File size: 10,669 Bytes
6550ac5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 | import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
# YaRN Rotary Position Embedding
class YaRNRoPE(nn.Module):
def __init__(
self,
head_dim: int,
original_max_seq_len: int = 4096,
factor: float = 1.0,
base: float = 10000.0,
beta_fast: int = 32,
beta_slow: int = 1,
):
super().__init__()
self.head_dim = head_dim
self.original_max_seq_len = original_max_seq_len
self.factor = factor
if factor > 1.0:
self.attention_factor = math.log(factor) * 0.1 + 1.0
t = torch.arange(head_dim // 2)
inv_freq = 1.0 / (base ** (2 * t.float() / head_dim))
wavelength = 2 * math.pi / inv_freq
low_freq_wavelen = original_max_seq_len / beta_slow
high_freq_wavelen = original_max_seq_len / beta_fast
ratio = (wavelength - high_freq_wavelen) / (low_freq_wavelen - high_freq_wavelen)
ratio = torch.clamp(ratio, 0.0, 1.0)
scale = 1 - ratio + ratio * factor
inv_freq = inv_freq / scale
else:
self.attention_factor = 1.0
inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim))
self.register_buffer("inv_freq", inv_freq)
self._set_cos_sin_cache(int(original_max_seq_len * factor))
def _set_cos_sin_cache(self, seq_len: int):
t = torch.arange(seq_len, device=self.inv_freq.device)
freqs = torch.outer(t, self.inv_freq)
emb = torch.cat((freqs, freqs), dim=-1)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :], persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :], persistent=False)
self.max_seq_len_cached = seq_len
def forward(self, x: torch.Tensor, seq_len: Optional[int] = None):
if seq_len is None:
seq_len = x.shape[-2]
if seq_len > self.max_seq_len_cached:
self._set_cos_sin_cache(seq_len)
cos = self.cos_cached[:, :, :seq_len, :]
sin = self.sin_cached[:, :, :seq_len, :]
x1, x2 = x[..., ::2], x[..., 1::2]
rotated = torch.stack(
[
x1 * cos[..., ::2] - x2 * sin[..., ::2],
x1 * sin[..., ::2] + x2 * cos[..., ::2],
],
dim=-1,
).flatten(-2)
return rotated * self.attention_factor
# Scaled Dot-Product Attention (with GQA support)
def scaled_dot_product_attention(
query: torch.Tensor,
key: torch.Tensor,
value: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
dropout: float = 0.0,
is_causal: bool = False,
scale: Optional[float] = None,
enable_gqa: bool = False,
) -> torch.Tensor:
B, Hq, L, E = query.shape
_, Hkv, S, _ = key.shape
if enable_gqa and Hq != Hkv:
assert Hq % Hkv == 0
n_rep = Hq // Hkv
key = key.unsqueeze(2).repeat(1, 1, n_rep, 1, 1).flatten(1, 2)
value = value.unsqueeze(2).repeat(1, 1, n_rep, 1, 1).flatten(1, 2)
if scale is None:
scale = E ** -0.5
scores = torch.matmul(query, key.transpose(-2, -1)) * scale
if is_causal and attention_mask is not None:
raise RuntimeError("is_causal and attention_mask cannot be set at the same time")
if is_causal:
causal_mask = torch.triu(torch.ones(L, S, dtype=torch.bool, device=query.device), diagonal=1)
scores = scores.masked_fill(causal_mask, float("-inf"))
if attention_mask is not None:
if attention_mask.dtype == torch.bool:
scores = scores.masked_fill(~attention_mask, float("-inf"))
else:
scores = scores + attention_mask
attn_weights = F.softmax(scores, dim=-1)
if dropout > 0.0:
attn_weights = F.dropout(attn_weights, p=dropout, training=True)
output = torch.matmul(attn_weights, value)
return output
# Grouped Query Attention
class GroupedQueryAttention(nn.Module):
def __init__(
self,
hidden_size: int,
num_heads: int,
num_key_value_heads: int,
head_dim: int,
max_seq_len: int,
):
super().__init__()
self.hidden_size = hidden_size
self.num_heads = num_heads
self.num_key_value_heads = num_key_value_heads
self.head_dim = head_dim
self.max_seq_len = max_seq_len
self.q_proj = nn.Linear(hidden_size, head_dim * num_heads, bias=False)
self.k_proj = nn.Linear(hidden_size, head_dim * num_key_value_heads, bias=False)
self.v_proj = nn.Linear(hidden_size, head_dim * num_key_value_heads, bias=False)
self.out_proj = nn.Linear(num_heads * head_dim, hidden_size, bias=False)
self.rope = YaRNRoPE(
head_dim=head_dim,
original_max_seq_len=max_seq_len,
factor=16.0,
)
def forward(self, query, key, value):
B, L_q, _ = query.size()
_, L_kv, _ = key.size()
q = self.q_proj(query).view(B, L_q, self.num_heads, self.head_dim).transpose(1, 2)
k = self.k_proj(key).view(B, L_kv, self.num_key_value_heads, self.head_dim).transpose(1, 2)
v = self.v_proj(value).view(B, L_kv, self.num_key_value_heads, self.head_dim).transpose(1, 2)
q_embed = self.rope(q)
k_embed = self.rope(k)
q_embed, k_embed = q_embed.to(q.dtype), k_embed.to(k.dtype)
attn_output = scaled_dot_product_attention(
q_embed, k_embed, v,
attention_mask=torch.ones(L_q, L_kv, dtype=torch.bool, device=q_embed.device),
dropout=0.0,
is_causal=False,
enable_gqa=True,
)
context = attn_output.transpose(1, 2).contiguous().view(B, L_q, self.num_heads * self.head_dim)
output = self.out_proj(context)
return output
# Gated GELU Feed-Forward Network
class GEGLU(nn.Module):
def __init__(self, hidden_size: int, intermediate_size: Optional[int] = None):
super().__init__()
if intermediate_size is None:
intermediate_size = int(8 / 3 * hidden_size)
self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
def forward(self, x):
gate = F.gelu(self.gate_proj(x))
value = self.up_proj(x)
hidden = gate * value
return self.down_proj(hidden)
# Transformer Decoder Layer
class TransformerDecoderLayer(nn.Module):
def __init__(
self,
hidden_size: int,
num_heads: int,
num_key_value_heads: int,
intermediate_size: int,
head_dim: int,
max_seq_len: int,
dropout: float,
):
super().__init__()
self.self_attn = GroupedQueryAttention(
hidden_size, num_heads, num_key_value_heads, head_dim, max_seq_len
)
self.ffn = GEGLU(hidden_size, intermediate_size)
self.input_layernorm = nn.RMSNorm(hidden_size)
self.post_attention_layernorm = nn.RMSNorm(hidden_size)
self.dropout = nn.Dropout(dropout)
def forward(self, hidden_states):
residual = hidden_states
hidden_states = self.input_layernorm(hidden_states)
attn_output = self.dropout(self.self_attn(hidden_states, hidden_states, hidden_states))
hidden_states = residual + attn_output
residual = hidden_states
hidden_states = self.post_attention_layernorm(hidden_states)
ffn_output = self.dropout(self.ffn(hidden_states))
hidden_states = residual + ffn_output
return hidden_states
# Decoder with Dense Layer Connections
class TransformerDecoder(nn.Module):
def __init__(
self,
hidden_size: int,
num_heads: int,
num_key_value_heads: int,
intermediate_size: int,
head_dim: int,
num_layers: int,
max_seq_len: int,
dropout: float,
):
super().__init__()
self.num_layers = num_layers
self.layers = nn.ModuleList([
TransformerDecoderLayer(
hidden_size, num_heads, num_key_value_heads,
intermediate_size, head_dim, max_seq_len, dropout
)
for _ in range(num_layers)
])
mask = torch.tril(torch.ones(num_layers, num_layers), diagonal=-1)
self.register_buffer("layer_weight_mask", mask)
self.layer_raw_weights = nn.Parameter(torch.randn(num_layers, num_layers) / 10)
def forward(self, hidden_states):
history = []
for idx_layer, layer in enumerate(self.layers):
layer_output = layer(hidden_states)
if history:
raw_weights = self.layer_raw_weights[idx_layer, :idx_layer]
masked_weights = raw_weights * self.layer_weight_mask[idx_layer, :idx_layer]
weights = F.softmax(masked_weights, dim=0)
hist_stack = torch.stack(history, dim=0)
residual = torch.einsum("lbtd,l->btd", hist_stack, weights)
hidden_states = layer_output + residual
else:
hidden_states = layer_output
history.append(hidden_states)
return hidden_states
# Classifier
class Classifier(nn.Module):
def __init__(
self,
hidden_size: int,
num_heads: int,
num_key_value_heads: int,
intermediate_size: int,
head_dim: int,
vocab_size: int,
num_layers: int,
max_seq_len: int,
dropout: float,
):
super().__init__()
self.token_embedding = nn.Embedding(vocab_size, hidden_size)
self.decoder = TransformerDecoder(
hidden_size, num_heads, num_key_value_heads,
intermediate_size, head_dim, num_layers, max_seq_len, dropout
)
self.final_layernorm = nn.RMSNorm(hidden_size)
self.lm_head = nn.Linear(hidden_size, 6, bias=False)
def forward(self, input_ids):
hidden_states = self.token_embedding(input_ids)
hidden_states = self.decoder(hidden_states)
hidden_states = self.final_layernorm(hidden_states)
logits = self.lm_head(hidden_states).mean(-2)
return logits
|