sgl-project / sgl-project/SpecForge

Why is the TPS of eagle3-qwen in the sglang inference of single-card H20 not as high as that of the original QWEN3 when the decoding algorithm is added

Open
#236 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
1.2k
Forks
347
Avg merge
4d 1h
Merged PRs (30d)
41

Description

Hello, I'm testing the speed of 100 tokens on a single H20. The original qwen3 has 200TPS during sglang inference, while the draft model eagle3 only has 130TPS. What's the reason for this

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by reproducing the reported 100-token comparison on a single H20: original Qwen3 at about 200 TPS versus eagle3-qwen at about 130 TPS during SGLang inference. Compare the two inference paths and decoding configuration, then document the measured cause of the throughput difference and a verified conclusion.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.