Use flash_attn 2.7.4.post1 precompiled wheel (torch 2.6 compatible)

requirements.txt

@@ -27,5 +27,6 @@ insightface==0.7.3
 transformers==4.52.0
 huggingface_hub
 ninja
-# flash_attn 暂时移除，因为没有 torch 2.6 的预编译版本
-# 模型会使用默认的 attention 实现
+# flash_attn 预编译 wheel (torch 2.6 + CUDA 12 + Python 3.10)
+# 参考: https://huggingface.co/spaces/fffiloni/Meigen-MultiTalk/blob/main/requirements.txt
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp310-cp310-linux_x86_64.whl