[skip ci][Docs] Add FlashAttention requirement for audio generation to prevent noise-only outputs in mimo-audio model (#2205)

Signed-off-by: 齐保元 <qibaoyuan@xiaomi.com>

examples/offline_inference/mimo_audio/README.md

@@ -29,6 +29,10 @@ The `MIMO_AUDIO_TOKENIZER_PATH` environment variable is mandatory due to the spe
 export MIMO_AUDIO_TOKENIZER_PATH="XiaomiMiMo/MiMo-Audio-Tokenizer"
 ```
 
+### Flash Attention (audio generation)
+
+For **audio generation** (e.g. TTS variants, multi-turn spoken dialogue with audio output), install the **`flash-attn`** package with a build that matches your **CUDA** and **PyTorch** versions. On GPU, omitting **`flash-attn`** can cause **generated audio to be noise-only or otherwise unusable**. See the [FlashAttention](https://github.com/Dao-AILab/flash-attention) project for installation options and prebuilt wheels.
+
 ## Quick Start
 
 Run a single sample for basic TTS:

examples/online_serving/mimo_audio/README.md

@@ -4,6 +4,9 @@
 
 Please refer to [README.md](../../../README.md)
 
+> ⚠️ **Important (audio generation)**  
+> For **audio generation** (TTS, responses that include synthesized audio, etc.), install **`flash-attn`** for your **CUDA** and **PyTorch** stack. Without it on GPU, **output audio may be noise-only or unusable**. See the [FlashAttention](https://github.com/Dao-AILab/flash-attention) repository for compatible builds.
+
 ## Run examples (MiMo-Audio)
 
 ### Launch the Server