xtuner — What is it?

InternLM/xtuner is a Python-based training engine designed for ultra-large MoE models, addressing the challenges of training and scaling these complex models efficiently.

⭐ 5,112 Stars 🍴 413 Forks Python Apache-2.0 Author: InternLM
Source: README View on GitHub →

Why it matters

XTuner is gaining attention due to its innovative approach to training MoE models, offering scalable and efficient solutions for large-scale language model training. Its unique dropless training and long sequence support features stand out, addressing the pain points of traditional training methods.

Source: Synthesis of README and project traits

Core Features

Dropless Training

XTuner's dropless training allows for scalable training of MoE models without the complexity of traditional 3D parallel training architectures, optimizing parallelism for efficiency.

Source: README
Long Sequence Support

The engine supports training on long sequences with memory-efficient design, enabling training of 200B MoE models on 64k sequence lengths without sequence parallelism.

Source: README
Superior Efficiency

XTuner achieves high training efficiency on Ascend NPU, surpassing traditional 3D parallel schemes for MoE models above 200B scale, and supports MoE training up to 1T parameters.

Source: README

Architecture

The architecture of XTuner is modular, with distinct components for training, algorithm, and inference. It leverages advanced memory optimization techniques and supports various hardware platforms, including GPUs and NPUs. The code structure reflects a clear separation of concerns, with dedicated modules for different functionalities.

Source: Code tree + dependency files

Project Knowledge Graph

Knowledge graph: project (center) + core features (inner hexagons) + key dependencies (outer chips) bitsandbytes mmengine transformers torch torchvision Dropless Training Long Sequence SupportLong Sequence Suppo… Superior Efficiency xtuner Project Core feature Key dependency

Center: project; inner ring: core feature modules; outer ring: key dependencies. Auto-generated from core_features and tech_stack.key_deps.

Tech Stack

LanguagePythonFrameworkPyTorch, DeepSpeed, MindSpeed
bitsandbytesmmenginetransformerstorchtorchvision
Docker
Source: Dependency files + code tree

Quick Start

pip install xtuner xtuner train --config path/to/config.yaml
Source: README Installation/Quick Start

Use Cases

XTuner is suitable for researchers and developers working on large-scale language models, particularly those focusing on MoE architectures. It is useful for pre-training, instruction fine-tuning, and reinforcement learning of ultra-large MoE models.

Source: README

Strengths & Limitations

Strengths

  • Strength 1: Efficient training of ultra-large MoE models
  • Strength 2: Scalable and memory-efficient design
  • Strength 3: Support for various hardware platforms

Limitations

  • Limitation 1: Limited documentation on certain advanced features
  • Limitation 2: Requires expertise in large-scale model training
Source: Synthesis of README, code structure and dependencies

Latest Release

v1.0.1 (2026-05-15): Bug fixes and improvements to the main branch v1.0.0rc0 (2025-11-18): Release candidate for version 1.0.0 v0.2.0 (2025-07-11): Initial release with support for pre-trained RM and bug fixes

Source: GitHub Releases

Verdict

InternLM/xtuner is a promising project for those involved in large-scale language model development, offering innovative solutions for training MoE models. It is particularly suited for teams with expertise in deep learning and large-scale model training.

Source: Synthesis
Transparency Notice
This page is auto-generated by AI (a large language model) from the following public materials: GitHub README, code tree, dependency files and release notes. Analyzed at: 2026-05-24 15:20. Quality score: 85/100.

Data sources: README, GitHub API, dependency files