All pull requests

Pull requests

• refactor(qwen35): adopt the shared step contract

  qwen35
  #1099·RicardoMin opened Sep 29, 2026Contributor·3 tasks done·

  ·
  35

• perf(deepseek-v2-lite): route small-row MoE on device

  dsv2lite
  #1089·Ke-Wng opened Sep 28, 2026Contributor·6 of 8 tasks·

  ·
  1

  14

• perf(qwen35): fuse the linear-attention decode projections into two GEMMs

  #1073·CAICAIIs opened Sep 23, 2026Collaborator·

  ·

• perf(qwen35): split the KV range across CTAs for small-batch decode attention

  #1072·CAICAIIs opened Sep 22, 2026Collaborator·

  ·
  1

• feat(qwen3): add native DFlash2 checkpoint and selector

  dsv2liteenhancementgemma4glm52qwen3qwen35
  #1064·RicardoMin opened Sep 18, 2026Contributor·

  ·
  8

• feat(frontend): expose cumulative prefix cache metrics through vllm

  stale
  #1045·xiaguan opened Sep 11, 2026Collaborator·

  ·
  1

• refactor(qwen35): split scheduler step drivers into scheduler/{steps,emit,telemetry}.rs

  needs-rebase
  #1037·CAICAIIs opened Sep 7, 2026Collaborator·

  ·
  5

• refactor(qwen35): split TP response contract into tp_executor/responses.rs

  needs-rebase
  #1036·CAICAIIs opened Sep 7, 2026Collaborator·

  ·
  3

• perf(k3): five cuts on the CP8 128k prefill superstep

  stale
  #1026·xiaguan opened Sep 1, 2026Collaborator·

  ·
  2

• test(kernels): reduce the duplicated CPU operator oracles in the HD256/HD512 QK-RoPE smokes

  needs-rebase
  #1002·BreezyB1n opened Aug 30, 2026Contributor·5 of 8 tasks·

  ·
  1

  4

• feat(k3): kernel mining + capsule substrate — vendored vLLM KDA/top-k cubins (+7%/+20% decode)

  stale
  #982·xiaguan opened Aug 28, 2026Collaborator·

  ·
  4

• feat(qwen35): launch only on the step contract

  needs-rebase
  #981·xiaguan opened Aug 28, 2026Collaborator·4 tasks·

  ·
  5

• devops: make serving benchmarks reproducible

  stale
  #979·xiaguan opened Aug 26, 2026Collaborator·

  ·
  3

• feat(observability): report real prefix-cache query/hit counters in /metrics

  #953·1571859588 opened Aug 23, 2026·

  ·
  24

• feat(higgs-audio): add checkpoint loading and validation foundation

  needs-rebase
  #864·ywh555hhh opened Aug 13, 2026Contributor·

  ·
  9

  

• feat(qwen35): add opt-in SM120 FlashInfer GDN prefill

  needs-rebase
  #862·qwzx-qwas opened Aug 10, 2026·7 of 9 tasks·

  ·
  1

  35

  

• feat(eagle3): Qwen3 EAGLE-3 drafter forward pass (single-step + chain rollout)

  stale
  #707·scatyf3 opened Jul 18, 2026Contributor·

  ·
  7

  

• fix(qwen3): proactive cancellation sweep for queued requests

  needs-rebase
  #646·sparkzky opened Jul 10, 2026Contributor·7 tasks done·

  ·
  1

  14