一、基于baichuan 7b模型进行sft,对其人类意图

二、sft数据是在开源MOSS数据中通过各个类别均衡采样15w数据进行sft

模型推理

Install package:

pip install transformers
pip install sentencepiece
pip install vllm

huggingface结合fastapi起服务,支持多轮对话

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
import uvicorn
from fastapi import FastAPI
import jsonlines

device = 'cuda'
model_name = 'mxmax/baichuan-7b-sft-001'
max_new_tokens = 500
top_p = 0.9
temperature = 0.35
repetition_penalty = 1.0
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    device_map={'': 0}#'auto'
).cuda()
# model = PeftModel.from_pretrained(model, adapter_name)
model.eval()
model = model.to(device)
# 输入模型的最大长度
history_max_len = 1024
def model_infer(user_input):
    history_token_ids = tokenizer('<s>', return_tensors="pt").input_ids
    user_input_ids = tokenizer(user_input, return_tensors="pt").input_ids
    history_token_ids = torch.concat((history_token_ids, user_input_ids[:, -history_max_len:]), dim=1)
    model_input_ids = history_token_ids.to(device)
    outputs = model.generate(
        input_ids=model_input_ids, max_new_tokens=max_new_tokens, do_sample=True, top_p=top_p,
        temperature=temperature, repetition_penalty=repetition_penalty, eos_token_id=tokenizer.eos_token_id
    )
    model_input_ids_len = model_input_ids.size(1)
    response_ids = outputs[:, model_input_ids_len:]
    response = tokenizer.batch_decode(response_ids)
    return response[0].strip().replace('</s>', "")

app = FastAPI()

@app.get('/')
async def root():
    return {"msg": "Hello World"}

@app.post('/baichuan_sft_001')
async def baichuan_sft_001(message: dict):
    prompt = ''
    for l in message['context']:
        prompt += 'human:'+l['human']+'\nassistant:'+l['assistant']+'</s>'
    result = model_infer(prompt)
    message['context'][-1]['assistant'] = result
    return {'model_ouput':result}

if __name__ == '__main__':
    uvicorn.run('model_serving:app',host="0.0.0.0", port=6006)

vllm结合fastapi起服务,加速推理,支持多轮对话

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
import uvicorn
from fastapi import FastAPI
import jsonlines
from vllm import LLM, SamplingParams

device = 'cuda'
model_name = 'mxmax/baichuan-7b-sft-001'
max_new_tokens = 512
top_p = 0.9
temperature = 0.35
repetition_penalty = 0.1
history_max_len = 1024
sampling_params = SamplingParams(temperature=temperature, top_p=top_p, max_tokens=max_new_tokens, presence_penalty=repetition_penalty)

# Create an LLM.
llm = LLM(model=model_name,trust_remote_code=True,dtype='float16')
file = jsonlines.open('chat_record.json','a')
app = FastAPI()

@app.get('/')
async def root():
    return {"msg": "Hello World"}

@app.post('/baichuan_sft_001')
async def baichuan_sft_001(message: dict):
    prompt = ''
    for l in message['context']:
        prompt += 'human:'+l['human']+'\nassistant:'+l['assistant']+'</s>'
    prompt = '<s>'+prompt[-history_max_len:]
    outputs = llm.generate([prompt], sampling_params)
    result = outputs[0].outputs[0].text
    message['context'][-1]['assistant'] = result
    return {'model_ouput':result}

if __name__ == '__main__':
    uvicorn.run('vllm_serving:app',host="0.0.0.0", port=6006)

模型效果展示

arch arch arch arch arch arch

联系方式

arch 加好友请备注:来自于huggingface网站交流技术+名字

qq群:621725172

引用

@misc{mxmax,
  title={baichuan_sft: baichuan-7b-sft-001},
  author={Ma Xin},
  year={2023},
  howpublished={\url{https://huggingface.co/mxmax/baichuan-7b-sft-001}},
}
Downloads last month
15
Inference Examples
Inference API (serverless) does not yet support model repos that contain custom code.