File size: 15,509 Bytes
fb67af8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 |
"""Transformer block with pre-normalization architecture and memory optimizations.
Critical implementation details:
1. Pre-normalization: RMSNorm BEFORE attention and FFN
2. Residual connections after each sub-layer
3. Modern component stack: RoPE + RMSNorm + SwiGLU
4. Gradient checkpointing support for memory efficiency
"""
import torch
import torch.nn as nn
from typing import Optional, Tuple, Dict, Any
from torch.utils.checkpoint import checkpoint
from .rmsnorm import RMSNorm
from .attention import MultiHeadAttention
from .swiglu import SwiGLU
class TransformerBlock(nn.Module):
"""Single transformer block with pre-normalization.
This follows the modern architecture used in LLaMA, Mistral, etc:
- Pre-normalization with RMSNorm
- Multi-head attention with RoPE
- SwiGLU activation in FFN
- Residual connections
- Gradient checkpointing support
"""
def __init__(
self,
d_model: int = 768,
n_heads: int = 12,
d_ffn: Optional[int] = None,
dropout: float = 0.1,
max_seq_len: int = 2048,
rope_base: int = 10000,
rope_percentage: float = 0.5,
rms_norm_eps: float = 1e-6,
use_flash_attention: bool = True,
use_gradient_checkpointing: bool = False,
):
super().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.use_gradient_checkpointing = use_gradient_checkpointing
# Pre-normalization layers
self.attn_norm = RMSNorm(d_model, eps=rms_norm_eps)
self.ffn_norm = RMSNorm(d_model, eps=rms_norm_eps)
# Multi-head attention with RoPE
self.attention = MultiHeadAttention(
d_model=d_model,
n_heads=n_heads,
dropout=dropout,
max_seq_len=max_seq_len,
rope_base=rope_base,
rope_percentage=rope_percentage,
use_flash_attention=use_flash_attention,
)
# SwiGLU FFN
# Default hidden dimension: 8/3 * d_model for parameter parity
if d_ffn is None:
d_ffn = int(8 * d_model / 3)
# Round to multiple of 256 for hardware efficiency
d_ffn = 256 * ((d_ffn + 255) // 256)
self.ffn = SwiGLU(
input_dim=d_model,
hidden_dim=d_ffn,
output_dim=d_model,
bias=False,
)
def _attention_block(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
use_cache: bool = False,
past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
"""Attention sub-block with pre-norm."""
# Pre-normalization
x_norm = self.attn_norm(x)
# Multi-head attention
attn_output, kv_cache = self.attention(
x_norm,
attention_mask=attention_mask,
position_ids=position_ids,
use_cache=use_cache,
past_kv=past_kv,
)
# Residual connection
return attn_output, kv_cache
def _ffn_block(self, x: torch.Tensor) -> torch.Tensor:
"""Feed-forward sub-block with pre-norm."""
# Pre-normalization
x_norm = self.ffn_norm(x)
# Feed-forward
ffn_output = self.ffn(x_norm)
return ffn_output
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
use_cache: bool = False,
past_kv: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, torch.Tensor]]]:
"""Forward pass of transformer block.
Args:
x: Input tensor [batch, seq_len, d_model]
attention_mask: Optional attention mask
position_ids: Optional position IDs for RoPE
use_cache: Whether to return KV cache
past_kv: Past key-value cache
Returns:
Output tensor and optional KV cache
"""
# Attention block with residual
if self.use_gradient_checkpointing and self.training:
# Use gradient checkpointing to save memory during training
def attention_fn(x_in):
attn_out, _ = self._attention_block(
x_in,
attention_mask=attention_mask,
position_ids=position_ids,
use_cache=False, # Can't use cache with checkpointing
past_kv=None,
)
return attn_out
attn_output = checkpoint(attention_fn, x, use_reentrant=False)
kv_cache = None
else:
attn_output, kv_cache = self._attention_block(
x,
attention_mask=attention_mask,
position_ids=position_ids,
use_cache=use_cache,
past_kv=past_kv,
)
# Add residual for attention
x = x + attn_output
# FFN block with residual
if self.use_gradient_checkpointing and self.training:
# Use gradient checkpointing for FFN as well
ffn_output = checkpoint(self._ffn_block, x, use_reentrant=False)
else:
ffn_output = self._ffn_block(x)
# Add residual for FFN
x = x + ffn_output
return x, kv_cache
class WikiMiniModel(nn.Module):
"""Complete WikiMini 95M language model.
Architecture:
- Token embeddings with weight tying
- Stack of transformer blocks
- Final RMSNorm
- LM head (tied with embeddings)
"""
def __init__(self, config: Dict[str, Any]):
super().__init__()
# Extract config values with defaults
self.vocab_size = config.get('vocab_size', 32000)
self.d_model = config.get('d_model', 768)
self.n_layers = config.get('n_layers', 12)
self.n_heads = config.get('n_heads', 12)
self.d_ffn = config.get('d_ffn', None)
self.max_seq_len = config.get('max_seq_len', 2048)
self.dropout = config.get('dropout', 0.1)
self.rope_percentage = config.get('rope_percentage', 0.5)
self.rope_base = config.get('rope_base', 10000)
self.rms_norm_eps = config.get('rms_norm_eps', 1e-6)
self.tie_embeddings = config.get('tie_embeddings', True)
self.use_flash_attention = config.get('use_flash_attention', True)
self.use_gradient_checkpointing = config.get('gradient_checkpointing', False)
# Token embeddings
self.token_embedding = nn.Embedding(self.vocab_size, self.d_model)
# Transformer blocks
self.blocks = nn.ModuleList([
TransformerBlock(
d_model=self.d_model,
n_heads=self.n_heads,
d_ffn=self.d_ffn,
dropout=self.dropout,
max_seq_len=self.max_seq_len,
rope_base=self.rope_base,
rope_percentage=self.rope_percentage,
rms_norm_eps=self.rms_norm_eps,
use_flash_attention=self.use_flash_attention,
use_gradient_checkpointing=self.use_gradient_checkpointing,
)
for _ in range(self.n_layers)
])
# Final normalization
self.final_norm = RMSNorm(self.d_model, eps=self.rms_norm_eps)
# Language modeling head
self.lm_head = nn.Linear(self.d_model, self.vocab_size, bias=False)
# Weight tying
if self.tie_embeddings:
self.lm_head.weight = self.token_embedding.weight
# Initialize weights
self._init_weights()
def _init_weights(self):
"""Initialize weights with scaled normal distribution."""
for module in self.modules():
if isinstance(module, nn.Linear):
torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
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=0.02)
def enable_gradient_checkpointing(self):
"""Enable gradient checkpointing for all transformer blocks."""
self.use_gradient_checkpointing = True
for block in self.blocks:
block.use_gradient_checkpointing = True
def disable_gradient_checkpointing(self):
"""Disable gradient checkpointing for all transformer blocks."""
self.use_gradient_checkpointing = False
for block in self.blocks:
block.use_gradient_checkpointing = False
def count_parameters(self) -> dict:
"""Count model parameters by component.
Returns:
Dictionary with parameter counts for each component
"""
# Count by component type
embedding_params = sum(p.numel() for p in self.token_embedding.parameters())
attention_params = 0
ffn_params = 0
norm_params = 0
for block in self.blocks:
# Attention parameters
attention_params += sum(p.numel() for p in block.attention.parameters())
# FFN parameters
ffn_params += sum(p.numel() for p in block.ffn.parameters())
# Norm parameters (attention + ffn norms)
norm_params += sum(p.numel() for p in block.attn_norm.parameters())
norm_params += sum(p.numel() for p in block.ffn_norm.parameters())
# Final norm
norm_params += sum(p.numel() for p in self.final_norm.parameters())
# LM head (only if not tied)
if not self.tie_embeddings:
lm_head_params = sum(p.numel() for p in self.lm_head.parameters())
else:
lm_head_params = 0 # Shared with embeddings
total_params = sum(p.numel() for p in self.parameters())
return {
'total': total_params,
'total_millions': total_params / 1e6,
'embedding': embedding_params,
'attention': attention_params,
'ffn': ffn_params,
'norm': norm_params,
'lm_head': lm_head_params,
}
def forward(
self,
input_ids: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
use_cache: bool = False,
past_key_values: Optional[list] = None,
) -> Dict[str, torch.Tensor]:
"""Forward pass of the model.
Args:
input_ids: Token IDs [batch, seq_len]
attention_mask: Optional attention mask
position_ids: Optional position IDs
labels: Optional labels for language modeling loss
use_cache: Whether to return KV cache
past_key_values: Past KV cache for inference
Returns:
Dictionary with 'logits' and optionally 'loss' and 'past_key_values'
"""
batch_size, seq_len = input_ids.shape
# Token embeddings
x = self.token_embedding(input_ids)
# Apply dropout to embeddings
x = nn.functional.dropout(x, p=self.dropout, training=self.training)
# Process through transformer blocks
past_key_values_out = [] if use_cache else None
for i, block in enumerate(self.blocks):
# Get past KV for this layer if available
past_kv = past_key_values[i] if past_key_values is not None else None
# Process through block
x, kv_cache = block(
x,
attention_mask=attention_mask,
position_ids=position_ids,
use_cache=use_cache,
past_kv=past_kv,
)
# Store KV cache if needed
if use_cache:
past_key_values_out.append(kv_cache)
# Final normalization
x = self.final_norm(x)
# Language modeling head
logits = self.lm_head(x)
# Prepare output
output = {'logits': logits}
# Calculate loss if labels provided
if labels is not None:
# Shift for next-token prediction
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten for cross-entropy
shift_logits = shift_logits.view(-1, self.vocab_size)
shift_labels = shift_labels.view(-1)
# Calculate cross-entropy loss
loss = nn.functional.cross_entropy(
shift_logits,
shift_labels,
ignore_index=-100, # Standard ignore index
)
output['loss'] = loss
# Add KV cache to output if requested
if use_cache:
output['past_key_values'] = past_key_values_out
return output
def create_model(config: Dict[str, Any]) -> WikiMiniModel:
"""Create a WikiMini model from configuration.
Args:
config: Model configuration dictionary
Returns:
WikiMiniModel instance
"""
return WikiMiniModel(config)
# Test the complete model
if __name__ == "__main__":
# Test configuration for ~95M parameters
config = {
'vocab_size': 32000,
'd_model': 768,
'n_layers': 12,
'n_heads': 12,
'd_ffn': 2048, # Adjusted for SwiGLU
'max_seq_len': 2048,
'dropout': 0.1,
'rope_percentage': 0.5,
'rope_base': 10000,
'rms_norm_eps': 1e-6,
'tie_embeddings': True,
'use_flash_attention': True,
'gradient_checkpointing': True, # Enable for memory efficiency
}
# Create model
model = WikiMiniModel(config)
# Count parameters
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"WikiMini Model:")
print(f" Total parameters: {total_params:,} ({total_params/1e6:.2f}M)")
print(f" Trainable parameters: {trainable_params:,} ({trainable_params/1e6:.2f}M)")
print(f" Layers: {model.n_layers}")
print(f" Hidden size: {model.d_model}")
print(f" Attention heads: {model.n_heads}")
# Test forward pass
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)
model.eval()
# Small test batch
batch_size = 2
seq_len = 128
input_ids = torch.randint(0, config['vocab_size'], (batch_size, seq_len), device=device)
# Enable gradient checkpointing
model.enable_gradient_checkpointing()
# Forward pass
with torch.no_grad():
outputs = model(input_ids=input_ids)
print(f"\nTest forward pass:")
print(f" Input shape: {input_ids.shape}")
print(f" Output logits shape: {outputs['logits'].shape}")
print(f" Device: {device}")
if torch.cuda.is_available():
print(f" Memory allocated: {torch.cuda.memory_allocated(device) / 1024**3:.2f} GB")
print("\n✓ Model test passed!") |