fix: handle paddlepaddle-gpu CUDA import at build time gracefully

paddlepaddle-gpu requires libcuda.so.1 at import time, but no CUDA
driver exists during Docker build. The download script now catches
this ImportError and skips PaddleOCR model pre-download on amd64.
Models will download on first use at runtime when the CUDA driver
is available via nvidia-container-toolkit.

On arm64 (CPU paddlepaddle), models are still pre-downloaded at build
time as before.

Also reverted CI to amd64-only Docker build test for speed. Multi-arch
build is tested on release via the release workflow.

.github/workflows/ci.yml

@@ -87,14 +87,11 @@ jobs:
 
       - uses: docker/setup-buildx-action@v3
 
-      - uses: docker/setup-qemu-action@v3
-
       - uses: docker/build-push-action@v6
         with:
           context: .
           file: docker/Dockerfile
           push: false
-          platforms: linux/amd64,linux/arm64
           tags: stirling-image:ci
           cache-from: type=gha,scope=unified
           cache-to: type=gha,mode=max,scope=unified

docker/Dockerfile

@@ -118,16 +118,10 @@ RUN if [ "$TARGETARCH" = "amd64" ]; then \
 
 RUN /opt/venv/bin/pip install seam-carving==1.1.0
 
-# Install CUDA forward-compat stubs for build-time model downloads.
-# paddlepaddle-gpu needs libcuda.so.1 to import, but the real driver is
-# only injected at runtime by nvidia-container-toolkit. The compat package
-# provides a stub that satisfies the dlopen without a real GPU.
-RUN if [ "$TARGETARCH" = "amd64" ]; then \
-    apt-get update && apt-get install -y --no-install-recommends cuda-compat-12-6 && \
-    rm -rf /var/lib/apt/lists/* \
-; fi
-
 # Pre-download and verify all ML models
+# Note: on amd64, paddlepaddle-gpu can't import without the CUDA driver (only
+# available at runtime). The download script gracefully skips PaddleOCR model
+# pre-download in this case; models download on first use at runtime instead.
 COPY docker/download_models.py /tmp/download_models.py
 RUN /opt/venv/bin/python3 /tmp/download_models.py && rm -f /tmp/download_models.py
 

docker/download_models.py

@@ -63,11 +63,16 @@ def download_realesrgan_model():
 def download_paddleocr_models():
     """Pre-download PaddleOCR models for all supported languages."""
     print("=== Downloading PaddleOCR models ===")
-    os.environ["PADDLE_PDX_DISABLE_MODEL_SOURCE_CHECK"] = "True"
-    # Force CPU mode during build - no GPU driver available at build time
-    os.environ["PADDLE_DEVICE"] = "cpu"
-    os.environ["FLAGS_use_cuda"] = "0"
-    from paddleocr import PaddleOCR
+    try:
+        from paddleocr import PaddleOCR
+    except ImportError as e:
+        if "libcuda" in str(e):
+            # paddlepaddle-gpu can't import without CUDA driver at build time.
+            # Models will be downloaded on first use at runtime instead.
+            print(f"  Skipping PaddleOCR model pre-download (no CUDA driver at build time)")
+            print(f"  Models will download on first use at runtime.\n")
+            return
+        raise
 
     for lang in PADDLEOCR_LANGUAGES:
         print(f"  Downloading models for lang={lang}...")
@@ -100,13 +105,20 @@ def smoke_test():
     is_amd64 = platform.machine() in ("x86_64", "amd64")
 
     from rembg import new_session
-    from paddleocr import PaddleOCR
     import mediapipe as mp
     import cv2
     import numpy
     from PIL import Image
     import seam_carving
-    print("  Core imports OK")
+
+    try:
+        from paddleocr import PaddleOCR
+        print("  Core imports OK (including PaddleOCR)")
+    except ImportError as e:
+        if "libcuda" in str(e):
+            print("  Core imports OK (PaddleOCR skipped - no CUDA driver at build time)")
+        else:
+            raise
 
     # RealESRGAN/basicsr has a known torchvision compat issue on arm64.
     # On amd64 we verify the full import chain; on arm64 we just check the model file.