[Bug] No speedup observed when accelerating Qwen3-235B-MoE model

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
• 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/SpecForge/discussions/new/choose Otherwise, it will be closed.
• 5. Please use English, otherwise it will be closed.

Describe the bug

I tried accelerating the Qwen3-235B-MoE and Qwen-30B-MOE model using this library, both with a custom trained speculative head and the official HuggingFace head. However, no significant speedup was observed. Is there a known issue or specific setup required to achieve acceleration for this model? Any guidance would be appreciated.

Reproduction

python3 -m sglang.launch_server --model Qwen3-30B-A3B --speculative-algorithm EAGLE3 --speculative-draft-model-path Qwen3-30B-A3B-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 --tp 2 --host 0.0.0.0 --port 7801

Environment

A800

Start with the reported sglang.launch_server command and the A800 environment, comparing its throughput with and without EAGLE3 for the stated Qwen model. Verify whether the command uses the intended model and draft head, then document the measured results and any confirmed setup or compatibility limitation.