Function main

tests/warmbench.py:4005–4040

()

        str(scenario_dir / "cpp.timing.log"),
        "--output-dir",
        str(scenario_dir / "cpp_audio"),
    ]
    if args.test_noise_file:
        cpp_command.extend(["--noise-file", args.test_noise_file])
    for option in args.cpp_session_option:
        cpp_command.extend(["--session-option", option])
    return python_command, cpp_command


def build_vibevoice_commands(
    config: dict[str, Any],
    backend: str,
    args: argparse.Namespace,
    scenario_dir: Path,
    requests: list[dict[str, Any]],
) -> tuple[list[str], list[str]]:
    request_sequence_json = json.dumps(requests, ensure_ascii=False, separators=(",", ":"))
    model_path = args.model or config["model"]
    python_env = str(config.get("python_conda_env", "qwen3-tts"))
    model_root = Path(model_path)
    if not model_root.is_absolute():
        model_root = REPO_ROOT / model_root
    model_config = json.loads((model_root / "config.json").read_text(encoding="utf-8"))
    latent_size = int(model_config["acoustic_vae_dim"])
    decoder_capacity = int(model_config["decoder_config"]["max_position_embeddings"])
    total_speakers = sum(max(1, len(request.get("voice_samples", []))) for request in requests)
    prompt_noise_count = total_speakers + total_speakers * decoder_capacity * latent_size
    prompt_noise_path = scenario_dir / "shared_prompt_noise.f32"
    prompt_rng = random.Random(args.seed + 17)
    prompt_values = array.array("f", (prompt_rng.gauss(0.0, 1.0) for _ in range(prompt_noise_count)))
    prompt_noise_path.parent.mkdir(parents=True, exist_ok=True)
    with prompt_noise_path.open("wb") as handle:
        prompt_values.tofile(handle)
    prompt_noise_file = str(prompt_noise_path)
    noise_file = args.test_noise_file
    if not noise_file:
        max_new_tokens = max([int(request.get("max_new_tokens", 0)) for request in requests] + [0])
        if max_new_tokens <= 0:
            max_new_tokens = decoder_capacity