Torch.fx.symbolic_trace fails for most encoder-decoder NLP models

I am new to TVM. I am trying to port models like “Helsinki-NLP/opus-mt-it-en” and whisper using tvm relax to mobile phones. Using TorchFX to trace a graph from the PyTorch module and getting error like " @compatibility(is_backward_compatible=True)

TraceError: symbolically traced variables cannot be used as inputs to control flow"

It happens due to control flow statement like if-else. I have used solution as mentioned in torch.fx.symbolic_trace fails on torch.arange with input-dependent size · Issue #51803 · pytorch/pytorch · GitHub, but still the problem persists. How to load the such model for relax? Is there any pointer toward such encoder-decoder foundational model within TVM framework?

1 Like