NVIDIA-NeMo/Megatron-Bridge

View on GitHub

Training library for Megatron-based models with bidirectional Hugging Face conversion capability

Stars: 920
Forks: 505
Open beginner issues: 0
Indexed issues: 263
Avg merge: 1d 16h
Merged PRs (30d): 253

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 18, 2026
Latest indexed: Sep 19, 2026
Contributing guide: Contributing guide
Code of conduct: No code of conduct
Beginner labels: good first issue help wanted

Beginner friendly Everything

Open All

263 open issues indexed

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

area:training bug community-request needs-triage

Difficulty 2/5 1-3 hours Newbie friendliness 84/100

NVIDIA-NeMo/Megatron-Bridge#6103 · Opened Sep 15, 2026
[bug] SafeTensorsStateSource._resolve_path ignores the requested revision; offline caches pinned to a commit sha fail with LocalEntryNotFoundError Open

area:ckpt bug community-request needs-triage

Difficulty 3/5 1-2 days Newbie friendliness 74/100

NVIDIA-NeMo/Megatron-Bridge#6104 · Opened Sep 15, 2026
[model] Add support for Qwen3.8-Flash-Next Open

area:model community-request feature waiting-on-maintainers

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#6060 · 1 comment · Opened Sep 11, 2026
[feature] Enable model-specific custom TaskEncoders in the Energon data path Open

area:data community-request feature waiting-on-customer

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#6048 · 2 comments · Opened Sep 10, 2026
[bug] NemotronH-56B 256-GPU BF16 perf recipes keep the 64-GPU global_batch_size (192), so config validation fails at DP=128 Open

area:perf bug needs-triage

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

NVIDIA-NeMo/Megatron-Bridge#6024 · 1 comment · Opened Sep 9, 2026
moe_a2a_overlap=false cannot disable A2A overlap when the recipe enables it Open

area:perf bug community-request needs-triage

Difficulty 2/5 1-3 hours Newbie friendliness 85/100

NVIDIA-NeMo/Megatron-Bridge#6006 · 1 comment · Opened Sep 8, 2026
[ckpt] GPU export overwrites GLM-5.2 `max_position_embeddings` (1048576 → training seq_length 8192) in the exported `config.json` (related: #5153) Open

area:ckpt bug community-request needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 56/100

NVIDIA-NeMo/Megatron-Bridge#5994 · 2 comments · Opened Sep 7, 2026
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 Open

area:build bug community-request needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA-NeMo/Megatron-Bridge#5993 · Opened Sep 7, 2026
[training] flex dispatcher downgrades an explicitly requested `deepep` to `alltoall` on GB200 (warning only, no error) — clarify GB200 policy or fail fast Open

area:training bug community-request waiting-on-maintainers

Difficulty 4/5 3-5 days Newbie friendliness 48/100

NVIDIA-NeMo/Megatron-Bridge#5992 · 1 comment · Opened Sep 7, 2026
[bug] temporary_distributed_context has a singleton TCP rendezvous race Open

area:ckpt bug community-request needs-triage

Difficulty 3/5 1-2 days Newbie friendliness 74/100

NVIDIA-NeMo/Megatron-Bridge#5967 · Opened Sep 4, 2026
[bug] Transformer Engine fused cross-entropy zeroes MTP acceptance metrics Open

area:training bug community-request waiting-on-maintainers

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA-NeMo/Megatron-Bridge#5947 · 3 comments · Opened Sep 3, 2026
[ALERT] Access to NVIDIA's Self-Hosted Runners is Expiring Open

area:ci ci nvidia-runners waiting-on-maintainers

NVIDIA-NeMo/Megatron-Bridge#5886 · 1 assignee · Opened Aug 30, 2026
[bug] _resolve_hf_tokenizer_revision calls snapshot_download without local_files_only, breaking revision-pinned tokenizers on pre-populated offline caches Open

area:training bug needs-triage

Difficulty 2/5 1-3 hours Newbie friendliness 82/100

NVIDIA-NeMo/Megatron-Bridge#5807 · Opened Aug 26, 2026
PoR: migrate HybridEP workloads to NCCL EP Open

area:perf area:training feature PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 30/100

NVIDIA-NeMo/Megatron-Bridge#5775 · Opened Aug 24, 2026
PoR: define ModelOpt refit and non-ModelOpt checkpoint boundaries Open

area:ckpt area:model area:quant feature PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5773 · Opened Aug 24, 2026
PoR: turn model verification cards into a navigable recipe catalog Open

area:misc feature PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 28/100

NVIDIA-NeMo/Megatron-Bridge#5772 · Opened Aug 24, 2026
[feature] Lift the expert-LoRA refusals for DoRA, shared-outer adapters, expert dropout, and grouped-adapter weight merge Open

area:peft community-request feature needs-triage

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#5636 · Opened 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) Open

area:peft bug community-request needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA-NeMo/Megatron-Bridge#5634 · Opened Aug 18, 2026
[bug] Mistral YaRN rope scaling crashes provider_bridge: MistralModelProvider has no scale_factor field Open

area:model bug community-request needs-triage

Difficulty 3/5 1-2 days Newbie friendliness 48/100

NVIDIA-NeMo/Megatron-Bridge#5627 · Opened Aug 17, 2026
[bug] Energon validation set is shuffled and never resets — validation loss is not reproducible Open

area:data bug community-request needs-triage

Difficulty 2/5 1-3 hours Newbie friendliness 78/100

NVIDIA-NeMo/Megatron-Bridge#5626 · Opened Aug 17, 2026
[model] Add Ling 3.0 Tiny and Flash checkpoint conversion support Open

area:model community-request feature needs-triage

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#5602 · Opened Aug 16, 2026
PoR: [model] tracking: GLM-5.2 production training qualification Open

area:ckpt area:model area:perf area:recipe area:training feature PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5476 · Opened Aug 10, 2026
[bug] MiniMax-M3 EP32 inference produces step-0 NaN logits with all-to-all Open

area:model blocked bug model-minimax

Difficulty 5/5 Over a week Newbie friendliness 32/100

NVIDIA-NeMo/Megatron-Bridge#5462 · 1 comment · Opened Aug 10, 2026
[feature] Integrate Megatron-Core Generalized Tensor Parallelism (GTP) Open

area:training feature tracking

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5434 · Opened Aug 7, 2026
[bug] BF16 sequence parallelism causes significant gradient norm drift with TP=4 on H200 Open

area:training bug community-request waiting-on-maintainers

Difficulty 5/5 Over a week Newbie friendliness 30/100

NVIDIA-NeMo/Megatron-Bridge#5411 · 3 comments · Opened Aug 7, 2026
[bug] Warmup LR is one step behind: OptimizerParamScheduler not advanced before the first optimizer step Open

area:training bug community-request waiting-on-maintainers

Difficulty 3/5 1-2 days Newbie friendliness 72/100

NVIDIA-NeMo/Megatron-Bridge#5399 · 2 comments · Opened Aug 7, 2026
[Tracking] Simplify Bridge conversion/modeling for standalone installation and RL integrations Open

area:ckpt feature PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5371 · 2 comments · 1 reaction · Opened Aug 6, 2026
[support] Qwen3.5-VL-35B-A3B MFU is less than 10% on B300. Open

area:perf community-request support waiting-on-maintainers

Difficulty 4/5 3-5 days Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#5318 · 5 comments · Opened Aug 4, 2026
[bug] DeepSeek V4 Pro FP8-MX pretraining gives gradient norm 'nan' on GB300 Open

area:recipe bug waiting-on-maintainers

NVIDIA-NeMo/Megatron-Bridge#5306 · 2 comments · 1 assignee · Opened Aug 4, 2026
[bug] Qwen35VLBridge MTP mappings use wrong Megatron prefix (language_model.mtp.* instead of mtp.*), crashes weight import Open

area:model bug community-request waiting-on-customer

Difficulty 2/5 1-3 hours Newbie friendliness 74/100

NVIDIA-NeMo/Megatron-Bridge#5252 · 1 comment · Opened Aug 2, 2026
[feature] Support DSpark speculative decoding draft training Open

area:model community-request feature waiting-on-maintainers

NVIDIA-NeMo/Megatron-Bridge#5230 · 5 comments · 1 assignee · Opened Aug 1, 2026
[support] How to determine if the DeepSeek V4 model conversion is correct? Open

area:ckpt community-request support waiting-on-customer

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5206 · 3 comments · Opened Jul 31, 2026
Unify checkpoint-derived HF config across CPU and GPU export backends Open

area:ckpt bug tracking

Difficulty 4/5 3-5 days Newbie friendliness 45/100

NVIDIA-NeMo/Megatron-Bridge#5153 · 1 comment · Opened Jul 28, 2026
[feature] Support SequentialMLP (non-fused MoE) AutoBridge Mappings for P0 models Open

area:model feature needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 55/100

NVIDIA-NeMo/Megatron-Bridge#5145 · Opened Jul 28, 2026
[NeMo FW 26.08 Release] MBridge 0.6.0 Roadmap Open

area:misc feature roadmap tracking

Difficulty 5/5 Over a week Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#5144 · 1 reaction · Opened Jul 28, 2026
[bug] [DeepEP][B200+B300][26.06.01] Timeout in internode_dispatch (CPU) during MoE token dispatch Open

area:perf bug waiting-on-maintainers

NVIDIA-NeMo/Megatron-Bridge#5067 · 10 comments · 1 assignee · Opened Jul 24, 2026
[bug] [HybridEP][B300][26.06.01] Segfault in RDMACoordinator::allocate_combine_buffers at training start Open

area:perf bug needs-triage

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#5051 · 2 comments · Opened Jul 23, 2026
[bug] [DeepEP][B300][26.04.01] NVSHMEM ibgda fails to initialize — "Unable to create ah" / DCT share error Open

area:perf bug needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 42/100

NVIDIA-NeMo/Megatron-Bridge#5050 · 2 comments · Opened Jul 23, 2026
[bug] 26.06.01 GPT-OSS 50% Performance Drop Open

area:perf bug performance tracking

NVIDIA-NeMo/Megatron-Bridge#5028 · 1 comment · 1 assignee · Opened Jul 22, 2026
Verify evals on Papers with Code Open

area:model community-request needs-triage support

Difficulty 3/5 1-2 days Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#5017 · Opened Jul 22, 2026
[support] Multi-node NCCL/RDMA training hangs appear related to forked DataLoader workers on H20 Open

area:data community-request needs-triage support

Difficulty 5/5 Over a week Newbie friendliness 28/100

NVIDIA-NeMo/Megatron-Bridge#5006 · Opened Jul 22, 2026
[support] Expert model parallel size allowed to exceed DP ranks. Open

area:training community-request support waiting-on-customer

Difficulty 4/5 3-5 days Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#4980 · 3 comments · Opened Jul 21, 2026
[bug] GPT-OSS HybridModel lacks EP-overlap schedule support after migration Open

area:model bug needs-triage

Difficulty 5/5 Over a week Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#4975 · Opened Jul 20, 2026
[model] Add verified Nemotron 3 Nano 4B CPT support Open

area:model feature model-fullrecipe model-nemotron needs-triage

Difficulty 5/5 Over a week Newbie friendliness 32/100

NVIDIA-NeMo/Megatron-Bridge#4967 · 1 comment · 1 reaction · Opened Jul 20, 2026
PoR: [model] tracking: Kimi K3 roadmap Open

area:model feature model-kimi PoR tracking

Difficulty 5/5 Over a week Newbie friendliness 18/100

NVIDIA-NeMo/Megatron-Bridge#4910 · 1 comment · Opened Jul 16, 2026
[Determinism] determinism diagnostic tool Open

area:perf Determinism feature tracking

NVIDIA-NeMo/Megatron-Bridge#4907 · 2 assignees · Opened Jul 16, 2026
[bug] [Determinism] determinism break of ultra v3 on large scale Open

area:model bug Determinism tracking

NVIDIA-NeMo/Megatron-Bridge#4906 · 2 assignees · Opened Jul 16, 2026
[determinism][CICD] add dsv4 determinism test for perf and verification Open

area:perf ci Determinism tracking

NVIDIA-NeMo/Megatron-Bridge#4905 · 1 assignee · Opened Jul 16, 2026
[determinism][CICD] add dsv3 determinism test for perf and verification Open

area:perf ci Determinism needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 25/100

NVIDIA-NeMo/Megatron-Bridge#4904 · Opened Jul 16, 2026
[determinism][CICD] add nemotron ultra determinism test for perf and verification Open

area:perf ci Determinism needs-triage

Difficulty 4/5 3-5 days Newbie friendliness 35/100

NVIDIA-NeMo/Megatron-Bridge#4903 · Opened Jul 16, 2026

NVIDIA-NeMo/Megatron-Bridge

Get new issues in your inbox