torch compile issues

[jdack41]

#1024 (comment)

I have written about some errors with torch compile above, but there are currently two issues with the torch compile option:

• When using einops version 0.6.1, an "unhashable type: non-singleton SymInt" error occurs during finetuning. This issue does not occur with version 0.7.0.
• When saving a model using torch compile, a prefix '_orig_mod.' is added to the saved model, which breaks the structure if saved as is.

I have been able to reproduce these issues in my local environment, but could someone please verify the situation in another environment?

[feffy380]

einops 0.6.1 requires you to manually register its operators as compilable, which sd-scripts fails to do.
einops 0.7 does this automatically.
https://github.com/arogozhnikov/einops/wiki/Using-torch.compile-with-einops

[jdack41]

einops 0.6.1 requires you to manually register its operators as compilable, which sd-scripts fails to do. einops 0.7 does this automatically. https://github.com/arogozhnikov/einops/wiki/Using-torch.compile-with-einops

Thank you for verification.
I understood why error occurred with 0.6.1.

[jdack41]

[2024-02-24 05:02:51,702] [33/29] torch._dynamo.variables.higher_order_ops: [ERROR] self mismatch for bound method

I’ｍ also concerned about the continued occurrence of the above error while exec sdxl_train_network.py. Is this occurring in other environments?

[jdack41]

When saving a model using torch compile, a prefix '_orig_mod.' is added to the saved model, which breaks the structure if saved as is.

It should be noted that this is incompatible with other applications such as a1111 as it is, or it should be modified to remove orig_mod as explained above.

[kohya-ss]

I will update the version of einops in the next release.