Spaces:
Running on Zero
Running on Zero
Commit History
2 days ago
-
fix: below 1024 px a control reached by Tab stops above the commit bar instead of under it: the page holds back the bar's height when it scrolls (round 4's N5, WCAG 2.4.11) (#20)
Unverified signatureBelow 1024 px the commit bar (the "Needs a prompt" line and Generate) is fixed to the window's bottom, 97 px tall. A browser scrolls a control reached by Tab only as far as the window's bottom edge, so it came to rest under the bar. The round-4 rig caught it at 390 on 1d8deb6: Tab from Advanced to the monitor's Play scrolled the page to 1104, and Play (y 800-844) lay entirely under the bar (y 747-844). That fails WCAG 2.2 SC 2.4.11, Focus Not Obscured (AA). On 25e35b7 the same Tab had left Play clear of the bar, at y 518-562; nothing held back the bar's height when the page scrolled, then or now. One rule in the below-1024 px block: html { scroll-padding-bottom } takes the same room the page already keeps below its content for the bar (var(--mr-bar-h) + the safe area + 16 px). A browser honours a scroll container's padding when it scrolls a focused control into view, so the control stops above the bar (WCAG technique C43). Desktop is untouched: its Generate sticks inside the rail. te…ea5b282 -
fix: the model's own sound reaches the writer unclamped: both runners build LTX-2.5's vocoder with the pinned forward minus its final clamp to [-1, 1], so sound over full scale is turned down as a whole by the writer instead of clipped inside the model (#14)
Unverified signatureThe lead's real-weights repro on the final backend (mode 05, 704x480x49, seed 7, evidence/local_runs/D2_S2_m05_49f/) still wrote a wav with 2 samples at full scale (int -32768) and no turn-down line. LTX-2.5's vocoder is diffusers' LTX2VocoderWithBWE (snapshot vocoder/config.json; final_act_fn and bwe_final_act_fn are null, so nothing earlier bounds it), and its forward ends `waveform = torch.clamp(residual + skip, -1, 1)` (pinned diffusers, pipelines/ltx2/vocoder.py:594). The model clipped its own sound to exactly full scale, so outputs._under_full_scale (peak > 1.0) never saw an over-range track. Deploy 1's 78 pinned samples were this clamp. audio.forward_without_final_clamp is the pinned forward line for line, minus that clamp: the stage-1 vocoder, the pad to the hop, the mel, the bandwidth extender, the resampled skip, the sum and the trim to the output length. audio.unclamp_vocoder sets it as the vocoder instance's forward. It is called right after each runner builds its componen…80f9b11 -
fix: an adapter that passes its check is staged again, and Remove, Remove all and its strength work again: the adapter list carries no change listener, so Gradio no longer hashes the staged adapters, which cannot be hashed; its own events collapse a failure card naming the adapters, when they changed the list (#20)
Unverified signatureSince minor 2's failure-card remedy (7896071), every form field carried a change listener so that a card collapses when a control its step names changes, and the adapter list (the field `loras`, a gr.State) carried one too. Gradio 6.24 hashes a State that has a change listener before and after every event that writes it (Blocks.process_api: get_state_ids_to_track, then utils.deep_hash), and a staged adapter carries the validator's report, a frozen dataclass holding dicts: deep_hash raises "unhashable type: 'dict'". So on the page, once an adapter passed its check, Check and add and the upload failed with "Error" where the accepted card belongs (nothing staged, the file left in the drop zone), and with an adapter staged Remove, Remove all and the strength failed before their handlers ran: on abc9150, 12 events. The handlers' own tests call them directly and never met Gradio's tracking; the round-4 rig's S8 did. The adapter list is the only State among the form fields (images, video, au…1d8deb6 -
fix: mode 11's first-frame line names the cause the round-4 verdict found: with a speech recording and no first-frame picture, some seeds come out as a black-and-white period film; add "in colour" to the prompt, or give a first-frame picture (#28)
Unverified signatureThe round-4 verdict found that the black-and-white result depends on the seed (the same recording and prompt gave a period film at seed 7 and colour at seed 21), not on how old the recording sounds, so "An old-sounding recording can come out as a period film" sent the user looking at the wrong cause. The sentence is now exactly: "With a speech recording and no first-frame picture, some seeds come out as a black-and-white period film: add “in colour” to the prompt, or give a first-frame picture." It stays where it was: in Audio to video's first-frame caption only, under the picture and above the prompt, the two remedies it names. UI_DESIGN's paragraph on mode 11's help says the same. Test (tests/test_ui_copy.py, re-pinned through the built page's own switch): in Audio to video exactly one shown component holds the new sentence, between the picture and the prompt, and the caption is "Optional. Used as the first frame. " followed by it, with no "old-sounding"; in every other mode no com…9887158 -
test: the access test pins the sample jump's desktop rules as they stand after M4: hidden beside the monitor at 1024 px and wider, shown there with mr-over-run while the monitor holds a run of another mode, and with no run the jump carries no mr-over-run (#20)
Unverified signaturee4fc218 -
style: this branch's new code and tests are laid out as ruff format lays out the files they sit in; no behaviour changes (#20)
Unverified signatureThe lines the earlier commits on this branch added to ui/blocks.py (already formatted at the base) and ui/render.py, and the test files they created, now pass ruff format. Lines of render.py the formatter would change that this branch did not write are left as they were.f16d35d -
fix: Restore settings says what it did, the seed included: a quiet line in Recent runs reads "Settings from 08:06 restored; the seed is random again, as it was for that run. Your current audio stays as it is." until the next run (#20)
Unverified signatureRound 4, minor 10. Restore writes the run's own settings back, its seed setting among them (usually -1, a new seed each run), and keeps the current media. On a desktop nothing moved and nothing said so (S11 and S12 are the same frame), and "Reuse seed" then "Restore settings" on one row silently went back to a random seed. Restore keeps writing the run's seed setting (it is one of that run's settings; Reuse seed is the seed's own button) and now says so, with the media it keeps, in a line under the Recent runs heading (elem_id recent-note, the page's quiet hint style); the next run's start takes it away. Tests (tests/test_ui_restore_says.py, through the built page's own steps): Reuse seed then Restore writes -1 back and says the seed is random again, naming no media in a mode with none; a run at seed 42 in Audio to video is restored with "with that run's seed, 42" and "Your current audio and image stay as they are."; the next run hides the line; the words.6150e0e -
fix: a LoRA rejection gives one next step for its kind ("Use an adapter made for LTX-2.5.", "Download the file again, then upload it.", ...), and off ZeroGPU neither the card nor the idle monitor speaks of spending (#20)
Unverified signatureRound 4, minor 3, without the operator decision. The rejection card read "Rejected: the adapter's shapes don't fit LTX-2.5. <file> was not staged. Nothing was spent." then the validator's words: no next step, against the page's own rule for a failure (what happened, then one step), and "Nothing was spent." on a machine with no meter. Each rejection kind now has its step (render.lora_step): another adapter made for LTX-2.5 for a version, layer-name, shape, target or attach refusal; the Hub id or an upload for a fetch failure; a fresh download for a cut or unreadable file; a standard LoRA for an IC LoRA; a smaller file for one over the limit; a report for an out-of-date layer table or an error the app has never seen. "Nothing was spent." is the Space's sentence alone (the handler words the card for its platform), and the idle monitor's line with no sample reads "Ready. Nothing runs until you press Generate." off ZeroGPU (on the Space it is unchanged). "spent" and "spend" join the meter …1583c93 -
fix: a failure card no longer outlives Show, a Restore or its own remedy within one mode: each turns it into its one line, "Last run: <mode>, <time>. <headline>." (#20)
Unverified signatureRound 4, minor 2 (R4's residue). R4 collapsed a failure card when the mode changed. Within one mode it stayed red and open: in the round-4 captures Enhance was already unticked and "The prompt enhancer is unavailable. Turn Enhance off to run with your own prompt." stood under Generate, through Show, Restore, LoRA staging and Upscale. Now the same line R4 writes replaces the card when Show puts another run on the monitor, when Restore writes a run's settings (in its own mode too), and when a control the card's step names changes: Enhance for the enhancer card; Upscale, Enhance or LoRA for the quota card; Enhance or Upscale for the per-run limit; size, length or Upscale for a timeout or memory card; LoRA for an adapter card; any field for a refused input (its detail names the field); the media for "Your inputs could not be read". A card whose step names no control (press Generate again) stays until the next run. render.step_names holds which fields each card's step names; each field's c…7896071 -
fix: a cost lamp's state is not in colour alone: an unlit lamp is hidden from assistive tech and the lamps' row is a named group, so what is read is the costs that apply (#20)
Unverified signatureRound 4, minor 12, the panel's part (the visual lens's N2). The lamps marked lit and unlit with a class only, and named their row with aria-label on a plain div, which ARIA does not allow and screen readers skip; a screen-reader user heard every lamp as if it applied, and off ZeroGPU a lit and an unlit lamp carry the same words ("Upscale", "LoRA"). WCAG 1.3.1. An unlit lamp now carries aria-hidden="true" and the row role="group" with its name, as the lens proposed. The picture is unchanged. Test (tests/test_ui_lamps_read.py, panels from handlers.review): at the defaults every lamp is unlit and hidden; with Enhance on, the Enhance lamp is lit and read and every unlit lamp is hidden; the row is a named group. tests/test_ui_parked_upscale.py's lamp reader now reads a lamp whatever its attributes.44edefb -
fix: the run summary's time for a run under half a second reads "under 1 s" (on ZeroGPU "under 1 s of GPU time, under 2 s of your quota"), never "Took 0 s" (#20)
Unverified signatureA sub-second run, as on the tiny checkpoint, read "Took 0 s on the CPU" in the Done card, and "It used 0 s" in the card for a result that could not be shown: a run that took no time at all, which is not what happened. took_words rounds as before, and a run that rounds to 0 s says "under 1 s"; on ZeroGPU the quota it drew is "under" the multiplier's seconds rather than "about 0 s". Half a second and more still reads "1 s". Tests: tests/test_ui_report.py adds the sub-second case on the CPU, unlabelled, on ZeroGPU and at 0.6 s; tests/test_ui_run_end.py's check that the card states what the run used now accepts "under 1 s" (its fake run is sub-second).74d0925 -
fix: each sample's caption says it was made earlier by this app from its example, with that example's inputs and seed, and no longer promises the same result from the same prompt and seed (#20)
Unverified signatureRound 4, minor 8. The samples were made on the operator's Mac (Q8 GGUF transformer); on the Space the same prompt and seed give another take (mode 05 at seed 12: two crewmen on deck there, the ship from above in the sample), yet the captions said "Same prompt and seed as “Storm on deck”" or "same seed", inviting a Space visitor to expect the sample from 360 s of free quota. Every caption now reads "Made earlier by this app from “<example>”, with its <prompt / picture / clip / recording> and seed: ...": what the sample is and where it came from, with no promise of a repeat. The words live in examples/manifest.json. Test (tests/test_ui_sample_captions.py): every shipped caption starts "Made earlier by this app from “" and says neither "Same prompt and seed" nor "same seed".b46dfc2 -
fix: the sample captions name the example each was made from, state the samples' real lengths ("4.7 s, the longest that fits the 5.0 s recording") and say "no upscale" in place of "single pass" (#20)
Unverified signatureRound 4, minor 11. Modes 10 and 11 called 4.7 s "the recording's length" for a 5.0 s recording (examples/media/ocean.wav measures 5.000 s; the run is the longest length on the frame grid that fits it, 113 frames), mode 11's "From the example's recording and prompt" did not say which of its two examples, and "single pass" is builder shorthand. Each caption now names its example in quotes ("Made by this app from “Waves, pictured”, with its recording, prompt and seed: LTX-2.5, no upscale."), the video modes say "no upscale", and the settings of modes 10 and 11 read "4.7 s ..., the longest that fits the 5.0 s recording". The words live in examples/manifest.json, beside the samples they describe. Tests (tests/test_ui_sample_captions.py, over the shipped manifest): every caption names exactly one of its mode's examples, whose seed is the sample's, and carries no "single pass" or "the example's"; every video mode's says "no upscale"; every length a sample's settings state is the file's own …9ab0812 -
fix: the Done card says "chosen by the model" only for a length strictly inside the range the user allowed, and at either end which end it is; "How quota works" gives the timeout card's step; no "legal" on the page (#20)
Unverified signatureRound 4, minor 14, three parts. The Length row. With "Let the model pick the length", diffusers clamps the duration head's figure to the shortest and longest the user allowed and snaps it to the frame grid; on the Space the head said 4.47 s, the range made it 5.0 s, and the card said "chosen by the model" (issue #16). The run record carries the frames rendered and the request the range; constraints.duration_head_frame_range gives the frame counts the head returns at each end, exactly as the pipeline clamps. A length strictly between them is the model's own figure, and only there does the card say "chosen by the model". At an end it says "the shortest you allowed" or "the longest you allowed" ("the only length you allowed" when both ends meet): true whether the range moved the figure or the model landed there, and the run's own note (#16) says what the model chose. The record has no field for the head's figure, so the card does not claim more than this. "How quota works" said a timed-…a8bf7a4 -
fix: "Use this prompt" turns Enhance off in the same step and says so ("Copied to your prompt; Enhance is off"), so the next run sends the enhanced prompt as it is; in a mode that takes no prompt it parks both instead of writing a hidden value (#20)
Unverified signatureRound 4, minor 9. After a run with Enhance on, the Done card offers "Use this prompt" with the prompt the model was sent. It wrote that prompt into the box and left Enhance on, so the next run enhanced the enhanced prompt again (on the Space, reserving the enhancer's seconds again): the next run did not use "this prompt". The click now writes the prompt and unticks Enhance together, and the button reads "Copied to your prompt; Enhance is off", disabled, until the next run offers it afresh. Found on the way: in a mode that takes no prompt (Image to video, after a switch) the click wrote into the hidden prompt box, which a run would then send; it now parks the prompt and Enhance, which come back with the prompt box, as a mode switch does. Tests (tests/test_ui_use_prompt.py, a run with a faked enhancer through handlers.run and the real generate(), the real finish step and click): the prompt is written and Enhance turned off, with the button's words; the next run offers the button afresh…8183336 -
fix: off CUDA a parked Upscale promises no size: while this machine cannot run it for the request, its label is the bare "Upscale ×2" and the cost panel shows no Upscale lamp; at a size it can run, both come back (#20)
Unverified signatureRound 4, minor 7. Where this machine cannot run Upscale ×2 for the request, the control is parked (unticked, disabled, the backend's reason as its line), but its label still read "Upscale ×2: returns 1408 × 960" over "Upscale ×2 needs a CUDA GPU; this machine runs on the CPU.", and the panel kept an unlit Upscale lamp. P12's rule, a size only when it will run, now holds here too. The review step, which already decides the parking, now writes the label with it: the bare label while parked, the returned size for a size the model takes otherwise. The size-hints step no longer writes the label, so two steps fired by one change cannot race over it. The panel drops the Upscale lamp while it is parked. Tests (tests/test_ui_parked_upscale.py, on the CPU platform, through the built page with its first paint): parked at the default size with no size in the label and no Upscale lamp; at 64 × 64 × 9 freed with "returns 128 × 128" and the lamp; at 704 parked again with the bare label; only the re…44c216d -
fix: arrow keys on a LoRA strength slider reach the run: the slider's input is wired as well as its release, last value wins, and the value it sends is held to the slider's own range (#20)
Unverified signatureRound 4, minor 1. The strength was wired to the slider's release only, and Gradio 6.24's slider sends release on pointer-up over the range or on leaving its number box; arrow keys on the range send input and change. A keyboard user saw 0.7 on the slider while the LoRA label, the cost panel's LoRA row and the run kept 1.0. Each row's slider now also listens to input (trigger_mode always_last), running the same step and the same label and review after it. Input fires on every keystroke in the number box too, and a typed value passes through numbers outside the range ("0" on the way to "0.7"), which Gradio's own slider preprocessing refuses with an error on screen; so the value input sends is held to the slider's min and max as the page holds them (read off the range element, elem_id lora-strength-N, since the LoRA package sets the range lazily). The pointer path is unchanged. Tests (tests/test_ui_lora_strength_keys.py, through the built page's own steps): an adapter staged through the …d468ab6 -
fix: the line beside Generate names what the run returns: the returned size, with "upscaled" when Upscale ×2 is on ("1408 × 960 upscaled, 5.0 s, needs … quota"), and "Sound" in a sound-only mode ("Sound, 5.0 s, …"), keeping the length and the quota or device words (#20)
Unverified signatureRound 4's M1. The one-line figure beside Generate (the desktop line over the button on a short window, and the phone and tablet bar's ready line) started from the render size: under Upscale ×2 it said "704 × 480, 5.0 s, needs 540 s quota" for a run that returns 1408 × 960, and in the three sound-only modes it named the 256 × 256 picture they render and throw away. The panel's Returns row, a few centimetres above, said otherwise. figure_line now starts from what the run returns, as the Returns row does: the returned size, with a no-break "upscaled" when stage 2 is on, or "Sound" when only sound comes back; then the length (or "the length of your clip" / "the length of your audio" where an input sets it, as the panel says), then the quota needed on ZeroGPU or where it runs elsewhere. commit_line's ready state is the same line. Tests (tests/test_ui_rail.py): the every-mode test that pinned the render size (test_the_mode_names_its_figure_in_every_mode) is rewritten to pin the returned si…05737f6 -
fix: mode 11 says an old-sounding recording can come out as a period film, and how to get colour: one sentence under its first-frame picture, above the prompt, in Audio to video only (#28)
Unverified signatureOn the Space the same prompt and recording gave a 1940s-film picture at seed 7 and colour at seed 21; with a first-frame picture the result is in colour. A user who expected colour got a period look and did not know why. The sentence is "An old-sounding recording can come out as a period film: say “in colour” in the prompt, or give a first-frame picture." It is in the caption of Audio to video's optional first-frame image, which sits between the picture input and the prompt box, the two remedies it names; the audio caption keeps its three sentences about lip-sync as round 3 pinned them. Test (tests/test_ui_copy.py, through the built page's own switch): in Audio to video exactly one shown component holds the sentence, the image caption, between the image and the prompt; in every other mode none does.9dcc63c -
fix: the decoder choices are named by what they do: "Auto", "Steadier (diffusion decoder)" and "Sharper (standard decoder)", in the choice, its line, the Upscale lock line, the Done card and the backend's refusals; the values sent stay auto, diffusion and conv (#27)
Unverified signatureOn the Space the same seed and prompt through both decoders gave the same scene: the diffusion decoder a little softer and steadier from frame to frame (fine detail 10.3 against 11.6, shimmer 1.47 against 1.88), the standard decoder sharper frame by frame with about 28% more shimmer (evidence/space/matrix/ab_decoder). The page called the softer one "Best quality" and the sharper one "Fast". Now the choices read "Auto" (still the vendor's default), "Steadier (diffusion decoder)" ("A little softer, steadier from frame to frame; takes a few seconds longer.") and "Sharper (standard decoder)" ("Sharper frame by frame, with a little more shimmer; faster."). Auto's line is true on every machine: "Diffusion decoder on an NVIDIA GPU that can run it, standard decoder elsewhere." Off CUDA, Auto always takes the standard decoder (pipelines.StageRunner._single_stage_decoder), so the old "where this machine can run it" was untrue on a Mac, and round 4 asked for the NVIDIA wording. The Upscale line …1c8a58d -
feat: a video + sound result offers its sound track as a quiet "Download the sound (WAV)" in the Done card's action row, beside Reuse seed, only when the run returned the picture and the sound as two files (#6)
Unverified signatureF30. Since 20155a1 every video + sound run also writes its sound as its own WAV and the public endpoint returns it, but the page played the video in one player and offered no way to save the sound alone. The action row that closes the Done card now holds a gr.DownloadButton (quiet, elem_id sound-download) set by the finish step to the run's WAV when the run returned a video and a separate sound file that is still on disk, and hidden otherwise: a sound-only or silent run, one file carrying both tracks, a failed run, and while a run is in flight. It is in the action row, not the monitor's corner toolbar, as round 4 asked (the corner clips focus rings, P17). Tests (tests/test_ui_sound_download.py): runs through handlers.run and the real generate() with the production writer (the pipeline faked), then the real finish step: a video with sound shows the link with the wav's path, beside Reuse seed; a sound-only and a silent run do not; one muxed file is not offered twice; a new run and a fa…99f710e -
fix: a request Gradio refuses before any handler runs can no longer leave the page on "Generating…": typed number fields declare no range, and a last failure listener that reads no input brings Generate back with a card (#5)
Unverified signatureF28, the hole left in F72. Two ways Gradio could fail the generation step while preparing its inputs, before any of the app's code ran: a seed typed below the field's minimum of -1, and an upload whose file Gradio's cache clean-up had removed after six hours. The failure step that brings Generate back takes the same inputs, so it failed the same way and the page stayed on "Generating…" with no card. Both reproduced through Gradio's own process_api: _on_generate Error then _finish_failed Error, and nothing after. The seed field no longer declares minimum=-1 (no typed number field declares a range now; sliders keep theirs), so a seed of -2 reaches the app's own gate, which refuses it before any GPU time in its own words ("seed=-2 is out of range: use a whole number from 0 to ..., or -1 for a random seed."), and the panel says so before Generate is pressed. After each failure step a last .failure listener (_finish_bare) takes no inputs, so nothing Gradio prepares can stop it, and sends …bfd58d1 -
fix: after a run every mode's sample can still be seen: while the monitor holds a run of another mode than the selected one, the button under the purpose line reads "See an example of <mode>" at every width (class mr-over-run), puts that sample on the monitor and lets later switches follow with their own, and reads "Back to your run" while a sample covers the run (#20)
Unverified signatureRound 4's M4: after one successful run no mode's sample could be seen again in the session. The monitor rightly kept the run when the mode changed (AC-06), the jump to the sample showed only while the monitor held a sample, and it was hidden at 1024 px and wider; a reload was the only way back, and it emptied Recent runs. The monitor state of a run now carries its mode, seed, files and whether it is an earlier run (Show). The button under the purpose line is decided from the selected mode and that state in one place (_jump_update): over a run of another mode whose mode has a sample it reads "See an example of Audio to audio" with the class mr-over-run (UI-A's stylesheet shows it on a desktop with that class); while a sample covers a run it reads "Back to your run", as long as the run's file is on disk; while the monitor shows a sample with no run under it, it is the phone's plain jump. A click puts the sample on the monitor with its usual caption and sets the state to "sample" with th…78036d2 -
fix: every video player on the page works without a hovering pointer, not only the monitor's: a loaded clip's preview shows its bar to keyboard focus, while paused and on a touch screen, names its Play and Full screen, lets only Enter and Space act, insets its corner buttons and keeps 44 px targets on a phone (#20)
Unverified signatureReview round 4, M3, extended. The page shows two Gradio video players, the monitor's (#video-out) and a loaded clip's preview in Video to video (#in-video), and the clip's had all of M3's defects: its bar stayed at opacity 0 with keyboard focus on Play, Play was named "play-pause-replay-button", and Tab toggled the clip's playback (the probe on the rebased head: paused, then playing, after one Tab). The bar's rules and the script now find a player by Gradio's own markup (a .wrap holding the picture's .mirror-wrap and the .controls bar) instead of #video-out: the bar shows to keyboard focus, while paused and under (hover: none); players() wires every `.mirror-wrap > video`; the key guard acts on any control it has named; and below 1024 px every player's controls are 44 px with 20 px glyphs. The corner inset names both blocks, :is(#video-out, #in-video), since a loaded clip's reset, trim and clear sat flush in the media zone's clipped corner too; a test checks that these are the only tw…fe49110 -
fix: on a phone a loaded audio's playback and editing controls and every media zone's source buttons are 44 px targets too (#20)
Unverified signatureReview round 4, minor 12, continued. The round-4 rig's B2, now also run with an audio file loaded (F5, Audio to audio, 390 x 844), found Gradio's audio editor controls 20-40 px wide: volume, playback speed, skip back, play, skip forward, reset and trim, and the zone's source buttons ("Upload file", "Record audio", 22 px). The 44 px rule had raised their height only. Below 1024 px they now get a 44 px minimum width, with the zone's download and clear buttons. Gradio sets the editor's controls out in two rows at 600 px and narrower (playback on its own row; volume and the editing buttons under it), which holds them at 44 px in the 320 px zone. The rule for the loaded editor's controls is kept to the audio zone: a loaded clip's trim controls (the video zone) were not seen by the rig.9e59de0 -
fix: on a phone a media zone's own buttons (download, clear) and a checkbox's label are 44 px targets like every other control (#20)
Unverified signatureReview round 4, minor 12 (the phone targets; visual N3 a and b). With a file loaded, the media toolbar's download and clear buttons were about 25 and 20 px wide (the 44 px rule raised their height only), and Enhance, Upscale and "Let the model pick the length" were a 16 px box in a 20 px label. Below 1024 px the toolbar's buttons get a 44 px minimum width and a checkbox's label a 44 px minimum height. The rig's B and B2 now also run in F5 (a loaded media zone) and P3, and test the checkboxes' labels.07beaa9 -
fix: Generate in flight is drawn busy, with a solid edge and a paper label, not like a refused Generate (#20)
Unverified signatureReview round 4, minor 13 (visual N10). "Generating…" had the refused button's dashed edge, slate fill and dust label, because the rule for a run in flight styled a data-running attribute that nothing set. The script now sets data-running while Generate is disabled with the start step's label ("Generating…", blocks.py's words, pinned in the test) and clears it when the label comes back; the rule draws a solid dust edge and a paper label. The unused .is-running selector is gone.700b8bc -
fix: under a run's report the monitor's caption keeps no blank reserve; the four-line reserve that keeps Generate level stays for browsing modes (#20)
Unverified signatureReview round 4, minor 5. The caption reserves four lines less the panel's notes so Generate does not move from mode to mode (P16), and kept that reserve after a run: a 57-69 px blank band under a one-line "Your run." caption that at 1440 x 790 left only the Done chip in view. The reserve now applies only while no run's report (the Done card, or a failure's card) is shown.81032ed -
fix: after a run the head of its report fits in the rail with the monitor, so following it no longer slides the monitor's top out of view: the monitor gives up the height instead (#20)
Unverified signatureReview round 4, minor 6. fit() sized the rail at scroll 0 for the monitor, caption, panel and Generate only; when a run ended, follow() scrolled the rail to bring the Done card's head (or a failure's next step) in, which cropped the top 51 px of the result at 1366 and 1280. fit() now counts that head, from Generate's foot in the rail's flow to the head's bottom plus its scroll margin, and the monitor gives up the height, letterboxed, down to its 180 px floor; follow() and fit() share reportTarget(), so they make room for, and follow, the same element. (The 89 px crop when Generate is pressed with the rail scrolled, S5b, needs the monitor scroll repeated once the start step's outputs land, in blocks.py.)0cf63ed -
fix: the idle rail no longer scrolls over nothing: an empty Reuse-seed row and an empty Recent runs take no room in it (#20)
Unverified signatureReview round 4, minor 4. At idle both are flex children of the rail with nothing in them, and each added the rail's 12 px gap after Generate: a 22-24 px scroll with a near-full-height scrollbar at 1366 x 657, 1280 x 680 and 1440 x 790, and with overscroll contained, a wheel over the rail nudged the sample instead of scrolling the page. They are display: none while they hold no shown button, as the empty Recent-runs rows already were.5a6543d -
fix: the cost travels with a stuck Generate: the figure line stands above it whenever the panel's Needs row (Runs off ZeroGPU) is out of view or under it, decided on every rail scroll, and only then; and on desktop Restore settings brings the cost panel into view (#20)
Unverified signatureReview round 4, M2 (with layout m1). The line was decided once, from the rail's layout at rail scroll 0 (the sum of the monitor, caption, panel and Generate), and scrolling the rail only moved the fades. So once Generate stuck at the rail's top (Recent runs read, six runs listed, after Restore) it carried no cost at all, while R2 made it pressable from anywhere in the rail; and at 1366 and 1280 on the Space the line showed although the Needs row stood in view above Generate, because the panel's outer box ran a few px under it, and its own 25 px then pushed the LoRA row under the fade. figureLine() reads where the figure row actually is: the line shows exactly when that row is not wholly inside the rail's visible part clear of Generate (above the rail's top, under Generate, or below the rail). It is judged from Generate's box as it would be without the line (stuck at the rail's foot the line grows Generate upward), and a shown line goes only once the row is clear by 6 px, so neither th…3b92225 -
fix: the jump to a mode's sample also shows at 1024 px and wider while it carries mr-over-run, the state in which the monitor holds a run from another mode and the jump is the only way to that mode's sample; it stays a quiet text button in the form column (#20)
Unverified signatureReview round 4, M4, the stylesheet's half of the contract with the behaviour change: the page gives #sample-jump the class mr-over-run (and its words, "See an example of <mode>" or "Back to your run") while the monitor holds a run from another mode. The desktop rule that hides the jump beside a monitor already in view stays as it is for every other state; a more specific rule shows the jump in that one state, with the button's own display (flex). It sits under the purpose line in the form column, so nothing in the rail moves and Generate keeps its place.f8393c1 -
fix: the monitor's player works without a hovering pointer: its bar shows to keyboard focus, while the video is paused and on a touch screen, in the well colour; Play is named "Play" or "Pause" and full screen "Full screen"; only Enter and Space act on them, so Tab no longer starts or stops a clip (#20)
Unverified signatureReview round 4, M3. Gradio 6.24's player (VideoPreview, pinned) keeps its control bar at opacity 0 unless the pointer hovers the player, names Play "play-pause-replay-button" and toggles playback on any keydown on it (a delegated handler that never looks at the key), so a keyboard user crossed three invisible tab stops on the monitor and started the clip, with sound, by tabbing past Play; on a touch screen an idle sample was a still picture with no sign that it plays. Nothing in Gradio is patched: - the stylesheet shows the bar while keyboard focus is inside the player, while the video is paused and always under (hover: none), and paints it in the well colour instead of Gradio's grey-800. Keyboard focus is `:has(:focus-visible)` rather than `:focus-within`, so a mouse click on Play does not pin the bar over a playing video; - the script sets .mr-paused on the player from the video's own play, pause, ended, emptied and loadedmetadata events, names Play and full screen, and a ca…93a7f52 -
test: the Enhance reservation's +80 s rests on the image preloading the enhancer, and now says so: its test checks the preload entry at the enhancer's pin, the note beside ENHANCE_WINDOW_SECONDS names the preload with the floor variable as the lever, and PRE_SECONDS stays at 25 s for a cold first call now that the decoder is built at start-up (#21)
Unverified signatureWith the enhancer preloaded by the image (preload_from_hub, the start-up prefetch as the fallback), a first enhanced call in a fresh container no longer pays the ~88 s Hub fetch inside its reservation, which the +80 s allowance was never sized for; the test fails if the preload entry goes, so the allowance cannot silently lose what it rests on. The 25 s before the first step was measured on a fresh worker that still built the diffusion decoder per run; that build now happens at start-up, and the 25 s is kept for a cold first call, the margin the Space runs were measured against.1b0022d -
fix: the GPU reservation follows each run's size and length, from the Space's measurements with a 1.3 margin, so a default run needs 120 s of quota to start instead of 360, and on the Space a run past the measured envelope is refused before any GPU time (#21)
Unverified signatureEvery run reserved 180 s (LTX25_GPU_DURATION_SECONDS) x1.5 with Upscale, +60 s with an adapter and +180 s with Enhance, whatever it rendered; ZeroGPU wants twice that free on xlarge, so a visitor with less than 360 s of quota could not start a 25-second default run, and nothing stopped a run too large to finish inside its reservation. zerogpu.estimate_duration now reads the request as the run will use it (a sound-only mode's pinned 256 x 256, no Upscale where the mode skips it, the duration head's longest pick, the length a recording or a clip set at the gate and the longest the app allows until it has, the tail of the schedule a clip variation runs) and models its window from the Space's measurements (six deploy-2 runs up to 1920 x 1088 at 10 s and the 24 deploy-1 calls): 25 s before the first step, steps x t(N1) with t(N) = 0.46 + 1.81e-4 N + 1.94e-9 N^2, then 1.0 s + 1.3e-7 s per returned pixel-frame; Upscale adds 3 s + 3 x t(N2) and its larger post; Enhance 60 s; any adapters 30 s…ac68de5 -
test: with the diffusion decoder built at start-up, and with its per-run fallback, a run's time line still has its own "decoding the frames" phase after sampling, and its split still adds up (#17)
Unverified signatureAfter the rebase onto the time-line work (#18), the run's "took ... s" diagnostic times the progress bar's phases. The decode phase is stamped by StageRunner._diffusion_decoded right before the decode call, which the start-up build does not change: what changes is that the runner's decoder lookup before sampling no longer loads anything, so the decode phase times the decode alone. On the tiny checkpoint, through generate() in the worker context, both runs with the start-up decoder and the run after a failed start-up build now also assert their time line, using #18's own parser (test_run_report_lines._adds_up / took_lines): one line, "decoding the frames" present and after "sampling", the phases adding up to the total. Seen locally: "took 0.9 s on cpu: loading the model 0.0 s, preparing the inputs 0.0 s, sampling 0.0 s, decoding the frames 0.7 s, writing the file 0.1 s". A confirmation of existing behaviour; it passes as written.5e06a43 -
fix: the image preloads the prompt enhancer too: a third preload_from_hub entry, generated like the others, fetches the 8 files its loader reads at its pinned commit (2.49 GB), so a first enhanced call in a fresh container no longer risks paying the ~88 s Hub fetch inside its GPU reservation; the start-up prefetch stays as the fallback (#23)
Unverified signatureThe reservation for Enhance is now sized for a warm enhancer (+80 s, #21). A first enhanced call in a container whose daemon prefetch (app.py, enhance.start_prefetch) had not finished would have spent the ~88 s download inside that reservation and could have been aborted. The enhancer repo, google/gemma-4-E2B-it-qat-mobile-transformers at dd693ff4, is not gated, so the Space's preload can fetch it at the same time as the model. One source. The repo id, the pin and the files with their sizes live in model_files (ENHANCER_REPO_ID, ENHANCER_REVISION, ENHANCER_SIZED_FILES / ENHANCER_FILES), and preload_from_hub_entries() emits a third entry from them, which the drift test keeps the README equal to. config.DEFAULT_PROMPT_ENHANCER_REPO and enhance.ENHANCER_REVISION now read those values instead of holding their own literals. The list lives in model_files rather than beside the loader in enhance.py because the architecture test (test_no_hardcoded_hub_strings_outside_model_files) admits a che…72b989d -
fix: start-up refuses a fallback download the disk cannot hold, before writing a byte: the missing files, ZeroGPU's packed copy still to be written and a 10 GB margin must fit in the free space (on ZeroGPU, also under its 150 GB cap), or start-up stops with one message naming the directory, the count and GB, the free GB and the remedy (#23)
Unverified signatureA download that fills the disk part-way is the worst failure a public Space can have; a clear refusal is better. ensure_weights now calls _require_room(missing bytes, count, cache directory, on_zerogpu) after it has found what is missing and before it says it is downloading. Needed: the missing bytes (from model_files' per-file sizes), a 10 GB margin and, on ZeroGPU only, the packed copy still to be written: ZEROGPU_PACKED_COPY_GB = 42.1 (40.6 GB measured on the 2026-11-13 boot, 41.24 derived, plus the 0.83 GB decoder now packed with it). A CUDA box that is not ZeroGPU packs nothing to disk, so it is not charged for one; it fetches the set with 81.9 GB free. Free: shutil.disk_usage of the cache directory (or its nearest existing parent). On ZeroGPU that is the host's filesystem -- the first deploy's report once read 1969.8 GB free of 8589.9 GB there while the workload was being evicted (pipelines.disk_footprint's docstring) -- so on its own it could never refuse. On ZeroGPU the room …ff62c55 -
fix: the start-up decoder is moved to its device only once it is fully built: the decode-pipeline wrapper is constructed before the move, so a build that fails at any step after the load has moved nothing and the per-run build still finds its file (#17)
Unverified signatureAt start-up ZeroGPU registers what is moved to cuda, packs it at launch and then prunes the file each packed tensor came from. In _decoder_part the LTX2VideoDiffusionDecodePipeline wrapper was built after decoder_model.to(device). Had that constructor raised, the decoder would have been moved (so packed, and its file pruned) but never cached, and each run's fallback build would then have found no file to load. The wrapper is now built first and the move is the last step before caching: load, NATTEN processor, tiling, the mmap-backing line, the wrapper, then the move. Wrapping before moving changes nothing else: the wrapper holds the same module, which moves in place. This also makes true what the first #17 commit's message said ("the move to the device is the last step before caching"); it was not, by one constructor. Test: tests/test_decoder_at_startup.py::test_a_failure_after_the_load_leaves_the_decoder_unmoved_and_the_run_builds_it (fake diffusers tree): the wrapper's first constr…d229cdc -
fix: the diffusion decoder is built once at start-up, beside the pipeline, so ZeroGPU packs it and every run's worker uses that instance instead of loading it again; a start-up build that fails logs one line and leaves the per-run build as before (#17)
Unverified signatureOn the Space the default decode is the diffusion decoder. It was built lazily inside each GPU worker and lost with it: every run loaded its 0.83 GB and resolved its NATTEN kernel over the network again (its load message appeared in every run's log of deploy 1). pipelines.build_decoder_at_startup(device=...) builds it with the existing _decoder_part, after the runners, and caches it where the runners' decoder factory already looks. app.py calls it at module scope in the CUDA branch, after the two get_pipeline calls, so its .to("cuda") is intercepted and registered for packing with the other components before demo.launch(); a forked worker's runner then finds the instance through _decoder_part's cache. get_pipeline itself still builds no decoder. Whether the NATTEN fetch needs a GPU at load time: it does not. diffusers' processor calls kernels.get_kernel("shi-labs/natten", version=1): the version-to-ref lookup and the variant download are network and disk; the build variant is chosen f…ed13532 -
fix: the Space's weights come from the image: preload_from_hub is live for exactly the 35 files the int8 Space loads (71.92 GB; the temporal upsampler, which nothing loads, is no longer fetched), every file carries its size in model_files, and start-up downloads only what the image lacks, saying how much (#23)
Unverified signatureThe Space is public now, so the gated repo can be preloaded; while it was private the preload step ran without the Space's secrets and got 401, and every cold start downloaded 72.19 GB itself (17 min 24 s on 2026-11-16, 14 min 55 s of it the download). model_files is the single source of the set and its sizes: each component lists its files with their byte sizes from the Hub's own listing at the pinned commits, size_gb is derived from them, and file_size / total_size_bytes answer per file and per set. temporal_latent_upsampler moves to the declared-but-not-fetched group: StageRunner wires only the spatial latent_upsampler and model_index.json does not list it, so it cost 0.26 GB on every start for nothing. The set goes from 37 files / 72,185,863,109 B to 35 files / 71,923,919,050 B. README front matter: preload_from_hub is live, generated by model_files.preload_from_hub_entries() (two entries, one per pinned commit: the decoder is pinned later), and its comment says what it does, why…41dcf05 -
fix: the time line times the bar's two newer words: a staged local load named by ProgressReporter.loading() and a decode inside the pipeline call are each a phase of their own, and the split still adds up (#18)
Unverified signatureSince aeab028 (#12) the bar has two more ways of naming a phase. The local runner names the loads of its staged pair with ProgressReporter.loading(), which shows "loading the model" at the bar's current position through _emit, not through phase(). A single pass that decodes inside its own pipeline call (the conv decoder, "Fast") names "decoding the frames" from the step callback after its last step, through decoding() and so phase(). The time line timed phase() alone, so on a staged run the load was booked to "preparing the inputs" (the split read loading the model, preparing the inputs, sampling, ... where the bar showed a second "loading the model" before "sampling"). The timing is now ProgressReporter._stamp(), taken out of phase(): phase() and loading() both call it, so a staged load is a "loading the model" phase of its own, in the bar's order, and the split still adds up to the total. The in-call decode needed nothing new: decoding() already goes through phase(). With nobody li…17782c8 -
feat: an API caller can see what ran: record.diagnostics carries one line per attached adapter and one time line split into the progress bar's own phases, and the time line is logged once per run at INFO (#18)
Unverified signature/on_generate returns the files, the notes and the status markdown whose Diagnostics block prints record.diagnostics. The page's Done card adds the adapter and the time from the run record, which an API caller never sees: on deploy 1, lora_m01_bts_w100/result.json named neither the adapter nor the time, and the harness could only bound GPU time from its own wall clock. record.diagnostics now carries, on every run: - one line per attached adapter, from the same (label, weight, attached) the record holds: "adapter bts at 0.8: attached to 1632 layers" (the weight as the page writes it, "layer" for one); - one time line: "took 23.4 s on cuda: loading the model 1.2 s, preparing the inputs 0.3 s, sampling 14.0 s, decoding the frames 4.3 s, writing the file 1.1 s". The total is record.elapsed_seconds and the device is the record's own word. The split is the phases the run reports to the progress bar, in the bar's labels and order (a two-pass run: sampling — pass 1 of 2, upscaling …0f7e71d -
fix: when the duration head's own figure lies outside the user's range, the note says what the model chose and what the range made of it, read off the head as it runs (#16)
Unverified signatureWith "Let the model pick the length" and a range of 5 to 10 s, the Space's duration head predicted 4.47 s; diffusers raised it to the shortest allowed length ("Duration prediction clamped: raw 4.47s outside [5.00s, 10.00s], using 5.04s (121 frames)") and the page said "The model chose the length: 5.0 s.", so the user was told the model wanted 5 s when the range decided. The raw figure exists only as a local inside diffusers' LTX2DurationHead.predict_num_frames: it calls the head (forward returns the predicted seconds), clamps, snaps to the frame grid and returns a frame count. Nothing keeps it. generate() now reads it with a forward hook on the head of the pipeline it calls (_DurationReading; the StageRunner's pipeline.duration_head, which the audio-conditioned derivative shares and the local runner keeps resident). The hook only reads the output and returns nothing, so what is sampled does not change; it is attached only under automatic duration and taken off in a finally after the p…9b280fb -
fix: the model's sound over full scale is turned down as a whole to a -0.1 dBFS peak instead of being clipped by the wav, the mp4 carries the same samples, and the run's diagnostics say by how many dB (#14)
Unverified signatureThree Space runs on deploy 1 wrote samples pinned at 0 dBFS (78, 30 and 14 in x1_m05_49f_single_pass, x2_m05_text_to_audio_video and enh_m05_text_to_audio_video): a 2 s shouted line took the model's float track over full scale. The wav is 16-bit PCM, and FFmpeg's float to 16-bit conversion saturated every sample beyond full scale. The mp4's AAC track was handed the same over-range floats and kept them, so the two files did not even carry the same sound: with a line peaking at 1.6 the wav was clipped flat at 0 dBFS (13,112 samples at or above 0.999) while the mp4 peaked at +3.6 dBFS. The model's float track becomes the written track in two places in outputs.py: _write_video, for a mode that exports video and sound (its planes are muxed and then written to the wav), and _write_audio, for a mode that exports sound only. In both, the model's planes now pass through _under_full_scale: when the peak is above 1.0 the whole track is multiplied by one factor that puts the peak at OVER_FULL_SCA…ed91409 -
fix: bitsandbytes' int8 cast warning reaches the log once per process instead of once per matrix product, so the Space's log keeps its start-up report and its real warnings (#15)
Unverified signaturebitsandbytes 0.50.2 logs "MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization" (logger.warning on bitsandbytes.autograd._functions) on every int8 matrix product. On deploy 1 that was 55,400 of the last 55,534 lines of the Space's log after eleven runs, and the platform keeps only the tail, so the start-up report and every real warning were pushed out. A logging.Filter on that logger (ltx25_studio/log_filter.py) lets the message through once per process and drops its repeats. It is keyed by the process id and the rendered message, so a forked worker records the fact once for itself and a different dtype is a different line. Every other record on that logger passes, the logger's level stays NOTSET so bitsandbytes' other warnings still reach the log, and no other logger is touched. The filter is attached by the logger's name, so bitsandbytes is not imported: its own getLogger(__name__) returns the same object, filter included, whenever it loads. install()…18bcc7a -
fix: the bar names what is happening: a staged local run's loads read "loading the model" at the bar's current position until the denoise really starts, and a single pass that decodes inside the pipeline call (the conv decoder, "Fast") reads "decoding the frames" after its last step (#12)
Unverified signatureMeasured on real weights: "sampling - 12%" for the first 60 s of a staged local run while the text encoder and the transformer loaded, and "sampling - 82%" for the last 34 s while the frames decoded; on the Space a Fast run went from "sampling" straight to "writing the file". The sampling callback, and with it the "sampling" announcement, was built while the stage graph put the call's kwargs together, before the runner's seam had made anything resident. StageRunner._call now builds it inside the seam, right before the pipeline call, so the local runner's text-encoder and transformer loads come first, and the local runner names them on the run's reporter as "loading the model" (ProgressReporter.loading): at the bar's current position, since that band is behind it by then, once per denoise and only when something actually loads. A repeated prompt with the transformer resident goes straight to "sampling". A single pass whose pipeline decodes inside its own call (the conv decoder) now sa…aeab028 -
fix: no entry can sample on an MPS PyTorch that fails the numerics self-test: build_local_runner applies the app's gate on mps, so scripts/local_generate.py --device mps exits refused (2) before loading anything, and the remedy names the README's floor, PyTorch 2.14 (#11)
Unverified signaturetorch 2.11 on MPS reuses the first scale scaled-dot-product attention saw, which gives wrong numbers without any error. The app's pipeline factory refused such a build through device.require_mps_numerics, but local_runtime.build_local_runner never called it, and scripts/local_generate.py builds the runner directly: on an affected torch it would have produced a wrong video. build_local_runner now calls require_mps_numerics on mps, after the watermark and availability checks and before any file is looked at, so every entry is covered by one check; LTX25_MPS_SELF_TEST=warn still lets an operator run unverified. The script counts MpsNumericsError among its refusals (exit 2). MPS_CLEAN_TORCH said 2.15 while the README states PyTorch 2.14 or newer (2.14.1 measured the same as 2.15.0 in the numerics checks), so the refusal and the unverified note now say 2.14, as the README does. Tests: the script test begun on wip/deploy2-w14 (ca91f07), unchanged; a runner-level test under enforce and warn…444bc2b -
fix: a bad MPS watermark pair stops the app's MPS start in its own words before the numerics self-test runs, and every numerics gate checks the pair first, so it is never reported as a PyTorch that computes wrongly (#10)
Unverified signatureWith PYTORCH_MPS_HIGH_WATERMARK_RATIO below 1.4 and the low ratio unset, PyTorch's MPS allocator refuses the setting and every GPU operation fails, the self-test's first one too. `LTX25_DEVICE=mps python app.py` then logged a failed self-test at start-up and refused Generate with MpsNumericsError, which sends the user to a newer PyTorch; the pair is the cause. preflight.run now reads the pair before the self-test on mps. Under a pair PyTorch refuses, the self-test is not run, and the start (strict, as every MPS start of the app is) raises memory.MpsWatermarkError after the start-up report, like any other setting that cannot work on this machine; a report-only run logs the problem instead. device.require_mps_numerics checks the pair first as well, whatever LTX25_MPS_SELF_TEST says, so the Generate gate, the pipeline factory and the local runner refuse it in the same words. A start that does not use MPS is not affected. Tests: the F61 tests begun on wip/deploy2-w14 (ca91f07). The app t…e664ad9
3 days ago
-
fix: the Space harness judges what a mode promises over the files a call returns, so a video-with-sound run passes with its mp4 carrying both tracks and the wav beside it judged as the sound on its own
Unverified signatureSince 20155a1 (F30) a mode that exports video and audio returns two files: the mp4 with both tracks, and output.wav, the same sound on its own. The harness still judged every returned file against the mode's whole export list (verify_streams.verify(mode, file) for each file), as when the mp4 came back in both outputs, so the wav "lacked" a video stream. On 2026-11-16 the Space returned m05_text_to_audio_video correctly (h264 704x480, 121 frames, AAC stereo 48 kHz, plus the wav), the harness called it a MISMATCH, and the plan stopped. Every call that returns video with sound would have done the same: modes 05, 06, 07 and 11, both Upscale calls, the enhancer, the model-chosen length and the operator's audio. What each output holds now comes from the registry (returned_files, from the mode's exports). The video output must carry every track the mode exports and is judged as before, strictly, with the geometry. Beside it, a mode that also exports audio returns its sound on its own. That f…a40a311