counts and drop derived gauges.

  #2615·wai-bao opened Oct 29, 2026Collaborator·

  ·
  1

• refactor: hoist draft weight sharing to the draft-model spec base.

  #2613·wai-bao opened Oct 29, 2026Collaborator·

  ·

• feat: add t5_split_qkv_append_cache npu op

  #2610·wisdom1710 opened Oct 29, 2026·

  ·
  4

• feat(npu): add DualAffineLayerNorm binding

  #2609·wisdom1710 opened Oct 29, 2026·

  ·
  5

• feat: add glm5 next mlu model and kpool foundations.

  #2607·xllm-org opened Oct 29, 2026Collaborator·10 of 17 tasks·

  ·

• docs: add T5 QKV cache operator design.

  #2604·wangzhao-11a opened Oct 28, 2026·8 of 17 tasks·

  ·
  2

• perf: optimize npu sampling and sparse dflash2 verification.

  #2603·AII12754 opened Oct 28, 2026Contributor·12 of 17 tasks·

  ·
  16

• bugfix: prewarm all python npu no-padding decode batches.

  #2599·XuZhang99 opened Oct 28, 2026Collaborator·7 of 12 tasks·

  ·
  3

• docs: add Huawei branding to Ascend NPU entry in README and README_zh

  #2585·yzeyu71 opened Oct 26, 2026·

  ·
  3

• perf: optimize NPU Qwen3.5 GDN prefill projections and packing.

  #2577·AII12754 opened Oct 25, 2026Contributor·8 of 13 tasks·

  ·
  4

• perf: fuse qwen python attention output gating.

  #2576·hswsplendid opened Oct 25, 2026Contributor·9 of 17 tasks·

  ·
  12

• perf: skip redundant filtering for greedy sampling rows.

  #2550·AII12754 opened Oct 22, 2026Contributor·6 of 11 tasks·

  ·
  1

• perf: batch seeded sampling on CUDA devices.

  #2542·AII12754 opened Oct 21, 2026Contributor·11 of 17 tasks·

  ·
  2

• feat: add dcpp support for pytorch glm5.3 models.

  #2533·XuZhang99 opened Oct 21, 2026Collaborator·9 of 17 tasks·

  ·
  2

• refactor: unify engine ownership under Engine and de-templatize schedulers.

  #2532·wai-bao opened Oct 20, 2026Collaborator·

  ·
  2

• feat: add olmo2 cuda inference and validation.

  #2531·hswsplendid opened Oct 20, 2026Contributor·10 of 14 tasks·

  ·
  3

• feat: load OLMo2 configuration without runtime factories.

  #2518·hswsplendid opened Oct 19, 2026Contributor·12 of 17 tasks·

  ·
  11

• Add MHCPreSinkhorn RMSNorm fused PyPTO-pro NPU kernel with docs

  #2501·tracozxc opened Oct 16, 2026·

  ·
  13

• refactor: prepare scoped kernel backends and validation.

  #2492·XuZhang99 opened Oct 15, 2026Collaborator·

  ·
  2

• bugfix: honor minimum tokens and seeds in native mtp decoding.

  #2479·XuZhang99 opened Oct 13, 2026Collaborator·9 of 13 tasks·

  ·
  6

• bugfix: specialize ACL graphs by LoRA route.

  #2466·cchh05 opened Oct 10, 2026Collaborator·

  ·
  7

• bugfix: treat temperature==0 as greedy regardless of top_k/top_p.

  #2465·cchh05 opened Oct 10, 2026Collaborator·

  ·
  2

• feat: add native stateless openai responses api.

  #2432·rhythm-agent-team opened Oct 5, 2026Contributor·12 of 18 tasks·

  ·
  1

• perf: route GLM Lightning Indexer through torch_npu.

  #2355·PixelFlat opened Sep 20, 2026Contributor·

  ·

• refactor: unify qwen image sequence parallel handling.

  #2322·moonseeker1 opened Sep 9, 2026·

  ·