Distributed training error using MXNet Gluon examples

I've tried run examples of MXNet Gluon scripts and notebooks ([sagemaker-python-sdk/mxnet_gluon_sentiment](https://github.com/aws/amazon-sagemaker-examples/blob/master/sagemaker-python-sdk/mxnet_gluon_sentiment/mxnet_sentiment_analysis_with_gluon.ipynb) and [sagemaker-python-sdk/mxnet_gluon_mnist](https://github.com/aws/amazon-sagemaker-examples/tree/master/sagemaker-python-sdk/mxnet_gluon_mnist)) with distributed training, and when the default framework version 1.6.0 (or 1.7.0, 1.8.0) is configured, the job failed with errors below:
```
[Epoch 1] Training: accuracy=0.909658
[Epoch 1] Validation: accuracy=0.812799
Vocabulary saved to "%s" /opt/ml/model/vocab.json
terminate called without an active exception
2020-12-03 03:58:54,540 sagemaker-training-toolkit ERROR ExecuteUserScriptError:
Command "/usr/local/bin/python3.6 sentiment.py --batch-size 8 --embedding-size 50 --epochs 2 --learning-rate 0.01 --log-interval 1000"
```

And change framework version to 1.4.1 would workaround this issue, SageMaker version is 2.* (2.1.0, 2.16.4, 2.17.0). Would like to know what is the cause and how to fix this using newer version like 1.6.0 and later.

Thanks.

Start by reproducing the failure in sagemaker-python-sdk/mxnet_gluon_sentiment and sagemaker-python-sdk/mxnet_gluon_mnist with framework versions 1.6.0 and later, comparing against 1.4.1. Trace the distributed training error from the example entry points and document or implement a verified fix for the newer versions.

Written by the indexing model from the issue text.