Skip to main content

Mojo module

grouped_tile_scheduler

Grouped tile scheduler for SM100 structured block-scaled GEMM.

This scheduler extends the SM100 TileScheduler to support grouped GEMM with variable problem sizes per group. It uses linear tile iteration instead of CLC (Cluster Launch Control) to map a global linear tile index to group-specific coordinates.

Key features:

  • GroupedWorkInfo: Extends WorkInfo with group_idx, k_tile_count, group_changed
  • delinearize_to_group(): Maps linear tile index to group + local coordinates
  • Supports variable M, N, K per group
  • Compatible with dynamic tensormap updates

Usage: var scheduler = GroupedTileScheduler[...](problem_sizes, tile_shape) var work_iter = scheduler.work_iterator() while work_iter.has_work(): with work_iter.next() as current: if current.group_changed: update_tensormaps(current.group_idx) process_tile(current)

Structs

Was this page helpful?