火山引擎GPU云服务器指南-部署基于DeepSpeed-Chat的行业大模型

export CUDA_HOME=/usr/local/cuda-11.4export PATH=$PATH:$CUDA_HOME/binexport LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CUDA_HOME/lib64

git clone https://github.com/GongFuXiong/Chinese-Medical-Question-Answering-Systemcd Chinese-Medical-Question-Answering-System/ChineseMed_QaData# tree.├── 1.txt├── answers.csv                # 回答文件├── dic4.txt├── model.txt└── questions.csv              # 问题文件

{"que_id":43823757,"ques_content":"头痛，低烧，检查是新型隐球菌脑膜炎？头痛，低烧，检查是新型隐球菌脑膜炎。","answers":[{"ans_content":"口服抗病毒药物和感冒冲剂试试.有炎症还是应该加上抗生素。口服药物不见效的，建议输液治疗为好。在当地医生指导下使用。有痰的加上鲜竹沥口服试试。如果发烧还需要适当加上额外的退烧药物。"}]},{"que_id":33890331,"ques_content":"刚出生的小孩感冒了，鼻子不通怎么办？","answers":[{"ans_content":"你好，刚出生的小孩感冒了，鼻子不通，多与受寒有关。服用小儿氨酚黄那敏颗粒等等看看，注意体温变化，有发热时，要及时就医。"}]},

class MedicalConsultationDataset(PromptRawDataset):def __init__(self, output_path, seed, local_rank, dataset_name):super().__init__(output_path, seed, local_rank, dataset_name)self.dataset_name = "jisoul_medical_consultation"self.dataset_name_clean = "jisoul_medical_consultation"def get_train_data(self):from .data_utils import get_raw_dataset_split_indexprint(self.raw_datasets)dataset = self.raw_datasets["train"]index = get_raw_dataset_split_index(self.local_rank, self.output_path,self.dataset_name_clean,self.seed, "train_eval", "10,1", 0,len(dataset))dataset = Subset(dataset, index)return datasetdef get_eval_data(self):from .data_utils import get_raw_dataset_split_indexdataset = self.raw_datasets["test"]index = get_raw_dataset_split_index(self.local_rank, self.output_path,self.dataset_name_clean,self.seed, "train_eval", "10,1", 1,len(dataset))dataset = Subset(dataset, index)return datasetdef get_prompt(self, sample):if sample['ques_content'] is not None:return " Human: " + sample['ques_content'] + " Assistant:"return Nonedef get_chosen(self, sample):if sample['answers'][0]['ans_content'] is not None:return " " + sample['answers'][0]['ans_content']return Nonedef get_rejected(self, sample):print(f"Warning: dataset {self.dataset_name} does not include rejected response.")return Nonedef get_prompt_and_chosen(self, sample):if sample['ques_content'] is not None and sample['answers'] is not None:if sample['ques_content'] is not None and sample['answers'][0]['ans_content'] is not None:return " Human: " + sample['ques_content'] + " Assistant: " + sample['answers'][0]['ans_content']return Nonedef get_prompt_and_rejected(self, sample):print(f"Warning: dataset {self.dataset_name} does not include rejected response.")return None

elif "jisoul/medical_consultation" in dataset_name:return raw_datasets.MedicalConsultationDataset(output_path, seed, local_rank,dataset_name)

#!/bin/bash# Copyright (c) Microsoft Corporation.# SPDX-License-Identifier: Apache-2.0# DeepSpeed Team# Note that usually LoRA needs to use larger learning rateOUTPUT=$1ZERO_STAGE=$2if [ "$OUTPUT" == "" ]; thenOUTPUT=./output/medical_consultationfiif [ "$ZERO_STAGE" == "" ]; thenZERO_STAGE=0fimkdir -p $OUTPUTdeepspeed --num_gpus 1 main.py \--data_path jisoul/medical_consultation \--model_name_or_path facebook/opt-125m \--gradient_accumulation_steps 4 \--zero_stage $ZERO_STAGE \--per_device_eval_batch_size 4 \--enable_tensorboard \--tensorboard_path $OUTPUT \--deepspeed --output_dir $OUTPUT

--language Chinese \--model_name_or_path_baseline facebook/opt-125m \--model_name_or_path_finetune /root/DeepSpeedExamples/applications/DeepSpeed-Chat/training/step1_supervised_finetuning/output/medical_consultation/

pip3 list | grep gradiopip3 list | grep transformers

import gradio as grimport torchfrom transformers import AutoModelForCausalLM, AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamerfrom threading import Threadtokenizer = AutoTokenizer.from_pretrained("/root/DeepSpeedExamples/applications/DeepSpeed-Chat/training/step1_supervised_finetuning/output/medical_consultation/")model = AutoModelForCausalLM.from_pretrained("/root/DeepSpeedExamples/applications/DeepSpeed-Chat/training/step1_supervised_finetuning/output/medical_consultation/")model = model.to('cuda:0')class StopOnTokens(StoppingCriteria):def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:stop_ids = [29, 0]for stop_id in stop_ids:if input_ids[0][-1] == stop_id:return Truereturn Falsedef predict(message, history):history_transformer_format = history + [[message, ""]]stop = StopOnTokens()messages = "".join(["".join(["\n<human>:"+item[0], "\n<bot>:"+item[1]])  #curr_system_message +for item in history_transformer_format])model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)generate_kwargs = dict(model_inputs,streamer=streamer,max_new_tokens=1024,do_sample=True,top_p=0.95,top_k=1000,temperature=1.0,num_beams=1,stopping_criteria=StoppingCriteriaList([stop]))t = Thread(target=model.generate, kwargs=generate_kwargs)t.start()partial_message  = ""for new_token in streamer:if new_token != '<':partial_message += new_tokenyield partial_messagegr.ChatInterface(predict).queue().launch(share=True)