Open Issues Need Help
View All on GitHubAI Summary: This issue proposes extending ONNX's symbolic shape inference capabilities to a wider range of operators. The goal is to improve handling of dynamic shapes, enable better graph optimizations, and enhance compatibility with models exported from frameworks like PyTorch and TensorFlow by reducing reliance on runtime shape inference.
AI Summary: The user is encountering a shape inference failure on a MatMul node within a Hugging Face ONNX model when using the onnxslim tool. This indicates a potential issue with how onnxslim handles matrix multiplication operations or with the model's structure itself.
AI Summary: The `input_shape_modification` function in OnnxSlim has a bug where it fails to parse input names containing colon characters. This is because the code uses `rsplit(":", 1)` which incorrectly splits the input name when it includes colons. A fix is needed to handle these cases properly.