PyTorch 2.0 and ONNX converter

[ForserX]

UnsupportedOperatorError: Exporting the operator 'aten::scaled_dot_product_attention' to ONNX opset version 14 is not supported.

More info: pytorch/pytorch#97262

[justinchuby]

@abock

[BowenBao]

cc @titaiwangms using atenlib for ts exporter

[patrickvonplaten]

cc @anton-l and @echarlaix here as well

[justinchuby]

For reference pytorch/pytorch#97262 is a transitional workaround until we expose the new dynamo interface

[echarlaix]

Was temporarily fixed in optimum : huggingface/optimum#888 (cc @fxmarty)

[ForserX]

Isn't it better to fix the converter script?

ForserX/StableDiffusionUI@195e21f

[fxmarty]

@justinchuby Out of curiosity, what do you mean by

expose the new dynamo interface

?

[justinchuby]

We are working on a new torch.onnx.dynamo_export API which will use dynamo for graph capturing instead of torch.jit. pytorch/pytorch#97920

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[titaiwangms]

Fixed by pytorch/pytorch#99658

[ForserX]

It remains to wait for the release of PyTorch 2.1.0

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[ashimatyuk]

Hi,
I'm facing "torch.onnx.errors.UnsupportedOperatorError: Exporting the operator 'aten::scaled_dot_product_attention' to ONNX opset version 14 is not supported" when using convert_stable_diffusion_checkpoint_to_onnx.py for converting sd-v1-5-inpainting.ckpt to onnx. As I can see it was fixed (as @titaiwangms mentioned above) but it still doesn't work. Am I doing smth wrong? The error is raised on the unet stage, text_encoder has been converted successfully.

Upd.: now on the VAE stage. On the unet 'pipeline.unet.set_attn_processor(CrossAttnProcessor())' helped, but same is not working for VAE.