yzhuang/MixtureOfInputs

1import subprocess2import threading3import os4import time5import spaces6 7def setup_mixinputs():8    # Step 1: Run mixinputs setup9    subprocess.run(["mixinputs", "setup"], check=True)10 11def launch_vllm_server(beta=1.0):12    # Step 2: Set environment variables13    env = os.environ.copy()14    env["MIXINPUTS_BETA"] = str(beta)15    env["VLLM_USE_V1"] = "1"16 17    # Step 3: Launch vLLM with custom options18    cmd = [19        "vllm", "serve",20        "Qwen/Qwen3-4B",21        "--tensor-parallel-size", "1",22        "--enforce-eager",23        "--max-model-len", "4096",24        "--max-seq-len-to-capture", "4096",25        "--max-num-seqs", "36",26        "--host", "0.0.0.0",27        "--port", "8000",28        "--enable-reasoning",29        "--reasoning-parser", "deepseek_r1",30        "--api-key", "EMPTY"31    ]32    _server_process = subprocess.Popen(cmd, env=env)33        34    # # Wait for the process to complete or be terminated35    # _server_process.wait()36 37# Step 1: Setup38setup_mixinputs()39 40# Step 2: Launch vLLM server in background41threading.Thread(target=launch_vllm_server, daemon=True).start()42