import torch | |
import torch.nn as nn | |
from transformers import LlamaModel | |
from transformers.models.llama.configuration_llama import LlamaConfig | |
class SuperTokenizer(LlamaModel): | |
def __init__(self, config: LlamaConfig): | |
super().__init__(config) | |
self.batch_size = 1 | |
self.my_pooler = nn.Linear(config.hidden_size, 4096) | |
def forward(self, input_ids, attention_mask, super_token_indices, **kwargs): | |
hidden_state = [] | |
for i in range(0, input_ids.shape[0], self.batch_size): | |
end_index = min(i + self.batch_size, input_ids.shape[0]) | |
output = super().forward( | |
input_ids[i:end_index], | |
attention_mask[i:end_index], | |
**kwargs, | |
) | |
for j in range(end_index - i): | |
hidden_state.append( | |
output.last_hidden_state[j][super_token_indices[i + j]] | |
) | |
embedding = torch.cat(hidden_state, dim=0).contiguous() | |
return embedding |