File size: 14,559 Bytes
922bb4b | 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 | """
Block-wise Sampler for SAD.
Instead of token-wise adaptive decoding, this sampler operates on blocks of tokens.
Given a context length of 512 and block size of 8, we have 64 blocks.
Block-wise adaptive: Resolve entire blocks at once based on aggregate confidence.
Per-level confidence is computed via h × prototype inner products, matching
the approach used in SADSampler and SADBlockSampler.
"""
from typing import Dict, List, Optional, Tuple
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import tqdm.auto as tqdm
from .sampler import compute_confidence
class BlockWiseAdaptiveSampler(nn.Module):
"""
Block-wise adaptive skip/backoff sampler.
Divides sequence into blocks (e.g., 512 / 8 = 64 blocks).
Each iteration:
1. Run model forward → leaf_logits, h
2. Compute per-level confidence via h × prototype inner products
3. Aggregate to block-level confidence
4. Resolve entire blocks that pass threshold
Args:
model: SADModel
hierarchy: SoftAncestorHierarchy or HardAncestorHierarchy
tokenizer: HuggingFace tokenizer
block_size: number of tokens per block (default 8)
confidence_metric: "neg_entropy" or "max_prob"
thresholds: per-level thresholds [tau_leaf, tau_l1, ...]
block_agg: how to aggregate token confidences to block confidence:
"mean", "min", or "max"
"""
def __init__(
self,
model: nn.Module,
hierarchy,
tokenizer,
block_size: int = 8,
confidence_metric: str = "neg_entropy",
thresholds: Optional[List[float]] = None,
block_agg: str = "mean",
freeze_resolved_blocks: bool = True,
):
super().__init__()
self.model = model
self.hierarchy = hierarchy
self.tokenizer = tokenizer
self.block_size = block_size
self.confidence_metric = confidence_metric
self.block_agg = block_agg
self.freeze_resolved_blocks = freeze_resolved_blocks
num_levels = hierarchy.num_levels
if thresholds is None:
thresholds = [0.8] + [0.5] * (num_levels - 1)
self.thresholds = thresholds
assert block_size > 0, "block_size must be positive"
assert block_agg in ["mean", "min", "max"], f"Unknown block_agg: {block_agg}"
def _aggregate_block_confidence(
self, token_conf: torch.Tensor
) -> torch.Tensor:
"""
Aggregate token-level confidences to block-level.
Args:
token_conf: [B, S] per-token confidence
Returns:
block_conf: [B, num_blocks] per-block confidence
"""
B, S = token_conf.shape
block_size = self.block_size
num_blocks = math.ceil(S / block_size)
# Pad if necessary
pad_len = num_blocks * block_size - S
if pad_len > 0:
token_conf = F.pad(token_conf, (0, pad_len), value=0.0)
# Reshape to [B, num_blocks, block_size]
token_conf = token_conf.reshape(B, num_blocks, block_size)
# Aggregate
if self.block_agg == "mean":
return token_conf.mean(dim=-1) # [B, num_blocks]
elif self.block_agg == "min":
return token_conf.min(dim=-1).values
elif self.block_agg == "max":
return token_conf.max(dim=-1).values
else:
raise ValueError(f"Unknown block_agg: {self.block_agg}")
def _get_block_resolution_level(
self, block_conf: torch.Tensor
) -> torch.Tensor:
"""
Determine which level each block should resolve to.
Args:
block_conf: dict mapping level to [B, num_blocks] confidence
Returns:
resolve_level: [B, num_blocks] int, 0=leaf, 1=level1, etc., -1=unresolved
"""
B, num_blocks = block_conf[0].shape
device = block_conf[0].device
# Start with -1 (unresolved)
resolve_level = torch.full((B, num_blocks), -1, dtype=torch.long, device=device)
# Check from leaf (level 0) to coarsest
for level in range(len(self.thresholds)):
conf = block_conf[level] # [B, num_blocks]
tau = self.thresholds[level]
# Mark unresolved blocks that meet threshold
unresolved = (resolve_level == -1)
should_resolve = unresolved & (conf >= tau)
resolve_level[should_resolve] = level
return resolve_level
@torch.no_grad()
def generate(
self,
num_samples: int,
num_steps: int,
max_length: int = 512,
device=None,
show_progress: bool = True,
random_coarse_init: bool = False,
) -> Tuple[torch.Tensor, Dict]:
"""
Generate sequences using block-wise adaptive decoding.
Returns:
token_ids: [B, S] final tokens
stats: dict with block-level statistics
"""
if device is None:
device = next(self.model.parameters()).device
B, S = num_samples, max_length
block_size = self.block_size
num_blocks = math.ceil(S / block_size)
mask_id = self.tokenizer.mask_token_id
vocab_size = self.model.vocab_size
# Initialize
if random_coarse_init and self.hierarchy.num_levels >= 2:
K_top = self.hierarchy.level_sizes[-1]
offset = sum(self.hierarchy.level_sizes[:-1])
rand_coarse = torch.randint(0, K_top, (B, S), device=device) + offset
current_ids = rand_coarse
else:
current_ids = torch.full((B, S), mask_id, dtype=torch.long, device=device)
# Track block resolution
block_resolved = torch.zeros(B, num_blocks, dtype=torch.bool, device=device)
block_exit_levels = torch.full((B, num_blocks), -1, dtype=torch.long, device=device)
token_exit_levels = torch.full((B, S), self.hierarchy.num_levels - 1,
dtype=torch.long, device=device)
# Pre-compute block boundaries for efficiency
block_boundaries = [(i * block_size, min((i + 1) * block_size, S))
for i in range(num_blocks)]
ts = torch.linspace(1.0 - self.t_eps, self.t_eps, num_steps, device=device)
# Check if hierarchy is soft (needs embeddings)
is_soft = hasattr(self.hierarchy, 'prototypes') and \
any(p.requires_grad for p in self.hierarchy.parameters())
resolved_over_steps = []
for step_i in tqdm.trange(num_steps, desc="Block-wise SAD", disable=not show_progress):
t_val = ts[step_i]
t_batch = t_val.expand(B)
# Forward pass
leaf_logits, _ = self.model(input_ids=current_ids, t=t_batch)
leaf_logits[..., mask_id] = float('-inf')
p_leaf = leaf_logits.softmax(dim=-1) # [B, S, V]
# Project upward
if is_soft:
leaf_emb = self.model.get_leaf_embeddings()
assignments = self.hierarchy.get_all_assignments(leaf_emb)
else:
assignments = self.hierarchy.get_all_assignments()
p_levels = self.hierarchy.project_upward(p_leaf, assignments=assignments)
# p_levels[l-1] = p^(l): [B, S, K_l]
# Compute per-token confidence at each level
conf_leaf = compute_confidence(p_leaf, self.confidence_metric) # [B, S]
conf_levels = [
compute_confidence(p_l, self.confidence_metric)
for p_l in p_levels
] # List of [B, S]
# Aggregate to block-level confidence
block_conf = {
0: self._aggregate_block_confidence(conf_leaf),
}
for li, conf_l in enumerate(conf_levels, start=1):
block_conf[li] = self._aggregate_block_confidence(conf_l)
# Determine resolution level for each block
resolve_level = self._get_block_resolution_level(block_conf) # [B, num_blocks]
# Update blocks
new_ids = current_ids.clone()
for block_idx in range(num_blocks):
if block_resolved[:, block_idx].all():
continue
start, end = block_boundaries[block_idx]
level = resolve_level[:, block_idx] # [B]
for b in range(B):
lvl = level[b].item()
if lvl < 0:
continue # Not confident enough
# Resolve this block at level lvl
if lvl == 0:
# Leaf level: sample/greedily select tokens
block_p = p_leaf[b, start:end] # [block_size, V]
block_tokens = block_p.argmax(dim=-1) # [block_size]
new_ids[b, start:end] = block_tokens
token_exit_levels[b, start:end] = 0
else:
# Intermediate level: use projected distribution
block_p = p_levels[lvl - 1][b, start:end] # [block_size, K_l]
block_ancestors = block_p.argmax(dim=-1) # [block_size]
# Offset to extended vocab
offset = sum(self.hierarchy.level_sizes[:lvl])
new_ids[b, start:end] = block_ancestors + offset
token_exit_levels[b, start:end] = lvl
block_resolved[b, block_idx] = True
block_exit_levels[b, block_idx] = lvl
current_ids = new_ids
resolved_over_steps.append(block_resolved.float().mean().item())
if block_resolved.all():
if show_progress:
print(f"All blocks resolved at step {step_i + 1}")
break
# Final pass: force unresolved to leaf
unresolved_blocks = ~block_resolved
if unresolved_blocks.any():
leaf_logits, _ = self.model(
input_ids=current_ids,
t=torch.full((B,), self.t_eps, device=device),
)
leaf_logits[..., mask_id] = float('-inf')
final_tokens = leaf_logits.argmax(dim=-1)
for b in range(B):
for block_idx in range(num_blocks):
if not block_resolved[b, block_idx]:
start, end = block_boundaries[block_idx]
current_ids[b, start:end] = final_tokens[b, start:end]
token_exit_levels[b, start:end] = 0
block_exit_levels[b, block_idx] = 0
stats = {
"block_exit_levels": block_exit_levels.cpu(),
"token_exit_levels": token_exit_levels.cpu(),
"block_resolved_over_steps": resolved_over_steps,
"num_blocks": num_blocks,
"block_size": block_size,
}
return current_ids.cpu(), stats
class BlockWiseAdjacentSampler(nn.Module):
"""
Baseline block-wise sampler without adaptive skipping.
Decodes block by block in a fixed order (left-to-right or random).
"""
def __init__(
self,
model: nn.Module,
tokenizer,
block_size: int = 8,
t_eps: float = 1e-4,
):
super().__init__()
self.model = model
self.tokenizer = tokenizer
self.block_size = block_size
self.t_eps = t_eps
@torch.no_grad()
def generate(
self,
num_samples: int,
num_steps: int,
max_length: int = 512,
device=None,
show_progress: bool = True,
) -> Tuple[torch.Tensor, Dict]:
"""Simple block-wise decoding without adaptive skip."""
if device is None:
device = next(self.model.parameters()).device
B, S = num_samples, max_length
block_size = self.block_size
num_blocks = math.ceil(S / block_size)
mask_id = self.tokenizer.mask_token_id
current_ids = torch.full((B, S), mask_id, dtype=torch.long, device=device)
ts = torch.linspace(1.0 - self.t_eps, self.t_eps, num_steps, device=device)
# Assign steps per block
steps_per_block = max(1, num_steps // num_blocks)
for step_i in tqdm.trange(num_steps, desc="Block-adjacent", disable=not show_progress):
t_val = ts[step_i]
t_batch = t_val.expand(B)
leaf_logits, _ = self.model(input_ids=current_ids, t=t_batch)
leaf_logits[..., mask_id] = float('-inf')
# Determine which block to update
block_idx = min(step_i // steps_per_block, num_blocks - 1)
start = block_idx * block_size
end = min(start + block_size, S)
# Update only current block
block_logits = leaf_logits[:, start:end] # [B, block_size, V]
block_tokens = block_logits.argmax(dim=-1) # [B, block_size]
# Only update if still masked
is_masked = (current_ids[:, start:end] == mask_id)
current_ids[:, start:end] = torch.where(
is_masked, block_tokens, current_ids[:, start:end]
)
# Final fill
is_masked = (current_ids == mask_id)
if is_masked.any():
leaf_logits, _ = self.model(
input_ids=current_ids,
t=torch.full((B,), self.t_eps, device=device),
)
leaf_logits[..., mask_id] = float('-inf')
final_tokens = leaf_logits.argmax(dim=-1)
current_ids = torch.where(is_masked, final_tokens, current_ids)
stats = {
"block_exit_levels": torch.zeros(B, num_blocks, dtype=torch.long),
"token_exit_levels": torch.zeros(B, S, dtype=torch.long),
"num_blocks": num_blocks,
"block_size": block_size,
}
return current_ids.cpu(), stats
|