Parameter-Efficient Fine-Tuning Steps

This section documents the parameter-efficient fine-tuning (PEFT) steps registered under src/nemotron/steps/peft/. Each step trains a low-rank adaptation (LoRA) adapter on top of a base model and produces a checkpoint_lora artifact. Merge that artifact with the base model by using the convert/merge_lora step when you need a standalone Hugging Face (HF) checkpoint for deployment.

Steps

peft/automodel

LoRA tuning with the NeMo AutoModel library against Hugging Face base models and JSON Lines chat datasets.

reference

peft/automodel

peft/megatron_bridge

LoRA tuning on top of NVIDIA Megatron-Bridge for distributed training of the Nemotron model family.

reference

peft/megatron_bridge

Related Documentation

• Nemotron Steps CLI Reference covers the shared option set, dotlist overrides, and passthrough arguments.

• Choose a PEFT Backend compares peft/automodel and peft/megatron_bridge.

• Configuration Conventions describes the per-step config/ layout.