Document the warm-up, the pipelined decode and the measured fps

Keep the ZeroGPU worker warm while the tab is open

Lock pacing on to the fastest observed block time and burn down the playback backlog

Overlap each block's VAE decode with the next block's sampling on a second CUDA stream

Only rank-5 decoder weights get the channels_last_3d layout

Run the Wan VAE decoder in channels_last_3d for cuDNN's fast NDHWC kernels

Split VAE decode and device->host copy in the block profile

Pace from the measured steady-state block time; split VAE/copy profiling

Ping /warmup on page load

Convert blocks to uint8 on the GPU; per-block dit/vae timing

Warm up the ZeroGPU worker before first use; per-block profiling; GPU-side uint8, threaded muxer, instant pacing lock-on

Update app.py

Load the AoTI-compiled WanAttentionBlock; document the speed work

UI: always-open image seed above prompt, fix blown-up HUD key chip, trim intro; perf: cache rotated keys, drop per-attention device syncs, hoist cross-attn KV, AoTI block hook

README: document the gradio.Server + WebSocket transport

index.html: Citrus-styled front-end for the gradio.Server transport

transport: swap the Gradio component stream for a gradio.Server + WebSocket backend

Swap the custom component transport for gradio.Server + WebSocket streaming

Document the gradio.Server serving layer

Pin websockets/python-multipart for the WebSocket transport

Serve the live stream from gradio.Server + WebSocket (Citrus-styled page)

README: document /steer

add /steer live-control endpoint + session id input

README: live streaming design

vae: incremental causal decoding

pipeline: stream blocks with a live control hook

Live streaming rollout + Citrus UI

Upload app.py with huggingface_hub

Upload app.py with huggingface_hub

Upload folder using huggingface_hub

initial commit