File size: 9,174 Bytes
3da2ee9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 | """
Model components optimized for CPU training.
Design rationale:
- RMSNorm instead of LayerNorm: simpler, faster (no mean computation)
- Rotary Position Embeddings (RoPE): no learned position embeddings needed,
saves parameters and generalizes better
- LoRA-style low-rank linear layers: dramatically reduces parameter count
while maintaining expressiveness
- All operations use float32 for CPU stability (no mixed precision)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from typing import Optional, Tuple
class RMSNorm(nn.Module):
"""
Root Mean Square normalization.
Why: ~30% faster than LayerNorm on CPU since it skips mean computation.
Empirically equivalent performance for transformers.
"""
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.eps = eps
self.weight = nn.Parameter(torch.ones(dim))
def forward(self, x: torch.Tensor) -> torch.Tensor:
norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
return x * norm * self.weight
class RotaryEmbedding(nn.Module):
"""
Rotary Position Embedding (RoPE).
Why:
- No learned parameters (saves memory)
- Relative position awareness without extra params
- Extrapolates better to unseen sequence lengths
- Computationally efficient on CPU (just sin/cos)
"""
def __init__(self, dim: int, max_seq_len: int = 512, base: float = 10000.0):
super().__init__()
self.dim = dim
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
self.register_buffer('inv_freq', inv_freq)
# Pre-compute for max_seq_len to avoid recomputation
self._build_cache(max_seq_len)
def _build_cache(self, seq_len: int):
t = torch.arange(seq_len, dtype=self.inv_freq.dtype)
freqs = torch.einsum('i,j->ij', t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer('cos_cached', emb.cos())
self.register_buffer('sin_cached', emb.sin())
def forward(self, seq_len: int) -> Tuple[torch.Tensor, torch.Tensor]:
if seq_len > self.cos_cached.size(0):
self._build_cache(seq_len)
return self.cos_cached[:seq_len], self.sin_cached[:seq_len]
def rotate_half(x: torch.Tensor) -> torch.Tensor:
"""Rotate 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)
def apply_rotary_pos_emb(q: torch.Tensor, k: torch.Tensor,
cos: torch.Tensor, sin: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""Apply rotary embeddings to queries and keys."""
# cos, sin: [seq_len, dim]
cos = cos.unsqueeze(0).unsqueeze(0) # [1, 1, seq_len, dim]
sin = sin.unsqueeze(0).unsqueeze(0)
q_embed = (q * cos) + (rotate_half(q) * sin)
k_embed = (k * cos) + (rotate_half(k) * sin)
return q_embed, k_embed
class LoRALinear(nn.Module):
"""
Low-Rank Adaptation linear layer.
Why: Instead of full d_in x d_out matrix, uses two smaller matrices:
d_in x rank + rank x d_out. For rank=16, d_in=d_out=256:
Full: 65,536 params
LoRA: 256*16 + 16*256 = 8,192 params (8x reduction!)
Still maintains good expressiveness for the tasks we need.
"""
def __init__(self, in_features: int, out_features: int, rank: int = 16, bias: bool = False):
super().__init__()
self.rank = rank
# If rank is large enough, just use full linear
if rank >= min(in_features, out_features) // 2:
self.use_lora = False
self.linear = nn.Linear(in_features, out_features, bias=bias)
else:
self.use_lora = True
self.down = nn.Linear(in_features, rank, bias=False)
self.up = nn.Linear(rank, out_features, bias=bias)
# Initialize to approximate identity-like behavior
nn.init.kaiming_uniform_(self.down.weight, a=math.sqrt(5))
nn.init.zeros_(self.up.weight)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.use_lora:
return self.up(self.down(x))
return self.linear(x)
class GatedMLP(nn.Module):
"""
SwiGLU-style gated MLP.
Why: Gated activation functions consistently outperform standard ReLU/GELU
in transformers, especially at small scale. The gate provides a learned
"feature selection" mechanism.
Uses LoRA projections to save parameters.
"""
def __init__(self, d_model: int, d_ff: int, rank: int = 16, dropout: float = 0.05):
super().__init__()
self.gate_proj = LoRALinear(d_model, d_ff, rank=rank)
self.up_proj = LoRALinear(d_model, d_ff, rank=rank)
self.down_proj = LoRALinear(d_ff, d_model, rank=rank)
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
gate = F.silu(self.gate_proj(x))
up = self.up_proj(x)
return self.dropout(self.down_proj(gate * up))
class MultiHeadAttention(nn.Module):
"""
Multi-Head Attention with RoPE and optional Grouped Query Attention.
Why these choices:
- Grouped Query Attention (GQA): shares KV heads, reducing memory and params
while maintaining quality. For 8 heads with 4 KV groups: 50% KV param reduction.
- Pre-computed causal mask: avoids recomputing each forward pass on CPU
- RoPE applied per-head: correct relative position encoding
"""
def __init__(self, d_model: int, n_heads: int, rank: int = 16,
dropout: float = 0.05, max_seq_len: int = 512,
n_kv_heads: Optional[int] = None):
super().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.n_kv_heads = n_kv_heads or n_heads
self.head_dim = d_model // n_heads
self.n_rep = n_heads // self.n_kv_heads # repetition factor for GQA
assert d_model % n_heads == 0
self.q_proj = LoRALinear(d_model, d_model, rank=rank)
self.k_proj = LoRALinear(d_model, self.n_kv_heads * self.head_dim, rank=rank)
self.v_proj = LoRALinear(d_model, self.n_kv_heads * self.head_dim, rank=rank)
self.o_proj = LoRALinear(d_model, d_model, rank=rank)
self.dropout = nn.Dropout(dropout)
self.rope = RotaryEmbedding(self.head_dim, max_seq_len)
# Pre-compute causal mask
mask = torch.triu(torch.ones(max_seq_len, max_seq_len), diagonal=1).bool()
self.register_buffer('causal_mask', mask)
def _repeat_kv(self, x: torch.Tensor) -> torch.Tensor:
"""Repeat KV heads to match Q heads for GQA."""
if self.n_rep == 1:
return x
bs, n_kv, seq_len, head_dim = x.shape
x = x[:, :, None, :, :].expand(bs, n_kv, self.n_rep, seq_len, head_dim)
return x.reshape(bs, self.n_heads, seq_len, head_dim)
def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
B, T, C = x.shape
q = self.q_proj(x).view(B, T, self.n_heads, self.head_dim).transpose(1, 2)
k = self.k_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(x).view(B, T, self.n_kv_heads, self.head_dim).transpose(1, 2)
# Apply RoPE
cos, sin = self.rope(T)
q, k = apply_rotary_pos_emb(q, k, cos, sin)
# Expand KV for GQA
k = self._repeat_kv(k)
v = self._repeat_kv(v)
# Attention
scale = math.sqrt(self.head_dim)
attn = torch.matmul(q, k.transpose(-2, -1)) / scale
# Apply causal mask
causal = self.causal_mask[:T, :T].unsqueeze(0).unsqueeze(0)
attn = attn.masked_fill(causal, float('-inf'))
if mask is not None:
# mask shape: [B, T] -> [B, 1, 1, T]
attn = attn.masked_fill(mask.unsqueeze(1).unsqueeze(2), float('-inf'))
attn = F.softmax(attn, dim=-1)
attn = self.dropout(attn)
out = torch.matmul(attn, v)
out = out.transpose(1, 2).contiguous().view(B, T, C)
return self.o_proj(out)
class TransformerBlock(nn.Module):
"""
Single transformer block with pre-norm architecture.
Why pre-norm: More stable training, especially at small scale.
Gradient flow is better since residual path is unimpeded.
"""
def __init__(self, d_model: int, n_heads: int, d_ff: int,
rank: int = 16, dropout: float = 0.05,
max_seq_len: int = 512, n_kv_heads: Optional[int] = None):
super().__init__()
self.attn_norm = RMSNorm(d_model)
self.attn = MultiHeadAttention(d_model, n_heads, rank, dropout, max_seq_len, n_kv_heads)
self.ff_norm = RMSNorm(d_model)
self.ff = GatedMLP(d_model, d_ff, rank, dropout)
def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = x + self.attn(self.attn_norm(x), mask)
x = x + self.ff(self.ff_norm(x))
return x |