Optimize Docker build to avoid OOM during compilation

- Reduce CUDA architectures to essential ones (7.5, 8.0, 8.6) for HF Spaces
- Set MAX_JOBS=1 to limit parallel compilation and reduce memory usage
- This should prevent exit code 137 (killed due to memory limits)
- Covers T4 (7.5), A100 (8.0), and A10G (8.6) GPUs commonly used in HF Spaces

Dockerfile

@@ -46,17 +46,18 @@ ENV HOME=/home/user \
 COPY --chown=user pyproject.docker.toml ./pyproject.toml
 COPY --chown=user gsplat/ ./gsplat/
 
-# Set CUDA compilation flags for HF Spaces GPU compatibility
-ENV TORCH_CUDA_ARCH_LIST="6.0;6.1;7.0;7.5;8.0;8.6;8.9;9.0" \
-    FORCE_CUDA=1
+# Set CUDA compilation flags optimized for HF Spaces (reduce memory usage)
+ENV TORCH_CUDA_ARCH_LIST="7.5;8.0;8.6" \
+    FORCE_CUDA=1 \
+    MAX_JOBS=1
 
 # 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
 
-# Build gsplat wheel for target architecture and reinstall it
+# Build gsplat wheel for target architecture and reinstall it (single-threaded to avoid OOM)
 RUN cd gsplat && \
-    uv run python setup.py bdist_wheel && \
+    MAX_JOBS=1 uv run python setup.py bdist_wheel && \
     mkdir -p ../wheels && \
     cp dist/*.whl ../wheels/ && \
     cd .. && \