| 56 | |
| 57 | |
| 58 | class T5LayerNorm(nn.Module): |
| 59 | |
| 60 | def __init__(self, dim, eps=1e-6): |
| 61 | super(T5LayerNorm, self).__init__() |
| 62 | self.dim = dim |
| 63 | self.eps = eps |
| 64 | self.weight = nn.Parameter(torch.ones(dim)) |
| 65 | |
| 66 | def forward(self, x): |
| 67 | x = x * torch.rsqrt(x.float().pow(2).mean(dim=-1, keepdim=True) + |
| 68 | self.eps) |
| 69 | if self.weight.dtype in [torch.float16, torch.bfloat16]: |
| 70 | x = x.type_as(self.weight) |
| 71 | return self.weight * x |
| 72 | |
| 73 | |
| 74 | class T5Attention(nn.Module): |