All pull requests

Pull requests

• refactor: share sequence-wise MoE auxiliary loss

  #1569·Er1ckPang opened Nov 13, 2026Contributor·1 of 5 tasks·

  ·

• feat: add standalone eval run mode with loss and perplexity metrics

  #1568·wang-hua-2019 opened Nov 13, 2026Contributor·4 of 5 tasks·

  ·

• Refactor MLA families and bind TP/CP execution

  #1567·xuxl1008 opened Nov 13, 2026Contributor·3 of 5 tasks·

  ·

• docs: standardize DeepSeek-V4.1 and Kimi-K2.6 training guides

  ci-pipeline-passedpr-check-pass
  #1565·Er1ckPang opened Nov 13, 2026Contributor·1 of 5 tasks·

  ·
  10

• feat: add chunked MoE training with DeepSeek and parent FSDP support

  #1564·xuxl1008 opened Nov 13, 2026Contributor·3 of 4 tasks·

  ·
  2

• feat(compile): support FSDP mixed precision in TextTrainer joint graphs

  ciflow/defaultciflow/pytorch-2.10ciflow/pytorch-2.11
  #1561·yu-chen-code opened Nov 12, 2026Contributor·2 of 5 tasks·

  ·

• chore: clarify contribution workflow and remove standalone lint tests

  ci-pipeline-passedpr-check-pass
  #1559·liuchongming74 opened Nov 12, 2026Contributor·2 of 5 tasks·

  ·
  8

• feat(torch): add opt-in head-chunk MLA with staged offload

  #1557·xuxl1008 opened Nov 12, 2026Contributor·2 of 5 tasks·

  ·

• feat(context-parallel): support sequence-sharded KDA

  #1555·UtopiaFable opened Nov 12, 2026·2 of 5 tasks·

  ·

• fix: release reduced gradients before root finalization

  ci-pipeline-failedpr-check-pass
  #1554·mxysfive opened Nov 12, 2026Contributor·2 tasks done·

  ·
  1

  29

• fix: correct V4.1 CP position and runtime validation

  ci-pipeline-failedpr-check-pass
  #1553·xuxl1008 opened Nov 12, 2026Contributor·

  ·
  10

• feat(optim): overlap Muon communication with Newton-Schulz compute

  pr-check-pass
  #1551·townwish4git opened Nov 12, 2026Contributor·3 of 5 tasks·

  ·
  11

• feat(mhc): replace omni mhc operators with CANN operators

  #1550·GitHuber1024 opened Nov 12, 2026·5 tasks·

  ·
  1

  4

• test: check recipe bundle parity and prune orphan bundles

  ci-pipeline-failedpr-check-pass
  #1549·shawny999 opened Nov 12, 2026Contributor·3 of 5 tasks·

  ·
  28

• feat(integration): add optional VeOmni support for Hyper DTensors

  ci-pipeline-failedpr-check-pass
  #1548·Cui-yshoho opened Nov 12, 2026Contributor·3 of 5 tasks·

  ·
  29

• feat(optimizer): add optional Gram Newton-Schulz backend

  pr-check-pass
  #1546·townwish4git opened Nov 12, 2026Contributor·2 of 5 tasks·

  ·
  1

  21

• feat(cp): add experimental one-sided KDA state pipeline

  #1544·xuxl1008 opened Nov 11, 2026Contributor·3 of 4 tasks·

  ·

• feat: training metrics monitoring

  #1542·wcrzlh opened Nov 11, 2026Contributor·5 tasks·

  ·

• fix(optimizer): use float32 step for fused AdamW

  pr-check-pass
  #1540·townwish4git opened Nov 11, 2026Contributor·5 tasks done·

  ·
  1

  6

• fix(checkpoint): 修复安全区域 SAC 策略与原地 AllReduce 重放

  #1538·xingleixu123 opened Nov 11, 2026Contributor·2 of 5 tasks·

  ·

• refactor(lint): remove obsolete backend import check

  ci-pipeline-passedpr-check-pass
  #1536·xingleixu123 opened Nov 11, 2026Contributor·

  ·
  17

• refactor: remove bundled multicore and use standalone HyperMulticore

  ci-pipeline-failedpr-check-pass
  #1535·forever-moe opened Nov 11, 2026Contributor·

  ·
  43

• fix(compile): enable DeepSeek V4.1 graph FSDP+EP+TP

  ci-pipeline-failedpr-check-pass
  #1534·honghaowang25 opened Nov 11, 2026Contributor·5 tasks done·

  ·
  1

  37

• feat(rl): 实现 Ascend 训推统一执行与跨并行数值一致性

  pr-check-pass
  #1533·haiyangxin opened Nov 11, 2026Contributor·3 of 5 tasks·

  ·
  18

• docs(optim): document Muon HSDP replica count

  pr-check-pass
  #1532·townwish4git opened Nov 11, 2026Contributor·

  ·
  21