← All questions
Training & Fine-tuningMedium· ZeRO· FSDP· Communication Overhead

What is the communication cost of ZeRO stage 3 compared...

The communication cost difference between ZeRO stage 3 and stage 1 is defined by the parameter sharding strategy. ZeRO stage 3 requires an extra P of communication because it performs a per-layer parameter AllGather during the forward pass to materialize the weights needed for computation. This represents a 1.5x increase in communication volume, which is the trade-off required to achieve N times parameter memory savings by sharding parameters across all ranks.

Code playground (Python)