mirror of https://github.com/THUDM/ChatGLM-6B
46 lines
1.4 KiB
Python
46 lines
1.4 KiB
Python
from langchain.llms.base import LLM
|
|
from typing import Optional, List, Mapping, Any
|
|
from langchain.llms.utils import enforce_stop_tokens
|
|
from transformers import AutoTokenizer, AutoModel
|
|
|
|
"""ChatGLM_G is a wrapper around the ChatGLM model to fit LangChain framework. May not be an optimal implementation"""
|
|
|
|
|
|
class ChatGLM_G(LLM):
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
|
"THUDM/chatglm-6b-int4", trust_remote_code=True
|
|
)
|
|
model = (
|
|
AutoModel.from_pretrained("THUDM/chatglm-6b-int4", trust_remote_code=True)
|
|
.half()
|
|
.cuda()
|
|
)
|
|
history = []
|
|
|
|
@property
|
|
def _llm_type(self) -> str:
|
|
return "ChatGLM_G"
|
|
|
|
def _call(self, prompt: str, stop: Optional[List[str]] = None) -> str:
|
|
response, updated_history = self.model.chat(
|
|
self.tokenizer, prompt, history=self.history, max_length=10000
|
|
)
|
|
print("history: ", self.history)
|
|
if stop is not None:
|
|
response = enforce_stop_tokens(response, stop)
|
|
self.history = updated_history
|
|
return response
|
|
|
|
def __call__(self, prompt: str, stop: Optional[List[str]] = None) -> str:
|
|
response, updated_history = self.model.chat(
|
|
self.tokenizer, prompt, history=self.history, max_length=10000
|
|
)
|
|
print("history: ", self.history)
|
|
|
|
if stop is not None:
|
|
response = enforce_stop_tokens(response, stop)
|
|
self.history = updated_history
|
|
|
|
return response
|