[Bug][Relax][ONNX] BitShift raises AttributeError: 'str' object has no attribute 'decode' when direction attribute is omitted #20950

Expected behavior

According to the ONNX specification for BitShift (opset 11+), when the direction attribute is omitted, it should default to "LEFT" (or b"LEFT"), converting the node to a relax.op.left_shift operation without crashing.

Actual behavior

When importing an ONNX model with a BitShift node that does not have an explicit direction attribute, from_onnx raises an AttributeError:

Error converting operator BitShift, with inputs: [x, y]
Traceback (most recent call last):
  File "repro.py", line 15, in <module>
    mod = from_onnx(model)
  File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 6864, in from_onnx
    return g.from_onnx(graph, opset)
  File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 6406, in from_onnx
    self._construct_nodes(graph)
  File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 6606, in _construct_nodes
    op = self._convert_operator(op_name, inputs, attr, self.opset)
  File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 6703, in _convert_operator
    sym = op_function(self.bb, inputs, attrs, [self._nodes, self._params])
  File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 917, in _impl_v11
    direction = attr.get("direction", "LEFT").decode("ascii")
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'str' object has no attribute 'decode'. Did you mean: 'encode'?

Environment

• TVM Version: 0.28.dev0 (commit cbb1e12fe on main)
• ONNX Version: 1.23.2
• Python Version: 3.12.2
• Operating System: Linux x86_64

Steps to reproduce

import onnx
from onnx import helper, TensorProto
from tvm.relax.frontend.onnx import from_onnx

# BitShift node with omitted direction attribute
node = helper.make_node("BitShift", inputs=["x", "y"], outputs=["z"])
graph = helper.make_graph(
    [node],
    "bitshift_graph",
    [
        helper.make_tensor_value_info("x", TensorProto.UINT32, [2, 2]),
        helper.make_tensor_value_info("y", TensorProto.UINT32, [2, 2]),
    ],
    [helper.make_tensor_value_info("z", TensorProto.UINT32, [2, 2])],
)
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 11)])
mod = from_onnx(model)

Root Cause Analysis

In python/tvm/relax/frontend/onnx/onnx_frontend.py:

@classmethod
def _impl_v11(cls, bb, inputs, attr, params):
    direction = attr.get("direction", "LEFT").decode("ascii")

In ONNX protobuf, string attributes are delivered as bytes. Throughout onnx_frontend.py (e.g., lines 2084, 3344, 3888, 4440, 5537), default values for string attributes are always passed as byte literals (e.g. attr.get("mode", b"constant").decode("utf-8")).

In BitShift, however, the fallback value was supplied as a Python string "LEFT". When direction is not in attr, attr.get("direction", "LEFT") returns "LEFT", and calling "LEFT".decode("ascii") fails with AttributeError because strings in Python 3 cannot be decoded.

Proposed Fix

Change the fallback default from "LEFT" to b"LEFT":

diff --git a/python/tvm/relax/frontend/onnx/onnx_frontend.py b/python/tvm/relax/frontend/onnx/onnx_frontend.py
--- a/python/tvm/relax/frontend/onnx/onnx_frontend.py
+++ b/python/tvm/relax/frontend/onnx/onnx_frontend.py
@@ -914,7 +914,7 @@ class BitShift(BitwiseBase):
 
     @classmethod
     def _impl_v11(cls, bb, inputs, attr, params):
-        direction = attr.get("direction", "LEFT").decode("ascii")
+        direction = attr.get("direction", b"LEFT").decode("ascii")
         if direction == "LEFT":
             cls.numpy_op = _np.left_shift
             cls.relax_op = relax.op.left_shift

Triage

• needs-triage
• relax:frontend