| import torch | |
| from transformers import PhiForCausalLM | |
| from .configuration_pruned_phi import PhiPrunedConfig | |
| import torch.nn as nn | |
| class PhiPrunedForCausalLM(PhiForCausalLM): | |
| config_class = PhiPrunedConfig | |
| def __init__(self, config: PhiPrunedConfig): | |
| super().__init__(config) | |
| for i in range(32): | |
| self.model.layers[i].self_attn.dense = nn.Linear(640, 2560, bias=True) | |
| self.model.layers[i].self_attn.hidden_size = 640 | |
| self.model.layers[i].self_attn.q_proj = nn.Linear(2560, 640, bias=True) | |
| self.model.layers[i].self_attn.k_proj = nn.Linear(2560, 640, bias=True) | |
| self.model.layers[i].self_attn.v_proj = nn.Linear(2560, 640, bias=True) | |
| self.model.layers[i].mlp.fc1 = nn.Linear(2560, 10240, bias=True) | |
| self.model.layers[i].mlp.fc2 = nn.Linear(10240, 2560, bias=True) | |
| for layer in self.model.layers: | |
| layer.self_attn.num_heads = layer.self_attn.q_proj.weight.data.shape[0] // layer.self_attn.head_dim | |
| layer.self_attn.num_key_value_heads = layer.self_attn.k_proj.weight.data.shape[ | |
| 0] // layer.self_attn.head_dim | |