weiweiw
/
Langchain-Chatchat
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

fix: 修复chatglm模型被复制，显存占用过多 

model作为类成员会在类实例化时进行一次复制
这导致每询问一个问题显存占用就会翻倍
通过将model改成全局变量修复这个问题
This commit is contained in:
myml 2023-04-05 01:05:06 +08:00
parent 1851b4f074
commit bed03a6ff1
1 changed files with 4 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
chatglm_llm.py
View File

@ -5,24 +5,14 @@ from transformers import AutoTokenizer, AutoModel
"""ChatGLM_G is a wrapper around the ChatGLM model to fit LangChain framework. May not be an optimal implementation""" """ChatGLM_G is a wrapper around the ChatGLM model to fit LangChain framework. May not be an optimal implementation"""
tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
model = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
class ChatGLM(LLM): class ChatGLM(LLM):
max_token: int = 10000 max_token: int = 10000
temperature: float = 0.1 temperature: float = 0.1
top_p = 0.9 top_p = 0.9
history = [] history = []
tokenizer = AutoTokenizer.from_pretrained(
"THUDM/chatglm-6b",
trust_remote_code=True
)
model = (
AutoModel.from_pretrained(
"THUDM/chatglm-6b",
trust_remote_code=True)
.half()
.cuda()
)
def __init__(self): def __init__(self):
super().__init__() super().__init__()
@ -34,8 +24,8 @@ class ChatGLM(LLM):
def _call(self, def _call(self,
prompt: str, prompt: str,
stop: Optional[List[str]] = None) -> str: stop: Optional[List[str]] = None) -> str:
response, updated_history = self.model.chat( response, updated_history = model.chat(
self.tokenizer, tokenizer,
prompt, prompt,
history=self.history, history=self.history,
max_length=self.max_token, max_length=self.max_token,