TrainerConfig¶

Abstract

This class cannot be instantiated directly. Use one of the variants listed below.

Module: fast_llm.engine.training.config

Variant of: RunnableConfig — select with type: train

Inherits from: PretrainedFastLLMModelConfig, ExperimentConfig, RunnableConfig

Fields¶

data — core

Type: DataConfig Default: (sub-fields optional)

Configuration for the dataset and model-independent preprocessing.

model — core

Type: FastLLMModelConfig Default: (sub-fields optional)

Configuration for the Fast-LLM model.

optimizer — core

Type: OptimizerConfig Default: (sub-fields optional)

Configuration for the training optimizer and learning rate schedule.

run — core

Type: RunConfig Default: (sub-fields optional)

Global properties for the experiment.

schedule — core

Type: ScheduleConfig Default: (sub-fields optional)

Configuration for the scheduling of each iteration.

training — core

Type: TrainingConfig Default: (sub-fields optional)

Configuration for the training phases and global properties.

callbacks — feature

Type: dict[str, TrainerCallbackConfig] Default: dict()

Configuration for training callbacks.

pretrained — feature

Type: CheckpointLoadConfig Default: (sub-fields optional)

Configuration for loading the configuration and state of a pretrained model.

reference_models — feature

Type: dict[str, PretrainedFastLLMModelConfig] Default: dict()

Auxiliary models used during training, ex. for knowledge distillation.

profiling — logging

Type: ProfilingConfig Default: (sub-fields optional)

Configuration for the optional profiling of GPU and CPU CUDA operations.

Select a variant by setting type: to one of the following values.

`type` value	Class	Description
`gpt`	GPTTrainerConfig	Top-level configuration for training a GPT model. Entry point for `fast-llm train gpt`
`multimodal`	MultiModalTrainerConfig