Torch profiler fails by default on pytorch==2.5.0

[dorian-K]

I set torch_profile = True in my config

The default args for torch.profiler.profiler.schedule include skip_first_wait=0, but this option is not available in torch 2.5.0: https://docs.pytorch.org/docs/2.5/profiler.html#torch.profiler.schedule

I would be good to have the default configuration not crash

[dorian-K]

Also
.../recipe/returnn/returnn/torch/engine.py:1779: FutureWarning: `export_memory_timeline` is deprecated and will be removed in a future version. Please use `torch.cuda.memory._record_memory_history` and `torch.cuda.memory._export_memory_snapshot` instead. self.profiler.export_memory_timeline("torch_memory_profile.html")

[albertz]

I'm personally not using PyTorch 2.5 anymore. Maybe just upgrade?

But if you want PyTorch 2.5 support here, feel free to do a PR about the schedule defaults.

About the export_memory_timeline, I don't remember that I have seen this warning, so maybe it is not deprecated anymore? At least it still seems to work in newer PyTorch versions. I would ignore this for now.

[dorian-K]

I upgraded to 2.10, and then I got that deprecation warning

[albertz]

I upgraded to 2.10, and then I got that deprecation warning

Ah then I misunderstood. But anyway, I would say to ignore this warning for now.