Skip to content

[cherry-pick to r2.6_aws_neuron] add torch_xla_graph_execution_check_level (default disabled) flag that emits warning(1) or throw error(2) during tensor sync and output the python frame #9077

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Conversation

aws-yyjau
Copy link
Contributor

Same commit (squashed and rebased) from #9057.
Please refer to the original PR for details and comments.

…t emits warning(1) or throw error(2) during tensor sync and output the python frame
@aws-yyjau aws-yyjau changed the title add torch_xla_graph_execution_check_level (default disabled) flag that emits warning(1) or throw error(2) during tensor sync and output the python frame [cherry-pick to r2.6_aws_neuron] add torch_xla_graph_execution_check_level (default disabled) flag that emits warning(1) or throw error(2) during tensor sync and output the python frame May 1, 2025
@aws-yyjau
Copy link
Contributor Author

I worked with @jeffhataws to test the commit in the BERT training example (Hugging Face BERT Pretraining Tutorial (Data-Parallel) — AWS Neuron Documentation)

With the commit,

LOG Tue May  6 20:54:57 2025 - (0, 128) step_loss : 8.4156 learning_rate : 2.56e-05 \
throughput : 4035.43 (w the commit)

Without the commit,

LOG Tue May  6 21:07:44 2025 - (0, 128) step_loss : 8.4156 learning_rate : 2.56e-05 \
throughput : 3955.75 (w/o the commit)

With the commit has 2% higher throughput than the baseline, which I think is within the fluctuation.
It validates that the commit does not cause performance change to the training process.

@aws-yyjau
Copy link
Contributor Author

This PR is closed since the feature is already supported. Please refer to the details in #9057.

@aws-yyjau aws-yyjau closed this May 14, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants