Studio: add the OpenAI videos API on /v1/videos - #9891

Classify references after reserving the model

When model is omitted or auto-switching is disabled, a direct model load can commit between this status() snapshot and begin_generate(). The reference is then routed according to the old H3 partition: a keyframe-to-Ref2VA change passes it as first_frame, while the reverse passes it in reference_images. Although begin_generate() deliberately reserves and validates against the new state, either case is rejected as unsupported conditioning even though the newly resident partition could accept the same image in the other field. Resolve the conditioning mode from the state actually reserved by begin_generate() rather than this earlier snapshot.

Useful? React with 👍 / 👎.

Fixed in 871b7e8. The route now passes input_reference to begin_generate without deciding between keyframe and Ref2VA conditioning. begin_generate classifies it inside the exact resident-state retry loop, so a partition swap cannot send the image through stale conditioning. Added a deterministic keyframe-to-Ref2VA reservation race test.

Track admitted keyless video requests in the media drain

When keyless API access admits POST /v1/videos without an Authorization header, LlamaKeepWarmMiddleware explicitly clears media_owner, so these new path entries do not increment the VIDEO tracker. However, maybe_auto_switch_media_model() always calls drain() as though the current request were counted, and drain() subtracts one from the tracker. If another authenticated video request has crossed the admission gate but has not yet called begin_generate(), a concurrent keyless auto-switch can therefore discount that other request, evict its resident model, and make it generate against the replacement. Track admitted keyless requests after authentication or pass whether the current request was actually counted into the drain logic.

Useful? React with 👍 / 👎.