LTX-2.5 guided denoiser: two comments name `RetakePipeline` as the one upstream caller that sets `force_uncond_pass`, and no upstream caller sets it at `fd4ded7f`

[localai-bot]

Two comments in the LTX-2.5 guided denoiser name RetakePipeline as an upstream caller that sets force_uncond_pass. No upstream caller sets it, at the pin or anywhere in the tree.

The two claims:

• include/vllm/model_executor/models/ltx2_denoisers.h:144-146 — "force_uncond_pass (:76, read at :102-103). RetakePipeline is the one upstream caller that sets it (retake.py:305-311)."
• src/vllm/model_executor/models/ltx2_denoisers.cpp:111-113 — ":102-109. force_uncond_pass adds the pass for a modality that is PRESENT even when its own guider does not ask (retake.py:305-311 is the one upstream caller that sets it)."

Both landed in daeff67f2 (row LTX25-GUIDED-VIDEO, #1092, PR #1102).

Measured at the pinned revision. Lightricks/LTX-2 at fd4ded7f2d88d3da713abcdd4ad41ecc4a9314ca, the revision .agents/specs/ltx25-retake.md:12 and .agents/specs/ltx25-guided-video.md:6 both pin, read at the local checkout /home/mudler/_git/LTX-2 whose git rev-parse HEAD is that SHA:

1. packages/ltx-pipelines/src/ltx_pipelines/retake.py:305-311 constructs GuidedDenoiser(v_context=..., a_context=..., video_guider=..., audio_guider=...) and passes no force_uncond_pass argument. GuidedDenoiser.__init__ declares it force_uncond_pass: bool = False (utils/denoisers.py:267), so retake takes the False default. grep -n GuidedDenoiser retake.py returns the import at :33 and that one construction at :305 — retake has no second site.
2. grep -rn 'force_uncond_pass' --include='*.py' over the whole upstream tree returns 10 hits, all inside utils/denoisers.py plus one in utils/samplers.py. Every denoisers.py hit is a declaration or a pass-through: :74 (the _guided_denoise keyword parameter), :102-103 (the two reads), :267,273,297 (GuidedDenoiser stores it and forwards it), :313,319,357 (FactoryGuidedDenoiser, the same). The samplers.py:692 hit is the substring "or force_uncond_pass=True and a negative_context." inside a raised error message, not an assignment. No upstream caller enables it.
3. Its actual purpose, per packages/ltx-pipelines/CLAUDE.md:76: GuidedDenoiser and FactoryGuidedDenoiser accept force_uncond_pass=True to run the uncond pass even when cfg_scale=1.0, "required by CFG++ when the guidance scale is 1 but the uncond prediction is still needed for the ODE derivative". It is a CFG++ affordance, and upstream ships the plumbing without turning it on.

A second, smaller error rides along: the header cites the declaration as :76, and :76 is the docstring. The declaration is denoisers.py:74.

Behavior is correct and no code change is owed. Ltx2GuidedDenoiseInputs::force_uncond_pass defaults to false (ltx2_denoisers.h:146), and grep -rn 'force_uncond_pass' src/ include/ tests/ finds exactly two reads (ltx2_denoisers.cpp:115,117) and no assignment anywhere. Our port therefore matches upstream, where nothing enables it either. This is a documentation defect only.

Why it matters. The false anchor is load-bearing rather than cosmetic. A reader who trusts it concludes that our retake path is missing a flag upstream sets, and "repairs" retake by setting it. That would create a real divergence: force_uncond_pass adds an uncond pass for every present modality, which doubles retake's DiT forward count from 8 to 16 and changes its output. The comment invites the exact defect it appears to describe.

Fix, in the same flow as this filing: both comments state what holds at fd4ded7f — the field mirrors upstream's declared-and-threaded-but-unused CFG++ plumbing, cited at denoisers.py:74, :102-103 and packages/ltx-pipelines/CLAUDE.md:76, with retake.py:305-311 named as a site that takes the False default rather than one that sets the flag. No behavior changes.

Specs checked and left alone: .agents/specs/ltx25-retake.md never mentions the field, and .agents/specs/ltx25-guided-video.md:198 states the condition (uncond when either guider asks or force_uncond_pass) without claiming a caller, which is true at the pin.

[localai-bot]

Fix written and pushed, and the merge is blocked by a gate this fix cannot honestly satisfy.

Branch row/LTX25-GUIDED-VIDEO-1528 at dafde0dc4, based on origin/main 2341bf31d. Comment text only, +15/-4 across three files: the two comments and the .agents/issue-index.md append.

scripts/agent-preflight.sh reports 90 gates ok and one FAIL:

FAIL doc-checkpoint range
     ERROR: commit dafde0dc4: changed user_usage but did not update docs/USAGE.md.
     This change alters what the project supports or how it is used, which its
     purpose-specific document has to reflect.

The trigger is a pure path match. scripts/check-doc-checkpoint.py:99 puts the whole include/vllm/ prefix in USER_USAGE_PREFIXES, and classify() at :345 tests path.startswith(...) with no content analysis, so any byte changed in a header under that tree owes a docs/USAGE.md edit. errors_for() at :384 only asks whether docs/USAGE.md is in the changed path set, so the demand is satisfied by touching the file at all.

This exact obstacle is already recorded. #1230 names it: "one of them lived in include/vllm/, a USER_USAGE_PREFIXES path in scripts/check-doc-checkpoint.py:99, which is a pure path match with no content analysis, so a comment-only anchor edit in a public header demanded a docs/USAGE.md edit the change did not owe." It is the same class as #595 (src/vllm/model_executor/models/ -> docs/FEATURES.md) and #515 (CMakeLists.txt -> docs/USAGE.md), and check-doc-checkpoint.py's own header calls classify-by-directory the defect its rewrite removed for src/, include/ and tests/.

There is nothing true to write in docs/USAGE.md. The page carries no LTX-2.5 guidance section: grep -n 'cfg-scale|cfg_scale' docs/USAGE.md returns nothing, and grep -n 'retake' docs/USAGE.md returns nothing. Its only LTX-2.5 content is four checkpoint-registry rows, inside <!-- checkpoint-registry:begin --> markers, none of which this change touches. docs/USAGE.md:165 routes the reader to models/README.md, and the guidance controls are documented at docs/models/ltx-2-5.md:188-194, which already states the one user-facing fact this audit bears on ("A CFG scale other than 1.0 requires negative conditioning"). Repeating it in docs/USAGE.md would break AGENTS.md's "each fact lives in one document", and answering the gate with prose is the failure check-doc-checkpoint.py's own docstring records for #1054/#1055.

Narrowing the prefix is not available in flow. It is a checker-semantics change, which AGENTS.md routes to its own row, spec, red-first test and fresh review, and explicitly excludes from the in-flow rule.

So the choice is between landing a docs/USAGE.md paragraph with nothing true to say, correcting only ltx2_denoisers.cpp and leaving the header asserting the opposite (the disagreement-inside-one-change harm #1230 exists to avoid), and merging with documentation-checkpoint red. Escalated rather than decided.

[localai-bot]

Blocked, not abandoned: the fix is written and pushed, and it is held by a gate it cannot honestly satisfy

The finding is confirmed and the correction is committed on
row/LTX25-GUIDED-VIDEO-1528 at dafde0dc4, based on origin/main
2341bf31d. No pull request is open, deliberately.

scripts/agent-preflight.sh on that commit reports 90 gates ok, 1 FAIL:

FAIL doc-checkpoint range
     ERROR: commit dafde0dc4: changed user_usage but did not update docs/USAGE.md.

scripts/check-doc-checkpoint.py:99 places the whole include/vllm/ prefix in
USER_USAGE_PREFIXES, and classify() at :345 is a pure startswith with no
content analysis. A comment-only edit in any public header therefore owes a
docs/USAGE.md edit it does not actually owe. The documentation-checkpoint
CI job (ci.yml:540) runs the same check, so the pull request would open red.

Why none of the three exits was taken

1. Write a docs/USAGE.md paragraph. There is nothing true to write.
   docs/USAGE.md mentions neither cfg_scale nor retake; its only LTX-2.5
   content is four generated checkpoint-registry rows, and :165 routes the
   reader to models/README.md, where docs/models/ltx-2-5.md:188-194 already
   states the one user-facing fact this audit touches. Repeating it there breaks
   each fact lives in one document, and answering a gate with prose written to
   satisfy it is the fix(qwen3.5): drop redundant AppleClang capture #1054/main is RED on check-public-doc-tables: #1054 pushed BENCHMARKS and FEATURES over their prose budgets, and the pre-push hook now blocks every branch #1055 failure the checker's own docstring records.
2. Land the .cpp alone. That leaves the header asserting the opposite of
   its own source file - precisely the disagreement-inside-one-change harm that
   #1230 argues makes a partial
   anchor correction strictly worse than none.
3. Narrow USER_USAGE_PREFIXES. This is a semantic checker change, which
   AGENTS.md excludes from the in-flow rule and which needs a spec, a
   red-before test and green-after evidence. It also makes a red gate green by
   widening what the gate accepts, which AGENTS.md prohibits. Doing that
   from inside an unrelated comment fix would loosen a gate for every concurrent
   branch without a spec arguing for it.

What this issue is waiting on

A row that owns the checker's path-match defect. #1230's index row already names
it verbatim as a consequence of that gate, and #1230's own 22-site sweep edits
include/vllm/ headers with comment-only changes, so that sweep must solve
this same gate anyway. The cheapest correct sequencing is for the two to share
one solution rather than for this issue to invent a second.

The branch is preserved and rebases cleanly; nothing here needs re-deriving.

Not a behaviour change, in either direction

grep -rn "force_uncond_pass" src/ include/ tests/ returns five lines: two
comment lines, the = false default initializer, and the two reads at
ltx2_denoisers.cpp:118,120. No assignment anywhere, which is exactly what
upstream does at fd4ded7f. The port is faithful; only its comment was wrong.