Custom Argument pass to Docker Container Azure ML inference

Question

Custom Argument pass to Docker Container Azure ML inference

khubaib Raza 1

Hello Team,

I'm trying to pass the arguments to Azure ML docker. I have created an environment like this.

env = Environment.from_conda_specification(name='pytorch-1.6-gpu', file_path='curated_env/conda_dependencies.yml' )

Am I passing the arguments correct?

DOCKER_ARGUMENTS = ["--shm-size","32G"]  # increase shared memory
env.docker.arguments = DOCKER_ARGUMENTS

The main goal of this project is to deploy a model on the AKS inference cluster. I have successfully deployed the model. When I try to get predictions from the model I got this error

It is possible that data loaders workers are out of shared memory. Please try to raise your shared memory limit

How can I do that if that's not the correct way to pass arguments?

Answer 1

romungi-MSFT 32,931 Microsoft Employee

@khubaib Raza To pass the argument for increasing the default "shm_size" you would have to use the DockerConfiguration object. Here is a sample to achieve this:

from azureml.core import Environment  
from azureml.core import ScriptRunConfig  
from azureml.core.runconfig import DockerConfiguration  
  
  
# Specify VM and Python environment:  
my_env = Environment.from_conda_specification(name='my-test-env', file_path=PATH_TO_YAML_FILE)  
my_env.docker.base_image = 'mcr.microsoft.com/azureml/openmpi3.1.2-cuda10.2-cudnn7-ubuntu18.04'  
  
docker_config = DockerConfiguration(use_docker=True,shm_size='32g')  
  
# Finally, use the environment in the ScriptRunConfig:  
src = ScriptRunConfig(source_directory=DEPLOY_CONTAINER_FOLDER_PATH,  
                      script=SCRIPT_FILE_TO_EXECUTE,  
                      arguments=EXECUTE_ARGUMENTS,  
                      compute_target=compute_target,  
                      environment=my_env,  
                      docker_runtime_config=docker_config)

If an answer is helpful, please click on or upvote which might help other community members reading this thread.

khubaib Raza 1

Thank you for your response @romungi-MSFT
Let me clarify for you, I'm not training the model. I'm just trying to deploy a pre-trained model to the inference cluster. I don't think ScriptRunConfig will help me anyway.

How can I pass the DockerConfiguration object to InferenceConfig or is there any other way?

here is my configuration it might help you to understand. What I'm trying to acheive.

inference_config = InferenceConfig(entry_script="score.py", environment=env , source_directory='.' )   
  
gpu_aks_config = AksWebservice.deploy_configuration(autoscale_enabled=False,  
                                                    num_replicas=1,  
                                                    cpu_cores=2,  
                                                    memory_gb=16)  
  
service = Model.deploy(workspace=ws,   
                           name=MODEL_NAME,   
                           models=[model],   
                           inference_config=inference_config,   
                           deployment_config=gpu_aks_config,  
                           deployment_target=aks_target,  
                           overwrite=True)  
  
service.wait_for_deployment(show_output=True)  
print(service.state)

romungi-MSFT 32,931 Reputation points Microsoft Employee

2021-09-29T09:03:59.783+00:00

You can use the InferenceConfig() extra_docker_file_steps parameter to run steps while setting up the image. I think this is basically a dockerfile to run the setup.

For the variables approach I think the steps mentioned in this SO post should work but I am not sure if this particular variable in the scoring script will help setup the image with increased size.

Custom Argument pass to Docker Container Azure ML inference

1 answer

Activity