File size: 7,454 Bytes
e2536b8
 
 
e17f601
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e2536b8
e17f601
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e2536b8
e17f601
 
 
 
 
 
 
 
 
 
 
 
 
e2536b8
e17f601
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e2536b8
 
 
 
d0e13a3
 
 
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
from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
from transformers import PreTrainedModel, AutoConfig, AutoModelForCausalLM
from .configuration_lumenspark import LumensparkConfig
from torch import nn
import torch
import math

# ----------------------------
# Low-Rank Linear Layer Implementation
# ----------------------------

class LowRankLinear(nn.Module):
    """

    A low-rank linear layer that factorizes a standard linear layer into two smaller ones.

    This allows for reduced parameter count and faster computation.

    """
    def __init__(self, in_features, out_features, rank, init_std=0.02):
        super().__init__()
        self.U = nn.Linear(in_features, rank, bias=False)
        self.V = nn.Linear(rank, out_features, bias=False)
        nn.init.normal_(self.U.weight, std=init_std)
        nn.init.normal_(self.V.weight, std=init_std)

    def forward(self, x):
        """

        Forward pass through two low-rank linear layers (U and V).

        """
        return self.V(self.U(x))
        
# ----------------------------
# Lumenspark Self-Attention Implementation
# ----------------------------

class LumensparkSelfAttention(nn.Module):
    """

    Custom self-attention mechanism for the Lumenspark model.

    It uses low-rank approximations to reduce computational cost and memory usage.

    """
    def __init__(self, embed_dim, num_heads, head_dim=None, dropout=0.0):
        super().__init__()
        assert (embed_dim % num_heads) == 0, 'Embedding dimension must be divisible by the number of heads'

        self.num_heads = num_heads
        self.embed_dim = embed_dim
        self.head_dim = head_dim if head_dim is not None else embed_dim // num_heads

        # Query, Key and Value transformations using LowRankLinear
        self.q_proj = nn.Linear(embed_dim, self.head_dim * num_heads)
        self.k_proj = nn.Linear(embed_dim, self.head_dim * num_heads)
        self.v_proj = nn.Linear(embed_dim, self.head_dim * num_heads)

        self.dropout_layer = nn.Dropout(dropout)
        self.output_transform = nn.Linear(self.head_dim * num_heads, embed_dim)

    def stable_softmax(self, x, dim=-1):
        # Subtract max for numerical stability
        x_max = torch.max(x, dim=dim, keepdim=True)[0]
        exp_x = torch.exp(x - x_max)
        return exp_x / (torch.sum(exp_x, dim=dim, keepdim=True) + 1e-6)
    
    def forward(self, inputs, attention_mask=None):
        batch_size, seq_len, _ = inputs.shape

        q = self.q_proj(inputs).view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        k = self.k_proj(inputs).view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
        v = self.v_proj(inputs).view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)

        attention_scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        
        if attention_mask is not None:
            attention_scores = attention_scores.masked_fill(attention_mask == 0, float('-inf'))
        
        attention_weights = self.stable_softmax(attention_scores, dim=-1)
        attention_weights = self.dropout_layer(attention_weights)

        attention_output = torch.matmul(attention_weights, v)
        attention_output = attention_output.transpose(1, 2).contiguous().view(batch_size, seq_len, self.embed_dim)
        return self.output_transform(attention_output)

# ----------------------------
# Define Lumenspark Model Wrapper
# ----------------------------

class LumensparkModel(PreTrainedModel):
    config_class = LumensparkConfig

    def __init__(self, config):
        super().__init__(config)
        self.config = config

        # Token and position embeddings
        self.token_embedding = nn.Embedding(config.vocab_size, config.embed_dim)
        self.position_embedding = nn.Embedding(config.seq_length, config.embed_dim)

        # Lumenspark transformer encoder layers with prenormalization and LayerScale
        self.layers = nn.ModuleList()
        for _ in range(config.depth):
            layer = nn.ModuleDict({
                "norm1": nn.LayerNorm(config.embed_dim),
                "attn": LumensparkSelfAttention(
                    embed_dim=config.embed_dim,
                    num_heads=config.heads,
                    head_dim=config.embed_dim // config.heads,
                    dropout=config.dropout
                ),
                "norm2": nn.LayerNorm(config.embed_dim),
                "ffn": nn.Sequential(
                    LowRankLinear(config.embed_dim, config.embed_dim * 4, rank=config.rank),
                    nn.GELU(),
                    nn.Dropout(config.dropout),
                    LowRankLinear(config.embed_dim * 4, config.embed_dim, rank=config.rank),
                    nn.Dropout(config.dropout)
                ),
            })
            layer.layer_scale_attn = nn.Parameter(torch.ones(config.embed_dim) * 1e-2)
            layer.layer_scale_ffn = nn.Parameter(torch.ones(config.embed_dim) * 1e-2)
            self.layers.append(layer)

        self.final_norm = nn.LayerNorm(config.embed_dim)
        self.fc_out = nn.Linear(config.embed_dim, config.vocab_size)
        self.dropout = nn.Dropout(config.dropout)

        # Call init_weights at the end to ensure proper initialization
        self.init_weights()

    def forward(self, input_ids, attention_mask=None, labels=None):
        batch_size, seq_length = input_ids.size()

        position_ids = torch.arange(0, seq_length, dtype=torch.long, device=input_ids.device)
        position_ids = position_ids.unsqueeze(0).expand(batch_size, seq_length)

        token_embeddings = self.token_embedding(input_ids)
        position_embeddings = self.position_embedding(position_ids)
        embeddings = token_embeddings + position_embeddings
        embeddings = self.dropout(embeddings)

        causal_mask = torch.tril(torch.ones((seq_length, seq_length), device=embeddings.device)).unsqueeze(0).unsqueeze(0)

        if attention_mask is not None:
            attention_mask = attention_mask[:, None, None, :].float()
            combined_mask = attention_mask * causal_mask
        else:
            combined_mask = causal_mask

        for layer in self.layers:
            embeddings_norm = layer["norm1"](embeddings)
            attn_output = layer["attn"](embeddings_norm, attention_mask=combined_mask)
            embeddings = embeddings + layer.layer_scale_attn * attn_output

            embeddings_norm = layer["norm2"](embeddings)
            ffn_output = layer["ffn"](embeddings_norm)
            embeddings = embeddings + layer.layer_scale_ffn * ffn_output

        embeddings = self.final_norm(embeddings)
        logits = self.fc_out(embeddings)

        loss = None
        if labels is not None:
            shift_logits = logits[:, :-1, :].contiguous().view(-1, self.config.vocab_size)
            shift_labels = labels[:, 1:].contiguous().view(-1)
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(shift_logits, shift_labels)

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=logits
        )
    
AutoConfig.register("lumenspark", LumensparkConfig)
AutoModelForCausalLM.register(LumensparkConfig, LumensparkModel)