domains / families / data / nemo-mbridge-perf-sequence-packing

nemo-mbridge-perf-sequence-packing

separate

Validate and use packed sequences and long-context training in Megatron-Bridge, including offline LLM packing, collate-time VLM packing, Energon online packing, and CP constraints.

When to use

• offline LLM packing
• collate-time VLM packing
• Energon online packing
• CP constraints
• long-context training
• packed sequence validation

When not to use

• MTP
• CUDA graphs
• Qwen3-VL DistTrain
• PP
• global-batch "batch" dataloader

Skill page

Source