[Waiting for internal sync]Support reasoner video input.#25

Open

foreverlms wants to merge 21 commits into

mainfrom

maoshengl/video_reasoner_inference

Commits on Jun 8, 2026

Add design spec: video input for reasoner model-mode inference

foreverlms
and
claude
committed
Add implementation plan: video input for reasoner model-mode

foreverlms
and
claude
committed
feat(reasoner): add video_* sampling fields + mutual-exclusion validation

foreverlms
and
claude
committed
refactor(reasoner): video_fps PositiveFloat + construction-time mutual-exclusion validator

foreverlms
and
claude
committed
feat(reasoner): add video_* defaults (null) to reasoner sample_args
foreverlms
committed
feat(reasoner): video branch in prepare_multimodal_reasoner_inputs
foreverlms
committed
feat(reasoner): accept video tensors in _impl_generate_reasoner_text

foreverlms
and
claude
committed
feat(reasoner): forward video tensors through generate_reasoner_text pass-throughs

foreverlms
and
claude
committed
fix(reasoner): revert out-of-scope param additions to Nemotron generate_reasoner_text stub

foreverlms
and
claude
committed
feat(reasoner): videos param + video chat block in OmniMoTModel.generate_reasoner_text

foreverlms
and
claude
committed
docs(reasoner): update generate_reasoner_text docstring for video path
foreverlms
committed
feat(reasoner): route mp4 vision_path to video conditioning in inference engine

foreverlms
and
claude
committed
docs(reasoner): document video input + add reasoner_video example

foreverlms
and
claude
committed
docs(reasoner): clarify vision_path comment covers video too

foreverlms
and
claude
committed
fix(reasoner): decode video frames for Qwen3VLProcessor; reduce knobs to video_fps
foreverlms
committed
fix(reasoner): decode video via torchvision.io + smart_nframes (drop decord/pkl_to_media dep)

foreverlms
and
claude
committed
fix(reasoner): emit reasoner_videos uniformly per-sample; mark design/plan superseded

foreverlms
and
claude
committed
chore(reasoner): untrack video-reasoner spec/plan docs (keep in-repo, untracked)

foreverlms
and
claude
committed
docs(reasoner): regenerate inference.md TOC for Reasoner section (pre-commit)

foreverlms
and
claude
committed

Commits on Jun 9, 2026

test(reasoner): cover video modality in get_sample_data tests; account for reasoner_videos key

foreverlms
and
claude
committed

Commits on Jun 11, 2026

Merge branch 'main' into maoshengl/video_reasoner_inference
foreverlms
authored