File size: 13,266 Bytes
49d2fa1 |
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 |
"""
model.py
========
Complete SmolLM2-135M model implementation
Architecture:
- 30 transformer blocks
- 576 hidden dimensions
- 9 query heads, 3 KV heads (Grouped Query Attention)
- SwiGLU feed-forward network
- RoPE position embeddings
- RMSNorm layer normalization
- Weight tying (embeddings = lm_head)
Total parameters: 134,515,008 (~135M)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from components import RMSNorm, TransformerBlock
from transformers import AutoConfig
class SmolLM2Model(nn.Module):
"""
SmolLM2-135M Language Model
A decoder-only transformer based on Llama architecture with:
- Grouped Query Attention (memory efficient)
- SwiGLU FFN (improved expressiveness)
- RoPE position embeddings (length extrapolation)
- RMSNorm (faster than LayerNorm)
Model configuration:
- Layers: 30
- Hidden size: 576
- Attention heads: 9 (Q) / 3 (KV)
- FFN size: 1536
- Vocab size: 49,152
- Context length: 2048
"""
def __init__(self, config):
"""
Initialize SmolLM2 model
Args:
config: Model configuration object with attributes:
- vocab_size: Size of vocabulary (49152)
- hidden_size: Model dimension (576)
- num_hidden_layers: Number of transformer blocks (30)
- tie_word_embeddings: Whether to tie input/output embeddings
- rms_norm_eps: Epsilon for RMSNorm
"""
super().__init__()
self.config = config
# Token embeddings
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
# Transformer blocks (30 layers)
self.layers = nn.ModuleList([
TransformerBlock(config) for _ in range(config.num_hidden_layers)
])
# Final layer normalization
self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
# Language modeling head (output projection)
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
# Weight tying: share embeddings with output projection
if config.tie_word_embeddings:
self.lm_head.weight = self.embed_tokens.weight
print(f"β
Model initialized with {config.num_hidden_layers} transformer blocks")
print(f"β
Weight tying: {config.tie_word_embeddings}")
def forward(self, input_ids, attention_mask=None, position_ids=None):
"""
Forward pass through the model
Args:
input_ids (torch.Tensor): Input token IDs [batch, seq_len]
attention_mask (torch.Tensor, optional): Attention mask
position_ids (torch.Tensor, optional): Position indices
Returns:
torch.Tensor: Logits over vocabulary [batch, seq_len, vocab_size]
"""
batch_size, seq_len = input_ids.shape
# Create position IDs if not provided
if position_ids is None:
position_ids = torch.arange(seq_len, device=input_ids.device)
# Embed tokens
hidden_states = self.embed_tokens(input_ids)
# Pass through all transformer blocks
for layer in self.layers:
hidden_states = layer(hidden_states, attention_mask, position_ids)
# Final normalization
hidden_states = self.norm(hidden_states)
# Project to vocabulary
logits = self.lm_head(hidden_states)
return logits
def generate(
self,
input_ids,
max_new_tokens=50,
temperature=1.0,
top_p=0.9,
top_k=None,
do_sample=True
):
"""
Generate text autoregressively
Supports multiple sampling strategies:
- Greedy decoding (temperature=0)
- Temperature sampling
- Nucleus (top-p) sampling
- Top-k sampling
Args:
input_ids (torch.Tensor): Input token IDs [batch, seq_len]
max_new_tokens (int): Number of tokens to generate
temperature (float): Sampling temperature (0 = greedy, >1 = more random)
top_p (float): Nucleus sampling threshold (0-1)
top_k (int, optional): Top-k sampling threshold
do_sample (bool): Whether to sample or use greedy decoding
Returns:
torch.Tensor: Generated token IDs [batch, seq_len + max_new_tokens]
"""
self.eval()
for _ in range(max_new_tokens):
with torch.no_grad():
# Forward pass
logits = self(input_ids)
# Get next token logits
next_token_logits = logits[:, -1, :]
# Apply temperature
if temperature > 0:
next_token_logits = next_token_logits / temperature
# Greedy decoding
if not do_sample or temperature == 0:
next_token = torch.argmax(next_token_logits, dim=-1, keepdim=True)
else:
# Top-k sampling
if top_k is not None:
top_k = min(top_k, next_token_logits.size(-1))
indices_to_remove = next_token_logits < torch.topk(next_token_logits, top_k)[0][..., -1, None]
next_token_logits[indices_to_remove] = float('-inf')
# Nucleus (top-p) sampling
if top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(next_token_logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above threshold
sorted_indices_to_remove = cumulative_probs > top_p
# Keep at least one token
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = False
# Scatter to original indexing
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
next_token_logits[indices_to_remove] = float('-inf')
# Sample from distribution
probs = F.softmax(next_token_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
# Append to sequence
input_ids = torch.cat([input_ids, next_token], dim=1)
return input_ids
def get_num_params(self, non_embedding=False):
"""
Count model parameters
Args:
non_embedding (bool): If True, exclude embedding parameters
Returns:
int: Number of parameters
"""
n_params = sum(p.numel() for p in self.parameters())
if non_embedding:
n_params -= self.embed_tokens.weight.numel()
# If weights are tied, don't double-count
if not self.config.tie_word_embeddings:
n_params -= self.lm_head.weight.numel()
return n_params
def initialize_weights(model, config):
"""
Initialize model weights using GPT-style initialization
Strategy:
- All weights: Normal(0, 0.02)
- Residual projections: Scaled by 1/sqrt(2 * num_layers)
- RMSNorm: Initialized to 1.0 (PyTorch default)
The residual scaling prevents variance explosion in deep networks.
Args:
model (SmolLM2Model): Model to initialize
config: Model configuration
"""
std = 0.02
num_layers = config.num_hidden_layers
# Residual scaling factor: 1/sqrt(2 * num_layers)
residual_scaling = 1.0 / math.sqrt(2 * num_layers)
print(f"Initializing weights with std={std}, residual_scaling={residual_scaling:.6f}")
# Initialize embeddings
nn.init.normal_(model.embed_tokens.weight, mean=0.0, std=std)
# Initialize each transformer block
for layer in model.layers:
# Attention projections
nn.init.normal_(layer.self_attn.q_proj.weight, mean=0.0, std=std)
nn.init.normal_(layer.self_attn.k_proj.weight, mean=0.0, std=std)
nn.init.normal_(layer.self_attn.v_proj.weight, mean=0.0, std=std)
# Output projection with residual scaling
nn.init.normal_(layer.self_attn.o_proj.weight, mean=0.0, std=std * residual_scaling)
# FFN projections
nn.init.normal_(layer.mlp.gate_proj.weight, mean=0.0, std=std)
nn.init.normal_(layer.mlp.up_proj.weight, mean=0.0, std=std)
# Output projection with residual scaling
nn.init.normal_(layer.mlp.down_proj.weight, mean=0.0, std=std * residual_scaling)
# RMSNorm weights are initialized to 1.0 by default (PyTorch)
print(f"β
Initialized {sum(1 for _ in model.parameters())} weight tensors")
def load_pretrained_weights(our_model, official_model, device='cuda'):
"""
Load weights from HuggingFace official model
Maps weight names from official model to our implementation:
- model.embed_tokens.weight -> embed_tokens.weight
- model.layers.{i}.* -> layers[i].*
- model.norm.weight -> norm.weight
- lm_head.weight (tied with embeddings)
Args:
our_model (SmolLM2Model): Our model to load weights into
official_model: HuggingFace official model
device (str): Device to load weights to
Returns:
int: Number of weight tensors loaded
"""
print("=" * 70)
print("LOADING PRETRAINED WEIGHTS")
print("=" * 70)
official_state = official_model.state_dict()
loaded_count = 0
# 1. Load token embeddings
our_model.embed_tokens.weight.data = official_state['model.embed_tokens.weight'].clone().to(device)
loaded_count += 1
# 2. Load all transformer blocks
num_layers = our_model.config.num_hidden_layers
for layer_idx in range(num_layers):
prefix = f'model.layers.{layer_idx}'
# Layer norms
our_model.layers[layer_idx].input_layernorm.weight.data = \
official_state[f'{prefix}.input_layernorm.weight'].clone().to(device)
our_model.layers[layer_idx].post_attention_layernorm.weight.data = \
official_state[f'{prefix}.post_attention_layernorm.weight'].clone().to(device)
# Attention projections
our_model.layers[layer_idx].self_attn.q_proj.weight.data = \
official_state[f'{prefix}.self_attn.q_proj.weight'].clone().to(device)
our_model.layers[layer_idx].self_attn.k_proj.weight.data = \
official_state[f'{prefix}.self_attn.k_proj.weight'].clone().to(device)
our_model.layers[layer_idx].self_attn.v_proj.weight.data = \
official_state[f'{prefix}.self_attn.v_proj.weight'].clone().to(device)
our_model.layers[layer_idx].self_attn.o_proj.weight.data = \
official_state[f'{prefix}.self_attn.o_proj.weight'].clone().to(device)
# FFN projections
our_model.layers[layer_idx].mlp.gate_proj.weight.data = \
official_state[f'{prefix}.mlp.gate_proj.weight'].clone().to(device)
our_model.layers[layer_idx].mlp.up_proj.weight.data = \
official_state[f'{prefix}.mlp.up_proj.weight'].clone().to(device)
our_model.layers[layer_idx].mlp.down_proj.weight.data = \
official_state[f'{prefix}.mlp.down_proj.weight'].clone().to(device)
loaded_count += 9 # 2 norms + 4 attn + 3 ffn
# 3. Load final norm
our_model.norm.weight.data = official_state['model.norm.weight'].clone().to(device)
loaded_count += 1
print(f"\nβ
Loaded {num_layers} transformer blocks")
print(f"β
Total loaded: {loaded_count} weight tensors")
print("=" * 70)
return loaded_count
if __name__ == "__main__":
"""Test model creation and parameter count"""
# Load config
config = AutoConfig.from_pretrained("HuggingFaceTB/SmolLM2-135M")
# Create model
model = SmolLM2Model(config)
# Count parameters
total_params = model.get_num_params()
print(f"\nTotal parameters: {total_params:,}")
print(f"Expected: 134,515,008")
print(f"Match: {total_params == 134_515_008}")
# Test forward pass
test_input = torch.randint(0, config.vocab_size, (1, 10))
output = model(test_input)
print(f"\nForward pass test:")
print(f" Input shape: {test_input.shape}")
print(f" Output shape: {output.shape}")
print(f" Expected: torch.Size([1, 10, 49152])")
# Test generation
generated = model.generate(test_input, max_new_tokens=5)
print(f"\nGeneration test:")
print(f" Generated shape: {generated.shape}")
print(f" Expected: torch.Size([1, 15])") |