Closed
[Documentation|文档反馈]: 补充aclnnQuantMatmulV5的MX与T-CG量化约束 #5368
sunduiyangAdd on  Aug 31Closed at  Aug 31
sunduiyang
sunduiyangCollaborators
Aug 31 Create

Document Link(文档链接)

https://gitcode.com/cann/ops-nn/blob/master/matmul/quant_batch_matmul_v4/docs/aclnnQuantMatmulV5.md

Issues Section(问题文档片段)

“约束说明”中的“MX量化场景约束”和“T-CG量化场景约束”。

Existing Issues(存在的问题)

现有约束未完整说明MX伪量化与T-CG量化场景下x1、x2的数据格式及转置要求:

  • MX伪量化场景下,当x2为FLOAT4_E2M1时,需明确k是8的倍数;x1仅支持非转置和ND格式;x2仅支持转置,支持ND、NZ格式,且NZ格式下n是8的倍数。
  • T-CG量化场景下,需明确k是32的倍数;x1仅支持非转置和ND格式;x2为ND格式时仅支持转置,为NZ格式时仅支持非转置。
likedislike
sunduiyangsunduiyangCollaborators
Aug 31 Added label: documentation
sunduiyangsunduiyangCollaborators
Aug 31 Associated pull request
sunduiyang
sunduiyangCollaborators
Aug 31 Comments:

/assign

likedislike
CANN-robotCANN-robotCollaborators
Aug 31 Assign sunduiyang as responsible person
CANN-robotCANN-robotCollaborators
Aug 31 Close issue
CANN-robotCANN-robotCollaborators
Aug 31 Added label: resolved