DeepSeek-R1-Distill-Qwen-7B-Eagle3

EAGLE3 Draft Head — DeepSeek-R1-Distill-Qwen-7B

A speculative decoding draft head for deepseek-ai/DeepSeek-R1-Distill-Qwen-7B, trained using the EAGLE3 method on Google Cloud TPU with the SpecJAX framework.

EAGLE3 draft heads accelerate autoregressive generation by proposing multiple tokens per step that a target model then verifies in parallel — typically achieving 2-3x throughput gains with no change in output quality.

Usage

SGLang (GPU)

Note: DeepSeek-R1-Distill-Qwen uses the Qwen2 architecture. EAGLE3 support requires a small patch to SGLang (adding set_eagle3_layers_to_capture() to the Qwen2 model). See the SpecJAX inference guide for details.

python -m sglang.launch_server \
    --model deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/DeepSeek-R1-Distill-Qwen-7B-Eagle3 \
    --speculative-num-steps 5 \
    --speculative-eagle-topk 4 \
    --dtype bfloat16

sglang-jax (TPU)

Note: Requires the same Qwen2 EAGLE3 patch applied to sglang-jax. The sglang-jax EAGLE3 pipeline is functional but not yet performance-optimized.

python -m sgl_jax.launch_server \
    --model-path deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/DeepSeek-R1-Distill-Qwen-7B-Eagle3 \
    --speculative-eagle-topk 1 \
    --speculative-num-steps 3 \
    --speculative-num-draft-tokens 4 \
    --tp-size 4 --dtype bfloat16

Python (SGLang client)

import sglang as sgl

llm = sgl.LLM(
    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
    speculative_algorithm="EAGLE3",
    speculative_draft_model_path="thoughtworks/DeepSeek-R1-Distill-Qwen-7B-Eagle3",
    speculative_num_steps=5,
    speculative_eagle_topk=4,
    dtype="bfloat16",
)

Training Details

Training Method

This model uses EAGLE3's Test-Time Training (TTT) objective with a rollout length of 7. At each training step, the draft head autoregressively proposes 7 tokens; the target model provides ground-truth hidden states and logits for all positions; a geometric loss (0.8^k weighting) trains the draft to match the target at each position.

Performance

Token acceptance rates on generic instruction-following data (ShareGPT-style prompts):

Measured on held-out evaluation data. Actual throughput gains depend on hardware, prompt distribution, and runtime version.

Model Architecture

The draft head is a single-layer transformer that operates on the target model's hidden states:

Limitations

• Trained on English-dominant instruction data; performance may degrade on non-English inputs or highly domain-specific content.
• Acceptance rates are measured on generic chat data and will vary by prompt distribution.
• This is a v1 checkpoint trained on generic data. A v2 with target-model-regenerated training data is planned.

License

This model is released under the MIT License. The base model (DeepSeek-R1-Distill-Qwen-7B) is subject to its own license terms.

References

@article{li2025eagle3,
  title={EAGLE3: Scalable Speculative Decoding with Training-Free Multi-Draft Speculation},
  author={Li, Yuhui and Wei, Fangyun and Zhang, Chao and Zhang, Hongyang},
  journal={arXiv preprint arXiv:2503.01840},
  year={2025}
}