The implementation is identical to `lit_gpt.model.Block` with the exception that we replace the attention layer where adaption is implemented.
| 123 | |
| 124 | |
| 125 | class Block(nn.Module): |
| 126 | """The implementation is identical to `lit_gpt.model.Block` with the exception that |
| 127 | we replace the attention layer where adaption is implemented.""" |
| 128 | |
| 129 | def __init__(self, config: Config, block_idx: int) -> None: |
| 130 | super().__init__() |
| 131 | self.norm_1 = config.norm_class(config.n_embd, eps=config.norm_eps) |
| 132 | self.attn = CausalSelfAttention(config, block_idx) |
| 133 | if not config.shared_attention_norm: |
| 134 | self.norm_2 = config.norm_class(config.n_embd, eps=config.norm_eps) |
| 135 | self.mlp = config.mlp_class(config) |
| 136 | |
| 137 | self.config = config |
| 138 | |
| 139 | def forward( |
| 140 | self, |
| 141 | x: torch.Tensor, |
| 142 | rope: RoPECache, |
| 143 | max_seq_length: int, |
| 144 | mask: Optional[torch.Tensor] = None, |
| 145 | input_pos: Optional[torch.Tensor] = None, |
| 146 | kv_cache: Optional[KVCache] = None, |
| 147 | adapter_kv_cache: Optional[KVCache] = None, |
| 148 | ) -> Tuple[torch.Tensor, Optional[KVCache], Optional[KVCache]]: |
| 149 | n_1 = self.norm_1(x) |
| 150 | h, new_kv_cache, new_adapter_kv_cache = self.attn( |
| 151 | n_1, rope, max_seq_length, mask, input_pos, kv_cache, adapter_kv_cache |
| 152 | ) |
| 153 | if self.config.parallel_residual: |
| 154 | n_2 = n_1 if self.config.shared_attention_norm else self.norm_2(x) |
| 155 | x = x + h + self.mlp(n_2) |
| 156 | else: |
| 157 | if self.config.shared_attention_norm: |
| 158 | raise NotImplementedError( |
| 159 | "No checkpoint amongst the ones we support uses this configuration" |
| 160 | " (non-parallel residual and shared attention norm)." |
| 161 | ) |
| 162 | x = x + h |
| 163 | x = x + self.mlp(self.norm_2(x)) |
| 164 | return x, new_kv_cache, new_adapter_kv_cache |
| 165 | |
| 166 | |
| 167 | class CausalSelfAttention(BaseCausalSelfAttention): |