how-do-llms-use-their-depth/downstream_task_options_analysis.py at main · Entelligence-AI/how-do-llms-use-their-depth · GitHub

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
from transformers import AutoModelForCausalLM, AutoTokenizer
import json
import torch
import torch.nn.functional as F
from datasets import load_dataset
import random
import os
import util.nethook as nethook
from util.tok_dataset import TokenizedDataset, dict_to_
from util.get_trace_layers import get_trace_layers
from dsets.downstream_text import PromptCompletionDataset
from tqdm import tqdm, trange
import pandas as pd
from tuned_lens.nn.lenses import TunedLens, LogitLens
from collections import defaultdict

class AttrDict(dict):
    def __init__(self, *args, **kwargs):
        super(AttrDict, self).__init__(*args, **kwargs)
        self.__dict__ = self

random.seed(42)

if __name__ == '__main__':
    # SELECT THE MODEL HERE

    model_name = 'gpt2-xl'
    hparams_filename = 'hparams/gpt2-xl.json'


    model_filename = model_name.split('/')[-1]

    # load params file
    f = open(hparams_filename)
    hparams = AttrDict(json.load(f))

    model = AutoModelForCausalLM.from_pretrained(model_name).cuda()
    tokenizer = AutoTokenizer.from_pretrained(model_name)

    #get necessary layers
    trace_layers, ln_1_layers, ln_2_layers, lm_head, ln_f, n_layers = get_trace_layers(model, hparams)

    model_output_name = model_filename
    skip_tokens = 0
    if model_filename == 'gpt2-xl':
        tuned_lens = TunedLens.from_model_and_pretrained(model, lens_resource_id = model_filename).cuda()
        embedding_matrix = nethook.get_module(model, "transformer.wte")
        def get_input_embeddings(input_ids):
            with torch.no_grad():
                return model.transformer.wte(input_ids)
    elif model_filename == 'Llama-2-7b-hf':
        print(model, 'meta-llama/' + model_filename)
        tuned_lens = TunedLens.from_model_and_pretrained(model, lens_resource_id = 'meta-llama/' + model_filename).cuda()
        embedding_matrix = model.lm_head
        def get_input_embeddings(input_ids):
            with torch.no_grad():
                return model.model.embed_tokens(input_ids)
        model_output_name = 'Llama-2-7b'
        skip_tokens = 1
    elif model_filename == 'Meta-Llama-3-8B':
        tuned_lens = TunedLens.from_model_and_pretrained(model, lens_resource_id = 'meta-llama/' + model_filename).cuda()
        embedding_matrix = model.lm_head
        def get_input_embeddings(input_ids):
            with torch.no_grad():
                return model.model.embed_tokens(input_ids)
        model_output_name = 'Llama-3-8B'
    elif model_filename == 'pythia-6.9b':
        tuned_lens = TunedLens.from_model_and_pretrained(model, lens_resource_id = 'EleutherAI/pythia-6.9b-deduped').cuda()
        embedding_matrix = nethook.get_module(model, "embed_out")
        def get_input_embeddings(input_ids):
            with torch.no_grad():
                return model.gpt_neox.embed_in(input_ids)

    print(f'Running model {model_output_name}')
    for downstream_task in ['MMLU', 'MRPC', 'NLI', 'SST']:
        dataset = PromptCompletionDataset(downstream_task=downstream_task, max_examples = 3_000)
        space = True
        answer_options = dataset.answer_options(tokenizer = tokenizer, space = space)
        output_df = []
        output_df_tuned = []

        print(f'Running on {len(dataset)} examples for {downstream_task} tag')
        print(f'Answer options: {answer_options}')

        for i in trange(len(dataset)):
            item = dataset.__getitem__(i)
            prompt = item['prompt']
            answer = item['answer']

            if answer is not None:
                answer_token = tokenizer.encode(answer, return_tensors='pt').cuda()[0]
                answer_token_w_space = tokenizer.encode(' ' + answer, return_tensors='pt').cuda()[0]

                answer_token_length = len(answer_token)
                answer_token_w_space_length = len(answer_token_w_space)

                first_answer_token = answer_token[skip_tokens]
                first_answer_token_w_space = answer_token_w_space[skip_tokens]

            def run_model(prompt):
                input_ids = tokenizer.encode(prompt, return_tensors='pt',
                                        truncation=True, max_length=1024).cuda()
                with nethook.TraceDict(
                    module=model,
                    layers=trace_layers,
                    retain_input=True,
                    retain_output=True,
                ) as tr:
                    output_ids = model.generate(
                        input_ids,
                        max_new_tokens=1,
                        num_return_sequences=1,
                        pad_token_id=tokenizer.eos_token_id
                    )

                last_token = output_ids[0][-1]
                output_text = tokenizer.decode(last_token, skip_special_tokens=True)
                input_embeddings = get_input_embeddings(input_ids)
                first_token_embedding = input_embeddings[:, -1, :][0]
                return output_text, last_token, tr, first_token_embedding

            output_text, last_token, tr, _ = run_model(prompt) # n+1 completion
            if output_text.isspace():
                prompt += ' '
                output_text, last_token, tr, _ = run_model(prompt)
                if output_text.isspace():
                    continue

            # print(last_token, output_text)

            def parse_data(df, tuned = True):
                for layer_name in tr:
                    with torch.no_grad():
                        cur_in = tr[layer_name].output
                        cur_in = cur_in[0] if isinstance(cur_in, tuple) else cur_in
                        if layer_name not in ('transformer.ln_f', 'model.norm', 'gpt_neox.final_layer_norm'):
                            if tuned:
                                layer = int(layer_name.split('.')[-1])
                                h = cur_in[0][-1]
                                cur_out = tuned_lens(h, layer)
                            else:
                                final_ln_output = ln_f(cur_in)
                                h = final_ln_output[0][-1] # [4096]
                                cur_out = embedding_matrix.weight @ h  # [32000, 4096] x [4096]
                        else:
                            continue

                        probabilities = torch.softmax(cur_out, dim=0)
                        result = {
                            'layer': layer_name,
                            'prompt index': i,
                            'prompt': prompt,
                            'model output': output_text.strip(),
                            'model token prediction': int(last_token.item()),
                            'answer': answer,
                        }

                        for option, token in answer_options:
                            if token is not None:
                                option_prob = probabilities[token].item()
                                rank = torch.sum(probabilities > option_prob).item() + 1

                                result[option + ' rank'] = rank
                                result[option + ' prob'] = option_prob

                        df.append(result)

            parse_data(output_df_tuned, tuned = True)
            parse_data(output_df, tuned = False)


            successful_examples = len(output_df) // (len(trace_layers) - 1)
            if successful_examples % 100 == 0:
                print(f'Completed {successful_examples}examples')
            if successful_examples > 10_000:
                print(f'Completed {successful_examples} examples')
                break

        output_df = pd.DataFrame(output_df)
        output_df.to_csv(f'{model_output_name}_{downstream_task}_options_analysis.csv', index = False)

        output_df_tuned = pd.DataFrame(output_df_tuned)
        output_df_tuned.to_csv(f'{model_output_name}_{downstream_task}_options_analysis_tuned.csv', index = False)