dFactory documentation

Quickstart

• Installation
  • Prerequisites
  • Clone the Repository
  • Environment Setup and Dependencies
  • Next Steps
• Preparation
  • Download and Merge Model Weights
  • Prepare Training Data
• Quickstart: SFT upon dFactory
  • VeOmni Best Practices
  • Model and Optimizer
  • Train Loop
• Configuration Design Guide
  • YAML Configuration Structure
  • Model Configuration
  • Data Configuration
  • Training Configuration
  • Configuration Patterns
  • Hardware Adaptation
  • Best Practices

Train

• Distributed Training
  • Single-Node, Multi-GPU Training
  • Instructions
  • Multi-Node, Multi-GPU Training
  • Prerequisites
  • Environment Variables
  • Example for a 2-Node Setup
• Profiler

Inference

• SGLang dLLM Inference Guide
  • Overview
  • Installation
  • Server Launch
  • API Usage
  • Additional Resources

Algorithms

• Discrete Diffusion Model
  • Random Masking Process (Forward Process)
  • Pre-training Objective
  • Supervised Fine-tuning (SFT)
  • Inference: The Reverse Denoising Process
  • Perplexity (PPL) Calculation
• Block Diffusion
  • Core Design Rationale
  • Methodology
  • Helper Function to Create Block Diffusion Mask
• Trainable Parallel Decoding
  • Overview
  • Path Distillation (Trajectory Compression)
  • DPARALLEL: Learnable Parallel Decoding

FAQ

• Frequently Asked Questions
  • Veomni related
  • Install related
  • Evaluation related