Repository navigation
[diffusion] CI: fix qwen cache-dit step 10/49 perf baselines - #42495
Merged
Merged
Conversation
…tep baselines - ltx_2_5_diffusion_decoder_2gpus: component-offloading the DiT made denoise randomly 1.4-1.9x slower per server launch. Keep the DiT and text encoder resident and re-baseline the case from resident CI runs. - qwen_image_t2i_cache_dit_enabled: steps 10 and 49 are first cached steps (~65 ms on every H100 runner); set them to the CI median.
niehen6174
requested review from
AgainstEntropy,
BBuf,
HaiShaw,
kevin-mii,
mickqian,
ping1jing2 and
yichiche
as code owners
October 4, 2026 12:03
niehen6174
added
diffusion
…ts timing tolerance The case exists to cover the offloaded DiT, so restore its original residency and baselines. Denoise is randomly 1.4-1.9x slower per launch with the DiT offloaded; a 1.0 timing_tolerance keeps the case green until that is fixed. The qwen cache-dit step 10/49 baseline fix is unchanged.
niehen6174
changed the title
niehen6174
marked this pull request as draft
October 4, 2026 15:30
niehen6174
marked this pull request as ready for review
October 5, 2026 07:59
The ltx_2_5 host-bound slowdown will be handled in a separate PR; this PR now only fixes the qwen cache-dit step 10/49 baselines.
niehen6174
changed the title
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivation
qwen_image_t2i_cache_dit_enabledfails its H100 per-step perf check at random. From 09-30 to 10-04, 71% of attempts went over the step-49 limit. Every final 1-GPU failure in that window came from this case.The cause, found in the debug runs in #42546, is how the step time is measured. It is not a regression.
Debug run on CI H100, 3 launches per arm, times in ms:
SGLANG_DIFFUSION_SYNC_STAGE_PROFILING=1With synced timing, steps 10 and 49 take about 7 ms, like every other cached step. #38584 enabled the fused epilogue on Hopper, which cut host launches from 3031 to 2007 per profiled window. The real GPU time got about 3% faster. But the host now enqueues faster too, so more GPU work spills into steps 10 and 49.
Modifications
perf_baselines/h100.json: steps 10 and 49 ofqwen_image_t2i_cache_dit_enabledare set to their CI medians, 64.58 / 66.38 ms (previously 48.7 / 45.11). The medians come from 328 H100 CI measurements between 09-30 and 10-04. They are the unsynced host wall times this check measures. No other baseline or tolerance changes.The
ltx_2_5_diffusion_decoder_2gpustiming tolerance from earlier commits is dropped. Its host-bound slowdown will be handled in a separate PR.Validation
PR CI on fe3b48d passed with the same qwen baselines.
🤖 Generated with Claude Code
CI States
Latest PR Test (Base): ✅ Run #37476814004
Latest PR Test (Extra): ❌ Run #37476813352
Latest PR Test (AMD ROCm 10): ❌ Run #37476814266