Build nvdiffrast from source at startup (fixes broken NVDR_CHECK_CPU in prebuilt Blackwell wheel); revert debug patch

app.py

@@ -19,6 +19,14 @@ subprocess.run("pip install spandrel==0.4.1 --no-deps", shell=True, check=True)
 # hides torch from diso's setup.py). FORCE_CUDA/TORCH_CUDA_ARCH_LIST are set
 # as Space variables since no GPU is visible in this (non-@spaces.GPU) process.
 subprocess.run("pip install diso==0.1.4 --no-build-isolation", shell=True, check=True)
+# The prebuilt Blackwell nvdiffrast 0.4.0 wheel has a broken NVDR_CHECK_CPU
+# device check (rejects tensors regardless of actual device - ABI skew vs.
+# the torch it's dynamically linked against). Build from source instead so
+# it's compiled against the exact torch/ABI installed in this env.
+subprocess.run(
+    "pip install git+https://github.com/NVlabs/nvdiffrast.git --no-build-isolation",
+    shell=True, check=True,
+)
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 DTYPE = torch.float16
@@ -86,19 +94,6 @@ NUM_VIEWS = 6
 from inference_ig2mv_sdxl import prepare_pipeline, preprocess_image, remove_bg
 from mvadapter.utils import get_orthogonal_camera, tensor_to_image, make_image_grid
 from mvadapter.utils.render import NVDiffRastContextWrapper, load_mesh, render
-# nvdiffrast's rasterize() defaults `ranges` to an empty CPU tensor when None,
-# but the Blackwell-built nvdiffrast 0.4.0 wheel's rasterize_fwd_cuda rejects
-# whatever it actually receives on this path. Force an explicit CPU tensor at
-# the wrapper boundary so the C++ device check always sees what it expects.
-_orig_ctx_rasterize = NVDiffRastContextWrapper.rasterize
-def _cpu_ranges_rasterize(self, pos, tri, resolution, ranges=None, grad_db=True):
-    if ranges is None:
-        ranges = torch.empty((0, 2), dtype=torch.int32, device=pos.device)
-    print(f"[rasterize debug] pos={pos.device},{pos.dtype},{tuple(pos.shape)} "
-          f"tri={tri.device},{tri.dtype},{tuple(tri.shape)} "
-          f"ranges={ranges.device},{ranges.dtype},{tuple(ranges.shape)}")
-    return _orig_ctx_rasterize(self, pos, tri, resolution, ranges=ranges, grad_db=grad_db)
-NVDiffRastContextWrapper.rasterize = _cpu_ranges_rasterize
 mv_adapter_pipe = prepare_pipeline(
     base_model="stabilityai/stable-diffusion-xl-base-1.0",
     vae_model="madebyollin/sdxl-vae-fp16-fix",