[BUG] Sparse initializers causes Type Error for SparseToDenseMatMul

[martinkjlarsson]

Describe the issue

I've created an ONNX model consisting solely of a SparseToDenseMatMul operator. A (data1) is a sparse initializer, and B (input1) is an input. Running the model results in the error:

This is an invalid model. Type Error: Type 'tensor(float)' of input parameter (data1) of operator (SparseToDenseMatMul) in node (SparseToDenseMatMul1) is invalid.

It says data1 is a tensor and not a sparse_tensor. However, when I inspect the model using Netron, it shows "category: Initializer" and "layout: sparse".

To reproduce

Attached are two models: sparse_matmul_julia.onnx created using my Julia code, and sparse_matmul.onnx created using some LLM-generated Python code below. The models are not identical but give the same error.
models.zip

import numpy as np
import onnx
from onnx import helper, TensorProto
import onnxruntime as ort

A_values = helper.make_tensor(
    name="data1",
    data_type=TensorProto.FLOAT,
    dims=[4],
    vals=[1.0, 2.0, 3.0, 4.0],
)

A_indices = helper.make_tensor(
    name="data1_indices",
    data_type=TensorProto.INT64,
    dims=[4, 2],
    vals=[
        0, 0,
        0, 2,
        1, 1,
        1, 3,
    ],
)

A = onnx.SparseTensorProto()
A.values.CopyFrom(A_values)
A.indices.CopyFrom(A_indices)
A.dims.extend([2, 4])

B = helper.make_tensor(
    name="input1",
    data_type=TensorProto.FLOAT,
    dims=[4, 3],
    vals=[
         1.0,  2.0,  3.0,
         4.0,  5.0,  6.0,
         7.0,  8.0,  9.0,
        10.0, 11.0, 12.0,
    ],
)

Y = helper.make_tensor_value_info(
    "output",
    TensorProto.FLOAT,
    [2, 3],
)

node = helper.make_node(
    "SparseToDenseMatMul",
    inputs=["data1", "input1"],
    outputs=["output"],
    name="SparseToDenseMatMul1",
    domain="com.microsoft",
)

graph = helper.make_graph(
    nodes=[node],
    name="SparseMatMulExample",
    inputs=[],
    outputs=[Y],
    initializer=[B],
    sparse_initializer=[A],
)

model = helper.make_model(
    graph,
    producer_name="python-test",
    opset_imports=[
        helper.make_operatorsetid("", 21),
        helper.make_operatorsetid("com.microsoft", 1),
    ],
)

# Check the protobuf structure.
onnx.checker.check_model(model)

onnx.save(model, "sparse_matmul.onnx")

# Load and run the model with onnxruntime
sess = ort.InferenceSession("sparse_matmul.onnx", providers=["CPUExecutionProvider"])
outputs = sess.run(None, {})
print(outputs[0])

Urgency

No urgency.

Platform

Windows

OS Version

Microsoft Windows 11 Home

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.24.4

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response