All pull requests

Pull requests

• fix(vllm): bind each data-parallel rank's KV events on its advertised port

  #196·llying-001 opened Oct 12, 2026Contributor·5 of 12 tasks·

  ·
  1

• Restyle the Pareto dashboard and draw measured runs over it

  #195·araina-amd opened Oct 11, 2026Collaborator·

  ·

• Add tiered KV placement below HBM and a Pareto sweep

  #194·araina-amd opened Oct 11, 2026Collaborator·

  ·

• Report the GEMM shapes a DES replay runs

  #193·araina-amd opened Oct 10, 2026Collaborator·6 of 7 tasks·

  ·

• fix: harden engine startup and Mooncake preflight validation

  #192·llying-001 opened Oct 10, 2026Contributor·8 tasks done·

  ·
  5

• Model the speculative drafter in simulate mode

  #191·araina-amd opened Oct 10, 2026Collaborator·

  ·

• Price hybrid-model layers from their own operations

  #190·araina-amd opened Oct 9, 2026Collaborator·

  ·

• Read the decode staircase only where the engine pads

  #189·araina-amd opened Oct 9, 2026Collaborator·

  ·

• feat(router): add opt-in Decode input-demand routing (R2)

  #188·llying-001 opened Oct 9, 2026Contributor·5 of 7 tasks·

  ·

• feat(build): carry Mooncake RDMA rail and ACK-timeout patches for SGLang

  #187·llying-001 opened Oct 9, 2026Contributor·6 of 7 tasks·

  ·

• feat(router): add opt-in session affinity with independent PD bindings

  #186·llying-001 opened Oct 8, 2026Contributor·

  ·
  5

• feat: add opt-out usage reporting for Python serving

  #182·jiejingzhangamd opened Oct 8, 2026Collaborator·4 of 6 tasks·

  ·

• Price pipeline-parallel hops, the token return, and cross-node stage boundaries

  #179·araina-amd opened Oct 1, 2026Collaborator·6 of 8 tasks·

  ·

• Attention-DP: apply the request split once in the profiler

  #177·ajassani opened Sep 23, 2026·1 of 2 tasks·

  ·

• EP AllToAll: size dispatch on the sequences a rank owns under attention-DP

  #175·ajassani opened Sep 22, 2026·2 tasks done·

  ·

• Refuse an out-of-regime anchor, and price sparse attention as selection plus indexing

  #174·araina-amd opened Sep 21, 2026Collaborator·2 tasks done·

  ·

• feat(agent_sys): an agent work system, and the five-stage LLM e2e optimisation flow

  #167·dorado269 opened Sep 14, 2026Collaborator·

  ·
  2

• [Compare only][Draft]fast bench script

  #164·dorado269 opened Sep 10, 2026Collaborator·12 tasks·

  ·

• examples(glm5.2 1P1D): accuracy evaluation, Mooncake trace replay, per-role torch profiling

  #140·ChengYao-amd opened Aug 27, 2026·

  ·

• fix(atom): relay KV events from data-parallel ranks

  #121·cquil11 opened Aug 17, 2026Contributor·

  ·
  2

• test(bench): replay the Mooncake FAST'25 trace the reporter pointed at

  #111·jiejingzhangamd opened Aug 9, 2026Collaborator·

  ·

• docs(bench): how to verify the short-prompt routing fix on a k8s fleet

  #110·jiejingzhangamd opened Aug 9, 2026Collaborator·

  ·

• fix(router): say so when a missing tokenizer disables cache-aware routing

  #108·jiejingzhangamd opened Aug 8, 2026Collaborator·

  ·