| 77 | """ |
| 78 | |
| 79 | def __init__( |
| 80 | self, |
| 81 | in_features: int, |
| 82 | out_features: int, |
| 83 | bias: bool = True, |
| 84 | gather_output: bool = True, |
| 85 | init_method: Callable[[torch.Tensor], torch.Tensor] = init.xavier_normal_, |
| 86 | stride: int = 1, |
| 87 | keep_master_weight_for_test: bool = False, |
| 88 | lora_rank=0 |
| 89 | ) -> None: |
| 90 | nn.Module.__init__(self) |
| 91 | |
| 92 | # Keep input parameters |
| 93 | self.in_features = in_features |
| 94 | self.out_features = out_features |
| 95 | self.gather_output = gather_output |
| 96 | # Divide the weight matrix along the last dimension. |
| 97 | world_size = get_model_parallel_world_size() |
| 98 | self.output_size_per_partition = divide_and_check_no_remainder(out_features, world_size) |
| 99 | |
| 100 | # Parameters. |
| 101 | # Note: torch.nn.functional.linear performs XA^T + b and as a result |
| 102 | # we allocate the transpose. |
| 103 | self.weight = Parameter(torch.Tensor(self.output_size_per_partition, self.in_features)) |
| 104 | if bias: |
| 105 | self.bias = Parameter(torch.Tensor(self.output_size_per_partition)) |
| 106 | # Always initialize bias to zero. |
| 107 | with torch.no_grad(): |
| 108 | self.bias.zero_() |
| 109 | else: |
| 110 | self.register_parameter("bias", None) |
| 111 | |
| 112 | # Initialize weight. |
| 113 | self.master_weight = _initialize_affine_weight( |
| 114 | self.weight, |
| 115 | self.out_features, |
| 116 | self.in_features, |
| 117 | self.output_size_per_partition, |
| 118 | 0, |
| 119 | init_method, |
| 120 | stride=stride, |
| 121 | return_master_weight=keep_master_weight_for_test, |
| 122 | ) |
| 123 | |
| 124 | self.lora_rank = lora_rank |
| 125 | if self.lora_rank > 0: |
| 126 | # if world_size > 1: |
| 127 | # raise NotImplemented("Lora with model parallel with change the original behavior, not yet supported") |
| 128 | self.lora_a = nn.Linear(self.in_features, self.lora_rank, bias=False) |
| 129 | # workaround because trunc_normal_ does not currently support bfloat16 |
| 130 | _ = init.trunc_normal_(self.lora_a.weight.data.to(torch.float32), std=.02) |
| 131 | self.lora_a.weight.data.copy_(_) |
| 132 | self.lora_b = ColumnParallelLinear(self.lora_rank, self.out_features, bias=False, gather_output=gather_output) |
| 133 | nn.init.zeros_(self.lora_b.weight) |
| 134 | else: |
| 135 | self.lora_a = None |
| 136 | self.lora_b = None |