Fix CUDA tensor type error: radii should be float32, not int32

- Debug output revealed radii=torch.int32 when it should be torch.float32
- Convert radii back to float32 after project_gaussians_2d_scale_rot
- Keep num_tiles_hit as int32 as required
- Apply fix to both main forward() and Gaussian ID visualization
- Based on debug: xy=float32✅, radii=int32❌, conics=float32✅, num_tiles_hit=int32✅, feat=float32✅

gradio_models.py

@@ -393,16 +393,21 @@ class GradioGaussianSplatting2D(nn.Module):
         )
         xy, radii, conics, num_tiles_hit = tmp
 
-        # Ensure num_tiles_hit is int32 tensor (required by CUDA backend)
-        num_tiles_hit = num_tiles_hit.to(dtype=torch.int32)
+        # Ensure correct tensor types for CUDA backend
+        num_tiles_hit = num_tiles_hit.to(dtype=torch.int32)  # Must be int32
+        radii = radii.to(dtype=torch.float32)  # Must be float32
 
         if not self.disable_tiles:
             enable_topk_norm = not self.disable_topk_norm
-            
+
             # Debug: Print tensor types to identify the Float tensor
-            self.worklog.info(f"DEBUG - Tensor types: xy={xy.dtype}, radii={radii.dtype}, conics={conics.dtype}, num_tiles_hit={num_tiles_hit.dtype}, feat={feat.dtype}")
-            self.worklog.info(f"DEBUG - Params: img_h={img_h} ({type(img_h)}), img_w={img_w} ({type(img_w)}), block_h={self.block_h} ({type(self.block_h)}), block_w={self.block_w} ({type(self.block_w)})")
-            
+            self.worklog.info(
+                f"DEBUG - Tensor types: xy={xy.dtype}, radii={radii.dtype}, conics={conics.dtype}, num_tiles_hit={num_tiles_hit.dtype}, feat={feat.dtype}"
+            )
+            self.worklog.info(
+                f"DEBUG - Params: img_h={img_h} ({type(img_h)}), img_w={img_w} ({type(img_w)}), block_h={self.block_h} ({type(self.block_h)}), block_w={self.block_w} ({type(self.block_w)})"
+            )
+
             out_image = rasterize_gaussians_sum(
                 xy,
                 radii,
@@ -482,8 +487,9 @@ class GradioGaussianSplatting2D(nn.Module):
             )
             xy, radii, conics, num_tiles_hit = tmp
 
-            # Ensure num_tiles_hit is int32 tensor (required by CUDA backend)
-            num_tiles_hit = num_tiles_hit.to(dtype=torch.int32)
+            # Ensure correct tensor types for CUDA backend
+            num_tiles_hit = num_tiles_hit.to(dtype=torch.int32)  # Must be int32
+            radii = radii.to(dtype=torch.float32)  # Must be float32
 
             if not self.disable_tiles:
                 enable_topk_norm = not self.disable_topk_norm