Axolotl微调Qwen2-7b开源大模型！AutoGen+Marker自动生成数据集！AutoGen Studio+GPT4评估Qwen2文章能力，vllm部署qwen

1.vllm部署

conda create -n myenv python=3.9 -y
conda activate myenv
pip install vllm

从huggingface下载模型并部署

python -m vllm.entrypoints.openai.api_server --model Qwen/Qwen2-7B --dtype auto --api-key token-abc123 --host 0.0.0.0 --trust-remote-code

#接口信息
        {
            "model": "Qwen/Qwen2-7B",
            "base_url": "http://64.247.196.36:8000/v1",
            "api_key": "token-abc123",
        },

 用本地模型部署

export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:86

python -m vllm.entrypoints.openai.api_server --model /home/cxh/Qwen-7B --dtype auto --api-key token-abc123 --host 0.0.0.0 --trust-remote-code --max_model_len=1024

import torch
torch.cuda.empty_cache()

如果是unsloth微调后的模型，用下面命令运行
python3 -m vllm.entrypoints.openai.api_server --model /home/cxh/Meta-Llama-3.1-8B-Instruct-bnb-4bit --load-format bitsandbytes --quantization bitsandbytes --enforce-eager --gpu-memory-utilization=0.85 --max_model_len=1024

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key='token-abc123',
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="/home/cxh/Meta-Llama-3.1-8B-Instruct-bnb-4bit",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)
content = chat_response.choices[0].message.content
print(content)

2.autogen studio 安装和启动

pip install autogenstudio

autogenstudio ui --port 8081 --host 0.0.0.0.

 导入模型刚才运行的模型
3.AutoGen+Marker生成数据集

 pip install marker-pdf -i https://pypi.tuna.tsinghua.edu.cn/simple

将pdf转换为md

marker_single GPT.pdf ./folder --batch_multiplier 2 --max_pages 52 --langs English

pip install pyautogen

pip install autogen[graph]

生成数据集

https://colab.research.google.com/drive/1xEqhuVUC89ZCeOn82mSo6qQfqc40xznD

4.用Axolotl 和生成的数据集微调qwen

# 从 GitHub 克隆 axolotl 仓库
git clone https://github.com/OpenAccess-AI-Collective/axolotl

# 切换到 axolotl 目录
cd axolotl

docker run --gpus '"all"' --rm -it winglian/axolotl:main-latest


# 使用 Accelerate 库启动 axolotl 训练脚本,使用 examples/openllama-3b/qlora.yml 配置文件
accelerate launch -m axolotl.cli.train examples/qwen2/qlora-fsdp.yaml

注意:

配置文件中fp16设为false bfp16设为true

如果数据集内容太少，需要将eval_sample_packing 设置为 False

qlora-fsdp.yaml配置

base_model: Qwen/Qwen2-7B
trust_remote_code: true
load_in_8bit: false
load_in_4bit: true
strict: false
datasets:
  - path: tatsu-lab/alpaca
    type: alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/out
sequence_len: 2048
sample_packing: true
eval_sample_packing: true
pad_to_sequence_len: true
adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.0002
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true
gradient_checkpointing: false
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: false
warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
special_tokens: