Refit the ZeroGPU duration estimate to measured AOTI timings (was reserving ~2x)

Keep the prompt-enhancer token cap lean (reverts the quota-inflating 2048 default)

Keep theme/css on launch() (Gradio 6) and add the Citrus dark-mode text-colour fix

Raise PE token cap to 2048; move theme/css onto Blocks + Citrus dark-mode fix

Trim the logo block's padding

Cap prompt enhancement at 1024 new tokens

Update app.py

Update app.py

README: AoTI kernels, tiling and the KV-cache budget

Turn the prefix KV cache off when it would not fit next to the resident weights

AoTI kernels for the cached decode blocks and the VAE decoder; NCII guard in a sidecar subprocess

Forward the live gr.Progress into the GPU function so the step counter works

Tile VAE decode at high res; tighten GPU budgets; non-thinking PE

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

initial commit