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

Describe the bug

When I run Fish-s2-pro on RTX 3090 using the example provided here: tts_s2pro.md, the process OOMs even though S2 Pro only needs 24GB VRAM. Running S2 Pro plainly as instructed on Fish Audio website here works fine, but it's slower than the benchmarks and I am trying to see if running it on sglang would give me better inference speeds.

Any idea how I can get sglang to reduce it's memory footprint and run S2 Pro on RTX 3090?

Reproduction

git clone https://github.com/sgl-project/sglang-omni.git
cd sglang-omni
uv venv .venv -p 3.12 && source .venv/bin/activate
uv pip install -v .
hf download fishaudio/s2-pro

sgl-omni serve
--model-path fishaudio/s2-pro
--config examples/configs/s2pro_tts.yaml
--port 8000

Environment

docker pull frankleeeee/sglang-omni:dev
docker run -it --shm-size 32g --gpus all frankleeeee/sglang-omni:dev /bin/zsh