cbalint13 opened a new pull request, #18207: URL: https://github.com/apache/tvm/pull/18207
This PR enforce ONNX constraint compliance for binary operations. --- #### Changes * Overides the behaviour of numpy precision widening, e.g. ```np.div(int64,in64)->float64``` * This **now complies** with the [onnx.Div](https://onnx.ai/onnx/operators/onnx__Div.html) output constraint: ```C: Result, has same element type as two inputs```. * Use an more appropriate public helper from API for the numpy type conversion (now it works with onnx 0.20 too). --- #### Failure example After shape sampling and division in the graph (excerpt from a custom onnx11n model) ```np.div(5,2)``` results in a ```float64``` (as per numpy behaviour) and then the input parameter of ```Slice``` fails upon its ```int64``` constraint. <img width="381" height="620" alt="onnx-div-widening-issue" src="https://github.com/user-attachments/assets/60fbd252-81f9-4ed7-842d-a2040f5923bd"; /> ``` tvm.error.InternalError: Check failed: (is_base_of_tuple_of_int64(sinfo)) is false: Operator Op(relax.strided_slice) requires the end argument to be a tuple of int64 PrimValues. However, in expression {...}, the end argument (R.prim_value(T.float64(2.5)),) has struct info R.Tuple(R.Prim(value=T.float64(2.5))) ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
