fix(qwentts): fix GPU detection race and add ROCm flash-attn support

2026-09-12 21:24:56 +02:00
parent 0b202e6409
commit 90ff24a041
1 changed files with 1 additions and 1 deletions
@@ -41,7 +41,7 @@ RUN pip install --no-cache-dir qwen-tts fastapi uvicorn python-multipart
 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) \
+        && (cd /tmp/flash-attention && GPU_ARCHS="gfx1032;gfx1150" python setup.py install) \
        || echo "flash-attn build failed or unsupported on this GPU, continuing without it"; \
        rm -rf /tmp/flash-attention; \
    fi