File size: 8,173 Bytes
488f6f3
f6a4b47
 
488f6f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6a4b47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
488f6f3
f6a4b47
 
 
 
 
 
 
 
488f6f3
 
f6a4b47
 
 
 
 
 
 
 
 
 
 
 
 
 
488f6f3
 
f6a4b47
 
 
 
 
 
488f6f3
f6a4b47
 
 
 
488f6f3
f6a4b47
488f6f3
f6a4b47
488f6f3
 
f6a4b47
 
 
488f6f3
 
 
 
 
 
f6a4b47
 
 
488f6f3
f6a4b47
 
488f6f3
f6a4b47
488f6f3
 
 
 
f6a4b47
 
488f6f3
f6a4b47
488f6f3
f6a4b47
488f6f3
 
 
 
 
 
 
 
 
 
f6a4b47
 
 
 
 
 
 
 
 
 
 
 
 
488f6f3
f6a4b47
 
 
 
488f6f3
 
 
 
 
 
 
 
 
 
f6a4b47
488f6f3
 
 
 
 
 
 
f6a4b47
 
 
 
 
 
 
488f6f3
 
 
 
f6a4b47
488f6f3
 
f6a4b47
 
488f6f3
f6a4b47
 
 
 
 
 
 
 
 
488f6f3
f6a4b47
 
 
488f6f3
f6a4b47
 
 
 
488f6f3
f6a4b47
 
 
 
 
488f6f3
f6a4b47
 
 
 
488f6f3
f6a4b47
 
 
 
 
 
 
 
 
 
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
import math 
import torch
import torch.nn as nn
import torch.nn.functional as F

def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
    t = torch.arange(end, device=freqs.device)
    freqs = torch.outer(t, freqs)
    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
    return freqs_cis

def reshape_for_broadcast(freqs_cis, x):
    batch_size, num_heads, seq_len, head_size = x.shape
    freqs_cis = freqs_cis[:seq_len]
    shape = [1, 1, seq_len, head_size // 2]
    return freqs_cis.view(*shape)
    
def apply_rope(x, position, freqs_cis):
    x_ = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
    freqs_cis = reshape_for_broadcast(freqs_cis, x)
    x_out = torch.view_as_real(x_ * freqs_cis).flatten(3)
    return x_out.type_as(x)

class RMSNorm(torch.nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.ones(dim))

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward(self, x):
        output = self._norm(x.float()).type_as(x)
        return output * self.weight

class Attention(nn.Module):
    """
    Multi-head Self-Attention with RoPE
    """

    def __init__(self, num_heads, head_size, num_embed):
        super().__init__()
        self.num_heads = num_heads
        self.head_size = head_size

        self.wq = nn.Linear(num_embed, num_heads * head_size, bias = False)
        self.wk = nn.Linear(num_embed, num_heads * head_size, bias = False)
        self.wv = nn.Linear(num_embed, num_heads * head_size, bias = False)
        self.wo = nn.Linear(num_heads * head_size, num_embed, bias = False)
        
    def forward(self, x, freqs_cis):
        B, T, C = x.shape

        mask = torch.triu(torch.full((T, T), float("-inf"), device=x.device), diagonal=1)

        xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)

        xq = xq.view(B, T, self.num_heads, self.head_size)
        xk = xk.view(B, T, self.num_heads, self.head_size)
        xv = xv.view(B, T, self.num_heads, self.head_size)

        xq = xq.transpose(1, 2)
        xk = xk.transpose(1, 2)
        xv = xv.transpose(1, 2)

        xq = apply_rope(xq, T, freqs_cis)
        xk = apply_rope(xk, T, freqs_cis)

        attn_weights = torch.matmul(xq, xk.transpose(2, 3)) / math.sqrt(self.head_size)
        attn_weights += mask
        attn_weights = F.softmax(attn_weights.float(), dim=-1).type_as(xq)
        output = torch.matmul(attn_weights, xv)
        output = output.transpose(1, 2).contiguous().view(B, T, C)
        return self.wo(output)

class MLP(nn.Module):
    def __init__(self, num_embed, dropout):
        super().__init__()
        self.num_embed = num_embed

        hidden_dim = 3 * int(num_embed * 2 / 3)

        self.linear1 = nn.Linear(num_embed, hidden_dim)
        self.linear2 = nn.Linear(hidden_dim, num_embed)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = self.linear1(x)
        x = F.silu(x)
        x = self.linear2(x)
        x = self.dropout(x)
        return x
        
class TransformerBlock(nn.Module):
    """
    This calss will group together MultiHead Attention and
    FeedForward NN, so that we can copy it in Transformer
    """

    def __init__(self, num_heads, num_embed, dropout):
        super().__init__()
        self.num_heads = num_heads
        self.num_embed = num_embed
        head_size = num_embed // num_heads
        self.sa = Attention(
            num_heads=num_heads,
            head_size=head_size,
            num_embed=num_embed
        )
        self.ffwd = MLP(num_embed=num_embed, dropout=dropout)
        # add the layer normalization
        self.ln1 = RMSNorm(num_embed)
        self.ln2 = RMSNorm(num_embed)

    def forward(self, x, freqs_cis):
        # "x +" is the skip (or residual) connection
        # it helps with optimization
        # also we apply layer normalization before self-attention
        # and feed-forward (a reshufle from original paper)
        x = x + self.sa(self.ln1(x), freqs_cis)
        x = x + self.ffwd(self.ln2(x))
        return x


class Transformer(nn.Module):
    def __init__(self, **kwargs):
        super().__init__()
        # a simple lookup table that stores embeddings of a fixed dictionary and size
        # each token directly reads off the logits for the next token from a lookup table
        # see more: https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html
        self.vocab_size = kwargs.get("vocab_size", 100)
        self.num_embed = kwargs.get("num_embed", 32)
        self.num_heads = kwargs.get("num_heads", 4)
        self.num_layers = kwargs.get("num_layers", 4)
        self.max_seq_len = kwargs.get("max_seq_len", 1024)
        self.dropout = kwargs.get("dropout", 0.2)
        # each token reads the logits for the next token from a lookup table
        self.token_embedding_table = nn.Embedding(self.vocab_size, self.num_embed)
        # each position from 0 to block_size-1 will get its embedding
        #self.position_embedding_table = nn.Embedding(self.block_size, self.num_embed)
        self.blocks = nn.ModuleList([
            TransformerBlock(
                num_heads=self.num_heads,
                num_embed=self.num_embed,
                dropout=self.dropout
            )
            for _ in range(self.num_layers)
        ])
        # we add the layer norm before the Linear layer
        self.lm_head = nn.Linear(self.num_embed, self.vocab_size)
        self.norm = RMSNorm(self.num_embed)

        self.freqs_cis = precompute_freqs_cis(
            self.num_embed//self.num_heads,
            self.max_seq_len * 2,
            500000,
        )

    def forward(self, idx, targets=None):
        B, T = idx.shape
        # idx and targets are (B,T) tensor of integers
        # the token_emb is (B, T, C), C = NUM_EMBED
        x = self.token_embedding_table(idx)

        freq = self.freqs_cis[:self.max_seq_len]
        # apply one head of self-attention
        for block in self.blocks:
            x = block(x, freq)

        x = self.norm(x)
            
        # (B, T, vocab_size)
        logits = self.lm_head(x)
        # compute the loss
        if targets != None:
            # cross_entropy accepts inputs in a (batch_size, num_classes)
            # so we need to reformat our logits dimensions to
            # (batch_size * time, dim_vocabulary), time = block_size
            loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
        else:
            loss = None
        return logits, loss

    def generate(self, idx: torch.Tensor, max_new_tokens: int, temperature: float = 0.7, top_p: float = 0.9):
        for _ in range(max_new_tokens):
            idx_crop = idx[:, -self.max_seq_len:]

            freq = self.freqs_cis[:self.max_seq_len]
            logits, loss = self.forward(idx_crop)
            logits = logits[:, -1, :]

            if temperature > 0:
                probs = F.softmax(logits / temperature, dim=-1)  
                idx_next = self.sample_top_p(probs, top_p)
            else:
                probs = F.softmax(logits, dim=-1)
                idx_next = torch.multinomial(probs, num_samples=1)
            idx = torch.cat((idx, idx_next), dim=1)  # (B, T+1)
        return idx[0]

    def sample_top_p(self, probs: torch.Tensor, top_p: float) -> torch.Tensor:
        sorted_probs, sorted_indices = torch.sort(probs, descending=True, dim=-1)
        cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
        
        # Create a mask for top-p filtering
        top_p_mask = cumulative_probs <= top_p
        top_p_mask[..., 1:] = top_p_mask[..., :-1].clone()
        top_p_mask[..., 0] = 1

        filtered_probs = sorted_probs * top_p_mask
        filtered_probs /= filtered_probs.sum(dim=-1, keepdim=True)  # Normalize filtered probabilities

        next_token = torch.multinomial(filtered_probs, num_samples=1)
        return torch.gather(sorted_indices, -1, next_token)