All pull requests

Pull requests

• Support custom layers in HybridModel

  complexity: medium
  #9116·Phlip79 opened Nov 8, 2026Member·1 task done·

  ·
  6

• Keep MFSDPv1 gradient buckets until their reduction finishes

  #9114·XiaodaNV opened Nov 8, 2026Contributor·1 task·

  ·
  2

• [Dev] Integrate Megatron FSDP v2 with partial CUDA Graphs

  community-request
  #9112·WEIXILIN0307 opened Nov 8, 2026·3 of 6 tasks·

  ·
  1

• chore: AUT-2775 nightly sync main into dev (2026-11-06)

  determinism-exemptDisable selective unit testsRun CICDRun full unit testsRun functional testsRun MBridge testsRun tests
  #9111·svcnemo-autobot opened Nov 8, 2026Collaborator·

  ·
  2

• [dev][kimi-k3] Fix: Split the KDA unit tests out of the broad ssm CI bucket

  #9110·yuzhongw-nvidia opened Nov 8, 2026Contributor·

  ·
  1

• feat(moe): add optional ep-invariant expert combine

  #9107·ko3n1g opened Nov 7, 2026Contributor·

  ·
  1

• feat(optimizer): integrate mfsdp v2 muon with the optimizer factory

  #9106·ko3n1g opened Nov 7, 2026Contributor·1 task·

  ·
  1

• fix(fsdp): end hfsdp prefetch windows at fsdp unit boundaries

  #9105·ko3n1g opened Nov 7, 2026Contributor·4 of 7 tasks·

  ·
  1

• fix(moe): keep global aux loss counters in int64 under float16module

  #9104·ko3n1g opened Nov 7, 2026Contributor·2 of 7 tasks·

  ·
  1

• fix nemo ckpt security bug

  #9103·dimapihtar opened Nov 7, 2026Contributor·1 of 7 tasks·

  ·
  1

• Handle stub DistributedOptimizer in checkpoint save/load

  complexity: low
  #9099·yueshen2016 opened Nov 7, 2026Contributor·4 of 6 tasks·

  ·
  6

• fix(tokenizers): don't end a Qwen3-Coder function at </function> inside a parameter value

  community-request
  #9097·pcbeingused333 opened Nov 7, 2026·3 of 6 tasks·

  ·
  1

  2

• Support config-defined HybridModel training and external layers

  #9094·Phlip79 opened Nov 7, 2026Member·4 of 7 tasks·

  ·
  1

• test: hold the virtual-expert histogram stable across CUDA graph replays

  community-request
  #9092·mvanhorn opened Nov 7, 2026·1 of 6 tasks·

  ·
  1

  2

• Align mHC layers with the shared residual stream recompute contract

  complexity: medium
  #9091·kwangjunahn opened Nov 7, 2026Contributor·

  ·
  1

  3

  

• fern conversion, cleaner

  #9088·megnvidia opened Nov 7, 2026Contributor·2 of 7 tasks·

  ·
  2

• fix(inference): int64 slot / block offsets in pool-indexing Triton kernels

  #9087·wdykas opened Nov 6, 2026Contributor·4 of 5 tasks·

  ·
  1

• Implement cuDNN backend support for Mamba layers in MambaMixer

  #9086·ghadiaravi13 opened Nov 6, 2026Contributor·1 of 7 tasks·

  ·
  2

• docs: Add Hugging Face MoE training handbook to latest news

  docs-only
  #9084·sbhavani opened Nov 6, 2026Contributor·2 tasks done·

  ·
  1

• Inference: Validate CPU placement during dynamic inference engine initialization

  complexity: lowRun functional testsRun tests
  #9082·mathemakitten opened Nov 6, 2026Contributor·1 of 7 tasks·

  ·
  3

• Combine mHC, Shortcut MoE, gigatoken, and inference GC changes

  Disable selective unit testsRun functional tests
  #9079·balasaajay opened Nov 6, 2026Contributor·

  ·
  6

• Remove Hugging Face import from the ModelOpt post-training examples

  complexity: medium
  #9078·kevalmorabia97 opened Nov 6, 2026Contributor·1 task·

  ·
  1

• Add elastic latent recurrence

  26.12
  #9077·janEbert opened Nov 6, 2026Contributor·1 task·

  ·
  2

• fix(cuda_graphs): refresh cached quantized weights in local CUDA graph replays

  #9076·wdykas opened Nov 6, 2026Contributor·4 of 7 tasks·

  ·
  1

• Ensure docstring quality

  complexity: mediumRun functional tests
  #9075·janEbert opened Nov 6, 2026Contributor·1 task done·

  ·
  5