Skip to content

Engine / Multi Stage

Classes

Class Description
CheckpointMetadata
FastLLMModelConfig (abstract) Abstract base configuration for a Fast-LLM model: base model, multi-stage layout, and distributed config
MultiStageConfig Configuration for the multi-stage model layout: layers per stage, ZeRO sharding, and buffer counts
PretrainedFastLLMModelConfig (abstract) Configuration wrapper that optionally loads model weights and config from a pretrained checkpoint
StageConfig Configuration for a single model stage: gradient precision, frozen weight storage, and debug logging