bridge.training.dpo_train

DPO training entry point: run-config guardrails, artifact validation, loop hand-off.

Module Contents

Functions

dpo_train	Run DPO training over a scored preference dataset.
validate_dpo_run_config	Fail fast if the run has no usable reference artifact for each split it trains or evaluates on.
_validate_artifact
_warn_on_expert_layout_mismatch	Warn, never fail, when the trainer’s expert layout (EP, ETP) differs from the scoring run’s.
expected_scoring_metadata	The scoring fingerprint this run’s config implies for split, compared against the artifact’s.

Data

logger

API

bridge.training.dpo_train.logger

‘getLogger(…)’

bridge.training.dpo_train.dpo_train(

config: megatron.bridge.training.config.ConfigContainer,

forward_step_func: megatron.bridge.training.forward_step_func_types.ForwardStepCallable,

callbacks: list[megatron.bridge.training.callbacks.Callback] | megatron.bridge.training.callbacks.CallbackManager | None = None,

) → None

Run DPO training over a scored preference dataset.

bridge.training.dpo_train.validate_dpo_run_config(

config: megatron.bridge.training.config.ConfigContainer,

) → None

Fail fast if the run has no usable reference artifact for each split it trains or evaluates on.

The loss contract itself (per-token loss, no collective averaging, CP=1, no MTP, even row counts) is enforced by ConfigContainer.validate for any DPO dataset.

bridge.training.dpo_train._validate_artifact(

config: megatron.bridge.training.config.ConfigContainer,

artifact_dir: str,

split: Literal[train, validation],

) → None

bridge.training.dpo_train._warn_on_expert_layout_mismatch(

config: megatron.bridge.training.config.ConfigContainer,

artifact_dir: str,

metadata: dict[str, Any],

) → None

Warn, never fail, when the trainer’s expert layout (EP, ETP) differs from the scoring run’s.

bridge.training.dpo_train.expected_scoring_metadata(

config: megatron.bridge.training.config.ConfigContainer,

split: Literal[train, validation] = 'train',

) → megatron.bridge.data.datasets.preference.ScoringFingerprint

The scoring fingerprint this run’s config implies for split, compared against the artifact’s.