Spaces:
Running on Zero
Running on Zero
Commit History
3 days ago
-
docs: every sentence stays true after the deploy: README and the release notes say this is deploy 2's code and point to issue #32 for what the Space shows, and no doc says the Space still runs deploy 1, is not deployed, or has not been redeployed since August (#32)
Unverified signatureThe commit we deploy is also what origin/main is fast-forwarded to, and nothing can be pushed after the deploy without restarting the public Space, so a sentence that a deploy makes false would stay false. README: "Where things stand, 2026-11-17" says the README describes deploy 2's code, which replaces deploy 1 on the Space, points to SPACE_MODE_MATRIX 10.1-10.3 and the release notes, and to issue #32 for deploy 2's own check on the Space. The cold-start time points to #32 instead of "has not been measured"; the longest accepted reservation is "not measured by these runs"; the browser runs of 2026-11-16 "did not finish in that session" instead of "have not been run ... yet". docs/RELEASE_NOTES.md: "Deploy 2 - 2026-11-17", the code of this commit's tree, the Space result on #32 (also under "On the Space"); deploy 1 "ran" as 0cd54f2 until deploy 2, and main "was fast-forwarded" to 1e9fd20 that day. The other docs, the same way: SPACE_MODE_MATRIX's header and 10 intro, 10.2 and 10.3; …abc9150 -
docs: the handoff's and ZEROGPU_DISK's status boxes no longer say the Space has run nothing since August: it runs deploy 1, whose Space record is SPACE_MODE_MATRIX §10.1, and the handoff points at the release notes (#29)
Unverified signature70ddbf2 -
docs: the weights come from the image: the README, PLAN and the handoff say preload_from_hub fetches the 35 model files (71.92 GB, no temporal upsampler) and the prompt enhancer's 8 (2.49 GB) before the app starts, and start-up downloads only what is missing, if the disk can hold it (#23)
Unverified signature41dcf05, ff62c55 and 72b989d made the preload live: the README's front matter (B3's) fetches the 35 files the int8 Space loads at their pinned commits and the enhancer's 8 at its own, and preflight.ensure_weights checks the cache at start-up, downloads only the missing files after one warning, and refuses a download the disk cannot hold (missing bytes + ZeroGPU's packed copy + a 10 GB margin, under the 150 GB cap) before writing a byte. The docs still said the Space downloads 72 GB at every start because the preload 401s on a private Space. README: the enhancer paragraph says the image brings its weights on the Space and the start-up prefetch (CUDA only, app.py) is the fallback; the ZeroGPU table gets a Weights row. No cold-start time is claimed: none is measured with the preload. docs/PLAN.md: 2.3 notes the 35-file set, the deploy-shape table's Space and Weights rows are corrected in place (public; from the image), 6.2 and 6.2.1 open with a 2026-11-17 correction, and the enhancer-pre…38e36e9
4 days ago
-
docs: the packed size is quoted as the platform's progress bar printed it, 40.6G
Unverified signatureThe status boxes added in 8e7494f wrote "40.6 GB". The ZeroGPU packing bar on the 2026-11-13 boot printed "40.6G" and does not say whether that is GB or GiB, so the boxes now quote it as printed rather than assert a unit.e6561d9 -
docs: each August record opens with what has been superseded since, and where the current truth is
Unverified signatureThe seven historical documents keep their bodies. Each gets a short "Status on 2026-11-16" box at the top naming the passages that later measurements overtook, with the corrected figure and its source: - the stage-2 "ceiling between 25 and 49 frames" was the 47.375 GiB `large` partition; 704x480x49 passes on `xlarge` (MIG_ALLOCATOR_SIGHTINGS.md §2); - "≤ ~675.5 GPU-s" for UTC 2026-08-25: the persistent ledger reads ≤ 1528.672 over 20 runner calls, ~1622 with the hand-bounded manual calls; - "82.2 GB of 150 [OK]" is a reading taken before ZeroGPU packs; the platform packed 40.6 GB on the 2026-11-13 boot and the disk peak was never measured; - TEST_PLAN.md and ZEROGPU_DISK.md never got the 2026-08-25 retraction of "all eleven modes generate, content-verified"; - deploying is a merge commit on the Space head pushed as a fast-forward, not a push of the work commit; the download is at startup, 13 min 59 s cold; - hinode is retired and its scripts are gone; `auto` is no longer the…8e7494f
-
docs: the app ran its whole life on half a GPU — `auto` cannot promote, and four documents said otherwise
Unverified signature`MIG_ALLOCATOR_SIGHTINGS.md` was a register of three NVML allocator asserts with no explanation and a ruling that nobody chase them until somebody read the headroom. `0a4ddbb` read it. The number was **47.375 GiB** — ZeroGPU's `large` — where every document in this repo assumed `xlarge`'s 94.971 GiB. ⛔ THE ROOT CAUSE, IN FOUR HOPS, ALL FROM THE PLATFORM'S OWN PACKAGE: * `spaces/zero/__init__.py:29-35` — `startup()` packs, clears the 42.00 GiB `xlarge_threshold`, and sets the module global `_default_gpu_size='xlarge'`. The promotion is real and it does fire. * `spaces/zero/client.py:109` — `get_duration_seconds()` is the **only** consumer of `default_gpu_size()` anywhere in the package, and it rebinds a **local**. * `spaces/zero/client.py:111-112` — that local feeds one branch: `if gpu_size != 'xlarge': duration_seconds *= duration_factor`. The promotion's entire observable effect is skipping a 1.5× multiplier on the *requested duration*. * `client.py:144` → `api.py:112-…caf1083
-
docs: every Space generation result was garbage — the invalidation, stated at the top
Unverified signatureThe deployed Space produced colour static for the entire project. Blanket int8 quantised the LTX-2.5 conditioning path (the eight LTX2AdaLayerNormSingle modulation heads plus the in/out projections), the denoiser stopped being steered, and the output became a correctly-shaped sample from the wrong distribution. Fixed at eb1aa9f, deployed e3b5333, verified by opening a frame. SPACE_MODE_MATRIX and SPACE_SESSION_2026-08-25 get a dated invalidation block at the very top, above existing content, per the operator's ruling. House rule honoured: nothing is rewritten, every measurement stands as history, the corrections are dated on top. Each block says the same six things: - Every generation result in the document was produced by a broken transformer. Space shas up to and including ff142d9 ran blanket int8. - MECHANICAL CLAIMS SURVIVE: all eleven modes execute; both fork-boundary directions carry their cargo; determinism and byte-reproducibility hold; tiled decode enables 1216x704x12…898dad2 -
docs: the 2026-08-25 session — byte identity sha256-verified, 1216x704 proven, the stage-2 ceiling found
Unverified signatureNew: docs/SPACE_SESSION_2026-08-25.md, the record of the day's campaign — 14 metered attempts plus two zero-GPU validates across two deploys (8e9f7a1, then ff142d9 carrying cdb3211/0be0014/b9ee967/2eaeb2c and the UI redesign), total <= ~675.5 GPU-s of 2400, the 1920 ceiling never approached. Every artifact is banked under a directory named by its sha256, and every hash in the doc was re-computed at write-up time. The headline closes the correction this morning's pass made: byte-level reproducibility, previously UNVERIFIED (no day-1 hash), is now VERIFIED in the happy direction — a fresh seed-1234 run hashed equal to the surviving 2026-08-22 artifact (b78803cc…) on a third container and a different code sha, then equal AGAIN through the redesigned UI on the tiled-decode build on a warm worker that had just carried an adapter. The condition pipeline reproduced its banked artifact byte-for-byte (ef8571e6…) — giving the conditioning discriminator a zero noise floor — and both LoRA routes …76c408d -
docs: the audit's stale claims annotated — measurements stand, corrections dated on top
Unverified signatureA four-agent audit swept every doc for claims events had falsified (~30, with file:line). Fixed per the house rule: history is annotated, never rewritten — dated blockquote corrections point at current reality with evidence. Six follow-up-table rows get in-cell appends with status flips (the file's own precedent for rows); no finding is deleted, no captured record reworded. The delicate one first. SPACE_MODE_MATRIX §5's 'byte-identical to the result banked the previous day' was never verified: no day-1 hash was ever recorded, the match was file size (39,146 B) plus every decoded statistic to three decimals, and the day-1 file is gone with its host — so byte identity is now unverifiable against it. The §5 caveat ('Do NOT upgrade this') was correct and reports upgraded past it; the error ran in that direction, not the doc's. Corrected claim: identical in size and in all decoded statistics, byte identity UNVERIFIED; a same-seed re-run sha256-compared against the surviving 2026-08-22 arti…fd85c87
-
docs: the fork-boundary fix is a pair, and half of it is a regression
Unverified signatureThree corrections to SPACE_MODE_MATRIX.md, plus four stale claims elsewhere. The fix for the fork boundary was recorded as one method: a `__reduce__` on LTXStudioError. Reading `spaces` rather than trying it shows that is wrong in both directions, and neither half works alone. `__reduce__` alone does nothing. ExceptionResult (types.py:28-31) carries only `traceback`, `error_cls` and `gradio_error`, and `exception_result` populates the third only when `isinstance(exc, gr.Error)`. Our exception object is therefore never transmitted at all, nothing is ever pickled, and `__reduce__` is never consulted. That also settles the open question of whether restoring picklability would restore the message: it would not, and no live run is needed to know it. A `gr.Error` subclass alone is worse than the status quo. It makes the object eligible for `gradio_error`, which must then cross a multiprocessing.Queue that our classes cannot pickle onto. `spaces` anticipates this at wrappers.py:272-275 and…f927c4e
-
docs: the size kwarg was borrowing the fork boundary's evidence
Unverified signatureOne bullet claimed both, under one tick: ✅ The `size` kwarg and the fork boundary -- reached; a worker forked, ran, and released cleanly (fail=false) The fork boundary earned that. The size kwarg did not, and bundling them let the weaker claim draw on the stronger one's evidence. Same shape as 60.0 GB carrying a steady-state label while being a peak, and as a warning marker sitting on a section that had been rewritten to say falsified. That pattern has been corrected three times today; it should not survive into the last commit. Split three ways, each on its own evidence: ✅ fork boundary -- crossed, worker forked, ran, released fail=false 200 ✅ size kwarg PLUMBING -- name survived gpu_decorator_kwargs' signature filter, no import-time TypeError, which is the risk PLAN 5.8 named ⏸️ an explicit size VALUE -- never passed app.py:107 sends size=None whenever LTX25_ZEROGPU_SIZE is auto, which is DEFAULT_ZEROGPU_SIZE and what the Space runs. So ZeroGPU chose, not us. I…eb760e3 -
docs: the handoff said the Space cannot run; it runs and generates
Unverified signatureFour stale claims, all false since the deploy, in the document a newcomer reads first. A wrong first impression propagates into every decision made afterwards, which is what made these worse than the rest of the doc debt. §1 line 22 "It does not run there today" (LINE-WRAPPED) §3 line 233 "ZEROGPU-PENDING" §3 line 235 "has evicted twice and has never run" §5 line 529 "Not currently possible" §7 line 751 "has evicted twice and has never run" Corrected without overcorrecting. The Space runs and generates, and that is ONE mode ONCE -- text_to_video at 256x256x9. Ten of eleven have never executed there, and no audio path has. Section 2's eleven-of-eleven is a hinode result and is left saying so; section 2 is untouched. Section 3 is retitled SPACE-PROVEN vs HOST-PROVEN rather than deleted. Its caution was right and only its framing was wrong -- the two evidence bases have diverged rather than merged, so the section is more useful now, not less. Its list of unanswerable …26e9259 -
docs: audit for claims that outran their evidence -- four found
Unverified signatureRe-read the handoff for the failure I had already made three times. Four more, all mine, all the same shape. 1. Section 2.3 carried the PRE-fix figures for H14, H15 and H16 as current while 2.2 carried post-fix and 2.5 listed those same numbers as the "before" column. I fixed the H14 prose paragraph for exactly this and did not look at the table above it. 2. Section 2.5's caveat said seeding was "verified byte-identical" without the scope I had just added to gap 8. The video path is unverified. 3. The lessons section carried the same unscoped sentence a THIRD time -- in the paragraph about refusing to write unmeasured claims. Now recorded there as its own lesson: a claim narrowed in one place does not narrow itself in the others, and a document is exactly where that goes wrong. 4. Gap 7 said "all twelve modes changed when it was corrected -- nothing silently survived it", and offered the magnitudes ordering by conditioning strength as corroboration. Both halves are wrong. The runs …9008276 -
docs: the byte-identical result is one mode, not the whole build
Unverified signature6fd3648 wrote "seed determines the output byte-for-byte, on GPU" as a general property. It is verified for audio_to_audio and nothing else. That is an overclaim by scope, written within minutes of recording a lesson about refusing to write claims I cannot support -- the sentence was true of the measurement I had and false of the sentence I wrote. host-runner caught the scope while reframing the text_to_video pair: it is no longer a residual hunt, because there is no residual, but the VIDEO path is still not shown to be bitwise reproducible and the entire twelve-mode pixel-diff gallery depends on that. If a video mode carries its own unseeded draw, those per-pair numbers stay unattributable however good the audio seeding is. Same run, better question -- and a better one than mine, which was looking for a defect that did not exist. Also recorded: git log -S 'generator' 4804da1..e4e8df1 returns 4ebb8b9 alone, so the seeding fix was complete when it landed rather than repaired afterward…eb7d791
-
docs: seeding is resolved and verified byte-identical
Unverified signaturehost-runner withdrew the 8.7% reproducibility residual: their control ran against a box at 4804da1, which predates the seeding fix in 4ebb8b9. They measured the pre-seeding tree. At the tip two runs at identical seed and config are byte-for-byte identical, md5 2988f011 both. So gap 8 moves from open to resolved, and the claim is STRONGER than either sentence I was reluctant to write: seed determines the output byte-for-byte, on GPU, with no deterministic-algorithm flags set. "Attached to nothing" and "seed-consistent but not bit-reproducible" were both drafted, both weaker, and both wrong. The remaining consequence is narrow rather than absent: 2.5's comparison table was measured before seeding worked, so its pixel diffs cannot attribute the change to the recipe fix. The seeded gallery supersedes it for attribution; the figures stay as the record of what the defect cost. Recorded as a lesson because it is the only time today source-reading and measurement CONFIRMED each other rather…6fd3648 -
docs: H14's proof is geometric, so drop the mixed-vintage figures
Unverified signatureThe H14 paragraph compared its post-fix size and time against H1's PRE-fix 886,265 B and 25.7 s -- the exact mixing of vintages the table beside it warns against, introduced by me when the corrected figures landed above it. The argument never needed them. 1408x960 against 704x480 is 2x in both dimensions and only the spatial upsampler produces that, which is the one claim in the document that neither vintage nor seeding can affect, since geometry is identical across every before/after pair. Removing the numbers makes the sentence both correct and stronger.835ad74 -
docs: eleven of eleven, and the evidence that still needs re-running
Unverified signatureSection 2 rebuilt on the corrected real-path figures, with H10 and H11 filled in -- audio_to_audio at -10.4 dBFS, audio_to_video at 193,825 B. Eleven of eleven modes produce artifacts with the expected layout on the real path, no bypass and no harness accommodation. The before/after pairs get their own section rather than mixed-vintage rows, because the pre-fix numbers ARE the evidence for what the defect cost -- 3x the compute, clipping on two modes, a 50 dB hole in a third. Deleting them would turn a measurement back into an assertion. Three notes a reader needs: H7 is the only artifact that grew and that is flagged rather than interpreted; H8 and H9 are byte-identical because uncompressed PCM at one duration always is, so their change is entirely in level and this table alone would call them unchanged; and everything is ~3x faster because unguided means one forward pass per step. But those pairs are NOT yet controlled -- the seed reached only the sidecar until 4ebb8b9, so both ha…e4e8df1 -
docs: re-attribute H9, record the intervention, and rule 9
Unverified signatureThe handoff said the video_to_audio silence was caused by the strength=0.8 default. It was not, and leaving that in would have been the single most misleading sentence in the document -- it names a lever that is now back at its default and a cause that was withdrawn. The cause was the distilled recipe reaching only the stage-2 path. Corrected, H9 runs at -17.8 dBFS against H8's -15.3, at the unchanged strength of 0.8: one mode 50 dB down and another driven into the rail were the same misconfiguration seen from opposite ends. The wrong diagnosis is kept rather than deleted, because how it went wrong is the useful part: a clean single-lever sweep was believed over a source citation that was correct throughout. Gap 7 records the intervention as a result. All twelve modes moved, with magnitudes ordering by conditioning strength, and the defect cost 3x the compute as well as the quality -- unguided means one forward pass per step instead of two or three, so the flagship 1216x704 with ste…5d73ea4 -
docs: we can prove a function ran, not that it ran with the right arguments
Unverified signatureThe gap this build ends on, and the operator asked for it in the PR body. DISTILLED_SIGMA_VALUES was defined, documented as mandatory, quoted in its module docstring and REFERENCED -- so every check we own passed. It was referenced from a function applied only when stage 2 was enabled. A reference check passes. A module-level reachability check passes. Even the function-level sweep passes unless someone names that particular function. Nothing checked the arguments the pipeline was actually called with. The pattern that works, used three times now, is to record the vendor's specification INDEPENDENTLY and assert against it -- MODEL_CARD_DISTILLED, REAL_AUDIO_VAE_CONFIG_KEYS, REAL_VOCODER_CONFIG_KEYS. Without that independence the test agrees with the code and with nothing else. Stated as a ladder: imported -> reached by the path that does the work -> this function ran -> the arguments were the vendor's. Only the last rung is still hand-built, one call site at a time. Also in this pa…1f07845 -
docs: structure is not content -- H9 came back near-silent
Unverified signatureAll twelve artifacts passed every structural check. Content was inspected for two of them and one looked wrong. That sentence is the honest state of this build and it is now the sentence the mode table is read through. video_to_audio produced roughly -55 dBFS: RMS 58.3 against text_to_audio's 18013.9 at the same settings, about 309x quieter, on a 32767 ceiling. Different md5 so it is not an echo of H8, but effectively inaudible where H8 is healthy. Recorded as OPEN -- defect, property of the conditioning clip, or correct is not established, and whether H5/H6/H7 are also quiet decides whether this is one defect or nine silent artifacts. Channels, rate, duration, non-emptiness and verify_streams exit 0 all pass on H9. Silence and a healthy signal have identical stream layout. This is the 'output looks right?' column failing exactly where it was predicted to, which is the one column no instrument can fill -- so the mode table now says these runs produce well-formed files rather than cor…ffe4f9a -
docs: the staffing conclusion, and stage 2 has run
Unverified signatureStaffing, as stated by the lead: two makers and a lead who verifies. The scarce resource was execution on real hardware, not review capacity -- at 1926 passing, six of eleven modes were broken and none of it was found by reviewing harder. Budget for another agent is better spent on a second executor with different hardware than on a second reviewer. And the default-REJECT verifier role was filled by reality rather than by an agent: where a real gate exists, prefer it, because an agent can be persuaded and a RuntimeError cannot. H14 has produced an artifact: 1408x960, 2,254,799 B, 84.3 s. Its proof is its RESOLUTION rather than a same-seed diff -- exactly 2x H1 in both dimensions is a size only the spatial upsampler can produce, which is stronger than showing that two runs differ. Stage 2 peaked at 83.897 GiB against 70.497 for single-stage, so it cannot run on a large 48 GB MIG slice at all and leaves ~11 GiB on xlarge before packing overhead. One of the few hinode numbers with a dir…569d803 -
docs: the operator handoff — what exists, what does not, what is next
Unverified signatureThe document that outlives the session. PLAN records how things were decided and TEST_PLAN records what was run; this one tells someone who was not here where the project actually stands. Every load-bearing number carries its provenance: measured on hinode, read off a pinned source, inferred, or not yet established. A number with no mark is a defect in the document, and it says so. It does not claim the Space works, and says outright that it has evicted twice and never run -- because the single most likely misreading of nine green modes is that they constitute a deployment. The ZeroGPU section points at ZEROGPU_DISK.md for the arithmetic rather than restating it, so there is no second place for that truth to live. H10/H11 and H14 are marked pending rather than softened, with the reason each matters: H10b is the acceptance test rather than H10, because noise_scale=1.0 discards the input by design and a bug that drops it entirely is indistinguishable from correct operation. connectors…635ce6e