File size: 12,434 Bytes
590a604
 
 
 
 
 
 
 
 
 
 
 
 
5a20c96
 
1ec7405
d18b34d
5a20c96
 
1ec7405
5a20c96
6e61c34
b43ba56
5a20c96
b43ba56
1ec7405
5a20c96
 
 
 
 
 
 
 
 
 
 
d18b34d
b43ba56
 
5a20c96
 
d18b34d
 
 
 
 
 
 
b43ba56
 
d18b34d
5a20c96
d18b34d
b43ba56
 
 
 
 
d18b34d
5a20c96
d18b34d
b43ba56
 
 
 
 
d18b34d
 
1ec7405
 
d18b34d
5a20c96
 
 
 
 
 
 
d18b34d
b43ba56
d18b34d
5a20c96
 
 
 
 
 
d18b34d
b43ba56
5a20c96
 
 
 
 
 
 
 
d18b34d
b43ba56
 
 
 
 
 
d18b34d
5a20c96
 
b43ba56
 
 
 
 
5a20c96
 
 
 
 
b43ba56
 
 
 
 
5a20c96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d18b34d
b43ba56
 
 
1ec7405
5a20c96
 
 
 
 
b43ba56
1ec7405
5a20c96
 
b43ba56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a20c96
b43ba56
 
5a20c96
 
 
d18b34d
 
 
 
 
 
 
b43ba56
 
d18b34d
 
 
5a20c96
 
1ec7405
 
5a20c96
 
 
 
 
 
 
 
 
590a604
 
 
5a20c96
d18b34d
5a20c96
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b43ba56
 
 
5a20c96
 
b43ba56
5a20c96
d18b34d
 
 
5a20c96
b43ba56
 
 
5a20c96
 
 
 
 
 
 
 
 
 
b43ba56
 
 
 
 
5a20c96
 
 
 
1ec7405
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a20c96
 
 
 
 
 
 
 
d18b34d
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
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
"""Transformer Encoder implementation (Pre-LN).

This module implements the encoder component of the Transformer architecture:
- TransformerEncoderLayer: Single encoder block with self-attention + FFN
- TransformerEncoder: Full stack with embeddings and positional encoding

Design notes:
- Pre-LN with RMSNorm for training stability
- Masks are boolean: True = attend, False = mask
- Supports T5-style relative position bias

Author: Oliver Perrin
Date: 2025-10-23
"""

from typing import List, Literal, Optional, Tuple, Union, cast

import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint

# Encoder implementation
from .attention import MultiHeadAttention, T5RelativePositionBias
from .feedforward import FeedForward
from .positional_encoding import LearnedPositionalEncoding, PositionalEncoding
from .t5_layer_norm import T5LayerNorm


class TransformerEncoderLayer(nn.Module):
    """
    Single Transformer encoder layer (Pre-LN).

    Args:
        d_model: model hidden size
        num_heads: number of attention heads
        d_ff: hidden dimension of the position-wise feed-forward network
        dropout: dropout probability applied to sublayer outputs
        quantization: optional quantization mode ("4bit", "8bit")
        activation: activation function for FFN ("gelu", "relu", or "swiglu")
        scale_attn_scores: Whether to scale attention scores by sqrt(d_k). T5 does NOT scale.
    """

    def __init__(
        self,
        d_model: int,
        num_heads: int,
        d_ff: int,
        dropout: float = 0.1,
        quantization: Optional[str] = None,
        activation: Literal["gelu", "relu", "swiglu", "gated-gelu"] = "gated-gelu",
        scale_attn_scores: bool = True,  # T5 uses False
    ):
        super().__init__()
        self.self_attn = MultiHeadAttention(
            d_model=d_model,
            num_heads=num_heads,
            dropout=0.0,
            quantization=quantization,
            scale_scores=scale_attn_scores,
        )
        # set MHA internal dropout to 0.0 and use dropout1/dropout2 in the layer
        self.ffn = FeedForward(
            d_model=d_model,
            d_ff=d_ff,
            dropout=dropout,
            activation=activation,
            quantization=quantization,
        )

        self.norm1 = T5LayerNorm(d_model)
        self.norm2 = T5LayerNorm(d_model)

        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

    def forward(
        self,
        x: torch.Tensor,
        mask: Optional[torch.Tensor] = None,
        collect_attn: bool = False,
        position_bias: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, Optional[torch.Tensor]]]:
        """
        Forward pass for the encoder layer.

        Args:
            x: (batch, seq_len, d_model) - input embeddings / representations
            mask: optional attention mask, shape either (batch, seq_q, seq_k) or (batch, 1, seq_q, seq_k)
            collect_attn: whether to return attention weights
            position_bias: optional (1, num_heads, seq_q, seq_k) T5-style relative position bias

        Returns:
            x: (batch, seq_len, d_model)
            If you want attention weights, set collect_attn externally (the encoder stack can collect them).
        """
        # Self-attention sublayer (Pre-LN)
        x_norm = self.norm1(x)  # Pre-LN
        # self_attn expects query, key, value; for encoder they are the same
        attn_out, attn_weights = self.self_attn(
            x_norm,
            x_norm,
            x_norm,
            mask,
            return_attn_weights=collect_attn,
            position_bias=position_bias,
        )
        x = x + self.dropout1(attn_out)

        # Clamp inf values for fp16/bf16 training stability (like HuggingFace T5)
        if x.dtype == torch.float16 or x.dtype == torch.bfloat16:
            clamp_value = torch.finfo(x.dtype).max - 1000
            x = torch.clamp(x, min=-clamp_value, max=clamp_value)

        # Feed-forward sublayer (Pre-LN)
        x_norm = self.norm2(x)
        ffn_out = self.ffn(x_norm)
        x = x + self.dropout2(ffn_out)

        # Clamp inf values for fp16/bf16 training stability
        if x.dtype == torch.float16 or x.dtype == torch.bfloat16:
            clamp_value = torch.finfo(x.dtype).max - 1000
            x = torch.clamp(x, min=-clamp_value, max=clamp_value)

        # Return output (and optionally attn_weights if caller wants to collect them)
        return x, attn_weights


class TransformerEncoder(nn.Module):
    """
    Full encoder: token embedding + positional encoding + N encoder layers.

    Args:
        vocab_size: vocabulary size (ignored if you always pass embeddings)
        d_model: model hidden size
        num_layers: number of encoder layers to stack
        num_heads: number of attention heads
        d_ff: hidden dimension in FFN
        dropout: dropout probability (applied in positional encoding & residuals)
        max_len: maximum sequence length for positional encoding
        pad_token_id: optional token id for padding; if provided and input is token ids,
                      a padding mask will be constructed automatically
    """

    def __init__(
        self,
        vocab_size: int,
        d_model: int = 512,
        num_layers: int = 6,
        num_heads: int = 8,
        d_ff: int = 2048,
        dropout: float = 0.1,
        max_len: int = 512,
        pad_token_id: Optional[int] = None,
        quantization: Optional[str] = None,
        use_learned_pos_enc: bool = False,
        activation: Literal["gelu", "relu", "swiglu", "gated-gelu"] = "gated-gelu",
        use_relative_position_bias: bool = False,  # T5-style relative position bias
        gradient_checkpointing: bool = False,
    ):
        super().__init__()
        self.vocab_size = vocab_size
        self.d_model = d_model
        self.pad_token_id = pad_token_id
        self.use_relative_position_bias = use_relative_position_bias
        self.gradient_checkpointing = gradient_checkpointing

        # Token embedding (only used if forward receives token ids)
        self.embedding = nn.Embedding(vocab_size, d_model, padding_idx=pad_token_id)

        # Positional encoding (disabled when using relative position bias for T5)
        self.relative_position_bias: Optional[T5RelativePositionBias] = None
        if use_relative_position_bias:
            # T5 uses relative position bias instead of absolute positional embeddings
            self.pos_encoder = None
            self.relative_position_bias = T5RelativePositionBias(
                num_heads=num_heads,
                num_buckets=32,
                max_distance=128,
                is_decoder=False,
            )
        elif use_learned_pos_enc:
            # T5 uses max_len=512 by default; we add buffer for special tokens
            self.pos_encoder = LearnedPositionalEncoding(
                d_model=d_model, max_len=max_len + 2, dropout=dropout
            )
        else:
            self.pos_encoder = PositionalEncoding(d_model=d_model, max_len=max_len, dropout=dropout)

        # T5 does NOT scale attention scores by sqrt(d_k), others do
        scale_attn_scores = not use_relative_position_bias

        # Encoder layers stack
        self.layers = nn.ModuleList(
            [
                TransformerEncoderLayer(
                    d_model=d_model,
                    num_heads=num_heads,
                    d_ff=d_ff,
                    dropout=dropout,
                    quantization=quantization,
                    activation=activation,
                    scale_attn_scores=scale_attn_scores,
                )
                for _ in range(num_layers)
            ]
        )

        # Final T5LayerNorm for Pre-LN stacks
        self.final_norm = T5LayerNorm(d_model)

        # Dropout applied after embedding + positional encoding (paper uses this)
        self.input_dropout = nn.Dropout(dropout)

    def _build_padding_mask(self, input_ids: torch.Tensor) -> torch.Tensor:
        """
        Build a 3D attention mask (batch, seq, seq) from input_ids and pad_token_id.
        True indicates valid positions; False indicates masked (pad).
        """
        assert self.pad_token_id is not None, (
            "pad_token_id must be set to build padding mask from ids."
        )
        # mask shape: (batch, seq) where True = token kept (non-pad)
        pad_mask = input_ids != self.pad_token_id
        # Convert to (batch, seq_q, seq_k) by outer product broadcasting
        # We want positions that are valid as both query and key
        attn_mask = pad_mask.unsqueeze(1) & pad_mask.unsqueeze(2)
        # attn_mask dtype should be bool
        return attn_mask

    def forward(
        self,
        inputs: torch.Tensor,
        mask: Optional[torch.Tensor] = None,
        collect_attn: bool = False,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, List[torch.Tensor]]]:
        """
        Forward through the encoder.

        Args:
            inputs: either
                - token ids: LongTensor of shape (batch, seq)
                - embeddings: FloatTensor of shape (batch, seq, d_model)
            mask: optional attention mask. If None and pad_token_id is set and inputs are token ids,
                  a padding mask will be created automatically with shape (batch, seq, seq).
                  The mask should be boolean where True indicates allowed attention.
            collect_attn: if True, returns (output, [attn_weights_per_layer]) where each entry is (batch, num_heads, seq, seq)

        Returns:
            output: (batch, seq, d_model)
            or (output, attn_list) if collect_attn True
        """
        # If inputs are token ids, embed them; otherwise assume they are embeddings
        if inputs.dim() == 2:  # token ids
            if self.embedding is None:
                raise ValueError("Encoder was not constructed with an embedding layer.")
            # T5/FLAN-T5 does NOT scale embeddings by sqrt(d_model)
            x = self.embedding(inputs)
            seq_len = inputs.size(1)
        elif inputs.dim() == 3:  # already embeddings
            x = inputs
            seq_len = inputs.size(1)
        else:
            raise ValueError(
                "inputs must be (batch, seq) token ids or (batch, seq, d_model) embeddings"
            )

        # Positional encoding + dropout (only if not using relative position bias)
        if self.pos_encoder is not None:
            x = self.pos_encoder(x)
        x = self.input_dropout(x)

        # Build mask if needed
        if mask is None and inputs.dim() == 2 and self.pad_token_id is not None:
            mask = self._build_padding_mask(inputs)

        # Ensure mask is boolean and on the same device
        if mask is not None:
            mask = mask.to(dtype=torch.bool, device=x.device)

        # Compute relative position bias if using T5-style
        position_bias = None
        if self.relative_position_bias is not None:
            position_bias = self.relative_position_bias(seq_len, seq_len, x.device)

        attn_weights_per_layer: List[torch.Tensor] = []

        # Pass through each encoder layer (optionally collect attn)
        for layer in self.layers:
            if self.gradient_checkpointing and self.training:
                # Gradient checkpointing requires the inputs to require grad
                # We use a lambda to pass keyword arguments
                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        return module(*inputs, mask=mask, collect_attn=collect_attn, position_bias=position_bias)
                    return custom_forward

                x, attn = cast(
                    Tuple[torch.Tensor, Optional[torch.Tensor]],
                    checkpoint(
                        create_custom_forward(layer),
                        x,
                        use_reentrant=False,
                    ),
                )
            else:
                x, attn = layer(x, mask=mask, collect_attn=collect_attn, position_bias=position_bias)
            
            if collect_attn:
                attn_weights_per_layer.append(attn)

        # Final normalization (Pre-LN stack)
        x = self.final_norm(x)

        if collect_attn:
            return x, attn_weights_per_layer
        return x