Spaces:
Sleeping
Sleeping
File size: 11,711 Bytes
a5e2743 | 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 | """
Speculative Decoding β Leviathan et al., 2023
"Fast Inference from Transformers via Speculative Decoding"
Core idea:
A small, fast "draft" model proposes K tokens autoregressively.
The large "target" model scores ALL K tokens in a single forward pass.
We accept/reject each proposed token using rejection sampling.
Expected speedup: ~K Γ acceptance_rate
Where acceptance_rate depends on how well the draft approximates the target.
Why it's LOSSLESS:
The rejection sampling scheme guarantees that the output distribution
is EXACTLY the same as standard autoregressive sampling from the target
model alone. This is not an approximation β it's mathematically exact.
Proof sketch:
- If p_target(x) >= p_draft(x): always accept β same as target
- If p_target(x) < p_draft(x): accept with prob p_target/p_draft
- On rejection: sample from residual distribution max(0, p_target - p_draft)
- This is classic rejection sampling from probability theory
Draft model strategy:
We use a "truncated" version of GPT-2: same architecture but only
the first N layers (e.g., 4 out of 12). This shares the same weights
and tokenizer, so there's no conversion needed.
"""
import torch
import torch.nn.functional as F
from torch import Tensor
from typing import Tuple, Optional, List
from dataclasses import dataclass, field
from model.transformer import GPT2, TransformerBlock
from model.layernorm import LayerNorm
from model.embedding import GPT2Embedding
from generation.sampling import sample_token
from config import ModelConfig
@dataclass
class SpeculativeMetrics:
"""Track speculative decoding statistics."""
total_proposed: int = 0
total_accepted: int = 0
total_steps: int = 0
@property
def acceptance_rate(self) -> float:
if self.total_proposed == 0:
return 0.0
return self.total_accepted / self.total_proposed
@property
def avg_accepted_per_step(self) -> float:
if self.total_steps == 0:
return 0.0
return self.total_accepted / self.total_steps
def summary(self) -> str:
return (
f"Spec decode: {self.acceptance_rate:.1%} acceptance rate, "
f"{self.avg_accepted_per_step:.1f} tokens/step avg "
f"({self.total_accepted}/{self.total_proposed} accepted)"
)
class DraftModel(GPT2):
"""
Draft model: first N layers of GPT-2.
Uses the same weights as the target model for shared layers.
Fewer layers = faster inference, but less accurate predictions.
The key insight is that early layers capture most of the "easy"
token predictions (common words, simple patterns). Later layers
refine the distribution for harder predictions. So a 4-layer model
can often predict the same token as a 12-layer model.
"""
@classmethod
def from_target_model(cls, target: GPT2, n_layers: int) -> "DraftModel":
"""
Create a draft model by taking the first n_layers from the target.
Shares weights β no additional memory for shared parameters.
"""
config = ModelConfig(
vocab_size=target.config.vocab_size,
d_model=target.config.d_model,
n_heads=target.config.n_heads,
n_layers=n_layers,
d_ff=target.config.d_ff,
max_seq_len=target.config.max_seq_len,
)
draft = cls(config)
# Share embedding weights
draft.embedding = target.embedding
# Copy first n_layers of transformer blocks
for i in range(n_layers):
draft.blocks[i] = target.blocks[i]
# Share final layernorm and LM head
draft.ln_f = target.ln_f
draft.lm_head = target.lm_head
draft.eval()
return draft
class SpeculativeDecoder:
"""
Speculative decoding engine.
Algorithm per step:
1. Draft model generates K tokens autoregressively (fast)
2. Target model scores all K tokens in ONE forward pass (parallel)
3. Accept/reject each token using rejection sampling
4. On first rejection: sample corrected token from residual distribution
5. Return all accepted tokens + corrected token
This produces 1 to K+1 tokens per "step" instead of exactly 1,
with the same distribution as standard autoregressive decoding.
"""
def __init__(
self,
target_model: GPT2,
draft_model: DraftModel,
K: int = 5, # Lookahead: propose K tokens
temperature: float = 1.0,
):
self.target = target_model
self.draft = draft_model
self.K = K
self.temperature = temperature
self.metrics = SpeculativeMetrics()
self.device = next(target_model.parameters()).device
@torch.no_grad()
def speculative_step(
self,
input_ids: Tensor, # [B, current_len]
target_caches: Optional[List], # KV caches for target model
draft_caches: Optional[List], # KV caches for draft model
) -> Tuple[Tensor, List, List]:
"""
One speculative decoding step.
Returns:
new_tokens: [B, n_accepted] β between 1 and K+1 tokens
updated target_caches
updated draft_caches
"""
B = input_ids.shape[0]
assert B == 1, "Speculative decoding currently supports batch_size=1"
# ============================================================
# PHASE 1: Draft model proposes K tokens (autoregressive, fast)
# ============================================================
draft_tokens = [] # K proposed token ids
draft_probs_list = [] # K probability distributions
draft_input = input_ids[:, -1:] # Start from last token
for k in range(self.K):
draft_logits, draft_caches = self.draft(
draft_input, kv_caches=draft_caches, use_cache=True
)
draft_logits_last = draft_logits[:, -1, :]
if self.temperature > 0:
draft_logits_last = draft_logits_last / self.temperature
draft_probs = F.softmax(draft_logits_last, dim=-1) # [B, vocab]
# Sample from draft distribution
token = torch.multinomial(draft_probs, 1) # [B, 1]
draft_tokens.append(token)
draft_probs_list.append(draft_probs)
draft_input = token
draft_token_ids = torch.cat(draft_tokens, dim=1) # [B, K]
# ============================================================
# PHASE 2: Target model scores ALL K tokens at once (parallel)
# ============================================================
# Feed the last real token + all K draft tokens to the target
# This is ONE forward pass instead of K sequential passes
target_input = torch.cat([input_ids[:, -1:], draft_token_ids], dim=1) # [B, K+1]
target_logits, target_caches = self.target(
target_input, kv_caches=target_caches, use_cache=True
)
# target_logits: [B, K+1, vocab]
# Position 0: logits for token AFTER the last real token
# Position k: logits for token AFTER draft_token[k-1]
if self.temperature > 0:
target_logits = target_logits / self.temperature
# ============================================================
# PHASE 3: Rejection sampling β accept/reject each draft token
# ============================================================
accepted_tokens = []
n_accepted = 0
for k in range(self.K):
# Target's probability for this position
p_target = F.softmax(target_logits[:, k, :], dim=-1) # [B, vocab]
# Draft's probability for this position
p_draft = draft_probs_list[k] # [B, vocab]
# The proposed token
token_k = draft_tokens[k] # [B, 1]
# Get probabilities for the specific proposed token
p_t = p_target.gather(-1, token_k).squeeze(-1) # [B]
p_d = p_draft.gather(-1, token_k).squeeze(-1) # [B]
# Acceptance criterion:
# Accept with probability min(1, p_target(x) / p_draft(x))
#
# Intuition:
# - If target likes this token MORE than draft: always accept
# - If target likes this token LESS than draft: accept proportionally
accept_prob = torch.clamp(p_t / (p_d + 1e-10), max=1.0)
# Stochastic acceptance
r = torch.rand(B, device=self.device)
if (r < accept_prob).all():
# ACCEPT: this token matches target's preferences
accepted_tokens.append(token_k)
n_accepted += 1
else:
# REJECT: sample from the residual distribution
# p_residual(x) = normalize(max(0, p_target(x) - p_draft(x)))
#
# This is the mathematical key to why speculative decoding
# is LOSSLESS: the residual distribution corrects for any
# mismatch between draft and target.
residual = torch.clamp(p_target - p_draft, min=0)
residual_sum = residual.sum(dim=-1, keepdim=True)
if residual_sum.item() > 1e-10:
residual = residual / residual_sum
corrected_token = torch.multinomial(residual, 1)
else:
# Fallback: sample from target distribution
corrected_token = torch.multinomial(p_target, 1)
accepted_tokens.append(corrected_token)
n_accepted += 0 # The corrected token doesn't count as "accepted"
break # Stop at first rejection
else:
# ALL K tokens accepted β sample one more from target
# This is the "bonus" token from the last position's logits
p_last = F.softmax(target_logits[:, self.K, :], dim=-1)
bonus_token = torch.multinomial(p_last, 1)
accepted_tokens.append(bonus_token)
# Update metrics
self.metrics.total_proposed += self.K
self.metrics.total_accepted += n_accepted
self.metrics.total_steps += 1
# Concatenate all accepted (+ corrected/bonus) tokens
new_tokens = torch.cat(accepted_tokens, dim=1) # [B, 1..K+1]
# IMPORTANT: We need to truncate the KV caches to only include
# the tokens we actually accepted. The target processed K+1 tokens
# but we may have only accepted fewer.
n_to_keep = new_tokens.shape[1]
n_to_remove = (self.K + 1) - n_to_keep
if n_to_remove > 0 and target_caches is not None:
# Trim the last n_to_remove positions from each layer's cache
target_caches = [
(k[:, :, :-n_to_remove, :], v[:, :, :-n_to_remove, :])
if k is not None else None
for k, v in target_caches
]
# Also reset draft caches β draft will re-run from the new position
draft_caches = None
return new_tokens, target_caches, draft_caches
def reset_metrics(self):
self.metrics = SpeculativeMetrics()
|