Possible problem on latest ONNX model version types

TVM compilation fails with models committed in the last month

Using TVM I attempted to compile some classification vision models from those uploaded lately (in the last month) - such as GoogleNet v12 and AlexNet v12, but I also tried much more. I received the following error to the platform:

Check failed: type_code_ == kDLFloat (8 vs. 2) : expected float but got Object.

However, older models (such as Googlenet v9, etc) work without any problems from the repo.

Now, this is an error raised to TVM, where there was an old issue report with no clear answer on their forum, which I re-raised. However, since this is something common to all models uploaded lately, I thought to ask here as well, in case there is a common model creation process that may introduce this error. I would suspect a type conversion and specific handling on a layer, but I do not have a clear picture on this.

Of course, I am also not 100% sure this is ONNX related, however I would like to hear your insight on this issue. Thank you!

Start by reproducing the reported TVM compilation failure with GoogleNet v12 and AlexNet v12, then compare them with older working versions such as GoogLeNet v9. Inspect whether a common model-creation process, type conversion, or layer difference explains the error; done means determining whether the issue is in the submitted models and documenting the cause or next action.