Skip to content

Dynamic sparse attention #22

@GandalfTea

Description

@GandalfTea

Implement and benchmark blocked dynamic sparse attention modules and metal kernels, as well as a custom paged KV-cache

Metadata

Metadata

Assignees

Labels

enhancementNew feature or requestperformancePerformance optimizations and resource efficiencywontfixThis will not be worked on

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions