Load model synchronously before Gradio launch (fix HF gcTimeout kill)

Fix stage-2 OOM: offload Lyra-2 net to CPU during DA3 GS recon

Refactor: load Lyra-2 model once at startup, reuse across requests

Shorter runtime copy

Update runtime copy to reflect real ~75min/request timing on A100

UI polish: remove attribution blurb below preview

Drop show_download_button kwarg (not supported by gr.Video in 6.13)

Fix preview mp4 (was filtered by .gitignore)

Inline sample-04 output preview for pre-filled input

Lock DMD toggle on (interactive=False) instead of hiding it

Remove cold-start warning banner

Soften macOS viewer copy to 'coming soon' until LyraViewerMac is public

Pre-fill first sample (04 amphitheater) image+caption on page load

README: raise sleep_time to 7200s (2h) — DCP load + inference needs ~1h on A100

Smarter download_checkpoints: skip download if /data/Lyra-2 already populated (direct model mount)

Heartbeat progress during long subprocess runs (keeps SSE stream alive)

Remove show_api kwarg (not supported in gradio 6.13)

Add wget to apt (HF injects OpenVSCode-server install)

Remove TE import from build sanity (triton autotune needs GPU; verify via metadata instead)

Skip GPU-dependent sanity imports at build time (HF builder has no GPU)

Ship pre-built CUDA wheels (TE+TE_torch/VIPE/DA3/gsplat) via Git LFS

Limit uv concurrency (downloads=2, builds=1, installs=2) to reduce peak RAM

MAX_JOBS=1 — MAX_JOBS=4 still OOMs TE compile on HF builder

Lower MAX_JOBS to 4 (OOM during TE compile) + verify .cpath non-empty

Bump tensorstore to 0.1.50 for cp312 wheels (0.1.45 source-build OOMs HF builder)

Rewrite Dockerfile with uv for Python 3.12 + deps

Drop python3.12-distutils (not packaged for 3.12; PEP 632)

Initial Dockerfile + Gradio UI

initial commit