Update app.py

examples: cache_examples=True, cache_mode=lazy; infer() args get defaults

cap queue at 160 frames (matches num_frames slider max 161 = 8*20+1)

simplify intro description

drop 3D preview; bridge Sana tqdm to gr.Progress; fix examples table

auto-sync num_frames to queue total (snapped to 8k+1)

examples: populate action_queue (the real input) instead of dsl_view

disable track_tqdm to avoid ZeroGPU subprocess iterable mismatch

build SanaWMPipeline at module level for ZeroGPU CUDA emulation contract

Update README.md

add scipy + pytz (top-level imports in action_overlay / logger)

add qwen-vl-utils (transitively imported by diffusion.model.builder)

stub flash_attn_2_cuda so flash-attn (no CUDA build) imports cleanly

install flash-attn>=2.7.0 at startup with FLASH_ATTENTION_SKIP_CUDA_BUILD=TRUE

give flash_attn stub a real ModuleSpec so diffusers find_spec works

import spaces before any CUDA-touching module

actually remove mmcv from requirements.txt

install mmcv at app startup; --no-build-isolation

Update requirements.txt

initial: SANA-WM ZeroGPU demo + vendored Sana (feat/sana-wm)

initial commit