Spaces:
Running on Zero
Running on Zero
Commit History
4 days ago
-
chore: stale comments say what is true now, and three LoRA tests and the test media no longer depend on the machine they run on
Unverified signature- generate.py: the forward mel transform reads only the audio VAE (true since 7cf1547); the VRAM unit no longer names the deleted host scripts. - The stale comments the docs review listed in config.py, pipelines.py, preflight.py, zerogpu.py, tests/verify_streams.py and tests/test_zerogpu.py: the transformers and bitsandbytes pins, the derived int8 packing peak, the 47.375 GiB partition the untiled decode was proven on, the falsified ~68 GB ceiling and the pre-pack 82.2 GB disk reading, the MPS remark, stage-2 parts built on first use, no RAM preflight in preflight.py, blobs that may not be freed rather than 72 GB that never is, and the retired host. - The LoRA message-length bounds count the adapter by its file name, not the temp root's length; the in-use stamp is compared with the clock read just before the load, not a 60 s wall-clock window. - tests/conftest.py puts the session's test media, and the seam test's clip folder, under pytest's base temp, which pytest p…7cb1c38 -
fix: the LoRA cache never evicts an adapter a session has staged or a run is loading, for an hour after its last use
Unverified signatureEviction ran after every fetch and upload with only the new adapter kept, ordered by a stamp that only fetch_lora touched. An adapter a session had staged was evicted before one fetched later and abandoned, so its Generate was refused ("no longer on disk"), and an eviction landing between the worker's gate and load_lora_weights failed the run inside the GPU window. lora.mark_used(path) stamps the cached adapter holding a path as in use (a second marker beside the last-used one; nothing outside a folder this cache stamped is touched). evict_lora_cache never removes an adapter stamped in use within EVICTION_GRACE_SECONDS (one hour: a session between Validate and Generate, a ZeroGPU queue wait and the run, each of which stamps it again), even when that leaves the cache over budget. staged_entry stamps every adapter a session stages, and apply_lora stamps it right before load_lora_weights opens it. A download or an upload copy that nobody staged is still evicted oldest first, as before.b8e93ea -
fix: a Hub refusal says what is known, so a missing repo or a rejected token is no longer reported as gated
Unverified signatureAny HTTP 401 was mapped to "is gated. Accept its terms", checked before the not-found branch. huggingface_hub 1.33 raises RepositoryNotFoundError for an anonymous listing of a missing or private repo (401, X-Error-Code RepoNotFound) and an HfHubHTTPError for an expired or invalid token (401, "Invalid credentials in Authorization header"), so both were told to accept the terms of a gated repo; at 445abcb the first said "not found". Only the Hub's own GatedRepo code (GatedRepoError) is now called gated, at either step. RepoNotFound keeps the not-found-or-private message, and a rejected token says the token is expired, revoked or mistyped. A bare 401 at the listing says the repo may be gated, private or misspelt; a bare 401 at the download, after the same repo listed fine, says it is most likely gated. The tests raise real responses through huggingface_hub's own hf_raise_for_status.4b994b5 -
test: an uploaded adapter is staged as the app's own copy in the LoRA cache, not as Gradio's upload file
Unverified signatureThe upload-direction test pinned staged path == the uploaded file. The UI now copies an upload into the LoRA cache before validating it (Gradio prunes its own upload cache), so the test pins that instead: the staged path is a different, existing file inside the cache with the same bytes; the label still names the upload; the Hub fetcher is never consulted; the same upload validated twice is staged once.15db7d8
5 days ago
-
fix: the LoRA validator checks shape fit with no weights, and tells the truth before a GPU is reserved
Unverified signatureStage 6.5 (does each low-rank factor fit its target module) only ran when a live transformer was passed. The UI never passes one, and Generate skips validation for an entry that already carries a report, so a UI user never got it. It was also wrong for the most common key convention: it stripped diffusion_model. without diffusers' native-to-diffusers rename, so the official LTX-2.5 adapter was told 26 of its 1660 modules were absent, "a normal version difference". - lora/transformer_linear_shapes.json lists the 1660 Linear weights of the pinned transformer (Lightricks/LTX-2.5-Diffusers @871165de), derived from the four shard headers, with that revision and a sha256 of its canonical JSON. `python -m ltx25_studio.lora.manifest` re-derives it from the Hub with header range reads only; an opt-in network test asserts equality. - lora/keys.py resolves adapter keys the way diffusers will: its own _convert_non_diffusers_ltx2_lora_to_diffusers where importable, the same rename as a…9d84791
-
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