mosaic_memory_profiling_tutorial.py failing against 2.11

[sekyondaMeta]

beginner_source/mosaic_memory_profiling_tutorial.py failing agains 2.11 RC

https://github.com/pytorch/tutorials/actions/runs/22152897277/job/64048844976?pr=3772#step:9:5472

cc @basilwong

[basilwong]

CI Failure in mosaic_memory_profiling_tutorial.py - Likely Transient, Please Retry

Hi,

The failure in mosaic_memory_profiling_tutorial.py appears to be a transient issue, likely caused by GPU memory pressure during the backward pass.

What happened

• The error SystemError: <built-in method dropout of type object at ...> returned NULL without setting an exception occurred during loss.backward() inside activation checkpointing recomputation
• This cryptic error is a known symptom of CUDA running out of memory but failing to surface a proper Python OOM exception

Why it's likely transient

• Memory availability on shared CI runners can vary depending on concurrent jobs
• Activation checkpointing recomputation temporarily spikes memory usage during backward pass
• No code changes would explain a sudden failure in this path

Recommended action

Please retry the job. If the failure persists across multiple attempts, we may need to investigate further (e.g., reduce batch size in the tutorial, check for PyTorch/CUDA version incompatibilities, or examine runner memory limits).

Let me know if it fails again!

[sekyondaMeta]

@basilwong Looks like this is still failing:

https://github.com/pytorch/tutorials/actions/runs/23771851846/job/69264923528?pr=3809#step:9:3937