-
Notifications
You must be signed in to change notification settings - Fork 1
/
eval_activations_open_source.py
154 lines (130 loc) · 6.72 KB
/
eval_activations_open_source.py
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
import os
import time
import json
import torch
import random
import argparse
import numpy as np
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
def measure_activations(model, tokenizer, prompts, states_path, keys_path, values_path, token_length=50, device=torch.device("cuda")):
num_layers = model.config.num_hidden_layers
num_heads = model.config.num_attention_heads
hidden_states_all_sample = []
values_all_sample = []
keys_all_sample = []
for prompt in tqdm(prompts):
inputs = tokenizer(prompt, return_tensors="pt").to(device)
for key in inputs.keys():
assert inputs[key].shape[1] >= token_length
inputs[key] = inputs[key][:, :token_length]
outputs = model.generate(
**inputs,
output_attentions=True,
output_hidden_states=True,
use_cache=True,
return_dict_in_generate=True,
max_new_tokens=1
)
attentions = outputs['attentions']
hidden_states = outputs["hidden_states"]
past_key_values = outputs["past_key_values"]
assert len(attentions) == 1
assert len(past_key_values) == num_layers and len(past_key_values[0]) == 2
values_all_layer = []
keys_all_layer = []
for l in range(num_layers):
# keys
keys_all_layer.append(past_key_values[l][0])
# values
values_all_layer.append(past_key_values[l][1])
# keys
keys_all_layer = torch.cat(keys_all_layer, dim=0)
keys_all_sample.append(keys_all_layer.unsqueeze(dim=0))
# values
values_all_layer = torch.cat(values_all_layer, dim=0)
values_all_sample.append(values_all_layer.unsqueeze(dim=0))
assert len(hidden_states) == 1
hidden_states_all_layer = []
for l in range(num_layers+1):
hidden_states_layer = hidden_states[0][l] # (num_samples, num_tokens, hidden_dim)
hidden_states_all_layer.append(hidden_states_layer)
hidden_states_all_layer = torch.cat(hidden_states_all_layer, dim=0)
hidden_states_all_sample.append(hidden_states_all_layer.unsqueeze(dim=0))
# attention_scores_all_sample = torch.cat(attention_scores_all_sample, dim=0) # (num_samples, num_layers, num_heads, num_tokens)
hidden_states_all_sample = torch.cat(hidden_states_all_sample, dim=0) # (num_samples, num_layers, num_tokens, hidden_dim)
keys_all_sample = torch.cat(keys_all_sample, dim=0)
values_all_sample = torch.cat(values_all_sample, dim=0)
# np.save(score_path, attention_scores_all_sample.cpu().numpy())
np.save(states_path, hidden_states_all_sample.cpu().numpy())
np.save(values_path, values_all_sample.cpu().numpy())
np.save(keys_path, keys_all_sample.cpu().numpy())
def compute_norm(states_path, device):
hidden_states = np.load(states_path)
num_samples, num_layers, num_tokens, dim = hidden_states.shape
hidden_states = torch.from_numpy(hidden_states).to(device)
split_size = 5
hidden_states_split = torch.split(hidden_states, split_size)
all_norms = []
for hidden_states in hidden_states_split:
norm = hidden_states.norm(p=2, dim=-1)
all_norms.append(norm)
all_norms = torch.cat(all_norms, dim=0)
return all_norms.mean(dim=0) # (num_samples, num_layers, num_tokens) -> (num_layers, num_tokens)
def compute_kv_norm(states_path, device):
hidden_states = np.load(states_path)
num_samples, num_layers, num_heads, num_tokens, dim = hidden_states.shape
hidden_states = torch.from_numpy(hidden_states).to(device)
split_size = 5
hidden_states_split = torch.split(hidden_states, split_size)
all_norms = []
for hidden_states in hidden_states_split:
norm = hidden_states.norm(p=2, dim=-1)
all_norms.append(norm)
all_norms = torch.cat(all_norms, dim=0)
return all_norms.mean(dim=(0, 2)) # (num_samples, num_layers, num_heads, num_tokens) -> (num_layers, num_tokens)
def measure_open_sourced_lms():
# load model family
device = torch.device("cuda")
os.makedirs("results", exist_ok=True)
########################################
gpt_family = ["openai-community/gpt2", "openai-community/gpt2-medium", "openai-community/gpt2-large", "openai-community/gpt2-xl"]
llama2_family = ["meta-llama/Llama-2-7b-hf", "meta-llama/Llama-2-13b-hf", "meta-llama/Llama-2-7b-chat-hf", "meta-llama/Llama-2-13b-chat-hf"]
llama3_family = ["meta-llama/Meta-Llama-3-8B", "meta-llama/Meta-Llama-3.1-8B", "meta-llama/Meta-Llama-3-8B-Instruct", "meta-llama/Meta-Llama-3.1-8B-Instruct"]
pythia_family = [f"EleutherAI/pythia-{size}" for size in ["14m", "31m", "70m", "160m", "410m", "1b", "1.4b", "2.8b", "6.9b", "12b"]]
opt_family = [f"facebook/opt-{size}" for size in ["125m", "350m", "1.3b", "2.7b", "6.7b", "13b"]]
mistral_family = [f"mistralai/Mistral-7B-v0.1", f"mistralai/Mistral-7B-Instruct-v0.1"]
model_pool = gpt_family + llama2_family + llama3_family + pythia_family + opt_family + mistral_family
########################################
for model_path in tqdm(model_pool):
model_name = model_path.split("/")[-1]
os.makedirs(f"results/{model_name}", exist_ok=True)
model = AutoModelForCausalLM.from_pretrained(
model_path,
attn_implementation="eager",
# torch_dtype=torch.bfloat16,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(
model_path
)
#########################################
#########################################
# load data and feed them into LLMs
file_path = "datasets/probe_valid_natural.jsonl"
token_length = 64
states_path = f"results/{model_name}/states_token{token_length}.npy"
values_path = f"results/{model_name}/values_token{token_length}.npy"
keys_path = f"results/{model_name}/keys_token{token_length}.npy"
with open(file_path, 'r') as f:
prompts = [json.loads(line)["text"] for line in f]
measure_activations(model, tokenizer, prompts, states_path, keys_path, values_path, token_length, device)
# analysis norm
all_norms = compute_norm(states_path, device)
print(f"model name: {model_name}, numerical norm: {all_norms}.") # (layer, token)
all_k_norms = compute_kv_norm(keys_path, device)
print(f"model name: {model_name}, numerical norm for keys: {all_k_norms}.") # (layer, token)
all_v_norms = compute_kv_norm(values_path, device)
print(f"model name: {model_name}, numerical norm for values: {all_v_norms}.") # (layer, token)
if __name__ == "__main__":
measure_open_sourced_lms()