fix: rewrite wrapper patch to download fresh file before patching

fix: revert to cudaMallocAsync to avoid NVML assertion in ZeroGPU

fix: upgrade stale from_config patch to bfloat16, fix allocator fragmentation

Fix: force bfloat16 on from_config to halve generator memory

Fix OOM: reduce temporal frames 32→16 to halve KV cache footprint

Fix: use cudaMallocAsync backend to avoid NVML calls in ZeroGPU

Fix: avoid CUDACachingAllocator NVML assert; enable streaming VAE

Fix: trailing comment was eating .to(device) in linalg.solve patch

Fix: move torch.linalg.solve to CPU to avoid cuSOLVER failure in ZeroGPU

Fix: replace broken attention.py patch with flash_attn stub module

Fix: SDP fallback for flash_attention when flash_attn not installed

Fix: pass prompt file path to data_path, not directory

Fix: use actual checkpoint filenames generator.pt and lora.pt

Fix: download Wan2.2-TI2V-5B config+encoders, patch from_pretrained→from_config

Fix: patch x_clip_loss import removido en transformers recientes

Simplify requirements.txt: remove heavy/optional deps (onnxruntime-gpu, CLIP, av, torchao)

Refactor: sdk: gradio + requirements.txt, elimina Dockerfile innecesario

Fix: torch==2.11.0 (2.8.0 no existe para cu128), flash-attn prebuilt

Initial deployment: LongLive-2.0-5B on ZeroGPU