nemo-mbridge-perf-moe-vlm-training

FSDP is typically preferred for simpler setups, while 3D-parallel offers finer control over memory and topology for large-scale MoE VLM training. This Skill compares both paths to help you evaluate trade-offs and select the best strategy.

To configure CUDA graphs for MoE VLM training, you must apply specific tuning tips for memory management and recompute considerations. This Skill provides guidance on avoiding MoE-specific pitfalls and optimizing dispatcher choices for production-ready configurations.

Yes, Megatron Bridge supports Qwen3-VL style MoE VLM training on GB200 or B200 hardware. This Skill details the required topology layouts, dispatcher choices, and memory management configurations for these specific platforms.

Common MoE-specific pitfalls during 3D-parallel VLM training involve memory management overhead and dispatcher misconfigurations. This Skill specifies requirements for production-ready configurations to help you avoid these layout and memory bottlenecks.

You plan a two-path MoE VLM training workflow by executing and evaluating both FSDP-first and 3D-parallel configurations in Megatron Bridge. This Skill guides you through documenting their trade-offs to select the best parallel strategy.