mindspore-ai / mindspore-ai/hyper-parallel
[Clean Code] 整改 Linear Attention GDN/KDA 代码扫描告警
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 53
- Forks
- 63
- Avg merge
- 23h 45m
- Merged PRs (30d)
- 63
Description
背景
代码扫描表中分配给作者的 Linear Attention GDN/KDA 告警,需要在不改变融合算子算法、Context Parallel 通信和训练数值的前提下完成低风险整改。GitCode PR #1477 与 #1393 尚未在 GitHub 主仓形成原生 PR,本 Issue 用于跟踪迁移后的独立整改。
范围
- GDN/KDA 模块和 Context Parallel host 侧低风险扫描整改。
- 修正
__all__、编码声明、显式异常、普通 Python 临时变量命名及可独立抽取的 shape/ShortConv helper。 - 在 KDA P2P backward 中及时释放最后消费者之后的重计算临时张量,使后续 backward 分配可以复用 storage。
- 保留 Triton JIT 参数、tile 符号、kernel 数量、launch shape、autograd 返回槽位和 CP 消息顺序。
本批不包含 FLA runtime adapter、第三方依赖切换、Triton-Ascend 编译器修复、kernel 算法重构或大 CP 实验代码;无法安全修改的 kernel 告警继续走精确豁免或后续独立整改。
验收
git diff --check、改动文件 compile/Pylint 和相关 CPU UT 通过。- GDN/KDA 的 P2P Triton、FSDP、activation checkpoint 多卡训练精度回归通过。
- KDA backward 临时张量释放完成相同 shape 的显存 A/B,且不引入稳定性能回退。
- PR 相对最新
master保持单一、可审核的业务 commit。
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the GDN/KDA modules and Context Parallel host-side code described in the issue, then review the existing GDN/KDA P2P backward and shape/ShortConv helper paths. Run git diff --check, compile/Pylint and related CPU unit tests first; completion requires the listed Triton, FSDP, activation-checkpoint multi-card accuracy regressions and the KDA temporary-tensor memory A/B without a stable performance regression.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- distributed-systems, machine-learning, performance
- Issue type
- Refactor
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Active
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100