FlashAttention2 not available in hipblas docker images

LocalAI version:

LocalAI v4.0.0 (8e8b7df715e620b64d07cddf6b73bff8f966dac5)

Environment, CPU architecture, OS, and Version:

AMD Strix Halo 395+
128Gb Unified Memory (96 VRAM)
Fedora 42

Describe the bug

I am unable to load either qwen3-tts or qwen3-asr with errors saying FlashAttention2 is not found.

To Reproduce

Docker Compose file:

  localai:
    container_name: localai
    image: localai/localai:latest-gpu-hipblas
    environment:
      DEBUG=true
      MODELS_PATH: /models
    ports:
      - "8080:8080"
    volumes:
      - ./backends:/backends
      - ./config:/config
      - ./models:/models
    devices:
      - /dev/dri:/dev/dri
      - /dev/kfd:/dev/kfd
    group_add:
      - "video" 
    restart: unless-stopped

After that install qwen3-tts and try to run the model.
Backend will download but fail load.

Expected behavior

Models should load without errors.

Logs

Additional context

Start by reproducing the failure with the localai/localai:latest-gpu-hipblas Docker Compose configuration and the qwen3-tts or qwen3-asr model. Inspect how the hipblas image is built and how its backend dependencies are installed. Done means the affected models load successfully in that image without the FlashAttention2 error.