Closed
[Requirement|需求建议]: engram正向训练算子量化优化 #7069
wuchenhao123Add on  12 days agoClosed at  12 days ago
wuchenhao123
wuchenhao123Collaborators
12 days ago Create

Backgroud(背景信息)

engram正向训练算子量化优化

Benefit / Necessity (价值/作用)

engram正向训练算子量化优化

Design(设计方案)

取消了批量拷贝的32b对齐限制

likedislike
wuchenhao123wuchenhao123Collaborators
12 days ago Added label: requirement
wuchenhao123wuchenhao123Collaborators
12 days ago Assign wuchenhao123 as responsible person
wuchenhao123wuchenhao123Collaborators
12 days ago Associated pull request
CANN-robotCANN-robotCollaborators
12 days ago Close issue
CANN-robotCANN-robotCollaborators
12 days ago Added label: resolved
周云飞
周云飞Collaborators
5 days ago Comments:

已合入 PR #15208,针对 Engram 训练正向量化场景进行了性能优化:

  • 在行宽未按 32B 对齐时,将 staging 聚合、接收散射和本地直取的逐行同步搬运改为分批搬运,以减少同步次数;超出 tile 容量等情况保留逐行兜底。
  • 在训练分支的 tiling 阶段增加行大小校验,避免 data 和 scale factor 对齐后的大小超过 UINT32_MAX,导致 kernel 内计算回绕。

关联 PR:PR #15208

likedislike