Page cover image


Checkpoints in LLMs are snapshots of the model's state, including its weights and other relevant parameters, saved at a specific point during training. Checkpoints are essential for various reasons, such as monitoring progress, resuming training, and performing model evaluation.

Concept of Gradient Checkpointing

Gradient checkpointing is a technique to reduce the memory consumption during the training of deep neural networks. This is especially relevant in scenarios where a network has many layers and storing all the intermediate activations (gradients) would consume an excessive amount of memory.

How It Works

  • Storing Key Points: Rather than storing all the intermediate activations, gradient checkpointing stores only a subset of them. These stored points are known as "checkpoints."

  • Recomputing Gradients: During the backpropagation phase, the missing gradients between these checkpoints are recomputed on-the-fly. This trade-off between computation and memory usage enables the training of much deeper models within the same memory constraints.

  • Memory-Computation Trade-off: While it helps in saving memory, gradient checkpointing adds extra computation overhead, as some of the gradients must be recalculated.


Imagine a long and winding hiking trail, and you want to remember the entire journey without taking photos at every single point.

Gradient checkpointing is like choosing specific scenic points to photograph (checkpoints) and then relying on your memory (recomputation) for the details in between. This allows you to save space on your camera (memory) at the cost of a bit more mental effort (computation).

Key Concepts Around Checkpointing

Checkpointing is a critical feature in machine learning and deep learning, particularly in scenarios involving extensive training periods or the risk of interruptions. The key concepts extracted from the provided document, generalized beyond specific programming languages or libraries, are as follows:

Checkpointing Basics

  • Checkpointing involves saving the state of a model, typically including its parameters and the state of the optimizer, at specific training steps or intervals. This is crucial for recovering the training process in case of interruptions like system failures or preemptions.

Single-Host vs. Multi-Process Environments

  • In a single-host environment, each process saves its checkpoint independently. This is straightforward but not suitable for environments where multiple processes need to save data to a common directory.

  • In a multi-process or distributed environment, special considerations are needed to ensure that checkpoints are saved correctly. Typically, a designated process (e.g., process 0) handles the main checkpoint file and manages the removal of old files.

File Management and Cleanup

  • The system maintains a specific number of past checkpoint files, deleting older or worse-performing checkpoints to manage disk space.

  • Overwrite options allow for the replacement of existing checkpoints if a new checkpoint at the current or a later step is created.

Asynchronous Saving and Callbacks

  • Asynchronous mechanisms can be employed to save checkpoints without blocking the main training process. This is particularly useful in single-host environments to maintain efficiency.

  • Callbacks or waiting functions might be necessary to ensure the finalization of asynchronous saving operations.

Restoration of Checkpoints

  • The restoration process involves loading the latest or best checkpoint from a set of saved checkpoints. This is critical for resuming training or for evaluation purposes.

  • Checkpoints are sorted naturally, and the system retrieves the highest-valued file, which represents the latest state.

Partial Restoration and Compatibility

  • Systems may allow partial restoration or compatibility with various formats, which is essential when dealing with complex models or when transitioning between different systems or versions.

Future-Proofing and Upgradability

  • Anticipating future changes in checkpointing mechanisms and ensuring compatibility with upcoming systems or standards is important. This includes the ability to migrate to new checkpointing systems as they become available.

In summary, effective checkpointing strategies are essential for efficient and resilient training processes in machine learning.

They must cater to different environments (single-host or distributed), manage file storage efficiently, and provide flexibility for asynchronous operations and future upgrades. These strategies enable practitioners to safeguard their progress and quickly recover from interruptions, making them a cornerstone of robust machine learning pipelines.

Last updated


Continuum - Accelerated Artificial Intelligence

Continuum WebsiteAxolotl Platform

Copyright Continuum Labs - 2023