Loading…
Loading…
We adapted the official speculative sampling training method, Eagle3, for training on Qwen3-30B-A3B
After implementing Eagle3, the inference performance of Qwen3-30B-Moe using the SGLang framework on 8*H200 GPU improved from 183 tokens/s to 325 tokens/s.
The TPS (tokens per second) improvement reached nearly 70%.
On a single RTX 5090, the TPS (transactions per second) of Qwen3-8B-Eagle3 increased from 164 to 268.
| model | gpu | tps |
|---|---|---|
| qwen3-30b_moe | h200 | 147 |
| qwen3-30b-moe_eagle3 | h200 | 231 |
| qwen3-30b_moe | 8*h200 | 183 |
| qwen3-30b_moe-eagle3 | 8*h200 | 325 |
| qwen3-30b_moe | 8*5090 | 164 |
| qwen3-30b_moe-eagle3 | 8*5090 | 268 |
Join our AI computing power cloud platform now and enjoy the best AI cloud service experience. The link is as follows: https://tenyunn.com/
To use Eagle3 with SGLang, first replace the qwen3_moe.py file in SGLang’s directory (sglang/python/sglang/srt/models/) with the qwen3_moe.py file from this project.
The launch command for using Eagle3 with SGLang is:
python3 -m sglang.launch_server --model Qwen/Qwen3-30B-A3B --speculative-algorithm EAGLE3 --speculative-draft-model-path Tengyunw/qwen3_30b_moe_eagle3 --speculative-num-steps 6 --speculative-eagle-topk 10 --speculative-num-draft-tokens 32 --mem-fraction 0.9 --cuda-graph-max-bs 2 --dtype bfloat16
Training Dataset: ultrachat_200k. Only the prompts from these datasets were utilized for data synthesis. This synthesized data is used to train the Eagle modules.
dataset nums: 600K samples,1B tokens
Evaluation Dataset: ShareGPT,GSM8K,HUAMEVAL,MT-BENCH,APLCA
Our Sharegpt test data is located in the eagle_data.jsonl file under this directory.
Qwen/Qwen3-8B