File size: 7,173 Bytes
f98cc3f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math
import torch
import torch.nn as nn
import torch.nn.functional as F

def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
    # Reshape position_ids to match q's shape
    position_ids = position_ids.unsqueeze(0).unsqueeze(0)  # [1, 1, seq_len]
    
    # Get the rotary embeddings for this position
    cos = cos.squeeze(0)  # [seq_len, dim]
    sin = sin.squeeze(0)  # [seq_len, dim]
    
    # Apply rotary embeddings
    q_embed = (q * cos) + (rotate_half(q) * sin)
    k_embed = (k * cos) + (rotate_half(k) * sin)
    
    return q_embed, k_embed

def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., :x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2:]
    return torch.cat((-x2, x1), dim=-1)

class LlamaRMSNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-5):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(hidden_size))
        self.eps = eps

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

class LlamaRotaryEmbedding(nn.Module):
    def __init__(self, dim, max_position_embeddings=2048, base=10000):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim//4, dtype=torch.float32) / dim))
        self.register_buffer("inv_freq", inv_freq)
        self.register_buffer("cos_cached", None, persistent=False)
        self.register_buffer("sin_cached", None, persistent=False)
        self.max_position_embeddings = max_position_embeddings

    def forward(self, x, seq_len):
        if self.cos_cached is not None and self.cos_cached.size(1) >= seq_len:
            return self.cos_cached[:, :seq_len, :], self.sin_cached[:, :seq_len, :]

        t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
        freqs = torch.einsum("i,j->ij", t, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        
        cos = torch.cos(emb)[None, :, :]
        sin = torch.sin(emb)[None, :, :]
        
        self.cos_cached = cos
        self.sin_cached = sin
        
        return cos, sin

class LlamaAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = 64  # Fixed head dimension to match saved model
        
        self.q_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)
        self.k_proj = nn.Linear(config.hidden_size, 64, bias=False)  # Single head dimension
        self.v_proj = nn.Linear(config.hidden_size, 64, bias=False)  # Single head dimension
        self.o_proj = nn.Linear(config.hidden_size, config.hidden_size, bias=False)

    def forward(self, hidden_states, rotary_emb=None):
        bsz, q_len, _ = hidden_states.size()
        
        q = self.q_proj(hidden_states)
        k = self.k_proj(hidden_states)
        v = self.v_proj(hidden_states)
        
        # Split q into heads before applying rotary embeddings
        q = q.view(bsz, q_len, self.num_heads, -1)  # -1 will be 64
        k = k.view(bsz, q_len, 1, -1)  # Keep k as single head
        v = v.view(bsz, q_len, 1, -1)  # Keep v as single head
        
        # Apply rotary embeddings if provided
        if rotary_emb is not None:
            position_ids = torch.arange(q_len, device=q.device)
            cos, sin = rotary_emb(v, q_len)
            # Split q and k in half for rotation
            q1, q2 = q[..., :32], q[..., 32:]
            k1, k2 = k[..., :32], k[..., 32:]
            # Apply rotation to first half
            q_embed = torch.cat([
                q1 * cos.unsqueeze(2) - q2 * sin.unsqueeze(2),
                q2 * cos.unsqueeze(2) + q1 * sin.unsqueeze(2)
            ], dim=-1)
            k_embed = torch.cat([
                k1 * cos.unsqueeze(2) - k2 * sin.unsqueeze(2),
                k2 * cos.unsqueeze(2) + k1 * sin.unsqueeze(2)
            ], dim=-1)
            q, k = q_embed, k_embed
        
        # Expand k and v to match number of heads
        k = k.expand(-1, -1, self.num_heads, -1)
        v = v.expand(-1, -1, self.num_heads, -1)
        
        # Scaled dot-product attention
        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        attn_weights = F.softmax(attn_weights, dim=-1)
        
        attn_output = torch.matmul(attn_weights, v)
        attn_output = attn_output.view(bsz, q_len, self.hidden_size)
        attn_output = self.o_proj(attn_output)
        
        return attn_output

class LlamaMLP(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.gate_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
        self.act_fn = nn.SiLU()

    def forward(self, x):
        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))

class LlamaDecoderLayer(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.self_attn = LlamaAttention(config)
        self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.mlp = LlamaMLP(config)

    def forward(self, hidden_states, rotary_emb=None):  # Add rotary_emb parameter
        residual = hidden_states
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.self_attn(hidden_states, rotary_emb=rotary_emb)
        hidden_states = residual + hidden_states
        
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        
        return hidden_states

class LlamaModel(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.config = config
        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
        self.rotary_emb = LlamaRotaryEmbedding(dim=64)  # This will create inv_freq of size 16
        self.layers = nn.ModuleList([LlamaDecoderLayer(config) for _ in range(config.num_hidden_layers)])
        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(self, input_ids):
        hidden_states = self.embed_tokens(input_ids)
        
        for layer in self.layers:
            hidden_states = layer(hidden_states, rotary_emb=self.rotary_emb)
            
        hidden_states = self.norm(hidden_states)
        return hidden_states

class LlamaForCausalLM(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.model = LlamaModel(config)
        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

    def forward(self, input_ids):
        hidden_states = self.model(input_ids)
        logits = self.lm_head(hidden_states)
        return logits, None