torch._ops.aten.bernoulli.p about executorch HOT 3 CLOSED

Thanks a lot for your quick reply to this issue and for the recommended solutions. @SS-JIA, the issue was related to eval mode. After I set the model to evaluation mode, it worked.

from executorch.

I would recommend to try disabling IR validity checking via EdgeCompileConfig. This should cause the check that is producing that error to be skipped.

One example of how to disable validity checking is:

# Assume that eager_model and sample_inputs have been defined above

exported_program = export(eager_model, sample_inputs)

edge_program = to_edge(
    exported_program,
    compile_config=exir.EdgeCompileConfig(_check_ir_validity=False, _use_edge_ops = False),
).transform(propagate_dynamic_shape())

Please try it out and see if it works!

from executorch.

However, even if you skip the check it's likely that your model may not be able to execute because non ATen canonical operators are not currently implemented in ExecuTorch's kernel libraries. This means that to execute your model, you may need to provide an implementation for the torch._ops.aten.bernoulli operator.

I would recommend that you try to execute the model while skipping the check, and if you run into an error complaining that there is no kernel available for the aten.bernoulli operator:

1. Please open an issue to request that a implementation of aten.bernoulli be added to the portable kernel library
2. Alternatively, you can try to provide an implementation for the operator yourself. If you decide to go this route, let me know and I can provide some more details on how to do this.

from executorch.