Adds INT2 (2-bit signed) and UINT2 (2-bit unsigned) data type

[vraspar]

Adds INT2 (2-bit signed) and UINT2 (2-bit unsigned) data types following ONNX spec PR #7446. These complement existing INT4/UINT4 support with similar packed representation.

Changes

Enums (_enums.py)

• Added INT2=25, UINT2=26 enum values
• Added ml_dtypes.int2/uint2 mappings and type properties (bitwidth=2, signed/unsigned)
• Added short names "i2"/"u2"

Packing (_type_casting.py)

• Implemented pack_2bitx4() / unpack_2bitx4() for 4-values-per-byte packing
• Handles non-multiple-of-4 sizes with padding

Core (_core.py)

• Updated PackedTensor to support 2-bit types (was 4-bit only)
• Added INT2/UINT2 to non-numpy-native types list
• Updated byte representation to pack 2-bit types

Serialization (serde.py)

• Added 2-bit unpacking from raw_data
• Added INT2/UINT2 to int32_data serialization paths

Torch adapters (tensor_adapters.py)

• Added torch.int2/uint2 dtype mappings for future compatibility
• PyTorch doesn't yet support creating tensors with these types

Example

import numpy as np
import ml_dtypes
import onnx_ir as ir

# Create and serialize INT2 tensor
array = np.array([-2, -1, 0, 1], dtype=ml_dtypes.int2)
tensor = ir.Tensor(array)
proto = ir.serde.to_proto(tensor)

# 4 elements packed into 1 byte
assert len(proto.raw_data) == 1

# Round-trip works
restored = ir.serde.from_proto(proto)
assert np.array_equal(restored.numpy(), array)

Original prompt

Helpe me create PR for this issue: #250

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[codecov]

❌ 2 Tests Failed:

Tests completed	Failed	Passed	Skipped
947	2	945	0

View the top 2 failed test(s) by shortest run time

::src.onnx_ir.serde_test

Stack Traces | 0s run time

src\onnx_ir\serde_test.py:146: in <module>
    class TensorProtoTensorTest(unittest.TestCase):
src\onnx_ir\serde_test.py:278: in TensorProtoTensorTest
    @parameterized.parameterized.expand(
.nox\test\Lib\site-packages\parameterized\parameterized.py:585: in parameterized_expand_wrapper
    raise ValueError(
E   ValueError: Parameters iterable is empty (hint: use `parameterized.expand([], skip_on_empty=True)` to skip this test when the input is empty)

src.onnx_ir.serde_test.TensorProtoTensorTest::test_tensor_proto_tensor_uint_5_INT2

Stack Traces | 0.003s run time

.nox\test_onnx_weekly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
src\onnx_ir\serde_test.py:336: in test_tensor_proto_tensor_uint
    np.testing.assert_array_equal(np.from_dlpack(tensor), tensor.numpy())
                                  ^^^^^^^^^^^^^^^^^^^^^^
src\onnx_ir\serde.py:351: in __dlpack__
    return self.numpy().__dlpack__(stream=stream)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E   BufferError: DLPack only supports signed/unsigned integers, float and complex dtypes.

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[justinchuby]

I realized we don’t need to handle endianess because these are subtbyte types. So endianess is irrelevant

[justinchuby]

@vraspar could you check the test errors? Thanks