File size: 20,892 Bytes
f881e8c |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 |
import os
import time
import math
import pickle
import random
import json
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import matplotlib.pyplot as plt
# We use Hugging Face’s transformers only for pretrained weight loading and tokenizer.
from transformers import GPT2LMHeadModel, GPT2Tokenizer
from dataclasses import dataclass
# ----------------------------
# Helper: ALiBi slopes computation
# ----------------------------
def get_alibi_slopes(n_head):
"""Compute ALiBi slopes for each head.
This implementation follows the approach used in several ALiBi implementations.
"""
def get_slopes_power_of_2(n):
start = 2 ** (-2 ** -(math.log2(n) - 3))
ratio = start
return [start * (ratio ** i) for i in range(n)]
if math.log2(n_head).is_integer():
slopes = get_slopes_power_of_2(n_head)
else:
closest_power_of_2 = 2 ** math.floor(math.log2(n_head))
slopes = get_slopes_power_of_2(closest_power_of_2)
extra_slopes = get_slopes_power_of_2(2 * closest_power_of_2)[0::2][: n_head - closest_power_of_2]
slopes.extend(extra_slopes)
return torch.tensor(slopes, dtype=torch.float32)
# ----------------------------
# Model Components
# ----------------------------
class LayerNorm(nn.Module):
"""LayerNorm with an optional bias."""
def __init__(self, ndim, bias: bool):
super().__init__()
self.weight = nn.Parameter(torch.ones(ndim))
self.bias = nn.Parameter(torch.zeros(ndim)) if bias else None
def forward(self, input):
return F.layer_norm(input, self.weight.shape, self.weight, self.bias, 1e-5)
class CausalSelfAttention(nn.Module):
def __init__(self, config):
super().__init__()
assert config.n_embd % config.n_head == 0
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd, bias=config.bias)
self.c_proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
self.attn_dropout = nn.Dropout(config.dropout)
self.resid_dropout = nn.Dropout(config.dropout)
self.n_head = config.n_head
self.n_embd = config.n_embd
self.dropout = config.dropout
self.use_rope = config.use_rope
self.rope_base = config.rope_base
# Existing APE support.
self.use_ape = getattr(config, 'use_ape', False)
# New: ALiBi support.
self.use_alibi = getattr(config, 'use_alibi', False)
if self.use_alibi and self.use_ape:
raise ValueError("Cannot use both ALiBi and APE simultaneously.")
# For APE, learn a parameter beta.
if self.use_ape:
self.beta = nn.Parameter(torch.tensor(1.0))
# Use Flash Attention if available (but disable when APE is enabled).
self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
if (not self.flash) or self.use_ape:
self.register_buffer("bias", torch.tril(torch.ones(config.block_size, config.block_size))
.view(1, 1, config.block_size, config.block_size))
def forward(self, x, return_attn_entropy=False, aggregate_heads=False):
"""
Args:
x: Input tensor [B, T, C]
return_attn_entropy (bool): If True, return attention entropy.
aggregate_heads (bool): If True, average entropy across heads.
Returns:
y: Output tensor [B, T, C] or (y, entropy)
"""
B, T, C = x.size()
q, k, v = self.c_attn(x).split(self.n_embd, dim=2)
head_dim = C // self.n_head
# Reshape to [B, n_head, T, head_dim]
q = q.view(B, T, self.n_head, head_dim).transpose(1, 2)
k = k.view(B, T, self.n_head, head_dim).transpose(1, 2)
v = v.view(B, T, self.n_head, head_dim).transpose(1, 2)
# Optionally, apply RoPE if enabled.
if self.use_rope:
hs = head_dim
d = hs // 2
if self.use_ape:
theta = 1.0 / (self.rope_base ** (2 * torch.arange(0, d, dtype=x.dtype, device=x.device) / hs))
else:
theta = 1.0 / (self.rope_base ** (2 * torch.arange(0, d, dtype=x.dtype, device=x.device) / hs))
t_pos = torch.arange(T, device=x.device, dtype=x.dtype)
freqs = torch.outer(t_pos, theta)
freqs_cos = torch.cos(freqs).unsqueeze(0).unsqueeze(0)
freqs_sin = torch.sin(freqs).unsqueeze(0).unsqueeze(0)
def apply_rope(tensor, cos, sin):
tensor = tensor.reshape(*tensor.shape[:-1], -1, 2)
x0 = tensor[..., 0]
x1 = tensor[..., 1]
x0_rot = x0 * cos - x1 * sin
x1_rot = x0 * sin + x1 * cos
return torch.stack([x0_rot, x1_rot], dim=-1).flatten(start_dim=-2)
q = apply_rope(q, freqs_cos, freqs_sin)
k = apply_rope(k, freqs_cos, freqs_sin)
# Compute scaled dot-product attention scores.
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(head_dim))
# --- Apply positional biases ---
if self.use_alibi:
slopes = get_alibi_slopes(self.n_head).to(x.device) # shape: (n_head,)
rel_positions = torch.arange(T, device=x.device).unsqueeze(0) - torch.arange(T, device=x.device).unsqueeze(1)
alibi_bias = slopes.view(1, self.n_head, 1, 1) * rel_positions.view(1, 1, T, T)
att = att - alibi_bias
elif self.use_ape:
pos_ids = torch.arange(T, device=x.device)
rel_dist = pos_ids.unsqueeze(0) - pos_ids.unsqueeze(1)
abs_rel = rel_dist.abs().float()
temp_matrix = 1.0 / (1.0 + abs_rel)
bias_matrix = - self.beta * torch.log(1.0 + abs_rel)
temp_matrix = temp_matrix.unsqueeze(0).unsqueeze(0)
bias_matrix = bias_matrix.unsqueeze(0).unsqueeze(0)
att = temp_matrix * att + bias_matrix
p_att = F.softmax(att, dim=-1)
entropy = -(p_att * torch.log(p_att + 1e-9)).sum(dim=-1) # [B, n_head, T, T]
if self.flash and not self.use_ape:
y = torch.nn.functional.scaled_dot_product_attention(
q, k, v,
attn_mask=None,
dropout_p=self.dropout if self.training else 0,
is_causal=True
)
else:
if T > self.bias.size(-1):
bias = torch.tril(torch.ones(T, T, device=x.device)).view(1, 1, T, T)
else:
bias = self.bias[:, :, :T, :T]
att = att.masked_fill(bias == 0, float('-inf'))
p_att = F.softmax(att, dim=-1)
entropy = -(p_att * torch.log(p_att + 1e-9)).sum(dim=-1)
att = self.attn_dropout(p_att)
y = att @ v # [B, n_head, T, head_dim]
y = y.transpose(1, 2).contiguous().view(B, T, C)
y = self.resid_dropout(self.c_proj(y))
if return_attn_entropy:
if aggregate_heads:
entropy = entropy.mean(dim=1) # [B, T, T]
return y, entropy
else:
return y
class MLP(nn.Module):
def __init__(self, config):
super().__init__()
self.c_fc = nn.Linear(config.n_embd, 4 * config.n_embd, bias=config.bias)
self.gelu = nn.GELU()
self.c_proj = nn.Linear(4 * config.n_embd, config.n_embd, bias=config.bias)
self.dropout = nn.Dropout(config.dropout)
def forward(self, x):
x = self.c_fc(x)
x = self.gelu(x)
x = self.c_proj(x)
x = self.dropout(x)
return x
class Block(nn.Module):
def __init__(self, config):
super().__init__()
self.ln_1 = LayerNorm(config.n_embd, bias=config.bias)
self.attn = CausalSelfAttention(config)
self.ln_2 = LayerNorm(config.n_embd, bias=config.bias)
self.mlp = MLP(config)
def forward(self, x, return_attn_entropy=False, aggregate_heads=False):
if return_attn_entropy:
attn_output, entropy = self.attn(self.ln_1(x), return_attn_entropy=True, aggregate_heads=aggregate_heads)
x = x + attn_output
x = x + self.mlp(self.ln_2(x))
return x, entropy
else:
attn_output = self.attn(self.ln_1(x), return_attn_entropy=False)
x = x + attn_output
x = x + self.mlp(self.ln_2(x))
return x
@dataclass
class GPTConfig:
block_size: int = 128
vocab_size: int = 50304 # For GPT-2
n_layer: int = 6
n_head: int = 6
n_embd: int = 384
dropout: float = 0.0
bias: bool = True
use_rope: bool = True
rope_base: float = 10000.0
use_ape: bool = False
lambda_temp: float = 0.1
use_alibi: bool = False
class GPT(nn.Module):
def __init__(self, config):
super().__init__()
assert config.vocab_size is not None and config.block_size is not None
self.config = config
# If using ALiBi, disable RoPE.
self.use_rope = config.use_rope and not config.use_alibi
print(f"Using RoPE in GPT init: {self.use_rope}")
self.transformer = nn.ModuleDict(dict(
wte = nn.Embedding(config.vocab_size, config.n_embd),
wpe = None if self.use_rope else nn.Embedding(config.block_size, config.n_embd),
drop = nn.Dropout(config.dropout),
h = nn.ModuleList([Block(config) for _ in range(config.n_layer)]),
ln_f = LayerNorm(config.n_embd, bias=config.bias),
))
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
self.transformer.wte.weight = self.lm_head.weight
self.apply(self._init_weights)
for pn, p in self.named_parameters():
if pn.endswith('c_proj.weight'):
torch.nn.init.normal_(p, mean=0.0, std=0.02/math.sqrt(2 * config.n_layer))
print("number of parameters: %.2fM" % (self.get_num_params()/1e6,))
def get_num_params(self, non_embedding=True):
n_params = sum(p.numel() for p in self.parameters())
if non_embedding and (not self.use_rope) and (self.transformer.wpe is not None):
n_params -= self.transformer.wpe.weight.numel()
return n_params
def _init_weights(self, module):
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 forward(self, idx, targets=None, return_attn_entropy=False, aggregate_heads=False):
device = idx.device
b, t = idx.size()
pos = torch.arange(0, t, dtype=torch.long, device=device)
tok_emb = self.transformer.wte(idx)
if self.use_rope or self.config.use_alibi:
x = self.transformer.drop(tok_emb)
else:
pos_emb = self.transformer.wpe(pos) if self.transformer.wpe is not None else 0
x = self.transformer.drop(tok_emb + pos_emb)
attn_entropies = []
for block in self.transformer.h:
if return_attn_entropy:
x, entropy = block(x, return_attn_entropy=True, aggregate_heads=aggregate_heads)
attn_entropies.append(entropy)
else:
x = block(x)
x = self.transformer.ln_f(x)
if targets is not None:
logits = self.lm_head(x)
loss = F.cross_entropy(logits.view(-1, logits.size(-1)), targets.view(-1), ignore_index=-1)
else:
logits = self.lm_head(x[:, [-1], :])
loss = None
if return_attn_entropy:
return logits, loss, attn_entropies
else:
return logits, loss
@torch.no_grad()
def generate_and_compute_perplexity(self, prompt, ground_truth, temperature=1.0, return_attn_entropy=False, aggregate_heads=False):
if return_attn_entropy:
_, _, attn_entropies = self(prompt, return_attn_entropy=True, aggregate_heads=aggregate_heads)
per_layer_avgs = [entropy.mean().item() for entropy in attn_entropies]
avg_entropy = np.mean(per_layer_avgs)
else:
avg_entropy = None
total_loss = 0.0
total_tokens = 0
prompt_length = prompt.size(1)
num_target_tokens = ground_truth.size(1) - prompt_length
idx = prompt.clone()
for i in range(num_target_tokens):
logits, _ = self(idx)
logits = logits[:, -1, :] / temperature
target = ground_truth[:, prompt_length + i]
loss = F.cross_entropy(logits, target, reduction='sum')
total_loss += loss.item()
total_tokens += target.numel()
target_token = target.unsqueeze(1)
idx = torch.cat((idx, target_token), dim=1)
avg_neg_log_likelihood = total_loss / total_tokens if total_tokens > 0 else float('inf')
perplexity = math.exp(avg_neg_log_likelihood)
return idx, perplexity, avg_entropy
@torch.no_grad()
def generate_until_end(self, idx, temperature=1.0, top_k=None, max_new_tokens=1000):
for i in range(max_new_tokens):
idx_cond = idx
logits, _ = self(idx_cond)
logits = logits[:, -1, :] / temperature
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
logits[logits < v[:, [-1]]] = -float('Inf')
probs = F.softmax(logits, dim=-1)
idx_next = torch.multinomial(probs, num_samples=1)
idx = torch.cat((idx, idx_next), dim=1)
if idx_next.item() == 50256:
break
return idx
# ----------------------------
# Utility Functions for Training & Evaluation
# ----------------------------
# Data Loader Functions
train_data_path = "/data1/home/nitinvetcha/Topics in AI/Streamlined/COLM2025/train_tinystories.bin"
val_data_path = "/data1/home/nitinvetcha/Topics in AI/Streamlined/COLM2025/val_tinystories.bin"
def get_batch(split):
data_path = train_data_path if split == 'train' else val_data_path
data = np.memmap(data_path, dtype=np.uint16, mode='r')
total_tokens = len(data)
max_ix = max(1, total_tokens - gptconf.block_size)
ix = torch.randint(0, max_ix, (batch_size,))
X = torch.stack([torch.from_numpy(data[i:i+gptconf.block_size].astype(np.int64)) for i in ix])
Y = torch.stack([torch.from_numpy(data[i+1:i+1+gptconf.block_size].astype(np.int64)) for i in ix])
return X.to(device), Y.to(device)
def evaluate_prompt_perplexity(model, token_file, prompt_length, num_trials, generation_params, device):
tokens = np.fromfile(token_file, dtype=np.uint16)
total_tokens = len(tokens)
perplexities = []
entropy_trials = []
max_new_tokens = generation_params.get("max_new_tokens", 50)
total_length = prompt_length + max_new_tokens
for trial in range(num_trials):
start_idx = random.randint(0, total_tokens - total_length)
sequence_tokens = tokens[start_idx : start_idx + total_length]
prompt_tokens = sequence_tokens[:prompt_length]
ground_truth_tokens = sequence_tokens
prompt_tensor = torch.tensor(prompt_tokens, dtype=torch.long).unsqueeze(0).to(device)
ground_truth_tensor = torch.tensor(ground_truth_tokens, dtype=torch.long).unsqueeze(0).to(device)
_, ppl, trial_entropy = model.generate_and_compute_perplexity(
prompt_tensor, ground_truth_tensor,
temperature=generation_params.get("temperature", 1.0),
return_attn_entropy=True, aggregate_heads=True
)
perplexities.append(ppl)
entropy_trials.append(trial_entropy)
print(f"Trial {trial+1}/{num_trials} for prompt length {prompt_length}: Perplexity = {ppl:.2f}, Avg Entropy = {trial_entropy:.4f}")
avg_ppl = np.mean(perplexities)
avg_entropy = np.mean(entropy_trials)
print(f"Prompt Length {prompt_length} - Avg Perplexity: {avg_ppl:.2f}, Avg Attention Entropy: {avg_entropy:.4f}\n")
return avg_ppl, avg_entropy
# ----------------------------
# Training Loop
# ----------------------------
# Training hyperparameters
batch_size = 12
max_iters = 25001
save_interval = 5000
learning_rate = 6e-4
weight_decay = 1e-1
grad_clip = 1.0
device = 'cuda' if torch.cuda.is_available() else 'cpu'
torch.manual_seed(1337)
# Model configuration: adjust these flags as needed.
model_args = dict(
n_layer=6,
n_head=6,
n_embd=384,
block_size=64, # You can change this as needed.
bias=False,
use_rope=True,
use_ape=True, # Set to True if you want APE.
use_alibi=False, # Set to True to use ALiBi.
rope_base=10000.0,
vocab_size=50304,
dropout=0.0
)
gptconf = GPTConfig(**model_args)
model = GPT(gptconf).to(device)
model.train()
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
iter_num = 0
start_time = time.time()
training_losses = []
validation_losses = []
save_iters = []
# Build a flag string for naming: e.g. "rope_ape" or "alibi" etc.
flag_parts = []
if gptconf.use_rope:
flag_parts.append("rope")
if gptconf.use_ape:
flag_parts.append("ape")
if gptconf.use_alibi:
flag_parts.append("alibi")
flag_str = "_".join(flag_parts) if flag_parts else "none"
weight_dir = f"weights_{flag_str}_{gptconf.block_size}"
os.makedirs(weight_dir, exist_ok=True)
while iter_num < max_iters:
X_train, Y_train = get_batch('train')
optimizer.zero_grad()
logits, loss_train = model(X_train, Y_train)
loss_train.backward()
if grad_clip > 0:
torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
optimizer.step()
training_losses.append(loss_train.item())
model.eval()
X_val, Y_val = get_batch('val')
with torch.no_grad():
logits_val, loss_val = model(X_val, Y_val)
validation_losses.append(loss_val.item())
model.train()
if iter_num % 100 == 0:
elapsed = time.time() - start_time
print(f"Iter {iter_num:5d}: train loss = {loss_train.item():.4f}, val loss = {loss_val.item():.4f}, time/iter = {elapsed/(iter_num+1):.4f}s")
if iter_num > 0 and iter_num % save_interval == 0:
save_iters.append(iter_num)
ckpt = {
'iter_num': iter_num,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'training_losses': training_losses,
'validation_losses': validation_losses,
'save_iters': save_iters,
}
ckpt_path = os.path.join(weight_dir, f"ckpt_{iter_num}.pt")
torch.save(ckpt, ckpt_path)
print(f"Checkpoint saved to {ckpt_path}")
iter_num += 1
print("Training complete.")
plt.figure(figsize=(10, 6))
plt.plot(range(len(training_losses)), training_losses, label="Training Loss")
plt.plot(range(len(validation_losses)), validation_losses, label="Validation Loss", alpha=0.7)
plt.xlabel("Iteration")
plt.ylabel("Loss")
plt.title("Training and Validation Loss per Iteration")
plt.legend()
plt.grid(True)
plt.show()
# ----------------------------
# Perplexity & Entropy Evaluation
# ----------------------------
token_file = val_data_path # Use validation data for evaluation.
prompt_lengths = [64, 128, 256, 512, 1024, 2048, 4096, 8192]
num_trials = 5
generation_params = {"temperature": 1.0, "max_new_tokens": 50}
avg_perplexities = []
avg_entropies = []
for pl in prompt_lengths:
print(f"Evaluating for prompt length: {pl}")
avg_ppl, avg_entropy = evaluate_prompt_perplexity(model, token_file, pl, num_trials, generation_params, device)
avg_perplexities.append(avg_ppl)
avg_entropies.append(avg_entropy)
results = {
"prompt_lengths": prompt_lengths,
"avg_perplexities": avg_perplexities,
"avg_entropies": avg_entropies
}
results_filename = f"results_{flag_str}_{gptconf.block_size}.json"
with open(results_filename, "w") as f:
json.dump(results, f)
print(f"Results saved to {results_filename}")
plt.figure(figsize=(8, 6))
plt.plot(prompt_lengths, avg_perplexities, marker='o')
plt.xlabel("Prompt Length")
plt.ylabel("Avg Generated Perplexity")
plt.title("Avg Generated Perplexity vs Prompt Length")
plt.grid(True)
plt.xscale('log')
plt.savefig(f"avg_generated_perplexity_{flag_str}_{gptconf.block_size}.png")
plt.show()
plt.figure(figsize=(8, 6))
plt.plot(prompt_lengths, avg_entropies, marker='o', color='red')
plt.xlabel("Prompt Length")
plt.ylabel("Avg Attention Entropy")
plt.title("Avg Attention Entropy vs Prompt Length\n(Averaged over Layers)")
plt.grid(True)
plt.xscale('log')
plt.savefig(f"avg_attention_entropy_{flag_str}_{gptconf.block_size}.png")
plt.show()
|