Issues

• [26.10.01.rc1] DeepSeek V4 Flash FP8-MX on B300 (128 GPUs): CUDA graph capture invalidated

  bugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6894 · Filter by author rsalagame-nvidia opened on Nov 12, 2026

• [bug] [26.10.01.rc1] Nemotron 3 Ultra FP8-MX on GB200 (256 GPUs): _LayerNormLinearBackward returned invalid gradient shape

  bugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6893 · Filter by author rsalagame-nvidia opened on Nov 12, 2026

• [26.10.01.rc1] DeepSeek V3 pretrain on VR NVL72 fails: deterministic dprob only implemented for SM100 (cuDNN grouped GEMM DGLU) for both 128 and 256x gpu scale

  bugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6892 · Filter by author rsalagame-nvidia opened on Nov 12, 2026

• 📣 [Announcement] Megatron Bridge/LM Consolidation Timeline

  NVIDIA-NeMo/Megatron-Bridge#6888 · Filter by author cuichenx opened on Nov 11, 2026

• [bug] Qwen3-VL: packed MoE padding mask is not sharded for sequence parallelism on later pipeline stages

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6880 · Filter by author zsombor-kaiko opened on Nov 11, 2026

• [feature] Publish the Megatron-LM commit pinned by nrl-r0.8.0 (NeMo RL v0.8.0) under an NVIDIA-owned ref

  community-request
  NVIDIA-NeMo/Megatron-Bridge#6876 · Filter by author mcmasruntotop opened on Nov 11, 2026

• [bug] AutoBridge import silently corrupts finegrained_fp8 checkpoints: raw fp8 codes become weights, every scale key is dropped

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6875 · Filter by author YancyLii opened on Nov 11, 2026

• [bug] Config-only HF export (from_auto_config) silently drops the Megatron-side RoPE and backfills the reference model's

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6873 · Filter by author YancyLii opened on Nov 11, 2026

• [bug] AutoBridge conversion with TP > 1 rejects checkpoints whose vocab size is not divisible by TP

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6872 · Filter by author YancyLii opened on Nov 11, 2026

• [bug] Mapping registry resolves */** wildcard captures in the wrong order when * precedes **

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6871 · Filter by author YancyLii opened on Nov 11, 2026
  2

• [bug] [Muse-Glimmer] Repeated vision RoPE casts prematurely round BF16 gradients

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6866 · Filter by author caohch-1 opened on Nov 10, 2026

• [bug] [Muse-Glimmer] BF16 model conversion truncates vision RoPE frequencies

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6864 · Filter by author caohch-1 opened on Nov 10, 2026

• [bug] Ministral3 Llama-4 attention scale uses the local buffer index instead of the token position (packed / CP)

  community-request
  NVIDIA-NeMo/Megatron-Bridge#6852 · Filter by author 0z5a opened on Nov 9, 2026

• [bug] Sarvam MoE export writes n_group/topk_group as null, so the exported model fails in HF

  community-request
  NVIDIA-NeMo/Megatron-Bridge#6851 · Filter by author 0z5a opened on Nov 9, 2026

• ERNIE 5.5: LoRA on the SD-MTP-aliased modules fails at model build in check_ernie55_mtp_aliases

  community-request
  NVIDIA-NeMo/Megatron-Bridge#6848 · Filter by author linmuchuiyang opened on Nov 9, 2026

• [bug] [MiMo-V2-Flash] Router logits default to BF16 instead of the reference FP32 computation

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6846 · Filter by author caohch-1 opened on Nov 9, 2026
  1

• [bug] [Ministral3] YaRN conversion does not preserve HF's default truncate=True

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6845 · Filter by author caohch-1 opened on Nov 9, 2026
  3

• [bug] [Qwen3.5-VL] Patch merger uses tanh GELU instead of the HF reference

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6844 · Filter by author caohch-1 opened on Nov 9, 2026

• [bug] [Step-3.7] Sliding-window conversion includes one extra token

  bugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6842 · Filter by author caohch-1 opened on Nov 9, 2026
  1

• [bug] Per-expert LoRA init repeats across EP ranks and changes with EP size

  area:peftbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6824 · Filter by author excepshenal opened on Nov 6, 2026

• [bug] MultiLoRAGroupedExpertLinear.reset_adapter inits expert LoRA A from the packed 3-D fans

  area:peftbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6822 · Filter by author excepshenal opened on Nov 6, 2026

• [bug] [MiMo-V2-Flash] fails with VPP due to missing vp_stage forwarding

  area:modelbugcommunity-requestneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6808 · Filter by author caohch-1 opened on Nov 6, 2026

• [bug] Nemotron 3 Ultra NCCL-EP fails during GIN bootstrap on B300 with RDMA retry exhaustion

  bugneeds-triage
  NVIDIA-NeMo/Megatron-Bridge#6790 · Filter by author rhmukundan opened on Nov 4, 2026

• [bug] SLURM master address loses suffix in multi-group hostlists

  community-request
  NVIDIA-NeMo/Megatron-Bridge#6785 · Filter by author 0z5a opened on Nov 4, 2026

• [tracking] Replace perf_recipes with library recipes and explicit benchmark overrides

  area:recipefeaturetracking
  NVIDIA-NeMo/Megatron-Bridge#6680 · Filter by author cuichenx opened on Oct 27, 2026