ChatGLM-6B/example_with_langchain_and_.../chatglm_llm.py

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