Clearer Tip + duration slider 2-10 with quota warning

Patch wan_vae.wrapped_encode + wrapped_decode to cast inputs to bf16 (fix I2V/timbre crash)

Default inference steps 30 → 25 (fits within 330s ZeroGPU duration)

Add PyAV (av==16.0.1, matches wanx) for write_video h264+aac muxing

Move wan_vae/audio_vae nn.Module to cuda explicitly + decode-stage timing logs

Fix VAE dtype (fp32) + torch.compile mode=default + SDPA flash priority

Fix VAE dtype mismatch + torch.compile DiT backbone for 1.3x speedup

gradio 6.x: queue(api_open=False) replaces launch(show_api=False)

Bump sdk_version to 6.16.0 (matches local wanx env); drop bool-schema monkey-patch

Monkey-patch gradio_client.utils to handle additionalProperties=true (bool)

Pin gradio_client==1.3.0 — fixes /api_info bool/dict bug under gradio 4.44.0

launch(show_api=False) — sidestep gradio 4.44.0 /api_info schema bug

Resolve nava_src/configs/model/*.json to absolute paths from site-packages

Drop gradio/spaces pins — Spaces injects gradio==4.44.0 & spaces==0.50.4 at build

Pin all versions to local wanx env (gradio 6.16, hub 0.35.3, flash-attn 2.8.3)

Pin torchvision/torchaudio to 2.8 family (avoid cu13 mismatch)

Fix flash-attn wheel URL: cu12torch2.8 (was cu128torch2.8 typo)

Fix Space frontmatter: shorten short_description, drop invalid hardware fields

Initial Space: NAVA on ZeroGPU (FP8 + VL caption + rewrite + retry)

initial commit