DeepSpec

DeepSpec is an open-source, full-stack codebase developed by DeepSeek for training and evaluating speculative decoding algorithms. It encompasses data preparation utilities, draft model implementations, training code, and evaluation scripts, facilitating the development and assessment of draft models for speculative decoding tasks.

DeepSpec is an open-source, full-stack codebase developed by DeepSeek for training and evaluating speculative decoding algorithms. It encompasses data preparation utilities, draft model implementations, training code, and evaluation scripts, facilitating the development and assessment of draft models for speculative decoding tasks.

• 01Open-source codebase for speculative decoding algorithms
• 02Includes data preparation utilities, draft model implementations, training code, and evaluation scripts
• 03Supports algorithms like DSpark, DFlash, and Eagle3
• 04Released under the MIT License
• 05Requires a single 8-GPU node and approximately 38 TB of storage for data preparation