All pull requests

Pull requests

• [Train] Fix Qwen3.5 text-only batch unpacking

  #1317·TianHe007 opened Oct 27, 2026·

  ·

• [CICD] Unify test policy, affected routing and execution evidence gates

  #1315·jfhx opened Oct 23, 2026Contributor·

  ·

• [Training][NPU] Support profiler lifecycle via plugin overrides

  #1314·cuizihua opened Oct 23, 2026·

  ·
  2

• [CICD] Unify runtime recovery across platforms and add opt-in infrastructure retry

  #1312·jfhx opened Oct 22, 2026Contributor·

  ·
  4

• feat: add HorizonKV support to FlagScale

  #1311·lemonadenn opened Oct 20, 2026·

  ·
  2

• Add basic multimodal sequence packing (qwen3.5)

  #1308·Baibaione opened Oct 14, 2026Contributor·

  ·

• feat(train): Add opt-in MegaLens runtime integration

  #1306·Chlience opened Oct 9, 2026·

  ·

• fix(train): Skip packed metadata broadcasts for dense batches

  #1305·Chlience opened Oct 9, 2026·

  ·

• fix(train): derive FlagCX backend config from registered capabilities

  #1302·MC952-arch opened Sep 24, 2026Contributor·

  ·
  1

• add Hygon DCU hardware health sampling

  #1299·lzy-edu opened Sep 18, 2026Contributor·

  ·

• [Train] Add NPU profiler support for Megatron training- #1294

  #1297·cuizihua opened Sep 17, 2026·

  ·
  4

• [Train] Add Python 3.10/3.11 compatibility for hybrid override

  #1293·lihongyang1990 opened Sep 13, 2026Contributor·

  ·

• Add multimodule for multimodal model

  #1292·wenqingqian opened Sep 10, 2026Contributor·

  ·

• fix: skip CUDA_DEVICE_MAX_CONNECTIONS check on non-CUDA platforms

  #1288·kolinweiwei opened Sep 8, 2026·

  ·
  1

  1

• Add MemRift weight and activation compression for training

  #1274·mngeyuu opened Aug 24, 2026·

  ·
  1

• support dreamzero

  #1259·legitnull opened Aug 4, 2026Collaborator·

  ·
  1

• Add NCCL H3 internal hang diagnostics

  #1257·lzy-edu opened Jul 30, 2026Contributor·

  ·
  1

• Add NCCL H2 collective and P2P mismatch diagnostics

  #1256·lzy-edu opened Jul 30, 2026Contributor·

  ·
  8

• Add CPU-side NCCL H1 missing-enter diagnostics

  #1255·lzy-edu opened Jul 30, 2026Contributor·

  ·
  2

• Support Qwen3.6 LLM Backbone, flash sparse attention and recursive transformer

  #1247·lxd-cumt opened Jul 23, 2026Collaborator·

  ·

• fix: restore GR00T N1.5 dataloader state

  #1236·yilongma0110 opened Jul 14, 2026Contributor·

  ·

• Support hipprof step profiling

  #1234·lihongyang1990 opened Jul 9, 2026Contributor·

  ·

• [NewModel] Add Bagel pretrain

  #1229·zhaoyinglia opened Jun 22, 2026Collaborator·

  ·

• [train][feat, fix]: Support mhc with ep overlapping. A bug fix of dsv4 when use_engram.

  #1218·LiJunscs opened Jun 9, 2026Collaborator·

  ·

• Add context parallel support for dsv4 sparse attention

  #1213·lxd-cumt opened Jun 3, 2026Collaborator·

  ·