StridedMemoryView.__dlpack__ raises BufferError for bfloat16 views #3337

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Runtime Error

Component

cuda.core

Describe the bug

A StridedMemoryView with an ml_dtypes.bfloat16 dtype cannot be exported over DLPack. _smv_dtype_numpy_to_dlpack (cuda/core/_memoryview.pyx) picks the type code from numpy.dtype.kind, and numpy.dtype(ml_dtypes.bfloat16).kind is "V", so bf16 falls through to the error. DLPack import already accepts kDLBfloat, the dtype property documents bfloat16, and _tensor_map.pyx already maps ml_dtypes.bfloat16 to kDLBfloat. So a bf16 tensor from PyTorch, CuPy or JAX can be viewed but not re-exported.

test_dlpack_export_roundtrip_dtypes[bfloat16] does not catch this because its NumPy consumer skips bf16.

How to Reproduce

cuda-core 1.2.2, ml_dtypes, numpy (host memory is enough):

import numpy as np, ml_dtypes
from cuda.core.utils import StridedMemoryView

bf16 = np.dtype(ml_dtypes.bfloat16)
src = np.zeros((2, 3), dtype=np.uint16)
view = StridedMemoryView.from_any_interface(src, stream_ptr=-1).view(dtype=bf16)
view.__dlpack__()

BufferError: Unsupported dtype for DLPack export: bfloat16

float16 and uint16 views of the same array export fine.

Expected behavior

The view exports with type code kDLBfloat and 16 bits, and StridedMemoryView.from_dlpack on the result gives a bfloat16 view of the same memory.

Operating System

Ubuntu 24.04, aarch64

nvidia-smi output

Not GPU-specific; host memory is enough.

Proposed fix

I have a fix ready: map ml_dtypes.bfloat16 to kDLBfloat before the kind dispatch, plus a host-memory round-trip test. With a source build of main and of the fix, the snippet above raises on main, and with the fix the export succeeds and re-imports as bfloat16 with the same pointer; an unsupported V4 dtype is still rejected. Happy to open a PR if you assign this to me.