diff --git a/containers/qwentts/entrypoint.sh b/containers/qwentts/entrypoint.sh
index 81e5b67..7b7de4f 100644
--- a/containers/qwentts/entrypoint.sh
+++ b/containers/qwentts/entrypoint.sh
@@ -6,9 +6,32 @@ set -e
 
 LOAD="${QWEN_TTS_LOAD:-customvoice,voiceclone}"
 
-echo "=== Qwen3-TTS Server (CPU) ==="
+# On ROCm images, torch.cuda.is_available() caches its result for the life
+# of the process (PyTorch decorates it with @lru_cache). If /dev/kfd isn't
+# fully initialized yet at container start, the very first check can return
+# False and the server gets stuck thinking it's CPU-only forever, even
+# though the GPU becomes visible moments later. Probe in throwaway
+# subprocesses (each gets its own fresh cache) until the GPU is actually
+# ready, so the real uvicorn process's first check succeeds.
+if [ -e /dev/kfd ]; then
+    echo "ROCm device detected, waiting for GPU to become available..."
+    for i in $(seq 1 15); do
+        if python3 -c "import sys, torch; sys.exit(0 if torch.cuda.is_available() else 1)" 2>/dev/null; then
+            echo "GPU is available."
+            break
+        fi
+        sleep 1
+    done
+fi
+
+DEVICE_REPORT="cpu"
+if [ -e /dev/kfd ] && python3 -c "import sys, torch; sys.exit(0 if torch.cuda.is_available() else 1)" 2>/dev/null; then
+    DEVICE_REPORT="rocm (gpu)"
+fi
+
+echo "=== Qwen3-TTS Server ==="
 echo "Loading: $LOAD"
-echo "Device: cpu"
+echo "Device: $DEVICE_REPORT"
 echo
 
 exec uvicorn server:app --host 0.0.0.0 --port 8000 --app-dir /app "$@"
diff --git a/containers/qwentts/qwentts.Containerfile b/containers/qwentts/qwentts.Containerfile
index 6a5cb22..35acf41 100644
--- a/containers/qwentts/qwentts.Containerfile
+++ b/containers/qwentts/qwentts.Containerfile
@@ -21,7 +21,7 @@ EXPOSE 8000
 # sox (system tool, required at runtime by qwen-tts for audio processing)
 # ffmpeg (optional but commonly needed for reference-audio format conversion)
 RUN apt-get update \
-    && apt-get install -y --no-install-recommends sox ffmpeg curl \
+    && apt-get install -y --no-install-recommends sox ffmpeg curl git \
     && apt-get autoremove -y \
     && apt-get clean -y \
     && rm -rf /var/lib/apt/lists/*
@@ -37,6 +37,15 @@ RUN if [ "$DEVICE" = "cpu" ]; then \
 # Install qwen-tts and a minimal API server
 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 \
+        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 && python setup.py install) \
+        || echo "flash-attn build failed or unsupported on this GPU, continuing without it"; \
+        rm -rf /tmp/flash-attention; \
+    fi
+
 # Models are downloaded at runtime into this directory (mount as a volume
 # to persist them across container restarts and avoid re-downloading).
 RUN mkdir -p /app/models
diff --git a/containers/qwentts/qwentts.container b/containers/qwentts/qwentts.container
index e80dda6..caf24bd 100644
--- a/containers/qwentts/qwentts.container
+++ b/containers/qwentts/qwentts.container
@@ -11,7 +11,9 @@ ContainerName=qwentts
 #Memory=8g
 #CPUs=4
 
-# ROCm 
+# ROCm
+# Use Triton flash-attn backend (CK backend targets MI200/300, not RDNA consumer cards)
+Environment=FLASH_ATTENTION_TRITON_AMD_ENABLE=TRUE
 AddDevice=/dev/kfd
 AddDevice=/dev/dri/renderD128
 PodmanArgs=--group-add=keep-groups --ipc=host --pids-limit=-1 --security-opt label=disable