All pull requests

Pull requests

• refactor: remove 61 unused env vars (production-audited, keep default behavior)

  auto-lgtm:needs-approvalreview:delayed
  #1622·LLLLKKKK opened Oct 31, 2026Collaborator·

  ·
  22

• feat(ascend): enable MTP speculative decoding on NPU

  auto-lgtm:needs-approvalreview:blocked
  #1621·sai841 opened Oct 31, 2026·

  ·
  6

• feat(ascend): add decode ACL graph (aclgraph) capture & replay

  auto-lgtm:needs-approvalreview:blocked
  #1620·QQzhouQ opened Oct 31, 2026·

  ·
  16

• perf: fuse SM120 Qwen mixed-quant operators and TP2 reductions

  auto-lgtm:needs-approvalreview:blocked
  #1618·hy-guo opened Oct 30, 2026·

  ·
  198

• fix: price dense slice at real width for expert-only quantization + V4 FP4 native detection

  auto-lgtm:needs-approvalreview:complete
  #1617·LLLLKKKK opened Oct 30, 2026Collaborator·6 tasks done·

  ·
  5

• feat: add GLM hidden-state capture for TorchSpec

  auto-lgtm:needs-approvalreview:blocked
  #1616·devillove084 opened Oct 30, 2026Collaborator·

  ·
  32

• feat: capture and replay GLM indexer tensors

  auto-lgtm:needs-approvalreview:complete
  #1615·devillove084 opened Oct 30, 2026Collaborator·

  ·
  6

• refactor(cache): centralize CP cache geometry and transfer assembly

  auto-lgtm:needs-approvalreview:delayed
  #1614·Adrenaline-S opened Oct 30, 2026Collaborator·

  ·
  33

• feat(minimax-m31): support MiniMax M3.1 text and VL models

  auto-lgtm:needs-approvalreview:blocked
  #1613·RuixuanZhang06 opened Oct 29, 2026Collaborator·

  ·
  129

• feat(ascend): add Qwen3.5 eager inference support

  auto-lgtm:needs-approvalreview:complete
  #1612·ningweikang opened Oct 29, 2026·

  ·
  55

• feat: support compact logprobs transport and rendering

  auto-lgtm:needs-approvalreview:blocked
  #1610·wanglining97 opened Oct 29, 2026Collaborator·

  ·
  19

• feat: add standalone target logprobs computation

  auto-lgtm:needs-approvalreview:complete
  #1609·wanglining97 opened Oct 29, 2026Collaborator·

  ·
  4

• Feature/flexlb ref kvcm

  auto-lgtm:needs-approvalreview:blocked
  #1608·jianglan89 opened Oct 29, 2026Collaborator·

  ·
  160

• fix(scheduler): wake on async cache load completion

  auto-lgtm:needs-approvalreview:complete
  #1606·HaiyuZhou opened Oct 28, 2026·

  ·
  5

• feat: add RDMA device health checks and kmonitor shutdown hooks

  auto-lgtm:needs-approvalreview:complete
  #1604·ennalalala opened Oct 28, 2026Collaborator·

  ·
  5

• perf: opt-in Qwen prefill tiling and draft graph buckets

  #1600·hy-guo opened Oct 27, 2026·

  ·

• perf: opt-in FlashInfer b12x for small-batch NVFP4

  auto-lgtm:needs-approvalreview:complete
  #1599·hy-guo opened Oct 27, 2026·

  ·
  12

• feat: independent physical KV-pool admission and PDFusion arbitration

  auto-lgtm:needs-approvalreview:complete
  #1598·hy-guo opened Oct 27, 2026·

  ·
  5

• P2p decode writeback

  auto-lgtm:needs-approvalreview:blocked
  #1596·renyuanming opened Oct 27, 2026Collaborator·

  ·
  112

• fix: freeze normal engine input placement at startup

  #1595·hy-guo opened Oct 27, 2026·

  ·

• fix: preserve fixed output budgets in chat responses

  #1594·hy-guo opened Oct 27, 2026·

  ·

• perf(dsv41): reuse physical decode indices across layers

  #1591·guluguluhhhh opened Oct 27, 2026Collaborator·

  ·
  1

• fix: repair CUDA13 loader and smoke baselines

  auto-lgtm:needs-approvalreview:blocked
  #1589·hy-guo opened Oct 26, 2026·

  ·
  7

• feat: integrate PPU Flash adaptation and isolated MegaMoE wheel

  auto-lgtm:needs-approvalreview:delayed
  #1587·fxp19950422 opened Oct 26, 2026Contributor·

  ·
  492

• feat(cache): report hard stops and RPC failures, split connector refs

  #1584·happyhyaa opened Oct 26, 2026Collaborator·

  ·
  1