Pull requests

• [Fix] Use HF sanitized chat IDs for GLM-5.3-Flash

  #5069·silencelamb opened Oct 26, 2026Collaborator·4 tasks done·

  ·

• fix(pytorch): fix cached tokens when preempted and reschedule

  #5068·caikun-pjlab opened Oct 26, 2026Collaborator·

  ·

• fix(pytorch): apply repetition penalty to the last generated token

  #5066·ViderStar opened Oct 26, 2026·

  ·

• fix(vl): read processor config as utf-8

  #5058·kodurd opened Oct 23, 2026·

  ·

• feat(pytorch): add opt-in unified compact state caching

  WIP
  #5057·RunningLeon opened Oct 23, 2026Collaborator·3 of 4 tasks·

  ·

• [Fix] Avoid duplicate generation config override kwargs

  #5055·kokyouu opened Oct 22, 2026·3 tasks done·

  ·

• fix(responses): keep one assistant turn in history

  #5053·kodurd opened Oct 22, 2026·

  ·

• refactor(pytorch): unify KDA prefill with optional FlashKDA checkpoints

  #5050·qescccczmr opened Oct 21, 2026Collaborator·4 tasks done·

  ·
  4

• fix(dlinfer): build the MoE expert-id map on the active device

  #5039·li-lizhe opened Oct 18, 2026Contributor·

  ·
  1

• [Fix] Preserve prompt indices in batched completion streams

  #5038·YueshenZ opened Oct 18, 2026·3 tasks done·

  ·

• docs: fix dead Llama 3.1 model-card link and ignore llama.com

  #5031·Evolian-o opened Oct 14, 2026·

  ·

• [ascend] support mooncake store KV offloading on Ascend NPU

  #5030·wanfengcxz opened Oct 14, 2026Collaborator·

  ·

• feat(serve): case-insensitive model-name matching and comma-separated aliases

  #5029·Evolian-o opened Oct 13, 2026·

  ·
  1

• feat: add ReplaySSM for GDN and KDA

  #5028·RunningLeon opened Oct 13, 2026Collaborator·

  ·
  9

• fix: forward request errors from n>1 chat completion streams

  #5025·hpdkhoa opened Oct 8, 2026·

  ·
  1

• [Docs] Fix invalid Python in the VLMEvalKit configuration example

  #5024·JimmyWang0417 opened Oct 8, 2026Contributor·3 tasks done·

  ·
  1

• Count reasoning_content in the restful chat benchmark

  #5020·David-Wu1119 opened Oct 3, 2026Contributor·

  ·
  1

• fix(turbomind): SM75 (pre-SM80) fixes for MTP speculative decoding [stacked on #5006]

  #5019·bltcn opened Oct 3, 2026Contributor·

  ·

• fix(turbomind): load per-channel FP8 compressed-tensors checkpoints (Ornith)

  #5018·bltcn opened Oct 3, 2026Contributor·

  ·

• Time each request separately in the non-stream pipeline benchmarks

  #5017·David-Wu1119 opened Oct 2, 2026Contributor·

  ·
  1

• [Fix] Avoid CUDA sync in DeepSeek-V4 prefill token metadata

  #5016·adenzhou1350 opened Oct 1, 2026Contributor·

  ·

• fix: keep developer messages in DeepSeek-V4 thinking drop

  #5015·longzhenren opened Oct 1, 2026·

  ·
  1

• fix: keep developer messages in DeepSeek-V3.2 thinking drop

  #5014·longzhenren opened Oct 1, 2026·

  ·
  1

• fix: proxy get_node_url raises TypeError when no node matches role

  #5013·Evolian-o opened Oct 1, 2026·

  ·
  2

• fix: raise ValueError instead of assert in DeepSeek DSML completion parsing

  #5012·pujitha24 opened Sep 30, 2026·

  ·
  1

  2