Skip to content

[feature] enable MLA down proj fusion if no regression #3396

@dingqingy-nv

Description

@dingqingy-nv

User problem

Merged MCore#3039
Not expect perf gain with current cudnn, but can enable if no regression.

Desired outcome

Test it's effect as well as combination with enable/disable cudnn LN.

Alternatives considered

No response

Affected area

area:perf

Urgency / use case

Nice to have

Extra context

No response

Metadata

Metadata

Assignees

Labels

26.04.01featureNew capabilities, enhancements, or enablement workperformance/releasePerformance items related with NeMo release

Projects

No projects

Relationships

None yet

Development

No branches or pull requests

Issue actions