| 56 | """ |
| 57 | |
| 58 | def __init__( |
| 59 | self, |
| 60 | hidden_size, |
| 61 | num_attention_heads, |
| 62 | layer_number, |
| 63 | fp16=True, |
| 64 | attention_softmax_in_fp32=True, |
| 65 | ): |
| 66 | super(SelfAttention, self).__init__() |
| 67 | self.hidden_size = hidden_size |
| 68 | self.num_attention_heads = num_attention_heads |
| 69 | self.fp16 = fp16 |
| 70 | self.attention_softmax_in_fp32 = attention_softmax_in_fp32 |
| 71 | self.layer_number = max(1, layer_number) |
| 72 | |
| 73 | assert self.hidden_size % self.num_attention_heads == 0 |
| 74 | self.hidden_size_per_attention_head = int(self.hidden_size // self.num_attention_heads) |
| 75 | |
| 76 | self.query = torch.nn.Linear(self.hidden_size, self.hidden_size) |
| 77 | self.key = torch.nn.Linear(self.hidden_size, self.hidden_size) |
| 78 | self.value = torch.nn.Linear(self.hidden_size, self.hidden_size) |
| 79 | |
| 80 | self.norm_factor = math.sqrt(self.hidden_size_per_attention_head) |
| 81 | self.softmax = torch.nn.Softmax(dim=-1) |
| 82 | |
| 83 | self.dense = torch.nn.Linear(self.hidden_size, self.hidden_size) |
| 84 | |
| 85 | def forward( |
| 86 | self, |