Deploy fine-tuned LLMs without compromising on control

import modal

vllm_image = (

    modal.Image.from_registry(f"nvidia/{tag}", add_python="3.12")

    .uv_pip_install("vllm==0.10.2", "torch==2.8.0")

)

model_cache = modal.Volume.from_name("huggingface-cache", create_if_missing=True)

app = modal.App("vllm-inference")

@app.function(image=vllm_image, gpu="H100", volumes={"/root/.cache/huggingface": model_cache})

@modal.web_server(port=8000)

def serve():

    import subprocess

    cmd = "vllm serve Qwen/Qwen3-8B-FP8 --port 8000"

    subprocess.Popen(cmd)