Issues

• [bug] LoRA under sequence parallelism drops the LoRA-A input gradient (default sequence_parallel_input_regather=False)

  area:peftbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6453 · Filter by author SumanthRH opened on Oct 14, 2026

• [tracking] Kimi K3: migrate Bridge model code to native Megatron-Core

  area:modelfeaturemlm-syncmodel-kimitracking
  NVIDIA-NeMo/Megatron-Bridge#6451 · Filter by author cuichenx opened on Oct 14, 2026

• [bug] Llama3.1 405B NVFP4 pretrain on 256x GB200: grad norm diverges to inf after step 4 (CP=1)

  area:trainingbugwaiting-on-customer
  NVIDIA-NeMo/Megatron-Bridge#6446 · Filter by author rsalagame-nvidia opened on Oct 13, 2026
  1

  

• [perf] Optimize Nemotron 3.5 Lightning FP8-MX performance on GB300 and GB200

  area:perffeatureneeds-triageperformance
  NVIDIA-NeMo/Megatron-Bridge#6361 · Filter by author rhmukundan opened on Oct 8, 2026

• [perf] NVFP4 performance worse than FP8-MX for DeepSeek V3 on GB200

  area:perfperformance
  NVIDIA-NeMo/Megatron-Bridge#6357 · Filter by author rhmukundan opened on Oct 8, 2026

• [perf] NVFP4 performance worse than FP8-MX for Nemotron 3 Super on GB200

  area:perfperformance
  NVIDIA-NeMo/Megatron-Bridge#6353 · Filter by author rhmukundan opened on Oct 8, 2026

• [perf] NVFP4 performance worse than FP8-MX for Nemotron 3 Ultra

  area:perfperformance
  NVIDIA-NeMo/Megatron-Bridge#6352 · Filter by author rhmukundan opened on Oct 8, 2026

• [feature] Energon cache pool is not exposed in the multimodal datamodule and dataset config

  area:datacommunity-requestfeatureneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6326 · Filter by author tkmtakada opened on Oct 6, 2026
  

• [bug] Mamba in_proj LoRA linear_out is gathered in per-rank order under TP>1 (merged and adapter export)

  area:peftbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6310 · Filter by author yoavf opened on Oct 4, 2026

• [bug] Qwen3-30B-A3B GB200 MXFP8 perf preset (full_iteration CUDA graphs + HybridEP) OOMs in first optimizer step on PyTorch >=2.14 / TE >=2.18 stacks

  area:perfbugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6307 · Filter by author janbernloehr opened on Oct 3, 2026
  1

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

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

• 📣 [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

• [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

• 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