Spaces:
Running on Zero
Running on Zero
Commit History
2 days ago
-
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
3 days ago
-
fix: everything a user reads after Generate is in the UI's words, and audio too short for the set length is refused in them
Unverified signatureThe notes above Diagnostics now state facts that change what the user got, in plain words. A silent mode says the sound the model made was left out and names the mode that keeps it (video to video names none, since it has none). Audio to audio says the sound is the model's, made from the user's, and how much it was changed; audio to video says the sound is the user's, unchanged. Lengths and the trim are said in seconds. A requested level is mentioned only when it was snapped to a different level, never as "you asked for 0.42 ... 0.42 is the nearest". The upscaled size, the decoder and the prompt sent are the run record's and are no longer repeated as notes. The decoder policy, evictions, and a track an encoder offered that the mode does not declare go to the diagnostics. The upscale phase reads "upscaling ×2", and Upscale off CUDA is refused as "Upscale ×2 needs a CUDA GPU; this machine runs on cpu." A set length longer than the audio is refused before dispatch with the longest lengt…8a7fa0c -
feat: every result carries a structured run record and every failure a category, so the UI renders facts instead of parsing sentences
Unverified signatureGenerationResult.record is a frozen RunRecord of builtins (it crosses the ZeroGPU fork): mode; requested and returned width/height; frames, frame rate and seconds produced; seed; whether stage 2 ran; the decoder that ran; the audio level used; where the length came from (request, model, or the user's audio); the prompt sent after enhancement; adapters as (label, weight, attached modules); worker elapsed seconds on a monotonic clock around the whole run; and peak reserved VRAM where the probe measured it. The notes stay as they were. Every LTXStudioError carries a stable category (quota, timeout, oom, input, adapter, enhancer, bug; constants in errors.py, set per family) and gpu_used, which generate() sets False exactly for a refusal raised before dispatch. errors.categorise(exc) classifies anything reaching the UI: ours by their attributes, torch's CUDA/MPS out-of-memory as oom, the ZeroGPU platform's gr.Errors by their words (quota exceeded / illegal duration as quota, GPU task abort…c08d3e8 -
feat: the decoder is a per-run choice, and auto decodes with the model's diffusion decoder wherever it can run
Unverified signatureSingle-stage runs returned the pipeline's own convolutional VAE decode; the diffusion decoder, which the vendor calls the native default from LTX-2.5 on, only ever ran after stage 2. The decode policy is auto | diffusion | conv, from a new optional request field (GenerationRequest.decoder, appended last), else LTX25_DECODER, default auto: - auto is the diffusion decoder on CUDA when its kernel processor builds and the video is at least the decoder's minimum (224x224x17 for the pinned config); otherwise conv, and the run notes say why. - an explicit diffusion that cannot run raises DecoderUnavailableError (cannot load, or off CUDA beyond the CPU decoder's size limit) or DecoderTooSmallError naming the smallest request; it is never downgraded quietly. - conv stays reachable for A/B; stage 2 always finishes with the diffusion decoder; the modes that discard their video always use conv. A single-stage diffusion decode asks for latents, decodes them with the existing LTX2VideoDi…174bb90 -
fix: stage 2 runs for every mode that offers it, and a video too small for the decoder is refused before stage 1
Unverified signatureThe condition pipeline builds its conditioning mask from height/width (512x768 by default) even when it is handed latents, so every image- and video-conditioned mode failed stage 2 with "Provided latents tensor has shape ...". Stage 2 is now given the upsampled latents' own size (latent H, W x the VAE's spatial ratio) and a frame count read off the latents, which also covers the duration head, where the request has none. text2video keeps upstream's recipe of no size. audio_to_video failed stage 2 with "got multiple values for keyword argument 'noise_scale'": the request already carried noise_scale and audio_latents and the stage-2 kwargs added both. Each call is now built by merging dicts, stage 2's own values winning. The diffusion decoder's neighbourhood windows need a minimum grid (224 px per side and 17 frames for the pinned config). The minimum is computed from the loaded decoder's kernels and checked before stage 1, raising DecoderTooSmallError with the closest request that wor…d2490e1
4 days ago
-
feat: upper bounds on size, length and duration, and a frame count derived from seconds
Unverified signatureWidth and height stop at 1920, num_frames at 481 (20 s at 24 fps, the vendor duration head's own default maximum) and max_seconds at 20. Over a bound is an OverLimitError that names the value, the limit and the nearest legal value, in the style of the existing DimensionError and FrameCountError. No stage-2 or pixel-count ceiling yet: those wait for measurements. Two pieces of frame arithmetic join them, both pure: largest_num_frames_within(seconds, frame_rate) is the longest legal 8k+1 clip that does not outlast a duration (clamped to [1, 481]), for runs whose length an audio input sets; duration_head_frame_range(min, max, fps) is the shortest and longest clip the vendor duration head can return for those bounds, mirroring its clamp, floor and snap. A test checks the range against diffusers' own predict_num_frames where torch is installed. MissingInputError is the typed refusal for a run missing an input its mode requires: it names the mode, each missing input in a creator's words, a…7d49d45
-
feat: prompt enhancement ships — a real toggle, in the web process, at zero GPU cost
Unverified signatureThe build refused this twice, and both reasons were wrong. ⛔ "The pack ships it as absent, so `enable_prompt_enhancement=True` fails on our subset." `model_index.json` does carry `prompt_enhancer: [None, None]`, and that is not what it means. Upstream types the component `Gemma4ForConditionalGeneration | None` (pipeline_ltx2.py:249) and lists it in `_optional_components` (:235) — a plain transformers class. The empty slot is there because the component is CONSTRUCTED, not resolved from the snapshot. So it is an ordinary runtime `from_pretrained` into the HF cache, and nothing in model_files.py / preflight.py / preload_from_hub needs to know it exists. ⛔ "ZeroGPU charges it twice at the packing peak, ~20.6 GB." A placement artefact, not a property of the model. Packing runs once per container, at `demo.launch()`, over tensors registered while the module scope ran. A model loaded lazily inside a request handler, on the CPU, is in neither window. It is charged once, as a download. ✅ An…2f36d8a
-
feat: wire the Hub repo-id direction of LoRA validation — bounded before download, fetched in the web process
Unverified signatureThe route existed as an interface and nothing else: validate_lora carried a fetcher seam (PLAN §4.6's designed entry point), the UI carried a repo-id textbox, and no fetcher was ever written — so every repo id anyone typed rejected with 'cannot be read', including the UI's own placeholder. This is the wiring. lora/fetch.py (new): resolve 'org/name' or 'org/name/file.safetensors' — exactly the two shapes the textbox promises; URLs are refused naming the accepted forms rather than half-parsed past their revision segments. The size cap binds BEFORE any download, from the Hub's file-metadata listing (HfApi.model_info(files_metadata=True)): an oversized adapter raises the existing LoRATooLargeError — whose message has said 'Refusing before download.' since the day it was written — and the download seam is provably never called. A sizeless listing refuses rather than downloading unbounded ('could not check' must not become 'checked and fine'). Downloads go through hf_hub_download's default …b9ee967
-
fix: the pipeline takes PipelineQuantizationConfig, not BitsAndBytesConfig
Unverified signatureThe Space deploy failed on a type, not on disk and not on RAM: ValueError: `quantization_config` must be an instance of `PipelineQuantizationConfig`. A bare BitsAndBytesConfig is what a COMPONENT's from_pretrained takes; DiffusionPipeline.from_pretrained rejects it. Read at the pinned SHA rather than guessed: quant_backend="bitsandbytes_8bit" is a key in both backend mappings, quantizers/auto.py:45 and :57. This is a rung past the one the last test was built for. reached -> correct argument -> ARGUMENT OF THE RIGHT TYPE FOR THE CALLEE. The kwargs test proved the config arrived and could not tell the two classes apart, because both arrive equally well. Bound B is now an ARGUMENT -- components_to_quantize=["transformer"] -- rather than an intent a source scan had to assert. That is strictly better: widening is a change to a value a test reads, not a change to what a call happens to produce. So test_only_the_transformer_is_quantised is retired, with its reason kept where it lived…b55f9af
-
test: deploy drift, ConfigError, zerogpu boundary, conditioning index trap, audio config
Closes the gaps the lead asked about. - Deploy drift: the gradio pin must EQUAL README sdk_version and be exact, the preload_from_hub entry must equal model_files.preload_from_hub_entry() including the revision SHA (without it the preload and the app resolve different cache keys and the Space re-downloads 72 GB), and every requirement needs a floor and a ceiling. - ConfigError: a misconfiguration is not a generation failure, so it must not live in that family — the UI tells an operator and a user different things. - zerogpu: a CUDA tensor anywhere in the returned structure is refused, since the failure mode is a silent hang; size= is built from the installed spaces signature; a duration= callable must match before schedule time. - conditioning: LTX2VideoCondition.index is a LATENT index. Pixel 9 is latent 2, and an out-of-range index raises instead of being silently skipped. - audio: every mel parameter comes from audio_vae.config, and audio.py may not contain a samp…2655c68 -
test: contract the three new LoRA compatibility errors; narrow the filename sweep
The coverage sweep did its job — LoRADimensionError, LoRAInContextError and LoRAModuleNotFoundError arrived without a stated message contract. Each now has one: a dimension error must name module, expected and actual. The filename sweep no longer flags a bare '.safetensors' used as a Gradio upload filter. It describes what a user may upload; no Lightricks rename touches it.b5f94e8 -
test: spec errors, constraints, model_files and the LoRA validator
Red-first specs for PLAN §4.3 (7-stage LoRA validator), §5.2 (single source of truth for Hub coordinates), §5.5 (frame/dimension constraints) and the error hierarchy. Runs with no GPU, no network and no model weights: safetensors fixtures are synthesised byte-for-byte in pure Python, which is also the only way to craft the truncated and corrupt-header files stage 3 has to distinguish.5861856