Skip to content

chore(EPv2): split dispatch/combine tuning schedules into independent tables - #533

Open
kawhil-amd wants to merge 2 commits into
mainfrom
epv2_tune
Open

chore(EPv2): split dispatch/combine tuning schedules into independent tables#533
kawhil-amd wants to merge 2 commits into
mainfrom
epv2_tune

Conversation

@kawhil-amd

Copy link
Copy Markdown
Contributor

Summary

  • Split the unified 5-tuple per-token schedule (max_tok, disp_block, disp_warp, comb_block, comb_warp) into two independent 3-tuple schedules:
    dispatch_schedule and combine_schedule, each (max_tok, block, warp)
  • Each phase now has its own per-shape lookup table (_DISP_TABLE / _COMB_TABLE), keyed independently by (world_size, hidden_dim, topk)
  • lookup() queries the two tables separately, enabling one phase to reuse a schedule across topk/dtype while the other varies
  • Refactored _pick() into a shared _pick_bucket() static method + independent dispatch/combine bucket selection
  • Trimmed measurement dump comments (GB/s, us, sweep methodology) from schedule definitions; kept structural notes and guardrails

@kawhil-amd kawhil-amd changed the title chore(EPv2): split dispatch/combine tuning schedules into independent chore(EPv2): split dispatch/combine tuning schedules into independent tables Aug 9, 2026
The previous commit split dispatch/combine tuning schedules but forgot
to define _MI355X_COMB_TABLE, causing ruff F821 (undefined name).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant