https://github.com/Dao-AILab/flash-attention/releases/

- 2.7.2.post1: cuda版本-torch版本-python版本


- ImportError: /lib64/libc.so.6: version `GLIBC_2.32' not found
    -- https://github.com/Dao-AILab/flash-attention/issues/1708 (falsh-attn版本问题)
- ImportError: cannot import name 'get_tcp_uri' from 'vllm.utils'
    -- https://github.com/volcengine/verl/issues/3638（貌似vllm和flash_attn都不干净）
    -- 看上去 flash_attn-2.7.4.post1 （cu12 / torch2.6.0-cu14 / python 3.11） + vllm 0.10.0 可以
    - 更新至0.10.0后又有以下错误
        - - vllm/_C.abi3.so: undefined symbol: _ZN3c106ivalue14ConstantString6createENSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEEE
    - 最后可能有效的版本(python 3.11 + torch 2.7.1(vllm required) + fa-2.6.0.post1 (abiTrue) + vllm 0.10.0)


- ModuleNotFoundError: No module named 'sglang.srt.weight_sync'
    -- https://github.com/volcengine/verl/issues/2921
        -- sglang版本应该为sglang[all]==0.4.9.post6 (flash-attn 2.8.2 / 看上去2.7.4.post1也能work)
- ModuleNotFoundError: No module named 'sgl_kernel.kvcacheio'
    -- https://github.com/sgl-project/sglang/issues/3687
- [sgl_kernel] CRITICAL: Could not load any common_ops library!


- ImportError: cannot import name 'layer_type_validation' from 'transformers.configuration_utils'
    -- https://github.com/huggingface/transformers/issues/38617 (更新至4.52.3 / 4.56.1)