All pull requests

Pull requests

• perf: avoid redundant copies when batching a single multimodal tensor.

  #2745·liutongxuan opened Nov 15, 2026Collaborator·17 tasks·

  ·
  1

• refactor: unify causal VLMs into the causal-lm framework.

  #2744·wai-bao opened Nov 14, 2026Collaborator·

  ·
  1

• docker: add ascend runtime and development dockerfile.

  #2742·XuZhang99 opened Nov 14, 2026Collaborator·7 of 17 tasks·

  ·

• refactor: inject scheduler and EPLB options across runtime boundaries.

  #2738·liutongxuan opened Nov 13, 2026Collaborator·17 tasks·

  ·
  1

• perf: capture the qwen3-vl vision encoder in an acl graph.

  #2737·9wutong opened Nov 13, 2026·11 of 16 tasks·

  ·
  28

• feat: optimize MiniMax-H3 Ref2VA on 16 Ascend A3 NPUs

  #2736·liuwenda4 opened Nov 13, 2026·

  ·
  2

• feat: add DeepSeek V4.1 Flash model support with DP safe decode graph.

  #2734·xllm-org opened Nov 13, 2026Collaborator·

  ·
  1

• feat: expand python vlm execution capabilities.

  #2732·XuZhang99 opened Nov 13, 2026Collaborator·

  ·
  7

• bugfix: support typed KV layouts and MLA graph execution.

  #2731·xllm-org opened Nov 12, 2026Collaborator·

  ·
  2

• perf: Prepare Qwen3-VL vision metadata on CPU.

  #2720·xagr opened Nov 11, 2026·8 of 11 tasks·

  ·
  3

• refactor: integrate npu normalization with kernel runtime.

  #2711·XuZhang99 opened Nov 11, 2026Collaborator·4 of 6 tasks·

  ·
  4

• perf: derive the decode graph batch ceiling from the speculative budget.

  #2709·9wutong opened Nov 11, 2026·10 of 17 tasks·

  ·
  11

• bugfix: preflight MiniMax FP8 checkpoint indexes.

  #2667·hswsplendid opened Nov 4, 2026Contributor·8 of 17 tasks·

  ·
  4

• bugfix: validate MiniMax FP8 dequantization inputs.

  #2665·hswsplendid opened Nov 4, 2026Contributor·8 of 17 tasks·

  ·
  3

• feat: Add standalone T5 QKV append operator.

  #2656·wangzhao-11a opened Nov 3, 2026·10 of 14 tasks·

  ·

• feat: add DualAffineLayerNorm operator.

  #2655·fbossing-123 opened Nov 2, 2026·

  ·
  9

• docs: add DualAffineLayerNorm operator design review

  #2654·fbossing-123 opened Nov 2, 2026·

  ·
  24

• bugfix: reconcile vmm cache usage after batch retirement.

  #2652·AII12754 opened Nov 2, 2026Contributor·2 tasks done·

  ·
  3

• feat: Add T5 QKV cache append wrapper.

  #2641·brandfucker opened Nov 1, 2026·6 of 8 tasks·

  ·

• feat: enable adaptive qwen mtp verification on npu.

  #2632·hswsplendid opened Oct 31, 2026Contributor·8 of 17 tasks·

  ·
  2

• feat: manage sliding window caches for block drafts.

  #2627·PixelFlat opened Oct 31, 2026Contributor·9 of 10 tasks·

  ·
  9

• feat: support measured graph costs for adaptive verification.

  #2623·hswsplendid opened Oct 30, 2026Contributor·11 of 17 tasks·

  ·
  1

• refactor: centralize dp rank placement in scheduler.

  #2621·PixelFlat opened Oct 30, 2026Contributor·7 of 12 tasks·

  ·
  8

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

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

  ·
  1

• docs: add T5 QKV cache operator design.

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

  ·
  3