File size: 8,591 Bytes
2eb3475
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Forward-pass cache shared by all cross-interaction features.

A single forward pass populates ``ForwardCache``; subsequent features read
from it without re-running the model. Captures:

- hidden_states[l]   for l in 0..L      (L+1 tensors, pre-LN residual)
- ln_inputs[l]       for l in 0..L-1    (input to attention LayerNorm at layer l)
- attn_weights[l]    for l in 0..L-1    (B, H, T, T)
- attn_outputs[l]    for l in 0..L-1    (B, T, H*Dh) raw before W_O
- mlp_pre_act[l]     for l in 0..L-1    (B, T, 4*D)  W_up @ LN(h)
- mlp_post_act[l]    for l in 0..L-1    (B, T, 4*D)  activation
- mlp_out[l]         for l in 0..L-1    (B, T, D)    W_down @ post_act
- head_writes[l]     for l in 0..L-1    (B, H, T, D) per-head contribution after W_O
- logits             (B, C)

Compatible with HuggingFace RoBERTa / ELECTRA encoder layouts:
  model.encoder.layer[l].attention.self
  model.encoder.layer[l].attention.output.dense          (W_O)
  model.encoder.layer[l].attention.output.LayerNorm      (post-attn LN)
  model.encoder.layer[l].intermediate.dense              (W_up)
  model.encoder.layer[l].intermediate.intermediate_act_fn
  model.encoder.layer[l].output.dense                    (W_down)
  model.encoder.layer[l].output.LayerNorm                (post-MLP LN)
"""
from __future__ import annotations

from dataclasses import dataclass, field
from typing import List, Optional

import torch
import torch.nn as nn


@dataclass
class ForwardCache:
    """Container for activations captured during a single forward pass."""
    hidden_states:   List[torch.Tensor] = field(default_factory=list)
    ln_inputs:       List[torch.Tensor] = field(default_factory=list)
    attn_weights:    List[torch.Tensor] = field(default_factory=list)
    attn_outputs:    List[torch.Tensor] = field(default_factory=list)
    mlp_pre_act:     List[torch.Tensor] = field(default_factory=list)
    mlp_post_act:    List[torch.Tensor] = field(default_factory=list)
    mlp_out:         List[torch.Tensor] = field(default_factory=list)
    head_writes:     List[torch.Tensor] = field(default_factory=list)
    logits:          Optional[torch.Tensor] = None
    # Per-layer LN params (gain/bias) — useful for LN-decomposition (LGAP)
    ln_gain:         List[torch.Tensor] = field(default_factory=list)
    ln_bias:         List[torch.Tensor] = field(default_factory=list)


def _find_encoder_layers(model: nn.Module) -> List[nn.Module]:
    """Locate the transformer encoder layer list for RoBERTa / ELECTRA."""
    if hasattr(model, "roberta"):
        return model.roberta.encoder.layer
    if hasattr(model, "electra"):
        return model.electra.encoder.layer
    if hasattr(model, "encoder") and hasattr(model.encoder, "layer"):
        return model.encoder.layer
    raise ValueError(f"Could not locate encoder layers in {type(model).__name__}")


def _attn_layout(layer: nn.Module):
    """Return (attention_self, attention_output_dense, attention_output_LN)."""
    return (layer.attention.self,
            layer.attention.output.dense,
            layer.attention.output.LayerNorm)


def _mlp_layout(layer: nn.Module):
    """Return (intermediate.dense, intermediate.intermediate_act_fn,
              output.dense, output.LayerNorm)."""
    return (layer.intermediate.dense,
            layer.intermediate.intermediate_act_fn,
            layer.output.dense,
            layer.output.LayerNorm)


@torch.no_grad()
def run_with_hooks(model: nn.Module,
                    input_ids: torch.Tensor,
                    attention_mask: torch.Tensor) -> ForwardCache:
    """Run a single forward pass, capturing activations into ForwardCache.

    The model is put in eval mode; gradients are disabled.
    """
    cache = ForwardCache()
    layers = _find_encoder_layers(model)
    L = len(layers)
    H = layers[0].attention.self.num_attention_heads
    Dh = layers[0].attention.self.attention_head_size

    handles = []

    # ---- Per-layer hooks --------------------------------------------------
    def make_attn_inp_hook(idx):
        def hook(module, inputs, output):
            # input[0] to attention.self is the post-LN normalised hidden state
            x = inputs[0].detach()
            while len(cache.ln_inputs) <= idx: cache.ln_inputs.append(None)
            cache.ln_inputs[idx] = x
        return hook

    def make_attn_self_hook(idx):
        def hook(module, inputs, output):
            # output: (context_layer, [attention_probs]) or context_layer
            if isinstance(output, tuple):
                ctx = output[0]
                # attention_probs are returned when output_attentions=True
                if len(output) > 1 and isinstance(output[1], torch.Tensor):
                    while len(cache.attn_weights) <= idx: cache.attn_weights.append(None)
                    cache.attn_weights[idx] = output[1].detach()
            else:
                ctx = output
            # ctx shape: (B, T, H*Dh)
            while len(cache.attn_outputs) <= idx: cache.attn_outputs.append(None)
            cache.attn_outputs[idx] = ctx.detach()
        return hook

    def make_int_hook(idx):
        def hook(module, inputs, output):
            # intermediate.dense output = W_up @ LN(h); pre activation
            while len(cache.mlp_pre_act) <= idx: cache.mlp_pre_act.append(None)
            cache.mlp_pre_act[idx] = output.detach()
        return hook

    def make_act_hook(idx):
        def hook(module, inputs, output):
            while len(cache.mlp_post_act) <= idx: cache.mlp_post_act.append(None)
            cache.mlp_post_act[idx] = output.detach()
        return hook

    def make_outdense_hook(idx):
        def hook(module, inputs, output):
            # output.dense projects post_act -> D
            while len(cache.mlp_out) <= idx: cache.mlp_out.append(None)
            cache.mlp_out[idx] = output.detach()
        return hook

    def make_post_attn_ln_hook(idx):
        def hook(module, inputs, output):
            # Capture the LayerNorm gain/bias and the input
            gain = module.weight.detach()
            bias = module.bias.detach()
            while len(cache.ln_gain) <= idx: cache.ln_gain.append(None)
            while len(cache.ln_bias) <= idx: cache.ln_bias.append(None)
            cache.ln_gain[idx] = gain
            cache.ln_bias[idx] = bias
        return hook

    for l, layer in enumerate(layers):
        attn_self, attn_dense, attn_LN = _attn_layout(layer)
        int_dense, act_fn, out_dense, out_LN = _mlp_layout(layer)
        handles.append(attn_self.register_forward_hook(make_attn_inp_hook(l)))
        handles.append(attn_self.register_forward_hook(make_attn_self_hook(l)))
        handles.append(int_dense.register_forward_hook(make_int_hook(l)))
        handles.append(act_fn.register_forward_hook(make_act_hook(l)))
        handles.append(out_dense.register_forward_hook(make_outdense_hook(l)))
        handles.append(attn_LN.register_forward_hook(make_post_attn_ln_hook(l)))

    try:
        model.eval()
        out = model(input_ids=input_ids,
                    attention_mask=attention_mask,
                    output_hidden_states=True,
                    output_attentions=True,
                    return_dict=True)
        cache.hidden_states = [h.detach() for h in out.hidden_states]
        # Some HF heads return attentions even when not hooked
        if hasattr(out, "attentions") and out.attentions is not None and not cache.attn_weights:
            cache.attn_weights = [a.detach() for a in out.attentions]
        cache.logits = out.logits.detach()
    finally:
        for h in handles:
            h.remove()

    # ---- Derived: head_writes -------------------------------------------
    # head_writes[l][b, h, t, :] = W_O_slice_h @ attn_outputs[l][b, t, h*Dh:(h+1)*Dh]
    for l, layer in enumerate(layers):
        attn_out = cache.attn_outputs[l]                       # (B, T, H*Dh)
        _, attn_dense, _ = _attn_layout(layer)
        W_O = attn_dense.weight.detach()                       # (D, H*Dh)
        b_O = attn_dense.bias.detach() if attn_dense.bias is not None else None
        B, T, _ = attn_out.shape
        D = W_O.shape[0]
        attn_h = attn_out.view(B, T, H, Dh)                    # (B, T, H, Dh)
        W_h = W_O.view(D, H, Dh).permute(1, 0, 2)              # (H, D, Dh)
        # head_writes[b, h, t, :] = sum_d (attn_h[b, t, h, d] * W_h[h, :, d])
        # = einsum("bthd, hpd -> bhtp", attn_h, W_h)
        hw = torch.einsum("bthd, hpd -> bhtp", attn_h, W_h)    # (B, H, T, D)
        cache.head_writes.append(hw)

    return cache