luyaor opened a new issue, #12305: URL: https://github.com/apache/tvm/issues/12305
## Description When compiling following model, TVM will produce wrong result. The model(with ONNX as frontend) with error is as follows, check bug.onnx in [reproduce.zip](https://github.com/apache/tvm/files/9259136/reproduce.zip) The semantic of the model is to sum up input tensors with five times. For example, the first element in the input tensor is -0.7491791248321533, so the first element in output should be -0.7491791248321533 * 5 = -3.74, however, the element in the output is -11.986866. <img width="1020" alt="image" src="https://user-images.githubusercontent.com/7541296/182831586-cf464fff-2d42-465d-9f9a-b19041879f98.png"> ## How to reproduce ### Expected behavior Output should be as follows: ``` [array([[[[-3.7458956 , 2.2319033 , 4.721547 ]], [[-2.5541332 , -1.010938 , -1.2561022 ]], [[-2.432028 , 2.1270378 , -2.2868755 ]]], [[[ 0.14414614, -3.7024972 , 3.4796305 ]], [[ 1.8737267 , 0.38758603, 1.1408547 ]], [[-2.7563024 , 3.4405584 , -3.7759078 ]]]], dtype=float32)] ``` ### Actual behavior This is the output of running model with TVM compiled. ``` [[[[-11.986866 7.142091 15.108952 ]] [[ -8.173226 -3.2350018 -4.019527 ]] [[ -7.78249 6.806521 -7.3180013 ]]] [[[ 0.46126765 -11.847991 11.134817 ]] [[ 5.9959254 1.2402753 3.6507351 ]] [[ -8.820168 11.009787 -12.082905 ]]]] ``` ### Environment Python3, with tvm, onnx, numpy, onnxruntime(optional) 1. Download [reproduce.zip](https://github.com/apache/tvm/files/9259136/reproduce.zip) 2. Run `python check.py`. To check with the actual output, you could check with the commented code. -- 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]
