docs/USAGE.md publishes an LTX-2.5 resolution envelope that two renders on 2026-08-16/17 have superseded. The page still tells a reader that 448x256 at 25 frames does not complete. It does, and so does a geometry 2.75x larger.
What the record says today
At cf4e672ca:
docs/USAGE.md:776 — | **Measured to complete on one GB10** | **320x192 at 25 frames** |docs/USAGE.md:777 — | Measured NOT to complete | 448x256 at 25 frames — the denoise finishes, then the decode loses about 59 GB in 24 s |docs/USAGE.md:960-961 — "320x192 at 25 frames completes through both distilled phases; 448x256 at 25 frames finishes its denoise and then loses about 59 GB in 24 seconds inside the decode and has to be stopped."docs/USAGE.md:966-968 — "Expect minutes, not seconds: most of a 320x192/25f render is spent in the host VAE decode at 0% GPU".docs/BENCHMARKS.md:487 — "SIZE: 320x192/25f completes on GB10, 448x256 does not".
What was measured
On dgx.casa, against main 0b0b8900f, which carries #1041 (threaded decode), #1032 (T2A) and #1036 (f32 decode accumulators). Container vllmcpp-build:gb10, Release, VLLM_CPP_CUDA=ON, arch 121a, TRITON=ON, CUTLASS absent so FlashAttention-2 was not built — like for like with the prior renders. VLLM_CPP_CPU_THREADS=20. NVFP4 transformer. No --allow-unported.
| Geometry |
Result |
Wall |
| 448x256 / 25 frames |
completed |
3085 s |
| 704x448 / 25 frames |
completed |
4231 s |
| 1024x576 / 25 frames |
not attempted to completion — the box was committed to another session, which is a scheduling fact and not an envelope result |
— |
The ~59 GiB cliff did not recur. A memory guard sampling at 2 s: the 448x256 rung floors MemAvailable at 38.96 GiB over 1289 samples, the 704x448 rung at 38.89 GiB over 1743 samples, and zero samples on either rung fall below 34 GiB. Peak used was 80 of 119 GiB. The box did not reboot.
Verification of the 704x448 artifact, which is the stronger of the two: global mean 90.34, std 60.54, per-frame variance 3630-3706, 0 near-uniform and 0 near-black frames; 25/25 distinct md5s, adjacent-frame mean-abs-diff 4.381 against a uniform-noise reference of 85.3 on the same shape, 0/24 zero-motion pairs; audio 48 kHz stereo, 1.010 s, RMS -37.29 dBFS, 20/20 windows above threshold. The mp4 is on the render host at benchmarks/media/ltx25-704x448-25f-audio.mp4 and is gitignored by .gitignore:35 (*.mp4); it is not committed.
What is not claimed
One run per geometry, on a shared box that was contended, with no oracle on either side. 704x448 is not a ceiling: the next rung was stopped by another session's claim on the box, not by the machine. Two rungs do not establish a scaling law.
The dominant cost moved
docs/USAGE.md:966 attributes most of a render to the host VAE decode. After #1041 that is no longer where the wall is: the same run measures a resolution-independent ~1731 s single-threaded phase (1731 s and 1732 s across two rungs whose voxel counts differ by 2.75x), which is 57-66% of wall. That is #1087, which is open and owns identifying the phase.
The 59 GB observation is kept, not deleted
It is the reason the old row gave, and AGENTS.md forbids deleting evidence to make a record look better. The original observation stands attributed to its own run: rung F1 in .agents/benchmark-record.md (the ladder table and the trace below it) was a prompt-embeds run with no text tower, and it ended in a watchdog kill at avail_kB=13774472 against an armed floor of 18 GiB, not in an engine failure. .agents/specs/ltx25-decode-speed.md §1.2 already recorded that the "inside the decode" half of that sentence is unsupported, and #1014 owns attributing the fall. This issue does not close #1014; it records that the fall did not reproduce on this build under a guard that would have seen it.
What this fixes
The envelope rows in docs/USAGE.md, the prose at :960 and :966, the docs/BENCHMARKS.md LTX cell, and §4 of .agents/specs/ltx25-resolution-envelope.md, which is the spec that published the envelope. Records only; nothing in src/, include/ or tests/ changes, so no build is involved.
FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]
docs/USAGE.mdpublishes an LTX-2.5 resolution envelope that two renders on 2026-08-16/17 have superseded. The page still tells a reader that 448x256 at 25 frames does not complete. It does, and so does a geometry 2.75x larger.What the record says today
At
cf4e672ca:docs/USAGE.md:776—| **Measured to complete on one GB10** | **320x192 at 25 frames** |docs/USAGE.md:777—| Measured NOT to complete | 448x256 at 25 frames — the denoise finishes, then the decode loses about 59 GB in 24 s |docs/USAGE.md:960-961— "320x192 at 25 frames completes through both distilled phases; 448x256 at 25 frames finishes its denoise and then loses about 59 GB in 24 seconds inside the decode and has to be stopped."docs/USAGE.md:966-968— "Expect minutes, not seconds: most of a 320x192/25f render is spent in the host VAE decode at 0% GPU".docs/BENCHMARKS.md:487— "SIZE: 320x192/25f completes on GB10, 448x256 does not".What was measured
On
dgx.casa, againstmain0b0b8900f, which carries #1041 (threaded decode), #1032 (T2A) and #1036 (f32 decode accumulators). Containervllmcpp-build:gb10,Release,VLLM_CPP_CUDA=ON, arch121a,TRITON=ON, CUTLASS absent so FlashAttention-2 was not built — like for like with the prior renders.VLLM_CPP_CPU_THREADS=20. NVFP4 transformer. No--allow-unported.The ~59 GiB cliff did not recur. A memory guard sampling at 2 s: the 448x256 rung floors
MemAvailableat 38.96 GiB over 1289 samples, the 704x448 rung at 38.89 GiB over 1743 samples, and zero samples on either rung fall below 34 GiB. Peak used was 80 of 119 GiB. The box did not reboot.Verification of the 704x448 artifact, which is the stronger of the two: global mean 90.34, std 60.54, per-frame variance 3630-3706, 0 near-uniform and 0 near-black frames; 25/25 distinct md5s, adjacent-frame mean-abs-diff 4.381 against a uniform-noise reference of 85.3 on the same shape, 0/24 zero-motion pairs; audio 48 kHz stereo, 1.010 s, RMS -37.29 dBFS, 20/20 windows above threshold. The mp4 is on the render host at
benchmarks/media/ltx25-704x448-25f-audio.mp4and is gitignored by.gitignore:35(*.mp4); it is not committed.What is not claimed
One run per geometry, on a shared box that was contended, with no oracle on either side. 704x448 is not a ceiling: the next rung was stopped by another session's claim on the box, not by the machine. Two rungs do not establish a scaling law.
The dominant cost moved
docs/USAGE.md:966attributes most of a render to the host VAE decode. After #1041 that is no longer where the wall is: the same run measures a resolution-independent ~1731 s single-threaded phase (1731 s and 1732 s across two rungs whose voxel counts differ by 2.75x), which is 57-66% of wall. That is #1087, which is open and owns identifying the phase.The 59 GB observation is kept, not deleted
It is the reason the old row gave, and AGENTS.md forbids deleting evidence to make a record look better. The original observation stands attributed to its own run: rung F1 in
.agents/benchmark-record.md(the ladder table and the trace below it) was a prompt-embeds run with no text tower, and it ended in a watchdog kill atavail_kB=13774472against an armed floor of 18 GiB, not in an engine failure..agents/specs/ltx25-decode-speed.md§1.2 already recorded that the "inside the decode" half of that sentence is unsupported, and #1014 owns attributing the fall. This issue does not close #1014; it records that the fall did not reproduce on this build under a guard that would have seen it.What this fixes
The envelope rows in
docs/USAGE.md, the prose at:960and:966, thedocs/BENCHMARKS.mdLTX cell, and §4 of.agents/specs/ltx25-resolution-envelope.md, which is the spec that published the envelope. Records only; nothing insrc/,include/ortests/changes, so no build is involved.FOLLOWING_AGENTS_PROTOCOL
Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:claude-opus-5 [claude-code]