Witllm/qwen/demo.py

import torch
from modelscope import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers.generation import GenerationConfig
from transformers import AutoConfig

from modeling_qwen import QWenLMHeadModel
from modeling_qwen import QwenRunner

seed = 4321
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)

model_dir = snapshot_download("qwen/Qwen-1_8B-Chat")
# model_dir = "/home/colin/.cache/modelscope/hub/qwen/Qwen-1_8B-Chat"

config, kwargs = AutoConfig.from_pretrained(
    "./",
    return_unused_kwargs=True,
    trust_remote_code=True,
    code_revision=None,
    _commit_hash=None,
)
model = QWenLMHeadModel(config)

print(model)

tokenizer = AutoTokenizer.from_pretrained(model_dir, trust_remote_code=True)
model = model.from_pretrained(model_dir).cuda()

model = model.eval()
# model = model.train()  # control by @torch.no_grad()

# 可指定不同的生成长度、top_p等相关超参
# model.generation_config = GenerationConfig.from_pretrained(
#     model_dir, trust_remote_code=True
# )

runner = QwenRunner(model)

# 第一轮对话
response, history, decode_tokens = runner.Chat(tokenizer, "东南亚国家日本的首都是什么市", "")
print(decode_tokens)
# <|im_start|>system
# You are a helpful assistant.<|im_end|>
# <|im_start|>user
# 东南亚国家日本的首都是什么市<|im_end|>
# <|im_start|>assistant
# 日本的首都东京。<|im_end|><|endoftext|>

# 第二轮对话

response, history, decode_tokens = runner.Chat(tokenizer, "给我讲一个年轻人奋斗创业最终取得成功的故事。", "")
print(decode_tokens)

if decode_tokens.split("\n")[-2] != """这个故事告诉我们，只要我们有决心和毅力，就一定能够克服困难，实现我们的梦想。<|im_end|>""":
    raise ()