File size: 7,184 Bytes
a2ffd07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Attention-pooling sequence probes over raw residual-stream hidden states.

Self-contained port of multilayer-sae's ``experiment/training/sequence_probe.py``
(no ``experiment.*`` dependencies) for use by ``train_probe_latent.py``.

Replaces the linear + max-pool probe (``probe/probing.py:LinearProbe`` trained by
``train_binary_probe``) with a per-layer probe that reads the *sequence* of
per-token hidden states directly, with no fixed pooling:

  * Operating on raw ``h_l`` (d_model) keeps the probe independent of any SAE
    dictionary.
  * A learned-query multi-head attention pool replaces max/mean pooling so no
    information is discarded by a hard reduction, and the probe co-adapts to
    *where* the concept lives in the token sequence (follows HyperSteer,
    arXiv:2506.03292).

``forward_logits`` gives one per-sequence logit per layer (for training on a
per-caption label); ``forward_token_logits`` gives a per-token logit per layer
(for detecting the concept live during generation) — both reuse the same weights.
"""
from __future__ import annotations

import torch
import torch.nn as nn


class _AttnPoolProbe(nn.Module):
    """Single-layer attention-pooling probe: (B, T, d_model) -> (B,) logit."""

    def __init__(
        self,
        d_model: int,
        d_probe: int = 256,
        n_heads: int = 4,
        n_ctx_blocks: int = 1,
        dropout: float = 0.0,
        spectral_norm: bool = False,
    ):
        super().__init__()
        self.in_proj = nn.Linear(d_model, d_probe)
        self.ctx_blocks = nn.ModuleList([
            nn.TransformerEncoderLayer(
                d_model=d_probe,
                nhead=n_heads,
                dim_feedforward=4 * d_probe,
                dropout=dropout,
                batch_first=True,
                norm_first=True,
            )
            for _ in range(n_ctx_blocks)
        ])
        # Learned query that attends over the token sequence (attention pooling).
        self.query = nn.Parameter(torch.zeros(1, 1, d_probe))
        nn.init.normal_(self.query, std=0.02)
        self.pool_attn = nn.MultiheadAttention(
            d_probe, n_heads, dropout=dropout, batch_first=True
        )
        self.norm = nn.LayerNorm(d_probe)
        head = nn.Linear(d_probe, 1)
        self.head = nn.utils.spectral_norm(head) if spectral_norm else head

    def forward(
        self, x: torch.Tensor, key_padding_mask: torch.Tensor | None
    ) -> torch.Tensor:
        # x: (B, T, d_model); key_padding_mask: (B, T) bool, True = ignore (PAD).
        x = self.in_proj(x)                                    # (B, T, d_probe)
        for blk in self.ctx_blocks:
            x = blk(x, src_key_padding_mask=key_padding_mask)
        B = x.shape[0]
        q = self.query.expand(B, -1, -1)                       # (B, 1, d_probe)
        pooled, _ = self.pool_attn(
            q, x, x, key_padding_mask=key_padding_mask, need_weights=False
        )                                                      # (B, 1, d_probe)
        pooled = self.norm(pooled.squeeze(1))                  # (B, d_probe)
        return self.head(pooled).squeeze(-1)                   # (B,)

    def token_logits(
        self, x: torch.Tensor, key_padding_mask: torch.Tensor | None
    ) -> torch.Tensor:
        """Per-token logits (no attention pooling): (B, T, d_model) -> (B, T).

        Reuses in_proj + context blocks + norm + head so a single probe module
        supports both the per-sequence (attention-pooled) and per-token readouts.
        Used to detect the concept live during generation.
        """
        x = self.in_proj(x)                                    # (B, T, d_probe)
        for blk in self.ctx_blocks:
            x = blk(x, src_key_padding_mask=key_padding_mask)
        return self.head(self.norm(x)).squeeze(-1)             # (B, T)


class SequenceLayerProbes(nn.Module):
    """One attention-pooling probe per monitored layer, run on raw hidden states."""

    def __init__(
        self,
        layer_indices: list[int],
        d_model: int,
        d_probe: int = 256,
        n_heads: int = 4,
        n_ctx_blocks: int = 1,
        dropout: float = 0.0,
        spectral_norm: bool = False,
    ):
        super().__init__()
        self.layer_indices = list(layer_indices)
        self.d_model = d_model
        self.probes = nn.ModuleList([
            _AttnPoolProbe(
                d_model, d_probe=d_probe, n_heads=n_heads,
                n_ctx_blocks=n_ctx_blocks, dropout=dropout,
                spectral_norm=spectral_norm,
            )
            for _ in layer_indices
        ])
        self._idx = {l: i for i, l in enumerate(self.layer_indices)}

    def forward_logits(
        self,
        feats_seq: dict[int, torch.Tensor],     # l -> (B, T, d_model)
        key_padding_mask: torch.Tensor,         # (B, T) bool, True = PAD/ignore
    ) -> list[torch.Tensor]:
        # Probe runs in fp32 for numerical stability regardless of model dtype.
        w_dtype = self.probes[0].in_proj.weight.dtype
        return [
            self.probes[self._idx[l]](
                feats_seq[l].to(w_dtype), key_padding_mask
            )
            for l in self.layer_indices
        ]

    def forward_token_logits(
        self,
        feats_seq: dict[int, torch.Tensor],     # l -> (B, T, d_model)
        key_padding_mask: torch.Tensor,         # (B, T) bool, True = PAD/ignore
    ) -> list[torch.Tensor]:
        """Per-token logits per layer: list of (B, T). For live-generation detection."""
        w_dtype = self.probes[0].in_proj.weight.dtype
        return [
            self.probes[self._idx[l]].token_logits(feats_seq[l].to(w_dtype), key_padding_mask)
            for l in self.layer_indices
        ]

    def forward(
        self, feats_seq: dict[int, torch.Tensor], key_padding_mask: torch.Tensor
    ) -> list[torch.Tensor]:
        return [torch.sigmoid(z) for z in self.forward_logits(feats_seq, key_padding_mask)]


def sequence_layer_probes_from_checkpoint(
    path: str,
    device: torch.device | str | None = None,
) -> SequenceLayerProbes:
    """Load a SequenceLayerProbes checkpoint saved by train_probe_latent.py.

    The checkpoint stores both the state_dict and the construction meta
    (layer_indices, d_model, probe_dim, ...) so the architecture is rebuilt
    exactly. Handles a ``module.`` DDP prefix.
    """
    # Checkpoint holds only tensors + a plain dict of ints/lists, so the safe
    # weights_only loader suffices (no arbitrary-object unpickling).
    ckpt = torch.load(path, map_location="cpu", weights_only=True)
    meta = ckpt["meta"]
    sd = ckpt["state_dict"]
    if any(k.startswith("module.") for k in sd):
        sd = {k.replace("module.", "", 1): v for k, v in sd.items()}
    probes = SequenceLayerProbes(
        meta["layer_indices"], meta["d_model"],
        d_probe=meta["d_probe"], n_heads=meta["n_heads"],
        n_ctx_blocks=meta["n_ctx_blocks"], spectral_norm=meta["spectral_norm"],
    )
    probes.load_state_dict(sd, strict=True)
    if device is not None:
        dev = device if isinstance(device, torch.device) else torch.device(device)
        probes = probes.to(dev)
    return probes