Repository navigation
[Diffusion] Optimize LTX-2 QKNorm and split RoPE on Hopper - #38396
Merged
Merged
Conversation
BBuf
requested review from
AgainstEntropy,
DarkSharpness,
HaiShaw,
HydraQYH,
celve,
mickqian,
ping1jing2,
yichiche,
yingluosanqian and
yuan-luo
as code owners
September 8, 2026 02:50
Collaborator
|
/tag-and-rerun-ci |
Collaborator
Author
This was referenced Sep 19, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
quality="high"or"extra-high"quality="lossless"on the current eager chain; keep the existing SM100+ default unchangedProfile evidence
Workload:
Lightricks/LTX-2, 768x512x121, 40 first-stage + 3 refinement steps, CFG parallel on 2x H200, eager mode.The source all-stage trace recorded 41,316 BF16 LayerNorm launches (271.51 ms) and 16,512 split-RoPE launches (85.37 ms). A two-step candidate trace records 1,152 fused kernel launches, zero split-RoPE launches, and removes a net 192 launches per step.
End-to-end result
Fixed H200 GPUs, two fresh-process repeats per side:
quality=highquality=highThe candidate repeats are byte deterministic. The fused operator differs from the Hopper eager BF16 chain by at most 0.0625 for Q and 0.125 for K, which is why this path is quality-gated instead of enabled for lossless requests. Candidate versus source high-mode video measures PSNR 25.5355 dB / FFmpeg SSIM 0.897576; visual review preserves the scene, composition, and broad motion. Source high mode itself differs more from lossless by frame metrics.
Validation
test_rope_ltx2.pyon H200: 2 passed, 5 Blackwell-only cases skippedtest_model_fast_paths.py -k ltx2on H200: 6 passedCI States
Latest PR Test (Base): ⏳ Run #34181428105
Latest PR Test (Extra): ❌ Run #34181428049
Latest PR Test (AMD ROCm 7.2): ❌ Run #34181428196