Annoucement

• [2025-10] 🎉🎉 Efficiency Report: We provide comprehensive Model FLOPs Utilization (MFU) metrics for various model architectures and training configurations. See MFU Reference for detailed benchmarks.
• [2025-10] 🚀🚀 LMMs-Engine v0.1 is here! a lean, efficient framework built to train unified multimodal model at scale.

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/LMMs-Lab/lmms-engine.git
cd lmms-engine

# Install editable packages
uv pip install -e ".[all]"
# or install as a packages
uv pip install -e .
# Install a stable release
uv pip install lmms-engine

# Install dependencies using uv sync
# For Linux systems (recommended - auto-detects platform):
bash uv_sync_linux.sh

# For other systems or if encountering errors:
uv sync
# If uv sync fails, try: 
uv pip install -r requirements.txt

# Optional: Performance optimizations
uv pip install flash-attn --no-build-isolation
uv pip install liger-kernel

Docker

We provide Docker images with pre-built environments including PyTorch, CUDA, and all necessary dependencies.

docker run --gpus all -it --rm \
  -v $(pwd):/workspace \
  -w /workspace \  
  fatbao55/lmms-engine:v1.0 \
  bash

Launch Training

Recommended: torchrun (native PyTorch)

torchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 \
  --master_addr=127.0.0.1 --master_port=12355 \
  -m lmms_engine.launch.cli config_yaml=examples/qwen3_vl/example_config.yaml

Alternative: Accelerate

accelerate launch --use_fsdp \
  -m lmms_engine.launch.cli config_yaml=examples/qwen3_vl/example_config.yaml

Single GPU

python -m lmms_engine.launch.cli config_yaml=examples/qwen3_vl/example_config.yaml

🔥 Featured Examples

Optimization Legend:

• FSDP2: Fully Sharded Data Parallel v2 for distributed training
• USP: Ulysses Sequence Parallel for long contexts
• Muon: Advanced optimizer with Newton-Schulz orthogonalization
• Liger: Triton fused kernels (CrossEntropy, RMSNorm, RoPE, SwiGLU) for 30% memory reduction
• Packing: First-fit bin packing for peaking at 35-40% MFU vs 20-25% (w/o in Qwen2.5-VL finetuning)
• NSA: Native Sparse Attention for efficient long-context processing
• EP: Expert Parallelism for Mixture-of-Experts models, sharding experts across GPUs

💡 Tip: Each run.sh file contains detailed setup instructions, prerequisites, and configuration options.

🤖 Model Support

20+ architectures spanning vision-language, diffusion, and language models.

Multimodal Models

• Qwen2.5-VL - SOTA level performance vision-language model
• Qwen3-VL - SOTA level performance vision-language model
• Qwen3-VL MoE - Vision-Language Mixture-of-Experts with Expert Parallelism and Sequence Parallelism support
• Qwen2.5-Omni - Unified vision + audio + text modalities
• Qwen3-Omni MoE - Multimodal Mixture-of-Experts with vision + audio + text and Expert Parallelism support
• LLaVA-OneVision - Fully open-source vision-language model
• Bagel - Unified multimodal model for visual understanding and generation
• Aero - Lightweight audio-language model

Diffusion & Generative Models

• dLLM (Qwen3) - Diffusion Language Model with masked prediction
• WanVideo (1.3B/14B) - Text/Image-to-Video generation (T2V/I2V/V2V)
• SiT (XL/2) - Scalable Interpolant Transformers for class-conditional image generation
• RAE-SigLip - Representation AutoEncoder with adversarial discriminator

Language Models

• Qwen2/2.5/3 series - Full Liger kernel support with fused operations
• Linear Attention Models - Recurrent architecture optimized for Muon; Please install FLA first.
• Custom architectures - Extensible via @register_model() decorator

⚡️ Optimizations

Production-grade efficiency from distributed training to kernel fusion.

Core Distributed Training

• FSDP2 - PyTorch 2.0+ DTensor-based sharding for parameters, gradients, and optimizer states. Improved composability over original FSDP enables flexible parallelism composition.

• Ulysses Sequence Parallel - Splits sequence dimension across GPUs for ultra-long contexts. Critical for vision-language models like Qwen3-VL with 10K+ visual tokens.

• Multi-dimensional Parallelism - Compose TP x PP × DP meshes for cluster-scale training.

Memory & Compute Optimizations

• Flash Attention + Unpadding - Tiled attention with use_rmpad eliminates all padding computation.

• Native Sparse Attention (NSA) - Hybrid attention mechanism combining compressed attention, topk sparse attention, and sliding window attention.

• Liger Kernel - Triton fused kernels (CrossEntropy, RMSNorm, RoPE, SwiGLU) achieve memory reduction by avoiding intermediate materializations.

• Monkey Patching System - Runtime kernel injection via lmms_engine/configs/monkey_patch/ for model-specific optimizations without code modification.

• Sequence Packing - Faster first-fit bin packing.

Advanced Optimizer

• Muon Optimizer - Newton-Schulz orthogonalization with Triton kernels, distributed via DTensor. Selective 2D-parameter application outperforms AdamW convergence.

Data Pipeline

• Streaming Datasets - IterableDataset for trillion-token pretraining without full data loading.

Configuration Examples

dataset_config:
  packing: true
  packing_strategy: first_fit
  packing_length: 32000

trainer_args:
  use_rmpad: true  # Requires flash-attn
  use_liger_kernel: true

trainer_args:
  use_liger_kernel: true

trainer_args:
  use_muon: true # enable muonwithadam optimizer
  adam_beta1: 0.9 # for the adam part in muonwithadam optimizer
  adam_beta2: 0.999 # for the adam part in muonwithadam optimizer
  adam_epsilon: 1.0e-8 # for the adam part in muonwithadam optimizer
  learning_rate: 0.001
  weight_decay: 0.01
  # ns_steps: 5  # Newton-Schulz iterations (default)

  # for some modules which the user hope to 

trainer_args:
  fsdp2: true
  fsdp_config:
    transformer_layer_cls_to_wrap: ["Qwen2VLDecoderLayer"]
    reshard_after_forward: false
    activation_checkpointing: true

trainer_args:
  sp_ulysses_degree: 2  # Sequence parallel degree

model_config:
  load_from_pretrained_path: "lmms-lab/BAGEL-7B-MoT-ver.LE"

monkey_patch:
  - type: nsa
    model_type: bagel
    kwargs:
      block_size: 64
      compress_type: "weightedpool"  # weightedpool, linear, avgpool
      kernel_size: 32
      kernel_stride: 16
      topk: 16
      init_blocks: 1
      local_blocks: 2
      window_size: 512

📖 Documentation

Step-by-Step Workflow

1. Process the dataset into OpenAI chat format (JSONL/JSON/Arrow/CSV)

   hf download kcz358/open-thoughts-debug --local-dir data/open_thoughts_debug --repo-type dataset

2. Prepare dataset YAML (optional for single data source)

   datasets:
     - path: data/open_thoughts_debug
       data_folder: ""
       data_type: arrow

3. Configure training - See examples/qwen3_vl/example_config.yaml or any model-specific config in examples/

Comprehensive Guides

Getting Started:

• Dataset Preparation - How to prepare and structure your data
• Dataset & Packing Guide - Detailed dataset implementations and packing strategies
• Training Guide - Comprehensive training walkthrough

Advanced Topics:

• Design Principles - Architectural patterns and philosophy
• API Reference - Detailed API documentation

🏗️ Codebase Architecture

Component Registry

Factory Pattern enables easy extensibility:

# Register a custom dataset
from lmms_engine.datasets import register_dataset, BaseDataset

@register_dataset("my_custom_dataset")
class MyCustomDataset(BaseDataset):
    def __init__(self, config):
        super().__init__(config)
        # Custom initialization

    def __getitem__(self, idx):
        # Custom data loading
        return item

# Register a custom processor
from lmms_engine.datasets.processor import register_processor

@register_processor("my_custom_processor")
class MyCustomProcessor:
    def __call__(self, raw_data):
        # Custom processing
        return processed_data

Training Pipeline

Builder Pattern for flexible composition:

from lmms_engine.train import TrainRunner

# Configuration defines the pipeline
runner = TrainRunner(config)
runner.build()  # Lazy initialization of components
runner.run()    # Execute training

Pipeline stages:

1. Model initialization - From pretrained or config
2. Dataset creation - With processor and collator
3. Monkey patching - Apply kernel optimizations
4. Trainer setup - FSDP2, DeepSpeed, or custom
5. Training execution - With checkpointing and logging

Supported Trainers

🎯 Use Cases

• Vision-Language Pretraining - Qwen-VL, LLaVA on large multimodal datasets
• Video Understanding - AERO on 3D video data
• Diffusion Models - DLLM, SiT, WanVideo for generation tasks
• Representation Learning - RAE for visual representations
• Language Model Pretraining - DGN, Qwen with Muon optimizer
• Multimodal Fine-tuning - Efficient SFT with sequence packing

🤝 Contributing

We welcome contributions! Please see our Design Principles for coding guidelines:

• Simplicity: Write simple, straightforward code
• Readability: Prioritize clarity over cleverness
• Testability: Create testable components
• Minimal Changes: Only modify code related to the task
• Less Code = Less Debt: Minimize code footprint

😊 Acknowledgement

Thanks to the following projects for their excellent work:

• axolotl
• LLaMA-Factory
• nanotron
• veScale
• veOmni

📝 Citation

If you use LMMs Engine in your research, please cite:

@software{lmms_engine2025,
  title={LMMs Engine: A simple, unified multimodal framework for pretraining and finetuning.},
  author={LMMs-Lab},
  year={2025},
  url={https://github.com/LMMs-Lab/lmms-engine}
}

📄 License

This project is licensed under the Apache 2.0 License - see the LICENSE file for details.

🔗 Links

• GitHub: https://github.com/EvolvingLMMs-Lab/lmms-engine
• LMMs-Lab: https://lmms-lab.com
• Documentation: docs/
• Issues: https://github.com/EvolvingLMMs-Lab/lmms-engine/issues

🎉 Awesome projects using LMMs-Engine

• LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling

• OpenMMReasoner: Pushing the Frontiers for Multimodal Reasoning with an Open and General Recipe