example_load.py

"""跑通这个发布包 —— 只需要 torch 和 piece_tokenizer。

    pip install torch
    pip install git+https://github.com/Ismantic/PieceTokenizer
    python example_load.py
"""
import torch

from model import Qwen3ForCausalLM
from tokenizer import PieceTokenizerWrapper

HERE = "."

tok = PieceTokenizerWrapper(HERE)
model = Qwen3ForCausalLM.from_pretrained(
    HERE, device="cuda" if torch.cuda.is_available() else "cpu",
    dtype=torch.bfloat16)

prompt = "机器翻译的基本任务是"
ids = [tok.bos_token_id] + tok.encode(prompt, add_special_tokens=False)  # 这个模型训练时每行都以 <bos> 开头,不加是分布外输入——会生成退化的重复内容
print(f"{prompt!r} -> {len(ids)} tokens")

# 贪心续写 40 步。没有 KV cache —— 每步重算前缀,短续写够用。
x = torch.tensor([ids], device=next(model.parameters()).device)
out = []
with torch.no_grad():
    for _ in range(40):
        nxt = int(model(x)[0, -1].argmax())
        if nxt in tok.stop_token_ids:
            break
        out.append(nxt)
        x = torch.cat([x, torch.tensor([[nxt]], device=x.device)], dim=1)

print(prompt + tok.decode(out))