File size: 8,664 Bytes
f42f624 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 |
#! /usr/bin/env python
"""
SmollmV2 model implementation
Author: Shilpaj Bhalerao
Date: 2025-01-19
"""
# Third-Party Imports
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
# Local Imports
from config import SmollmConfig, RoPEConfig
class RoPEAttention:
"""
Rotary Position Embedding attention with support for different Q/K dimensions
"""
def __init__(self, head_dim, kv_dim, base=RoPEConfig.base):
"""
Initialize rotary embeddings
Args:
head_dim: Dimension of query head
kv_dim: Dimension of key/value head
base: Base for the angle calculations (default: 10000)
"""
super().__init__()
# Generate theta parameter for rotary embeddings for both Q and K dimensions
inv_freq_k = 1.0 / (base ** (torch.arange(0, kv_dim, 2).float() / kv_dim))
self.register_buffer('inv_freq_k', inv_freq_k)
self.head_dim = head_dim
self.kv_dim = kv_dim
self.seq_len_cached = None
self.cos_cached = None
self.sin_cached = None
def _update_cos_sin_cache(self, x, seq_len):
"""Update cached cos and sin values for given sequence length"""
if seq_len != self.seq_len_cached:
self.seq_len_cached = seq_len
t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq_k)
freqs = torch.einsum('i,j->ij', t, self.inv_freq_k)
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
self.cos_cached = emb.cos()[None, None, :, :]
self.sin_cached = emb.sin()[None, None, :, :]
def _rotate_half(self, x):
"""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 __call__(self, q, k):
"""
Apply rotary embeddings to input queries and keys
Args:
q: Query tensor of shape (batch, n_head, seq_len, head_dim)
k: Key tensor of shape (batch, n_head, seq_len, kv_dim)
Returns:
q_rot: Rotated query tensor
k_rot: Rotated key tensor
"""
seq_len = q.shape[2]
self._update_cos_sin_cache(k, seq_len)
# Apply rotary embeddings to keys
k_cos = self.cos_cached[..., :self.kv_dim]
k_sin = self.sin_cached[..., :self.kv_dim]
k_rot = (k * k_cos) + (self._rotate_half(k) * k_sin)
# For queries, we only apply rotation to the part that interacts with keys
q_part = q[..., :self.kv_dim]
q_cos = self.cos_cached[..., :self.kv_dim]
q_sin = self.sin_cached[..., :self.kv_dim]
q_rot_part = (q_part * q_cos) + (self._rotate_half(q_part) * q_sin)
# Combine rotated part with unrotated parts for query
q_rot = torch.cat([q_rot_part, q[..., self.kv_dim:]], dim=-1)
return q_rot, k_rot
def register_buffer(self, name, tensor):
"""Helper function to register a buffer"""
setattr(self, name, tensor)
class CausalSelfAttention(nn.Module):
"""
Causal self-attention mechanism with reduced KV dimensions and RoPE
"""
def __init__(self, config):
super().__init__()
assert config.n_embd % config.n_head == 0
# Calculate dimensions
self.head_dim = config.n_embd // config.n_head # 576/9 = 64
self.n_head = config.n_head
self.n_embd = config.n_embd
# Make kv_dim divisible by n_head (189 is closest to 192 that's divisible by 9)
self.kv_dim = 189 # 189 = 9 * 21, closest to 192 that's divisible by 9
self.kv_dim_per_head = self.kv_dim // self.n_head # 21
# Separate projections with reduced dimensions for k,v
self.q_proj = nn.Linear(config.n_embd, config.n_embd, bias=False)
self.k_proj = nn.Linear(config.n_embd, self.kv_dim, bias=False) # 189 dimensions
self.v_proj = nn.Linear(config.n_embd, self.kv_dim, bias=False) # 189 dimensions
# output projection
self.o_proj = nn.Linear(config.n_embd, config.n_embd, bias=False)
# rotary embeddings
self.rope = RoPEAttention(self.head_dim, self.kv_dim_per_head)
def forward(self, x):
B, T, C = x.size()
# calculate query, key, values
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
# reshape with exact dimensions
q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.n_head, self.kv_dim_per_head).transpose(1, 2)
v = v.view(B, T, self.n_head, self.kv_dim_per_head).transpose(1, 2)
# apply rotary embeddings
q, k = self.rope(q, k)
# pad k and v to match q dimension for attention
k_pad = torch.zeros_like(q)
v_pad = torch.zeros_like(q)
k_pad[..., :self.kv_dim_per_head] = k
v_pad[..., :self.kv_dim_per_head] = v
# flash attention
y = F.scaled_dot_product_attention(q, k_pad, v_pad, is_causal=True)
# reshape back
y = y.transpose(1, 2).contiguous().view(B, T, C)
# output projection
y = self.o_proj(y)
return y
class MLP(nn.Module):
"""
MLP (Multi-Layer Perceptron) layer with gate/up/down projection structure
"""
def __init__(self, config):
super().__init__()
hidden_dim = int(config.n_embd * config.mlp_ratio) - 1
self.gate_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
self.up_proj = nn.Linear(config.n_embd, hidden_dim, bias=False)
self.down_proj = nn.Linear(hidden_dim, config.n_embd, bias=False)
self.down_proj.NANOGPT_SCALE_INIT = 1
def forward(self, x):
# SwiGLU activation as used in PaLM, Llama, etc.
gate = self.gate_proj(x)
up = self.up_proj(x)
x = F.silu(gate) * up
x = self.down_proj(x)
return x
class Block(nn.Module):
"""
Transformer block
"""
def __init__(self, config):
super().__init__()
self.ln_1 = nn.LayerNorm(config.n_embd, bias=False)
self.attn = CausalSelfAttention(config)
self.ln_2 = nn.LayerNorm(config.n_embd, bias=False)
self.mlp = MLP(config)
def forward(self, x):
x = x + self.attn(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class SmollmV2(nn.Module):
"""
SmollmV2 model
"""
def __init__(self, config=SmollmConfig()):
super().__init__()
self.config = config
self.transformer = nn.ModuleDict(dict(
wte = nn.Embedding(config.vocab_size, config.n_embd),
h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
ln_f = nn.LayerNorm(config.n_embd, bias=False),
))
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
# weight sharing
self.transformer.wte.weight = self.lm_head.weight
# weight initialization
self.apply(self._init_weights)
# Compile the model if torch version supports it
if hasattr(torch, 'compile'):
self.forward = torch.compile(self.forward)
def _init_weights(self, module):
if isinstance(module, nn.Linear):
std = 0.02
if hasattr(module, 'NANGPT_SCALE_INIT'):
std *= (2 * self.config.n_layer) ** -0.5
torch.nn.init.normal_(module.weight, mean = 0.0, std = std)
if module.bias is not None:
torch.nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
torch.nn.init.normal_(module.weight, mean=0.0, std = 0.04)
def forward(self, idx, targets=None):
# idx is of shape (B, T)
B, T = idx.size()
assert T <= self.config.block_size, f"Cannot forward sequence of length {T}, block size is only {self.config.block_size}"
# forward the token and posisition embeddings
tok_emb = self.transformer.wte(idx) # token embeddings of shape (B, T, n_embd)
x = tok_emb
# forward the blocks of the transformer
for block in self.transformer.h:
x = block(x)
# forward the final layernorm and the classifier
x = self.transformer.ln_f(x)
logits = self.lm_head(x) # (B, T, vocab_size)
loss = None
if targets is not None:
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1))
return logits, loss
|