| 25 | class MPTDecoderLayer(Module): |
| 26 | |
| 27 | def __init__(self, config: PretrainedConfig, layer_idx: int): |
| 28 | super().__init__() |
| 29 | self.layer_idx = layer_idx |
| 30 | self.config = config |
| 31 | |
| 32 | hidden_size = config.hidden_size |
| 33 | dtype = config.dtype |
| 34 | tp_size = config.mapping.tp_size |
| 35 | tp_rank = config.mapping.tp_rank |
| 36 | tp_group = config.mapping.tp_group |
| 37 | layernorm_epsilon = config.norm_epsilon |
| 38 | |
| 39 | self.input_layernorm = LayerNorm(normalized_shape=hidden_size, |
| 40 | eps=layernorm_epsilon, |
| 41 | bias=False, |
| 42 | dtype=dtype) |
| 43 | |
| 44 | layers_range = config.mapping.pp_layers(config.num_hidden_layers) |
| 45 | local_layer_idx = layer_idx - layers_range[0] |
| 46 | self.attention = Attention( |
| 47 | local_layer_idx=local_layer_idx, |
| 48 | hidden_size=hidden_size, |
| 49 | num_attention_heads=config.num_attention_heads, |
| 50 | num_kv_heads=config.num_key_value_heads, |
| 51 | attention_mask_type=AttentionMaskType.causal, |
| 52 | dtype=dtype, |
| 53 | tp_group=tp_group, |
| 54 | tp_size=tp_size, |
| 55 | tp_rank=tp_rank, |
| 56 | bias=config.bias, |
| 57 | position_embedding_type=PositionEmbeddingType.alibi, |
| 58 | quant_mode=config.quant_mode, |
| 59 | clip_qkv=config.clip_qkv, |
| 60 | alibi_bias_max=config.alibi_bias_max) |
| 61 | |
| 62 | self.mlp = MLP(hidden_size=hidden_size, |
| 63 | ffn_hidden_size=hidden_size * 4, |
| 64 | hidden_act=config.hidden_act, |
| 65 | dtype=dtype, |
| 66 | bias=config.bias, |
| 67 | tp_group=tp_group, |
| 68 | tp_size=tp_size, |
| 69 | quant_mode=config.quant_mode) |
| 70 | |
| 71 | self.post_layernorm = LayerNorm(normalized_shape=hidden_size, |
| 72 | eps=layernorm_epsilon, |
| 73 | bias=False, |
| 74 | dtype=dtype) |
| 75 | |
| 76 | def forward(self, |
| 77 | hidden_states: Tensor, |