如何使用 Cloud Run 作业微调 LLM

如何使用 Cloud Run 作业微调 LLM

关于此 Codelab

subject上次更新时间:6月 3, 2025
account_circleGoogle 员工编写

1. 简介

概览

在此 Codelab 中,您将使用 Cloud Run 作业微调 Gemma 模型,然后使用 vLLM 在 Cloud Run 上提供结果。

在本 Codelab 中,您将使用文本到 SQL 数据集,以便 LLM 在收到自然语言问题时以 SQL 查询的形式进行回答。

学习内容

  • 如何使用 Cloud Run 作业 GPU 进行微调
  • 如何使用 Cloud Run 搭配 vLLM 提供模型
  • 如何为 GPU 作业使用直接 VPC 配置,以更快地上传和分发模型

2. 准备工作

启用 API

在开始使用此 Codelab 之前,请通过运行以下命令启用以下 API:

gcloud services enable run.googleapis.com \
    compute.googleapis.com \
    run.googleapis.com \
    cloudbuild.googleapis.com \
    secretmanager.googleapis.com \
    artifactregistry.googleapis.com

GPU 配额

为受支持的区域申请增加配额。配额为 nvidia_l4_gpu_allocation_no_zonal_redundancy,位于 Cloud Run Admin API 下。

注意:如果您使用的是新项目,则可能需要几分钟时间,API 配额才会显示在此页面中。

Hugging Face

此 Codelab 使用托管在 Hugging Face 上的模型。如需获取此模型,请请求具有“读取”权限的 Hugging Face 用户访问令牌。在下文中,我们会用 YOUR_HF_TOKEN 来引用它。

您还需要同意使用条款才能使用该模型:https://huggingface.co/google/gemma-2b

3. 设置和要求

设置以下资源:

  • IAM 服务账号和关联的 IAM 权限,
  • 用于存储 Hugging Face 令牌的 Secret Manager Secret
  • Cloud Storage 存储分区(用于存储经过微调的模型),以及
  • Artifact Registry 代码库,用于存储您将构建的用于微调模型的映像。
  1. 为此 Codelab 设置环境变量。我们已为您预先填充了一些变量。指定您的项目 ID、区域和 Hugging Face 令牌。
    export PROJECT_ID=<YOUR_PROJECT_ID>
    export REGION=<YOUR_REGION>
    export HF_TOKEN=<YOUR_HF_TOKEN>

    export AR_REPO=codelab-finetuning-jobs
    export IMAGE_NAME=finetune-to-gcs
    export JOB_NAME=finetuning-to-gcs-job
    export BUCKET_NAME=$PROJECT_ID-codelab-finetuning-jobs
    export SECRET_ID=HF_TOKEN
    export SERVICE_ACCOUNT="finetune-job-sa"
    export SERVICE_ACCOUNT_ADDRESS=$SERVICE_ACCOUNT@$PROJECT_ID.iam.gserviceaccount.com
  2. 运行以下命令来创建服务账号:
    gcloud iam service-accounts create $SERVICE_ACCOUNT \
     
    --display-name="Service account for fine-tuning codelab"
  3. 使用 Secret Manager 存储 Hugging Face 访问令牌:
    gcloud secrets create $SECRET_ID \
         
    --replication-policy="automatic"

    printf $HF_TOKEN
    | gcloud secrets versions add $SECRET_ID --data-file=-
  4. 向您的服务账号授予 Secret Manager Secret Accessor 角色:
    gcloud secrets add-iam-policy-binding $SECRET_ID \
     
    --member serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
     
    --role='roles/secretmanager.secretAccessor'
  5. 创建一个用于托管经过微调的模型的存储分区:
    gcloud storage buckets create -l $REGION gs://$BUCKET_NAME
  6. 向您的服务账号授予对存储分区的访问权限:
    gcloud storage buckets add-iam-policy-binding gs://$BUCKET_NAME \
     
    --member=serviceAccount:$SERVICE_ACCOUNT_ADDRESS \
     
    --role=roles/storage.objectAdmin
  7. 创建 Artifact Registry 制品库以存储容器映像:
    gcloud artifacts repositories create $AR_REPO \
       
    --repository-format=docker \
       
    --location=$REGION \
       
    --description="codelab for finetuning using CR jobs" \
       
    --project=$PROJECT_ID

4. 创建 Cloud Run 作业映像

在下一步中,您将创建以下代码:

  • 从 Hugging Face 导入 Gemma 模型
  • 使用 Hugging Face 中的数据集对模型进行微调。该作业使用单个 L4 GPU 进行微调。
  • 将名为 new_model 的精调模型上传到您的 Cloud Storage 存储分区
  1. 为微调作业代码创建一个目录。
    mkdir codelab-finetuning-job
    cd codelab
    -finetuning-job
  2. 创建一个名为 finetune.py 的文件
    # Copyright 2024 Google LLC
    #
    # Licensed under the Apache License, Version 2.0 (the "License");
    # you may not use this file except in compliance with the License.
    # You may obtain a copy of the License at
    #
    #      http://www.apache.org/licenses/LICENSE-2.0
    #
    # Unless required by applicable law or agreed to in writing, software
    # distributed under the License is distributed on an "AS IS" BASIS,
    # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    # See the License for the specific language governing permissions and
    # limitations under the License.

    import os
    import torch
    from datasets import load_dataset
    from transformers import (
       
    AutoModelForCausalLM,
       
    AutoTokenizer,
       
    BitsAndBytesConfig,
       
    TrainingArguments,

    )
    from peft import LoraConfig, PeftModel

    from trl import SFTTrainer

    # Cloud Storage bucket to upload the model
    bucket_name = os.getenv("BUCKET_NAME", "YOUR_BUCKET_NAME")

    # The model that you want to train from the Hugging Face hub
    model_name = os.getenv("MODEL_NAME", "google/gemma-2b")

    # The instruction dataset to use
    dataset_name = "b-mc2/sql-create-context"

    # Fine-tuned model name
    new_model = os.getenv("NEW_MODEL", "gemma-2b-sql")

    ################################################################################
    # QLoRA parameters
    ################################################################################

    # LoRA attention dimension
    lora_r = int(os.getenv("LORA_R", "4"))

    # Alpha parameter for LoRA scaling
    lora_alpha = int(os.getenv("LORA_ALPHA", "8"))

    # Dropout probability for LoRA layers
    lora_dropout = 0.1

    ################################################################################
    # bitsandbytes parameters
    ################################################################################

    # Activate 4-bit precision base model loading
    use_4bit = True

    # Compute dtype for 4-bit base models
    bnb_4bit_compute_dtype = "float16"

    # Quantization type (fp4 or nf4)
    bnb_4bit_quant_type = "nf4"

    # Activate nested quantization for 4-bit base models (double quantization)
    use_nested_quant = False

    ################################################################################
    # TrainingArguments parameters
    ################################################################################

    # Output directory where the model predictions and checkpoints will be stored
    output_dir = "./results"

    # Number of training epochs
    num_train_epochs = 1

    # Enable fp16/bf16 training (set bf16 to True with an A100)
    fp16 = True
    bf16 = False

    # Batch size per GPU for training
    per_device_train_batch_size = int(os.getenv("TRAIN_BATCH_SIZE", "1"))

    # Batch size per GPU for evaluation
    per_device_eval_batch_size = int(os.getenv("EVAL_BATCH_SIZE", "2"))

    # Number of update steps to accumulate the gradients for
    gradient_accumulation_steps = int(os.getenv("GRADIENT_ACCUMULATION_STEPS", "1"))

    # Enable gradient checkpointing
    gradient_checkpointing = True

    # Maximum gradient normal (gradient clipping)
    max_grad_norm = 0.3

    # Initial learning rate (AdamW optimizer)
    learning_rate = 2e-4

    # Weight decay to apply to all layers except bias/LayerNorm weights
    weight_decay = 0.001

    # Optimizer to use
    optim = "paged_adamw_32bit"

    # Learning rate schedule
    lr_scheduler_type = "cosine"

    # Number of training steps (overrides num_train_epochs)
    max_steps = -1

    # Ratio of steps for a linear warmup (from 0 to learning rate)
    warmup_ratio = 0.03

    # Group sequences into batches with same length
    # Saves memory and speeds up training considerably
    group_by_length = True

    # Save checkpoint every X updates steps
    save_steps = 0

    # Log every X updates steps
    logging_steps = int(os.getenv("LOGGING_STEPS", "50"))

    ################################################################################
    # SFT parameters
    ################################################################################

    # Maximum sequence length to use
    max_seq_length = int(os.getenv("MAX_SEQ_LENGTH", "512"))

    # Pack multiple short examples in the same input sequence to increase efficiency
    packing = False

    # Load the entire model on the GPU 0
    device_map = {'':torch.cuda.current_device()}

    # Set limit to a positive number
    limit = int(os.getenv("DATASET_LIMIT", "5000"))

    dataset = load_dataset(dataset_name, split="train")
    if limit != -1:
       
    dataset = dataset.shuffle(seed=42).select(range(limit))


    def transform(data):
       
    question = data['question']
       
    context = data['context']
       
    answer = data['answer']
       
    template = "Question: {question}\nContext: {context}\nAnswer: {answer}"
       
    return {'text': template.format(question=question, context=context, answer=answer)}


    transformed = dataset.map(transform)

    # Load tokenizer and model with QLoRA configuration
    compute_dtype = getattr(torch, bnb_4bit_compute_dtype)

    bnb_config = BitsAndBytesConfig(
       
    load_in_4bit=use_4bit,
       
    bnb_4bit_quant_type=bnb_4bit_quant_type,
       
    bnb_4bit_compute_dtype=compute_dtype,
       
    bnb_4bit_use_double_quant=use_nested_quant,
    )

    # Check GPU compatibility with bfloat16
    if compute_dtype == torch.float16 and use_4bit:
       
    major, _ = torch.cuda.get_device_capability()
       
    if major >= 8:
           
    print("=" * 80)
           
    print("Your GPU supports bfloat16")
           
    print("=" * 80)

    # Load base model
    model = AutoModelForCausalLM.from_pretrained(
       
    model_name,
       
    quantization_config=bnb_config,
       
    device_map=device_map,
       
    torch_dtype=torch.float16,
    )
    model.config.use_cache = False
    model.config.pretraining_tp = 1

    # Load LLaMA tokenizer
    tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    # Load LoRA configuration
    peft_config = LoraConfig(
       
    lora_alpha=lora_alpha,
       
    lora_dropout=lora_dropout,
       
    r=lora_r,
       
    bias="none",
       
    task_type="CAUSAL_LM",
       
    target_modules=["q_proj", "v_proj"]
    )

    # Set training parameters
    training_arguments = TrainingArguments(
       
    output_dir=output_dir,
       
    num_train_epochs=num_train_epochs,
       
    per_device_train_batch_size=per_device_train_batch_size,
       
    gradient_accumulation_steps=gradient_accumulation_steps,
       
    optim=optim,
       
    save_steps=save_steps,
       
    logging_steps=logging_steps,
       
    learning_rate=learning_rate,
       
    weight_decay=weight_decay,
       
    fp16=fp16,
       
    bf16=bf16,
       
    max_grad_norm=max_grad_norm,
       
    max_steps=max_steps,
       
    warmup_ratio=warmup_ratio,
       
    group_by_length=group_by_length,
       
    lr_scheduler_type=lr_scheduler_type,
    )

    trainer = SFTTrainer(
       
    model=model,
       
    train_dataset=transformed,
       
    peft_config=peft_config,
       
    dataset_text_field="text",
       
    max_seq_length=max_seq_length,
       
    tokenizer=tokenizer,
       
    args=training_arguments,
       
    packing=packing,
    )

    trainer.train()

    trainer.model.save_pretrained(new_model)

    # Reload model in FP16 and merge it with LoRA weights
    base_model = AutoModelForCausalLM.from_pretrained(
       
    model_name,
       
    low_cpu_mem_usage=True,
       
    return_dict=True,
       
    torch_dtype=torch.float16,
       
    device_map=device_map,
    )
    model = PeftModel.from_pretrained(base_model, new_model)
    model = model.merge_and_unload()

    # push to Cloud Storage

    file_path_to_save_the_model = '/finetune/new_model'
    model.save_pretrained(file_path_to_save_the_model)
    tokenizer.save_pretrained(file_path_to_save_the_model)
  3. 创建 requirements.txt 文件:
    accelerate==0.34.2
    bitsandbytes
    ==0.45.5
    datasets
    ==2.19.1
    transformers
    ==4.51.3
    peft
    ==0.11.1
    trl
    ==0.8.6
    torch
    ==2.3.0
  4. 创建 Dockerfile
    FROM nvidia/cuda:12.6.2-runtime-ubuntu22.04

    RUN apt-get update && \
        apt-get -y --no-install-recommends install python3-dev gcc python3-pip git && \
        rm -rf /var/lib/apt/lists/*

    COPY requirements.txt /requirements.txt

    RUN pip3 install -r requirements.txt --no-cache-dir

    COPY finetune.py /finetune.py

    ENV PYTHONUNBUFFERED 1

    CMD python3 /finetune.py --device cuda
  5. 在 Artifact Registry 代码库中构建容器:
    gcloud builds submit \
      --tag $REGION-docker.pkg.dev/$PROJECT_ID/$AR_REPO/$IMAGE_NAME \
      --region $REGION

5. 部署和执行作业

在此步骤中,您将为作业创建 YAML 配置,并使用直接 VPC 出站流量,以便更快地将数据上传到 Google Cloud Storage。

请注意,此文件包含您将在后续步骤中更新的变量。

  1. 创建名为 finetune-job.yaml.tmpl 的文件:
    apiVersion: run.googleapis.com/v1
    kind: Job
    metadata:
      name: $JOB_NAME
      labels:
        cloud.googleapis.com/location: $REGION
      annotations:
        run.googleapis.com/launch-stage: ALPHA
    spec:
      template:
        metadata:
          annotations:
            run.googleapis.com/execution-environment: gen2
            run.googleapis.com/network-interfaces: '[{"network":"default","subnetwork":"default"}]'
        spec:
          parallelism: 1
          taskCount: 1
          template:
            spec:
              serviceAccountName: $SERVICE_ACCOUNT_ADDRESS
              containers:
              - name: $IMAGE_NAME
                image: $REGION-docker.pkg.dev/$PROJECT_ID/$AR_REPO/$IMAGE_NAME
                env:
                - name: MODEL_NAME
                  value: "google/gemma-2b"
                - name: NEW_MODEL
                  value: "gemma-2b-sql-finetuned"
                - name: BUCKET_NAME
                  value: "$BUCKET_NAME"
                - name: LORA_R
                  value: "8"
                - name: LORA_ALPHA
                  value: "16"
                - name: GRADIENT_ACCUMULATION_STEPS
                  value: "2"
                - name: DATASET_LIMIT
                  value: "1000"
                - name: LOGGING_STEPS
                  value: "5"
                - name: HF_TOKEN
                  valueFrom:
                    secretKeyRef:
                      key: 'latest'
                      name: HF_TOKEN
                resources:
                  limits:
                    cpu: 8000m
                    nvidia.com/gpu: '1'
                    memory: 32Gi
                volumeMounts:
                - mountPath: /finetune/new_model
                  name: finetuned_model
              volumes:
              - name: finetuned_model
                csi:
                  driver: gcsfuse.run.googleapis.com
                  readOnly: false
                  volumeAttributes:
                    bucketName: $BUCKET_NAME
              maxRetries: 3
              timeoutSeconds: '3600'
              nodeSelector:
                run.googleapis.com/accelerator: nvidia-l4
  2. 运行以下命令,将 YAML 中的变量替换为您的环境变量:
    envsubst < finetune-job.yaml.tmpl > finetune-job.yaml
  3. 创建 Cloud Run 作业:
    gcloud alpha run jobs replace finetune-job.yaml
  4. 执行作业:
    gcloud alpha run jobs execute $JOB_NAME --region $REGION --async

该作业大约需要 10 分钟才能完成。您可以使用上一个命令的输出中提供的链接来查看状态。

6. 使用 Cloud Run 服务通过 vLLM 提供经过微调的模型

在此步骤中,您将部署一个 Cloud Run 服务。此配置使用直接 VPC 通过专用网络访问 Cloud Storage 存储分区,以加快下载速度。

请注意,此文件包含您将在后续步骤中更新的变量。

  1. 创建 service.yaml.tmpl 文件:
    apiVersion: serving.knative.dev/v1
    kind: Service
    metadata:
      name: serve-gemma-sql
      labels:
        cloud.googleapis.com/location: $REGION
      annotations:
        run.googleapis.com/launch-stage: BETA
        run.googleapis.com/ingress: all
        run.googleapis.com/ingress-status: all
    spec:
      template:
        metadata:
          labels:
          annotations:
            autoscaling.knative.dev/maxScale: '1'
            run.googleapis.com/cpu-throttling: 'false'
            run.googleapis.com/gpu-zonal-redundancy-disabled: 'true'
            run.googleapis.com/network-interfaces: '[{"network":"default","subnetwork":"default"}]'
        spec:
          containers:
          - name: serve-finetuned
            image: us-docker.pkg.dev/vertex-ai/vertex-vision-model-garden-dockers/pytorch-vllm-serve:20250505_0916_RC00
            ports:
            - name: http1
              containerPort: 8000
            resources:
              limits:
                cpu: 8000m
                nvidia.com/gpu: '1'
                memory: 32Gi
            volumeMounts:
            - name: fuse
              mountPath: /finetune/new_model
            command: ["python3", "-m", "vllm.entrypoints.api_server"]
            args:
            - --model=/finetune/new_model
            - --tensor-parallel-size=1
            env:
            - name: MODEL_ID
              value: 'new_model'
            - name: HF_HUB_OFFLINE
              value: '1'
          volumes:
          - name: fuse
            csi:
              driver: gcsfuse.run.googleapis.com
              volumeAttributes:
                bucketName: $BUCKET_NAME
          nodeSelector:
            run.googleapis.com/accelerator: nvidia-l4
  2. 使用您的存储分区名称更新 service.yaml 文件。
    envsubst < service.yaml.tmpl > service.yaml
  3. 部署 Cloud Run 服务:
    gcloud alpha run services replace service.yaml

7. 测试微调后的模型

在此步骤中,您将提示模型测试微调。

  1. 获取 Cloud Run 服务的服务网址:
    SERVICE_URL=$(gcloud run services describe serve-gemma-sql --platform managed --region $REGION --format 'value(status.url)')
  2. 为模型创建提示。
    USER_PROMPT="Question: What are the first name and last name of all candidates? Context: CREATE TABLE candidates (candidate_id VARCHAR); CREATE TABLE people (first_name VARCHAR, last_name VARCHAR, person_id VARCHAR)"
  3. 使用 C网址 调用您的服务以提示模型:
    curl -X POST $SERVICE_URL/generate \
      -H "Content-Type: application/json" \
      -H "Authorization: bearer $(gcloud auth print-identity-token)" \
      -d @- <<EOF
    {
        "prompt": "${USER_PROMPT}"
    }
    EOF

您应该会看到如下所示的响应:

{"predictions":["Prompt:\nQuestion: What are the first name and last name of all candidates? Context: CREATE TABLE candidates (candidate_id VARCHAR); CREATE TABLE people (first_name VARCHAR, last_name VARCHAR, person_id VARCHAR)\nOutput:\n CREATE TABLE people_to_candidates (candidate_id VARCHAR, person_id VARCHAR) CREATE TABLE people_to_people (person_id VARCHAR, person_id VARCHAR) CREATE TABLE people_to_people_to_candidates (person_id VARCHAR, candidate_id"]}

8. 恭喜!

恭喜您完成此 Codelab!

建议您参阅 Cloud Run 文档。

所学内容

  • 如何使用 Cloud Run 作业 GPU 进行微调
  • 如何使用 Cloud Run 搭配 vLLM 提供模型
  • 如何为 GPU 作业使用直接 VPC 配置,以更快地上传和分发模型

9. 清理

为避免意外产生费用,例如,如果 Cloud Run 服务的意外调用次数超出了免费层级的 Cloud Run 调用次数配额,您可以删除在第 6 步中创建的 Cloud Run 服务。

如需删除 Cloud Run 服务,请前往 Cloud Run Cloud 控制台 (https://console.cloud.google.com/run),然后删除 serve-gemma-sql 服务。

如需删除整个项目,请前往管理资源,选择您在第 2 步中创建的项目,然后选择“删除”。如果您删除该项目,则需要在 Cloud SDK 中更改项目。您可以通过运行 gcloud projects list 来查看所有可用项目的列表。