[RFC]: Updates for float16/bfloat16 and for dtypes that are lacking full support in libraries

从 Array API standard 2025.12 的 data_types 页面开始，将其中的 dtype 定义与所链接的 NumPy、PyTorch、JAX、CuPy、MLX、ndonnx、DPCTL 和 cuTile 参考资料进行比较。明确哪些文档需要更新，保留 float16 和 bfloat16 这两个名称，并记录仅存储支持或部分支持；完成的修改应准确反映比较结果，并处理所列出的注意事项。

This issue is meant to provide context for changes to dtype support in the next version of the standard.

Overview of data types implemented in various array libraries:

• Array API standard v2025.12: https://data-apis.org/array-api/2025.12/API_specification/data_types.html
• NumPy: https://numpy.org/devdocs/user/basics.types.html#relationship-between-numpy-data-types-and-c-data-types
• PyTorch: https://docs.pytorch.org/docs/stable/tensor_attributes.html#torch-dtype
• JAX: dtype names in https://docs.jax.dev/en/latest/jax.numpy.html plus bfloat16 in https://docs.jax.dev/en/latest/jax.dtypes.html. See note on float64 in https://docs.jax.dev/en/latest/default_dtypes.html
• CuPy: dtype names in https://docs.cupy.dev/en/stable/reference/comparison.html, plus minimal bfloat16 support (see https://docs.cupy.dev/en/stable/upgrade.html#minimal-support-for-bfloat16)
• MLX: https://ml-explore.github.io/mlx/build/html/python/data_types.html
• ndonnx: https://ndonnx.readthedocs.io/en/latest/datatypes/datatypes.html
• DPCTL: https://intelpython.github.io/dpctl/latest/api_reference/dpctl/tensor.data_types.html
• cuTile: https://docs.nvidia.com/cuda/cutile-python/data.html#data-types
• ml_dtypes: https://github.com/jax-ml/ml_dtypes/blob/main/README.md

Summary of dtype support across array libraries

Legend: ✓ = full support, ○ = partial support, ✗ = no support

Notes:

1. PyTorch uint16/uint32/uint64 are "shell dtypes" with limited op and backend support; primarily for use with torch.compile.
2. MLX float64 only works with CPU operations; using float64 on the GPU raises an exception.
3. NumPy bfloat16 support is available via the ml_dtypes package, not natively.
4. CuPy has minimal bfloat16 support via ml_dtypes.bfloat16; some gaps remain especially in cupyx.
5. PyTorch has torch.complex32 defined, but operator coverage is limited.
6. JAX 64-bit dtypes require opt-in via jax.config.update('jax_enable_x64', True) or the JAX_ENABLE_X64 env var; disabled by default, and 64-bit values are silently truncated to 32-bit without it.
7. DPCTL float64/complex128 are device-dependent; only available when the target device's has_aspect_fp64 property is True.
8. DPCTL float16 is device-dependent; only available when the target device's has_aspect_fp16 property is True.

Conclusions

• Only a few dtypes have no caveats at all: bool, int8/int16/int32, uint8, float32
• Support for float64 is the biggest issue: very important for scientific computing and other fields that require high accuracy, not available at all or CPU-only on several deep learning-focused libraries
• Complex dtype support is generally spotty
• float16 and bfloat16 don't have universal support yet, but are consistently named

Next steps

Open a PR for discussion which brings documentation on data more in line with reality, reserves the float16/bfloat16 names, and says something about dtype support that is storage-only or partial.