File size: 9,017 Bytes
c7f1373
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""
Worker: run hijack intervention on layer 1 for a single checkpoint and plot heatmaps.
Usage: python hijack_layer1_worker.py <checkpoint.pt> --output-dir <dir>
"""
import argparse
import os
import sys
import types
import numpy as np
import torch
import torch.nn.functional as F

import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt

sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', 'grid-run'))
from model_analysis import GPT, GPTConfig

BIN_SIZE = 8
N_BINS = 256 // BIN_SIZE
INTENSITY = 10.0
LAYER = 1
N_TRIALS = 2000


def remap_state_dict(sd):
    new_sd = {}
    for key, val in sd.items():
        new_key = key
        for i in range(10):
            new_key = new_key.replace(f'transformer.h.{i}.attn.', f'transformer.h.{i}.c_attn.')
            new_key = new_key.replace(f'transformer.h.{i}.mlp.', f'transformer.h.{i}.c_fc.')
        new_sd[new_key] = val
    return new_sd


def load_model(ckpt_path, device):
    ckpt = torch.load(ckpt_path, map_location='cpu')
    mc = ckpt['model_config']
    vocab_size = mc['vocab_size'] - 1
    block_size = mc['block_size']
    with_layer_norm = mc.get('use_final_LN', True)

    config = GPTConfig(block_size=block_size, vocab_size=vocab_size,
                       with_layer_norm=with_layer_norm)
    model = GPT(config)

    sd = remap_state_dict(ckpt['model_state_dict'])
    grid_wpe_size = block_size * 4 + 1
    if 'transformer.wpe.weight' in sd and sd['transformer.wpe.weight'].shape[0] > grid_wpe_size:
        sd['transformer.wpe.weight'] = sd['transformer.wpe.weight'][:grid_wpe_size]
    keys_to_skip = [k for k in sd if k.endswith('.c_attn.bias') and 'c_attn.c_attn' not in k]
    for k in keys_to_skip:
        del sd[k]
    if 'lm_head.weight' in sd:
        del sd['lm_head.weight']

    model.load_state_dict(sd, strict=False)
    model.to(device).eval()
    return model, config


def get_batch(vocab_size, block_size, device='cpu'):
    x = torch.randperm(vocab_size)[:block_size]
    vals, _ = torch.sort(x)
    return torch.cat((x, torch.tensor([vocab_size]), vals), dim=0).unsqueeze(0).to(device)


def compute_hijack(model, config, device):
    bs = config.block_size
    vs = config.vocab_size
    attn_module = model.transformer.h[LAYER].c_attn
    records = []

    for trial in range(N_TRIALS):
        idx = get_batch(vs, bs, device)
        unsorted = idx[0, :bs]
        sorted_part = idx[0, bs + 1: 2 * bs + 1]

        with torch.no_grad():
            _, _ = model(idx)
        raw_attn = attn_module.raw_attn.clone()

        for p in range(bs - 1):
            location = bs + 1 + p
            current_num = sorted_part[p].item()
            correct_next = idx[0, location + 1].item()

            next_loc_in_unsorted = (unsorted == correct_next).nonzero(as_tuple=True)[0]
            if len(next_loc_in_unsorted) == 0:
                continue
            next_loc = next_loc_in_unsorted[0].item()
            main_attn_val = raw_attn[location, next_loc].item()

            candidates = [i for i in range(bs) if unsorted[i].item() != correct_next]
            if not candidates:
                continue

            boost_idx = candidates[torch.randint(len(candidates), (1,)).item()]
            boosted_number = unsorted[boost_idx].item()

            def make_new_forward(loc, bidx, mav):
                def new_forward(self_attn, x, layer_n=-1):
                    B, T, C = x.size()
                    qkv = self_attn.c_attn(x)
                    q, k, v = qkv.split(self_attn.n_embd, dim=2)
                    q = q.view(B, T, self_attn.n_heads, C // self_attn.n_heads).transpose(1, 2)
                    k = k.view(B, T, self_attn.n_heads, C // self_attn.n_heads).transpose(1, 2)
                    v = v.view(B, T, self_attn.n_heads, C // self_attn.n_heads).transpose(1, 2)
                    attn = q @ k.transpose(-1, -2) * 0.1 / (k.size(-1)) ** 0.5
                    attn[:, :, loc, bidx] = mav + INTENSITY
                    attn = attn.masked_fill(self_attn.bias[:, :, :T, :T] == 0, float('-inf'))
                    attn = F.softmax(attn, dim=-1)
                    y = attn @ v
                    y = y.transpose(1, 2).contiguous().view(B, T, C)
                    y = self_attn.c_proj(y)
                    return y
                return new_forward

            old_forward = attn_module.forward
            attn_module.forward = types.MethodType(
                make_new_forward(location, boost_idx, main_attn_val), attn_module)

            with torch.no_grad():
                logits, _ = model(idx)
            predicted = torch.argmax(logits, dim=-1)[0, location].item()

            attn_module.forward = old_forward
            records.append((current_num, boosted_number, predicted, correct_next))

    return np.array(records, dtype=np.int32) if records else np.empty((0, 4), dtype=np.int32)


def plot_heatmaps(data, plot_dir, tag):
    if len(data) == 0:
        print("No data to plot!")
        return

    current = data[:, 0]; boosted = data[:, 1]
    predicted = data[:, 2]; correct = data[:, 3]
    broken = (predicted != correct).astype(np.float64)
    hijacked = (predicted == boosted).astype(np.float64)
    cur_bin = np.clip(current // BIN_SIZE, 0, N_BINS - 1)
    bst_bin = np.clip(boosted // BIN_SIZE, 0, N_BINS - 1)

    break_map = np.full((N_BINS, N_BINS), np.nan)
    hijack_map = np.full((N_BINS, N_BINS), np.nan)
    count_map = np.zeros((N_BINS, N_BINS), dtype=int)
    for cb in range(N_BINS):
        for bb in range(N_BINS):
            mask = (cur_bin == cb) & (bst_bin == bb)
            n = mask.sum()
            count_map[cb, bb] = n
            if n >= 5:
                break_map[cb, bb] = broken[mask].mean()
                hijack_map[cb, bb] = hijacked[mask].mean()

    tick_labels = [f'{i * BIN_SIZE}' for i in range(0, N_BINS, 4)]
    tick_positions = list(range(0, N_BINS, 4))

    for arr, cmap, label, fname in [
        (break_map, 'YlOrRd', 'Breaking Rate',
         f'hijack_breaking_rate_heatmap_layer{LAYER}.png'),
        (hijack_map, 'YlOrRd', 'Hijack Rate',
         f'hijack_hijack_rate_heatmap_layer{LAYER}.png'),
    ]:
        fig, ax = plt.subplots(figsize=(10, 8.5))
        im = ax.imshow(arr, aspect='auto', cmap=cmap, vmin=0, vmax=1,
                       interpolation='nearest', origin='lower')
        ax.set_xlabel('Intervened-toward Number (binned)', fontsize=12)
        ax.set_ylabel('Current Number (binned)', fontsize=12)
        title_map = {'Breaking Rate': 'Breaking Rate: P(pred \u2260 correct)',
                     'Hijack Rate': 'Hijack Rate: P(pred == intervened target)'}
        ax.set_title(f'{title_map[label]}\n{tag}  layer={LAYER}  intensity={INTENSITY}',
                     fontsize=12, fontweight='bold')
        ax.set_xticks(tick_positions); ax.set_xticklabels(tick_labels, fontsize=8)
        ax.set_yticks(tick_positions); ax.set_yticklabels(tick_labels, fontsize=8)
        plt.colorbar(im, ax=ax, label=label, shrink=0.85)
        fig.tight_layout()
        fig.savefig(os.path.join(plot_dir, fname), dpi=200, bbox_inches='tight')
        plt.close()
        print(f"Saved: {fname}")

    fig, ax = plt.subplots(figsize=(10, 8.5))
    im = ax.imshow(count_map, aspect='auto', cmap='viridis',
                   interpolation='nearest', origin='lower')
    ax.set_xlabel('Intervened-toward Number (binned)', fontsize=12)
    ax.set_ylabel('Current Number (binned)', fontsize=12)
    ax.set_title(f'Sample Count per (current, target) bin\n{tag}  layer={LAYER}  intensity={INTENSITY}',
                 fontsize=11, fontweight='bold')
    ax.set_xticks(tick_positions); ax.set_xticklabels(tick_labels, fontsize=8)
    ax.set_yticks(tick_positions); ax.set_yticklabels(tick_labels, fontsize=8)
    plt.colorbar(im, ax=ax, label='Count', shrink=0.85)
    fig.tight_layout()
    fname = f'hijack_sample_count_heatmap_layer{LAYER}.png'
    fig.savefig(os.path.join(plot_dir, fname), dpi=200, bbox_inches='tight')
    plt.close()
    print(f"Saved: {fname}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('checkpoint', type=str)
    parser.add_argument('--output-dir', type=str, required=True)
    args = parser.parse_args()

    device = 'cuda'
    os.makedirs(args.output_dir, exist_ok=True)

    print(f"Loading {os.path.basename(args.checkpoint)} ...", flush=True)
    model, config = load_model(args.checkpoint, device)

    print(f"Running hijack layer {LAYER} ({N_TRIALS} trials) ...", flush=True)
    data = compute_hijack(model, config, device)
    print(f"Collected {len(data)} records", flush=True)

    bn = os.path.basename(args.checkpoint).replace('.pt', '')
    parts = bn.split('__')
    ckpt_type = parts[1] if len(parts) > 1 else 'final'
    itr = int(ckpt_type.replace('ckpt', '')) if ckpt_type.startswith('ckpt') else 1000000
    tag = f"V=256  B=16  lr=0.03  iters={itr}  dseed=1337  iseed=1337"

    plot_heatmaps(data, args.output_dir, tag)


if __name__ == '__main__':
    main()