import torch import torch.nn as nn from model.config import SentinelConfig class SentinelModel(nn.Module): def __init__(self, config=None): super().__init__() self.config = config or SentinelConfig() self.token_embedding = nn.Embedding(self.config.vocab_size, self.config.hidden_size, padding_idx=self.config.pad_token_id) self.position_embedding = nn.Embedding(self.config.max_position_embeddings, self.config.hidden_size) layer = nn.TransformerEncoderLayer( d_model=self.config.hidden_size, nhead=self.config.num_heads, dim_feedforward=self.config.intermediate_size, dropout=self.config.dropout, activation="gelu", batch_first=True, norm_first=True ) self.transformer = nn.TransformerEncoder(layer, num_layers=self.config.num_layers) self.norm = nn.LayerNorm(self.config.hidden_size) self.lm_head = nn.Linear(self.config.hidden_size, self.config.vocab_size, bias=False) self.lm_head.weight = self.token_embedding.weight self.apply(self._init_weights) def _init_weights(self, module): if isinstance(module, nn.Linear): nn.init.normal_(module.weight, mean=0.0, std=0.02) if module.bias is not None: nn.init.zeros_(module.bias) elif isinstance(module, nn.Embedding): nn.init.normal_(module.weight, mean=0.0, std=0.02) def forward(self, input_ids, labels=None): _, seq_len = input_ids.shape if seq_len > self.config.max_position_embeddings: raise ValueError(f"Sequence length {seq_len} exceeds context length {self.config.max_position_embeddings}") pos = torch.arange(seq_len, device=input_ids.device).unsqueeze(0) x = self.token_embedding(input_ids) + self.position_embedding(pos) causal = torch.triu(torch.ones(seq_len, seq_len, device=input_ids.device, dtype=torch.bool), diagonal=1) padding = input_ids.eq(self.config.pad_token_id) x = self.transformer(x, mask=causal, src_key_padding_mask=padding) logits = self.lm_head(self.norm(x)) loss = None if labels is not None: loss = nn.CrossEntropyLoss(ignore_index=self.config.pad_token_id)( logits[:, :-1, :].contiguous().view(-1, self.config.vocab_size), labels[:, 1:].contiguous().view(-1) ) return {"logits": logits, "loss": loss}