You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
ColossalAI/applications/ColossalQA/colossalqa/local/llm.py

196 lines
6.3 KiB

"""
API and LLM warpper class for running LLMs locally
Usage:
import os
model_path = os.environ.get("ZH_MODEL_PATH")
model_name = "chatglm2"
colossal_api = ColossalAPI(model_name, model_path)
llm = ColossalLLM(n=1, api=colossal_api)
TEST_PROMPT_CHATGLM="续写文章:惊蛰一过,春寒加剧。先是料料峭峭,继而雨季开始,"
logger.info(llm(TEST_PROMPT_CHATGLM, max_new_tokens=100), verbose=True)
"""
from typing import Any, List, Mapping, Optional
import torch
from colossalqa.local.utils import get_response, post_http_request
from colossalqa.mylogging import get_logger
from langchain.callbacks.manager import CallbackManagerForLLMRun
from langchain.llms.base import LLM
from transformers import AutoModelForCausalLM, AutoTokenizer
logger = get_logger()
class ColossalAPI:
"""
API for calling LLM.generate
"""
__instances = dict()
def __init__(self, model_type: str, model_path: str, ckpt_path: str = None) -> None:
"""
Configure model
"""
if model_type + model_path + (ckpt_path or "") in ColossalAPI.__instances:
return
else:
ColossalAPI.__instances[model_type + model_path + (ckpt_path or "")] = self
self.model_type = model_type
self.model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype=torch.float16, trust_remote_code=True)
if ckpt_path is not None:
state_dict = torch.load(ckpt_path)
self.model.load_state_dict(state_dict)
self.model.to(torch.cuda.current_device())
# Configure tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
self.model.eval()
@staticmethod
def get_api(model_type: str, model_path: str, ckpt_path: str = None):
if model_type + model_path + (ckpt_path or "") in ColossalAPI.__instances:
return ColossalAPI.__instances[model_type + model_path + (ckpt_path or "")]
else:
return ColossalAPI(model_type, model_path, ckpt_path)
def generate(self, input: str, **kwargs) -> str:
"""
Generate response given the prompt
Args:
input: input string
**kwargs: language model keyword type arguments, such as top_k, top_p, temperature, max_new_tokens...
Returns:
output: output string
"""
if self.model_type in ["chatglm", "chatglm2"]:
inputs = {
k: v.to(torch.cuda.current_device()) for k, v in self.tokenizer(input, return_tensors="pt").items()
}
else:
inputs = {
"input_ids": self.tokenizer(input, return_tensors="pt")["input_ids"].to(torch.cuda.current_device())
}
output = self.model.generate(**inputs, **kwargs)
output = output.cpu()
prompt_len = inputs["input_ids"].size(1)
response = output[0, prompt_len:]
output = self.tokenizer.decode(response, skip_special_tokens=True)
return output
class VllmAPI:
def __init__(self, host: str = "localhost", port: int = 8077) -> None:
# Configure api for model served through web
self.host = host
self.port = port
self.url = f"http://{self.host}:{self.port}/generate"
def generate(self, input: str, **kwargs):
output = get_response(post_http_request(input, self.url, **kwargs))[0]
return output[len(input) :]
class ColossalLLM(LLM):
"""
Langchain LLM wrapper for a local LLM
"""
n: int
api: Any
kwargs = {"max_new_tokens": 100}
@property
def _llm_type(self) -> str:
return "custom"
def _call(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
logger.info(f"kwargs:{kwargs}\nstop:{stop}\nprompt:{prompt}", verbose=self.verbose)
for k in self.kwargs:
if k not in kwargs:
kwargs[k] = self.kwargs[k]
generate_args = {k: kwargs[k] for k in kwargs if k not in ["stop", "n"]}
out = self.api.generate(prompt, **generate_args)
if isinstance(stop, list) and len(stop) != 0:
for stopping_words in stop:
if stopping_words in out:
out = out.split(stopping_words)[0]
logger.info(f"{prompt}{out}", verbose=self.verbose)
return out
@property
def _identifying_params(self) -> Mapping[str, int]:
"""Get the identifying parameters."""
return {"n": self.n}
def get_token_ids(self, text: str) -> List[int]:
"""Return the ordered ids of the tokens in a text.
Args:
text: The string input to tokenize.
Returns:
A list of ids corresponding to the tokens in the text, in order they occur
in the text.
"""
# use the colossal llm's tokenizer instead of langchain's cached GPT2 tokenizer
return self.api.tokenizer.encode(text)
class VllmLLM(LLM):
"""
Langchain LLM wrapper for a local LLM
"""
n: int
api: Any
kwargs = {"max_new_tokens": 100}
@property
def _llm_type(self) -> str:
return "custom"
def _call(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
for k in self.kwargs:
if k not in kwargs:
kwargs[k] = self.kwargs[k]
logger.info(f"kwargs:{kwargs}\nstop:{stop}\nprompt:{prompt}", verbose=self.verbose)
generate_args = {k: kwargs[k] for k in kwargs if k in ["n", "max_tokens", "temperature", "stream"]}
out = self.api.generate(prompt, **generate_args)
if len(stop) != 0:
for stopping_words in stop:
if stopping_words in out:
out = out.split(stopping_words)[0]
logger.info(f"{prompt}{out}", verbose=self.verbose)
return out
def set_host_port(self, host: str = "localhost", port: int = 8077, **kwargs) -> None:
if "max_tokens" not in kwargs:
kwargs["max_tokens"] = 100
self.kwargs = kwargs
self.api = VllmAPI(host=host, port=port)
@property
def _identifying_params(self) -> Mapping[str, int]:
"""Get the identifying parameters."""
return {"n": self.n}