Inductor cpp wrapper: clean up the hard-coded schema for fusion OPs #112552
Description
🐛 Describe the bug
Inductor cpp wrapper uses hard-coded cpp_op_schema for fusion OPs including:
- MKLPackedLinear
- LinearUnary
- LinearBinary
- ConvolutionUnary
- ConvolutionBinary
- ConvolutionBinaryInplace
- ConvolutionTransposeUnary
- QConvPointWisePT2E
- QConvPointWiseBinaryPT2E
- QLinearPointwisePT2E
- QLinearPointwiseBinaryPT2E
As suggested in #112373 (comment), we need to clean up these hard-coded strings and add a general way to create the cpp_op_schema for these OPs.
Error logs
No response
Minified repro
No response
Versions
PyTorch version: 2.2.0a0+git17eaa44
cc @ezyang @msaroufim @bdhirsh @anijain2305 @zou3519 @chauhang @wconstab