Hooks
ablate_pos_keep_self_attn_hook(module, args, kwargs, output, ablation_queries)
Hook function to ablate the tokens in the attention mask but keeping the self attn weigths. It will set to 0 the row of tokens to ablate except for the las position
Source code in easyroutine/interpretability/hooks.py
620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 |
|
attention_pattern_head(module, args, kwargs, output, token_indexes, layer, cache, head='all', act_on_input=False, attn_pattern_avg='none', attn_pattern_row_partition=None)
Hook function to extract the attention pattern of the heads. It will extract the attention pattern. As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
-
|
args
|
the input args of the hook function |
required |
-
|
kwargs
|
the input kwargs of the hook function |
required |
-
|
output
|
the output of the hook function |
required |
-
|
token_indexes(List[Tuple])
|
the indexes of the tokens to extract the attention pattern |
required |
-
|
layer (int
|
the layer of the model |
required |
-
|
cache (ActivationCache
|
the cache where to save the activations |
required |
-
|
head (Union[str, int]
|
the head of the model. If "all" is passed, it will extract all the heads of the layer |
required |
-
|
attn_pattern_avg (Literal["mean", "sum", "baseline_ratio", "none"]
|
the method to average the attention pattern |
required |
-
|
attn_pattern_row_partition (List[int]
|
the indexes of the tokens to partition the attention pattern |
required |
Avg strategies
If the attn_pattern_avg is not "none", the attention pattern is divided in blocks and the average value of each block is computed, using the method specified in attn_pattern_avg. The idea is to partition the attention pattern into groups of tokens, and then compute a single average value for each group. The pattern is divided into len(attn_pattern_row_partition) x len(token_indexes) blocks, where each block B_ij is defined to have the indeces of the rows in attn_pattern_row_partition[i] and the columns in token_indexes[j]. If attn_pattern_row_partition is None, then the rows are the same as token_indexes.
0| a_00 0 0 0 0 0 0 token_indexes = [(1,3), (4)] 1| a_10 a_11 0 0 0 0 0 attn_pattern_row_partition = [(0,1)] 2| a_20 a_21 a_22 0 0 0 0 3| a_30 a_31 a_32 a_33 0 0 0 4| a_40 a_41 a_42 a_43 a_44 0 0 5| a_50 a_51 a_52 a_53 a_54 a_55 0 6| a_60 a_61 a_62 a_63 a_64 a_65 a_66
0 1 2 3 4 5 6
- Block B_00:
- Rows: 0,1
- Columns: 1,2,3
- Block: [a_01, a_02, a_03, a_11, a_12, a_13]
- Block B_01:
- Rows: 0,1
- Columns: 4
- Block: [a_04, a_14]
If attn_pattern_avg is "mean", the average value for each block is computed as the mean of the block, so the output will be: batch n_row_blocks n_col_blocks So in this case, the output will be: batch 1 2 where the first value is the average of the first block and the second value is the average of the second block.
The method to compute a single value for each block is specified by the attn_pattern_avg parameter, and can be one of the following: - "mean": Compute the mean of the block. - "sum": Compute the sum of the block. - "baseline_ratio": Compute the ratio of the observed average attention to the expected average attention. The expected average attention is computed by assuming that attention is uniform across the block. So, for each row in attn_pattern_row_partition, we compute the fraction of allowed keys that belong to token_indexes. The expected average attention is the sum of these fractions divided by the number of rows. The final ratio is the observed average attention divided by the expected average attention.
Source code in easyroutine/interpretability/hooks.py
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 |
|
avg_attention_pattern_head(module, args, kwargs, output, token_indexes, layer, attn_pattern_current_avg, batch_idx, cache, avg=False, extract_avg_value=False, act_on_input=False)
Hook function to extract the average attention pattern of the heads. It will extract the attention pattern and then average it. As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
-
|
b
|
the input of the hook function. It's the output of the attention pattern of the heads |
required |
-
|
s
|
the state of the hook function. It's the state of the model |
required |
-
|
layer
|
the layer of the model |
required |
-
|
head
|
the head of the model |
required |
-
|
attn_pattern_current_avg
|
the current average attention pattern |
required |
Source code in easyroutine/interpretability/hooks.py
972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 |
|
avg_hook(module, args, kwargs, output, cache, cache_key, last_image_idx, end_image_idx)
It save the activations of the residual stream in the cache. It will save the activations in the cache (a global variable out the scope of the function)
Source code in easyroutine/interpretability/hooks.py
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 |
|
compute_statistics(tensor, dim=-1, keepdim=True, eps=1e-06)
Computes the mean, variance, and second moment of a given tensor along a specified dimension.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
tensor
|
Tensor
|
Input tensor. |
required |
dim
|
int
|
Dimension along which to compute statistics (default: -1). |
-1
|
keepdim
|
bool
|
Whether to keep the reduced dimension (default: True). |
True
|
eps
|
float
|
Small constant for numerical stability. |
1e-06
|
Returns:
Name | Type | Description |
---|---|---|
tuple |
(mean, variance, second_moment) |
Source code in easyroutine/interpretability/hooks.py
121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 |
|
create_dynamic_hook(pyvene_hook, **kwargs)
DEPRECATED: pyvene is not used anymore. This function is used to create a dynamic hook. It is a wrapper around the pyvene_hook function.
Source code in easyroutine/interpretability/hooks.py
76 77 78 79 80 81 82 83 84 85 86 |
|
embed_hook(module, args, kwargs, output, token_indexes, cache, cache_key)
Hook function to extract the embeddings of the specified tokens and save them in the cache. Args: module: The module being hooked. args: Positional arguments. kwargs: Keyword arguments. output: Output from the module. token_indexes: List of token indexes to extract. cache: The cache object to store results. cache_key: The key under which to store the embeddings.
Source code in easyroutine/interpretability/hooks.py
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 |
|
input_embedding_hook(module, args, kwargs, output, cache, cache_key, token_indexes, keep_gradient=False, avg=False)
Hook to capture the output of the embedding layer, enable gradients, and store it in the cache.
Source code in easyroutine/interpretability/hooks.py
1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 |
|
intervention_attn_mat_hook(module, args, kwargs, output, q_positions, k_positions, head, multiplication_value, patching_values=None, apply_softmax=False)
Hook function to ablate the tokens in the attention mask. It will set to 0 the value vector of the tokens to ablate
Source code in easyroutine/interpretability/hooks.py
489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 |
|
intervention_heads_hook(module, args, kwargs, output, ablation_queries)
Hook function to ablate the heads in the attention mask. It will set to 0 the output of the heads to ablate
Source code in easyroutine/interpretability/hooks.py
598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 |
|
intervention_query_key_value_hook(module, args, kwargs, output, token_indexes, head, head_dim, num_key_value_groups, num_attention_heads, patching_values=None)
Hook function to intervene on the query, key and value vectors. It first unpack the vectors from the output of the module and then apply the intervention and then repack the vectors.
Source code in easyroutine/interpretability/hooks.py
324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 |
|
intervention_resid_hook(module, args, kwargs, output, token_indexes, patching_values=None)
Hook function to ablate the tokens in the residual stream. It will set to 0 the value vector of the tokens to ablate
Source code in easyroutine/interpretability/hooks.py
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
|
layernom_hook(module, args, kwargs, output, token_indexes, cache, cache_key, avg=False)
Compute and save mean, variance, and second moment for the specified token indexes. If avg is True, computes statistics for each token group; otherwise, flattens indexes. Args: module: The module being hooked. args: Positional arguments. kwargs: Keyword arguments. output: Output from the module. token_indexes: List of token index groups. cache: The cache object to store results. cache_key: The key under which to store the statistics. avg: Whether to average over each token group separately.
Source code in easyroutine/interpretability/hooks.py
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 |
|
multiply_pattern(tensor, multiplication_value)
Set the attention values to zero
Source code in easyroutine/interpretability/hooks.py
480 481 482 483 484 485 |
|
process_args_kwargs_output(args, kwargs, output)
Extract the main tensor from output, args, or kwargs. Prioritizes output (first element if tuple), then first arg, then kwargs['hidden_states'] if present. Args: args: Positional arguments from the hook. kwargs: Keyword arguments from the hook. output: Output from the hooked function. Returns: The main tensor to be processed by the hook.
Source code in easyroutine/interpretability/hooks.py
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 |
|
projected_key_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)
Hook function to extract the key vectors of the heads and project them through the attention output matrix. This shows the contribution that keys in each position could have to the residual stream through the attention mechanism.
Like other hooks, it saves the activations in the cache.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
b
|
the input of the hook function (output of the key vectors) |
required | |
layer
|
the layer of the model |
required | |
head
|
Union[str, int]
|
the head of the model. If "all" is passed, it will extract all the heads of the layer |
'all'
|
expand_head
|
bool
|
bool to expand the head dimension when extracting the keys vectors |
True
|
Source code in easyroutine/interpretability/hooks.py
775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 |
|
projected_query_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)
Hook function to extract the query vectors of the heads and project them through the attention output matrix. This shows the contribution that queries in each position could have to the residual stream through the attention mechanism.
Like other hooks, it saves the activations in the cache.
Parameters:
Name | Type | Description | Default |
---|---|---|---|
b
|
the input of the hook function (output of the query vectors) |
required | |
layer
|
the layer of the model |
required | |
head
|
Union[str, int]
|
the head of the model. If "all" is passed, it will extract all the heads of the layer |
'all'
|
expand_head
|
bool
|
bool to expand the head dimension when extracting the query vectors |
True
|
Source code in easyroutine/interpretability/hooks.py
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 |
|
projected_value_vectors_head(module, args, kwargs, output, layer, cache, token_indexes, num_attention_heads, num_key_value_heads, hidden_size, d_head, out_proj_weight, out_proj_bias, head='all', act_on_input=False, expand_head=True, avg=False)
Hook function to extract the values vectors of the heads. It will extract the values vectors and then project them with the final W_O projection As the other hooks, it will save the activations in the cache (a global variable out the scope of the function)
Parameters:
Name | Type | Description | Default |
---|---|---|---|
b
|
the input of the hook function. It's the output of the values vectors of the heads |
required | |
s
|
the state of the hook function. It's the state of the model |
required | |
layer
|
the layer of the model |
required | |
head
|
Union[str, int]
|
the head of the model. If "all" is passed, it will extract all the heads of the layer |
'all'
|
expand_head
|
bool
|
bool to expand the head dimension when extracting the values vectors and the attention pattern. If true, in the cache we will have a key for each head, like "value_L0H0", "value_L0H1", ... while if False, we will have only one key for each layer, like "value_L0" and the dimension of the head will be taken into account in the tensor. |
True
|
Source code in easyroutine/interpretability/hooks.py
666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 |
|
query_key_value_hook(module, args, kwargs, output, cache, cache_key, token_indexes, layer, head_dim, num_key_value_groups, num_attention_heads, head='all', avg=False)
Same as save_resid_hook but for the query, key and value vectors, it just have a reshape to have the head dimension.
Source code in easyroutine/interpretability/hooks.py
254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 |
|
restore_same_args_kwargs_output(b, args, kwargs, output)
Restore the structure of output, args, and kwargs after modification. Args: b: The new tensor to insert. args: Original positional arguments. kwargs: Original keyword arguments. output: Original output from the hooked function. Returns: The updated output, or (args, kwargs) if output is None.
Source code in easyroutine/interpretability/hooks.py
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 |
|
save_resid_hook(module, args, kwargs, output, cache, cache_key, token_indexes, avg=False)
Save the activations of the residual stream for the specified token indexes in the cache. If avg is True, saves averaged activations for each token group. Args: module: The module being hooked. args: Positional arguments. kwargs: Keyword arguments. output: Output from the module. cache: The cache object to store results. cache_key: The key under which to store the activations. token_indexes: List of token index groups. avg: Whether to average over each token group separately.
Source code in easyroutine/interpretability/hooks.py
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 |
|