Minor - Fix CUDA OOM for 16GB VRAM GPUs

[changhaowuwu]

Fix CUDA OOM for 16GB GPUs
Problem
telestylevideo_inference.py consistently hit torch.OutOfMemoryError on GPUs with 16GB VRAM (e.g. RTX 4080). Two issues:

1.Both models on GPU simultaneously — The VAE (3.5 GB) and transformer (3.5 GB) were both loaded to GPU at init, leaving insufficient headroom for activation tensors during the transformer forward pass (340 MiB for patch_embedding2) and VAE encoding (8.7 GiB for 129-frame 3D convolutions).
2.Stale GPU processes — Interrupted runs (Ctrl+C, OOM kills → exit code 137) left zombie processes holding 11+ GiB of VRAM, starving subsequent runs.

Solution
Model offloading — Only one model resides on GPU at any time:

Additional optimizations:

• VAE tiling (enable_tiling()) — tiles 3D convolutions spatially so encoding 129 frames doesn't require one massive allocation
• Latent-only pipeline output — pipeline returns raw latents (output_type="latent") so VAE decode happens after the transformer is offloaded
• Automatic stale process cleanup — _kill_stale_gpu_processes() queries nvidia-smi at startup and kills any leftover GPU processes
• Graceful exit handlers — atexit + SIGINT/SIGTERM handlers call torch.cuda.empty_cache()
• PYTORCH_ALLOC_CONF=expandable_segments:True set programmatically to reduce fragmentation

Testing
Verified end-to-end on a 16GB GPU with the default 129-frame, 720×1248 configuration. All 25 diffusion steps complete at ~77s/step with peak VRAM usage under 12 GiB.