torch.svd support

[SydCS]

Hi!
When running diff_erank, I came across "NotImplementedError: "svd_cuda_gesvdj" not implemented for 'Half'" or "NotImplementedError: "svd_cuda_gesvdj" not implemented for 'BFloat16'".
Does the same problem ever occur to you? Or is there any probable fix?
Thanks if someone could answer!

[waltonfuture]

Thanks for your interest to our work!
You can try this way:

def cal_erank(A):
    with torch.no_grad():
        # Convert to float32 before SVD
        A_float = A.to(torch.float32)
        eig_val = torch.svd(A_float / torch.trace(A_float))[1] 
        entropy = - (eig_val * torch.log(eig_val)).nansum().item()
        erank = math.exp(entropy)
    return erank