sglang on Ascend 910C runs less than half the speed of vllm-ascend

[weiyangdaren]

Hi, I encountered a significant performance gap when running sglang on Ascend 910C compared to vllm-ascend.

Test Results

Hardware	Framework	model	Speed (tokens/sec)
NVIDIA 4090D	vLLM	Qwen3-8B	~1055
NVIDIA 4090D	sglang	Qwen3-8B	~1109
Ascend 910C	vllm-ascend	Qwen3-32B	~567
Ascend 910C	sglang	Qwen3-32B	~244

On 4090D, vLLM and sglang have very similar performance.
But on Ascend 910C, sglang is less than half the speed of vllm-ascend.

What could be the potential factors causing this performance difference on Ascend 910C?

Thank you for your help!

Below is a screenshot of the experimental results.

---