Final optimization: use pre-built wheel with correct GPU architectures

- Built gsplat wheel locally with TORCH_CUDA_ARCH_LIST='7.5;8.0;8.6' for HF Spaces
- Uploaded optimized wheel to blanchon/image-gs-models-utils repository
- Simplified Dockerfile to download wheel (no compilation in Docker)
- Use direct .venv/bin/python execution for fastest startup
- This avoids both memory issues and CUDA kernel compatibility problems

Dockerfile

@@ -42,27 +42,18 @@ ENV HOME=/home/user \
     GRADIO_SERVER_PORT=7860 \
     UV_CACHE_DIR=/home/user/.cache/uv
 
-# Copy dependency files and gsplat source
+# Copy dependency files
 COPY --chown=user pyproject.docker.toml ./pyproject.toml
-COPY --chown=user gsplat/ ./gsplat/
 
-# 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
+# Download optimized gsplat wheel (built with HF Spaces GPU architectures)
+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"
 
-# Create virtual environment and install dependencies
+# Install dependencies with optimized wheel
 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 (single-threaded to avoid OOM)
-RUN cd gsplat && \
-    MAX_JOBS=1 uv run python setup.py bdist_wheel && \
-    mkdir -p ../wheels && \
-    cp dist/*.whl ../wheels/ && \
-    cd .. && \
-    uv pip install wheels/gsplat-0.1.0-cp310-cp310-linux_x86_64.whl --force-reinstall
-
 # Copy the rest of the application
 COPY --chown=user . .
 
@@ -72,5 +63,5 @@ COPY --chown=user pyproject.docker.toml ./pyproject.toml
 # Expose port 7860 (default for HF Spaces)
 EXPOSE 7860
 
-# Launch the Gradio app using uv run (but packages already installed, so no reinstall)
-CMD ["uv", "run", "python", "gradio_app.py"]
+# Launch the Gradio app directly with venv python (no uv overhead)
+CMD [".venv/bin/python", "gradio_app.py"]

pyproject.docker.toml

@@ -24,7 +24,7 @@ dependencies = [
 # We use python 3.10 and cu124
 [tool.uv.sources]
 fused_ssim = { git = "https://github.com/rahul-goel/fused-ssim/" }
-gsplat = { path = "gsplat" }
+gsplat = { path = "wheels/gsplat-0.1.0-cp310-cp310-linux_x86_64.whl" }
 torch = [
     { index = "pytorch-cu124", marker = "sys_platform == 'linux'" },
 ]