File size: 14,405 Bytes
00ecf49 |
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 |
"""
Speculative Decoding Module for MiniMind Max2
Use small draft model to accelerate large model inference.
"""
from dataclasses import dataclass
from typing import List, Optional, Dict, Any, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
import time
@dataclass
class SpeculativeConfig:
"""Configuration for speculative decoding."""
# Speculation settings
num_speculative_tokens: int = 5 # Number of tokens to speculate
max_speculation_length: int = 8
# Acceptance settings
acceptance_method: str = "rejection" # rejection, nucleus
temperature: float = 1.0
top_p: float = 0.95
# Performance tuning
adaptive_speculation: bool = True # Adjust speculation based on acceptance rate
min_speculative_tokens: int = 2
max_speculative_tokens: int = 10
target_acceptance_rate: float = 0.8
class DraftModel:
"""
Wrapper for draft model in speculative decoding.
Typically a smaller, faster model (e.g., max2-nano for max2-pro).
"""
def __init__(
self,
model: nn.Module,
tokenizer = None,
device: str = "cuda",
):
self.model = model
self.tokenizer = tokenizer
self.device = device
self.model.eval()
@torch.no_grad()
def speculate(
self,
input_ids: torch.Tensor,
num_tokens: int = 5,
temperature: float = 1.0,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Generate speculative tokens.
Args:
input_ids: Current input sequence [batch, seq_len]
num_tokens: Number of tokens to speculate
temperature: Sampling temperature
Returns:
Tuple of (speculated_tokens, speculated_probs)
"""
batch_size = input_ids.shape[0]
speculated_tokens = []
speculated_probs = []
current_ids = input_ids
for _ in range(num_tokens):
# Forward pass
_, logits, _, _ = self.model(current_ids)
next_logits = logits[:, -1, :] / temperature
# Sample
probs = F.softmax(next_logits, dim=-1)
next_token = torch.multinomial(probs, num_samples=1)
# Get probability of selected token
token_prob = probs.gather(1, next_token)
speculated_tokens.append(next_token)
speculated_probs.append(token_prob)
# Append to sequence
current_ids = torch.cat([current_ids, next_token], dim=1)
# Stack results
speculated_tokens = torch.cat(speculated_tokens, dim=1) # [batch, num_tokens]
speculated_probs = torch.cat(speculated_probs, dim=1) # [batch, num_tokens]
return speculated_tokens, speculated_probs
class SpeculativeDecoder:
"""
Speculative decoding for accelerated generation.
Uses a small draft model to propose tokens, verified by target model.
"""
def __init__(
self,
target_model: nn.Module,
draft_model: nn.Module,
config: Optional[SpeculativeConfig] = None,
device: str = "cuda",
):
self.target = target_model
self.draft = DraftModel(draft_model, device=device)
self.config = config or SpeculativeConfig()
self.device = device
# Statistics
self.total_generated = 0
self.total_accepted = 0
self.speculation_lengths = []
def _rejection_sampling(
self,
draft_probs: torch.Tensor,
target_probs: torch.Tensor,
draft_tokens: torch.Tensor,
) -> Tuple[torch.Tensor, int]:
"""
Rejection sampling for token acceptance.
Returns:
Tuple of (accepted_mask, num_accepted)
"""
batch_size, num_tokens = draft_tokens.shape
# Compute acceptance probability: min(1, target_p / draft_p)
acceptance_probs = torch.min(
torch.ones_like(draft_probs),
target_probs / (draft_probs + 1e-10),
)
# Sample uniform for rejection test
uniform = torch.rand_like(acceptance_probs)
accepted = uniform < acceptance_probs
# Find first rejection point
accepted_mask = torch.cumprod(accepted.float(), dim=1).bool()
num_accepted = accepted_mask.sum(dim=1).min().item()
return accepted_mask, num_accepted
@torch.no_grad()
def generate_step(
self,
input_ids: torch.Tensor,
num_speculative: Optional[int] = None,
) -> Tuple[torch.Tensor, Dict[str, Any]]:
"""
Single speculative generation step.
Args:
input_ids: Current sequence [batch, seq_len]
num_speculative: Number of tokens to speculate (uses config if None)
Returns:
New tokens and statistics
"""
num_spec = num_speculative or self.config.num_speculative_tokens
# Phase 1: Draft model speculation
draft_tokens, draft_probs = self.draft.speculate(
input_ids,
num_tokens=num_spec,
temperature=self.config.temperature,
)
# Phase 2: Target model verification (single forward pass)
spec_input = torch.cat([input_ids, draft_tokens], dim=1)
_, target_logits, _, _ = self.target(spec_input)
# Get target probabilities for draft tokens
target_probs = F.softmax(target_logits[:, -num_spec-1:-1, :] / self.config.temperature, dim=-1)
target_probs_selected = target_probs.gather(2, draft_tokens.unsqueeze(-1)).squeeze(-1)
# Phase 3: Rejection sampling
accepted_mask, num_accepted = self._rejection_sampling(
draft_probs,
target_probs_selected,
draft_tokens,
)
# Accept verified tokens
if num_accepted > 0:
new_tokens = draft_tokens[:, :num_accepted]
else:
new_tokens = torch.empty(input_ids.shape[0], 0, dtype=torch.long, device=self.device)
# Sample one more token from target if not all accepted
if num_accepted < num_spec:
# Resample from target distribution at rejection point
next_logits = target_logits[:, input_ids.shape[1] + num_accepted - 1, :]
next_probs = F.softmax(next_logits / self.config.temperature, dim=-1)
bonus_token = torch.multinomial(next_probs, num_samples=1)
new_tokens = torch.cat([new_tokens, bonus_token], dim=1)
# Statistics
self.total_generated += new_tokens.shape[1]
self.total_accepted += num_accepted
self.speculation_lengths.append(num_spec)
stats = {
"num_speculated": num_spec,
"num_accepted": num_accepted,
"num_generated": new_tokens.shape[1],
"acceptance_rate": num_accepted / num_spec if num_spec > 0 else 0,
}
return new_tokens, stats
@torch.no_grad()
def generate(
self,
input_ids: torch.Tensor,
max_new_tokens: int = 100,
eos_token_id: Optional[int] = None,
) -> Tuple[torch.Tensor, Dict[str, Any]]:
"""
Full speculative generation.
Args:
input_ids: Initial input [batch, seq_len]
max_new_tokens: Maximum tokens to generate
eos_token_id: EOS token to stop generation
Returns:
Generated sequence and statistics
"""
self.target.eval()
generated = input_ids.clone()
total_stats = {
"steps": 0,
"tokens_generated": 0,
"acceptance_rates": [],
}
start_time = time.time()
num_speculative = self.config.num_speculative_tokens
while total_stats["tokens_generated"] < max_new_tokens:
# Speculative step
new_tokens, step_stats = self.generate_step(generated, num_speculative)
if new_tokens.shape[1] == 0:
break
generated = torch.cat([generated, new_tokens], dim=1)
# Update stats
total_stats["steps"] += 1
total_stats["tokens_generated"] += new_tokens.shape[1]
total_stats["acceptance_rates"].append(step_stats["acceptance_rate"])
# Check for EOS
if eos_token_id is not None and (new_tokens == eos_token_id).any():
break
# Adaptive speculation
if self.config.adaptive_speculation:
avg_acceptance = sum(total_stats["acceptance_rates"][-5:]) / min(5, len(total_stats["acceptance_rates"]))
if avg_acceptance > self.config.target_acceptance_rate:
num_speculative = min(num_speculative + 1, self.config.max_speculative_tokens)
elif avg_acceptance < self.config.target_acceptance_rate - 0.1:
num_speculative = max(num_speculative - 1, self.config.min_speculative_tokens)
end_time = time.time()
total_stats["time_seconds"] = end_time - start_time
total_stats["tokens_per_second"] = total_stats["tokens_generated"] / total_stats["time_seconds"]
total_stats["avg_acceptance_rate"] = sum(total_stats["acceptance_rates"]) / max(1, len(total_stats["acceptance_rates"]))
total_stats["avg_tokens_per_step"] = total_stats["tokens_generated"] / max(1, total_stats["steps"])
return generated, total_stats
def get_statistics(self) -> Dict[str, float]:
"""Get overall statistics."""
return {
"total_generated": self.total_generated,
"total_accepted": self.total_accepted,
"overall_acceptance_rate": self.total_accepted / max(1, self.total_generated),
"avg_speculation_length": sum(self.speculation_lengths) / max(1, len(self.speculation_lengths)),
}
def reset_statistics(self):
"""Reset statistics counters."""
self.total_generated = 0
self.total_accepted = 0
self.speculation_lengths = []
class TreeSpeculativeDecoder(SpeculativeDecoder):
"""
Tree-based speculative decoding for higher acceptance rates.
Generates multiple speculation branches.
"""
def __init__(
self,
target_model: nn.Module,
draft_model: nn.Module,
num_branches: int = 3,
config: Optional[SpeculativeConfig] = None,
device: str = "cuda",
):
super().__init__(target_model, draft_model, config, device)
self.num_branches = num_branches
@torch.no_grad()
def generate_tree(
self,
input_ids: torch.Tensor,
depth: int = 3,
) -> List[Tuple[torch.Tensor, torch.Tensor]]:
"""
Generate tree of speculative tokens.
Returns:
List of (tokens, probs) tuples for each branch
"""
branches = []
# Generate multiple branches from draft model
for _ in range(self.num_branches):
tokens, probs = self.draft.speculate(
input_ids,
num_tokens=depth,
temperature=self.config.temperature,
)
branches.append((tokens, probs))
return branches
@torch.no_grad()
def generate_step(
self,
input_ids: torch.Tensor,
num_speculative: Optional[int] = None,
) -> Tuple[torch.Tensor, Dict[str, Any]]:
"""Tree-based speculative step."""
num_spec = num_speculative or self.config.num_speculative_tokens
# Generate tree of speculations
branches = self.generate_tree(input_ids, num_spec)
best_tokens = None
best_accepted = 0
# Verify each branch and pick best
for draft_tokens, draft_probs in branches:
spec_input = torch.cat([input_ids, draft_tokens], dim=1)
_, target_logits, _, _ = self.target(spec_input)
target_probs = F.softmax(
target_logits[:, -num_spec-1:-1, :] / self.config.temperature, dim=-1
)
target_probs_selected = target_probs.gather(2, draft_tokens.unsqueeze(-1)).squeeze(-1)
_, num_accepted = self._rejection_sampling(
draft_probs,
target_probs_selected,
draft_tokens,
)
if num_accepted > best_accepted:
best_accepted = num_accepted
best_tokens = draft_tokens[:, :num_accepted]
if best_tokens is None or best_tokens.shape[1] == 0:
# Fallback: sample from target
_, logits, _, _ = self.target(input_ids)
probs = F.softmax(logits[:, -1, :] / self.config.temperature, dim=-1)
best_tokens = torch.multinomial(probs, num_samples=1)
best_accepted = 0
stats = {
"num_speculated": num_spec * self.num_branches,
"num_accepted": best_accepted,
"num_generated": best_tokens.shape[1],
"acceptance_rate": best_accepted / num_spec if num_spec > 0 else 0,
"num_branches": self.num_branches,
}
return best_tokens, stats
def benchmark_speculative_decoding(
target_model: nn.Module,
draft_model: nn.Module,
input_ids: torch.Tensor,
num_tokens: int = 100,
device: str = "cuda",
) -> Dict[str, Any]:
"""
Benchmark speculative decoding vs standard generation.
"""
import time
# Standard generation
target_model.eval()
start = time.time()
with torch.no_grad():
standard_output = target_model.generate(
input_ids,
max_new_tokens=num_tokens,
)
standard_time = time.time() - start
# Speculative generation
decoder = SpeculativeDecoder(target_model, draft_model, device=device)
start = time.time()
spec_output, spec_stats = decoder.generate(
input_ids,
max_new_tokens=num_tokens,
)
spec_time = time.time() - start
return {
"standard": {
"time": standard_time,
"tokens_per_second": num_tokens / standard_time,
},
"speculative": {
"time": spec_time,
"tokens_per_second": spec_stats["tokens_per_second"],
"acceptance_rate": spec_stats["avg_acceptance_rate"],
"avg_tokens_per_step": spec_stats["avg_tokens_per_step"],
},
"speedup": standard_time / spec_time,
}
|