[Bug][Relax][ONNX] Dynamic MaxPool SAME_UPPER attempts int(SizeVar)

[Nanmur]

Summary

A PP-OCRv6-derived ONNX graph contains a 2-D MaxPool whose spatial width is symbolic and whose attributes are:

kernel_shape = [2, 2]
strides      = [1, 1]
auto_pad     = SAME_UPPER

relax.frontend.onnx.from_onnx fails before producing a Relax module because the MaxPool importer converts every input shape dimension to a Python int. This is invalid when a dimension is a symbolic TIR Var.

The failure is reproducible with the minimal public graph below; no model file is required.

Expected behavior

The Relax ONNX frontend should import this valid ONNX MaxPool with symbolic spatial dimensions, without materializing symbolic dimensions as Python integers.

Actual behavior

Error converting operator MaxPool, with inputs: [x]
...
File "python/tvm/relax/frontend/onnx/onnx_frontend.py", line 3996, in _get_input_spatial_shape
    return _np.array([int(d) for d in tensor.ty.shape], dtype="int64")[2:]
TypeError: int() argument must be a string, a bytes-like object or a real number, not 'Var'

Minimal reproduction

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

x = helper.make_tensor_value_info("x", TensorProto.FLOAT, ["N", 3, 48, "W"])
y = helper.make_tensor_value_info("y", TensorProto.FLOAT, ["N", 3, 48, "W"])
node = helper.make_node(
    "MaxPool",
    ["x"],
    ["y"],
    kernel_shape=[2, 2],
    strides=[1, 1],
    auto_pad="SAME_UPPER",
)
graph = helper.make_graph([node], "dynamic_same_maxpool", [x], [y])
model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 11)])

from_onnx(model)

Environment

TVM: v0.26.0 (c7b458e946bc4266915da582457476bdcd9705ae)
ONNX: 1.22.0
OS: Windows
Frontend: tvm.relax.frontend.onnx.from_onnx

Related issues

• This is distinct from [Bug] [Unity] TypeError encountered when converting ONNX model with MaxPool operator #15895, which was an earlier MaxPool default-strides issue.
• [Bug][Relax] Import failures on PaddleOCR and FasterRCNN-style models: Squeeze axes, shape Gather, and dynamic TopK #19965 also covers PP-OCRv6 import limitations, but its reported failures are Squeeze/shape-graph/TopK related rather than this symbolic MaxPool padding path.

Triage

• needs-triage
• type: bug