Align SDK/python to the org ZeroGPU baseline (gradio 5.49.1 / py3.10)

Resolution-aware ZeroGPU duration estimate + budget guard

Adaptive batch size + OOM guard, calibrated ZeroGPU duration, safer slider ceilings

RynnValue-8B demo: reference prefix-uniform pipeline, examples, gradio 5.50

Upload folder using huggingface_hub

Cast value heads to bf16 (fixes mat1/mat2 dtype mismatch)

Prefix-sweep RynnValue demo (official inference.py path) with repo example clips

Remove stale pycache and superseded example asset

RynnValue-8B demo: remaining-time value curve + analysis

Tighten ZeroGPU duration to measured worst case (30s)

Drop unused example asset from superseded app version

Plot predicted value and clip-end reference on separate axes (matches plot_utils.py)

Single-pass RynnValue demo: 5 real example clips + allocator fix for long videos

Remove example assets from superseded app version

Restore verified prefix-uniform RynnValue-8B demo (measured 21s default / 72s max)

Fix allocator fragmentation on long clips (expandable_segments); coerce radio value

Remove leftover example asset from earlier attempt

Pin gradio 5.50.0 + transformers 4.57.6 (huggingface_hub<1 constraint)

RynnValue-8B remaining-time value model demo

Tighten ZeroGPU duration to 100s (measured worst case 72s)

Delete files examples/put_box_in_drawer.mp4 examples/put_box_in_drawer_partial.mp4 with huggingface_hub

Pin gradio 5.50.0 + transformers 4.57.6 (huggingface_hub<1 constraint)

RynnValue-8B ZeroGPU demo

Add reference prefix-uniform scoring mode (default)

Pin gradio 5.50.0 (transformers 4.57 needs huggingface-hub<1)

RynnValue-8B temporal-distance demo

initial commit