[RFC]: 删除platform

动机（Motivation）.

1. 背景与现状

hyper_parallel/platform/ 当前的架构是 双实现 + 单一分发器：

• platform/platform.py（约 67KB）定义抽象的 Platform 基类，契约面高达 100+ 个属性（collective、differentiable_* 异步族、tensor 工厂、checkpoint/swap、stream/random、param 自省、custom_ops 等）。
• platform/mindspore/platform.py（约 83KB）与 platform/torch/platform.py（约 64KB）是两套各自独立、平行实现的具体平台类。
• 分发逻辑 get_platform()（platform/platform.py:99-126）返回全局单例，默认 MindSpore，可被环境变量 HYPER_PARALLEL_PLATFORM 覆盖，失败时 ImportError 回退 torch：

# platform.py
try:
    return get_mindspore_platform()
except ImportError:
    return get_torch_platform()

• 整个代码库（core/、trainer/、models/、collectives/、integration/、dmodule/）几乎全部通过 from hyper_parallel import get_platform 拿到单例，再在模块顶层绑定 Tensor = platform.Tensor、platform.get_rank() 等使用。core/ 层不直接 import torch 或 import mindspore，而是通过接缝与框架解耦。

2. 本次任务目标

删除platform，代码直接使用torch进行实现。

目标设计.

删除platform，代码直接使用torch进行实现。

需要设计评审DFX建议.

NA

相关的RFCs和API

NA

完整的反馈期限.

CC List.

其他补充说明.

Thanks for contributing 🎉! The hyper-parallel core team hosts a biweekly RFC review session, while most RFCs can be discussed online, you can optionally sign up for a slot to discuss your RFC online.

Before submitting a new issue...

• Make sure you already searched for previous RFCs.

schema_version: 1
source: gitcode
gitcode_repo: mindspore/hyper-parallel
gitcode_issue: 377
source_url: https://gitcode.com/mindspore/hyper-parallel/issues/377

Start with platform/platform.py and the parallel implementations under platform/mindspore/ and platform/torch/, then trace get_platform usage across core/, trainer/, models/, collectives/, integration/, and dmodule/. Done means the platform abstraction and dispatcher are removed, callers use torch directly, and the affected functionality remains covered by the repository's existing checks.