training_method/sequence_probe.py

"""Attention-pooling sequence probes over raw model hidden states.

Replaces the SAE-feature + max-pool probe (``finetune_adv.LayerProbes``) with a
per-layer probe that reads the *sequence* of per-token hidden states directly,
with no fixed pooling. Motivation:

  * Operating on raw ``h_l`` (d_model) instead of SAE latents removes the
    dependence on whether the concept is well represented in the SAE dictionary
    (the relation-to-relation variance we observed).
  * A learned-query multi-head attention pool replaces ``masked_max_pool`` /
    ``masked_mean_pool`` so no information is discarded by a hard reduction, and
    the probe can co-adapt to *where* the concept lives in the sequence. This
    follows HyperSteer (arXiv:2506.03292), whose best variant reads the base-LM
    residual-stream sequence via attention rather than a pooled summary.

The probe stays fully differentiable w.r.t. the hidden states, so the
adversarial suppression gradient still flows from the probe logits into the
LoRA parameters during the model step.

API mirrors ``LayerProbes`` so the training loop can swap between them:
    probes.forward_logits(feats_seq, key_padding_mask) -> list[(B,)]
where ``feats_seq[l]`` is (B, T, d_model) and ``key_padding_mask`` is (B, T)
bool with True at PAD positions to ignore.
"""
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 by the general per-token hallucination probe.
        """
        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 the general probe."""
        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,
    layer_indices: list[int],
    d_model: int,
    device: torch.device | str | None = None,
    d_probe: int = 256,
    n_heads: int = 4,
    n_ctx_blocks: int = 1,
    spectral_norm: bool = False,
) -> SequenceLayerProbes:
    """Load a SequenceLayerProbes state_dict (handles a ``module.`` DDP prefix)."""
    sd = torch.load(path, map_location="cpu", weights_only=True)
    if any(k.startswith("module.") for k in sd):
        sd = {k.replace("module.", "", 1): v for k, v in sd.items()}
    use_sn = spectral_norm or any("weight_orig" in k for k in sd)
    probes = SequenceLayerProbes(
        layer_indices, d_model, d_probe=d_probe, n_heads=n_heads,
        n_ctx_blocks=n_ctx_blocks, spectral_norm=use_sn,
    )
    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