fix: Add ml.p5e.48xlarge and ml.p5.48xlarge to EFA instance lists

[srujithpoondla03]

Description

Add ml.p5e.48xlarge to SM_EFA_NCCL_INSTANCES and SM_EFA_RDMA_INSTANCES.
Add ml.p5.48xlarge to SM_EFA_RDMA_INSTANCES (was missing).

Without these entries, NCCL hangs during distributed training initialization on P5e instances due to missing EFA environment variables (FI_PROVIDER, FI_EFA_USE_DEVICE_RDMA, RDMAV_FORK_SAFE).

Related

• Fixes Add ml.p5e.48xlarge to EFA instance lists in sagemaker-train and sagemaker-core #5491
• sagemaker-training-toolkit issue: Add ml.p5e.48xlarge to EFA instance lists (SM_EFA_NCCL_INSTANCES and SM_EFA_RDMA_INSTANCES) sagemaker-training-toolkit#240
• sagemaker-training-toolkit PR: feat: Add ml.p5e.48xlarge and ml.p5.48xlarge to EFA instance lists sagemaker-training-toolkit#241

Testing

• Added unit test for P5/P5e EFA instance list membership

[srujithpoondla03]

Is there a specific process for testing EFA/instance-specific changes on actual hardware (e.g., P5e instances) before merging? Or are unit tests sufficient for this type of change?

We've verified the fix works in production with a manual workaround (setting the EFA env vars directly), but wanted to confirm if there's an integration test process for these hardware-specific configurations.

[mrpic-amzn]

We've verified the fix works in production with a manual workaround

Can you add more details about how this was tested in production. For example, the training job arns, the exact commands you used to test, the CUDA versions that were used in the training algo.