All pull requests

Pull requests

• perf(qwen4): tune six MoE/PLE/QSA kernels bit-identically

  qwen4
  #1264·sysuTobo opened Nov 15, 2026Contributor·4 of 5 tasks·

  ·

• perf(qwen35): cap packed prefill steps at 16 live decoders

  qwen35
  #1263·CAICAIIs opened Nov 14, 2026Collaborator·

  ·

• feat(rocm): add ROCm 10.1 Qwen3 BF16 eager backend

  dsv2liteenhancementgemma4glm52kimi-k2needs-rebaseqwen3qwen4qwen35
  #1255·zihaomu opened Nov 12, 2026·5 of 8 tasks·

  ·
  13

• perf(qwen4): group the MoE expert GEMMs and speed up the route fold

  qwen3qwen4
  #1252·sysuTobo opened Nov 12, 2026Contributor·

  ·
  9

• perf(qwen35): overlap the requests of one prefill chunk

  qwen35
  #1237·CAICAIIs opened Nov 7, 2026Collaborator·

  ·
  18

• perf(kimi-k2): split decode MLA attention over KV tiles

  kimi-k2stale
  #1195·CAICAIIs opened Oct 29, 2026Collaborator·

  ·
  7

• feat(qwen3): run DSpine drafters on the DFlash speculative path

  enhancementneeds-rebaseqwen3
  #1191·sysuTobo opened Oct 27, 2026Contributor·

  ·
  9

• fix(eval): bind rerun recovery to current inputs

  bugstale
  #1183·DivyamTalwar opened Oct 23, 2026·

  ·
  2

• fix(bench): compare retained serving configurations

  bugdsv2litestale
  #1182·DivyamTalwar opened Oct 23, 2026·

  ·
  2

• fix(bench): rebuild reports against their retained workload

  bugdsv2litestale
  #1181·DivyamTalwar opened Oct 23, 2026·

  ·
  2

• fix(glm52): prevent partial router streams passing the smoke gate

  bugglm52stale
  #1180·DivyamTalwar opened Oct 23, 2026·

  ·
  2

• fix(profiling): protect SQLite trace inputs

  bugstale
  #1179·DivyamTalwar opened Oct 23, 2026·5 of 6 tasks·

  ·
  2

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

  needs-rebasestale
  #1045·xiaguan opened Sep 11, 2026Collaborator·

  ·
  3

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

  needs-rebasestale
  #1026·xiaguan opened Sep 1, 2026Collaborator·

  ·
  4

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

  needs-rebasestale
  #982·xiaguan opened Aug 28, 2026Collaborator·

  ·
  6