What it is
Full-stack codebase for preparing data, training draft models, and evaluating speculative decoding against target LLMs. Supports DSpark, DFlash, and Eagle3. Train and benchmark draft models for speculative decoding with an end-to-end pipeline covering target-output caching, multi-GPU training, and benchmark evaluation.
Key results
- Supports 3 algorithms: DSpark, DFlash, and Eagle3
- Evaluates on 9 benchmarks including GSM8K, MATH500, HumanEval, MBPP
- Default target cache for Qwen3-4B is ~38 TB
Shows that speculative decoding research can be run end-to-end in one reproducible stack across multiple draft algorithms and target LLM families.