Repository navigation
update(perf): record the wrapped-window Gemma 3 decode measurement - #2275
Merged
Merged
Conversation
The decode profile script hardcoded the 512-token prompt shape, so profiling a wrapped sliding window meant editing the script. --prompt-tokens and --max-tokens take the issue-2192 isolation shape (pp2048/tg128) without an edit.
Repeating the issue-2239 three-arm measurement on the current tree does not reproduce the pp2048 regression: the pipelined arm is -0.13 percent against the pre-epic loop, inside the null arm's spread, and a build of origin/main at e24fb7b is at parity too (+0.29 percent), so the commits merged since the Oct 8 session (#2271, #2272) recovered the gap. MLXCEL_FORCE_SYNC on the same binary now measures below the pipelined arm where Oct 8 measured it above, so the synchronous step is no longer the faster choice in this bench. The page and the session scripts record the four-arm run, the pre-fix comparison and the pp512 spot check.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Docs-only measurement record for issue #2239: the reported pipelined B=1 decode regression on Gemma 3 (gfx1151, pp2048/tg128) does not reproduce on the current tree, so no code change ships.
docs/benchmark_results/wrapped-ring-lookahead-gfx1151-2026-10-12.mdrecords the paired measurements, with session scripts and raw CSV underdocs/benchmark_results/data/.Measurements
Four-arm interleaved session (pre-epic
4c44e317base, branch pipelined,MLXCEL_FORCE_SYNC=1, base noise floor), 12 runs, all CLEAN under the GPU guard: base median 62.61 tok/s, pipelined branch 62.63 (paired median -0.13 percent), sync 61.10 (-2.41 percent), noise floor +0.00 percent (range -0.18 to +0.14). A follow-up arm on origin/main ate24fb7b2without the branch's changes measures 62.79 (+0.29 percent), so the Oct 8 -1.3 percent gap was already recovered by #2271 and #2272 before this branch existed. The pp512 cell stays within the 1.0 percent threshold (+0.78 percent). A wrapped-window synchronous handover prototype carried during measurement was dropped: it never routes this bench path and its destination measures below the pipeline today.Test plan
make verify-rocm-held): 12264 passed, 0 failed, 404 ignoredengine-parity: 14 pairs, 13 identical, 1 n/aCloses #2239