Issues

• [feature] Support CP & PP for Gemma4 31B

  area:modelblockedfeature
  NVIDIA-NeMo/Megatron-Bridge#6276 · Filter by author sophiayyya opened on Oct 1, 2026
  1

• [feature] Support pipeline-stage-local HF export for RL weight synchronization

  area:ckptfeatureneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6268 · Filter by author sbhavani opened on Sep 30, 2026
  1

• get_model_config() returns incorrect configs for families without a config mapping

  area:modelbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6242 · Filter by author jayzuccarelli opened on Sep 29, 2026

• QKV split helpers create CPU index tensors for CUDA weights

  area:ckptbugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6229 · Filter by author sbhavani opened on Sep 28, 2026

• 📣 [Announcement] Megatron Bridge will be consolidated with Megatron LM

  area:miscdocstracking
  NVIDIA-NeMo/Megatron-Bridge#6172 · Filter by author cuichenx opened on Sep 22, 2026

• [bug] DeepSeek-V4 SFT recipes default use_fused_mhc=True on Blackwell, but the fused backward aborts on variable-length micro-batches

  area:recipebugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6161 · Filter by author linmuchuiyang opened on Sep 20, 2026

• [bug] gpt_step does not pass loss_mask to the model, so the MTP loss (and its gradient) is computed over padding

  area:trainingbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6103 · Filter by author linmuchuiyang opened on Sep 15, 2026

• [model] Add support for Qwen3.8-Flash-Next

  area:modelcommunity-requestfeaturewaiting-on-maintainers
  NVIDIA-NeMo/Megatron-Bridge#6060 · Filter by author NINGBENZHE opened on Sep 11, 2026
  1

• [bug] NemotronH-56B 256-GPU BF16 perf recipes keep the 64-GPU global_batch_size (192), so config validation fails at DP=128

  area:perfbugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6024 · Filter by author janbernloehr opened on Sep 9, 2026
  1

• moe_a2a_overlap=false cannot disable A2A overlap when the recipe enables it

  area:perfbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6006 · Filter by author rutayan-nv opened on Sep 8, 2026
  1

• [ckpt] GPU export overwrites GLM-5.2 max_position_embeddings (1048576 → training seq_length 8192) in the exported config.json (related: #5153)

  area:ckptbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5994 · Filter by author linmuchuiyang opened on Sep 7, 2026
  2

• nemo:26.08 aarch64 image: PyTorch built against cuDNN 9.23.0 but only libcudnn 9.21.1 is installed — torch.backends.cudnn.version() raises

  area:buildbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5993 · Filter by author linmuchuiyang opened on Sep 7, 2026

• [bug] Transformer Engine fused cross-entropy zeroes MTP acceptance metrics

  area:trainingbugwaiting-on-maintainers
  NVIDIA-NeMo/Megatron-Bridge#5947 · Filter by author milank94 opened on Sep 3, 2026
  5

• [ALERT] Access to NVIDIA's Self-Hosted Runners is Expiring

  area:cicinvidia-runnerswaiting-on-maintainers
  NVIDIA-NeMo/Megatron-Bridge#5886 · Filter by author nvidia-runner-mgmt opened on Aug 30, 2026
  1

  

• [bug] _resolve_hf_tokenizer_revision calls snapshot_download without local_files_only, breaking revision-pinned tokenizers on pre-populated offline caches

  area:trainingbugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5807 · Filter by author janbernloehr opened on Aug 26, 2026

• PoR: migrate HybridEP workloads to NCCL EP

  area:perfarea:trainingfeaturePoRtracking
  NVIDIA-NeMo/Megatron-Bridge#5775 · Filter by author yaoyu-33 opened on Aug 24, 2026

• PoR: define ModelOpt refit and non-ModelOpt checkpoint boundaries

  area:ckptarea:modelarea:quantfeaturePoRtracking
  NVIDIA-NeMo/Megatron-Bridge#5773 · Filter by author yaoyu-33 opened on Aug 24, 2026

• PoR: turn model verification cards into a navigable recipe catalog

  area:miscfeaturePoRtracking
  NVIDIA-NeMo/Megatron-Bridge#5772 · Filter by author yaoyu-33 opened on Aug 24, 2026

• [feature] Lift the expert-LoRA refusals for DoRA, shared-outer adapters, expert dropout, and grouped-adapter weight merge

  area:peftcommunity-requestfeatureneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5636 · Filter by author tigist-far opened on Aug 18, 2026

• [bug] Expert LoRA computes wrong deltas and gradients at expert_tensor_parallel_size > 1 (and at TP > 1 under moe_shared_expert_overlap)

  area:peftbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5634 · Filter by author tigist-far opened on Aug 18, 2026

• [bug] Mistral YaRN rope scaling crashes provider_bridge: MistralModelProvider has no scale_factor field

  area:modelbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5627 · Filter by author lonexreb opened on Aug 17, 2026

• [bug] Energon validation set is shuffled and never resets — validation loss is not reproducible

  area:databugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5626 · Filter by author psycherun-kaiko opened on Aug 17, 2026

• [model] Add Ling 3.0 Tiny and Flash checkpoint conversion support

  area:modelcommunity-requestfeatureneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#5602 · Filter by author fkuner opened on Aug 16, 2026

• PoR: [model] tracking: GLM-5.2 production training qualification

  area:ckptarea:modelarea:perfarea:recipearea:trainingfeaturePoRtracking
  NVIDIA-NeMo/Megatron-Bridge#5476 · Filter by author yaoyu-33 opened on Aug 10, 2026

• [bug] MiniMax-M3 EP32 inference produces step-0 NaN logits with all-to-all

  area:modelblockedbugmodel-minimax
  NVIDIA-NeMo/Megatron-Bridge#5462 · Filter by author yaoyu-33 opened on Aug 10, 2026
  1