首先可以参考modelScope社区给出的使用文档，已经足够全面
但在按照文档中步骤部署时，还是有些错误问题发生，可以搜索参考的解决方式不多，所以记录下来
个人电脑部署
这里不太建议使用自己的笔记本部署通义千问模型，因为实在是太耗资源，我使用的M2芯片的MacBook Pro即使运行起来了，但模型回答一个问题都需要四五分钟的时间，内存全部占满，其他应用程序也都强制退出了。所以还是使用社区提供的免费资源，或者有更高配置的服务器来部署模型。而且期间还有各种问题，搜了很多github上的问答才解决，耗时耗力，这里就不记录了，很不推荐这种方式。
免费算力服务器
打开modelScope社区后，点击登录注册可以看到免费赠送算力的活动
注册完成后在对应模型里可以看到，随时都能启用的服务器
这里CPU环境的服务器勉强可以跑起来模型，但运行效果感人，而且配置过程中有各种问题需要修改，而GPU环境启动模型可以说是非常流畅，体验效果也很好
CPU环境启动
社区提供的服务器配置已经很高了，8核32G，但因为是纯CPU环境，启动过程中还是有些问题
安装依赖包
第一行命令不需要运行，服务器已经自带了modelscope包
只需要新建一个Terminal窗口来执行第二条命令
启动代码
直接运行文档提供的代码会报错，这里是因为纯CPU环境导致的
错误 1
RuntimeError: “addmm_impl_cpu_” not implemented for ‘Half’Hide Error Details
ini复制代码RuntimeError: "addmm_impl_cpu_" not implemented for 'Half'
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
Cell In[1], line 8
5 model = AutoModelForCausalLM.from_pretrained("qwen/Qwen-7B-Chat", revision = 'v1.0.5',device_map="auto", trust_remote_code=True,fp16 = True).eval()
6 model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat",revision = 'v1.0.5', trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
----> 8 response, history = model.chat(tokenizer, "你好", history=None)
9 print(response)
10 response, history = model.chat(tokenizer, "浙江的省会在哪里？", history=history)
File ~/.cache/huggingface/modules/transformers_modules/Qwen-7B-Chat/modeling_qwen.py:1010, in QWenLMHeadModel.chat(self, tokenizer, query, history, system, append_history, stream, stop_words_ids, **kwargs)
1006 stop_words_ids.extend(get_stop_words_ids(
1007 self.generation_config.chat_format, tokenizer
1008 ))
1009 input_ids = torch.tensor([context_tokens]).to(self.device)
-> 1010 outputs = self.generate(
1011 input_ids,
1012 stop_words_ids = stop_words_ids,
1013 return_dict_in_generate = False,
1014 **kwargs,
1015 )
1017 response = decode_tokens(
1018 outputs[0],
1019 tokenizer,
(...)
1024 errors='replace'
1025 )
1027 if append_history:
错误 2
ValueError: The current device_map had weights offloaded to the disk. Please provide an offload_folder for them. Alternatively, make sure you have safetensors installed if the model you are using offers the weights in this format.Hide Error Details
ini复制代码ValueError: The current `device_map` had weights offloaded to the disk. Please provide an `offload_folder` for them. Alternatively, make sure you have `safetensors` installed if the model you are using offers the weights in this format.
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[2], line 5
2 from modelscope import GenerationConfig
4 tokenizer = AutoTokenizer.from_pretrained("qwen/Qwen-7B-Chat", revision = 'v1.0.5',trust_remote_code=True)
----> 5 model = AutoModelForCausalLM.from_pretrained("qwen/Qwen-7B-Chat", revision = 'v1.0.5',device_map="auto", trust_remote_code=True,fp16 = True).eval()
6 model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat",revision = 'v1.0.5', trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
7 model.float()
File /opt/conda/lib/python3.8/site-packages/modelscope/utils/hf_util.py:98, in get_wrapped_class.<locals>.ClassWrapper.from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs)
95 else:
96 model_dir = pretrained_model_name_or_path
---> 98 model = module_class.from_pretrained(model_dir, *model_args,
99 **kwargs)
100 model.model_dir = model_dir
101 return model
解决方式
首先确保torch 2.0.1版本，然后在代码中添加这两行，即可运行
model.float()
offload_folder=”offload_folder”,
python复制代码from modelscope import AutoModelForCausalLM, AutoTokenizer
from modelscope import GenerationConfig
import datetime
print("启动时间：" + str(datetime.datetime.now()))
tokenizer = AutoTokenizer.from_pretrained("qwen/Qwen-7B-Chat", revision = 'v1.0.5',trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("qwen/Qwen-7B-Chat", revision = 'v1.0.5',device_map="auto",offload_folder="offload_folder", trust_remote_code=True,fp16 = True).eval()
model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat",revision = 'v1.0.5', trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
model.float()
print("开始执行：" + str(datetime.datetime.now()))
response, history = model.chat(tokenizer, "你好", history=None)
print(response)
print("第一个问题处理完毕：" + str(datetime.datetime.now()))
response, history = model.chat(tokenizer, "浙江的省会在哪里？", history=history)
print(response)
print("第二个问题处理完毕：" + str(datetime.datetime.now()))
response, history = model.chat(tokenizer, "它有什么好玩的景点", history=history)
print(response)
print("第三个问题处理完毕：" + str(datetime.datetime.now()))
运行起来之后速度实在感人，每回答一个问题都需要 5 分钟左右，还有一定概率直接启动失败
启动模型过程中会出现这种报错，点击OK重新执行就好了，可能是服务器负载太高
