File size: 9,484 Bytes
1620846
 
 
 
 
 
 
aabde66
1620846
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aabde66
 
 
1620846
aabde66
1620846
aabde66
 
 
 
 
 
 
 
 
 
 
1620846
aabde66
 
1620846
 
 
 
 
 
 
 
 
aabde66
1620846
aabde66
 
1620846
 
 
 
 
 
aabde66
 
 
 
 
 
 
1620846
aabde66
 
1620846
 
 
 
 
 
 
 
aabde66
 
 
 
 
 
 
 
 
1620846
 
 
 
9c2509e
 
 
1620846
 
 
9c2509e
 
1620846
9c2509e
 
 
 
aabde66
 
 
 
 
 
1620846
 
 
 
9c2509e
 
 
 
1620846
 
 
 
9c2509e
1620846
9c2509e
1620846
 
 
 
 
 
aabde66
 
 
 
1620846
 
 
 
 
 
 
 
 
 
 
 
 
 
aabde66
 
 
1620846
 
 
 
 
 
 
 
 
 
 
 
 
aabde66
 
 
 
 
1620846
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aabde66
1620846
aabde66
 
1620846
 
 
aabde66
 
 
 
 
1620846
 
 
aabde66
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Tuple

class ScaledDotProductAttention(nn.Module):
    """Scaled Dot-Product Attention mechanism with numerical stability"""
    
    def __init__(self, temperature: float = 1.0, dropout: float = 0.1):
        super().__init__()
        self.temperature = temperature
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            q: Query tensor [batch_size, n_heads, seq_len, d_k]
            k: Key tensor [batch_size, n_heads, seq_len, d_k]
            v: Value tensor [batch_size, n_heads, seq_len, d_k]
            mask: Mask tensor [batch_size, 1, seq_len, seq_len] or [batch_size, 1, 1, seq_len]
        
        Returns:
            output: Attention output [batch_size, n_heads, seq_len, d_k]
            attention: Attention weights [batch_size, n_heads, seq_len, seq_len]
        """
        # Calculate attention scores with temperature scaling
        d_k = q.size(-1)
        scores = torch.matmul(q, k.transpose(-2, -1)) / (self.temperature * math.sqrt(d_k))
        
        # Apply mask if provided - using fp16-safe value
        if mask is not None:
            # Determine safe mask value based on dtype
            if scores.dtype == torch.float16:
                mask_value = -65504.0  # Max negative value for fp16
            else:
                mask_value = -1e9  # Original value for fp32
            
            # Use torch.finfo for more robust dtype handling
            mask_value = torch.finfo(scores.dtype).min if hasattr(torch, 'finfo') else mask_value
            scores = scores.masked_fill(mask == 0, mask_value)
        
        # Apply softmax with numerical stability
        attention = F.softmax(scores, dim=-1)
        
        # Apply dropout
        attention = self.dropout(attention)
        
        # Apply attention to values
        output = torch.matmul(attention, v)
        
        return output, attention


class MultiHeadAttention(nn.Module):
    """Multi-Head Attention mechanism with improved stability"""
    
    def __init__(self, d_model: int, n_heads: int, dropout: float = 0.1, 
                 use_bias: bool = True, pre_norm: bool = False):
        super().__init__()
        assert d_model % n_heads == 0, "d_model must be divisible by n_heads"
        
        self.d_model = d_model
        self.n_heads = n_heads
        self.d_k = d_model // n_heads
        self.pre_norm = pre_norm
        
        # Linear projections with optional bias
        self.W_q = nn.Linear(d_model, d_model, bias=use_bias)
        self.W_k = nn.Linear(d_model, d_model, bias=use_bias)
        self.W_v = nn.Linear(d_model, d_model, bias=use_bias)
        self.W_o = nn.Linear(d_model, d_model, bias=use_bias)
        
        # Initialize weights using Xavier uniform
        self._init_weights()
        
        # Attention
        self.attention = ScaledDotProductAttention(temperature=1.0, dropout=dropout)
        
        # Dropout
        self.dropout = nn.Dropout(dropout)
        
        # Layer normalization
        self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
        
    def _init_weights(self):
        """Initialize weights with Xavier uniform distribution"""
        for module in [self.W_q, self.W_k, self.W_v, self.W_o]:
            nn.init.xavier_uniform_(module.weight)
            if module.bias is not None:
                nn.init.zeros_(module.bias)
    
    def forward(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
        """
        Args:
            query: Query tensor [batch_size, seq_len_q, d_model]
            key: Key tensor [batch_size, seq_len_k, d_model]
            value: Value tensor [batch_size, seq_len_v, d_model]
            mask: Mask tensor
        
        Returns:
            output: Multi-head attention output [batch_size, seq_len_q, d_model]
            attention: Attention weights [batch_size, n_heads, seq_len_q, seq_len_k]
        """
        batch_size = query.size(0)
        seq_len_q = query.size(1)  # Query sequence length
        seq_len_k = key.size(1)     # Key sequence length (can be different!)
        seq_len_v = value.size(1)   # Value sequence length (same as key)
        
        # Pre-norm variant (if enabled)
        if self.pre_norm:
            query = self.layer_norm(query)
            key = self.layer_norm(key)
            value = self.layer_norm(value)
        
        # Store residual
        residual = query
        
        # Linear projections - FIXED: Use correct sequence lengths
        Q = self.W_q(query).view(batch_size, seq_len_q, self.n_heads, self.d_k).transpose(1, 2)
        K = self.W_k(key).view(batch_size, seq_len_k, self.n_heads, self.d_k).transpose(1, 2)
        V = self.W_v(value).view(batch_size, seq_len_v, self.n_heads, self.d_k).transpose(1, 2)
        
        # Apply attention
        attn_output, attention_weights = self.attention(Q, K, V, mask)
        
        # Concatenate heads - use seq_len_q for output
        attn_output = attn_output.transpose(1, 2).contiguous().view(
            batch_size, seq_len_q, self.d_model
        )
        
        # Final linear projection
        output = self.W_o(attn_output)
        output = self.dropout(output)
        
        # Add residual and normalize
        output = output + residual
        if not self.pre_norm:
            output = self.layer_norm(output)
        
        return output, attention_weights

def create_padding_mask(seq: torch.Tensor, pad_idx: int = 0) -> torch.Tensor:
    """
    Create padding mask for attention
    
    Args:
        seq: Input sequence [batch_size, seq_len]
        pad_idx: Padding index
    
    Returns:
        mask: Padding mask [batch_size, 1, 1, seq_len]
    """
    # Create boolean mask
    mask = (seq != pad_idx).unsqueeze(1).unsqueeze(2)
    return mask.to(torch.bool)


def create_look_ahead_mask(size: int, device: torch.device) -> torch.Tensor:
    """
    Create look-ahead mask for decoder self-attention
    
    Args:
        size: Sequence length
        device: Device to create mask on
    
    Returns:
        mask: Look-ahead mask [1, 1, size, size]
    """
    # Create upper triangular matrix
    mask = torch.triu(torch.ones(size, size, device=device, dtype=torch.bool), diagonal=1)
    # Invert it (1 for allowed positions, 0 for masked)
    mask = ~mask
    return mask.unsqueeze(0).unsqueeze(0)


def create_masks(src: torch.Tensor, tgt: torch.Tensor, 
                 pad_idx: int = 0) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Create all masks needed for transformer
    
    Args:
        src: Source sequence [batch_size, src_len]
        tgt: Target sequence [batch_size, tgt_len]
        pad_idx: Padding index
    
    Returns:
        src_mask: Source padding mask
        tgt_mask: Target mask (padding + look-ahead)
        memory_mask: Memory mask for decoder cross-attention
    """
    # Source mask (padding only)
    src_mask = create_padding_mask(src, pad_idx)
    
    # Target padding mask
    tgt_pad_mask = create_padding_mask(tgt, pad_idx)
    
    # Target look-ahead mask
    tgt_len = tgt.size(1)
    tgt_look_ahead_mask = create_look_ahead_mask(tgt_len, tgt.device)
    
    # Combine padding and look-ahead masks for target
    # Both masks should be True where attention is allowed
    tgt_mask = tgt_pad_mask & tgt_look_ahead_mask
    
    # Memory mask (same as source mask)
    memory_mask = src_mask
    
    return src_mask, tgt_mask, memory_mask


# Optional: Flash Attention wrapper (if available)
try:
    from torch.nn.functional import scaled_dot_product_attention
    FLASH_ATTENTION_AVAILABLE = True
except ImportError:
    FLASH_ATTENTION_AVAILABLE = False

class FlashAttention(nn.Module):
    """Flash Attention wrapper for better performance (if available)"""
    
    def __init__(self, dropout: float = 0.1):
        super().__init__()
        self.dropout = dropout
        
    def forward(self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
                mask: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, None]:
        """
        Uses PyTorch's scaled_dot_product_attention if available (includes Flash Attention)
        """
        if FLASH_ATTENTION_AVAILABLE and mask is None:
            # Use efficient implementation when no mask
            output = scaled_dot_product_attention(
                q, k, v, 
                dropout_p=self.dropout if self.training else 0.0,
                is_causal=False
            )
            return output, None
        else:
            # Fallback to standard implementation
            d_k = q.size(-1)
            scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(d_k)
            
            if mask is not None:
                mask_value = torch.finfo(scores.dtype).min
                scores = scores.masked_fill(mask == 0, mask_value)
            
            attention = F.softmax(scores, dim=-1)
            if self.training and self.dropout > 0:
                attention = F.dropout(attention, p=self.dropout)
            
            output = torch.matmul(attention, v)
            return output, attention