fix(qwentts): fix GPU detection race, make ROCm flash-attn build opt-in

containers/qwentts/build-container.sh

@@ -13,10 +13,19 @@ case "$DEVICE" in
         ;;
 esac
 
+BUILD_FLASH_ATTN=false
+if [ "$DEVICE" = "rocm" ]; then
+    read -r -p "Compilare anche flash-attn da sorgente? Richiede molto tempo (decine di minuti). [y/N] " reply
+    case "$reply" in
+        [yY]|[yY][eE][sS]) BUILD_FLASH_ATTN=true ;;
+    esac
+fi
+
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
 podman build \
     --build-arg DEVICE="$DEVICE" \
+    --build-arg BUILD_FLASH_ATTN="$BUILD_FLASH_ATTN" \
     -t "localhost/qwentts:$DEVICE" \
     -t localhost/qwentts:latest \
     -f "$SCRIPT_DIR/qwentts.Containerfile" \

containers/qwentts/qwentts.Containerfile

@@ -6,6 +6,7 @@
 ### BUILD (rocm): ./build-container.sh rocm
 ### RUN:   podman run --rm -p 8000:8000 -v /path/to/models:/app/models qwentts:latest
 ARG DEVICE=cpu
+ARG BUILD_FLASH_ATTN=false
 ARG BASE_IMAGE_CPU=python:3.12-slim
 ARG BASE_IMAGE_ROCM=docker.io/rocm/pytorch-nightly
 
@@ -14,6 +15,7 @@ FROM ${BASE_IMAGE_ROCM} AS base-rocm
 
 FROM base-${DEVICE}
 ARG DEVICE
+ARG BUILD_FLASH_ATTN
 
 USER root
 EXPOSE 8000
@@ -38,7 +40,7 @@ RUN if [ "$DEVICE" = "cpu" ]; then \
 RUN pip install --no-cache-dir qwen-tts fastapi uvicorn python-multipart
 
 # Optional flash-attn for ROCm, built from source (PyPI wheel is CUDA-only); best-effort, skipped on failure
-RUN if [ "$DEVICE" = "rocm" ]; then \
+RUN if [ "$DEVICE" = "rocm" ] && [ "$BUILD_FLASH_ATTN" = "true" ]; then \
         pip install --no-cache-dir ninja \
         && git clone --depth 1 https://github.com/Dao-AILab/flash-attention.git /tmp/flash-attention \
         && (cd /tmp/flash-attention && GPU_ARCHS="gfx1150;gfx1151" python setup.py install) \