All pull requests

Pull requests

• feat(recording): record batch schedules and embedded request timing v…

  auto-lgtm:needs-approvalreview:running
  #1638·HaiyuZhou opened Nov 4, 2026Collaborator·

  ·
  1

• fix(server): make cooperative process reclamation bounded and leak-free

  auto-lgtm:needs-approvalreview:delayed
  #1637·LLLLKKKK opened Nov 4, 2026Collaborator·

  ·
  3

• feat(sm120): support non-PP DeepSeek V4 inference

  auto-lgtm:needs-approvalreview:running
  #1636·Chiiisen opened Nov 3, 2026Collaborator·

  ·
  2

• (feat): support sparse kvoffload for dsa-like kvoffload

  auto-lgtm:needs-approvalreview:blocked
  #1634·renyuanming opened Nov 3, 2026Collaborator·

  ·
  20

• fix: validate decode graph buckets and propagate P/D cancellation

  auto-lgtm:needs-approvalreview:complete
  #1633·john-kaiii opened Nov 3, 2026Collaborator·

  ·
  6

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

  auto-lgtm:needs-approvalreview:delayed
  #1632·Adrenaline-S opened Nov 3, 2026Collaborator·

  ·
  2

• feat(flexlb): add worker-oriented scheduling and NON_BATCH_V2 queues

  auto-lgtm:needs-approvalreview:blocked
  #1629·sunmiaozju opened Nov 2, 2026Collaborator·

  ·
  452

• build(xpu): add --config=xpu and an XPU Python requirements lock

  auto-lgtm:needs-approvalreview:blocked
  #1628·SKRohit opened Nov 2, 2026·

  ·
  21

• feat(xpu): recognize Intel XPU devices

  auto-lgtm:needs-approvalreview:blocked
  #1627·SKRohit opened Nov 2, 2026·

  ·
  9

• fix: platform-neutral fixes for non-CUDA/ROCm (generic) build paths

  auto-lgtm:needs-approvalreview:complete
  #1626·SKRohit opened Nov 2, 2026·

  ·
  3

• feat: optimize beam search and fuse FP8 activation

  auto-lgtm:needs-approvalreview:running
  #1624·siluzhou opened Nov 2, 2026Collaborator·

  ·
  149

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

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

  ·
  236

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

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

  ·
  10

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

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

  ·
  148

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

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

  ·
  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·

  ·
  90

• feat: capture and replay GLM indexer tensors

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

  ·
  7

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

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

  ·
  215

• feat: support compact logprobs transport and rendering

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

  ·
  24

• 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, 2026Collaborator·

  ·
  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, 2026Collaborator·

  ·