[https://nvbugs/6572800][fix] Update visual generation LPIPS goldens

[yibinl-nvidia]

Dev Engineer Review

• Updated SHA-256 values in seven visual-generation LPIPS golden JSON files.
• JSON structure remains unchanged.
• Removed 13 obsolete skip entries from tests/integration/test_lists/waives.txt.
• No code, API, performance, or error-handling changes.
• Verdict: sufficient.

QA Engineer Review

• No test-db/ or qa/ files were modified.
• Removed 13 obsolete entries from waives.txt: one Nemotron NVFP4 test, 11 visual-generation accuracy or LPIPS tests, and one MiniMax multi-node evaluation test.
• CBTS coverage data is unavailable.
• Verdict: needs follow-up.

Description

Test Coverage

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[coderabbitai]

Note

Reviews paused

It looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the reviews.auto_review.auto_pause_after_reviewed_commits setting.

Use the following commands to manage reviews:

• @coderabbitai resume to resume automatic reviews.
• @coderabbitai review to trigger a single review.

Use the checkboxes below for quick actions:

• ▶️ Resume reviews
• 🔍 Trigger review

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d89b2947-590c-4aab-b099-87dd21088b92

📥 Commits

Reviewing files that changed from the base of the PR and between 1db2819 and 3ad123d.

📒 Files selected for processing (1)

• tests/integration/test_lists/waives.txt

Included review availability: Your plan provides up to 12 included reviews per hour; 9 remain after this review.

---

Walkthrough

Seven visual-generation LPIPS golden artifacts now contain updated SHA-256 values. Obsolete integration test waiver entries were removed, while specified LPIPS skips remain.

Changes

Visual generation validation updates

Layer / File(s)	Summary
Update golden artifact checksums tests/integration/defs/examples/visual_gen/golden/visual_gen_lpips/*.json	Updated SHA-256 values for Flux.2, LTX-2, QwenImage, and Wan 2.2 golden artifacts.
Update integration test waivers tests/integration/test_lists/waives.txt	Removed obsolete Nemotron, visual-generation feature-accuracy, FastWan, and MiniMax multi-node skip entries. Existing visual-generation LPIPS skips remain.

Estimated code review effort: 1 (Trivial) | ~2 minutes

Merge Risk: 🔵 Low · up to 3ad12

This change refreshes visual-generation LPIPS goldens and removes related test waivers. An incorrect waiver removal could alter which CI checks run, so the affected checks should be confirmed before merge; otherwise the change is mergeable with owner awareness.

Possibly related PRs

• NVIDIA/TensorRT-LLM#17476: Updates visual-generation LPIPS golden checksums and related test coverage.

Suggested reviewers: bowenfu

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description contains only the template and does not explain the change, issue, solution, or test coverage.	Add a concise issue and solution summary, list relevant test coverage, and complete the applicable checklist items.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title follows the required format and clearly identifies the LPIPS golden updates for NVBug 6572800.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches 💡 1

🛠️ Fix failing CI checks 💡

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Note

GitHub couldn't provide a complete incremental comparison for this pull request, so CodeRabbit is performing a full review instead. This review may take a little longer.

[karljang]

I pulled both LFS media objects (old 6901191670…, new 1264656ea1…) and diffed the zip
members to see what actually changed behind the pointer.

What checks out: 42 members in both, none added or removed, 34 byte-identical. All seven
updated sha256 values match their new media exactly — I verified each hash against the zip
contents rather than taking the diff at face value.

Four things I'd like resolved before this merges.

1. fastwan_lpips_golden_video.mp4 was replaced, but its JSON was not

Eight media members changed; only seven JSONs were updated. The eighth is fastwan
(2,420,782 → 1,886,253 bytes, −22%), and fastwan_lpips_golden_video.json is untouched.

That golden is not a self-golden:

"source": "FastVideo VideoGenerator (trusted reference)"

So: which stack produced the new fastwan_lpips_golden_video.mp4?

If it came from FastVideo, please say so in the JSON (version/commit) — it's a reference
re-cut and the file should record it. If it came from TRT-LLM, then this test quietly stops
being a cross-stack correctness gate and becomes a self-golden regression gate, while the
file still claims FastVideo produced it. That's a meaningful change in what
test_fastwan_lpips_against_golden proves, and it should be a deliberate, stated decision
rather than a side effect.

Worth noting this file has no sha256 field at all, which is why nothing flagged the swap.
Adding one while it's being re-cut would close that gap.

2. Three tests are unwaived with byte-identical goldens

cosmos3[nvfp4], flux[flux1-nvfp4] and wan[wan21-nvfp4] come off the waiver list, but
their media is unchanged in the new zip (verified by hash). The PR's remedy is "update the
goldens", and for these three no golden was updated.

So what makes them pass now? If they were fixed by something already on main, that's fine —
but it should be named, because the waiver removal is the only thing carrying them.

3. The updated goldens still record the old build

All seven keep:

"torch_version": "2.12.0a0+5aff3928d8.nv26.05",
"tensorrt_llm_version": "1.3.0rc21",
"tensorrt_llm_commit": "b2131b181f5be6717cd302a0b53c22c6a70c65b3",
"container_image": "urm.nvidia.com/...@sha256:475999862b89..."

The media is new, so these fields now attribute freshly-cut media to an old build and an old
container. Those fields exist precisely so a future reviewer can reproduce a golden; as it
stands they point at something that did not produce these bytes. Please refresh them to the
run that actually cut the media.

4. This collides with #17780, which is still open

Both PRs replace visual_gen_lpips_golden_media.zip starting from the same base oid
(6901191670…), and both edit the same block of waives.txt.

A binary LFS pointer does not merge. Whoever lands second has to re-cut their zip on top of
the other's, not just resolve the pointer. If the conflict is settled by picking one oid,
the other PR's re-cut media is silently discarded — and because the diff is a one-line
pointer, nothing in review would show it. #17780 re-cuts seven Cosmos3 members; this one
re-cuts eight others; the two sets are disjoint, so the correct resolution keeps all fifteen.

There's a semantic collision too: this PR unwaives
test_cosmos3_feature_accuracy_against_golden[nvfp4], which #17780 deliberately keeps waived
and deliberately leaves unpinned — its reasoning being that the NVFP4 golden isn't re-cut
there, so it should keep generating under the host default until it is. If this PR unwaives it
while #17780 lands the unpinned path, that test runs unpinned against a golden nobody re-cut.
Worth syncing with @ishovkun on ordering before either merges.

---

Two smaller notes:

• The description is still the unfilled template — no root cause, no test evidence, checklist
  unchecked — for a change that unwaives 11 tests against nvbugs/6572800. A line on what was
  re-run and where would make this reviewable on its own terms.
• The auto-generated summary says 13 waiver removals including a Nemotron NVFP4 test and a
  MiniMax multi-node test. The diff removes exactly 11, all under examples/visual_gen/.
  Please don't let that part drive follow-up work.

[karljang]

@yibinl-nvidia , please check my claude's comment above 😅