import torch from torch import nn import torch.nn.functional as F from typing import List, Optional, Tuple, Union, Dict from transformers.modeling_utils import PreTrainedModel from transformers.modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions from transformers.generation import GenerationMixin from torch.utils.checkpoint import checkpoint from .configuration_stldec import STLDecoderConfig class STLPreTrainedModel(PreTrainedModel): config_class = STLDecoderConfig base_model_prefix = "model" def _init_weights(self, module): """Migliorata con Xavier Uniform per evitare gradienti esplosivi nelle fasi iniziali.""" if isinstance(module, nn.Linear): torch.nn.init.xavier_uniform_(module.weight) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=0.02) class STLAttention(nn.Module): def __init__(self, config): super().__init__() self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.scaling = self.head_dim ** -0.5 self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False) self.k_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False) self.v_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False) self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False) self.dropout = nn.Dropout(config.attention_dropout) def forward(self, hidden_states, encoder_hidden_states=None, past_key_value=None, attention_mask=None): bsz, q_len, _ = hidden_states.size() key_value_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states kv_len = key_value_states.size(1) q = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) * self.scaling k = self.k_proj(key_value_states).view(bsz, kv_len, self.num_heads, self.head_dim).transpose(1, 2) v = self.v_proj(key_value_states).view(bsz, kv_len, self.num_heads, self.head_dim).transpose(1, 2) if past_key_value is not None and encoder_hidden_states is None: k = torch.cat([past_key_value[0], k], dim=2) v = torch.cat([past_key_value[1], v], dim=2) present_kv = (k, v) if encoder_hidden_states is None else None attn_weights = torch.matmul(q, k.transpose(-1, -2)) if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_probs = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype) attn_output = torch.matmul(self.dropout(attn_probs), v) attn_output = attn_output.transpose(1, 2).contiguous().view(bsz, q_len, self.hidden_size) return self.out_proj(attn_output), present_kv class STLDecoderBlock(nn.Module): def __init__(self, config): super().__init__() self.ln1 = nn.LayerNorm(config.hidden_size) self.self_attn = STLAttention(config) self.ln_cross = nn.LayerNorm(config.hidden_size) self.cross_attn = STLAttention(config) self.ln2 = nn.LayerNorm(config.hidden_size) self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size) self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size) self.dropout = nn.Dropout(config.dropout) self.gradient_checkpointing = False def forward(self, hidden_states, encoder_hidden_states=None, past_key_value=None, attention_mask=None): if self.training and self.gradient_checkpointing: return checkpoint( self.internal_forward, hidden_states, encoder_hidden_states, past_key_value, attention_mask, use_reentrant=False ) return self.internal_forward(hidden_states, encoder_hidden_states, past_key_value, attention_mask) def internal_forward(self, hidden_states, encoder_hidden_states=None, past_key_value=None, attention_mask=None): """Modificata in Pre-Norm per garantire la stabilità del gradiente.""" # 1. Self-Attention residual = hidden_states hidden_states = self.ln1(hidden_states) # LN PRIMA hidden_states, pkv = self.self_attn(hidden_states, past_key_value=past_key_value, attention_mask=attention_mask) hidden_states = residual + self.dropout(hidden_states) # 2. Cross-Attention if encoder_hidden_states is not None: residual = hidden_states hidden_states = self.ln_cross(hidden_states) # LN PRIMA hidden_states, _ = self.cross_attn(hidden_states, encoder_hidden_states=encoder_hidden_states) hidden_states = residual + self.dropout(hidden_states) # 3. Feed-Forward residual = hidden_states hidden_states = self.ln2(hidden_states) # LN PRIMA hidden_states = self.fc2(F.gelu(self.fc1(hidden_states))) hidden_states = residual + self.dropout(hidden_states) return hidden_states, pkv class STLDecoderModel(STLPreTrainedModel, GenerationMixin): def __init__(self, config): super().__init__(config) self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, config.pad_token_id) # Posizionali Sinusoidali rimossi dal __init__ perché calcolati dinamicamente nel forward self.layers = nn.ModuleList([STLDecoderBlock(config) for _ in range(config.num_hidden_layers)]) self.norm = nn.LayerNorm(config.hidden_size) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.post_init() def get_sinusoidal_embeddings(self, seq_len, d_model, device): """Genera posizioni matematiche stabili, evitando errori di indice della tabella fixed.""" inv_freq = 1.0 / (10000 ** (torch.arange(0, d_model, 2).float() / d_model)).to(device) pos = torch.arange(seq_len, device=device).type_as(inv_freq) sin_inp = torch.einsum("i,j->ij", pos, inv_freq) emb = torch.cat((sin_inp.sin(), sin_inp.cos()), dim=-1) return emb[None, :, :] @property def supports_gradient_checkpointing(self): return True def _set_gradient_checkpointing(self, module, value=False): if isinstance(module, STLDecoderBlock): module.gradient_checkpointing = value def get_input_embeddings(self): return self.embed_tokens def set_input_embeddings(self, v): self.embed_tokens = v def get_output_embeddings(self): return self.lm_head def forward(self, input_ids=None, encoder_hidden_states=None, past_key_values=None, labels=None, use_cache=None, return_dict=None, **kwargs): return_dict = return_dict if return_dict is not None else self.config.use_return_dict use_cache = use_cache if use_cache is not None else self.config.use_cache bsz, seq_len = input_ids.size() past_len = 0 if past_key_values is not None: past_len = past_key_values[0][0].shape[2] hidden_states = self.embed_tokens(input_ids) # Sostituzione con sinusoidali (più robusti dello Script 2) pos_emb = self.get_sinusoidal_embeddings(seq_len, self.config.hidden_size, input_ids.device) hidden_states = hidden_states + pos_emb[:, :seq_len, :] # Maschera causale ottimizzata causal_mask = torch.full((seq_len, seq_len + past_len), float("-inf"), device=input_ids.device, dtype=hidden_states.dtype) causal_mask.triu_(diagonal=past_len + 1) causal_mask = causal_mask[None, None, :, :] next_cache = () if use_cache else None for i, layer in enumerate(self.layers): pk = past_key_values[i] if past_key_values is not None else None hidden_states, pkv = layer(hidden_states, encoder_hidden_states=encoder_hidden_states, past_key_value=pk, attention_mask=causal_mask) if use_cache: next_cache += (pkv,) hidden_states = self.norm(hidden_states) logits = self.lm_head(hidden_states) loss = None if labels is not None: shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # --- MODIFICA DI SICUREZZA --- # Prendiamo il vocab_size dai logits correnti, non dal config fisso, # per evitare l'Assertion t < n_classes se hai fatto un resize_token_embeddings. current_vocab_size = logits.size(-1) loss = F.cross_entropy( shift_logits.view(-1, current_vocab_size), shift_labels.view(-1), ignore_index=-100 # Standard HF per ignorare padding trasformato ) if not return_dict: return (loss, logits, next_cache) if loss is not None else (logits, next_cache) return CausalLMOutputWithCrossAttentions(loss=loss, logits=logits, past_key_values=next_cache) @staticmethod def _reorder_cache(past_key_values, beam_idx): return tuple(tuple(s.index_select(0, beam_idx.to(s.device)) for s in layer) for layer in past_key_values)