Megatron Bridge — 26.06 Release Plan (Public)

Code freeze: June 1, 2026

This is a community-facing snapshot of the work targeted for the 26.06 release. Only items with a tracking PR or issue in NVIDIA-NeMo/Megatron-Bridge are listed.

1. Model Support

LLM

VLM / Multimodal

2. Training Features

Work in progress

• ModelOpt: pure eval loop + return val LM loss from evaluate_and_print_results — issue Support pure eval loop given a model and ConfigContainer #2033
• Qwen3.5-VL MegatronMIMO example — PR Add Qwen3.5-VL MegatronMIMO examples #4071 (depends on merged MCore PR [mimo] Support bridge fan-out for variable modality tokens NVIDIA/Megatron-LM#5062)

Merged

• Energon v7 support — PR NVIDIA-NeMo/Megatron-Bridge#4090
• Low-precision bridge (FP8 / MXFP4) — PR [feat] Support quantization before weight resharding #2737 (community @hy2826), issue Need a FP8 (low precision) page in Doc #1147
• Eval-time context parallelism via decentralized PG rebinding — PR [training, perf] feat: Add eval-time context parallelism via decentralized PG rebinding #3755
• MegatronMIMO non-collocated training (phases 2–5) — PR NVIDIA-NeMo/Megatron-Bridge#2004, PR NVIDIA-NeMo/Megatron-Bridge#2007, PR NVIDIA-NeMo/Megatron-Bridge#2869, PR NVIDIA-NeMo/Megatron-Bridge#2870; rename PR NVIDIA-NeMo/Megatron-Bridge#3340
• Evaluator backend integration — PR NVIDIA-NeMo/Megatron-Bridge#2990
• DisTrain for Qwen3-VL — PR NVIDIA-NeMo/Megatron-Bridge#2367
• LoRA × ModelOpt × DeepSeek arch — PR NVIDIA-NeMo/Megatron-Bridge#3612
• MoE / MTP metrics → MLFlow & Comet — PR [training] feat: forward MoE/MTP metrics to MLFlow and Comet #3647 (community @lonexreb)
• PEFT target-module validation with deferred init — PR [peft] feat: validate PEFT target modules with deferred init #3587
• MLflow run lifecycle (FAILED on crash, KILLED on SIGTERM) — PR NVIDIA-NeMo/Megatron-Bridge#3822 (community @rob-luke)
• ModelOpt: save custom Megatron model back to HF format — PR NVIDIA-NeMo/Megatron-Bridge#2778 (community @ShiftyBlock); issue Support saving custom megatron model (e.g. pruned variant) back to Hugging Face format #2036

3. Performance

26.06-labeled tracking issues:

• enable 0SM PP — issue [feature] enable 0SM PP #3394
• enable full-iteration cudagraph — issue [feature] enable full-iteration cudagraph #3393
• FA4 bump — issue [feature] FA4 bump #3392

4. UX & Infrastructure

Merged

• Integrate Megatron Inference into Bridge — refactor example inference scripts to use the MCore Inference Engine; CLI for pure LLM inference; inference_optimized path — PR [Inference] Add MCore inference examples and model wrappers #3897
• Remove model-specific size providers — PR [models] refactor: Remove size-specific provider classes #3854
• RL API refactoring — PR NVIDIA-NeMo/Megatron-Bridge#3813
• Tokenizer unification MCore ↔ Bridge — MCore PR convert tokenizer args to config NVIDIA/Megatron-LM#4406, Bridge PR upstream MCore tokenizers config #3451
• Per-model docs pages (docs taxonomy refactor) — PR [docs] refactor: Organize model docs by family #3908

5. Examples & Tutorials

Work in progress

• Long-context extension example — 128k target (Qwen3-30B P0, Moonlight-16B P1); YaRN example with Qwen3 0.6B — issue CP + Seq packing in LLM & VLM #1926, PR NVIDIA-NeMo/Megatron-Bridge#3316
• VLM dataset benchmark / support — vision balancing, new data pipeline, Energon + hi-res / multi-image, stress tests; Qwen3-VL or Qwen3.5-VL; phase 1 merged (PR VLM Energon updates for videos, multiple images #3691) — issue [data] tracking: VLM SFT benchmark — high-res & interleaved multi-image dataset support #3133
• Megatron FSDP — tracking/benchmark matrix across model sizes (Llama 405B, DSv3), LLM/VLM, pretrain/SFT/LoRA/THD, RL; convergence vs 5D + speed — issue Tune best performance for FSDP on GB300 for LLAMA31 405B #2118, issue [bug] Torch FSDP2 does not work #2785

Merged

• HF ↔ Megatron-FSDP conversion bridge — PR NVIDIA-NeMo/Megatron-Bridge#2529, PR NVIDIA-NeMo/Megatron-Bridge#1910; MCore docs PR Unify and refactor Megatron-FSDP documentation. NVIDIA/Megatron-LM#4418
• Dynamic CP example — underlying feature enablement PR [training, perf] feat: Add eval-time context parallelism via decentralized PG rebinding #3755; cross-feature example PR [examples] Add dynamic context parallel example #3892