All pull requests

Pull requests

• fix: align reported KV token capacity with full cache groups

  #1465·Adrenaline-S opened Sep 27, 2026Collaborator·

  ·
  4

• feat: remove flashinfer cpp, use native flashinfer-python

  #1464·moui0 opened Sep 26, 2026Collaborator·

  ·
  33

• refactor: share engine lifecycle for sleep/wake and bounded graceful shutdown

  #1463·MMadhatter opened Sep 25, 2026Collaborator·

  ·
  30

• feat: add context token delta metrics

  #1462·jhuangxue opened Sep 24, 2026·

  ·
  50

• fix: enforce FP8 KV startup metadata contracts

  #1461·Zn0425 opened Sep 24, 2026·

  ·
  6

• feat: add chunked prefill for standard, MLA, and hybrid attention

  #1460·Jonathan-hwx opened Sep 23, 2026Collaborator·

  ·
  27

• perf: optimize beam search output and KV cache

  #1459·siluzhou opened Sep 23, 2026Collaborator·

  ·
  14

• feat(multimodal): add cache-aware ViT routing and shared serving framework

  #1458·ySingularity opened Sep 23, 2026Collaborator·

  ·
  70

• feat(flexlb): add KVCM and multi-engine runtime support

  #1457·Martin7-1 opened Sep 23, 2026Collaborator·

  ·
  79

• fix: avoid ATREX buffer limit for large ROCm KV caches

  #1456·zhangjianning-zjn opened Sep 23, 2026Collaborator·

  ·
  3

• feat(rocm): single-launch verify attention, DFlash drafting, deterministic FP8 MoE and TorchSpec DSpark

  #1453·yzyDavid opened Sep 22, 2026Collaborator·

  ·
  60

• feat(dsv4): integrate standalone decode MoE front

  #1452·zsnoob opened Sep 21, 2026·

  ·
  30

• fix(openai): release the parked think tail when generation stops inside a think block

  #1451·Linxiushen opened Sep 21, 2026·

  ·
  18

• feat: add Pro5000 quantization A/B switches

  #1449·hy-guo opened Sep 21, 2026·

  ·
  10

• feat: add Master HTTP batch fanout and batch placement

  #1448·CHW0218 opened Sep 21, 2026Collaborator·

  ·
  108

• feat(p2p): integrate planned decode handoff

  #1443·ZhihanYan opened Sep 18, 2026Collaborator·

  ·
  98

• feat: migrate RTP-LLM to Python Native packaging and pytest execution

  #1441·baohengyi opened Sep 17, 2026·

  ·
  73

• fix(trace): prevent false FetchResponse cancellation after normal completion

  #1438·SAzwj opened Sep 17, 2026Collaborator·

  ·
  6

• feat(multimodal): add standalone KVCM EMB Python client

  #1436·charpty opened Sep 16, 2026Collaborator·

  ·
  75

• feat: support DeepSeek V4 Flash PPU PD and MTP

  #1434·fxp19950422 opened Sep 16, 2026·

  ·
  230

• feat: Minimax-M3 model integration on main

  #1433·qqbbiu opened Sep 16, 2026Collaborator·

  ·
  269

• fix(think): align think handling with the template anchor and the seq budget

  #1424·hy-guo opened Sep 12, 2026·

  ·
  143

• feat(rocm): upgrade AITER and migrate kernels to Triton 3.8

  #1421·junna2016 opened Sep 11, 2026Collaborator·

  ·
  21

• fix(rocm): moriep support hipgraph

  #1420·jacobwin-ai opened Sep 11, 2026Collaborator·

  ·
  57

• fix: handle ROCm all-reduce barrier epoch wraparound

  #1418·yzyDavid opened Sep 10, 2026Collaborator·

  ·
  5