Fix attribute error for detr

[meenakshiramanathan1]

No description provided.

[nvukobratTT]

Let's try without model overrides. It would be best to fix the root cause of the issue without fine-tuning compiler behavior.

Do you have a wider context of the problem solved with these flags?

[meenakshiramanathan1]

zeros op is binded as global_var and passed as input params to tvmgen_default_forge_main.
def @main(%pixel_values: Tensor[(1, 3, 427, 640), float32] /* ty=Tensor[(1, 3, 427, 640), float32] span=aten::size_0.pixel_values:0:0 */, %model.query_position_embeddings.weight: %0 = zeros(shape=[1, 100, 256], dtype="float32") /* ty=Tensor[(1, 100, 256), float32] span=C.graph: aten::zeros_like, jit._trace.TopLevelTracedModule: %1 = @tvmgen_default_forge_main_0(%0, %model.query_position_embeddings.weight, %model.decoder.layers.0.

This behaviour is observed after tvm transform.PartitionGraph , hence adding these flags was binding the zeros op into tvm module.

[nvukobratTT]

Hmm, do you have an example from reportify of how this looks like?

[nvukobratTT]

And do you know where are these zeros coming from? I can't infer precise info based on TVM IR you provided..

[meenakshiramanathan1]

Updated the changes now by bringing in a workaround in forge_utils mentioned here.