All pull requests

Pull requests

• [kvmeta] add variable-size embedding cache and registered GPU transport

  ai reviewed
  #440·charpty opened Nov 8, 2026Collaborator·

  ·
  6

• [config] verify Valkey control-plane compatibility

  #439·affonsov opened Nov 8, 2026·

  ·
  4

• [py_connector] submit save sessions on steps without a forward pass

  #437·grance199981 opened Nov 7, 2026·

  ·
  5

• [integration_test] stop classifying every qwen3_* model_type as hybrid in is_hybrid_model

  #436·grance199981 opened Nov 7, 2026·

  ·
  4

• [client] preserve buffer safety in transfer errors

  ai reviewed
  #435·GrayXu opened Nov 6, 2026Collaborator·

  ·
  4

• [hisim] fix prefix dataset random access

  #434·bingdongni opened Nov 6, 2026·

  ·
  1

• [service] unify exit signal handling and graceful shutdown

  ai reviewed
  #433·wangxiyu191 opened Nov 5, 2026Collaborator·

  ·
  13

• [metrics] expose node lease and election phase observations

  ai reviewed
  #431·kevinJobs2012 opened Nov 4, 2026Collaborator·

  ·
  7

• [metrics] probe metadata and coordination Redis health independently

  ai reviewed
  #430·kevinJobs2012 opened Nov 4, 2026Collaborator·

  ·
  4

• [service] dispatch Meta requests through a shared worker pool

  ai reviewed
  #428·shaohuaxi opened Nov 4, 2026Collaborator·

  ·
  3

• [metrics] expose node roles and Redis health, and clear stale follower metrics

  ai reviewed
  #423·kevinJobs2012 opened Nov 3, 2026Collaborator·

  ·
  13

• [manager] support storage-specific replica requirements

  ai reviewed
  #422·shaohuaxi opened Nov 2, 2026Collaborator·

  ·
  4

• [py_connector] state the backend's host segments before the client is created

  #416·lpdink opened Oct 28, 2026Collaborator·

  ·
  1

• [py_connector] let the vLLM connector run one pipeline stage per data plane

  #413·lpdink opened Oct 28, 2026Collaborator·

  ·
  1

• [py_connector] add a pluggable transfer backend seam to the vLLM connector

  ai reviewed
  #412·lpdink opened Oct 28, 2026Collaborator·

  ·
  4

• [py_connector] sglang: make the v1 hit count and the v2 commit agree with what was served and written

  #405·lpdink opened Oct 26, 2026Collaborator·11/11·

  ·
  4

• [py_connector] sglang: pin registered specs to the specs the ranks actually use

  #404·lpdink opened Oct 26, 2026Collaborator·10/11·

  ·
  1

• [py_connector] sglang: guard the uri/buffer alignment outside the assert machinery

  #403·lpdink opened Oct 26, 2026Collaborator·9/11·

  ·
  1

• [client] share the host registration classification with the pybind layer

  #400·lpdink opened Oct 22, 2026Collaborator·8/11·

  ·
  1

• [py_connector] sglang: broadcast from the attention-TP leader and join Gloo groups locally

  #399·lpdink opened Oct 22, 2026Collaborator·7/11·

  ·
  1

• [py_connector] sglang: report only the pages a v1 GET actually filled

  #398·lpdink opened Oct 22, 2026Collaborator·6/11·

  ·
  2

• [py_connector] sglang: hand every host pool segment to the client before Create

  #394·lpdink opened Oct 21, 2026Collaborator·5/11·

  ·
  13

• [client] carry every host memory registration through the client channel

  #393·lpdink opened Oct 21, 2026Collaborator·4/11·

  ·
  23

• [py_connector] sglang: freeze the location spec set once registration completes

  #392·lpdink opened Oct 21, 2026Collaborator·3/11·

  ·
  5

• [py_connector] sglang: back any derived side pool with its own location spec

  #391·lpdink opened Oct 21, 2026Collaborator·2/11·

  ·
  13