RFC: add `"mutable arrays"` to `capabilities`

未关闭

#845 创建于 2024年9月23日 21 条评论 4 个 reaction 已指派 0 人在 GitHub 查看

还没有人认领这个 Issue。

评估

难度: 5/5
预计耗时: 一周以上
新手友好度: 35/100

Issue 类型: 功能
描述清晰度: 需要澄清
活跃度: 冷清
技术栈: python
领域: documentation

调研方向

先阅读链接的 array_api.info.capabilities 文档，然后阅读包含 21 条评论的 RFC 讨论串，以了解围绕可变性 capability 尚未解决的设计问题。当项目就该 capability 作出决定，并将达成一致的行为记录在 Array API 标准中时，即视为完成。

由索引模型根据 Issue 内容生成。

描述

API extension Needs Discussion RFC topic: Inspection

Several parts of the Array API standard assume that array objects are mutable. Some array API implementations (notably JAX) do not support mutating array objects. This has led to array API implementations currently being developed in scipy and sklearn to be entirely unusable in JAX.

Given this, downstream implementations have a few choices:

Use mutability semantics, excluding libraries like JAX.
Avoid mutability semantics to support libraries like JAX.
Explicitly special-case arrays of type jax.numpy.Array, changing the implementation logic for that case.

(1) is a bad choice, because it means JAX will not be supported. (2) is a bad choice, because for libraries like NumPy, it leads to excessive copying of buffers, worsening performance. (3) is a bad choice because it hard-codes the presence of specific implementations in a context that is supposed to be implementation-agnostic.

One way the Array API standard could address this is by adding "mutable arrays" or something similar to the existing capabilities dict. Then downstream implementations could use strategy (3) without special-casing particular implementations.

主要语言: Python
星标: 281
派生: 52
PR 合并指标: 30 天内没有已合并 PR