does it support p5 instance when do data parallelism?

[abdalgader-a]

accelerate/src/accelerate/utils/constants.py

Line 35 in d6d3e03

SAGEMAKER_PARALLEL_EC2_INSTANCES = ["ml.p3.16xlarge", "ml.p3dn.24xlarge", "ml.p4dn.24xlarge"]

Hi, does this support running on p5 instances on parallelism mode (model or data)?

is it related to: https://github.com/aws/sagemaker-python-sdk/blob/daa55180f1e4f6f1451eaa352a2d7940b3424c2d/src/sagemaker/estimator.py#L4064C23-L4064C56

Thanks!

[abdalgader-a]

@BenjaminBossan any thoughts?

[BenjaminBossan]

Sorry, I have no experience with SageMaker. It would also help if you could elaborate your question: Did you try and run into an error?

Note that the constant SAGEMAKER_PARALLEL_EC2_INSTANCES that you're referring to is just there to provide some standard answers when users create the accelerate config, it does not mean that those are the only supported options.

[abdalgader-a]

Thanks @BenjaminBossan for getting back on this.

Yes, ran a python script via accelerate with Sagmaker-sdk integration and set ec2_instance_type: ml.p5.48xlarge run into the error of:
"ValueError: SMModelParallel and SMDataParallel currently do not support p5 instances." that I point it above.

So, I'm wondering why it not support p5 instances?

[BenjaminBossan]

The error message looks very unambiguous to me, so the answer is it's not supported. I don't think this is an accelerate issue but rather on AWS SageMaker to add the support. I can see that you already opened a ticket there, I don't think we can add anything here.

[abdalgader-a]

Thanks @BenjaminBossan! I'll close the issue as you it confirmed the issue is on AWS SageMaker SDK side.