LTX-2.5 sigma schedule: Ltx2SigmaSchedule(1, ...) divides 0/0 and returns {-nan, 0}

[localai-bot]

Found while repairing LTX-2.5 phase L5 (#435, PR #437) after a sweep over step counts was added. Reported rather than repaired, deliberately: the fix moves every sigma on every arm, so it needs its own spec and fresh review.

The defect

Ltx2SigmaSchedule(1, ...) returns {-nan, 0} where upstream returns {0.10000002, 0}.

Root cause, measured

Upstream's shift is a Python scalar divided by a tensor, and torch evaluates that as scalar * reciprocal(tensor). At sigma == 1 that yields 0.99999994, not 1.0.

At steps == 1 that one-ulp residue is the entire stretch anchor. Our port computes the same expression in f32 and gets exactly 1.0, so the stretch divides 0/0.

Why it was invisible

Two layers:

1. The fixture exercised steps in {8,6,5,1,4,7} with no sweep, and
2. MaxAbsDiff dropped NaN (MaxAbsDiff helpers are NaN-blind: a golden comparing all-NaN against correct values reports max|diff| = 0 and PASSES #449), so the OneStep arm compared NaN against a correct golden and reported worst = 0.0.

Both are now closed: L5's repair added a sweep asserting exact-0 termination over steps 1..200 (which carves out steps == 1 explicitly, so the defect is recorded in the tree rather than only in a report), and #449 hardened the helper.

Wants

Mirror upstream's scalar * reciprocal(tensor) evaluation so the anchor carries the same one-ulp residue, then regenerate the affected goldens. Note the ordering trap: the goldens must be regenerated from upstream, never adjusted to match the port. Expect every sigma on every arm to move by ~1 ulp, so this is a whole-phase re-gate, not a local edit.

Related: #449 (the NaN-blind helper that hid it), .agents/specs/ltx-2-5.md §7.0(c) (the fixture that could not discriminate).

🤖 Generated with Claude Code

[localai-bot]

Ordering constraint with #449, recorded so it is not discovered the hard way

tests/vllm/models/test_ltx2_pipeline.cpp:157 still carries the NaN-blind if (d > worst) worst = d; form. It was NOT converted by #449's hardening, because that work branched before phase L5 landed on row/MODEL-DIFFUSION-LTX25. Both are now merged (3fa3976a), so the blind helper and the hardened tests/support/max_abs_diff.h coexist in the tree.

This defect is currently invisible only because that helper is blind. The L5 re-review demonstrated it directly: making MaxAbsDiff NaN-aware, changing nothing else, turns the suite red on the OneStep arm:

CHECK( worst <= kRoundOff ) is NOT correct!
  values: CHECK( inf <= 5e-06 )
  logged: LTX2Scheduler arm = OneStep max|diff| = inf

So suite greenness at steps == 1 rests on two things: the blind helper, and L5's explicit steps >= 2 carve-out on the sweep's finiteness check.

What that means for sequencing

Fix this issue before routing test_ltx2_pipeline.cpp through the hardened helper. Doing it the other way round makes the gate red on a known, already-tracked defect, which invites someone to widen the carve-out to get green again, and that is precisely the move AGENTS.md forbids.

Concretely:

1. Fix this (scalar * reciprocal(tensor) mirroring, regenerate goldens from upstream, never adjusted to match the port; expect every sigma on every arm to move ~1 ulp, so it is a whole-phase re-gate).
2. Then convert test_ltx2_pipeline.cpp to tests/support/max_abs_diff.h.
3. Then remove the steps >= 2 carve-out and confirm the sweep is green at steps == 1 for the right reason.

A related suggestion from the re-review

The carve-out has no assertion pinning the defect it excludes, so once this lands nothing forces its removal and it would quietly persist as dead permissiveness. A CHECK that steps == 1 is currently non-finite would make the carve-out self-retiring. Worth folding into step 1.

Also worth carrying (#451 territory)

The re-review found the sweep's upper bound of 200 is arbitrary where the property is not: a LinspaceF32 correct through steps = 200 and naive above passes the whole suite green, while --steps 209 terminates at 0.1 instead of 0. Low exploitability, since 23 counts inside the range fire, but the bound deserves a stated reason or an extension.

🤖 Generated with Claude Code