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: Qwen2.5 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.

bash

python -m sglang.launch_server \
    --model Qwen/Qwen2.5-7B-Instruct \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/Qwen2.5-7B-Instruct-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.

bash

python -m sgl_jax.launch_server \
    --model-path Qwen/Qwen2.5-7B-Instruct \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/Qwen2.5-7B-Instruct-Eagle3 \
    --speculative-eagle-topk 1 \
    --speculative-num-steps 3 \
    --speculative-num-draft-tokens 4 \
    --tp-size 4 --dtype bfloat16

Python (SGLang client)

python

import sglang as sgl

llm = sgl.LLM(
    model="Qwen/Qwen2.5-7B-Instruct",
    speculative_algorithm="EAGLE3",
    speculative_draft_model_path="thoughtworks/Qwen2.5-7B-Instruct-Eagle3",
    speculative_num_steps=5,
    speculative_eagle_topk=4,
    dtype="bfloat16",
)

Training Details

Parameter	Value
Framework	SpecJAX — pure JAX, no Flax/PyTorch
Hardware	Google Cloud TPU v4-32 (4 hosts x 4 chips, TP=4, DP=4)
Dataset	54K mixed: ShareGPT (45%) + UltraChat-200K (35%) + Open-PerfectBlend (20%)
Epochs	3
Steps	9,966 total
Optimizer	AdamW, cosine LR decay, 3% warmup
Learning rate	8e-4
Batch size	B=4, sequence length T=1024, gradient accumulation 4
TTT length	7 (multi-step speculative rollout)
Training time	~5.5 hours
Precision	bfloat16

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):

Position	Acceptance Rate
acc_0 (1st draft token)	61.8%
acc_1	56.9%
acc_2	54.6%

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:

Parameter	Value
Architecture	`LlamaForCausalLM` (1 decoder layer)
Hidden size	3584
Attention heads	28 (GQA: 4 KV heads)
Vocabulary size	152,064 (full target vocab)
Draft vocab size	32,000 (top tokens by training frequency)
Parameters	~300M

Auxiliary Layer Indices

This head uses multi-layer feature fusion from layers {1, 13, 24} of the 28-layer Qwen2.5-7B model (SpecForge convention: {1, L//2-1, L-4}).

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 Apache License 2.0, consistent with the base model's license.

References

bibtex

@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}
}