"""Cloud GPU command-line chat. The original CLI is vision_chat.py.""" import argparse from chatcode.runtime import answer, load_engine if __name__ == "__main__": parser = argparse.ArgumentParser(description="PocketVision Chat & Code (text only, cloud GPU)") parser.add_argument("prompt") parser.add_argument("--adapter", help="A matching chat/code adapter; omit to use the pretrained base") parser.add_argument("--max-new-tokens", type=int, default=768) args = parser.parse_args() try: tokenizer, model = load_engine(args.adapter) print(answer(tokenizer, model, args.prompt, max_new_tokens=args.max_new_tokens)) except (ValueError, RuntimeError) as error: parser.exit(1, f"{error}\n")