Test using bfloat16 and model input/output

[josel-amd]

Bug Report

Is the issue related to model conversion?

No

Describe the bug

Tests that use bfloat16 as model input/ouput (listed in [0]) are disabled in most backends (see e.g. here). On one side, this can be problematic for other backends and, on the other, it becomes unclear what is the level of support for bfloat16.

A quick look at the test runner shows that there is support forTensorProto.BFLOAT16 where the data - in that format - gets converted to float32 through numpy_helper.bfloat16_to_float32. The problem with the test cases mentioned before is that the output_*.pb files seem to be read the reference data as TensorProto.UINT16, which then never triggers the expected conversion to float32.

I'm not sure if converting to float32 is the expected behavior but I'm assuming so, for the purpose of this discussion.

Related issue

[0]

test_castlike_BFLOAT16_to_FLOAT
test_castlike_FLOAT_to_BFLOAT16
test_cast_FLOAT_to_BFLOAT16

System information

• OS Platform and Distribution (e.g. Linux Ubuntu 20.04): Linux Ubuntu 22.04
• ONNX version (e.g. 1.13): 1.16.1
• Python version: 3.10
• Protobuf version: 5.26.1

Expected behavior

Unsure. By looking at the code, the output for the tests that return bfloat16 should be in np.float32 not in np.uint16 format

[justinchuby]

Thanks. We can enable bfloat16 support with the ml_dtypes package in the reference evaluator.