====== paddlematerials常见问题 ======

===== cuda13.0安装不上 paddle_scatter 的问题  =====
paddlematerials 里 链接引用了 https://github.com/PFCCLab/paddle_scatter/tree/8efe9af96bf2f98be9c4bc7ecf5dff471875149a  没有更新到最新

paddle_scatter  的代码里没有兼容cuda13.x，导致安装有点问题。 咱们也不贸然升级版本，最新的main分支是支持了 13.x ，免得有问题，直接补丁一下

找到setup_ops.py 文件修订如下：

<code python>
def set_cuda_archs():
    major, _ = paddle.version.cuda_version.split(".")
    if int(major) >= 13:
        # CUDA 13.0 minimum is sm_75 (Turing), supports up to Blackwell (sm_120/121)
        # Supported: 75, 80, 86, 87, 88, 89, 90, 100, 103, 110, 120, 121
        paddle_known_gpu_archs = [75, 80, 86, 87, 89, 90, 100, 120]
    elif int(major) >= 12:
        paddle_known_gpu_archs = [50, 60, 61, 70, 75, 80, 86, 89, 90]
    elif int(major) >= 11:
        paddle_known_gpu_archs = [50, 60, 61, 70, 75, 80, 86]
    elif int(major) >= 10:
        paddle_known_gpu_archs = [50, 52, 60, 61, 70, 75]
    else:
        raise ValueError("Not support cuda version.")

    os.environ["PADDLE_CUDA_ARCH_LIST"] = ",".join(
        [str(arch) for arch in paddle_known_gpu_archs]
    )

</code>