Glossary: Fine-tuning LLMs

Gradient checkpointing

Also written: gradient_checkpointing

Recomputing activations in the backward pass instead of storing them, saving memory at the cost of speed. The small SmolLM2 model doesn't need it.