小小林熬夜学编程

这个屌丝很懒，什么也没留下！

热门标签

ChuanhuChatGPT集成百川大模型

作者：小小林熬夜学编程 | 2024-04-28 03:18:12

踩

搭建步骤：

拷贝本地模型，把下载好的Baichuan2-7B-Chat拷贝到models目录下
修改modules\models\base_model.py文件，class ModelType增加Baichuan

Baichuan = 16

elif "baichuan" in model_name_lower:

model_type = ModelType.Baichuan
修改modules\models\models.py文件，get_model方法增加ModelType.Baichuan

elif model_type == ModelType.Baichuan:

from .Baichuan import Baichuan_Client

model = Baichuan_Client(model_name, user_name=user_name)

增加modules\models\Baichuan.py文件

from modelscope import snapshot_download, AutoModelForCausalLM, AutoTokenizer,GenerationConfig

from transformers import AutoModelForCausalLM, AutoTokenizer

from transformers.generation import GenerationConfig

import logging

import colorama

from ..index_func import *

from ..presets import *

from ..utils import *

from .base_model import BaseLLMModel

from ..presets import MODEL_METADATA

from datetime import datetime

class Baichuan_Client(BaseLLMModel):

def __init__(self, model_name, user_name="") -> None:

super().__init__(model_name=model_name, user=user_name)

import torch

from transformers import AutoModel, AutoTokenizer

global CHATGLM_TOKENIZER, CHATGLM_MODEL

print("__init__ Baichuan_Client")

if CHATGLM_TOKENIZER is None or CHATGLM_MODEL is None:

model_path = None

if os.path.exists("models"):

model_dirs = os.listdir("models")

if model_name in model_dirs:

model_path = f"models/{model_name}"

if model_path is not None:

model_source = model_path

else:

model_source = snapshot_download(f"baichuan-inc/{model_name}", revision='v1.0.4')

CHATGLM_TOKENIZER = AutoTokenizer.from_pretrained(

model_source, device_map="auto", trust_remote_code=True, torch_dtype=torch.float16

)

quantified = False

if "int4" in model_name:

quantified = True

model = AutoModelForCausalLM.from_pretrained(

model_source, device_map="auto", trust_remote_code=True, torch_dtype=torch.float16

)

model.generation_config = GenerationConfig.from_pretrained(model_source)

model = model.eval()

CHATGLM_MODEL = model

def _get_glm_style_input(self):

print("_get_glm_style_input")

print(f"the history is: {self.history}")

history = [x["content"] for x in self.history]

query = history.pop()

print(f"the message is: {query}")

return history, query

def get_answer_at_once(self):

print("get_answer_at_once")

history,query = self._get_glm_style_input()

messages = []

messages.append({'role': 'user', 'content': query})

now = datetime.now()

print("get_answer_at_once start"+"++++++++"+ now.strftime("%Y-%m-%d %H:%M:%S"))

response = CHATGLM_MODEL.chat(

CHATGLM_TOKENIZER, messages)

now = datetime.now()

print("get_answer_at_once end"+"++++++++"+ now.strftime("%Y-%m-%d %H:%M:%S"))

print(f"the response is: {response}")

return response, len(response)

def get_answer_stream_iter(self):

history,query = self._get_glm_style_input()

messages = []

messages.append({'role': 'user', 'content': query})

result = ""

now = datetime.now()

print("get_answer_stream_iter start"+"++++++++"+ now.strftime("%Y-%m-%d %H:%M:%S"))

for response in CHATGLM_MODEL.chat(

CHATGLM_TOKENIZER,

messages

):

print(f"the response is: {response}")

result += response

yield result

now = datetime.now()

print("get_answer_stream_iter end"+"++++++++"+ now.strftime("%Y-%m-%d %H:%M:%S"))

答案回调开关控制get_answer_at_once、get_answer_stream_iter方法调用选择
执行效果

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/小小林熬夜学编程/article/detail/500164