What is Gradient checkpointing?
Gradient checkpointing is a training-stage concept in modern language models. It is part of the vocabulary teams use when they compare models, design prompts or review an AI system.
Definition
Gradient checkpointing is a training-stage concept in modern language models. It is part of the vocabulary teams use when they compare models, design prompts or review an AI system. The term is used in training discussions, where precision matters: two teams can use the same word to mean slightly different things. In this glossary, gradient checkpointing is treated as a practical concept rather than a formal specification.
Why it matters
Training terms cover how models learn from data, how preferences are shaped and how behaviour is adjusted after pre-training. They matter because most capability and safety differences are set during training, not at inference time.
Example
Suppose you are comparing Nemotron Nano 7B with another model for a production task. Gradient checkpointing is one of the factors behind the numbers you see: Nemotron Nano 7B lists a 8,192-token context window, costs $0.35 per million input tokens and scores 31.3 on AIME 2025. Understanding the term helps you decide which of those differences actually matter for your workload.
How to evaluate it
- Check whether gradient checkpointing is defined the same way in the documentation you are reading.
- Ask which measurement or trade-off the term is describing before comparing numbers.
- Separate vendor claims from independently observed behaviour on your own data.
- Write down the definition your team uses so reviews and evaluations stay consistent.
Related terms
Related guides
FAQ
What does Gradient checkpointing mean?
Gradient checkpointing is a training-stage concept in modern language models. It is part of the vocabulary teams use when they compare models, design prompts or review an AI system.
Why does Gradient checkpointing matter?
Training terms cover how models learn from data, how preferences are shaped and how behaviour is adjusted after pre-training. They matter because most capability and safety differences are set during training, not at inference time.
How is Gradient checkpointing used in practice?
Suppose you are comparing Nemotron Nano 7B with another model for a production task. Gradient checkpointing is one of the factors behind the numbers you see: Nemotron Nano 7B lists a 8,192-token context window, costs $0.35 per million input tokens and scores 31.3 on AIME 2025. Understanding the term helps you decide which of those differences actually matter for your workload.