INNER CODE UNIT · Python

build_parser

invergent-ai/surogate · surogate/serve/tools/reference/qwen3_5_moe/cli.py:57

def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--weights", required=True, help="hybrid MoE .sinfer artifact")
    prompt = parser.add_mutually_exclusive_group(required=True)
    prompt.add_argument("--prompt", help="single user message rendered by the artifact template")
    prompt.add_argument("--ids", help="comma/space-separated prompt token IDs")
    prompt.add_argument("--messages", help="Qwen messages JSON with optional image/video content")
    parser.add_argument(
        "--thinking",
        action=argparse.BooleanOptionalAction,
        default=True,
        help="enable or disable the Qwen thinking generation prompt",
    )
    parser.add_argument("--decode", type=int, default=512, help="maximum generated tokens")
    parser.add_argument("--device", default="cuda")
    parser.add_argument("--gpu-memory", default="auto")
    parser.add_argument("--headroom", default="2GiB")
    parser.add_argument("--prefill-chunk", type=int, default=CFG.prefill_chunk)

View source record →