[ONNX] Fix initializer prefix stripping (#20323)

The Relax ONNX importer used `str.strip("onnx::")` when
`keep_params_in_input=True`. Since `strip` treats its argument as a set
of characters, initializer names that do not have the prefix can still
lose leading or trailing `o`, `n`, `x`, or `:` characters. For example,
`neck...` became `eck...`.

This change uses `str.removeprefix("onnx::")` so only the exact
PyTorch-generated prefix is removed. The regression test covers both a
prefixed initializer and an ordinary initializer beginning with `n`.

Fixes #20290

Tests:
- `python -m pytest tests/python/relax/test_frontend_onnx.py -k
'params_names_start_with_onnx or
initializer_name_only_removes_onnx_prefix or concat_with_param or
multi_ops_with_same_params' -q`
- `python -m ruff check python/tvm/relax/frontend/onnx/onnx_frontend.py
tests/python/relax/test_frontend_onnx.py`
- `python -m ruff format --check
python/tvm/relax/frontend/onnx/onnx_frontend.py
tests/python/relax/test_frontend_onnx.py`

Local full-file note: the complete ONNX test file reached 28 passing
tests before hitting the unrelated Windows JIT export assertion `Target
triple should not be empty`.

Author: Nanmur

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

@@ -6354,7 +6354,7 @@ def _parse_graph_initializers(self, graph: onnx.onnx_ml_pb2.GraphProto):
             # Create variables for constants.
             if self._keep_params_in_input:
                 # Pytorch sometimes inserts silly weight prefix. Remove it.
-                var_name = init_tensor.name.strip("onnx::")
+                var_name = init_tensor.name.removeprefix("onnx::")
                 init_var = self._new_var(var_name, shape=array.shape, dtype=array.dtype)
                 self._nodes[init_tensor.name] = init_var
                 # We need to keep track of both the real value and variable for this variable.

tests/python/relax/test_frontend_onnx.py

@@ -11797,6 +11797,36 @@ def main(
     tvm.ir.assert_structural_equal(tvm_model, Expected)
 
 
+@pytest.mark.parametrize(
+    ("initializer_name", "expected_name"),
+    [
+        ("onnx::weight", "weight"),
+        (
+            "neck.lateral_convs.2.conv2.weight_quantized",
+            "neck.lateral_convs.2.conv2.weight_quantized",
+        ),
+    ],
+)
+def test_initializer_name_only_removes_onnx_prefix(initializer_name, expected_name):
+    graph = helper.make_graph(
+        [helper.make_node("Add", ["input", initializer_name], ["output"])],
+        "test_initializer_name_only_removes_onnx_prefix",
+        inputs=[helper.make_tensor_value_info("input", TensorProto.FLOAT, [1])],
+        initializer=[numpy_helper.from_array(np.ones([1], dtype="float32"), initializer_name)],
+        outputs=[helper.make_tensor_value_info("output", TensorProto.FLOAT, [1])],
+    )
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 14)])
+    model.ir_version = 8
+
+    tvm_model = from_onnx(
+        model,
+        keep_params_in_input=True,
+        sanitize_input_names=False,
+    )
+
+    assert tvm_model["main"].params[-1].name == expected_name
+
+
 def test_shape_dim_string_expression_graph_add():
     identity_node = helper.make_node("Identity", ["x"], ["y"])