Match reference inference dtype handling, drop autocast

Reference script (scripts/inference_detailgen3d.py) casts surface to
model dtype and runs inference without autocast. app.py was leaving
surface as float32 while the pipeline is bfloat16, then wrapping the
call in @torch .autocast(device_type=cuda) which defaults to fp16 — the
fp32/bf16/fp16 mismatch surfaces as torch.AcceleratorError on the GPU
matmul. Mirror the reference: cast surface to (DEVICE, DTYPE) and drop
the autocast decorators.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

app.py

@@ -97,12 +97,11 @@ def load_mesh(mesh_path, num_pc=20480):
     ind = rng.choice(surface.shape[0], num_pc, replace=False)
     surface = torch.FloatTensor(surface[ind])
     normal = torch.FloatTensor(normal[ind])
-    surface = torch.cat([surface, normal], dim=-1).unsqueeze(0).cuda()
+    surface = torch.cat([surface, normal], dim=-1).unsqueeze(0).to(DEVICE, dtype=DTYPE)
 
     return surface
 
 @torch.no_grad()
-@torch.autocast(device_type=DEVICE)
 def run_detailgen3d(
     pipeline,
     image,
@@ -140,7 +139,6 @@ def run_detailgen3d(
 
 @spaces.GPU(duration=180)
 @torch.no_grad()
-@torch.autocast(device_type=DEVICE)
 def run_refinement(
     rgb_image: Any,
     mesh: Any,