torch.jit.trace fails with very simple conv layer on tpu but succeeded on both mps and cpu

[BitPhinix]

Minimal reproduction example:

import torch
import torch_xla
import torch_xla.core.xla_model as xm

class SimpleModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.conv_in = torch.nn.Conv2d(3, 64, kernel_size=3, padding=1)

    def forward(self, sample):
        return self.conv_in(sample)


def trace_model(device):
    model = SimpleModel().to(device)
    model.eval()  # Set the model to evaluation mode

    # Create a sample input
    sample = torch.randn(1, 3, 32, 32, device=device)

    # Attempt to trace the model
    try:
        with torch.no_grad():  # Disable gradient computation
            traced_model = torch.jit.trace(model, sample)
        print(f"Tracing successful on {device}")

        # Test the traced model
        test_output = traced_model(sample)
        print(f"Test output shape: {test_output.shape}")

    except Exception as e:
        print(f"Tracing failed on {device}: {str(e)}")


def main():
    # Test on CPU
    print("Testing on CPU:")
    trace_model(torch.device("cpu"))

    # Test on XLA device
    print("\nTesting on XLA device:")
    xla_device = xm.xla_device()
    trace_model(xla_device)


if __name__ == "__main__":
    main()

Works fine on CPU but fails with [ XLAFloatType{1,64,32,32} ]) of traced region did not have observable data dependence with trace inputs; this probably indicates your program cannot be understood by the tracer.

Tested on a TPU v4 pod.

Especially annoying because tracing on CPU and loading it onto the xla device isn't possible because of pytorch/pytorch#96448