Issues

• 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-requestwaiting-on-maintainers
  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-request
  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

• --deterministic-mode without pinning NCCL_ALGO: proposal for an opt-in, with results on 32 B200 nodes (256 GPUs)

  community-requestwaiting-on-maintainers
  NVIDIA/Megatron-LM#8138 · Filter by author KeitaW opened on Oct 12, 2026
  2

  

• [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

• Explicit --ckpt-step can be ignored or load a different checkpoint

  bugmodule: rl
  NVIDIA/Megatron-LM#8110 · Filter by author sbhavani opened on Oct 10, 2026

• [Question] GLM-5.3-Flash training with Megatron FSDP: support plan and validation scope

  community-requestwaiting-on-customer
  NVIDIA/Megatron-LM#8106 · Filter by author gerayking opened on Oct 10, 2026
  1

  

• DistributedOptimizer.load_state_dict copies the allocated optimizer state onto itself, which can OOM during checkpoint load

  community-request
  NVIDIA/Megatron-LM#8046 · Filter by author gavinqii opened on Oct 9, 2026

• Upgrade cudNN to the latest version

  NVIDIA/Megatron-LM#8016 · Filter by author NVSkySC opened on Oct 9, 2026
  

• [ENHANCEMENT] Add opt-in semantic per-head Muon for fused projections

  community-request
  NVIDIA/Megatron-LM#8003 · Filter by author Hollyming opened on Oct 9, 2026
  3

  

• Clarify BOS token handling in Nemotron SFT tokenizer round-trip

  bugcommunity-requestwaiting-on-customer
  NVIDIA/Megatron-LM#8000 · Filter by author lk137095576 opened on Oct 9, 2026
  2

• Failed NIXL transfers are polled indefinitely and retain cache storage

  NVIDIA/Megatron-LM#7996 · Filter by author jthomson04 opened on Oct 8, 2026

• Add a Kimi Delta Attention (KDA) linear-attention layer to Megatron Core

  community-request
  NVIDIA/Megatron-LM#7983 · Filter by author Aravind-11 opened on Oct 8, 2026
  1

• Enable PAO for non-delayed FP8 recipes with BF16 storage

  enhancement
  NVIDIA/Megatron-LM#7968 · Filter by author CarlosGomes98 opened on Oct 8, 2026

• [BUG] Hybrid NVLS batch-invariant prefill differs from full forward on the 26.09 CI stack

  bug
  NVIDIA/Megatron-LM#7958 · Filter by author balasaajay opened on Oct 8, 2026

• [Design Proposal] Custom dataloader contract and checkpoint integration

  NVIDIA/Megatron-LM#7951 · Filter by author maanug-nv opened on Oct 7, 2026
  1

• [inference] Offloaded payload copies all conversation media to host under expanded-prefix stitching; carry only suffix media

  NVIDIA/Megatron-LM#7944 · Filter by author lauradang opened on Oct 7, 2026
  

• loader_mixtral_hf.py fails with transformers 5.x (version check and MoE expert layout)

  NVIDIA/Megatron-LM#7934 · Filter by author janbernloehr opened on Oct 7, 2026

• CI flake: worker TCPStore EADDRINUSE during functional-test startup

  bug
  NVIDIA/Megatron-LM#7920 · Filter by author balasaajay opened on Oct 6, 2026