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/sglang/discussions/new/choose Otherwise, it will be closed.
• 5. Please use English, otherwise it will be closed.

Describe the bug

Accuracy on gsm8k dataset is decreased for EP MoE.
cc: @xiaobochen123

Reproduction

EP:

python3 -m sglang.launch_server --model-path neuralmagic/DeepSeek-Coder-V2-Instruct-FP8 --disable-radix-cache --trust-remote-code --tp 8 --enable-ep-moe --disable-cuda-graph
python3 benchmark/gsm8k/bench_sglang.py --num-questions 1400 --parallel 1400

Accuracy: 0.540
Invalid: 0.005
Latency: 205.758 s
Output throughput: 1017.681 token/s

TP:

python3 -m sglang.launch_server --model-path neuralmagic/DeepSeek-Coder-V2-Instruct-FP8 --disable-radix-cache --trust-remote-code --tp 8 --disable-cuda-graph
python3 benchmark/gsm8k/bench_sglang.py --num-questions 1400 --parallel 1400

Accuracy: 0.930
Invalid: 0.000
Latency: 196.344 s
Output throughput: 1011.191 token/s

Environment

• sglang: main branch (0.4.0.post1)
• torch: 2.5.1
• triton: 3.1.0