AI Training Checkpoint Storage with Data-Type-Aware Caching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The existing technology for caching checkpoint files during AI model training is inefficient, leading to prolonged writing times and suspension of the training process due to high buffer management overheads and low demand for reading checkpoint files.

Innovation Solution

A system and method that identifies the type of data being stored, using a buffer mechanism for datasets and direct storage for checkpoint files in local hard disks, optimizing storage strategies to improve reading and writing efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If a buffer mechanism is used to cache datasets for frequent reading, then dataset reading efficiency is improved, but checkpoint file writing time increases due to buffer management overhead

Engineering Contradiction:
Improvedataset reading efficiencyVSAvoidcheckpoint file writing time
Core Design Contradiction:
SpeedVSLoss of time

Solution Approach 1:

The patent segments the storage system into two distinct paths: a buffer-based caching path for datasets and a direct storage path for checkpoint files. This segmentation allows each data type to be handled by the most appropriate storage mechanism, avoiding the overhead of buffer management for checkpoint files while maintaining efficient dataset reading through buffer caching.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary classification of incoming data to determine whether it is a dataset or a checkpoint file. Based on this classification, the appropriate storage path is selected in advance: datasets are routed to the buffer caching path for efficient reading, while checkpoint files are routed to the direct storage path for fast writing without buffer overhead.

Inventive Principle:
Principle #10Preliminary action

2Device complexity

If checkpoint files are stored using the same buffer mechanism as datasets, then storage management is simplified, but writing efficiency decreases due to unnecessary buffer management overhead

Engineering Contradiction:
Improvestorage management complexityVSAvoidcheckpoint writing efficiency
Core Design Contradiction:
Device complexityVSProductivity

Solution Approach 1:

The storage management system is segmented into two distinct handling paths based on data type. Checkpoint files follow a simplified direct storage path that bypasses buffer management overhead, while datasets follow a buffer-based path for efficient caching. This segmentation maintains operational simplicity while optimizing writing efficiency for checkpoint files.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different storage strategies are applied to different data types based on their specific requirements. Checkpoint files, which are written infrequently but require fast writing, receive direct storage treatment. Datasets, which are read frequently, receive buffer caching treatment. This local quality approach optimizes each data type's storage handling according to its access patterns.

Inventive Principle:
Principle #3Local quality

3Reliability

If the training process waits for checkpoint writing to complete, then data consistency is ensured, but training efficiency decreases due to frequent suspensions

Engineering Contradiction:
Improvecheckpoint data consistencyVSAvoidtraining efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The direct storage path acts as an intermediary mechanism that enables fast checkpoint file writing to local storage without requiring the training process to wait for completion. The system can implement asynchronous writing or confirmation mechanisms that ensure data consistency while minimizing training suspension time, thus maintaining both reliability and productivity.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentUS20260073291A1Systems for training artificial intelligence model and checkpoint file storage methods
Publication Date: 2026.03.12 ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
  • US20260073291A1 patent drawing
  • US20260073291A1 patent drawing
  • US20260073291A1 patent drawing

AI summary

This specification provides methods, systems, and computer-readable media for training an artificial intelligence model and storing a checkpoint file. An example system includes a computing node and a first cache node. The computing node is configured to: read a dataset for training from the first cache node to execute a training task of an artificial intelligence model; and in a process of executing the training task, generate a checkpoint file and send it to the first cache node. The first cache node is configured to: identify a type of obtained to-be-stored data; if the type of the to-be-stored data is a dataset, first write the to-be-stored data into a local buffer, and then store the to-be-stored data in a local hard disk from the local buffer; or if the type of the to-be-stored data is a checkpoint file, directly store the to-be-stored data in the local hard disk.