Closed
[Requirement|需求建议]: mla_prolog算子支持RoPE开关 #4133
zhengyuhao3Add on Aug 4Closed at Aug 18
zhengyuhao3
Aug 4 Comments:
Aug 4 Comments:
/assign


Aug 4 Assign zhengyuhao3 as responsible person
Aug 8 Associated pull request
Zzhengyuhao3
Aug 13 Change title to "[Requirement|需求建议]: mla_prolog算子支持RoPE开关", originally titled "[Requirement|需求建议]: mla_prolog_v3算子支持RoPE开关".
Aug 13 Change title to "[Requirement|需求建议]: mla_prolog算子支持RoPE开关", originally titled "[Requirement|需求建议]: mla_prolog_v3算子支持RoPE开关".
Aug 13 Modify description of issue
Aug 18 Modify description of issue
Aug 18 Close issue
Aug 18 Added label: resolved
Aug 20 Associated pull request
Aug 20 Associated pull request
Aug 21 Modify description of issue
Aug 21 Associated pull request
Aug 24 Associated pull request
Thanks for sending an requirement! Please fill in the following template to help quickly solve your problem.
Backgroud(背景信息)
Kimi-K3模型MLA前置操作去除了rope部分,需要在原有的MLAPO算子基础上增加rope部分的开关
Benefit / Necessity (价值/作用)
价值:通过开关使能rope操作,Kimi-K3模型性能提升
应用场景:Kimi-K3相关模型部署场景。
Design(设计方案)
新增aclnn接口名:aclnn_mla_prolog_v4_weight_nz
mla_prolog_v3 算子def文件新增可选开关属性值 do_rope
torch_extension: mla_prolog
实现方案:
新增aclnn接口aclnn_mla_prolog_v4_weight_nz,较v3版本上新增doRope参数,用于开关计算流中的rope操作
kernel内部:
通过模板参数doRope=true/false,决定RoPE操作进行/不进行