File size: 29,871 Bytes
921c3d4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 | """AAM Diffusion LLM β Mirror Speculative Decoder
Uses the SAME model as both draft and verifier with different denoising
step counts, eliminating the need for a separate smaller draft model.
Standard Speculative Decoding:
βββββββββββ draft ββββββββββββ verify
β Small β βββββββββΊ β Large β βββββββββΊ accept/reject
β Model β β Model β
βββββββββββ ββββββββββββ
(separate model) (separate model)
Mirror Speculative Decoding (this module):
ββββββββββββββββββββββββββββββββββββ
β SAME Model β
β ββββββββββββ ββββββββββββββββ β
β β 1-step β β 3-step β β
β β denoise β β denoise β β
β β (draft) β β (verify) β β
β ββββββ¬ββββββ ββββββββ¬ββββββββ β
β β β β
β βββββββΊ accept/reject βββββββ
ββββββββββββββββββββββββββββββββββββ
Why Mirror Speculative for AAM?
- No separate draft model needed (saves memory and complexity)
- Graph conditioning provides strong priors, making 1-step draft
surprisingly accurate for sentence arrangement
- Works well with anchored diffusion (draft starts from meaningful
prediction, not random noise)
- The denoising trajectory is continuous: 1-step and 3-step are
points on the same trajectory, so they are inherently consistent
- For AAM, the "draft" is essentially the anchored prediction with
minimal refinement, while "verify" adds the full coherence pass
Architecture:
Draft Phase (1 diffusion step):
anchor_prediction β 1-step refine β draft logits β sample tokens
Verify Phase (3 diffusion steps):
anchor_prediction β 3-step refine β verify logits β compare
Accept/Reject:
For each draft token:
- If P_verify(token) >= P_draft(token) * threshold β ACCEPT
- Otherwise β REJECT, sample from verify distribution
Continue from first rejection point.
Speedup estimation:
If acceptance_rate = Ξ± and draft generates k tokens per verify pass,
speedup β k / (1 + (1-Ξ±) * k)
For Ξ±=0.8 and k=5: speedup β 2.8x
"""
from __future__ import annotations
import math
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
@dataclass
class MirrorSpeculativeConfig:
"""Configuration for Mirror Speculative Decoder.
Attributes:
draft_steps: Number of diffusion denoising steps for the draft
pass. Default=1 (single step from anchor prediction).
This is fast but approximate β suitable for generating
candidate tokens quickly.
verify_steps: Number of diffusion denoising steps for the
verification pass. Default=3 (full coherence refinement).
This is slower but more accurate, catching errors from
the fast draft pass.
acceptance_threshold: Probability threshold for accepting draft
tokens. A draft token is accepted if:
P_verify(token) >= P_draft(token) * (1 - threshold)
Lower threshold = more accepting (faster but less accurate).
Higher threshold = more rejecting (slower but more accurate).
Range: [0, 1). Default=0.1 (accept unless verify strongly
disagrees).
max_draft_tokens: Maximum number of tokens to generate per
draft pass. Higher values = potentially more speedup but
also more wasted computation on rejection.
temperature: Sampling temperature for token generation.
Higher = more diverse, lower = more deterministic.
d_model: Model hidden dimension (must match the diffusion model).
d_vocab: Vocabulary size (must match the tokenizer).
use_graph_conditioning: Whether to use graph encoder output
as additional conditioning during both draft and verify
passes. This is AAM-specific: the graph provides strong
structural priors that make even 1-step drafts accurate.
resample_rejected: Whether to resample rejected tokens from
the verify distribution (True) or simply use the verify
model's top-1 prediction (False).
"""
draft_steps: int = 1
verify_steps: int = 3
acceptance_threshold: float = 0.1
max_draft_tokens: int = 5
temperature: float = 1.0
d_model: int = 768
d_vocab: int = 32000
use_graph_conditioning: bool = True
resample_rejected: bool = True
class DraftVerifyHead(nn.Module):
"""Shared projection head for draft and verify logits.
Both draft and verify passes use the SAME projection weights
(this is the "mirror" aspect). The difference is only in the
number of denoising steps applied to the hidden states before
projection.
Architecture:
hidden_states β RMSNorm β Linear(d_model, d_vocab)
The RMSNorm ensures stable logit magnitudes regardless of the
number of denoising steps that produced the hidden states.
"""
def __init__(self, d_model: int, d_vocab: int) -> None:
super().__init__()
self.norm = nn.RMSNorm(d_model)
self.proj = nn.Linear(d_model, d_vocab, bias=False)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
"""Project hidden states to vocabulary logits.
Args:
hidden_states: Denoised hidden states of shape
(batch, seq_len, d_model).
Returns:
Logits of shape (batch, seq_len, d_vocab).
"""
return self.proj(self.norm(hidden_states))
class RefinementStep(nn.Module):
"""Single denoising refinement step for the mirror decoder.
Each refinement step takes the current hidden state estimate
and produces a refined version. This is essentially a
lightweight denoising network that operates in the model's
representation space.
For AAM, the refinement incorporates graph conditioning:
refined = x + gate * MLP(cat(x, graph_context_proj(x), step_emb(t)))
This allows each refinement step to leverage the knowledge
graph structure for more informed denoising.
"""
def __init__(
self,
d_model: int,
d_refine: Optional[int] = None,
use_graph_conditioning: bool = True,
) -> None:
super().__init__()
self.d_model = d_model
self.d_refine = d_refine or d_model * 2
self.use_graph_conditioning = use_graph_conditioning
# Step embedding
self.step_embed = nn.Sequential(
nn.Linear(d_model, d_model),
nn.SiLU(),
nn.Linear(d_model, d_model),
)
# Graph conditioning projection (AAM-specific)
if use_graph_conditioning:
self.graph_proj = nn.Sequential(
nn.Linear(d_model, d_model, bias=False),
nn.SiLU(),
nn.Linear(d_model, d_model, bias=False),
)
# Refinement MLP
input_dim = d_model * 3 if use_graph_conditioning else d_model * 2
self.refine_mlp = nn.Sequential(
nn.Linear(input_dim, self.d_refine, bias=False),
nn.SiLU(),
nn.Linear(self.d_refine, d_model, bias=False),
)
# Gating mechanism
self.gate = nn.Sequential(
nn.Linear(d_model, 1, bias=False),
nn.Sigmoid(),
)
self.norm = nn.RMSNorm(d_model)
@staticmethod
def sinusoidal_step_embedding(
step: int, d_model: int, device: torch.device
) -> torch.Tensor:
"""Generate sinusoidal embedding for the current step index."""
half_dim = d_model // 2
emb = math.log(10000) / (half_dim - 1)
emb = torch.exp(torch.arange(half_dim, device=device, dtype=torch.float) * -emb)
emb = torch.tensor([step], device=device, dtype=torch.float) * emb
emb = torch.cat([emb.sin(), emb.cos()], dim=-1)
if d_model % 2 == 1:
emb = F.pad(emb, (0, 1))
return emb.unsqueeze(0) # (1, d_model)
def forward(
self,
x: torch.Tensor,
step: int,
graph_context: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Apply one refinement step.
Args:
x: Current hidden state estimate (batch, seq_len, d_model).
step: Current refinement step index (0-based).
graph_context: Optional graph encoder output for AAM conditioning.
Returns:
Refined hidden state (batch, seq_len, d_model).
"""
batch_size, seq_len, _ = x.shape
# Step embedding
step_emb = self.sinusoidal_step_embedding(step, self.d_model, x.device)
step_emb = self.step_embed(step_emb)
step_emb = step_emb.expand(batch_size, seq_len, -1)
if self.use_graph_conditioning:
# Project graph context (use zeros if not provided)
if graph_context is not None:
graph_proj = self.graph_proj(graph_context)
# Mean-pool graph context if it has different seq dim
if graph_proj.shape[1] != seq_len:
graph_proj = graph_proj.mean(dim=1, keepdim=True).expand(
-1, seq_len, -1
)
else:
# Zero placeholder maintains consistent input dimension
graph_proj = torch.zeros(
batch_size, seq_len, self.d_model,
device=x.device, dtype=x.dtype,
)
# Concatenate all inputs
refine_input = torch.cat([x, step_emb, graph_proj], dim=-1)
else:
refine_input = torch.cat([x, step_emb], dim=-1)
# Refinement
refinement = self.refine_mlp(refine_input)
# Gated residual
gate = self.gate(x)
x = self.norm(x + gate * refinement)
return x
class MirrorSpeculativeDecoder(nn.Module):
"""Mirror Speculative Decoder for AAM Diffusion LLM.
Uses the same model with different denoising step counts:
- Draft: `draft_steps` diffusion steps (fast, approximate)
- Verify: `verify_steps` diffusion steps (slower, accurate)
- Accept/reject based on consistency between draft and verify
This is more efficient for AAM because:
- No need for a separate draft model (saves memory)
- Graph conditioning provides strong priors, making even
1-step drafts accurate for sentence arrangement
- Works well with anchored diffusion (draft starts from
a meaningful prediction, not random noise)
The decoder maintains shared refinement steps and a shared
projection head. The only difference between draft and verify
is how many refinement steps are applied.
Example usage:
>>> config = MirrorSpeculativeConfig(draft_steps=1, verify_steps=3)
>>> decoder = MirrorSpeculativeDecoder(config)
>>> # anchor_hidden from graph-conditioned initial prediction
>>> tokens, info = decoder(anchor_hidden, graph_context=graph_out)
>>> print(f"Acceptance rate: {info['acceptance_rate']:.2%}")
Args:
config: Mirror speculative decoder configuration.
"""
def __init__(self, config: Optional[MirrorSpeculativeConfig] = None) -> None:
super().__init__()
self.config = config or MirrorSpeculativeConfig()
self.draft_steps = self.config.draft_steps
self.verify_steps = self.config.verify_steps
self.max_draft_tokens = self.config.max_draft_tokens
self.temperature = self.config.temperature
self.acceptance_threshold = self.config.acceptance_threshold
# Max refinement steps needed (for verify pass)
max_steps = max(self.draft_steps, self.verify_steps)
# Shared refinement steps (the "mirror" β same weights for draft/verify)
self.refinement_steps = nn.ModuleList([
RefinementStep(
d_model=self.config.d_model,
use_graph_conditioning=self.config.use_graph_conditioning,
)
for _ in range(max_steps)
])
# Shared logits head (same weights for draft and verify)
self.logits_head = DraftVerifyHead(
d_model=self.config.d_model,
d_vocab=self.config.d_vocab,
)
def draft(
self,
anchor_hidden: torch.Tensor,
graph_context: Optional[torch.Tensor] = None,
n_tokens: Optional[int] = None,
temperature: Optional[float] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Generate draft tokens using minimal diffusion steps.
The draft pass applies only `draft_steps` refinement steps
(typically 1) to the anchor prediction, then samples tokens.
Args:
anchor_hidden: Initial hidden state prediction from the
graph-conditioned model, of shape (batch, seq_len, d_model).
This is the "anchor" β the model's best guess before
any refinement.
graph_context: Optional graph encoder output for AAM conditioning.
n_tokens: Number of draft tokens to generate. Defaults to
max_draft_tokens from config.
temperature: Sampling temperature. Defaults to config value.
Returns:
Tuple of:
- draft_token_ids: Sampled token IDs, shape (batch, n_tokens)
- draft_log_probs: Log probabilities of sampled tokens,
shape (batch, n_tokens)
- draft_hidden: Refined hidden states after draft steps,
shape (batch, seq_len, d_model)
"""
n_tokens = n_tokens or self.max_draft_tokens
temperature = temperature or self.temperature
# Apply draft_steps refinement steps
x = anchor_hidden
for step_idx in range(self.draft_steps):
x = self.refinement_steps[step_idx](x, step=step_idx, graph_context=graph_context)
# Project to logits
logits = self.logits_head(x) # (batch, seq_len, d_vocab)
# Sample tokens from the last n_tokens positions
# (or from the full sequence if it's shorter)
draft_logits = logits[:, -n_tokens:, :] # (batch, n_tokens, d_vocab)
# Temperature scaling and sampling
scaled_logits = draft_logits / temperature
log_probs = F.log_softmax(scaled_logits, dim=-1)
probs = torch.exp(log_probs)
# Sample tokens
draft_token_ids = torch.multinomial(
probs.reshape(-1, self.config.d_vocab), 1
).reshape(probs.shape[0], probs.shape[1])
# Gather log probs for sampled tokens
draft_log_probs = log_probs.gather(
-1, draft_token_ids.unsqueeze(-1)
).squeeze(-1)
return draft_token_ids, draft_log_probs, x
def verify(
self,
anchor_hidden: torch.Tensor,
draft_token_ids: torch.Tensor,
graph_context: Optional[torch.Tensor] = None,
temperature: Optional[float] = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""Re-score draft tokens using full diffusion steps.
The verify pass applies `verify_steps` refinement steps
(typically 3) to the same anchor prediction, then evaluates
the probability of the draft tokens under the more refined
distribution.
Args:
anchor_hidden: Same initial hidden state used for draft.
draft_token_ids: Draft token IDs to verify, shape
(batch, n_tokens).
graph_context: Optional graph encoder output.
temperature: Sampling temperature. Defaults to config value.
Returns:
Tuple of:
- verify_log_probs: Log probabilities of draft tokens
under the verified distribution, shape (batch, n_tokens)
- verify_logits: Full verified logits for the draft
positions, shape (batch, n_tokens, d_vocab)
"""
temperature = temperature or self.temperature
# Apply verify_steps refinement steps (more than draft)
x = anchor_hidden
for step_idx in range(self.verify_steps):
x = self.refinement_steps[step_idx](x, step=step_idx, graph_context=graph_context)
# Project to logits
logits = self.logits_head(x) # (batch, seq_len, d_vocab)
# Extract logits for draft positions
n_draft = draft_token_ids.shape[1]
verify_logits = logits[:, -n_draft:, :] # (batch, n_draft, d_vocab)
# Temperature scaling
scaled_logits = verify_logits / temperature
verify_log_probs_full = F.log_softmax(scaled_logits, dim=-1)
# Gather log probs for draft tokens
verify_log_probs = verify_log_probs_full.gather(
-1, draft_token_ids.unsqueeze(-1)
).squeeze(-1)
return verify_log_probs, verify_logits
def accept_reject(
self,
draft_token_ids: torch.Tensor,
draft_log_probs: torch.Tensor,
verify_log_probs: torch.Tensor,
verify_logits: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""Compare draft vs verify probabilities and accept matching tokens.
The acceptance criterion follows the standard speculative decoding
rule, adapted for the mirror setting:
Accept token t if:
P_verify(t) >= P_draft(t) * (1 - acceptance_threshold)
This is equivalent to:
exp(verify_log_prob - draft_log_prob) >= (1 - threshold)
When a token is rejected:
- If resample_rejected: sample from the verify distribution
at that position
- Otherwise: take argmax of the verify distribution
Acceptance stops at the first rejection β all subsequent tokens
are discarded, and generation continues from that position.
Args:
draft_token_ids: Draft token IDs, shape (batch, n_tokens).
draft_log_probs: Draft log probabilities, shape (batch, n_tokens).
verify_log_probs: Verify log probabilities at draft token
positions, shape (batch, n_tokens).
verify_logits: Full verify logits, shape (batch, n_tokens, d_vocab).
Returns:
Tuple of:
- accepted_tokens: Final accepted token sequence,
shape (batch, n_tokens). Positions after first rejection
are filled with the resampled/argmax token from verify.
- accepted_mask: Boolean mask of accepted positions,
shape (batch, n_tokens). True = accepted from draft.
- first_rejection_pos: Index of first rejection per batch,
shape (batch,). Equals n_tokens if all accepted.
"""
batch_size, n_tokens = draft_token_ids.shape
# Compute acceptance probability
log_ratio = verify_log_probs - draft_log_probs
ratio = torch.exp(log_ratio) # P_verify / P_draft
accept_threshold = 1.0 - self.acceptance_threshold
# Per-position accept decision
per_position_accept = ratio >= accept_threshold # (batch, n_tokens)
# Find first rejection position per batch item
# If all accepted, first_rejection = n_tokens
rejection_mask = ~per_position_accept # True where rejected
if rejection_mask.any():
# For each batch, find first rejection
first_rejection_pos = torch.full(
(batch_size,), n_tokens, dtype=torch.long, device=draft_token_ids.device
)
for b in range(batch_size):
rejected_positions = rejection_mask[b].nonzero(as_tuple=True)[0]
if len(rejected_positions) > 0:
first_rejection_pos[b] = rejected_positions[0].item()
else:
first_rejection_pos = torch.full(
(batch_size,), n_tokens, dtype=torch.long, device=draft_token_ids.device
)
# Build accepted mask: accept all positions before first rejection
# At the first rejection position, we use the verify distribution
position_indices = torch.arange(n_tokens, device=draft_token_ids.device).unsqueeze(0)
accepted_mask = position_indices < first_rejection_pos.unsqueeze(1) # (batch, n_tokens)
# Also include the first rejection position (resampled from verify)
rejection_position_mask = position_indices == first_rejection_pos.unsqueeze(1)
included_mask = accepted_mask | rejection_position_mask # (batch, n_tokens)
# Resample rejected positions from verify distribution
verify_probs = F.softmax(verify_logits, dim=-1) # (batch, n_tokens, d_vocab)
if self.config.resample_rejected:
# Sample from adjusted verify distribution at rejection positions
resampled_ids = torch.multinomial(
verify_probs.reshape(-1, self.config.d_vocab), 1
).reshape(batch_size, n_tokens)
else:
# Argmax from verify distribution
resampled_ids = verify_logits.argmax(dim=-1) # (batch, n_tokens)
# Combine: use draft tokens where accepted, resampled at first rejection
accepted_tokens = torch.where(
accepted_mask,
draft_token_ids,
resampled_ids,
)
# Zero out positions after first rejection + 1
valid_mask = included_mask
accepted_tokens = accepted_tokens * valid_mask.long()
return accepted_tokens, accepted_mask, first_rejection_pos
def forward(
self,
anchor_hidden: torch.Tensor,
graph_context: Optional[torch.Tensor] = None,
n_iterations: int = 1,
) -> Tuple[torch.Tensor, Dict[str, object]]:
"""Full mirror speculative decoding loop.
Iterates:
1. Generate draft tokens (fast, draft_steps)
2. Verify against full model (slow, verify_steps)
3. Accept matching prefix, reject divergent tokens
4. Continue from first rejection point
5. Repeat until max_draft_tokens reached or n_iterations exhausted
Args:
anchor_hidden: Initial hidden state prediction from
graph-conditioned model, shape (batch, seq_len, d_model).
graph_context: Optional graph encoder output for AAM
conditioning, shape (batch, n_nodes, d_model).
n_iterations: Number of draft-verify iterations. Each
iteration generates up to max_draft_tokens draft tokens.
Returns:
Tuple of:
- all_accepted_tokens: Concatenated accepted tokens across
all iterations, shape (batch, total_accepted).
- info: Dictionary containing:
- "acceptance_rate": Fraction of draft tokens accepted
- "total_draft_tokens": Total draft tokens generated
- "total_accepted_tokens": Total tokens accepted
- "speedup_estimate": Estimated speedup vs autoregressive
- "iterations": Per-iteration statistics
"""
batch_size = anchor_hidden.shape[0]
device = anchor_hidden.device
all_accepted_tokens: List[torch.Tensor] = []
iteration_stats: List[Dict[str, float]] = []
total_draft = 0
total_accepted = 0
current_hidden = anchor_hidden
for iteration in range(n_iterations):
# Step 1: Draft
draft_ids, draft_log_probs, draft_hidden = self.draft(
current_hidden, graph_context=graph_context
)
# Step 2: Verify
verify_log_probs, verify_logits = self.verify(
current_hidden, draft_ids, graph_context=graph_context
)
# Step 3: Accept/Reject
accepted_tokens, accepted_mask, first_rejection = self.accept_reject(
draft_ids, draft_log_probs, verify_log_probs, verify_logits
)
# Count accepted tokens per batch item
n_accepted_per_item = first_rejection + 1 # include first rejection point
n_draft_tokens = draft_ids.shape[1]
# Collect accepted tokens (up to first rejection + 1)
# For simplicity, use the minimum across batch
n_to_keep = n_accepted_per_item.min().item()
n_to_keep = min(n_to_keep, n_draft_tokens)
all_accepted_tokens.append(accepted_tokens[:, :n_to_keep])
# Track stats
iter_accepted = n_to_keep
total_draft += n_draft_tokens
total_accepted += iter_accepted
iter_rate = iter_accepted / max(n_draft_tokens, 1)
iteration_stats.append({
"iteration": iteration,
"n_draft": n_draft_tokens,
"n_accepted": iter_accepted,
"acceptance_rate": iter_rate,
"first_rejection": first_rejection.float().mean().item(),
})
# Step 4: Update hidden state for next iteration
# In a full implementation, we would append accepted tokens
# and re-encode. Here we use the draft hidden as a placeholder.
current_hidden = draft_hidden
# Early exit: if all draft tokens accepted, we can continue
# If none accepted, something is wrong β still continue
# Concatenate all accepted tokens
if all_accepted_tokens:
all_tokens = torch.cat(all_accepted_tokens, dim=1)
else:
all_tokens = torch.zeros(
batch_size, 0, dtype=torch.long, device=device
)
# Compute overall statistics
overall_acceptance_rate = total_accepted / max(total_draft, 1)
speedup = compute_acceptance_rate(
acceptance_rate=overall_acceptance_rate,
n_draft_tokens=self.max_draft_tokens,
)
info: Dict[str, object] = {
"acceptance_rate": overall_acceptance_rate,
"total_draft_tokens": total_draft,
"total_accepted_tokens": total_accepted,
"speedup_estimate": speedup,
"iterations": iteration_stats,
"draft_steps": self.draft_steps,
"verify_steps": self.verify_steps,
}
return all_tokens, info
def compute_acceptance_rate(
acceptance_rate: float,
n_draft_tokens: int = 5,
verify_cost_ratio: Optional[float] = None,
) -> float:
"""Compute estimated speedup from mirror speculative decoding.
The speedup from speculative decoding depends on:
1. The acceptance rate (Ξ±) β fraction of draft tokens accepted
2. The number of draft tokens per iteration (k)
3. The cost ratio between draft and verify passes
For mirror speculative decoding, the cost ratio is simply:
verify_steps / draft_steps (e.g., 3/1 = 3x)
Speedup formula (adapted from Leviathan et al., 2023):
If acceptance rate = Ξ±, draft tokens = k, cost ratio = Ξ³:
- Expected tokens per iteration: 1 + Ξ± * k / (1 - Ξ±)
(the "1" accounts for the always-verified rejection token)
- Cost per iteration: Ξ³ + 1 (verify is Ξ³x more expensive than draft)
- Speedup = expected_tokens / cost_per_iteration
For the mirror case where both passes use the same model:
Ξ³ = verify_steps / draft_steps
Args:
acceptance_rate: Fraction of draft tokens accepted (0 to 1).
n_draft_tokens: Number of draft tokens per iteration.
verify_cost_ratio: Cost ratio of verify vs draft pass. If None,
computed as verify_steps/draft_steps (mirror assumption).
Returns:
Estimated speedup factor. For example, 2.0 means the
speculative decoding is 2x faster than autoregressive.
Examples:
>>> compute_acceptance_rate(0.8, n_draft_tokens=5)
2.8 # approximately
>>> compute_acceptance_rate(1.0, n_draft_tokens=5)
5.0 # perfect acceptance = full draft speedup
>>> compute_acceptance_rate(0.0, n_draft_tokens=5)
0.5 # all rejected, just adding overhead
"""
alpha = max(0.0, min(1.0, acceptance_rate))
k = max(1, n_draft_tokens)
# Default mirror cost ratio: verify_steps / draft_steps = 3/1
if verify_cost_ratio is None:
verify_cost_ratio = 3.0 # verify_steps=3, draft_steps=1
gamma = max(1.0, verify_cost_ratio)
if alpha >= 1.0:
# Perfect acceptance: all k tokens accepted, cost = 1 + gamma
# But we get k tokens per iteration
return k / (1.0 + gamma)
if alpha <= 0.0:
# No acceptance: we only get 1 token (from verify at rejection)
# Cost = 1 (draft) + gamma (verify)
return 1.0 / (1.0 + gamma)
# Expected tokens accepted per iteration
# Following the geometric distribution from speculative decoding theory:
# E[tokens] = (1 + alpha * k) / (1 - alpha + 1/k)
# Simplified approximation:
expected_tokens = 1.0 + alpha * k
# Cost per iteration: 1 draft pass + 1 verify pass
cost = 1.0 + gamma
speedup = expected_tokens / cost
return round(speedup, 2)
|