File size: 7,865 Bytes
b9c4adf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Baseline implementations for fair comparison.

Baselines:
  1. Standard Transformer: Dense MLP FFN, no TT, no quantum.
  2. Distilled: Smaller transformer trained with KD.
  3. Pruned: Magnitude-based structured pruning.
  4. TT-Only: Tensor network FFN without quantum or adaptive rank.
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional


class StandardTransformer(nn.Module):
    """
    Basic transformer decoder (GPT-style) with dense MLP FFN.

    Reference baseline — matches Q-TensorFormer architecture
    exactly except for TT decomposition and quantum layers.
    """

    def __init__(self, vocab_size: int = 10000, d_model: int = 128,
                 n_heads: int = 4, n_layers: int = 2, ff_mult: int = 4,
                 max_seq_len: int = 128, dropout: float = 0.1):
        super().__init__()
        self.d_model = d_model
        self.config = type("config", (), {
            "d_model": d_model, "n_heads": n_heads, "n_layers": n_layers,
            "ff_multiplier": ff_mult, "max_seq_len": max_seq_len,
            "vocab_size": vocab_size, "dropout": dropout,
        })()

        self.embedding = nn.Embedding(vocab_size, d_model)
        self.pos_encoding = _PositionalEncoding(d_model, max_seq_len, dropout)

        self.blocks = nn.ModuleList([
            _StandardBlock(d_model, n_heads, ff_mult, dropout, max_seq_len)
            for _ in range(n_layers)
        ])

        self.ln_f = nn.LayerNorm(d_model)
        self.lm_head = nn.Linear(d_model, vocab_size, bias=False)
        self.lm_head.weight = self.embedding.weight

    def forward(self, input_ids, attention_mask=None, return_stats=False):
        x = self.embedding(input_ids)
        x = self.pos_encoding(x)

        for block in self.blocks:
            x = block(x, mask=attention_mask)

        x = self.ln_f(x)
        logits = self.lm_head(x)

        if return_stats:
            return logits, []
        return logits

    @property
    def total_params(self) -> int:
        return sum(p.numel() for p in self.parameters())


class DistilledTransformer(nn.Module):
    """
    Smaller transformer trained via knowledge distillation.

    Designed to match Q-TensorFormer parameter counts.
    """

    def __init__(self, vocab_size: int = 10000, d_model: int = 96,
                 n_heads: int = 4, n_layers: int = 2, ff_mult: int = 3,
                 max_seq_len: int = 128, dropout: float = 0.1):
        super().__init__()
        self.d_model = d_model
        self.config = type("config", (), {
            "d_model": d_model, "n_heads": n_heads, "n_layers": n_layers,
            "ff_multiplier": ff_mult, "max_seq_len": max_seq_len,
            "vocab_size": vocab_size, "dropout": dropout,
        })()

        self.embedding = nn.Embedding(vocab_size, d_model)
        self.pos_encoding = _PositionalEncoding(d_model, max_seq_len, dropout)

        self.blocks = nn.ModuleList([
            _StandardBlock(d_model, n_heads, ff_mult, dropout, max_seq_len)
            for _ in range(n_layers)
        ])

        self.ln_f = nn.LayerNorm(d_model)
        self.lm_head = nn.Linear(d_model, vocab_size, bias=False)
        self.lm_head.weight = self.embedding.weight

    def forward(self, input_ids, attention_mask=None, return_stats=False):
        x = self.embedding(input_ids)
        x = self.pos_encoding(x)

        for block in self.blocks:
            x = block(x, mask=attention_mask)

        x = self.ln_f(x)
        logits = self.lm_head(x)

        if return_stats:
            return logits, []
        return logits

    @property
    def total_params(self) -> int:
        return sum(p.numel() for p in self.parameters())


class PrunedTransformer(nn.Module):
    """
    Magnitude-pruned standard transformer.

    Prunes FFN weights globally to match Q-TensorFormer parameter count.
    Applies structured pruning (zeroing channels) for efficiency.
    """

    def __init__(self, base_model: StandardTransformer,
                 prune_ratio: float = 0.5):
        super().__init__()
        self.base = base_model
        self.prune_ratio = prune_ratio
        self.config = base_model.config
        self._prune()

    def _prune(self):
        """Apply structured magnitude pruning to FFN layers."""
        all_weights = []
        for block in self.base.blocks:
            for weight in [block.ffn[0].weight, block.ffn[2].weight]:
                all_weights.append(weight.flatten())

        # Compute global threshold
        flat = torch.cat(all_weights)
        k = int(len(flat) * self.prune_ratio)
        threshold = torch.topk(flat.abs(), k, largest=False).values[-1]

        # Apply structured pruning (zero rows/cols)
        for block in self.base.blocks:
            for layer in [block.ffn[0], block.ffn[2]]:
                mask = (layer.weight.abs() > threshold).float()
                # Zero small rows entirely
                row_norms = mask.sum(dim=1)
                dead_rows = row_norms < layer.weight.size(1) * 0.1
                mask[dead_rows] = 0
                layer.weight.data *= mask

    def forward(self, *args, **kwargs):
        return self.base(*args, **kwargs)

    @property
    def total_params(self) -> int:
        return sum(p.numel() for p in self.parameters())


class _StandardBlock(nn.Module):
    """Standard transformer decoder block."""

    def __init__(self, d_model, n_heads, ff_mult, dropout, max_seq_len):
        super().__init__()
        self.ln1 = nn.LayerNorm(d_model)
        self.attn = _CausalAttention(d_model, n_heads, dropout, max_seq_len)
        self.ln2 = nn.LayerNorm(d_model)
        self.ffn = nn.Sequential(
            nn.Linear(d_model, d_model * ff_mult),
            nn.GELU(),
            nn.Linear(d_model * ff_mult, d_model),
            nn.Dropout(dropout),
        )
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, mask=None):
        x = x + self.dropout(self.attn(self.ln1(x), mask=mask))
        x = x + self.ffn(self.ln2(x))
        return x


class _CausalAttention(nn.Module):
    """Causal multi-head attention."""

    def __init__(self, d_model, n_heads, dropout, max_seq_len):
        super().__init__()
        assert d_model % n_heads == 0
        self.n_heads = n_heads
        self.head_dim = d_model // n_heads
        self.scale = math.sqrt(self.head_dim)

        self.qkv = nn.Linear(d_model, 3 * d_model, bias=False)
        self.out_proj = nn.Linear(d_model, d_model, bias=False)
        self.dropout = nn.Dropout(dropout)

        self.max_seq_len = max_seq_len

    def forward(self, x, mask=None):
        B, T, C = x.shape
        qkv = self.qkv(x).reshape(B, T, 3, self.n_heads, self.head_dim)
        q, k, v = qkv.unbind(dim=2)
        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)

        attn = (q @ k.transpose(-2, -1)) / self.scale
        causal = torch.triu(torch.ones(T, T, device=x.device) * float("-inf"), diagonal=1)
        attn = attn + causal

        if mask is not None:
            attn = attn + mask.unsqueeze(1).unsqueeze(2) * float("-inf")

        attn = F.softmax(attn, dim=-1)
        attn = self.dropout(attn)

        out = (attn @ v).transpose(1, 2).reshape(B, T, C)
        return self.out_proj(out)


class _PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_len, dropout):
        super().__init__()
        self.dropout = nn.Dropout(dropout)
        pe = torch.zeros(max_len, d_model)
        pos = torch.arange(max_len).unsqueeze(1).float()
        div = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
        pe[:, 0::2] = torch.sin(pos * div)
        pe[:, 1::2] = torch.cos(pos * div)
        self.register_buffer("pe", pe.unsqueeze(0))

    def forward(self, x):
        return self.dropout(x + self.pe[:, :x.size(1)])