GitHub DeepSeek-AI/DeepSpec

DeepSeek-AI released DeepSpec, an open-source codebase for training and evaluating draft models for speculative decoding, supporting three draft model algorithms (DSpark, DFlash, Eagle3) and requiring up to 38 TB of storage for target cache preparation. The project, licensed under MIT, aims to accelerate inference by enabling efficient speculative decoding with target models like Qwen3 and Gemma.

DeepSpec is a full-stack codebase for training and evaluating draft models for speculative decoding. It contains data preparation utilities, draft model implementations, training code, and evaluation scripts. Install the Python dependencies: python -m pip install -r requirements.txt Data preparation additionally requires an inference engine to serve the target model when regenerating answers; see scripts/data/README.md /deepseek-ai/DeepSpec/blob/main/scripts/data/README.md for details. Run the stages in order — each stage's output feeds the next: Data Preparation — download prompts, regenerate target answers, and build the target cache. Training — train a draft model against the cached target outputs. Evaluation — measure speculative-decoding acceptance on benchmark tasks. See scripts/data/README.md /deepseek-ai/DeepSpec/blob/main/scripts/data/README.md for the step-by-step data pipeline: - download and split training data, - regenerate answers, - prepare the target cache storage warning: this can be very large — roughly 38 TB for the default Qwen/Qwen3-4B setting . bash scripts/train/train.sh train.sh launches train.py , which spawns one worker per visible GPU. Select the algorithm and target model by pointing config path at one of the configs under config/ /deepseek-ai/DeepSpec/blob/main/config e.g. config/dspark/dspark qwen3 4b.py ; see the script header for the full list of configs, how to override config path / target cache dir , and how to use --opts to override individual config fields. Checkpoints are written to ~/checkpoints/<project name /<exp name /step . Hardware: the default configs and scripts assume a single node with 8 GPUs. For fewer GPUs, reduce CUDA VISIBLE DEVICES . bash scripts/eval/eval.sh eval.sh runs eval.py against a trained draft checkpoint over the speculative-decoding benchmarks in eval datasets/ /deepseek-ai/DeepSpec/blob/main/eval datasets gsm8k, math500, aime25, humaneval, mbpp, livecodebench, mt-bench, alpaca, arena-hard-v2 . Set: target name or path — the target model the draft was trained against e.g. Qwen/Qwen3-4B , draft name or path — the draft checkpoint, e.g. ~/checkpoints/deepspec/dspark block8 qwen3 4b/step latest . Currently, DeepSpec includes three draft models: DSpark /deepseek-ai/DeepSpec/blob/main/DSpark paper.pdf , DFlash https://arxiv.org/abs/2602.06036 and Eagle3 https://arxiv.org/abs/2503.01840 . DeepSpec is released under the MIT License /deepseek-ai/DeepSpec/blob/main/LICENSE . It includes code adapted from third-party projects under their own licenses; see NOTICE /deepseek-ai/DeepSpec/blob/main/NOTICE for the full attribution. DeepSpec builds on the ideas and code of several excellent open-source projects: SpecForge https://github.com/sgl-project/SpecForge Apache-2.0 — the overall training framework and Eagle3 implementation; portions of the Eagle3 modeling, loss, optimizer, attention, and evaluation code are adapted from it. Adapted files carry an in-file attribution comment, and the full notice is recorded in NOTICE /deepseek-ai/DeepSpec/blob/main/NOTICE . DFlash https://github.com/z-lab/dflash MIT — the DFlash draft-model design and training recipe. Qwen3 https://github.com/QwenLM/Qwen3 and Gemma https://github.com/google-deepmind/gemma — the target model families supported in this repo. We thank the authors and maintainers of these projects. Contributions of new algorithms are welcome.