Auto-detect best GPU (RTX 5060 preferred) + build CuPy from source for sm_120

Dockerfile

@@ -1,4 +1,4 @@
-FROM nvidia/cuda:11.8.0-devel-ubuntu22.04
+FROM nvidia/cuda:12.4.0-devel-ubuntu22.04
 
 ENV DEBIAN_FRONTEND=noninteractive
 ENV TZ=Europe/Paris
@@ -45,15 +45,15 @@ RUN pip3 install --no-cache-dir \
     pillow-avif-plugin \
     cmcrameri
 
-# Install CuPy for GPU acceleration (optional - will fallback to numpy if not available)
-# We use CuPy 13.4 (CUDA 11.x wheel) because CuPy 14.x dropped JIT compilation
-# support.  With CUPY_CUDA_COMPILE_WITH_CACHE=1, CuPy 13.4 compiles kernels at
-# runtime for GPU architectures not in the pre-built wheel (e.g. sm_89 / RTX 4060 Ti).
-# The devel image includes nvcc, required for JIT compilation.
-# nvcc stays in PATH for the 'lidar' user.
-ENV CUPY_CUDA_COMPILE_WITH_CACHE=1
-ENV PATH=/usr/local/cuda/bin:${PATH}
-RUN pip3 install --no-cache-dir cupy-cuda11x==13.4.0 || echo "CuPy not available - GPU acceleration disabled"
+# Build CuPy from source with nvcc, targeting sm_120 (RTX 5060) and nearby archs.
+# Pre-built wheels (cupy-cuda12x 14.x) don't include sm_120, so we compile.
+# This step takes ~30 min the first time; the image is cached after that.
+RUN apt-get update && apt-get install -y --no-install-recommends git && \
+    git clone --depth 1 --branch v14.0.0 https://github.com/cupy/cupy.git /tmp/cupy-src && \
+    cd /tmp/cupy-src && \
+    CUPY_NVCC_GENERATE_CODE='sm_89;sm_90;sm_120' \
+    pip3 install --no-cache-dir -e . && \
+    rm -rf /tmp/cupy-src
 
 # Copy and install the pipeline package
 COPY setup.py .