python -m pip install -U alipai
# 配置鉴权密钥，PAI工作空间等信息
python -m pai.toolkit.config

from pai.common.utils import random_str
from pai.session import get_default_session
from pai.modelscope import ModelScopeModel
sess = get_default_session()
# 创建 ModelScopeModel 对象。
model = ModelScopeModel(
    # 推理服务启动命令
    command=(
        "python webui/webui_server.py --model-path=Qwen/Qwen1.5-7B-Chat "
        "--model-type=qwen2 --backend=vllm "
        # vLLM engine arguments
        " --max-model-len=4096 --gpu-memory-utilization=0.95"
    ),
    # 推理服务镜像
    image_uri=f"eas-registry-vpc.{sess.region_id}.cr.aliyuncs.com/pai-eas/chat-llm-webui:3.0.3-vllm",
)
# 部署创建推理服务
predictor = model.deploy(
    # 服务名称
    service_name=f"ms_serving_{random_str(6)}",
    # 服务使用的机型规格
    instance_type="ecs.gn7i-c8g1.2xlarge",
    options={
        "metadata.enable_webservice": True,
        "metadata.rpc.keepalive": 300000,
        "features.eas.aliyun.com/extra-ephemeral-storage": "30Gi",
    }
)

# 大语言模型推理服务支持OpenAI API风格
openai_api = predictor.openai()
# 通过OpenAI SDK调用推理服务
resp = openai_api.chat.completions.create(
    model="default",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "What is the meaning of life?"},
    ],
    max_tokens=1024,
)
print(resp.choices[0].message.content)
# 注意测试完成后删除部署的服务，避免产生额外费用
# predictor.delete_service()

from pai.modelscope.estimator import ModelScopeEstimator
 
# 模型训练超参
hyperparameters = {
    "model_type": "qwen1half-0_5b-chat",
    "sft_type": "lora",
    "tuner_backend": "swift",
    # 模型输出地址，请勿修改，只有该地址下的输出文件才会被保存
    "output_dir": "/ml/output/model/",
    "dtype": "AUTO",
    "dataset": "ms-bench",
    "train_dataset_sample": "5000",
    "num_train_epochs": "2",
    "max_length": "1024",
    "check_dataset_strategy": "warning",
    "lora_rank": "8",
    "lora_alpha": "32",
    "lora_dropout_p": "0.05",
    "lora_target_modules": "ALL",
    "gradient_checkpointing": "True",
    "batch_size": "1",
    "weight_decay": "0.01",
    "learning_rate": "1e-4",
    "gradient_accumulation_steps": "16",
    "max_grad_norm": "0.5",
    "warmup_ratio": "0.03"
}
# 创建 ModelScopeEstimator 对象
est = ModelScopeEstimator(
    # 指定训练脚本的启动命令，通过 $PAI_USER_ARGS 传入所有超参信息，请参考：
    # https://help.aliyun.com/zh/pai/user-guide/submit-a-training-job
    command="swift sft $PAI_USER_ARGS",
    # 任务使用的机型规格
    instance_type="ecs.gn6e-c12g1.3xlarge",
    # 使用相应 modelscope_version 的训练镜像
    modelscope_version="1.13.3",
    hyperparameters=hyperparameters,
    base_job_name="modelscope-sdk-train",
    # 第三方依赖包
    requirements=["ms_swift==2.0.3.post1"],
)
# 提交创建一个训练任务（产生PAI-DLC账单费用）
est.fit(wait=False)
# 创建并打开一个TensorBoard实例
tb = est.tensorboard()
print(tb.app_uri)
# 等待训练作业执行完成
est.wait()
# 查看训练任务所产出的模型地址
print(est.model_data())