Force SDPA in GS-LRM attention (xformers unusable on Blackwell)

Fix real OOM: 4-D SDPA layout; cu130 is ZeroGPU-native, drop cu128 cruft

Make pip nvcc detection robust (glob, not hardcoded path)

Revert cu130 forcing; use ZeroGPU-native cu128 torch + pip cu128 nvcc

Install cu130 CUDA runtime libs (drop --no-deps) to fix spurious OOM

Add GPU memory/build diagnostic before inference

Use SDPA in MVAttnProcessor to fix OOM on multi-view attention

Disable expandable_segments instead of the whole caching allocator

Disable CUDA caching allocator to dodge ZeroGPU NVML assert

Disable Gradio SSR mode to fix ZeroGPU NVML allocator assert

Drop enable_xformers call; use PyTorch SDPA on Blackwell GPU

Move torch stack to 2.11.0 to resolve xformers conflict

Force cu130 torch at startup before importing torch

Make cu130 the primary pip index, keep torch pin bare

Pin torch/torchvision to explicit +cu130 local version

Move torch stack to cu130 to match ZeroGPU CUDA 13.0 toolkit

Use --no-build-isolation for diff-gaussian-rasterization

Pin sdk_version to 5.49.1 to match gradio requirement

Bump torch to 2.8.0 for ZeroGPU compatibility

Update README.md

change readme

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

demo

Change thumnail

Update demo

Update demo

Update demo

Update demo