Scaling

Guides for running ESPnet3 at larger scale — across multiple nodes, over large datasets, and through batch inference pipelines.

Multi-node training

Configure trainer.num_nodes, strategy, and launcher settings for multi-node distributed training.

Large-scale data

Shard large datasets, tune the data pipeline, and avoid bottlenecks when training on hundreds of hours.

Dataset sharding

Implement ShardedDataset, understand the shard rotation formula, and wire up total_shards in YAML.

Inference at scale

Run batch inference with the provider/runner layer across a cluster or local multi-GPU setup.