{"full_name":"NVIDIA-NeMo/Megatron-Bridge","owner":"NVIDIA-NeMo","name":"Megatron-Bridge","shard":"1","url":"https://github.com/NVIDIA-NeMo/Megatron-Bridge","description":"Training library for Megatron-based models with bidirectional Hugging Face conversion capability","homepage":"https://docs.nvidia.com/nemo/megatron-bridge/latest/","topics":[],"tag_pages":[],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/213689629?v=4","created_at":"2025-05-21T21:45:06Z","pushed_at":"2026-11-16T08:43:38Z","metrics":{"stars":997,"forks":583,"watchers":9,"issues":985,"prs":5947,"releases":20,"issues_open":284,"issues_closed":701,"prs_open":261,"prs_closed":1582,"prs_merged":4104,"discussions":5},"growth":{"day":{"stars":1,"forks":1,"base_day":"2026-11-15"}},"history":{"dates":["2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[990,991,992,996,996,997],"forks":[577,579,579,582,582,583],"watchers":[8,8,8,9,9,9],"issues_open":[279,283,283,284,284,284],"prs_open":[247,253,255,257,253,261]},"boards":{},"same_owner":[{"full_name":"NVIDIA-NeMo/Speech","name":"Speech","shard":"1","boards":{}}],"issues":[{"number":4214,"title":"Compare Axolotl Qwen vs MBridge Qwen vs MBridge Nemotron Super LoRA SFT performance","comments":14,"created_at":"2026-06-08T20:05:46Z"},{"number":1255,"title":"Nan grad norm in MXFP8 training","comments":14,"created_at":"2025-11-10T05:06:37Z"},{"number":3555,"title":"[bug] There seems to be a consistent increase in CPU memory while using MBridge","comments":11,"created_at":"2026-04-28T15:56:39Z"},{"number":5067,"title":"[bug] [DeepEP][B200+B300][26.06.01] Timeout in internode_dispatch (CPU) during MoE token dispatch","comments":10,"created_at":"2026-07-24T05:23:27Z"},{"number":4609,"title":"[feature] intra-microbatch reordering for MegatronMIMO (+ sequence packing, scalable DP)","comments":10,"created_at":"2026-07-01T06:15:59Z"},{"number":4351,"title":"[feature] Support offline tokenized + index-only packed mode for SFT datasets","comments":10,"created_at":"2026-06-14T10:35:08Z"},{"number":2785,"title":"[bug] Torch FSDP2 does not work","comments":10,"created_at":"2026-03-13T07:17:24Z"},{"number":1986,"title":"Support and switching mechanism for multiple LoRA adapters","comments":10,"created_at":"2026-01-19T10:34:35Z"}],"discussions":[{"number":4957,"title":"Accelerating GLM-5.2 Long-Context Training with cuDNN Fused DSA Kernels and THD-Packed Context Parallelism using Megatron Bridge","comments":0,"category":"Announcements","created_at":"2026-07-19T22:31:44Z"},{"number":3460,"title":"Low-Precision Training with Megatron-Bridge: FP8 (Current Scaling), MXFP8, NVFP4 vs BF16","comments":0,"category":"Announcements","created_at":"2026-04-21T23:08:40Z"},{"number":2935,"title":"Dapt traing with 8xb200","comments":0,"category":"Q&A","created_at":"2026-03-21T00:48:21Z"},{"number":1609,"title":"TildeOpen30B distilation","comments":3,"category":"Q&A","created_at":"2025-12-05T13:44:31Z"},{"number":1320,"title":"What is the status of DeepEP/HybridEP Support in Megatron Bridge and what's their hardware requirement?","comments":1,"category":"General","created_at":"2025-11-12T18:42:08Z"}],"latest_release":{"name":"NVIDIA Megatron-Bridge 0.7.0","tag":"v0.7.0","published_at":"2026-10-30T03:39:25Z","body":"## **Highlights**\r\n\r\n### **Model Collection Support**\r\n\r\n#### **LLM**\r\n\r\n* **DeepSeek V4** — enhanced support with MCore `HybridModel` on main branch; library configs modernized; multi-node inference launcher ([PR\\#5860](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5860?ref=relnotes), [PR\\#5554](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5554?ref=relnotes), [PR\\#6163](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6163?ref=relnotes))  \r\n* **GLM-5.2** — enhanced support with MCore `HybridModel`, with shared GLM-5.2/5.3 bridge compatibility ([PR\\#6158](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6158?ref=relnotes), [PR\\#6117](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6117?ref=relnotes))  \r\n* **Kimi K3** — language backbone ([PR\\#5130](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5130?ref=relnotes))  \r\n* **Muse Glimmer 30B** ([PR\\#5479](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5479?ref=relnotes), [PR\\#6255](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6255?ref=relnotes))  \r\n* **K-EXAONE 2.0** ([PR\\#5191](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5191?ref=relnotes), community @eagle705)  \r\n* **Muse Spark Open 816B-A42B** and **Beam 1.0 501B-A23B** are on `main`, not in 0.7.0. Verified with Megatron Bridge `main` in the NeMo 26.10 container plus Transformers 5.20.1; see their verification cards.\r\n\r\n#### **VLM / Multimodal**\r\n\r\n* **Nemotron 3.5 Super VL** — support, text-only conversion, MoE fine-tuning state and verification refresh ([PR\\#6100](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6100?ref=relnotes), [PR\\#6130](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6130?ref=relnotes), [PR\\#6220](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6220?ref=relnotes))  \r\n* **BAGEL-7B-MoT** — T2I / editing / VLM WebDataset conversion, BAGEL MIMO construction, flow-matching schedule, H100 pretrain and finetune recipes ([PR\\#5605](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5605?ref=relnotes), community @ympcMark)\r\n\r\n#### **Diffusion**\r\n\r\n* **DiffusionGemma** — bridge ([PR\\#6284](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6284?ref=relnotes), community @BenDGeist)\r\n\r\n### **Training & Functionality**\r\n\r\n* **Multi-LoRA and MoE PEFT**  \r\n  * Multi-LoRA for Megatron models ([PR\\#4218](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/4218?ref=relnotes), community @mathewjhan), with seeded slot re-initialization ([PR\\#6068](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6068?ref=relnotes), community @yushengsu-thu)  \r\n  * LoRA on grouped MoE expert linears ([PR\\#5155](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5155?ref=relnotes), community @yushengsu-thu), fused expert LoRA output merge ([PR\\#6041](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6041?ref=relnotes), community @j316chuck)  \r\n  * LoRALinear wrap of the `nn.Linear` LoRA path ([PR\\#5310](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5310?ref=relnotes), community @HollowMan6), merge-before-quantize ([PR\\#5384](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5384?ref=relnotes), community @HollowMan6), `expand_shared_outer` export ([PR\\#5588](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5588?ref=relnotes), community @HollowMan6), LoRA+ LR split ([PR\\#5589](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5589?ref=relnotes), community @HollowMan6), 3D-MoE LoRA export ([PR\\#5939](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5939?ref=relnotes), community @HollowMan6)  \r\n  * LoRA branch input no longer pre-quantized ([PR\\#6342](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6342?ref=relnotes))  \r\n* **SFT data: in-batch packing and multi-source blending** — GPT SFT in-batch packing ([PR\\#5493](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5493?ref=relnotes)), native Energon packing for Qwen3-VL ([PR\\#5285](https://github.com/NVIDIA-NeMo/Megatron-Bridge/pull/5285?ref=relnotes)), direct-HF and JSONL blending ([PR\\#5438](https://gith","is_prerelease":false},"readme":true,"readme_zh":false,"news":[],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}