Qwen大模型微调

戊辰happy

已于 2024-09-23 14:33:41 修改

阅读量418

点赞数 8

文章标签： python 深度学习人工智能

于 2024-08-20 17:23:51 首次发布

本文链接：https://blog.csdn.net/weixin_42569775/article/details/141362820

版权

1.Qwen基本问答demo2

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
model_name = "Qwen/Qwen2-7B-Instruct"
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "是否需要预约才能拜访楼上的公司？"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print("response:", response)

2.微调

下载LLaMA-Factory

python 3.11 CUDA 12.1

cd LLaMA-Factory
pip install -e ".[torch,metrics]"

Use pip install --no-deps -e . to resolve package conflicts.

python src/webui.py

根据自己要微调的模型和数据进行选择

微调数据格式

[
  {
    "instruction": "user instruction (required)",
    "input": "user input (optional)",
    "output": "model response (required)",
    "system": "system prompt (optional)",
    "history": [
      ["user instruction in the first round (optional)", "model response in the first round (optional)"],
      ["user instruction in the second round (optional)", "model response in the second round (optional)"]
    ]
  }
]

数据格式及代码

import pandas as pd
import json

# 读取 Excel 文件
excel_file_path = 'C:\\Users\\Administrator\\Desktop\\知识库V0.1-英文.xlsx'
df = pd.read_excel(excel_file_path)

# 假设 Excel 文件有两列：'Question' 和 'Answer'
# 如果列名不同，请相应地修改
questions = df['Question']
answers = df['Answer']

# 转换为 alpaca 格式
alpaca_data = []
for question, answer in zip(questions, answers):
    alpaca_item = {
        "instruction": question,
        "input": "",
        "output": answer,
        "system": "",
        "history": []
    }
    alpaca_data.append(alpaca_item)

# 将结果写入 JSON 文件
json_file_path = 'C:\\Users\\Administrator\\Desktop\\data.json'
with open(json_file_path, 'w', encoding='utf-8') as f:
    json.dump(alpaca_data, f, ensure_ascii=False, indent=4)

print(f"转换完成，结果已保存到 {json_file_path}")

在data/dataset_info.json添加自己的数据新增自己的数据，在webui界面可选训练数据集

“our_data": {
"file_name": "data.json"
},

也可直接运行以下命令训练

llamafactory-cli train     --stage sft     --do_train True     --model_name_or_path Qwen/Qwen1.5-7B-Chat     --preprocessing_num_workers 16     --finetuning_type lora     --template qwen     --flash_attn auto     --dataset_dir data     --dataset our_data     --cutoff_len 1024     --learning_rate 5e-05     --num_train_epochs 1000.0     --max_samples 100000     --per_device_train_batch_size 2     --gradient_accumulation_steps 8     --lr_scheduler_type cosine     --max_grad_norm 1.0     --logging_steps 5     --save_steps 100     --warmup_steps 0     --optim adamw_torch     --packing False     --report_to none     --output_dir saves/Qwen1.5-7B-Chat/lora/train_2024-08-16-17-21-38     --bf16 True     --plot_loss True     --ddp_timeout 180000000     --include_num_input_tokens_seen True     --lora_rank 8     --lora_alpha 16     --lora_dropout 0     --lora_target all

微调后测试，可使用基座大模型和微调lora模型

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
model_name = "Qwen/Qwen2-7B-Instruct"
#model_name = "/media/dgh/LLaMA-Factory/saves/Qwen2-7B-Chat/lora/train_2024-08-10-14-15-57/checkpoint-4400"
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

model = PeftModel.from_pretrained(model, model_id = "/media/dgh/LLaMA-Factory/saves/Qwen2-7B-Chat/lora/train_2024-08-10-14-15-57/checkpoint-4400")

prompt = "是否需要预约才能拜访楼上的公司？"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print("response:", response)

模型合并同时也可运行合并后的模型测试

CUDA_VISIBLE_DEVICES=0 llamafactory-cli export \
    --model_name_or_path  Qwen/Qwen1.5-7B-Chat\
    --adapter_name_or_path /media/dgh/LLaMA-Factory/saves/Qwen1.5-7B-Chat/lora/train_2024-08-16-17-21-38/checkpoint-7300 \
    --template qwen \
    --finetuning_type lora \
    --export_dir /media/dgh/Qwen2-main/save \
    --export_size 2 \
    --export_legacy_format False

编译llama.cpp

git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make

guuf格式生成

python convert_hf_to_gguf.py /media/dgh/Qwen2-main/save --outfile /media/dgh/Qwen2-main/7b_guuf/qwen2-7b-instruct-fp16.gguf

量化GGUF模型： q2_k 、 q3_k_m 、 q4_0 、 q4_k_m 、 q5_0 、 q5_k_m 、 q6_k 和 q8_0 。了解更多信息，请访问 llama.cpp 。

q4_0

./llama-quantize /home/dgh/LLaMA-Factory/saves/Qwen1.5-14b_guuf/Qwen1.5-14B-Chat-F16.gguf /home/dgh/LLaMA-Factory/saves/qwen1.5-14b-q4_0_gguf/qwen1.5-14b-q4_0.gguf q4_0

q5_k_m

./llama-quantize /home/dgh/LLaMA-Factory/saves/Qwen1.5-14b_guuf/Qwen1.5-14B-Chat-F16.gguf /home/dgh/LLaMA-Factory/saves/qwen1.5-14b-q5_k_m_gguf/qwen1.5-14b-q5_k_m.gguf q5_k_m

run

./llama-cli -m /home/dgh/LLaMA-Factory/saves/qwen1.5-14b-q5_k_m_gguf/qwen1.5-14b-q5_k_m.gguf \
-n 512 -co -i -if -f prompts/chat-with-qwen.txt \
--in-prefix "<|im_start|>user\n" \
--in-suffix "<|im_end|>\n<|im_start|>assistant\n" \
-ngl 80 -fa

阿里百炼在线大模型调用大模型服务平台百炼

在阿里百炼可以零代码创建自己的智能体，其中调用API如下：

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    base_url="https://dashscope.aliyuncs.com/compatible-mode/v1",
)
completion = client.chat.completions.create(
    model="qwen-turbo",
    messages=[
        {'role': 'system', 'content': 'You are a helpful assistant.'},
        {'role': 'user', 'content': '你是谁？'}
    ],
    temperature=0.8
)

print(completion.choices[0].message.content)

创建自己的智能体也可根据给定API调用，具体参考官方使用指导教程

下载最新的LLaMA-Factory 支持微调Qwen-VL

遇到报错：

ValueError: The checkpoint you are trying to load has model type `qwen2_vl` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date

原因：pip中transformers还没有更新，用github上的版本更新
解决方法：pip install git+https://github.com/huggingface/transformers

采用swift框架微调Qwen2-VL

参考Qwen2-VL-阿里云开发者社区

采用的是python 3.8。CUDA 12.1

git clone https://github.com/modelscope/swift.git
cd swift
pip install -e .[llm]
pip install pyav qwen_vl_utils

采用官方指定数据集训练

CUDA_VISIBLE_DEVICES=0,1,2,3 NPROC_PER_NODE=4 swift sft \
  --model_type qwen2-vl-7b-instruct \
  --model_id_or_path qwen/Qwen2-VL-7B-Instruct \
  --sft_type lora \
  --dataset coco-en-mini#20000 \
  --deepspeed default-zero2

采用自己的数据集训练

  --dataset train.jsonl \
  --val_dataset val.jsonl \

数据集格式

{"query": "<image>55555", "response": "66666", "images": ["image_path"]}
{"query": "eeeee<image>eeeee<image>eeeee", "response": "fffff", "history": [], "images": ["image_path1", "image_path2"]}
{"query": "EEEEE", "response": "FFFFF", "history": [["query1", "response2"], ["query2", "response2"]], "images": []}

微调后推理，并合并模型

CUDA_VISIBLE_DEVICES=0 swift infer \
    --ckpt_dir output/qwen2-vl-7b-instruct/vx-xxx/checkpoint-xxx \
    --load_dataset_config true --merge_lora true

合并之后的模型可直接调用

from PIL import Image
import torch
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info
from modelscope import snapshot_download
model_dir = "/Qwen2-VL-2B-Instruct/output/v4-20240923/checkpoint-1000-merged"
# Load the model in half-precision on the available device(s)
model = Qwen2VLForConditionalGeneration.from_pretrained(model_dir, device_map="auto", torch_dtype = torch.float16)
min_pixels = 256*28*28
max_pixels = 1280*28*28
processor = AutoProcessor.from_pretrained(model_dir, min_pixels=min_pixels, max_pixels=max_pixels)
messages = [{"role": "user", "content": [{"type": "image", "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"}, {"type": "text", "text": "Describe this image."}]}]
# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt")
inputs = inputs.to('cuda')
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(output_text)

多图推理

# Messages containing multiple images and a text query
messages = [{"role": "user", "content": [{"type": "image", "image": "file:///path/to/image1.jpg"}, {"type": "image", "image": "file:///path/to/image2.jpg"}, {"type": "text", "text": "Identify the similarities between these images."}]}]
# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt")
inputs = inputs.to('cuda')
# Inference
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(output_text)

视频理解

# Messages containing a video and a text query
messages = [{"role": "user", "content": [{"type": "video", "video": "file:///path/to/video1.mp4", 'max_pixels': 360*420, 'fps': 1.0}, {"type": "text", "text": "Describe this video."}]}]
# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(text=[text], images=image_inputs, videos=video_inputs, padding=True, return_tensors="pt")
inputs = inputs.to('cuda')
# Inference
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
output_text = processor.batch_decode(generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)
print(output_text)