RunConfig¶

Module: fast_llm.engine.config_utils.run

Fields¶

experiment_dir — core

Type: Path or None Default: None

Directory where every checkpoint, artifact, etc., will be saved.

experiment_name — feature

Type: str or None Default: None

A custom name for the experiment. Default: the experiment directory name or 'default'

triton_linear_kernels — performance

Type: bool Default: False

Global switch to use triton kernels for linear layers. These may be slightly slower than the defaults.

enable_all_loggers — logging

Type: bool Default: False

Enable all existing loggers, including those external to Fast-LLM, by setting their level to info.

log_timestamps — logging

Type: bool Default: True

Add a timestamp to every Fast-LLM (structured) log.

model_debug_level — logging

Type: int Default: 0

Debugging level for the model, ex. for printing intermediate model states.

structured_logs — logging

Type: bool Default: True

Configure logging to the Fast-LLM format.

tensor_logs — logging

Type: TensorLogsConfig Default: (sub-fields optional)

Configuration for debug tensor logs.

enable_triton_kernels — expert

Type: bool Default: True

Global switch to allow disabling triton kernels. This parameter may be ignored when no alternative is available.

torch_dynamo_enable — expert

Type: bool Default: True

Set to False to disable torch compile entirely. Not recommended unless there is a good reason to do so.