File size: 17,547 Bytes
518db7a | 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 | """
Text Decoder Module for SEM V6 - Multi-Token Prediction.
Converts latent representations to text using Llama-3 tokenizer.
Uses multi-token prediction (MTP) for faster training and better representations.
Architecture:
latent (256) → N prediction heads → vocab_logits[N, vocab_size] → tokens → text
Reference:
- Multi-Token Prediction (Meta 2024): https://arxiv.org/abs/2404.19737
- Llama 3: https://llama.meta.com/
"""
import os
import glob
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
from transformers import AutoTokenizer
class MultiTokenDecoder(nn.Module):
"""
Multi-token prediction decoder for SEM V6 with TIED EMBEDDINGS.
Predicts N future tokens simultaneously from a latent vector,
enabling faster training and richer representations.
Architecture (per prediction head):
latent (latent_dim)
→ shared_trunk (latent_dim → hidden_dim)
→ head_i (hidden_dim → embed_dim)
→ vocab_embedding.T (embed_dim → vocab_size) # TIED WEIGHTS
The key innovation is TIED EMBEDDINGS: the vocabulary embedding matrix
is shared between input and output projections. This:
1. Reduces parameters by 256x (from hidden→vocab to hidden→embed + embed→vocab shared)
2. Provides semantic guidance - vocabulary learns meaningful relationships
3. Matches standard LLM practice (GPT-2, Llama, etc.)
Each head predicts token at position i (0 = next token, 1 = token after, etc.)
Example:
>>> decoder = MultiTokenDecoder(latent_dim=256, num_predict=4)
>>> z = torch.randn(1, 256)
>>> logits = decoder(z) # (1, 4, vocab_size)
>>> text = decoder.generate(z, max_length=100)
"""
def __init__(
self,
latent_dim: int = 256,
hidden_dim: int = 512,
num_predict: int = 4, # Number of future tokens to predict
tokenizer_name: str = "meta-llama/Meta-Llama-3-8B",
device: str = "cuda",
embed_dim: int = 512, # Embedding dimension for tied weights
) -> None:
"""
Initialize multi-token decoder with tied embeddings.
Args:
latent_dim: Dimension of input latent vectors
hidden_dim: Hidden layer dimension
num_predict: Number of future tokens to predict (default 4)
tokenizer_name: HuggingFace tokenizer (default: Llama-3.2-1B)
device: Computation device
embed_dim: Vocabulary embedding dimension (default 512)
Tied between input/output for semantic learning
"""
super().__init__()
self.latent_dim = latent_dim
self.hidden_dim = hidden_dim
self.num_predict = num_predict
self.device = torch.device(device)
self.embed_dim = embed_dim
# Resolve tokenizer path for offline use (prefer local cache)
tokenizer_path = self._resolve_tokenizer_path(tokenizer_name)
# Load tokenizer
# Offline-safe: rely on local HF cache when network is unavailable.
self.tokenizer = AutoTokenizer.from_pretrained(
tokenizer_path,
trust_remote_code=True,
local_files_only=True,
)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
# Use len(tokenizer) not vocab_size - includes special tokens (128256 vs 128000)
self.vocab_size = len(self.tokenizer)
# TIED EMBEDDING: Shared vocabulary embedding matrix
# This is the key fix - vocabulary learns semantic relationships
# Output projection uses embedding.weight.T (tied weights)
self.vocab_embedding = nn.Embedding(self.vocab_size, embed_dim)
# Scale init for 512-dim bottleneck: std=0.02 gives σ_logit≈0.32 (near-uniform softmax)
# Target σ_logit≈3.0 for meaningful discrimination over 128k vocab
embed_init_std = 3.0 / (embed_dim ** 0.5) # ≈0.13 for embed_dim=512
nn.init.normal_(self.vocab_embedding.weight, std=embed_init_std)
# Shared trunk: latent → hidden representation
self.trunk = nn.Sequential(
nn.LayerNorm(latent_dim),
nn.Linear(latent_dim, hidden_dim),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
)
# Separate prediction heads for each future position
# Head i predicts token at position i
# NOW: hidden_dim → embed_dim (NOT vocab_size directly)
self.heads = nn.ModuleList(
[
nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(
hidden_dim, embed_dim
), # Output embed_dim, not vocab_size
)
for _ in range(num_predict)
]
)
# Output bias (optional, for vocab projection)
self.output_bias = nn.Parameter(torch.zeros(self.vocab_size))
# Embedding trunk: embed_dim → hidden for direct embedding path
# This is used by forward_from_embedding() for learnable encoder training
# Must be defined in __init__ (not lazily) so weights can be loaded from checkpoint
self._embed_trunk = nn.Sequential(
nn.LayerNorm(embed_dim),
nn.Linear(embed_dim, hidden_dim),
nn.GELU(),
nn.Dropout(0.1),
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
)
# Initialize head output layers
for head in self.heads:
nn.init.normal_(head[-1].weight, std=0.02)
nn.init.zeros_(head[-1].bias)
# Initialize _embed_trunk layers
for layer in self._embed_trunk:
if isinstance(layer, nn.Linear):
nn.init.normal_(layer.weight, std=0.02)
nn.init.zeros_(layer.bias)
@staticmethod
def _resolve_tokenizer_path(tokenizer_name: str) -> str:
"""Prefer local tokenizer snapshot to avoid network calls."""
# Explicit override via env
env_path = os.getenv("SEM_TOKENIZER_PATH")
if env_path and os.path.isdir(env_path):
return env_path
# If tokenizer_name is already a local path, use it
if os.path.isdir(tokenizer_name):
return tokenizer_name
# Try to resolve from HF cache
cache_root = os.path.expanduser("~/.cache/huggingface/hub")
model_dir = f"models--{tokenizer_name.replace('/', '--')}"
snapshot_glob = os.path.join(cache_root, model_dir, "snapshots", "*")
snapshots = sorted(glob.glob(snapshot_glob), key=os.path.getmtime, reverse=True)
if snapshots:
return snapshots[0]
# Fallback to original name (may fail offline)
return tokenizer_name
def forward(self, z: torch.Tensor) -> torch.Tensor:
"""
Predict multiple future tokens from latent using TIED EMBEDDINGS.
Args:
z: Latent tensor (batch, latent_dim) or (latent_dim,)
Returns:
logits: (batch, num_predict, vocab_size) or (num_predict, vocab_size)
"""
squeeze = z.dim() == 1
if squeeze:
z = z.unsqueeze(0)
# Shared representation
h = self.trunk(z) # (batch, hidden_dim)
# Predict from each head using TIED EMBEDDINGS
all_logits = []
for head in self.heads:
# Head outputs embed_dim representation
embed = head(h) # (batch, embed_dim)
# Project to vocab using TIED embedding weights
# logits = embed @ vocab_embedding.weight.T + bias
logits = F.linear(embed, self.vocab_embedding.weight, self.output_bias)
all_logits.append(logits) # (batch, vocab_size)
# Stack: (batch, num_predict, vocab_size)
output = torch.stack(all_logits, dim=1)
if squeeze:
output = output.squeeze(0)
return output
def embed_tokens(self, token_ids: torch.Tensor) -> torch.Tensor:
"""
Embed token IDs using the shared vocabulary embedding.
This enables using the same embeddings for both input and output,
which is the key to tied embeddings learning meaningful semantics.
Args:
token_ids: Token IDs (batch, seq_len) or (seq_len,)
Returns:
embeddings: Token embeddings (batch, seq_len, embed_dim) or (seq_len, embed_dim)
"""
return self.vocab_embedding(token_ids)
def forward_from_embedding(self, embed: torch.Tensor) -> torch.Tensor:
"""
Predict tokens directly from embedding space (bypasses latent→hidden trunk).
This provides a direct path for text generation when using learnable encoder
with tied embeddings. The embedding is projected to vocab using tied weights.
Args:
embed: Embedding tensor (batch, embed_dim) or (embed_dim,)
This should be the pooled embedding from LearnableEncoder.encode_to_embedding()
Returns:
logits: (batch, num_predict, vocab_size) or (num_predict, vocab_size)
"""
squeeze = embed.dim() == 1
if squeeze:
embed = embed.unsqueeze(0)
# Shared representation from embedding
# Note: _embed_trunk is now created in __init__ so weights load from checkpoint
h = self._embed_trunk(embed) # (batch, hidden_dim)
# Predict from each head using TIED EMBEDDINGS
all_logits = []
for head in self.heads:
head_embed = head(h) # (batch, embed_dim)
logits = F.linear(head_embed, self.vocab_embedding.weight, self.output_bias)
all_logits.append(logits)
output = torch.stack(all_logits, dim=1)
if squeeze:
output = output.squeeze(0)
return output
def decode_tokens(self, token_ids: torch.Tensor) -> str:
"""Decode token IDs to text."""
if token_ids.dim() > 1:
token_ids = token_ids.flatten()
return self.tokenizer.decode(token_ids.tolist(), skip_special_tokens=True)
def encode_text(self, text: str) -> torch.Tensor:
"""Encode text to token IDs."""
tokens = self.tokenizer.encode(text, return_tensors="pt")
return tokens.to(self.device)
@torch.no_grad()
def generate(
self,
z: torch.Tensor,
max_length: int = 100,
temperature: float = 0.8,
top_k: int = 50,
top_p: float = 0.9,
) -> str:
"""
Generate text using multi-token prediction.
Generates num_predict tokens at a time for faster generation.
Args:
z: Latent vector (latent_dim,) or (1, latent_dim)
max_length: Maximum tokens to generate
temperature: Sampling temperature
top_k: Top-k filtering
top_p: Nucleus sampling
Returns:
Generated text string
"""
if z.dim() == 1:
z = z.unsqueeze(0)
generated_ids = []
current_z = z
safe_temperature = max(temperature, 1e-4)
while len(generated_ids) < max_length:
# Get multi-token prediction
logits = self.forward(current_z) # (1, num_predict, vocab_size)
# Sample from each position
for i in range(self.num_predict):
if len(generated_ids) >= max_length:
break
pos_logits = logits[0, i] / safe_temperature
# Top-k filtering
if top_k > 0:
top_k_vals, _ = torch.topk(
pos_logits, min(top_k, pos_logits.size(-1))
)
pos_logits[pos_logits < top_k_vals[-1]] = float("-inf")
# Top-p filtering
if top_p < 1.0:
sorted_logits, sorted_idx = torch.sort(pos_logits, descending=True)
cumsum = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
mask = cumsum > top_p
mask[1:] = mask[:-1].clone()
mask[0] = False
sorted_logits[mask] = float("-inf")
pos_logits = sorted_logits.scatter(
0, sorted_idx.argsort(), sorted_logits
)
if torch.isneginf(pos_logits).all():
pos_logits = torch.zeros_like(pos_logits)
pos_logits = torch.clamp(pos_logits, -50.0, 50.0)
probs = F.softmax(pos_logits, dim=-1)
probs = torch.where(
torch.isfinite(probs), probs, torch.zeros_like(probs)
)
probs = probs + 1e-8
total = probs.sum()
if not torch.isfinite(total) or total <= 0:
probs = torch.ones_like(probs) / probs.numel()
else:
probs = probs / total
token_id = torch.multinomial(probs, 1).item()
# Stop at EOS
if token_id == self.tokenizer.eos_token_id:
return self.tokenizer.decode(
generated_ids, skip_special_tokens=True
)
generated_ids.append(token_id)
# For continuous generation, we'd update z based on generated tokens
# This requires feeding tokens back through the encoder
# For now, we continue with same z (limitation - proper impl needs feedback)
return self.tokenizer.decode(generated_ids, skip_special_tokens=True)
def multi_token_prediction_loss(
z_pred: torch.Tensor,
target_text: list[str],
decoder: MultiTokenDecoder,
max_length: int = 128,
loss_weight: float = 1.0,
position_decay: float = 1.0,
) -> tuple[torch.Tensor, dict]:
"""
Multi-token prediction loss following Meta's MTP approach.
Trains decoder to predict N future tokens from latent representation.
Each head predicts a different future position.
Reference:
- Meta (2024): Better & Faster LLMs via Multi-token Prediction
https://arxiv.org/abs/2404.19737
- NVIDIA Megatron: Average across depths, apply scaling factor
https://docs.nvidia.com/megatron-core/developer-guide/latest/api-guide/multi_token_prediction.html
Args:
z_pred: Predicted latent (batch, latent_dim)
target_text: List of target text strings
decoder: MultiTokenDecoder module
max_length: Max tokenization length
loss_weight: Scaling factor for MTP loss (default 1.0, use 0.1 when
combining with main task loss per NVIDIA Megatron)
position_decay: Exponential decay factor for later positions (default 1.0 = no decay)
Use <1.0 (e.g., 0.9) to weight earlier positions more heavily
Returns:
loss: Combined cross-entropy loss (averaged across positions, scaled by loss_weight)
metrics: Dict with per-position accuracy and loss breakdown
"""
batch_size = z_pred.size(0)
device = z_pred.device
num_predict = decoder.num_predict
# Tokenize targets
encoded = decoder.tokenizer(
target_text,
padding=True,
truncation=True,
max_length=max_length,
return_tensors="pt",
)
target_ids = encoded["input_ids"].to(device) # (batch, seq_len)
seq_len = target_ids.size(1)
# Get multi-token predictions
logits = decoder(z_pred) # (batch, num_predict, vocab_size)
# Compute loss for each prediction position
# Per Meta: average across all depths
# Per NVIDIA: compute average of MTP losses across all depths, multiply by scaling factor
total_loss = 0.0
total_weight = 0.0
accuracies = []
per_pos_losses = []
num_positions = min(num_predict, seq_len)
for i in range(num_positions):
pos_logits = logits[:, i] # (batch, vocab_size)
pos_targets = target_ids[:, i] # (batch,)
# Cross-entropy for this position
pos_loss = F.cross_entropy(pos_logits, pos_targets)
# Optional position weighting (earlier positions = higher weight)
pos_weight = position_decay**i
total_loss = total_loss + pos_loss * pos_weight
total_weight += pos_weight
per_pos_losses.append(pos_loss.item())
# Accuracy
with torch.no_grad():
preds = pos_logits.argmax(dim=-1)
acc = (preds == pos_targets).float().mean().item()
accuracies.append(acc)
# Average loss across positions (weighted if position_decay < 1.0)
avg_loss = total_loss / total_weight if total_weight > 0 else total_loss
# Apply loss weight (scaling factor per NVIDIA Megatron)
scaled_loss = avg_loss * loss_weight
metrics = {
"mtp_loss": scaled_loss.item(),
"mtp_loss_unscaled": avg_loss.item(),
"avg_accuracy": sum(accuracies) / len(accuracies) if accuracies else 0.0,
**{f"acc_pos_{i}": acc for i, acc in enumerate(accuracies)},
**{f"loss_pos_{i}": loss for i, loss in enumerate(per_pos_losses)},
}
return scaled_loss, metrics
# Alias for backward compatibility
TextDecoder = MultiTokenDecoder
text_generation_loss = multi_token_prediction_loss
|