All pull requests

Pull requests

• [DO NOT MERGE] Test CI approval orchestration

  #2375·leo-automation opened Sep 23, 2026·

  ·
  6

• ci: avoid restarting tests on PR approval

  #2374·leo-automation opened Sep 23, 2026·

  ·
  1

  1

• ci: check container model mount in atom test

  ci:atom
  #2372·gyohuangxin opened Sep 23, 2026·

  ·
  3

• [WIP][PD][DCP] Reorder mla cache in P to the dcp layout in D

  #2371·MengqingCao opened Sep 23, 2026·

  ·
  1

• ci(atomesh): add GLM-5.2 1P2D CPP4+DPA4 agentic benchmark suites

  #2370·Jasen2201 opened Sep 23, 2026·

  ·
  4

• fix(plugin): hold the deferred-save clock on the slotted SeqView

  #2369·PerryZhang01 opened Sep 23, 2026·

  ·
  6

• fix(mtp): install the draft work plan when warming draft graphs

  #2367·Jasen2201 opened Sep 23, 2026·

  ·
  5

• [feature] support pp>1 for Kimi-K3

  #2365·gbyu-amd opened Sep 23, 2026·

  ·
  3

• feat: add Ulysses sequence parallelism for native GQA prefill

  #2364·whx-sjtu opened Sep 23, 2026·

  ·
  3

• [Bugfix][vLLM] Only engage the EAGLE3 heterogeneous KV pool for real MLA+MHA draft pairs

  bug
  #2362·npoulad1 opened Sep 22, 2026·

  ·
  3

  

• [Bugfix][vLLM] Fill padded GDN decode state slots with NULL_BLOCK_ID, not PAD_SLOT_ID

  bug
  #2361·npoulad1 opened Sep 22, 2026·

  ·
  3

  

• [Bugfix][vLLM] Forward unknown kwargs through the Harmony parser wrapper

  bug
  #2359·npoulad1 opened Sep 22, 2026·

  ·
  2

  

• chore: add security scanning workflows

  #2358·haribabug opened Sep 22, 2026·

  ·
  2

  

• feat(plugin): run DeepSeek-V4.1-Flash text path on the vLLM plugin

  #2357·PerryZhang01 opened Sep 22, 2026·

  ·
  1

  

• CI: allow overriding ATOM inference watchdog threshold

  #2355·gyohuangxin opened Sep 22, 2026·

  ·
  7

  

• feat(offload): add MP-only save admission budget

  #2353·yhl-amd opened Sep 22, 2026·

  ·
  1

  

• feat(mesh): add Envoy ext-proc inference ingress

  #2350·Yuechguo opened Sep 22, 2026·

  ·
  3

  

• Add MiMo V2.5 Pro support to the SGLang plugin

  #2349·yucshen opened Sep 22, 2026·

  ·
  1

  

• fp4 indexer: build the DCP prefill staging indices once per forward, in one launch

  ci:atom
  #2348·ZhiweiYan-96 opened Sep 22, 2026·

  ·
  45

• feat(moe): prefill-only mega combine wire

  ci:atom
  #2347·yanboshao opened Sep 22, 2026·

  ·
  2

  

• ci: add cache-aware V4 P/D DPA benchmarks without offload

  #2346·ZhangLirong-amd opened Sep 22, 2026·

  ·
  1

• (agentic) update glm5.2 recipe for agentX

  #2345·zhuyuhua-v opened Sep 22, 2026·

  ·
  1

• (doc)GLM-5.3 LMCache: concurrency sweep (8/16/32) and HBM/tier/ceiling split

  #2344·PerryZhang01 opened Sep 22, 2026·

  ·
  1

  

• Enable GLM-5.3-Flash under vllm serve on the ATOM plugin.- #2299

  #2340·sajandhy opened Sep 22, 2026·

  ·
  11

• fix(offload): fence dense LMCache saves

  #2339·NidhoggD1 opened Sep 21, 2026·

  ·
  10