Support Megatron training backend for MoE training

[tyler-griggs]

To support large-scale multi-node MoE training, we will integrate the Megatron training backend.

Tasks:

• Initial Megatron Support for GRPO on TP + PP for dense models: [trainer] Initial Megatron TP + PP Support #223
• Sequence packing + context parallel support: [trainer][megatron] Sequence packing + Context Parallel for Megatron #274
• Checkpointing support [Megatron] Add checkpointing support #298
• EP support for MoE models [trainer][megatron] Enable expert model and expert tensor parallel for MoE models #285
• migration to Megatron-Bridge [megatron] upgrade from mbridge -> Megatron-Bridge (breaking change) #453
• LoRA support via Megatron-Bridge
• unify gradient checkpointing flag
• Add guide for debugging and getting around OOM
• enable megatron fsdp
• Disaggregated Training
• PPO support (just need to add Critic and test)
• Add megatron optimizer/scheduler pass through config options
• Virtual Pipeline Parallel Support
• Dynamic Batch sizing
• Optimizing for efficiency (optimized kernels/bucketing weight updates) [megatron] Added non cuda ipc wt sync to megatron workers #635
• Support FlashRL + Megatron integration
• Documentation

[erictang000]

• Megatron DP Convergence (see [trainer] Initial Megatron TP + PP Support #223 (comment))

[benrhodes26]

Is this checklist up-to-date?