[ROADMAP][Updated on Jan 26] Megatron Core MoE Roadmap

[yanring]

Description

The focus for Megatron Core MoE is to provide comprehensive support for latest MoE architectures, advanced parallelism strategies, and performance optimizations for Blackwell. This is a tentative roadmap and subject to change.

🎉 This Roadmap is based on the dev branch; please see the details in its README.

---

Model Support

• ✅ DeepSeek
  • ✅ DeepSeek-V2
  • ✅ DeepSeek-V3, including MTP
  • ✅ DeepSeek-V3.2 [dev] DeepSeek V3.2 support #2154
• ✅ Qwen
  • ✅ Qwen2-57B-A14B
  • ✅ Qwen3-235B-A22B
  • ✅ (🚀New!) Qwen3-Next
• ✅ Mixtral

Core MoE Functionality

• ✅ Token dropless MoE - Advanced routing without token dropping
• ✅ Top-K Router with flexible K selection
• ✅ Load balancing losses for expert load balancing optimization

Advanced Parallelism

• ✅ Expert Parallel (EP) with 3D parallelism integration
• ✅ Full parallelism combo: EP + DP + TP + PP + SP support
• ✅ Context Parallel (CP) for long sequence MoE training
• ✅ Parallel Folding Heterogeneous Parallelism Mappings for Efficient Large-Scale MoE Model Training
• ✅ Distributed Optimizer for MoE (ZeRO-1 equivalent)
• ✅ (🚀New!) Megatron FSDP/HSDP with full expert parallel support

Optimizations

• ✅ Memory Efficient token permutation
• ✅ Fine-grained Recomputations (mla, moe, mlp, moe_act, norm)
• ✅ GroupedGEMM and Gradient Accumulation Fusion
• ✅ DP/PP/TP/EP Communication Overlapping
• ✅ Advanced fusions for Router; Permutation; MLA Rope; FP8 casting, etc
• ✅ cuDNN fused Attention and FlashAttn integration
• ✅ (🚀New!) 1F1B EP A2A Overlap - Hiding Expert Parallel Communication with 1F1B Pipeline Schedule
• ✅ (🚀New!) Muon and Layer-wise distributed optimizer
• ✅ (🚀New!) Pipeline-aware fine-grained activation offloading [Dev] feat(moe): Fine-grained activation offloading #1912
• ✅ (🚀New!) Production-ready cudaGraph support for MoE

Precision Support

• ✅ GroupedGEMM including FP8/MXFP8 support
• ✅ FP8 weights with BF16 optimizer states
• ✅ FP8 training full support

Optimized Expert Parallel Communication Support

• ✅ DeepEP support for H100 and B200
• ✅ (🚀New!) HybridEP for GB200

Developer Experience

• ✅ MoE Model Zoo with pre-training best practices
• ✅ MCore2HF Converter for ecosystem compatibility in megatron-bridge
• ✅ Distributed Checkpointing Support
• ✅ Runtime Upcycling Support for efficient model scaling
• ✅ Layer-wise logging for detailed monitoring

---

Next Release Roadmap (MCore v0.17)

Performance & Kernel Optimizations

• Split-K Indexer Kernels - Avoid materializing [seqlen_q, seqlen_k] tensor with split-K kernels [WIP Feat] Split-K Indexer Kernels #2869 (draft)
• Absorbed MLA - MLA computation optimization for DSA Add absorbed-mla & fused dsa #3044
• A2A Overlap Refinement - Refine A2A overlap under CUDA_DEVICE_MAX_CONNECTIONS=1 [Dev](moe):Refine A2A overlap under CUDA_DEVICE_MAX_CONNECTIONS=1 #2730
• HybridEP preprocess optimization

Long Context & Context Parallel

• Hybrid CP Part 2 - Enhanced hybrid data x context parallelism [Dev] feat: Dynamic CP (part 2) #2000
• THD Format E2E Support - End-to-end THD format support Add E2E support for THD format #2924 (draft)

Model & Architecture

• Manifold Hyper Connection (mHC) - New architecture feature for improved loss convergence [dev] feat(mHC): Add basic pytorch implementation of manifold hyper connection(mHC). #2943 (draft)
• GDN THD Support - Packed sequence support for gated delta net [dev] feat(moe): Support packed sequence for gated delta net (GDN) #2644
• GDN Refinement - Refine gated delta net implementation [dev] perf(moe): Refine gated delta net implementation #3040

Advanced Functionality

• Router replay support for RL training (in progress) feat: add routing replay for Mcore #2693
• Megatron FSDP performance optimization for MoE training

CUDA Graph Enhancements

• MoE ECHO - Elastic Cloning for sync-free, full CUDA-graph dropless MoE training MoE ECHO: Unlocking Sync-Free, Full CUDA-Graph Support for Dropless MoE via Elastic Cloning #2368 (draft)
• Paged Stashing - Dynamic tensor support for dropless MoE with CUDA graph Paged Stashing #2690 (draft)
• CUDA Graph + Offloading - Support CUDA Graph capture with offloading modules Support CUDA Graph capture offloading modules #2437 (draft)
• Optimizer CUDA Graph - Enable CUDA graph for ADAM optimizer Enable Optimizer CUDA graph for ADAM optimizer #2931 (draft)

Ongoing Long-term Features

• E2E Performance optimization for DeepSeek-V3, Qwen-3 and other fine-grained MoEs
• Sync-Free and Full-Iter cudaGraph MoE Training
  • MoE ECHO for dropless MoE load balancing MoE ECHO: Unlocking Sync-Free, Full CUDA-Graph Support for Dropless MoE via Elastic Cloning #2368
  • Paged Stashing for dynamic tensor support Paged Stashing #2690
• CPU Overhead Optimizations for Blackwell Performance
• MLA Optimizations
  • Absorbed MLA Add absorbed-mla & fused dsa #3044
  • MLA CP 2.0 - MLA CP Enhancement for Longer Sequence Training
• THD and Long Context
  • THD Format E2E Support - End-to-end THD format support Add E2E support for THD format #2924
  • Dynamic Context Parallel for Imbalanced Long-Sequence Training
    • [Dev] Hybrid Data x Context Parallelism Feature #2054
    • [Dev] feat: Dynamic CP (part 2) #2000
• Megatron FSDP Performance Optimization for MoE Training
• Kernel fusions and optimizations for MoE models from TE MoE training optimization TransformerEngine#2438
• New Architecture Support
  • Manifold Hyper Connection (mHC) [dev] feat(mHC): Add basic pytorch implementation of manifold hyper connection(mHC). #2943

---

v0.16 Update Highlights

Performance & Memory

• 🚀 Fused Linear and Cross Entropy - Fuse lm_head and CE to avoid materializing intermediate logits, reducing memory [Dev] Feature: linear cross entropy fusion #2256
• 🚀 Optimizer State Offloading - Offload optimizer states and master weights to CPU for significant GPU memory savings [Dev] [Reapply] Optimizer State and Master Weight Offloading #2987
• 🚀 MTP Standalone Stages - Support placing MTP layers into standalone pipeline stages for better VPP balance [Dev] feat(moe): Support placing MTP layers into standalone stages #1916
• 🚀 DeepSeek V3.2 Support - Performance optimizations for DeepSeek V3.2 [dev] DeepSeek V3.2 support #2154
• DeepSeek V3 Pre-training Performance Guide (~960 TFLOPS/GPU on 256 GB200s) [Dev] A Guide to Reproduce DeepSeek-V3 Pre-training Performance on GB200 #1996

CUDA Graph

• 🚀 cuda_graph_scope Refactoring [Dev] feat(MoE): Refactor cuda_graph_scope - part2 #2353 [Dev] TE cudagraph recompute #2694
• 🚀 Partial CUDA Graph for EP Overlap - Release CPU pressure within selected scope for EP A2A overlap [Dev](Reapply) Partial CUDA Graph support for EP Overlap #2810
• TE cudagraph input memory optimization - Reuse static input memory buffer among microbatches [Dev] Optimize TE cudagraph input memory #2391
• CudaGraph compatibility with FP8 params (tensorwise & blockwise) [DEV] Make CUDA graph compatible with FP8 params (tensorwise & blockwise). #2087
• NVFP4 MOE CUDA Graph support with 128 zero padding [DEV][NVFP4][MOE] 128 Zero Padding for Grouped Quantization kernels and Cuda Graph Support  #2654

Model & Parallelism

• 🚀 Qwen3-Next Enhancements - QK layernorm weight decay support and Gated Delta Net CP for long context [dev] feat(moe): Support apply wd to qk layernorm for Qwen3-Next #2825 [Dev] Feat(moe): Gated delta net context parallel (CP) #2614
• 🚀 Hybrid Data x Context Parallelism - New parallelism strategy combining DP and CP [Dev] Hybrid Data x Context Parallelism Feature #2054
• 🚀 Router Replay - Deterministic routing mechanism for debugging and RL training feat: add routing replay for Mcore #2693
• 🚀 Fake Distributed Process Group - Skip all distributed comm ops with --fake-process-group for profiling [DEV] Add support of fake distributed process group #2254
• Remove padding token in MoE routing loss - Improve aux loss correctness and efficiency [Dev] Remove calculation of padding token in moe routing loss #2121
• Context parallel support for eager attention implementation [Community][Dev] feat(moe): Adding context parallel support to eager attention implementation #1859
• Packed sequence support in MTP module [Dev] Support packed seq in MTP #2043

Fine-grained Activation Offloading Enhancement

• 🚀 OOP Refactoring - Object-oriented redesign for fine-grained activation offloading [Dev]feat(moe): code refactor for fine grained activation offloading #2905
• Fix accuracy mismatch when offloading and recomputing same module [Dev] fix(offloading): Accuracy mismatch when offloading and recomputing same module #2123
• Bug fix for fine-grained activation offloading in evaluate() [Dev] [fix] Bug fix for fine-grained activation offloading in evaluate() #3041

Megatron-FSDP

• 🚀 FP8 Params Support - MXFP8/Blockwise FP8 params for Megatron-FSDP [dev] Reapply fsdp mxfp8 #2828
• 🚀 HSDP Support - Hybrid Sharded Data Parallel with EP submesh registration Fix HSDP Registering Device Mesh #2388
• Megatron-FSDP user guide documentation [Dev] docs(megatron-fsdp): add Megatron-FSDP user guide #2397

Communication

• 🚀 Hybrid-EP Upgrade - Latest Hybrid-EP with kernel optimizations for EP64 and NVL8+IB [Dev] Use the latest Hybrid-EP #2424
• HybridEP memory overhead reduction for 1F1B A2A overlap [Dev] fix(moe): Support HybridEP and reduce memory overhead for 1F1B A2A overlap #2201

Optimizer

• 🚀 LayerWise DistOpt - LayerWiseDistributedOptimizer with torch_dist checkpoint format and Muon support [Dev] Support LayerWiseDistributedOptimizer with torch_dist checkpoint format #1928 [DEV] Update emerging optimizers #2261

Critical Bug Fixes

• Megatron-FSDP hang fix - Resolve hang caused by non-deterministic reduce-scatter [Dev] fix(megatron-fsdp): Resolve hang caused by non-deterministic reduce-scatter #2252
• EP Overlap correctness - Fix missing final layernorm in EP overlap [Dev] Fix ep overlap missing final layernorm #2691
• Hybrid-EP hotfix - Fix bug of hybrid-ep backend in flex-dispatcher [DEV] [HOT FIX] Fix bug of hybrid-ep backend in flex-dispatcher #2287
• CUDA RNG Tracker - Fix RNG tracker to use expert-parallel-rng correctly [Dev] Fix CUDA RNG Tracker #2640

---

Call for Community Contributions

• Model implementations - Additional MoE model variants
• Performance testing - Performance tests across different platforms and workloads
• Documentation and tutorials - Best practices and optimization guides
• Bug fixes

---

This roadmap reflects the collective efforts of NVIDIA and our collaborators

Credits: MCore MoE Team and @sbhavani

Labels: roadmap, moe, call-for-contribution

[Skylion007]

No upstreaming of distributed MuON on the roadmap for Q3/Q4? Open to external contribution/PR?

[Skylion007]

Also FSDP2 has EP support now in PyTorch nightly and is used for such in TorchTitan. We probably should integrate that here too.

Custom FSDP - NVIDIA-optimized FSDP implementation with full expert parallel support

[ericharper]

No upstreaming of distributed MuON on the roadmap for Q3/Q4? Open to external contribution/PR?

This roadmap was for MoE only, we're planning on adding a more general roadmap for all of Core soon.

[jeromeku]

@yanring

Very interested in learning more about the following:

• MCore2HF Converter
• Sync-Free MoE

Where can I find more info on these?

Thanks!

[switiz]

No upstreaming of distributed MuON on the roadmap for Q3/Q4? Open to external contribution/PR?

This roadmap was for MoE only, we're planning on adding a more general roadmap for all of Core soon.

Please consider distributed Muon support as a priority.
We are currently evaluating its performance across several open-source vendors such as Kimi and Zipu.

[sbhavani]

@switiz We are prioritizing Muon support and will share some of our early results soon

[yanring]

Hi @jeromeku, Thanks for reaching out. Some updates on your questions:

• MCore2HF Converter: Please check https://github.com/NVIDIA-NeMo/Megatron-Bridge
• Sync-Free MoE: A significant challenge with MoE is that dynamic shapes lead to D2H transfers and synchronization, which in turn creates considerable CPU overhead. We are currently developing a lossless solution to address this. While we can't disclose the specifics just yet, we hope to share more updates with you soon.

[yucc-leon]

Hi, what is the "mixture of linear attention"? MLA is famous for a variant of multi-head attention, and I can't find this "MLA" elsewhere using search engine...

[sbhavani]

@yucc-leon that's a typo, should be multi-head latent attention

[JazzeYoung]

Hi, what is the "mixture of linear attention"? MLA is famous for a variant of multi-head attention, and I can't find this "MLA" elsewhere using search engine...

Is this Multi-Latent Attention?