docker gpu arm update

src/paddle_ocr/Dockerfile.build-paddle

@@ -92,8 +92,16 @@ RUN mkdir -p build
 WORKDIR /build/Paddle/build
 
 # Configure CMake for ARM64 + CUDA build
-# Note: Adjust CUDA_ARCH_NAME based on your GPU architecture
-# Common values: Auto, Ampere, Ada, Hopper
+#
+# CUDA_ARCH is auto-detected from host GPU and passed via docker-compose.
+# To detect: nvidia-smi --query-gpu=compute_cap --format=csv,noheader
+# Example: 12.1 -> use "90" (Hopper, closest supported), 9.0 -> use "90"
+#
+# Build time: ~30-60 min with single arch vs 2-4 hours with all archs
+
+ARG CUDA_ARCH=90
+RUN echo "Building for CUDA architecture: sm_${CUDA_ARCH}"
+
 RUN cmake .. \
     -GNinja \
     -DCMAKE_BUILD_TYPE=Release \
@@ -107,7 +115,9 @@ RUN cmake .. \
     -DON_INFER=OFF \
     -DWITH_PYTHON=ON \
     -DWITH_AVX=OFF \
-    -DCUDA_ARCH_NAME=Auto \
+    -DCUDA_ARCH_NAME=Manual \
+    -DCUDA_ARCH_BIN="${CUDA_ARCH}" \
+    -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCH}" \
     -DCMAKE_EXPORT_COMPILE_COMMANDS=ON
 
 # Build PaddlePaddle (this takes 2-4 hours)