File size: 11,477 Bytes
fd448dd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 | """Standard Transformer language model implementation."""
import math
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from taoTrain.core import BaseModel
from taoTrain.config import ModelConfig
from .registry import register_architecture
# ============================================================================
# Components
# ============================================================================
class PositionalEmbedding(nn.Module):
"""Sinusoidal positional embeddings."""
def __init__(self, dim: int, max_seq_length: int = 2048):
"""Initialize positional embeddings."""
super().__init__()
self.dim = dim
self.max_seq_length = max_seq_length
# Precompute positional embeddings
pe = torch.zeros(max_seq_length, dim)
pos = torch.arange(0, max_seq_length, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, dim, 2).float() * (-math.log(10000.0) / dim))
pe[:, 0::2] = torch.sin(pos * div_term)
if dim % 2 == 1:
pe[:, 1::2] = torch.cos(pos * div_term[:-1])
else:
pe[:, 1::2] = torch.cos(pos * div_term)
self.register_buffer("pe", pe, persistent=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Add positional embeddings to input.
Args:
x: Input tensor (batch, seq_len, hidden_dim)
Returns:
Input + positional embeddings
"""
seq_len = x.shape[1]
return x + self.pe[:seq_len]
class Attention(nn.Module):
"""Multi-head self-attention using scaled dot-product attention."""
def __init__(self, config: ModelConfig):
"""Initialize attention."""
super().__init__()
self.hidden_dim = config.hidden_dim
self.num_heads = config.num_heads
self.head_dim = config.head_dim
assert self.hidden_dim % self.num_heads == 0
# Linear projections
self.q_proj = nn.Linear(self.hidden_dim, self.hidden_dim)
self.k_proj = nn.Linear(self.hidden_dim, self.hidden_dim)
self.v_proj = nn.Linear(self.hidden_dim, self.hidden_dim)
self.out_proj = nn.Linear(self.hidden_dim, self.hidden_dim)
self.dropout_p = config.dropout
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Forward pass using scaled_dot_product_attention.
Args:
x: Shape (batch, seq_len, hidden_dim)
attention_mask: Shape (batch, seq_len)
Returns:
Output: Shape (batch, seq_len, hidden_dim)
"""
batch_size, seq_len, _ = x.shape
# Project to Q, K, V
q = self.q_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim)
k = self.k_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim)
v = self.v_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim)
# Transpose for attention: (batch, num_heads, seq_len, head_dim)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
# NOTE: PyTorch's scaled_dot_product_attention does NOT support both
# explicit attn_mask AND is_causal=True together.
# When is_causal=True, PyTorch handles causal masking automatically.
# Padding positions are handled separately via loss computation (labels=-100).
# See: https://github.com/pytorch/pytorch/issues/96099
# Compute attention using scaled_dot_product_attention
# is_causal=True automatically applies causal masking
# We do NOT pass attn_mask when is_causal=True
out = F.scaled_dot_product_attention(
q, k, v,
attn_mask=None, # Must be None when is_causal=True
dropout_p=self.dropout_p if self.training else 0.0,
is_causal=True,
scale=None # Uses default scale of 1/sqrt(head_dim)
) # (batch, num_heads, seq_len, head_dim)
# Transpose back and reshape
out = out.transpose(1, 2).contiguous() # (batch, seq_len, num_heads, head_dim)
out = out.reshape(batch_size, seq_len, self.hidden_dim)
# Output projection
out = self.out_proj(out)
return out
class SwiGLU(nn.Module):
"""Swish Gated Linear Unit activation."""
def __init__(self, in_dim: int, out_dim: int, dropout: float = 0.0):
"""
Initialize SwiGLU.
Args:
in_dim: Input dimension
out_dim: Intermediate/hidden dimension
dropout: Dropout rate
"""
super().__init__()
# Project to 2x the intermediate dimension (for value and gate)
self.fc1 = nn.Linear(in_dim, 2 * out_dim)
self.fc2 = nn.Linear(out_dim, in_dim) # Project back to input dimension
self.dropout = nn.Dropout(dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Forward pass with SwiGLU activation.
Args:
x: Input tensor
Returns:
Gated activation output (same dimension as input)
"""
# Project to 2x intermediate dimension
x = self.fc1(x)
# Split into value and gate
x, gate = x.chunk(2, dim=-1)
# SwiGLU: value * swish(gate) = value * gate * sigmoid(gate)
x = x * F.silu(gate) # SiLU is Swish: x * sigmoid(x)
x = self.dropout(x)
x = self.fc2(x) # Project back to input dimension
return x
class FeedForward(nn.Module):
"""Feed-forward network with SwiGLU activation."""
def __init__(self, config: ModelConfig):
"""Initialize FFN with SwiGLU."""
super().__init__()
self.swiglu = SwiGLU(
in_dim=config.hidden_dim,
out_dim=config.intermediate_dim,
dropout=config.dropout
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""Forward pass with SwiGLU activation."""
return self.swiglu(x)
class TransformerBlock(nn.Module):
"""Single transformer block with attention and FFN."""
def __init__(self, config: ModelConfig):
"""Initialize transformer block."""
super().__init__()
self.norm1 = nn.LayerNorm(config.hidden_dim)
self.attn = Attention(config)
self.norm2 = nn.LayerNorm(config.hidden_dim)
self.ffn = FeedForward(config)
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Forward pass with pre-norm residual connections."""
# Attention with residual
x = x + self.attn(self.norm1(x), attention_mask=attention_mask)
# FFN with residual
x = x + self.ffn(self.norm2(x))
return x
# ============================================================================
# Transformer LM
# ============================================================================
@register_architecture("transformer")
class TransformerLM(BaseModel):
"""Standard Transformer language model."""
def __init__(self, config: ModelConfig):
"""Initialize Transformer LM."""
super().__init__(config)
# Embeddings
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_dim)
self.pos_embed = PositionalEmbedding(config.hidden_dim, max_seq_length=config.max_seq_length)
self.dropout = nn.Dropout(config.dropout)
# Transformer blocks
self.blocks = nn.ModuleList([
TransformerBlock(config) for _ in range(config.num_layers)
])
# Final layer norm
self.final_norm = nn.LayerNorm(config.hidden_dim)
# Output projection (shared with input embeddings for efficiency)
self.lm_head = nn.Linear(config.hidden_dim, config.vocab_size, bias=False)
# Weight tying (optional)
self.lm_head.weight = self.embed_tokens.weight
# Initialize weights
self._init_weights()
def _init_weights(self):
"""Initialize model weights."""
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, std=self.config.init_std)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, std=self.config.init_std)
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
pixel_values: Optional[torch.Tensor] = None,
) -> dict[str, torch.Tensor]:
"""
Forward pass.
Args:
input_ids: (batch_size, seq_len)
attention_mask: (batch_size, seq_len)
labels: (batch_size, seq_len) for loss computation
Returns:
Dict with 'logits' and optionally 'loss'
"""
if inputs_embeds is None:
if input_ids is None:
raise ValueError("Either input_ids or inputs_embeds must be provided")
batch_size, seq_len = input_ids.shape
x = self.embed_tokens(input_ids)
else:
batch_size, seq_len, _ = inputs_embeds.shape
x = inputs_embeds
# Add positional embeddings
x = self.pos_embed(x)
x = self.dropout(x)
# Transformer blocks
for block in self.blocks:
x = block(x, attention_mask=attention_mask)
# Final normalization
x = self.final_norm(x)
# LM head
logits = self.lm_head(x) # (batch, seq_len, vocab_size)
# Loss computation
loss = None
if labels is not None:
# Flatten for loss computation
logits_flat = logits.view(-1, logits.size(-1)) # (batch * seq_len, vocab_size)
labels_flat = labels.view(-1)
valid_label_mask = labels_flat != -100
if not torch.any(valid_label_mask):
raise ValueError(
"All labels are masked out (-100), so loss cannot be computed. "
"This usually indicates a dataset parsing or masking bug."
)
# Only compute loss on valid targets (ignore -100 tokens)
loss = F.cross_entropy(
logits_flat,
labels_flat,
reduction='mean',
ignore_index=-100
)
return {
'logits': logits,
'loss': loss,
}
|