Functions — github.com/boyiwei/alignment-attribution-code

CodeHub MCPcopy Create free account

hub / github.com/boyiwei/alignment-attribution-code / functions

Functions99 in github.com/boyiwei/alignment-attribution-code

⨍ Functions 99 ◇ Types & classes 14

↓ 20 callersFunctionfind_layers

Recursively find the layers of a certain type in a module. Args: module (nn.Module): PyTorch module. layers (list): List of

lib/prune.py:17

↓ 14 callersFunctionreturn_given_alpha

(alpha, sort_res, W_metric, tmp_metric, sum_before)

lib/prune.py:140

↓ 13 callersFunctioneval_attack

Evaluate the attack performance of a given model on AdvBench. Args: model (object): The model object to be evaluated. tokeni

lib/eval.py:257

↓ 13 callersFunctionget_loaders

( name, nsamples=128, seed=0, seqlen=2048, tokenizer=None, disentangle=False )

lib/data.py:114

↓ 7 callersMethodadd_batch

(self, inp, out)

lib/ablate.py:29

↓ 5 callersFunctionadd_batch

lib/prune.py:332

↓ 5 callersFunctionmake_Act

(model, verbose=False)

lib/model_wrapper.py:91

↓ 5 callersFunctionprepare_calibration_input

(model, dataloader, device, nsamples)

lib/prune.py:93

↓ 4 callersFunctioneval_ppl

(args, model, tokenizer, device=torch.device("cuda:0"))

↓ 3 callersFunction_prune_core

lib/model_wrapper.py:321

↓ 3 callersFunctionapply_prompt_template

Apply a prompt template to a dataset of plain queries. Add system prompt, user prompt, <<SYS>> tags, [INST] tags, etc.. Args: pro

lib/prompt_utils.py:87

↓ 3 callersFunctioneval_zero_shot

( model_name, model, tokenizer, task_list=[ "boolq", "rte", "hellaswag

lib/eval.py:205

↓ 3 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

↓ 3 callersFunctionmake_Act

(model, verbose=False)

main_low_rank_diff.py:77

↓ 3 callersFunctionmake_Act

(model, verbose=False)

lib/model_wrapper_low.py:76

↓ 3 callersFunctionmake_low_rank

( args, model, tokenizer, device=torch.device("cuda:0"), prune_data="wikitext" )

lib/model_wrapper_low.py:128

↓ 3 callersFunctionrevert_Act_to_Linear

Reverts ActLinear modules back to their original nn.Linear layers.

lib/model_wrapper.py:115

↓ 2 callersFunctionadd_batch_extra

lib/prune.py:775

↓ 2 callersFunctionclear_act_buffer

main_low_rank_diff.py:123

↓ 2 callersFunctionclear_act_buffer

lib/model_wrapper.py:137

↓ 2 callersFunctionclear_act_buffer

lib/model_wrapper_low.py:122

↓ 2 callersMethodclear_act_buffer

main_low_rank_diff.py:29

↓ 2 callersMethodclear_act_buffer

lib/model_wrapper_low.py:28

↓ 2 callersMethodfasterprune

( self, args, sparsity, mask=None, prune_n=0, prune_m=0,

lib/ablate.py:86

↓ 2 callersMethodfree

lib/ablate.py:187

↓ 2 callersFunctionget_align

(nsamples, seed, seqlen, tokenizer, disentangle=False, mode="base")

↓ 2 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

rewind_ft_model.py:27

↓ 2 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

main_low_rank_diff.py:334

↓ 2 callersFunctionnot_matched

↓ 2 callersFunctionprune_wandg

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/model_wrapper.py:241

↓ 1 callersMethod__init__

(self, model, mask)

main_low_rank_diff.py:62

↓ 1 callersMethod__init__

(self, model, mask)

lib/model_wrapper.py:76

↓ 1 callersMethod__init__

(self, model, mask)

lib/model_wrapper_low.py:61

↓ 1 callersFunctioncheck_sparsity

lib/prune.py:41

↓ 1 callersMethodclear_act_buffer

lib/model_wrapper.py:31

↓ 1 callersFunctioneval_ppl_wikitext

(model, testenc, bs=1, device=None)

lib/eval.py:154

↓ 1 callersFunctionget_alpaca

(nsamples, seed, seqlen, tokenizer, disentangle=False, dataset="alpaca")

↓ 1 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

main_low_rank.py:40

↓ 1 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

lib/model_wrapper.py:534

↓ 1 callersFunctionget_llm

(model_name, cache_dir="llm_weights")

lib/model_wrapper_low.py:281

↓ 1 callersMethodget_mag_mask

(self, sparsity, prunen, prunem)

lib/ablate.py:67

↓ 1 callersFunctionget_mask

Save mask for the unstructured pruned model (for ft-attack evaluation). `neg_prune`: - if `args.neg_prune` is False (bottom pruning),

lib/prune.py:2041

↓ 1 callersFunctionget_prompt_template

(prompt_template_style="base")

lib/prompt_utils.py:7

↓ 1 callersMethodget_wanda_mask

(self, sparsity, prunen, prunem)

lib/ablate.py:48

↓ 1 callersFunctionget_wikitext2

(nsamples, seed, seqlen, tokenizer)

↓ 1 callersFunctionmain

rewind_ft_model.py:39

↓ 1 callersFunctionmain

main_low_rank.py:59

↓ 1 callersFunctionmain

↓ 1 callersFunctionmake_low_rank

prune_data_pos: retain most useful (total_rank - rank_pos) ranks prune_data_neg: remove most useful (total_rank - rank_neg) ranks final_

main_low_rank_diff.py:129

↓ 1 callersFunctionpattern_match

(patterns, source_list)

lib/eval.py:224

↓ 1 callersFunctionprune_ablate

(args, model, tokenizer, dev, prune_n=0, prune_m=0)

lib/prune.py:1910

↓ 1 callersFunctionprune_attention_head

Prune the attention_heads based on the probing results. Still not supporting reover from base. Only support Llama-2-7b-chat-hf Args: args

lib/prune.py:2068

↓ 1 callersFunctionprune_magnitude

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/prune.py:199

↓ 1 callersFunctionprune_random

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/prune.py:151

↓ 1 callersFunctionprune_sparsegpt

(args, model, tokenizer, dev, prune_n=0, prune_m=0)

lib/prune.py:1791

↓ 1 callersFunctionprune_wanda

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/prune.py:255

↓ 1 callersFunctionprune_wanda_decouple_activations

Compute wanda score based on the difference between the align activation and misalign activation (In an online way, do not need to load wanda sco

lib/prune.py:641

↓ 1 callersFunctionprune_wanda_v2

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/model_wrapper.py:143

↓ 1 callersFunctionprune_wandg_set_difference

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/prune.py:1612

↓ 1 callersFunctionrevert_Act_to_Linear

Reverts ActLinear modules back to their original nn.Linear layers.

main_low_rank_diff.py:101

↓ 1 callersFunctionrevert_Act_to_Linear

Reverts ActLinear modules back to their original nn.Linear layers.

lib/model_wrapper_low.py:100

Method__enter__

main_low_rank_diff.py:50

Method__enter__

main_low_rank_diff.py:66

Method__enter__

lib/model_wrapper.py:64

Method__enter__

lib/model_wrapper.py:80

Method__enter__

lib/model_wrapper_low.py:49

Method__enter__

lib/model_wrapper_low.py:65

(self, exc_type, exc_val, exc_tb)

main_low_rank_diff.py:55

(self, exc_type, exc_val, exc_tb)

main_low_rank_diff.py:71

(self, exc_type, exc_val, exc_tb)

lib/model_wrapper.py:69

(self, exc_type, exc_val, exc_tb)

lib/model_wrapper.py:85

(self, exc_type, exc_val, exc_tb)

lib/model_wrapper_low.py:54

(self, exc_type, exc_val, exc_tb)

lib/model_wrapper_low.py:70

(self, base: nn.Linear)

main_low_rank_diff.py:23

main_low_rank_diff.py:47

lib/ablate.py:14

lib/prune.py:111

(self, base: nn.Linear)

lib/model_wrapper.py:21

lib/model_wrapper.py:61

(self, input_ids)

lib/sparsegpt.py:15

(self, base: nn.Linear)

lib/model_wrapper_low.py:22

lib/model_wrapper_low.py:46

(self, layer, layer_id=0, layer_name="none")

lib/layerwrapper.py:11

Methodadd_batch

(self, inp, out)

lib/sparsegpt.py:28

Methodadd_batch

tar: batch_size * seq_len, inp corresponding to the position where tar == -100 will be ignored

lib/layerwrapper.py:25

Functioncheck_sparsity_layerwise

lib/prune.py:68

Functioneval_ppl_wikitext_train

(model, trainloader, bs=1, device=None)

lib/eval.py:100

Methodfasterprune

(self, sparsity, prune_n=0, prune_m=0, blocksize=128, percdamp=0.01)

lib/sparsegpt.py:43

main_low_rank_diff.py:32

(self, inp, **kwargs)

lib/prune.py:115

lib/model_wrapper.py:35

lib/model_wrapper_low.py:31

lib/sparsegpt.py:126

Functionget_sentence_embedding

(model, tokenizer, sentence)

Functionprune_wanda_decouple_activation_norms

Compute wanda score based on the difference between tow activation norms (In an online way, do not need to load wanda score from file) Compu

lib/prune.py:1125

Functionprune_wandg_v1

( args, model, tokenizer, model_base=None, device=torch.device("cuda:0"), prune_n=0,

lib/model_wrapper.py:191

Functionset_seed

lib/prune.py:333