All pull requests

Pull requests

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

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

  ·
  1

• [CICD] Add PPU 810E training CI support

  #1300·jfhx opened Sep 22, 2026·

  ·
  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·

  ·
  3

• [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

• Fix diagnostics stop lifecycle

  #1271·lzy-edu opened Aug 19, 2026Contributor·

  ·
  1

• Fix process heartbeat false positives during checkpointing

  #1268·lzy-edu opened Aug 17, 2026Contributor·

  ·
  3

• Fix vpp

  #1266·cjwystz opened Aug 14, 2026Contributor·

  ·

• 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·

  ·
  1

• 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·

  ·

• Feat/qwen3.5 fp8toint8

  #1212·1235v opened Jun 2, 2026·

  ·
  5

• [CICD] Add MetaX image build workflow and Fix containers

  #1211·BrianPei opened May 28, 2026Collaborator·

  ·
  13

• [CICD] Expand unit test coverage on CUDA

  #1210·BrianPei opened May 28, 2026Collaborator·

  ·
  10

• [CICD] Add Debian + RPM packaging

  #1205·shiptux opened May 21, 2026·

  ·