Spaces:
Running on Zero
Running on Zero
File size: 20,067 Bytes
b8c861f | 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 | # Copyright 2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from __future__ import annotations
from dataclasses import dataclass
import torch
from ..configuration_utils import ConfigMixin, register_to_config
from ..utils import BaseOutput
from .scheduling_utils import SchedulerMixin
@dataclass
class BlockRefinementSchedulerOutput(BaseOutput):
"""
Output class for block refinement scheduling.
Args:
prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`):
Updated block tokens after the current refinement step.
transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`):
Boolean mask indicating which tokens were committed (mask-filling).
editing_transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`):
Boolean mask indicating which tokens were edited (non-mask replacement).
sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`):
Sampled token IDs from the model logits.
sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`):
Probabilities of the sampled tokens.
"""
prev_sample: torch.LongTensor
transfer_index: torch.BoolTensor
editing_transfer_index: torch.BoolTensor
sampled_tokens: torch.LongTensor
sampled_probs: torch.Tensor
class BlockRefinementScheduler(SchedulerMixin, ConfigMixin):
"""
Scheduler for block-wise iterative refinement (commit-by-confidence).
At each step, the scheduler samples candidate tokens from model logits and commits those with the highest
confidence. The number of tokens to commit per step is determined by evenly distributing the block length across
the number of refinement steps.
Optionally supports editing: after all mask tokens are resolved, tokens can be replaced if the model predicts a
different token with confidence above a positive `editing_threshold` (`None`, `0.0`, or negative disables editing).
"""
order = 1
@register_to_config
def __init__(
self,
block_length: int = 32,
num_inference_steps: int = 32,
threshold: float = 0.95,
editing_threshold: float | None = None,
minimal_topk: int = 1,
):
self.num_inference_steps = num_inference_steps
self.timesteps = torch.arange(self.num_inference_steps - 1, -1, -1, dtype=torch.long)
self._transfer_schedule: torch.LongTensor | None = None
def set_timesteps(
self,
num_inference_steps: int,
device: str | torch.device | None = None,
block_length: int | None = None,
) -> None:
if num_inference_steps <= 0:
raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.")
if block_length is None:
block_length = self.config.block_length
elif block_length <= 0:
raise ValueError(f"`block_length` must be > 0, got {block_length}.")
self.num_inference_steps = num_inference_steps
self.timesteps = torch.arange(self.num_inference_steps - 1, -1, -1, device=device, dtype=torch.long)
self._transfer_schedule = self.get_num_transfer_tokens(block_length, self.num_inference_steps).to(
device=device if device is not None else "cpu"
)
def get_num_transfer_tokens(self, block_length: int, num_inference_steps: int) -> torch.LongTensor:
"""Evenly distribute `block_length` token commits across `num_inference_steps` steps."""
if num_inference_steps <= 0:
return torch.zeros((0,), dtype=torch.long)
base = block_length // num_inference_steps
remainder = block_length % num_inference_steps
out = torch.full((num_inference_steps,), base, dtype=torch.long)
out[:remainder] += 1
return out
# --- SAR sampling utilities ---
@staticmethod
def _top_p_filtering(logits: torch.Tensor, top_p: float | None) -> torch.Tensor:
"""Nucleus (top-p) logit filtering."""
if top_p is None or top_p >= 1.0:
return logits
if not (0.0 < top_p <= 1.0):
raise ValueError(f"`top_p` must be in (0, 1], got {top_p}.")
sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
sorted_probs = torch.softmax(sorted_logits, dim=-1)
cumulative_probs = sorted_probs.cumsum(dim=-1)
sorted_indices_to_remove = cumulative_probs > float(top_p)
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
sorted_logits = sorted_logits.masked_fill(sorted_indices_to_remove, torch.finfo(sorted_logits.dtype).min)
filtered = logits.scatter(-1, sorted_indices, sorted_logits)
return filtered
@staticmethod
def _top_k_filtering(logits: torch.Tensor, top_k: int | None) -> torch.Tensor:
"""Top-k logit filtering."""
if top_k is None or top_k <= 0:
return logits
if top_k >= logits.shape[-1]:
return logits
values, _ = torch.topk(logits, k=top_k, dim=-1)
min_keep = values[..., -1, None]
return logits.masked_fill(logits < min_keep, torch.finfo(logits.dtype).min)
@staticmethod
def _sample_from_logits(
logits: torch.Tensor,
*,
temperature: float,
top_k: int | None,
top_p: float | None,
generator: torch.Generator | None,
use_multinomial: bool,
) -> tuple[torch.LongTensor, torch.Tensor]:
"""Sample tokens from logits with temperature scaling, top-k, and top-p."""
if temperature < 0:
raise ValueError(f"`temperature` must be >= 0, got {temperature}.")
vocab_size = logits.shape[-1]
flat_logits = logits.reshape(-1, vocab_size)
if temperature == 0.0 or not use_multinomial:
probs = torch.softmax(flat_logits.float(), dim=-1)
token = flat_logits.argmax(dim=-1, keepdim=True)
token_prob = torch.gather(probs, -1, token)
return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1])
scaled = flat_logits
if temperature != 1.0:
scaled = flat_logits / temperature
filtered = BlockRefinementScheduler._top_k_filtering(scaled, top_k=top_k)
filtered = BlockRefinementScheduler._top_p_filtering(filtered, top_p=top_p)
probs = torch.softmax(filtered.float(), dim=-1)
token = torch.multinomial(probs, num_samples=1, generator=generator)
token_prob = torch.gather(probs, -1, token)
return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1])
def step(
self,
model_output: torch.Tensor,
timestep: int | torch.Tensor,
sample: torch.LongTensor,
*,
mask_token_id: int,
temperature: float = 0.0,
top_p: float | None = None,
top_k: int | None = None,
sampling_method: str = "auto",
threshold: float | None = None,
editing_threshold: float | None = None,
minimal_topk: int | None = None,
prompt_mask: torch.BoolTensor | None = None,
generator: torch.Generator | None = None,
return_dict: bool = True,
) -> (
BlockRefinementSchedulerOutput
| tuple[torch.LongTensor, torch.BoolTensor, torch.BoolTensor, torch.LongTensor, torch.Tensor]
):
"""
Perform a single refinement step: sample from logits, commit confident tokens, and optionally edit existing
ones.
Args:
model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`):
Raw logits from the model for the current block.
timestep (`int` or `torch.Tensor`):
Current step index within the block's refinement schedule.
sample (`torch.LongTensor` of shape `(batch_size, block_length)`):
Current block token IDs (contains mask tokens for uncommitted positions).
mask_token_id (`int`):
Token ID used for masked positions.
temperature (`float`):
Sampling temperature.
top_p (`float`, *optional*):
Nucleus sampling cutoff.
top_k (`int`, *optional*):
Top-k sampling cutoff.
sampling_method (`str`):
Sampling method (`auto`, `greedy`, `multinomial`).
threshold (`float`, *optional*):
Confidence threshold for committing tokens. Defaults to config value.
editing_threshold (`float`, *optional*):
Confidence threshold for editing non-mask tokens; must be positive to enable editing. Defaults to
config value.
minimal_topk (`int`, *optional*):
Minimum tokens to commit per step. Defaults to config value.
prompt_mask (`torch.BoolTensor`, *optional*):
Boolean mask of shape `(block_length,)` where `True` marks prompt (non-editable) positions.
generator (`torch.Generator`, *optional*):
RNG for sampling.
return_dict (`bool`):
Whether to return a `BlockRefinementSchedulerOutput` or a tuple.
"""
if threshold is None:
threshold = float(self.config.threshold)
if editing_threshold is None:
editing_threshold = self.config.editing_threshold
if minimal_topk is None:
minimal_topk = self.config.minimal_topk
# Sample from logits
use_multinomial = sampling_method == "multinomial" or (sampling_method == "auto" and temperature != 0.0)
sampled_tokens, sampled_probs = self._sample_from_logits(
model_output,
temperature=temperature,
top_k=top_k,
top_p=top_p,
generator=generator,
use_multinomial=use_multinomial,
)
batch_size, block_length = sample.shape
active_block = sample == mask_token_id
masks_remaining = active_block.any()
if isinstance(timestep, torch.Tensor):
step_index = int(timestep.item())
else:
step_index = int(timestep)
# --- Mask-filling transfer ---
transfer_index = torch.zeros_like(sampled_tokens, dtype=torch.bool)
if masks_remaining and self._transfer_schedule is not None:
clamped_step = min(step_index, len(self._transfer_schedule) - 1)
num_to_transfer = int(self._transfer_schedule[clamped_step].item())
confidence = torch.where(
active_block,
sampled_probs.to(dtype=torch.float32),
torch.full_like(sampled_probs, -torch.inf, dtype=torch.float32),
)
for b in range(batch_size):
high_conf = confidence[b] > threshold
if high_conf.sum().item() >= num_to_transfer:
transfer_index[b] = high_conf
else:
k = min(num_to_transfer, int(active_block[b].sum().item()))
if k > 0:
_, idx = torch.topk(confidence[b], k=k)
transfer_index[b, idx] = True
# --- Editing transfer (non-mask, non-prompt positions) ---
editing_enabled = editing_threshold is not None and editing_threshold > 0.0
editing_transfer_index = torch.zeros_like(sampled_tokens, dtype=torch.bool)
if editing_enabled:
if prompt_mask is None:
prompt_mask = torch.zeros(block_length, device=sample.device, dtype=torch.bool)
editable = (~active_block) & (~prompt_mask.unsqueeze(0))
editing_conf = torch.where(
editable,
sampled_probs.to(dtype=torch.float32),
torch.full_like(sampled_probs, -torch.inf, dtype=torch.float32),
)
high_conf_edit = editing_conf > float(editing_threshold)
token_changed = sampled_tokens != sample
editing_transfer_index = high_conf_edit & token_changed & editable
# Apply transfers
final_transfer = transfer_index | editing_transfer_index
prev_sample = sample.clone()
if final_transfer.any():
prev_sample[final_transfer] = sampled_tokens[final_transfer]
if not return_dict:
return prev_sample, transfer_index, editing_transfer_index, sampled_tokens, sampled_probs
return BlockRefinementSchedulerOutput(
prev_sample=prev_sample,
transfer_index=transfer_index,
editing_transfer_index=editing_transfer_index,
sampled_tokens=sampled_tokens,
sampled_probs=sampled_probs,
)
@staticmethod
def check_eos_finished(
cur_x: torch.LongTensor,
sampled_tokens: torch.LongTensor,
final_transfer: torch.BoolTensor,
finished: torch.BoolTensor,
eos_token_id: int,
mask_token_id: int,
prompt_length: int,
) -> torch.BoolTensor:
"""
Update per-batch finished flags when EOS tokens are committed.
Args:
cur_x (`torch.LongTensor` of shape `(batch_size, seq_len)`):
Current full sequence including all blocks up to the current window.
sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`):
Tokens sampled by the scheduler in this step.
final_transfer (`torch.BoolTensor` of shape `(batch_size, block_length)`):
Combined mask of committed and edited positions.
finished (`torch.BoolTensor` of shape `(batch_size,)`):
Current per-batch finished flags.
eos_token_id (`int`):
EOS token ID.
mask_token_id (`int`):
Mask token ID.
prompt_length (`int`):
Number of prompt tokens at the start of the sequence.
Returns:
`torch.BoolTensor`: Updated finished flags.
"""
batch_size = cur_x.shape[0]
for b in range(batch_size):
if finished[b]:
continue
eos_in_commits = (sampled_tokens[b][final_transfer[b]] == eos_token_id).any().item()
if not eos_in_commits:
continue
eos_pos = (cur_x[b] == eos_token_id).nonzero(as_tuple=True)
if len(eos_pos[0]) == 0:
continue
eos_pos = int(eos_pos[0][0].item())
# The first generated token sits at index `prompt_length`; allow EOS there.
if eos_pos < prompt_length:
continue
if (cur_x[b, prompt_length:eos_pos] != mask_token_id).all().item():
finished[b] = True
return finished
def check_block_should_continue(
self,
step_idx: int,
masks_remaining: bool,
editing_enabled: bool,
editing_transfer_index: torch.BoolTensor,
post_steps: int,
max_post_steps: int,
finished: torch.BoolTensor,
) -> bool:
"""
Determine whether the inner refinement loop should continue for the current block.
Args:
step_idx (`int`):
Current refinement step index within this block.
masks_remaining (`bool`):
Whether any mask tokens remain in the block.
editing_enabled (`bool`):
Whether editing mode is active.
editing_transfer_index (`torch.BoolTensor`):
Which tokens were edited in this step.
post_steps (`int`):
Number of post-mask editing steps taken so far.
max_post_steps (`int`):
Maximum allowed post-mask editing steps.
finished (`torch.BoolTensor`):
Per-batch finished flags (from EOS detection).
Returns:
`bool`: `True` if refinement should continue, `False` to break.
"""
if finished.all():
return False
if not masks_remaining and not editing_enabled:
return False
if not masks_remaining and not editing_transfer_index.any():
return False
if masks_remaining and step_idx >= self.num_inference_steps:
return False
if not masks_remaining and post_steps > max_post_steps:
return False
return True
def add_noise(
self,
original_samples: torch.LongTensor,
attention_mask: torch.LongTensor,
*,
prompt_length: int,
block_length: int,
mask_token_id: int,
generator: torch.Generator | None = None,
) -> tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor]:
"""
Apply the forward (noising) process for semi-autoregressive block masking.
For each block after the prompt, a random fraction of valid (non-padding) tokens are replaced with
`mask_token_id`. Two complementary views are returned: `noisy` and `noisy_rev`, where the masked positions in
one are the unmasked positions in the other.
Args:
original_samples (`torch.LongTensor` of shape `(batch_size, seq_len)`):
Clean token IDs.
attention_mask (`torch.LongTensor` of shape `(batch_size, seq_len)`):
Padding mask (1 for valid, 0 for padding).
prompt_length (`int`):
Number of leading prompt tokens to keep unmasked.
block_length (`int`):
Block size for masking.
mask_token_id (`int`):
Token ID to use for masked positions.
generator (`torch.Generator`, *optional*):
RNG for reproducibility.
Returns:
`tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor]`:
`(noisy, noisy_rev, masked, masked_rev)` — the two complementary noisy sequences and their
corresponding boolean masks.
"""
batch_size, seq_len = original_samples.shape
device = original_samples.device
noisy = original_samples.clone()
noisy_rev = original_samples.clone()
masked = torch.zeros_like(original_samples, dtype=torch.bool)
masked_rev = torch.zeros_like(original_samples, dtype=torch.bool)
valid = attention_mask.to(dtype=torch.bool)
for block_start in range(prompt_length, seq_len, block_length):
block_end = min(seq_len, block_start + block_length)
seg_len = block_end - block_start
if seg_len <= 0:
continue
p_mask = torch.rand((batch_size, 1), device=device, generator=generator)
seg = torch.rand((batch_size, seg_len), device=device, generator=generator) < p_mask
seg = seg & valid[:, block_start:block_end]
seg_rev = (~seg) & valid[:, block_start:block_end]
masked[:, block_start:block_end] = seg
masked_rev[:, block_start:block_end] = seg_rev
noisy = torch.where(masked, torch.full_like(noisy, mask_token_id), noisy)
noisy_rev = torch.where(masked_rev, torch.full_like(noisy_rev, mask_token_id), noisy_rev)
return noisy, noisy_rev, masked, masked_rev
__all__ = ["BlockRefinementScheduler", "BlockRefinementSchedulerOutput"]
|