Create components.py
Browse files- components.py +388 -0
components.py
ADDED
|
@@ -0,0 +1,388 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
components.py
|
| 3 |
+
=============
|
| 4 |
+
Architectural components for SmolLM2-135M implementation
|
| 5 |
+
|
| 6 |
+
Components:
|
| 7 |
+
- RMSNorm: Root Mean Square Layer Normalization
|
| 8 |
+
- RotaryEmbedding: Rotary Position Embeddings (RoPE)
|
| 9 |
+
- GroupedQueryAttention: Grouped Query Attention (9 Q heads, 3 KV heads)
|
| 10 |
+
- SwiGLU_FFN: SwiGLU Feed-Forward Network
|
| 11 |
+
- TransformerBlock: Complete transformer block with pre-norm architecture
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
import torch
|
| 15 |
+
import torch.nn as nn
|
| 16 |
+
import torch.nn.functional as F
|
| 17 |
+
import math
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
class RMSNorm(nn.Module):
|
| 21 |
+
"""
|
| 22 |
+
Root Mean Square Layer Normalization
|
| 23 |
+
|
| 24 |
+
Simpler and faster than LayerNorm:
|
| 25 |
+
- No mean centering
|
| 26 |
+
- No bias term
|
| 27 |
+
- 10-15% faster than LayerNorm
|
| 28 |
+
|
| 29 |
+
Formula: output = input * rsqrt(mean(input²) + eps) * weight
|
| 30 |
+
"""
|
| 31 |
+
|
| 32 |
+
def __init__(self, hidden_size, eps=1e-5):
|
| 33 |
+
"""
|
| 34 |
+
Args:
|
| 35 |
+
hidden_size (int): Dimension of the input
|
| 36 |
+
eps (float): Small constant for numerical stability
|
| 37 |
+
"""
|
| 38 |
+
super().__init__()
|
| 39 |
+
self.eps = eps
|
| 40 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 41 |
+
|
| 42 |
+
def forward(self, x):
|
| 43 |
+
"""
|
| 44 |
+
Args:
|
| 45 |
+
x (torch.Tensor): Input tensor of shape [batch, seq_len, hidden_size]
|
| 46 |
+
|
| 47 |
+
Returns:
|
| 48 |
+
torch.Tensor: Normalized tensor of same shape as input
|
| 49 |
+
"""
|
| 50 |
+
# Calculate variance (mean of squares)
|
| 51 |
+
variance = x.pow(2).mean(-1, keepdim=True)
|
| 52 |
+
|
| 53 |
+
# Normalize: x / sqrt(variance + eps)
|
| 54 |
+
x = x * torch.rsqrt(variance + self.eps)
|
| 55 |
+
|
| 56 |
+
# Scale by learned weight
|
| 57 |
+
return self.weight * x
|
| 58 |
+
|
| 59 |
+
|
| 60 |
+
class RotaryEmbedding(nn.Module):
|
| 61 |
+
"""
|
| 62 |
+
Rotary Position Embedding (RoPE)
|
| 63 |
+
|
| 64 |
+
Encodes position by rotating Q and K vectors in 2D subspaces.
|
| 65 |
+
Enables relative position encoding and extrapolation to longer sequences.
|
| 66 |
+
|
| 67 |
+
Key properties:
|
| 68 |
+
- Applied only to Q and K, not V
|
| 69 |
+
- Different rotation frequencies for different dimension pairs
|
| 70 |
+
- Enables length extrapolation beyond training sequences
|
| 71 |
+
"""
|
| 72 |
+
|
| 73 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000.0):
|
| 74 |
+
"""
|
| 75 |
+
Args:
|
| 76 |
+
dim (int): Dimension of each attention head (typically hidden_size / num_heads)
|
| 77 |
+
max_position_embeddings (int): Maximum sequence length
|
| 78 |
+
base (float): Base for inverse frequency calculation (theta)
|
| 79 |
+
"""
|
| 80 |
+
super().__init__()
|
| 81 |
+
self.dim = dim
|
| 82 |
+
self.max_position_embeddings = max_position_embeddings
|
| 83 |
+
self.base = base
|
| 84 |
+
|
| 85 |
+
# Calculate inverse frequencies for rotation
|
| 86 |
+
# inv_freq[i] = 1 / (base^(2i/dim)) for i in [0, dim/2)
|
| 87 |
+
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float() / self.dim))
|
| 88 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 89 |
+
|
| 90 |
+
def forward(self, x, position_ids):
|
| 91 |
+
"""
|
| 92 |
+
Args:
|
| 93 |
+
x (torch.Tensor): Input tensor (used for device/dtype)
|
| 94 |
+
position_ids (torch.Tensor): Position indices [batch, seq_len] or [seq_len]
|
| 95 |
+
|
| 96 |
+
Returns:
|
| 97 |
+
tuple: (cos, sin) embeddings of shape [batch, seq_len, dim]
|
| 98 |
+
"""
|
| 99 |
+
# Ensure position_ids has batch dimension
|
| 100 |
+
if position_ids.dim() == 1:
|
| 101 |
+
position_ids = position_ids.unsqueeze(0)
|
| 102 |
+
|
| 103 |
+
# Calculate rotation angles: position_ids × inv_freq
|
| 104 |
+
# Shape: [batch, seq_len, dim/2]
|
| 105 |
+
freqs = torch.einsum('bi,j->bij', position_ids.float(), self.inv_freq)
|
| 106 |
+
|
| 107 |
+
# Duplicate frequencies for both sin and cos
|
| 108 |
+
# Shape: [batch, seq_len, dim]
|
| 109 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 110 |
+
|
| 111 |
+
# Return cos and sin, preserving input dtype
|
| 112 |
+
return emb.cos().to(x.dtype), emb.sin().to(x.dtype)
|
| 113 |
+
|
| 114 |
+
|
| 115 |
+
def rotate_half(x):
|
| 116 |
+
"""
|
| 117 |
+
Rotate half the hidden dimensions
|
| 118 |
+
|
| 119 |
+
For RoPE, we rotate pairs of dimensions. This function rearranges
|
| 120 |
+
the tensor to prepare for rotation.
|
| 121 |
+
|
| 122 |
+
Args:
|
| 123 |
+
x (torch.Tensor): Input of shape [..., dim]
|
| 124 |
+
|
| 125 |
+
Returns:
|
| 126 |
+
torch.Tensor: Rotated tensor where second half is negated and moved to first
|
| 127 |
+
"""
|
| 128 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 129 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 130 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
def apply_rotary_pos_emb(q, k, cos, sin):
|
| 134 |
+
"""
|
| 135 |
+
Apply rotary position embeddings to queries and keys
|
| 136 |
+
|
| 137 |
+
Rotation formula:
|
| 138 |
+
q_rotated = q * cos + rotate_half(q) * sin
|
| 139 |
+
k_rotated = k * cos + rotate_half(k) * sin
|
| 140 |
+
|
| 141 |
+
Args:
|
| 142 |
+
q (torch.Tensor): Query tensor [batch, num_heads, seq_len, head_dim]
|
| 143 |
+
k (torch.Tensor): Key tensor [batch, num_heads, seq_len, head_dim]
|
| 144 |
+
cos (torch.Tensor): Cosine embeddings [batch, seq_len, head_dim]
|
| 145 |
+
sin (torch.Tensor): Sine embeddings [batch, seq_len, head_dim]
|
| 146 |
+
|
| 147 |
+
Returns:
|
| 148 |
+
tuple: (q_rotated, k_rotated) with rotary embeddings applied
|
| 149 |
+
"""
|
| 150 |
+
# Add dimensions for broadcasting
|
| 151 |
+
# cos/sin: [batch, seq_len, dim] -> [batch, 1, seq_len, dim]
|
| 152 |
+
if cos.dim() == 2:
|
| 153 |
+
cos = cos.unsqueeze(0)
|
| 154 |
+
sin = sin.unsqueeze(0)
|
| 155 |
+
if cos.dim() == 3:
|
| 156 |
+
cos = cos.unsqueeze(1)
|
| 157 |
+
sin = sin.unsqueeze(1)
|
| 158 |
+
|
| 159 |
+
# Apply rotation
|
| 160 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 161 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 162 |
+
|
| 163 |
+
return q_embed, k_embed
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
class GroupedQueryAttention(nn.Module):
|
| 167 |
+
"""
|
| 168 |
+
Grouped Query Attention (GQA)
|
| 169 |
+
|
| 170 |
+
Memory-efficient attention where multiple query heads share KV heads.
|
| 171 |
+
SmolLM2-135M uses 9 query heads and 3 KV heads (3:1 ratio).
|
| 172 |
+
|
| 173 |
+
Benefits:
|
| 174 |
+
- Reduces KV cache memory by 66% vs full MHA
|
| 175 |
+
- Maintains most of multi-head attention's expressiveness
|
| 176 |
+
- Used in Llama 2, Mistral, and other modern LLMs
|
| 177 |
+
|
| 178 |
+
Architecture:
|
| 179 |
+
- 9 query heads (each head_dim=64)
|
| 180 |
+
- 3 KV heads (each head_dim=64)
|
| 181 |
+
- Each KV head is repeated 3 times to serve 3 query heads
|
| 182 |
+
"""
|
| 183 |
+
|
| 184 |
+
def __init__(self, config):
|
| 185 |
+
"""
|
| 186 |
+
Args:
|
| 187 |
+
config: Model configuration with attributes:
|
| 188 |
+
- hidden_size: Model dimension (576)
|
| 189 |
+
- num_attention_heads: Number of query heads (9)
|
| 190 |
+
- num_key_value_heads: Number of KV heads (3)
|
| 191 |
+
- max_position_embeddings: Max sequence length
|
| 192 |
+
- rope_theta: RoPE base frequency
|
| 193 |
+
"""
|
| 194 |
+
super().__init__()
|
| 195 |
+
self.hidden_size = config.hidden_size # 576
|
| 196 |
+
self.num_heads = config.num_attention_heads # 9
|
| 197 |
+
self.num_kv_heads = config.num_key_value_heads # 3
|
| 198 |
+
self.num_kv_groups = self.num_heads // self.num_kv_heads # 3
|
| 199 |
+
self.head_dim = self.hidden_size // self.num_heads # 64
|
| 200 |
+
|
| 201 |
+
assert self.hidden_size % self.num_heads == 0, "hidden_size must be divisible by num_heads"
|
| 202 |
+
assert self.num_heads % self.num_kv_heads == 0, "num_heads must be divisible by num_kv_heads"
|
| 203 |
+
|
| 204 |
+
# Projections (no bias in any linear layers)
|
| 205 |
+
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
|
| 206 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
|
| 207 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_kv_heads * self.head_dim, bias=False)
|
| 208 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
| 209 |
+
|
| 210 |
+
# Rotary embeddings
|
| 211 |
+
self.rotary_emb = RotaryEmbedding(
|
| 212 |
+
self.head_dim,
|
| 213 |
+
max_position_embeddings=config.max_position_embeddings,
|
| 214 |
+
base=config.rope_theta
|
| 215 |
+
)
|
| 216 |
+
|
| 217 |
+
def forward(self, hidden_states, attention_mask=None, position_ids=None):
|
| 218 |
+
"""
|
| 219 |
+
Forward pass of grouped query attention
|
| 220 |
+
|
| 221 |
+
Args:
|
| 222 |
+
hidden_states (torch.Tensor): Input [batch, seq_len, hidden_size]
|
| 223 |
+
attention_mask (torch.Tensor, optional): Attention mask
|
| 224 |
+
position_ids (torch.Tensor, optional): Position indices
|
| 225 |
+
|
| 226 |
+
Returns:
|
| 227 |
+
torch.Tensor: Output [batch, seq_len, hidden_size]
|
| 228 |
+
"""
|
| 229 |
+
batch_size, seq_len, _ = hidden_states.size()
|
| 230 |
+
|
| 231 |
+
# Create position IDs if not provided
|
| 232 |
+
if position_ids is None:
|
| 233 |
+
position_ids = torch.arange(seq_len, device=hidden_states.device)
|
| 234 |
+
|
| 235 |
+
# Q, K, V projections
|
| 236 |
+
query_states = self.q_proj(hidden_states) # [batch, seq_len, 576]
|
| 237 |
+
key_states = self.k_proj(hidden_states) # [batch, seq_len, 192]
|
| 238 |
+
value_states = self.v_proj(hidden_states) # [batch, seq_len, 192]
|
| 239 |
+
|
| 240 |
+
# Reshape to separate heads
|
| 241 |
+
# Q: [batch, seq_len, 9, 64] -> [batch, 9, seq_len, 64]
|
| 242 |
+
query_states = query_states.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 243 |
+
# K, V: [batch, seq_len, 3, 64] -> [batch, 3, seq_len, 64]
|
| 244 |
+
key_states = key_states.view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
|
| 245 |
+
value_states = value_states.view(batch_size, seq_len, self.num_kv_heads, self.head_dim).transpose(1, 2)
|
| 246 |
+
|
| 247 |
+
# Apply RoPE to Q and K
|
| 248 |
+
cos, sin = self.rotary_emb(value_states, position_ids)
|
| 249 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
|
| 250 |
+
|
| 251 |
+
# Repeat K and V for GQA (3 KV heads -> 9 to match Q heads)
|
| 252 |
+
# Each KV head is repeated 3 times: [batch, 3, seq, 64] -> [batch, 9, seq, 64]
|
| 253 |
+
key_states = key_states.repeat_interleave(self.num_kv_groups, dim=1)
|
| 254 |
+
value_states = value_states.repeat_interleave(self.num_kv_groups, dim=1)
|
| 255 |
+
|
| 256 |
+
# Scaled dot-product attention (PyTorch 2.0+ optimized)
|
| 257 |
+
# Equivalent to ~80% of Flash Attention performance
|
| 258 |
+
attn_output = F.scaled_dot_product_attention(
|
| 259 |
+
query_states,
|
| 260 |
+
key_states,
|
| 261 |
+
value_states,
|
| 262 |
+
attn_mask=attention_mask,
|
| 263 |
+
dropout_p=0.0,
|
| 264 |
+
is_causal=True # Causal masking for autoregressive generation
|
| 265 |
+
)
|
| 266 |
+
|
| 267 |
+
# Reshape back: [batch, 9, seq_len, 64] -> [batch, seq_len, 576]
|
| 268 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 269 |
+
attn_output = attn_output.view(batch_size, seq_len, self.hidden_size)
|
| 270 |
+
|
| 271 |
+
# Output projection
|
| 272 |
+
attn_output = self.o_proj(attn_output)
|
| 273 |
+
|
| 274 |
+
return attn_output
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
class SwiGLU_FFN(nn.Module):
|
| 278 |
+
"""
|
| 279 |
+
SwiGLU Feed-Forward Network
|
| 280 |
+
|
| 281 |
+
Uses Swish-Gated Linear Units instead of standard FFN.
|
| 282 |
+
Formula: FFN(x) = down_proj(SiLU(gate_proj(x)) ⊙ up_proj(x))
|
| 283 |
+
|
| 284 |
+
Key differences from standard FFN:
|
| 285 |
+
- 3 linear projections instead of 2 (gate, up, down)
|
| 286 |
+
- Element-wise gating mechanism (⊙)
|
| 287 |
+
- 50% more parameters but better performance
|
| 288 |
+
- Used in Llama, PaLM, and most modern LLMs
|
| 289 |
+
"""
|
| 290 |
+
|
| 291 |
+
def __init__(self, config):
|
| 292 |
+
"""
|
| 293 |
+
Args:
|
| 294 |
+
config: Model configuration with attributes:
|
| 295 |
+
- hidden_size: Model dimension (576)
|
| 296 |
+
- intermediate_size: FFN intermediate dimension (1536)
|
| 297 |
+
"""
|
| 298 |
+
super().__init__()
|
| 299 |
+
self.hidden_size = config.hidden_size # 576
|
| 300 |
+
self.intermediate_size = config.intermediate_size # 1536
|
| 301 |
+
|
| 302 |
+
# Three projections (no bias)
|
| 303 |
+
self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 304 |
+
self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
|
| 305 |
+
self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
|
| 306 |
+
|
| 307 |
+
# Swish/SiLU activation
|
| 308 |
+
self.act_fn = nn.SiLU()
|
| 309 |
+
|
| 310 |
+
def forward(self, x):
|
| 311 |
+
"""
|
| 312 |
+
Forward pass: down(SiLU(gate) * up)
|
| 313 |
+
|
| 314 |
+
Args:
|
| 315 |
+
x (torch.Tensor): Input [batch, seq_len, hidden_size]
|
| 316 |
+
|
| 317 |
+
Returns:
|
| 318 |
+
torch.Tensor: Output [batch, seq_len, hidden_size]
|
| 319 |
+
"""
|
| 320 |
+
# Gate path: apply SiLU activation
|
| 321 |
+
gate = self.act_fn(self.gate_proj(x))
|
| 322 |
+
|
| 323 |
+
# Up path: linear transformation
|
| 324 |
+
up = self.up_proj(x)
|
| 325 |
+
|
| 326 |
+
# Element-wise multiplication (gating)
|
| 327 |
+
gated = gate * up
|
| 328 |
+
|
| 329 |
+
# Down projection
|
| 330 |
+
return self.down_proj(gated)
|
| 331 |
+
|
| 332 |
+
|
| 333 |
+
class TransformerBlock(nn.Module):
|
| 334 |
+
"""
|
| 335 |
+
Complete Transformer Block with Pre-Norm Architecture
|
| 336 |
+
|
| 337 |
+
Architecture:
|
| 338 |
+
1. x -> RMSNorm -> Attention -> Add residual
|
| 339 |
+
2. x -> RMSNorm -> FFN -> Add residual
|
| 340 |
+
|
| 341 |
+
Pre-norm (norm before sublayer) is standard in modern transformers
|
| 342 |
+
as it provides better gradient flow in deep networks.
|
| 343 |
+
"""
|
| 344 |
+
|
| 345 |
+
def __init__(self, config):
|
| 346 |
+
"""
|
| 347 |
+
Args:
|
| 348 |
+
config: Model configuration
|
| 349 |
+
"""
|
| 350 |
+
super().__init__()
|
| 351 |
+
|
| 352 |
+
# Layer normalization (pre-norm)
|
| 353 |
+
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 354 |
+
|
| 355 |
+
# Self-attention
|
| 356 |
+
self.self_attn = GroupedQueryAttention(config)
|
| 357 |
+
|
| 358 |
+
# Post-attention layer norm
|
| 359 |
+
self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 360 |
+
|
| 361 |
+
# Feed-forward network
|
| 362 |
+
self.mlp = SwiGLU_FFN(config)
|
| 363 |
+
|
| 364 |
+
def forward(self, hidden_states, attention_mask=None, position_ids=None):
|
| 365 |
+
"""
|
| 366 |
+
Forward pass through transformer block
|
| 367 |
+
|
| 368 |
+
Args:
|
| 369 |
+
hidden_states (torch.Tensor): Input [batch, seq_len, hidden_size]
|
| 370 |
+
attention_mask (torch.Tensor, optional): Attention mask
|
| 371 |
+
position_ids (torch.Tensor, optional): Position indices
|
| 372 |
+
|
| 373 |
+
Returns:
|
| 374 |
+
torch.Tensor: Output [batch, seq_len, hidden_size]
|
| 375 |
+
"""
|
| 376 |
+
# Self-attention with residual connection
|
| 377 |
+
residual = hidden_states
|
| 378 |
+
hidden_states = self.input_layernorm(hidden_states)
|
| 379 |
+
hidden_states = self.self_attn(hidden_states, attention_mask, position_ids)
|
| 380 |
+
hidden_states = residual + hidden_states
|
| 381 |
+
|
| 382 |
+
# FFN with residual connection
|
| 383 |
+
residual = hidden_states
|
| 384 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 385 |
+
hidden_states = self.mlp(hidden_states)
|
| 386 |
+
hidden_states = residual + hidden_states
|
| 387 |
+
|
| 388 |
+
return hidden_states
|