All pull requests

Pull requests

• perf(dsv41): optimize DeepSelect and Engram FP8 lookup

  auto-lgtm:authorizedreview:running
  #1582·linzehaina opened Oct 26, 2026Contributor·

  ·
  1

• fix: prepare MTP verification positions before async attention setup

  #1581·hy-guo opened Oct 26, 2026·

  ·

• fix: keep FlashInfer prefill CUDA graph KV writes in bounds

  #1579·hy-guo opened Oct 26, 2026·

  ·

• fix: avoid publishing uncomputed recurrent prefix snapshots

  #1578·hy-guo opened Oct 26, 2026·

  ·

• implement k3 pr

  auto-lgtm:needs-approvalreview:running
  #1577·Jonathan-hwx opened Oct 26, 2026Collaborator·

  ·
  3

• fix: load Qwen tokenizer JSON with legacy Transformers

  auto-lgtm:needs-approvalreview:running
  #1575·hy-guo opened Oct 26, 2026·

  ·
  1

• feat(cache): add reusable tail retention and proportional tier pools

  auto-lgtm:needs-approvalreview:running
  #1574·xiqi-lq opened Oct 26, 2026Collaborator·

  ·
  1

• fix: correct serving completion, client errors and resource cleanup

  #1573·hy-guo opened Oct 26, 2026·

  ·

• perf: use execAllGather for TP embedding and logits

  #1572·hy-guo opened Oct 26, 2026·

  ·

• perf: copy active FlashInfer MLA metadata only

  #1571·hy-guo opened Oct 26, 2026·

  ·

• fix: include MTP lookahead in prefix cache keys

  #1570·hy-guo opened Oct 26, 2026·

  ·

• feat: unify OpenAI generation config and model routing

  #1569·hy-guo opened Oct 26, 2026·

  ·

• feat: opt-in XGrammar whitespace handling

  #1568·hy-guo opened Oct 26, 2026·

  ·

• fix: align CUDA sampling penalty semantics

  #1567·hy-guo opened Oct 26, 2026·

  ·

• fix: round fused GDN recurrent state consistently

  #1566·hy-guo opened Oct 26, 2026·

  ·

• fix: preserve stream cache state across RPC teardown and P/D transfer

  #1565·hy-guo opened Oct 26, 2026·

  ·

• P2p decode writeback

  auto-lgtm:needs-approvalreview:delayed
  #1564·renyuanming opened Oct 26, 2026Collaborator·

  ·
  1

• feat: precision-gated dynamic decode attention

  auto-lgtm:needs-approvalreview:delayed
  #1563·SAzwj opened Oct 26, 2026Collaborator·

  ·
  1

• [WIP] Reuse partial prefill KV cache tails and exact common prefixes

  #1562·Adrenaline-S opened Oct 26, 2026Collaborator·

  ·

• feat: isolate DeepSeek V4 Flash PPU PD and MTP support

  #1561·fxp19950422 opened Oct 26, 2026Contributor·

  ·

• feat: restore async prefill KV preparation and allocator warmup

  auto-lgtm:needs-approvalreview:delayed
  #1560·wanglining97 opened Oct 26, 2026Collaborator·

  ·
  2

• feat: restore GLM startup and request compatibility

  auto-lgtm:needs-approvalreview:delayed
  #1559·wanglining97 opened Oct 26, 2026Collaborator·

  ·
  2

• feat: add compact logprobs and request accounting to serving

  auto-lgtm:needs-approvalreview:delayed
  #1558·wanglining97 opened Oct 26, 2026Collaborator·

  ·
  2

• feat: separate prefill query and KV token budgets

  auto-lgtm:needs-approvalreview:delayed
  #1557·wanglining97 opened Oct 26, 2026Collaborator·

  ·
  2

• feat: monitor xgrammar update latency and disable grammar on timeout

  auto-lgtm:needs-approvalreview:running
  #1556·alibaba-miji opened Oct 25, 2026Collaborator·

  ·
  2