viggle-mine-workflow / VALIDATION.md
ZIP Space Deployer
Deploy ZIP — 2026-10-14 15:17 UTC
18bfeb9
|
Raw History Blame Contribute Delete
11 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade

Validation record · 2026-10-14

Status: consolidated implementation candidate; not a claim of complete live ZeroGPU or visual parity.

Latest user evidence: Standard completed successfully on the experimental Space; its e.data[0] error followed premature Stop. Workflow's missing OAuth scope warning is addressed in the README. TAP's 270-second request is consistent with the SDK's Blackwell multiplier on the unchanged 180-second default. See docs/OAUTH_AND_GPU_DURATION.md.

Startup repair: the deployed HTTPS root redirected to an absolute HTTP URL. The updated middleware uses a relative redirect; see docs/STARTUP_FIX.md. This fixes a confirmed HTTPS landing defect locally. The reported component KeyError: 142 has not been reproduced or proven resolved.

Pass A integration repair: docs/PASS_A_STALL.md. The corrected bridge preserves the actual queued Gradio request/event/context, bounds waiting and rejects overlapping jobs after cancellation or timeout. Real queued CPU tests prove that cancelled results no longer commit a take. Duplicate Gradio route/middleware construction and canvas load registration are also corrected. These are verified integration defects; the exact cause of the reported GPU worker stall is not established and successful Pass A/B ZeroGPU inference has not been demonstrated with this build.

Evidence produced in this run

Check Result What it establishes / does not establish
Canonical file hashes PASS Backend, TAP, SAM, camera, media, planning, full original app, requirements/packages, handoff and license/notice match the attached V22 bytes
Canonical orchestration extraction PASS v22_core.py exactly equals the canonical prefix before TRANSLATIONS; runtime adapter replaces only its UI/thread bridge, preserving the decorated GPU function
Native 6.30.0 graph parser and topology PASS Schema loads, is acyclic, ports/edge types agree, bind signatures match, output indices align
Graph structure PASS 20 nodes / 20 edges; three media references; Standard has only source/reference ancestors; B does not consume final A; no direct operator→operator edge
Coordinate/artifact contracts PASS Prepared IDs, dimensions, frames/times and finite in-bounds points checked; wrong kinds/spaces rejected; A stabilized output handed to B
CPU algorithms PASS Real FFmpeg source normalization, 48-frame clip, anchor indices/timestamps, real V22 camera solve/stabilized encoding/reference crop from synthetic tracks, exact stabilized Recovery frame extraction
Adapter orchestration/cache PASS with spies Saved points passed once to tracker; A/B called separately; identical takes replayed; changed completed takes rejected; B settings inheritance/overrides; stale Recovery rejected. Spies do not generate video or validate a GPU
Native HTTP application PASS Real FastAPI/Gradio config, bound endpoint registration, picker/legacy/Recovery routes, media serving, canvas landing and cookie
Native queued REST round trip PASS Actual local Gradio 6.30.0 server, media upload, queued Prepare Source with completed SSE response, three served anchors, saved/reloaded points
Pytest 49 tests + 6 subtests passed Three expected Gradio beta warnings; no model inference; context propagation, timeout isolation, cancellation, startup and reference-A forwarding included
Queued cancellation/error lifecycle PASS Real Gradio client/server: request/event/blocks reach thread; exception reaches caller; cancellation ends coordinator, protects active worker and never caches late result; retry succeeds after worker exits
GPU registration/startup PASS on CPU Registration branch retains one set of canvas load events; one CORS/Brotli layer; configured component IDs resolve. No real GPU decorator was run in this construction test
GPU decorator / fusion order PASS by source/hash inspection Original allocation boundaries and merge-before-inverse code preserved; not evidence that the scheduler allocated successfully

Machine-readable evidence is in docs/validation/. Tests include original canonical planning/camera/SAM-guard CPU checks. The CPU loader selects unchanged canonical CPU function ASTs and makes every GPU entry point raise an explicit “CPU preview only” error. It never substitutes generated-looking dummy inference outputs into the demo.

The local UI/HTTP environment used Python 3.12, Gradio 6.30.0 and huggingface-hub 1.29.0. The full pinned torch/transformers/Diffusers inference stack and large model weights were not installed or loaded locally. A successful CPU server is therefore not a successful production model boot.

Blocked / not performed

Gate Actual status
HF Job dependency/import check The connected Hugging Face account was available, but Jobs inspect, ps, and an attempted run returned UNAVAILABLE: MCP tool hf_jobs was not returned by tools/list. No Job was launched or billed by this work.
Previous ALL_IMPORTS_OK Job The user reported success at 6acf5eaf5ccb47ba0f799bc5. The tool failure prevented inspecting its logs here. This is inherited evidence, not a newly verified result.
Production eager model startup Not run with this candidate's exact deployment image. Canonical pre-fork strategy is unchanged and statically checked.
TAP / SAM on actual ZeroGPU Previously validated by the user for the experimental baseline. Not rerun through this candidate.
Standard / Pass A / Pass B actual ZeroGPU User reports a successful Standard run. Pass A/B remain unverified; no full parity, output-quality, quota or cancellation claim is made.
Actual rendered candidate graph Partial inspection: the deployed direct ?ui=canvas URL rendered the empty graph at 100% and 32% fit-to-view in the browser. The corrected build's embedded HF page, populated cards and full visual acceptance remain unverified.
Native Markdown previews / clickable link rendering Server-generated markup and routes checked; actual browser rendering remains unverified.
New picker touch interaction / mobile canvas Not exercised in a browser. The old working standalone picker is retained byte-for-byte. New page/bridge has CPU HTTP and validation checks only.
EN/FR parity English canvas labels only. Canonical EN/FR Blocks UI preserved in v22_app.py, not integrated into native Workflow.

Native UI limitations, including the full-pipeline App/global Run behavior and required explicit refresh after external selections, are described in docs/WORKFLOW_API.md. The graph has a designed initial arrangement, but visual excellence and minimal wire crossings remain unconfirmed. This package must not be described as a visually validated final release.

Reproducible image check after deployment

The archive includes tools/check_deployed_image.py. It asserts the exact pinned dependency versions, Diffusers VCS revision d6726f3, the requested five imports, canonical hashes and this candidate's file manifest. It prints ALL_IMPORTS_OK only after the real imports succeed. --server adds the real CPU Gradio/HTTP smoke test. It does not import the large V22 backend, download its models, or emulate ZeroGPU.

After the new Space build has completed, from an authenticated HF CLI environment:

hf jobs run --flavor cpu-basic --timeout 20m \
  hf.co/spaces/fffiloni/viggle-mine-workflow \
  python /home/user/app/tools/check_deployed_image.py --server

The path shown is the normal Gradio Space app directory; if the deployed image uses a different working directory, use its actual path. A Job always sees the currently built Space image. Before deployment, this file will not exist there, and testing the old image cannot validate the unpublished ZIP. If /home/user/app has changed after the build or the owner saved an edited workflow, the release file-manifest comparison intentionally reports that difference.

A regular HF GPU Job can test CUDA imports/inference but does not establish successful dynamic ZeroGPU scheduler allocation. The existing Space itself is the final environment for that gate.

Compact acceptance matrix for the one candidate deployment

Scenario Required observation
Fresh desktop viewer Default canvas; main settings visible, advanced settings collapsed; all 20 cards positioned coherently; no overlapping expanded cards or wires hiding actions
Desktop at 30–40% zoom Execution order and optional Recovery/independent Standard readable; inspect long Reference A fan-out specifically; correct layout before declaring visual approval
Mobile Open Point Picker from result, session already loaded, tap 3–12 points on all anchors, undo/clear/save, return and run Track without copying JSON; no queue/SSE picker failure
Prepare → TAP Original 24-fps dimensions/timestamps/normalized pixel points; one large tracker job; preview follows the intended subject
SAM off / on Off performs no SAM job; on one large job and guarded camera output; no mask used as final compositing matte
Camera Source-side Reference A transformed at the chosen original-source timestamp; stable input/frame count and camera limits match V22
Standard Run directly from source/reference with Smart stages unexecuted; output and audio match V22 for identical settings
A forward-only and backward option All technical windows/branches logged in one xlarge allocation per A; final and stabilized-generated intermediate both exist with correct frame counts
Recovery Candidate extracted from stabilized input, uploaded B preserves geometry; B logs a separate xlarge allocation; no A allocation repeats; B input is stabilized A generated output, not recomposed final
B fusion Long backward computation independent from short visible transition; A/B fused before exactly one B-final inverse; compare identity/camera/audio/frame-count behavior to V22
Cache / edits Re-run unchanged A returns saved take; changed A settings with same take fail; increment Take deliberately regenerates; changed points require Track rerun; changed B selection requires Prepare Recovery rerun
Refresh / errors Same browser keeps completed take protection; expired state gives actionable error; quota errors/failed jobs do not create false completed checkpoints

The Recovery browser enforces the existing V22 generation constraints before extraction: at least one second of backward context, and a frame before the last frame. Clips shorter than 26 frames disable candidate extraction. This prevents selecting an endpoint frame that V22 would internally clamp to a different generation anchor; the inference code itself is unchanged. HTTP tests reject these invalid endpoints.