ir_version issue in parsing

General Discussion
1

Hello everyone,

I am currently working on compiling a vit_large_patch16_224_Opset17.onnx model, which I downloaded from the official ONNX GitHub Model Zoo, for deployment on a Hailo-8L accelerator.

However, when I try to parse the model with the Hailo Dataflow Compiler, I encounter the following error (image can be provided):

image
image1257×418 21.8 KB

Hello,

I am trying to compile the vit_large_patch16_224_Opset17 model from the ONNX GitHub to use it on a Hailo-8L. However, an error like the one below occurs. (I have a picture)

So, I tried to change the model’s opset to 13, but I ran into this error: D:\a\onnx\onnx\onnx/version_converter/adapters/no_previous_version.h:26: adapt: Assertion 'false' failed: No Previous Version of LayerNormalization exists.

Is the hardware incompatible with the model’s ir_version? I have no idea what the problem is or what I should do.

To begin with, is it impossible to run a somewhat large model that isn’t on the hailo-github, even if it’s split?

Thank you so much.

2

Hey @ko_kyung_taek,

I wanted to let you know about an issue with compiling ViT models. You won’t be able to compile ViT models that use LayerNormalization unless you take one of these approaches:

• Manually replace the LayerNormalization with a combination of supported operators like Sub, Pow, ReduceMean, Div, etc.
• Use the alls and yaml files to handle the replacement automatically

If you’re working with the exact same model from the model zoo, you’ll need to use these specific files:

Hope this helps!