Add debug logging and fix Docker venv setup

- Add debug logging to identify which tensor has wrong type in CUDA call
- Fix Dockerfile to use proper venv instead of uv run (avoids package reinstalls)
- Use direct .venv/bin/python execution for faster startup
- This should prevent fused-ssim reinstallation on every launch

Dockerfile

@@ -50,9 +50,10 @@ RUN mkdir -p wheels && \
     wget -O wheels/gsplat-0.1.0-cp310-cp310-linux_x86_64.whl \
     "https://huggingface.co/blanchon/image-gs-models-utils/resolve/main/gsplat-0.1.0-cp310-cp310-linux_x86_64.whl"
 
-# Install dependencies with uv (using pre-built wheel - much faster!)
+# Create virtual environment and install dependencies
 RUN --mount=type=cache,target=/tmp/uv-cache,sharing=locked,uid=1000,gid=1000 \
-    UV_CACHE_DIR=/tmp/uv-cache uv sync --no-dev
+    UV_CACHE_DIR=/tmp/uv-cache uv sync --no-dev && \
+    uv venv .venv
 
 # Copy the rest of the application
 COPY --chown=user . .
@@ -63,5 +64,5 @@ COPY --chown=user pyproject.docker.toml ./pyproject.toml
 # Expose port 7860 (default for HF Spaces)
 EXPOSE 7860
 
-# Launch the Gradio app
-CMD ["uv", "run", "python", "gradio_app.py"]
+# Activate venv and launch the Gradio app directly
+CMD [".venv/bin/python", "gradio_app.py"]

gradio_models.py

@@ -392,12 +392,17 @@ class GradioGaussianSplatting2D(nn.Module):
             xy, scale, rot, int(img_h), int(img_w), tile_bounds
         )
         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)
 
         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)})")
+            
             out_image = rasterize_gaussians_sum(
                 xy,
                 radii,
@@ -476,7 +481,7 @@ class GradioGaussianSplatting2D(nn.Module):
                 xy, scale, rot, int(self.img_h), int(self.img_w), self.tile_bounds
             )
             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)