The ONNX parser parses GEMM operators without checking if the matrix A is transposed. This is useful for training purposes where the activation tensor is transposed (as depicted). In ZigZag, the GEMM operator is parsed correctly.
I can provide a small ONNX reproducer and a PR that should work with a fix using the ZigZag ONNX Parser. https://github.com/JeremyMorlier/stream/tree/checkGemm
The ONNX parser parses GEMM operators without checking if the matrix A is transposed. This is useful for training purposes where the activation tensor is transposed (as depicted). In ZigZag, the GEMM operator is parsed correctly.
I can provide a small ONNX reproducer and a PR that should work with a fix using the ZigZag ONNX Parser. https://github.com/JeremyMorlier/stream/tree/checkGemm