xlm.commands.push_to_hub

PushToHubConfig dataclass

Configuration for pushing model to Hugging Face Hub.

instantiate_model(cfg, datamodule, tokenizer)

Instantiate a model from checkpoint for pushing to Hub.

Uses Harness.from_checkpoint(apply_ema=True) when a full Lightning checkpoint is available. This applies EMA weights AFTER load_state_dict completes, avoiding a Lightning quirk where on_load_checkpoint modifications to the state_dict are overwritten by the subsequent load_state_dict call.

Falls back to load_model_for_inference for model-only checkpoints (which have no EMA state to apply).

Parameters:

Name	Type	Description	Default
cfg	DictConfig	Hydra config	required
datamodule	Any	Datamodule instance	required
tokenizer	Any	Tokenizer instance	required

Returns:

Name	Type	Description
Harness	Harness	The instantiated model ready to push to Hub

main(cfg)

Main function for Push to Hub.