All pull requests

Pull requests

• ci: Update transformers to latest version 5.18.0

  #4116·svcnvidia-nemo-ci opened Oct 2, 2026Contributor·

  ·
  1

• fix(peft): groundwork for boundary-owned PEFT prefix handling

  community-request
  #4115·stanley1208 opened Oct 2, 2026Contributor·2 of 3 tasks·

  ·
  1

• feat(optim): train grouped MoE experts with Muon

  #4114·HuiyingLi opened Oct 2, 2026Contributor·

  ·
  3

• feat(diffusion): native MoE transformers with expert parallelism

  #4113·HuiyingLi opened Oct 2, 2026Contributor·

  ·
  1

• fix(peft): convert merged export state dicts on host memory

  #4112·pstjohn opened Oct 1, 2026Contributor·

  ·
  1

• chore: CI proxy for #4110

  #4111·akoumpa opened Oct 1, 2026Contributor·

  ·
  2

• fix(glm5_next, kimi_linear): use chunk_kda whenever gradients are needed

  community-requestwaiting-on-customer
  #4110·waynel96 opened Oct 1, 2026Contributor·

  ·
  1

  2

• fix(moe): retain router work in balanced routing benchmarks

  #4107·akoumpa opened Oct 1, 2026Contributor·3 tasks done·

  ·
  1

  3

• chore: CI proxy for #4103

  #4104·akoumpa opened Oct 1, 2026Contributor·

  ·
  3

• fix(moe): avoid retaining routed inputs in chunked checkpoints

  community-request
  #4103·yh0903 opened Oct 1, 2026·9 tasks done·

  ·
  1

  1

• docs(model-cards): enforce recipes, content limits, and checkpoint filenames

  docs-only
  #4102·akoumpa opened Oct 1, 2026Contributor·4 tasks done·

  ·
  9

• feat(checkpoint): support quantized saves for Kimi K2.5 and K3

  #4081·HuiyingLi opened Oct 1, 2026Contributor·5 of 6 tasks·

  ·
  1

• perf(deepseek-v41): clone + slice write instead of torch.cat in _apply_rope

  #4080·yisongbetter opened Oct 1, 2026Contributor·3 tasks done·

  ·
  2

• perf(deepseek-v41): opt-in torch.compile cores for mHC and fp32 RMSNorm

  #4079·yisongbetter opened Oct 1, 2026Contributor·3 tasks done·

  ·
  6

• fix(peft): freeze router biases in LoRA recipes

  #4076·yuhezhang-ai opened Sep 30, 2026Contributor·8 of 10 tasks·

  ·
  3

• fix(deepseek-v4): make LoRA checkpoints reload and resume

  #4072·yuhezhang-ai opened Sep 30, 2026Contributor·8 tasks done·

  ·
  3

• fix(deepseek-v4): shard each fp32 island once under AC [AMINT-368]

  #4068·HuiyingLi opened Sep 30, 2026Contributor·

  ·
  2

• feat(optim): add Muown with FSDP2 and MoE support

  #4065·HuiyingLi opened Sep 29, 2026Contributor·

  ·
  22

• fix(datasets): count skipped batches in the diffusion sampler position

  community-requestwaiting-on-maintainers
  #4059·MohammedAlkindi opened Sep 29, 2026Contributor·

  ·
  1

• fix(moe): map FSDP expert shards with torch.chunk layout

  community-requestwaiting-on-maintainers
  #4058·MohammedAlkindi opened Sep 29, 2026Contributor·

  ·
  5

• feat(mimo): add MiMo-V2.6-Pro support

  #4054·HuiyingLi opened Sep 29, 2026Contributor·

  ·
  1

  9

• fix(qwen3_5_moe): build M-RoPE position ids before the HF VL forward

  community-request
  #4053·naraysa opened Sep 29, 2026·2 of 3 tasks·

  ·
  4

• fix(recipes): reduce seq-cls validation loss over all samples and DP ranks

  community-request
  #4043·MohammedAlkindi opened Sep 28, 2026Contributor·

  ·
  8

• fix(pipeline): pool 1f1b receive buffers by default per stage

  #4041·akoumpa opened Sep 28, 2026Contributor·

  ·
  18

• fix(build): pin C++20-compatible CUDA dependencies

  #4036·akoumpa opened Sep 28, 2026Contributor·4 tasks done·

  ·
  5