File size: 9,395 Bytes
922bb4b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Noisy state builder for block-AR SAD training.

Noising levels per token (uniformly sampled):
  0        = clean  (keep leaf embedding)
  1..L-1   = ancestor level l (sample from LUT, use learnable ancestor embedding)
  L        = mask token

AncestorTable provides:
  - Fixed LUT (indices + probs): which ancestor cluster each token maps to
  - Learnable ancestor embeddings: the actual embedding used as noisy input
"""

from typing import List, Optional, Tuple

import torch
import torch.nn as nn

from src.diffusion.ancestor_table import AncestorTable


class NoisyStateBuilder(nn.Module):
    """
    Builds noisy embeddings for vectorized block-AR training.

    Args:
        vocab_size:    V
        mask_token_id: id of [MASK] in the leaf vocabulary
        temp:          kept for API compat, unused (temp is baked into LUT probs)
        top_k_per_level: kept for API compat, unused
    """

    def __init__(
        self,
        vocab_size: int,
        mask_token_id: int,
        temp: float = 1.0,
        top_k_per_level: Optional[List[int]] = None,
        use_soft_expected: bool = True,
    ):
        super().__init__()
        self.vocab_size = vocab_size
        self.mask_token_id = mask_token_id
        # temp / top_k_per_level kept for API compat but not used

    def sample_levels_uniform(
        self, B: int, S: int, num_total_states: int, device: torch.device
    ) -> torch.Tensor:
        """
        Sample per-token levels uniformly from {0, 1, ..., num_total_states-1}.

        num_total_states = num_ancestor_levels + 1 + 1
                        = (mask_level)  where mask_level = num_ancestor_levels + 1

        Returns:
            levels: [B, S] int64
        """
        return torch.randint(0, num_total_states, (B, S), device=device)

    @staticmethod
    def sample_t(
        B: int,
        device: torch.device,
        eps: float = 1e-3,
        low_discrepancy: bool = False,
        rank: int = 0,
        world_size: int = 1,
    ) -> torch.Tensor:
        """
        Sample per-sequence noise level t ~ U[eps, 1-eps]  (shape [B]).

        Change 3 (MDLM): low-discrepancy stratified sampling.  A single shared
        phase u is drawn per step; the batch covers the grid {u, u+1/B, ...}.
        In DDP, the phase is offset by rank/(world_size*B) so that all ranks
        together cover finer strata (analogous to Kingma's VDM).

        Change 2 (BD3-LMs / Soft-Masked Diffusion): clamp to [eps, 1-eps] to
        avoid high-variance ELBO gradients at extreme mask rates (t→0 or t→1).
        """
        if low_discrepancy:
            # Fresh shared phase every step — must not be cached across steps.
            u = torch.rand((), device=device)
            if world_size > 1:
                # Offset phase per rank so disjoint strata across the global batch.
                u = (u + rank / (world_size * B)) % 1.0
            t = torch.arange(B, device=device, dtype=torch.float32) / B
            t = (t + u) % 1.0
        else:
            t = torch.rand(B, device=device)
        # Clamp to [eps, 1-eps]: avoid high-variance ELBO at extreme mask rates.
        return (1 - 2 * eps) * t + eps

    def sample_levels_hdlm(
        self,
        t: torch.Tensor,
        S: int,
        num_ancestor_levels: int,
        gamma: float = 1.0,
    ) -> torch.Tensor:
        """
        HDLM 3-state (clean / ancestor / mask) hierarchical schedule,
        γ=1 form from HierarchicalDiffusion.get_alpha_betapi:

            α_t = 1 - t                       # P(clean, keep original)
            c_t = -(1-t) * log(1-t)           # P(ancestor state)
            m_t = t + (1-t) * log(1-t)        # P(mask)

        The three probabilities sum to 1 (with tiny numerical renormalization).

        Each sequence has its own t (shape [B]); each token within the sequence
        is drawn i.i.d. from the 3-way categorical. If the token lands on the
        ancestor state and num_ancestor_levels > 1, c_t is split evenly among
        levels 1..L.

        Args:
            t:                   [B] in (0, 1), from sample_t()
            S:                   sequence length
            num_ancestor_levels: L = ancestor_table.num_levels (typically 1)
            gamma:               only γ=1 supported

        Returns:
            levels: [B, S] int64, values in {0, 1..L, L+1=mask_level}
        """
        if gamma != 1.0:
            raise NotImplementedError("sample_levels_hdlm only supports γ=1")
        assert num_ancestor_levels >= 1

        B = t.shape[0]
        device = t.device
        eps = 1e-8

        one_m_t = (1.0 - t).clamp(min=eps)                # [B]
        log_1m_t = one_m_t.log()                          # [B] <= 0
        alpha_t = one_m_t                                 # P(clean)
        c_t = -one_m_t * log_1m_t                         # P(ancestor-any)  >= 0
        m_t = (t + one_m_t * log_1m_t).clamp(min=0.0)     # P(mask)         >= 0

        # Renormalize for safety (should already sum to ~1)
        total = alpha_t + c_t + m_t
        alpha_t = alpha_t / total
        c_t = c_t / total
        m_t = m_t / total

        # Split c_t evenly among ancestor levels 1..L
        c_per_level = c_t / num_ancestor_levels           # [B]

        # Build per-sample probability over categories [clean, anc_1..anc_L, mask]
        probs = torch.stack(
            [alpha_t] + [c_per_level] * num_ancestor_levels + [m_t],
            dim=-1,
        )                                                 # [B, 2+L]

        # Per-token multinomial. Category index i maps 1-to-1 to level value:
        #   0      -> clean (level 0)
        #   1..L   -> ancestor levels 1..L
        #   L+1    -> mask_level (= num_ancestor_levels + 1 in NoisyStateBuilder)
        probs_flat = (
            probs.unsqueeze(1).expand(B, S, -1).reshape(B * S, -1)
        )
        sampled = torch.multinomial(probs_flat, num_samples=1).squeeze(-1)
        return sampled.view(B, S).long()

    def build_noisy_embeddings(
        self,
        input_ids: torch.Tensor,
        levels: torch.Tensor,
        ancestor_table: AncestorTable,
        leaf_embeddings: torch.Tensor,
        mask_embedding: torch.Tensor,
        # kept for API compat (ignored):
        hierarchy=None,
    ) -> Tuple[torch.Tensor, torch.Tensor, List[Optional[torch.Tensor]], torch.Tensor]:
        """
        Build noisy embeddings from LUT-based ancestor table.

        For each position:
          level 0:    leaf_embeddings[token_id]   (clean)
          level 1..L: learnable ancestor embedding sampled via LUT
          level L+1:  mask_embedding              (fully masked)

        Args:
            input_ids:        [B, S] int64
            levels:           [B, S] int64  (0 .. num_ancestor_levels+1)
            ancestor_table:   AncestorTable
            leaf_embeddings:  [V, d]
            mask_embedding:   [d]
            hierarchy:        ignored (kept for API compat)

        Returns:
            noisy_embs:             [B, S, d]
            ancestor_log_probs:     [B, S]  log-prob of chosen ancestor (0 at clean/mask)
            ancestor_probs_per_lvl: list of None (not needed downstream)
            corrupt_mask:           [B, S] bool – True at positions level >= 1
        """
        B, S = input_ids.shape
        dtype = leaf_embeddings.dtype
        device = input_ids.device

        # mask level = num_ancestor_levels + 1
        mask_level = ancestor_table.num_levels + 1

        # Start with clean embeddings everywhere
        noisy_embs = leaf_embeddings[input_ids].clone()   # [B, S, d]

        ancestor_log_probs = torch.zeros(B, S, device=device, dtype=dtype)
        corrupt_mask = torch.zeros(B, S, dtype=torch.bool, device=device)
        ancestor_probs_per_lvl: List[Optional[torch.Tensor]] = []

        # Apply mask
        mask_pos = (levels == mask_level)
        if mask_pos.any():
            noisy_embs[mask_pos] = mask_embedding.to(dtype)
            corrupt_mask[mask_pos] = True

        for l in range(1, mask_level):  # ancestor levels 1..mask_level-1
            pos_l = (levels == l)
            if not pos_l.any():
                ancestor_probs_per_lvl.append(None)
                continue

            flat_ids = input_ids[pos_l]                      # [N]
            N = flat_ids.shape[0]

            # Sample ancestor index via LUT multinomial
            lut_idx = ancestor_table.lut_indices(l)[flat_ids]    # [N, top_k]
            lut_prob = ancestor_table.lut_probs(l)[flat_ids]     # [N, top_k]

            sampled_local = torch.multinomial(lut_prob, num_samples=1).squeeze(1)  # [N]
            sampled_global = lut_idx[
                torch.arange(N, device=device), sampled_local
            ]  # [N]

            # Learnable ancestor embedding
            anc_emb = ancestor_table.ancestor_embeddings(l)  # [K, d]
            noisy_embs[pos_l] = anc_emb[sampled_global].to(dtype)

            # Log-prob of chosen ancestor (from LUT probs)
            chosen_lp = lut_prob[
                torch.arange(N, device=device), sampled_local
            ].clamp(min=1e-8).log()
            ancestor_log_probs[pos_l] = chosen_lp.to(dtype)

            corrupt_mask[pos_l] = True
            ancestor_probs_per_lvl.append(None)  # not needed downstream

        return noisy_embs, ancestor_log_probs, ancestor_probs_per_lvl, corrupt_mask