Use float16 instead of bfloat16 to match reference inference

torch_cluster.fps (called inside the VAE encoder for FPS sub-sampling)
has CUDA kernels for fp16/fp32/fp64 but not bf16, so the bf16 surface
tensor raised NotImplementedError: "_" not implemented for 'BFloat16'.
The reference scripts/inference_detailgen3d.py uses dtype=torch.float16,
which works.

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

app.py

@@ -23,7 +23,7 @@ sys.path.append(os.path.dirname(os.path.abspath(__file__)))
 # Constants
 MAX_SEED = np.iinfo(np.int32).max
 TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
-DTYPE = torch.bfloat16
+DTYPE = torch.float16
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"