from pathlib import Path
import torch
from qwen35_standalone import load_standalone

HERE = Path(__file__).resolve().parent

if __name__ == "__main__":
    model, tokenizer = load_standalone(HERE)
    prompt = tokenizer.apply_chat_template(
        [{"role": "user", "content": "Explain sparse neural networks in one paragraph."}],
        tokenize=False,
        add_generation_prompt=True,
    )
    inputs = tokenizer(prompt, return_tensors="pt").to(next(model.parameters()).device)
    with torch.inference_mode():
        out = model.generate(**inputs, max_new_tokens=128, do_sample=False)
    print(tokenizer.decode(out[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))