Megatron Bridge — 26.08 Release Plan (Public)

Code freeze: ~Aug 1, 2026

This is a community-facing snapshot of the work targeted for the 26.08 release. Only items with a tracking PR or issue in NVIDIA-NeMo/Megatron-Bridge are listed. Larger multi-PR efforts (DeepSeek V4, MegatronMIMO, the MCore HybridModel migration, etc.) are tracked in their own issues — progress is updated there, and those issues are linked below rather than duplicated here.

1. Model Support

LLM

VLM / Multimodal

Diffusion / Masked-Diffusion

Tracked separately (progress updated in the linked issue)

• DeepSeek V4 — enhanced support & verification — issue PoR: Deepseek v4 Roadmap #4633
• Model framework: migrate to MCore HybridModel (cross-model refactor) — issue PoR: migrate to MCore HybridModel #4510
• Model support verification levels (full-model conversion + inference validation) — issue POR: Track model support verification levels #4509

2. Training Features

Merged

• Unified dataset configuration, processing, collation & packing architecture — shared Config + Builder paths across text and multimodal datasets, common masking/padding/packing contracts, model-owned VLM collators, direct THD packing — PR refactor(data): adopt config and builder dataset architecture #4685, PR refactor(data)!: unify multimodal dataset builders #4774, PR [data] fix: isolate and unify Nemotron Omni collation #4776
• Masked-LM HF-pretrained wrapper support — AutoBridge selects PreTrainedMaskedLM for *ForMaskedLM architectures — PR feat(model): add masked lm support. #4735 (community @bbuschkaemper)
• Checkpoint, resume & export usability — save HF weights alongside native Megatron checkpoints (PR [training, ckpt] feat: Save HF weights alongside native Megatron checkpoints #3825, community @dancingpipi); Energon dataloader save/resume (PR [ckpt, data] feat: complete Energon dataloader checkpointing save/res… #4521, community @zsombor-kaiko); distributed-optimizer resume memory fix (PR fix(ckpt): reduce distributed optimizer resume memory #4936)
• PEFT / LoRA scaling — distributed adapter export (PR feat(conversion): support distributed adapter export #4221); EP-reshardable per-expert LoRA (PR Changes to support EP reshardable per-expert LoRA #3951); sequence-parallel LoRA activation-memory reduction (PR [PEFT] feat: reduce LoRA sequence-parallel activation memory #4581); shared outer LoRA (PR Add shared outer lora #3958, community @nanjiangwill)
• Fine-tuning & evaluation usability — epoch-based training duration (PR feat(training): support epoch-based training duration #4147, community @muyihao); averaged eval losses + pure-eval example (PR feat(training): return eval losses and add pure eval example #4125)
• Quantized HF export foundations (ModelOpt) — NVFP4 HF export path (PR Add ModelOpt NVFP4 HF weight export support #4102, community @HollowMan6); W4A16/W4A4 NVFP4 grouped-MoE export (PR support modelopt W4A16 & W4A4 NVFP4 export for grouped MoE weights #4566, community @HollowMan6); KV-cache amax mapping (PR feat(quant): Add modelopt KV cache amax mapping. #4591)
• VLM data pipeline refactor (processing / batching / packing / masking) — issue PoR: [VLM] Refactor: clean up VLM data pipeline (processing / batching / packing / masking) #4041

Tracked separately (progress updated in the linked issue)

• MegatronMIMO 2026 Q3–Q4 roadmap (DistTrain + MIMO extensions) — issue PoR: MegatronMIMO 2026 Q3-Q4 Roadmap #4392
• Dynamic (Hybrid) Context Parallelism — issue [support] Dynamic (Hybrid) Context Parallelism: enablement status and end-to-end verification in Megatron-Bridge #4586
• Training-loop upstreaming to Megatron Core — issue PoR: align Megatron-LM and Megatron-Bridge checkpoints #4595
• SFT dataset unification / upstreaming with Megatron Core — issue [tracking] SFT dataset unification / upstreaming with Megatron Core #4596
• Merge MLM dev-branch MoE recipes into Bridge recipes — issue PoR: Merge MLM dev branch MoE recipes into Bridge training recipes #4471

3. Performance & Recipes

Merged (foundation)

• Unified flat performance recipes for all model families — PR [recipe][perf] feat: unified flat performance recipes for all model families #2803; H100 library recipe namespace (PR refactor(recipes): add h100 recipe namespace #4622); legacy perf-config removal (PR refactor(performance): remove legacy perf configs #4623); recipe environment-variable support (PR feat(recipes): support recipe environment variables #4682); perf recipes routed by name through train.sh --recipe (PR feat(training): route performance recipes by name #4945)

Tracked separately (progress updated in the linked issue)

• Merge performance recipes and model recipes with convergence validation — issue PoR: Merge performance recipes and model recipes with convergence validation #4582
• Megatron FSDP in Qwen3-VL 30B-A3B — issue PoR: Megatron FSDP in Qwen3-VL 30B-A3B #4514

4. UX & Infrastructure

Merged

• Unified training & checkpoint-conversion CLI entry points — supported scripts/training/train.sh launcher (PR feat(training): add shared recipe launcher #4856, PR feat(training): route performance recipes by name #4945); scripts/conversion/convert.sh and distributed roundtrip (PR feat(conversion): add shared checkpoint launcher #4873, PR feat(ckpt): add roundtrip conversion launcher #4886)

5. Examples & Tutorials

Merged

• Nemotron 3 Ultra recipes and examples — conversion, inference, DCLM pretraining, packed OpenMath SFT/PEFT, packing scripts/docs — PR feat(nemotronh): add Nemotron 3 Ultra recipes and examples #4224
• Qwen3.5-VL MegatronMIMO non-colocated examples — conversion, HF-data SFT, Slurm launch, tutorial docs — PR Add Qwen3.5-VL MegatronMIMO examples #4071, PR Add Qwen3.5-VL MegatronMIMO non-colocated SFT tutorial #4239
• MoE recipe notebook — PR [tutorial] feat: add MoE notebook example #3890 (community @karinseve)

Tracked separately (progress updated in the linked issue)

• VLM dataset benchmark / support (hi-res & interleaved multi-image) — issue [data] tracking: VLM SFT benchmark — high-res & interleaved multi-image dataset support #3133
• Extended Megatron FSDP example: verify combinations & document caveats — issue [26.08] Extended Megatron FSDP example: verify combinations & document caveats #4246