Issues

• Minor Redesign of config serialization+deserialization

  Initiative

  NVIDIA/Megatron-LM#8459 · Filter by author maanug-nv opened on Oct 23, 2026
  

• Make TrainState available in CallbackContext for callback system

  enhancement
  Task

  NVIDIA/Megatron-LM#8438 · Filter by author maanug-nv opened on Oct 22, 2026

• Log ConfigContainer to WandB and TensorBoard

  Task

  NVIDIA/Megatron-LM#8437 · Filter by author maanug-nv opened on Oct 22, 2026

• Review and clean up unit tests after Bridge migration is complete

  Task

  NVIDIA/Megatron-LM#8436 · Filter by author maanug-nv opened on Oct 22, 2026

• Config validation and finalization cleanup

  Task

  NVIDIA/Megatron-LM#8435 · Filter by author maanug-nv opened on Oct 22, 2026

• Determine if tokenizer can be removed from MCore DatasetConfigs

  Task

  NVIDIA/Megatron-LM#8433 · Filter by author maanug-nv opened on Oct 22, 2026

• MegaMoE Support Status and Roadmap

  Enhancement

  NVIDIA/Megatron-LM#8427 · Filter by author QiZhangNV opened on Oct 22, 2026
  

• DeepSeek-V4.1-Flash optimizer support

  26.12enhancement
  NVIDIA/Megatron-LM#8416 · Filter by author janEbert opened on Oct 21, 2026
  

• HybridEP retains dispatch metadata after combine

  NVIDIA/Megatron-LM#8409 · Filter by author jepio opened on Oct 21, 2026

• Batch the vision-encoder forward across waiting multimodal requests in the dynamic inference engine

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8407 · Filter by author going-song opened on Oct 21, 2026

• [BUG] MLA recomputation can read offloaded inputs before their reload finishes

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8396 · Filter by author LitLeo opened on Oct 21, 2026
  3

  

• GTP + Kimi K3 perf optimization

  Enhancement

  NVIDIA/Megatron-LM#8339 · Filter by author fanshiqing opened on Oct 19, 2026
  

• [BUG] Megatron-FSDP v1 local Linear overwrites accumulated wgrads in unsharded gradient buffers

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8332 · Filter by author tang-t21 opened on Oct 19, 2026

• NeMo flash-attention drops sequence lengths, and the packed-projector and decode errors name the wrong piece

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8325 · Filter by author Aravind-11 opened on Oct 18, 2026

• MiMo HF Whisper attends through the padded mel tail

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8324 · Filter by author Aravind-11 opened on Oct 18, 2026

• Revert CG+CP fix once NVIDIA/TransformerEngine#3205 lands

  bug
  NVIDIA/Megatron-LM#8271 · Filter by author tdene opened on Oct 15, 2026

• CI: hybrid MFSDP CUDA-graph validation receives only 3 of 50 metric steps

  bug
  NVIDIA/Megatron-LM#8218 · Filter by author balasaajay opened on Oct 14, 2026
  

• [Tracking] Disaggregated inference: remaining limitations and follow-up work

  NVIDIA/Megatron-LM#8200 · Filter by author wdykas opened on Oct 14, 2026
  2

• Dataset index cache: rank-0 crash mid-numpy.save leaves torn .npy that passes the isfile cache hit-test — restarted job crashes in dataloader

  community-request
  NVIDIA/Megatron-LM#8189 · Filter by author rkoivunen-sw opened on Oct 13, 2026
  1

• Support expert deployment tensor redistribution across EP layouts in mLite

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8186 · Filter by author lidongdongbuaa opened on Oct 13, 2026
  1

  

• [QUESTION] Why is SM120 excluded from thd_layout_kernels?

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8183 · Filter by author felixffl opened on Oct 13, 2026
  

• [BUG] MFSDP v2 combined 1F1B prematurely reduces gradients with tied embedding/output weights

  bugcommunity-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8179 · Filter by author zpguo23 opened on Oct 13, 2026
  3

  

• [BUG] GDN dynamic inference: illegal memory access on prefill/mixed CUDA graph replay

  bug
  NVIDIA/Megatron-LM#8151 · Filter by author anlthms opened on Oct 12, 2026
  

• [BUG] FP32 MoE router has TE/fallback numerical mismatch on H200 hidden by test tolerances; downstream GLM parity diverges

  community-requestwaiting-on-customer
  NVIDIA/Megatron-LM#8126 · Filter by author tang-t21 opened on Oct 12, 2026
  1

• [ENHANCEMENT] Opt-in PyTorch grouped weight gradients for SM90 BF16 MoE experts

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8117 · Filter by author ggboooy opened on Oct 11, 2026