TrainResult

TrainResult

from modal_training_gym.common.train_result import TrainResult

One completed training run’s checkpoint handle.

Constructor

TrainResult(app_name, framework, run_id, checkpoint_dir, base_model, model_class='', checkpoints_volume_name='', checkpoints_mount_path='', iteration_prefix='', wandb_project='', wandb_entity='', wandb_run_id='', extra=<factory>)

Parameter	Type	Default	Description
app_name	str	required
framework	str	required
run_id	str	required
checkpoint_dir	str	required
base_model	str	required
model_class	str	""
checkpoints_volume_name	str	""
checkpoints_mount_path	str	""
iteration_prefix	str	""
wandb_project	str	""
wandb_entity	str	""
wandb_run_id	str	""
extra	dict[str, Any]	<factory>

Attributes

Attribute	Type	Default	Description
app_name	str
framework	str
run_id	str
checkpoint_dir	str
base_model	str
model_class	str	""
checkpoints_volume_name	str	""
checkpoints_mount_path	str	""
iteration_prefix	str	""
wandb_project	str	""
wandb_entity	str	""
wandb_run_id	str	""
extra	dict[str, Any]	{}

Methods

build_serve_app(self, *, served_model_name: 'str | None' = None, checkpoint_path: 'str | None' = None, **vllm_kwargs: 'Any') -> "'App'"

Build a vLLM serving app pointing at a trained checkpoint.

dashboard_url(self) -> 'str'

URL for browsing the checkpoints volume in the Modal dashboard.

latest_checkpoint_path(self) -> 'str'

Absolute in-volume path of the latest checkpoint.

list_checkpoints(self) -> 'list[str]'

Return per-iteration checkpoint directory names under

list_runs(app_name: 'str') -> 'list[str]'

Return all run_ids saved for app_name, sorted oldest

load(app_name: 'str', run_id: 'str | None' = None) -> "'TrainResult'"

Load a completed run’s result from the shared store.

save(self) -> 'None'

Persist this result to the shared :class:modal.Dict.

volume(self) -> "'Volume'"

Return a handle to the checkpoints :class:modal.Volume.

wandb_metrics(self, keys: 'list[str] | None' = None, samples: 'int' = 500) -> 'list[dict[str, Any]]'

Fetch training metrics from W&B.

wandb_summary(self) -> 'dict[str, Any]'

Fetch the W&B run summary (final metric values).

wandb_url(self) -> 'str | None'

Return the W&B run URL, or None if W&B info is not set.

Source: modal_training_gym/common/train_result.py