Skip to content

[Feat]: Add dispatch & combine tuning configuration#772

Open
yanboshao wants to merge 16 commits into
mainfrom
yanbo/combine_perf_tuning
Open

[Feat]: Add dispatch & combine tuning configuration#772
yanboshao wants to merge 16 commits into
mainfrom
yanbo/combine_perf_tuning

Conversation

@yanboshao

@yanboshao yanboshao commented Jun 30, 2026

Copy link
Copy Markdown
Contributor

Motivation

Technical Details

  1. Added tuning configs for dispatch and combine operators
  2. Optimized combine performance when zero-copy = False
  3. Refactored dispatch and combine using high-level FlyDSL APIs

Test Plan

Test Result

Submission Checklist

yanboshao and others added 11 commits June 29, 2026 07:28
…fall

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 1d56f32)
…mized kernel

Co-authored-by: Cursor <cursoragent@cursor.com>
(cherry picked from commit 099f1c1)
Profile/cudagraph path resolved combine geometry via the op tuning table and
never forwarded --combine-block-num/--combine-warp-per-block, making config
sweeps no-ops. Pin the CLI geometry (when both are given) onto every FlyDSL
combine() call; default (unset) still falls back to the op tuning table.

(cherry picked from commit 265aed0)
…Vec/shuffle/filled, dedup fp8 pack, clean unused)
@yanboshao yanboshao changed the title combine perf tuning [Feat]: Add dispatch & combine tuning configuration Jun 30, 2026
Refresh r1/pro/flash bf16 combine block_num/warp_per_block from the
full profiler+cudagraph sweep; add missing num_tokens buckets
(pro/flash 512/1024/2048, r1 8192).
@yanboshao yanboshao marked this pull request as ready for review July 1, 2026 22:24
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant