File size: 15,244 Bytes
9e3b8ca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2026 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

import math
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 DiscreteDDIMSchedulerOutput(BaseOutput):
    """
    Output class for the discrete DDIM scheduler.

    Args:
        prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`):
            Updated block tokens after the current denoising step.
        sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`):
            Token IDs sampled from the model logits, i.e. the predicted clean tokens `x0`.
        sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`):
            Probabilities of the sampled tokens.
        pred_logits (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`):
            The denoiser logits, passed through for self-conditioning the next step.
    """

    prev_sample: torch.LongTensor
    sampled_tokens: torch.LongTensor
    sampled_probs: torch.Tensor
    pred_logits: torch.Tensor


class DiscreteDDIMScheduler(SchedulerMixin, ConfigMixin):
    """
    Discrete DDIM scheduler for the uniform corruption process, following "Structured Denoising Diffusion Models in
    Discrete State-Spaces" (D3PM, https://huggingface.co/papers/2107.03006).

    On the linear schedule the survival probability of a clean token at time `t` is `alpha(t) = 1 - t`. One denoising
    step from time `t` to `s < t` samples every block position from the exact posterior `q(x_s | x_t, x0)`, which for
    the uniform kernel decomposes into three routes: jump to the predicted clean token `x0`, stay on the current token,
    or jump to a uniformly random token. Unlike masked diffusion, there is no mask token; uncommitted positions carry
    random tokens.

    An optional predictor-corrector mode follows "Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and
    Absorbing State Reformulation" via the leave-one-out (LOO) denoiser (https://huggingface.co/papers/2605.22765).
    When `corrector_steps > 0`, the pipeline runs that many Gibbs corrector sweeps after each predictor step (see
    [`~DiscreteDDIMScheduler.step_correct`]), resampling the least-confident positions from the one-coordinate
    conditional `Cat(alpha_s * x0_loo + (1 - alpha_s) / K)` while holding the rest fixed, which leaves the marginal
    `p_s` invariant and improves generation at no training cost.

    Args:
        num_inference_steps (`int`, defaults to 32):
            The number of denoising steps, defining the linear time grid the posterior is evaluated on.
        corrector_steps (`int`, defaults to 0):
            Number of Gibbs corrector sweeps run after each predictor step. `0` recovers plain ancestral DDIM sampling.
        corrector_k (`int`, defaults to 1):
            Number of positions resampled per corrector sweep.
        corrector_selection (`str`, defaults to `"lowest_log_margin"`):
            How the resampled positions are chosen: `"lowest_log_margin"`, `"lowest_maxprob"`, `"lowest_current_prob"`,
            or `"random"`.
        corrector_selection_tau (`float`, defaults to 1.0):
            Temperature of the Gumbel-top-k position selection (lower is greedier).
    """

    order = 1

    @register_to_config
    def __init__(
        self,
        num_inference_steps: int = 32,
        corrector_steps: int = 0,
        corrector_k: int = 1,
        corrector_selection: str = "lowest_log_margin",
        corrector_selection_tau: float = 1.0,
    ):
        self.num_inference_steps = num_inference_steps
        self.timesteps = torch.arange(num_inference_steps, dtype=torch.long)

    def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None) -> None:
        if num_inference_steps <= 0:
            raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.")
        self.num_inference_steps = num_inference_steps
        self.timesteps = torch.arange(num_inference_steps, device=device, dtype=torch.long)

    @staticmethod
    def _sample_from_logits(
        logits: torch.Tensor,
        *,
        temperature: float,
        generator: torch.Generator | None,
    ) -> tuple[torch.LongTensor, torch.Tensor]:
        """Sample one token per position with optional temperature, returning tokens and their probabilities."""
        if temperature < 0:
            raise ValueError(f"`temperature` must be >= 0, got {temperature}.")

        vocab_size = logits.shape[-1]
        flat_logits = logits.reshape(-1, vocab_size)
        probs = torch.softmax(flat_logits.float(), dim=-1)

        if temperature == 0.0:
            token = flat_logits.argmax(dim=-1, keepdim=True)
        else:
            scaled_probs = torch.softmax(flat_logits.float() / temperature, dim=-1)
            token = torch.multinomial(scaled_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 _alpha(self, step_index: int) -> float:
        """Survival probability `alpha = 1 - t` of a clean token at the time grid point `step_index`."""
        return step_index / self.num_inference_steps

    @staticmethod
    def _to_loo_logits(logits: torch.Tensor, tokens: torch.LongTensor, alpha: float) -> torch.Tensor:
        """
        Convert plain-denoiser logits to the leave-one-out posterior for the uniform kernel.

        Subtracts `log(1 + K * alpha / (1 - alpha))` from the observed token's logit (eq. 13 of
        https://huggingface.co/papers/2605.22765); renormalization happens in the following softmax.
        """
        if alpha <= 0.0 or alpha >= 1.0:
            return logits
        delta = math.log1p(logits.shape[-1] * alpha / (1.0 - alpha))
        shifted = logits.clone()
        src = torch.full((*tokens.shape, 1), -delta, dtype=shifted.dtype, device=shifted.device)
        shifted.scatter_add_(-1, tokens.unsqueeze(-1), src)
        return shifted

    def step(
        self,
        model_output: torch.Tensor,
        timestep: int | torch.Tensor,
        sample: torch.LongTensor,
        *,
        temperature: float = 0.0,
        generator: torch.Generator | None = None,
        return_dict: bool = True,
    ) -> DiscreteDDIMSchedulerOutput | tuple[torch.LongTensor, torch.LongTensor, torch.Tensor]:
        """
        Sample the next block from the posterior `q(x_s | x_t, x0)` of the uniform corruption process.

        With `a = alpha_t / alpha_s` (survival probability from `s` to `t`) and `b = alpha_s`, the posterior mass of
        each route is

            clean: `b * (1 - a) / K + a * b * 1[x_t = x0]`, stay: `a * (1 - b) / K`, noise: `(1 - a) * (1 - b) / K`,

        so the last step (`b = 1`) deterministically commits the predicted clean tokens.

        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 denoising schedule, in `[0, num_inference_steps - 1]`.
            sample (`torch.LongTensor` of shape `(batch_size, block_length)`):
                Current block token IDs `x_t`.
            temperature (`float`):
                Sampling temperature applied to the logits when drawing `x0`.
            generator (`torch.Generator`, *optional*):
                RNG for sampling.
            return_dict (`bool`):
                Whether to return a [`DiscreteDDIMSchedulerOutput`] or a plain tuple.
        """
        if isinstance(timestep, torch.Tensor):
            step_index = int(timestep.item())
        else:
            step_index = int(timestep)

        sampled_tokens, sampled_probs = self._sample_from_logits(
            model_output, temperature=temperature, generator=generator
        )

        vocab_size = model_output.shape[-1]
        num_steps = self.num_inference_steps
        # `step_index` counts up from 0 to `num_inference_steps - 1`: alpha(t) = 1 - t increases towards the clean end,
        # with alpha_s = 1 on the final step so the predicted clean tokens are committed deterministically.
        alpha_t = step_index / num_steps
        alpha_s = (step_index + 1) / num_steps
        survival = alpha_t / alpha_s

        same = (sample == sampled_tokens).float()
        clean_mass = alpha_s * (1 - survival) / vocab_size + survival * alpha_s * same
        stay_mass = survival * (1 - alpha_s) / vocab_size * torch.ones_like(same)
        noise_mass = (1 - survival) * (1 - alpha_s) / vocab_size * torch.ones_like(same)

        route_probs = torch.stack([clean_mass, stay_mass, noise_mass], dim=-1)
        route_probs = route_probs / route_probs.sum(dim=-1, keepdim=True)
        routes = torch.multinomial(route_probs.view(-1, 3), num_samples=1, generator=generator).view_as(sample)

        random_tokens = torch.randint(
            low=0, high=vocab_size, size=sample.shape, device=sample.device, generator=generator
        )
        prev_sample = torch.where(routes == 0, sampled_tokens, sample)
        prev_sample = torch.where(routes == 2, random_tokens, prev_sample)

        if not return_dict:
            return prev_sample, sampled_tokens, sampled_probs, model_output
        return DiscreteDDIMSchedulerOutput(
            prev_sample=prev_sample,
            sampled_tokens=sampled_tokens,
            sampled_probs=sampled_probs,
            pred_logits=model_output,
        )

    def _select_positions(
        self, sample: torch.LongTensor, cond_log_probs: torch.Tensor, generator: torch.Generator | None
    ) -> torch.LongTensor:
        """Pick `corrector_k` positions per row to resample, least-confident first (Gumbel-top-k without replacement)."""
        selection = self.config.corrector_selection
        batch_size, seq_len = sample.shape
        k_eff = min(max(1, int(self.config.corrector_k)), seq_len)

        if selection == "random":
            scores = torch.rand(batch_size, seq_len, device=sample.device, generator=generator)
            return torch.topk(scores, k=k_eff, dim=-1).indices

        if selection == "lowest_maxprob":
            confidence = -cond_log_probs.max(dim=-1).values
        elif selection == "lowest_current_prob":
            confidence = -torch.gather(cond_log_probs, -1, sample.unsqueeze(-1)).squeeze(-1)
        elif selection == "lowest_log_margin":
            log_current = torch.gather(cond_log_probs, -1, sample.unsqueeze(-1)).squeeze(-1)
            alt = cond_log_probs.clone().scatter_(-1, sample.unsqueeze(-1), float("-inf"))
            confidence = -(log_current - alt.max(dim=-1).values)
        else:
            raise ValueError(f"Unknown `corrector_selection`: {selection!r}.")

        keys = confidence / float(self.config.corrector_selection_tau)
        u = torch.rand(keys.shape, device=keys.device, generator=generator).clamp_(1e-12, 1.0 - 1e-12)
        keys = keys + (-torch.log(-torch.log(u)))
        return torch.topk(keys, k=k_eff, dim=-1).indices

    def step_correct(
        self,
        model_output: torch.Tensor,
        timestep: int | torch.Tensor,
        sample: torch.LongTensor,
        *,
        generator: torch.Generator | None = None,
        return_dict: bool = True,
    ) -> DiscreteDDIMSchedulerOutput | tuple[torch.LongTensor, torch.LongTensor, torch.Tensor]:
        """
        Run one Gibbs corrector sweep at the post-predictor time `s`, following the leave-one-out predictor-corrector
        of https://huggingface.co/papers/2605.22765.

        The model logits (recomputed on the current `sample`) are converted to the LOO denoiser, the one-coordinate
        conditional `p_s(x^l | x^{-l}) = Cat(alpha_s * x0_loo + (1 - alpha_s) / K)` is formed, the least-confident
        `corrector_k` positions are selected, and those positions are resampled while the rest are held fixed. The
        sweep preserves `p_s`, so it refines the sample without changing its marginal and needs no extra training.

        Args:
            model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`):
                Raw logits from the model recomputed on the current (post-predictor) `sample`.
            timestep (`int` or `torch.Tensor`):
                The predictor step index just completed; the corrector runs at the following grid point `s`.
            sample (`torch.LongTensor` of shape `(batch_size, block_length)`):
                Current block token IDs to refine.
            generator (`torch.Generator`, *optional*):
                RNG for sampling.
            return_dict (`bool`):
                Whether to return a [`DiscreteDDIMSchedulerOutput`] or a plain tuple.
        """
        if isinstance(timestep, torch.Tensor):
            step_index = int(timestep.item())
        else:
            step_index = int(timestep)

        # The corrector acts at the cleaner time `s` reached by the predictor.
        alpha_s = self._alpha(step_index + 1)
        vocab_size = model_output.shape[-1]

        # Match the reference corrector, which forms the conditional in float64 (the LOO correction reaches ~log(K)).
        loo_logits = self._to_loo_logits(model_output.double(), sample, alpha_s)
        loo_log_probs = torch.log_softmax(loo_logits, dim=-1)
        log_uniform = math.log1p(-alpha_s) - math.log(vocab_size)
        cond_log_probs = torch.logaddexp(
            math.log(alpha_s) + loo_log_probs, torch.full_like(loo_log_probs, log_uniform)
        )

        positions = self._select_positions(sample, cond_log_probs, generator)
        rows = torch.arange(sample.shape[0], device=sample.device).unsqueeze(-1).expand_as(positions)
        chosen_probs = cond_log_probs[rows, positions].exp()
        resampled = torch.multinomial(
            chosen_probs.reshape(-1, vocab_size), num_samples=1, generator=generator
        ).view_as(positions)

        prev_sample = sample.clone()
        prev_sample[rows, positions] = resampled
        sampled_probs = torch.gather(chosen_probs, -1, resampled.unsqueeze(-1)).squeeze(-1)

        if not return_dict:
            return prev_sample, resampled, sampled_probs, model_output
        return DiscreteDDIMSchedulerOutput(
            prev_sample=prev_sample,
            sampled_tokens=resampled,
            sampled_probs=sampled_probs,
            pred_logits=model_output,
        )


__all__ = ["DiscreteDDIMScheduler", "DiscreteDDIMSchedulerOutput"]