thoughtworks/Qwen3-32B-Eagle3

EAGLE3 Draft Head — Qwen3-32B

A speculative decoding draft head for Qwen/Qwen3-32B, 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.

This is the first EAGLE3 draft head trained with TP=8 tensor parallelism spanning multiple hosts.

Usage

SGLang (GPU)

Qwen3 EAGLE3 is natively supported in SGLang.

python -m sglang.launch_server \
    --model Qwen/Qwen3-32B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/Qwen3-32B-Eagle3 \
    --speculative-num-steps 5 \
    --speculative-eagle-topk 4 \
    --dtype bfloat16

Thinking mode

Qwen3 supports an optional thinking mode (/think and /no_think tokens). This draft head was trained on generic instruction-following data and is compatible with both modes:

# Disable thinking mode for pure instruction-following workloads
python -m sglang.launch_server \
    --model Qwen/Qwen3-32B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/Qwen3-32B-Eagle3 \
    --speculative-num-steps 5 \
    --speculative-eagle-topk 4 \
    --dtype bfloat16 \
    --chat-template qwen3-instruct-no-thinking

sglang-jax (TPU)

Qwen3 EAGLE3 is natively supported in sglang-jax. Note: sglang-jax's EAGLE3 pipeline is functional but not yet performance-optimized.

python -m sgl_jax.launch_server \
    --model-path Qwen/Qwen3-32B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path thoughtworks/Qwen3-32B-Eagle3 \
    --speculative-eagle-topk 1 \
    --speculative-num-steps 3 \
    --speculative-num-draft-tokens 4 \
    --tp-size 8 --dtype bfloat16

Python (SGLang client)

import sglang as sgl

llm = sgl.LLM(
    model="Qwen/Qwen3-32B",
    speculative_algorithm="EAGLE3",
    speculative_draft_model_path="thoughtworks/Qwen3-32B-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=8, DP=2)
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	1.5e-4
Batch size	B=1, sequence length T=2048, gradient accumulation 8
TTT length	7 (multi-step speculative rollout)
Training time	~12 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.

Qwen3's architecture includes per-head QK RMSNorm and tied word embeddings. The draft head is trained to match Qwen3's output distribution at every speculative position.

Note on learning rate: The standard LR of 3e-4 used for smaller models diverged for 32B. A reduced LR of 1.5e-4 (linear batch-size scaling) was required for stable convergence.

Performance

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

Position	Acceptance Rate
acc_0 (1st draft token)	59.3%
acc_1	55.7%
acc_2	53.9%

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	5120
Attention heads	40 (GQA: 8 KV heads)
Vocabulary size	151,936 (full target vocab)
Draft vocab size	32,000 (top tokens by training frequency)
Parameters	~530M

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 (non-thinking mode) and may differ under extended thinking prompts.
• 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

@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}
}