Hello, I’m currently working on optimizing an integer-only convolution network using TVM. I recently discovered that in TVM’s library, there are some models implemented as Relay IR . (I’m using get_workload from tvm.relay.testing.resnet) These models seems to support various data types, including int8, int32, and uint8. However, I encountered errors during optimization for some operations.
I suspect there are two possible reasons for these errors:
- The implemented Relay code themselves might be incorrect and require modification(like datatype for convolution output).
- Some operators used in the model may not be supported on the target hardware.
To check the first case, I need to validate the inference results at the Relay IR level. Do you have any examples or guidance on how to do this?
thanks always.