Sagemaker部署PyTorch模型调用predict时遇SSL协议错误求助

部署预训练PyTorch模型到Sagemaker Notebook环境后，执行response = predictor.predict(serialized_data)调用推理时，遇到如下SSL错误：

SSLError: SSL validation failed for https://runtime.sagemaker.us-east-1.amazonaws.com/endpoints/pytorch-inference-2024-06-11-13-39-50-210/invocations EOF occurred in violation of protocol (_ssl.c:2426)

预期行为 #

未部署模型时，直接从S3桶加载权重，执行以下代码可正常得到推理结果：

input_data = (interaction_data, mt_data)
results = predict_fn(input_data, model)

输出示例：Iteration at 0: auc 0.964, map 0.439

当前情况 #

模型定义与部署步骤均正常完成：

pytorch_model = PyTorchModel(model_data=f's3://{model_bucket}/{model_key}', 
                             role=role, 
                             entry_point='inference.py', 
                             framework_version='1.8.1', 
                             py_version='py3', 
                             sagemaker_session=sagemaker_session)
predictor = pytorch_model.deploy(instance_type='ml.m5.large', initial_instance_count=1)

但调用predictor.predict(serialized_data)时触发上述SSL错误。

复现步骤 #

• 定义数据路径：

  interaction_data = "s3://path_to_pkl/interaction.pkl"
  auxiliary_data = "s3://path_to_pkl/auxiliary.pkl"
  

• 定义模型存储桶：

  model_bucket = 'path_to_model_bucket'
  model_key = 'Model-Structure/model.tar.gz'
  

• 处理数据：

  with open("interaction.pkl", 'rb') as f:
      data1 = CPU_Unpickler(f).load()
  with open("auxiliary.pkl", 'rb') as f:
      data2= CPU_Unpickler(f).load()
  serialized_data = pickle.dumps({ 'data1': data1, 'data2': data2 })
  

• 定义并部署模型：

  pytorch_model = PyTorchModel(model_data=f's3://{model_bucket}/{model_key}', 
                               role=role, 
                               entry_point='inference.py', 
                               framework_version='1.8.1', 
                               py_version='py3', 
                               sagemaker_session=sagemaker_session)
  predictor = pytorch_model.deploy(instance_type='ml.m5.large', initial_instance_count=1)
  

• 调用推理：

  response = predictor.predict(serialized_data)
  

已使用Sagemaker环境通用的inference.py文件进行模型评估和响应获取。

内容的提问来源于stack exchange，提问作者Dhruv Shah