AngelSpec Documentation

AngelSpec is a torch-native framework for training speculative-decoding draft models, covering both autoregressive MTP drafting and the block-parallel DFlash family. Inference and training are disaggregated: inference engines generate target-model hidden states and stream them — via the Mooncake store over RDMA — to distributed training workers, so each side scales independently.

AngelSpec supports 6 draft architectures (Eagle3, DFlash, DFlare, DFly, DSpark, MTP) across multiple inference backends (vLLM, SGLang, HuggingFace).

Get Started

• About AngelSpec
• Installation
• Quickstart

Concepts

• Speculative Decoding
  • Overview
  • How it works
  • Feature-space draft models
• Disaggregated Architecture
  • Why disaggregate
  • Components
  • Data flow
  • Related pages
• The Draft-Model Family
  • Comparison
  • Shared training pipeline
  • Selecting an architecture
  • Architecture pages

Basic Usage

• Data Preparation
• Training
• Inference Backends
• Checkpoint Conversion

Advanced Features

• Multi-Node Training
• Long-Sequence Training (USP)
• Sequence Packing
• Performance Metrics
• Customization

Examples

• Examples