What this is
Row LTX25-PROMPT-ADALN / LTX25-ADALN-CLAIMS (#644) rests its headline
conclusion on the shipped LTX-2.5 checkpoint. None of that evidence runs in
CI. Filing it so it is visible debt rather than something the next person
discovers mid-campaign. This issue is a record, not a request to wire
checkpoints into CI in that row.
Measured
LTX2_CHECKPOINT_ROOT appears in no workflow. grep -rn CHECKPOINT_ROOT .github/ exits 1 with zero hits, while the same pattern matches in tests/
and .agents/ (positive control run in the same command, so this is not an
assertion from a failed grep).- Consequently
test_ltx2_video's TEST_CASE("ltx2 video: the SHIPPED Lightricks checkpoints parse and load") skips on every CI run. Measured on
the same binary: 30 cases / 502 assertions with the variable unset,
30 / 8734 with it set. CI therefore executes 502 of 8734 = 5.7% of that
suite's assertions, at an unchanged case count of 30 — so a green CI row and a
green local row are indistinguishable from each other and from a run that never
opened a checkpoint.
scripts/measure-ltx2-prompt-adaln.py, which produces the row's
shipped-weights numbers (rms|term|/rms|table| 1347% video / 1583% audio,
max|term|/max|table| 7119% / 2817%), is a manual tool no gate invokes:
grep -rn measure-ltx2-prompt-adaln hits only the script's own usage string
and the spec.
Why it is recorded rather than fixed here
The skip is already announced in four places (the test's own MESSAGE, the
comment block above the case, .agents/specs/ltx25-prompt-adaln.md, and
.agents/porting-inventory.md), which is the right disposition: a gate that
quietly does nothing when its input is absent is how "we tested the real
checkpoint" becomes untrue. What was missing is a tracked statement that the
row's checkpoint-derived evidence is manual and host-local — reproducible
only on a box with the 18.72 GB NVFP4 DiT and the 23 GB FP8 DiT under
$CHECKPOINT_ROOT (here: the NAS at /mnt/nas_share/checkpoints/ltx-2.5).
What would close it
A decision, not necessarily a runner: either a self-hosted lane that mounts the
checkpoints and sets LTX2_CHECKPOINT_ROOT, or an explicit recorded acceptance
that this evidence stays host-local and is re-run by hand at named points. Either
way the assertion count must be quoted with its configuration, never bare.
Refs #644.
What this is
Row
LTX25-PROMPT-ADALN/LTX25-ADALN-CLAIMS(#644) rests its headlineconclusion on the shipped LTX-2.5 checkpoint. None of that evidence runs in
CI. Filing it so it is visible debt rather than something the next person
discovers mid-campaign. This issue is a record, not a request to wire
checkpoints into CI in that row.
Measured
LTX2_CHECKPOINT_ROOTappears in no workflow.grep -rn CHECKPOINT_ROOT .github/exits 1 with zero hits, while the same pattern matches intests/and
.agents/(positive control run in the same command, so this is not anassertion from a failed grep).
test_ltx2_video'sTEST_CASE("ltx2 video: the SHIPPED Lightricks checkpoints parse and load")skips on every CI run. Measured onthe same binary: 30 cases / 502 assertions with the variable unset,
30 / 8734 with it set. CI therefore executes 502 of 8734 = 5.7% of that
suite's assertions, at an unchanged case count of 30 — so a green CI row and a
green local row are indistinguishable from each other and from a run that never
opened a checkpoint.
scripts/measure-ltx2-prompt-adaln.py, which produces the row'sshipped-weights numbers (
rms|term|/rms|table|1347% video / 1583% audio,max|term|/max|table|7119% / 2817%), is a manual tool no gate invokes:grep -rn measure-ltx2-prompt-adalnhits only the script's own usage stringand the spec.
Why it is recorded rather than fixed here
The skip is already announced in four places (the test's own
MESSAGE, thecomment block above the case,
.agents/specs/ltx25-prompt-adaln.md, and.agents/porting-inventory.md), which is the right disposition: a gate thatquietly does nothing when its input is absent is how "we tested the real
checkpoint" becomes untrue. What was missing is a tracked statement that the
row's checkpoint-derived evidence is manual and host-local — reproducible
only on a box with the 18.72 GB NVFP4 DiT and the 23 GB FP8 DiT under
$CHECKPOINT_ROOT(here: the NAS at/mnt/nas_share/checkpoints/ltx-2.5).What would close it
A decision, not necessarily a runner: either a self-hosted lane that mounts the
checkpoints and sets
LTX2_CHECKPOINT_ROOT, or an explicit recorded acceptancethat this evidence stays host-local and is re-run by hand at named points. Either
way the assertion count must be quoted with its configuration, never bare.
Refs #644.