-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy patheval.py
182 lines (142 loc) · 8.32 KB
/
eval.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
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
import time
import json
from dataclasses import dataclass
import torch
import gc
from transformers import AutoTokenizer
import warnings
from TuringLLM.inference import TuringLLMForInference
from SAE.SAE_TopK import SAE
subset_layer_latent_count = 82
num_tokens_per_sequence = 64
sequences_per_batch = 256
device = 'cuda' if torch.cuda.is_available() else 'cpu'
@dataclass
class TuringLLMConfig:
block_size: int = 1024
vocab_size: int = 50304
n_layer: int = 12
n_head: int = 16
n_embd: int = 1024
hidden_size: int = 4096
norm_eps: float = 1e-5
sae_dim = 40960
turing_sae_latent_values = torch.zeros((12, subset_layer_latent_count))
warnings.filterwarnings("ignore")
def run():
print("")
torch.cuda.empty_cache()
gc.collect()
print("Loading Turing-LLM...")
max_length = 64 + 1
turing = TuringLLMForInference(collect_latents=True, max_length=max_length)
# Tokenizer
tokenizer_model_id = "microsoft/Phi-3-mini-4k-instruct"
try:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_model_id, use_fast=True, local_files_only=True, _fast_init=True)
except:
tokenizer = AutoTokenizer.from_pretrained(tokenizer_model_id, use_fast=True, _fast_init=True)
print("")
space_token = tokenizer.encode(" ")[0]
for layer_index in range(TuringLLMConfig.n_layer):
print(f"Layer {layer_index+1}")
# Load SAE
sae_model = SAE(TuringLLMConfig.n_embd, sae_dim, 128, only_encoder=True).to(device)
sae_model = torch.compile(sae_model)
sae_model.load(f"./SAE/sae/sae_layer_{layer_index}.pth")
if layer_index == 0:
sae_model.k = 128 + (4 * 16)
else:
sae_model.k = 128 + (layer_index * 16)
# Load Data
layer_eval_data = []
with open(f'./latent_data/{layer_index}/eval_inputs.jsonl', 'r') as f:
for i, line in enumerate(f):
layer_eval_data.append(json.loads(line))
# Tokenize Text
for i, layer_eval_input in enumerate(layer_eval_data):
layer_eval_input["text_tokens"] = tokenizer.encode(layer_eval_input["text"])[:64]
if len(layer_eval_input["text_tokens"]) == 0:
layer_eval_data[i] = None
continue
if len(layer_eval_input["text_tokens"]) < 64:
while len(layer_eval_input["text_tokens"]) < 64:
layer_eval_input["text_tokens"] = layer_eval_input["text_tokens"] + layer_eval_input["text_tokens"]
layer_eval_input["text_tokens"] = layer_eval_input["text_tokens"][:64]
if layer_eval_input["type"] == "top-token":
layer_eval_input["text_tokens"] = layer_eval_input["text_tokens"][:62] + [space_token] + [layer_eval_input["token"]]
elif layer_eval_input["type"] == "connecting-tokens":
layer_eval_input["text_tokens"] = layer_eval_input["text_tokens"][:60] + [space_token] + [layer_eval_input["tokens"][0]] + [space_token] + [layer_eval_input["tokens"][1]]
layer_eval_data = [layer_eval_input for layer_eval_input in layer_eval_data if layer_eval_input is not None]
num_activation_values_added = 0
for batch_index in range(0, len(layer_eval_data), sequences_per_batch):
batch_start_time = time.time()
batch_layer_eval_data = layer_eval_data[batch_index:batch_index+sequences_per_batch]
batch_length = len(batch_layer_eval_data)
print(f" Processing Latents {batch_index+1}-{batch_index+batch_length} / {len(layer_eval_data)} | Turing Inference ", end="\r")
with torch.no_grad():
max_length = num_tokens_per_sequence + 1
logits, latents = turing.generate_batch([input["text_tokens"] for input in batch_layer_eval_data], max_length=max_length, tokenize=False, decode=False, ignore_end=True)
print(f" Processing Latents {batch_index+1}-{batch_index+batch_length} / {len(layer_eval_data)} | SAE Inference ", end="\r")
with torch.no_grad():
activations = [torch.split(layer_latents.view(-1, 1024), num_tokens_per_sequence * batch_length)[0] for layer_latents in latents[0]]
x = activations[layer_index].to(device)
sae_latents, _, _ = sae_model.encode(x)
sae_latents = sae_latents.view(batch_length, -1, sae_dim).to(device)
for i in range(batch_length):
if (i+batch_index) >= len(layer_eval_data):
continue
latent_index = layer_eval_data[i+batch_index]["latent_index"]
text_tokens = layer_eval_data[i+batch_index]["text_tokens"]
eval_data_type = layer_eval_data[i+batch_index]["type"]
if eval_data_type == "top-token":
# Find token
token = layer_eval_data[i+batch_index]["token"]
if token not in text_tokens:
layer_eval_data[i+batch_index]["tokens_not_found"] = True
continue
token_index = text_tokens.index(token)
layer_eval_data[i+batch_index]["activation_value"] = sae_latents[i][token_index][latent_index].item()
elif eval_data_type == "connecting-tokens":
# Find tokens and get average act
tokens = layer_eval_data[i+batch_index]["tokens"]
if tokens[0] not in text_tokens or tokens[1] not in text_tokens:
layer_eval_data[i+batch_index]["tokens_not_found"] = True
continue
token_0_index = text_tokens.index(tokens[0])
token_1_index = text_tokens.index(tokens[1])
activation_token_0 = sae_latents[i][token_0_index][latent_index].item()
activation_token_1 = sae_latents[i][token_1_index][latent_index].item()
layer_eval_data[i+batch_index]["activation_value"] = (activation_token_0 + activation_token_1) / 2
elif eval_data_type == "detecting-dataset-topic":
# Get top value
layer_eval_data[i+batch_index]["activation_value"] = max(sae_latents[i][token_index][latent_index].item() for token_index in range(64))
for token_sae_latents in sae_latents[i]:
turing_sae_latent_values[layer_index] += torch.tensor(token_sae_latents[0:subset_layer_latent_count], device="cpu")
num_activation_values_added += 1
print(f" Processing Latents {batch_index+1}-{batch_index+batch_length} / {len(layer_eval_data)} | Duration: {time.time()-batch_start_time:.2f}s ")
# Get Average Latent Values
turing_sae_latent_values[layer_index] = turing_sae_latent_values[layer_index] / num_activation_values_added
# Process Eval Data
for i, layer_eval_dict in enumerate(layer_eval_data):
if "tokens_not_found" in layer_eval_dict and layer_eval_dict["tokens_not_found"] is True:
continue
latent_index = layer_eval_dict["latent_index"]
latent_activation_value = layer_eval_dict["activation_value"]
latent_avg_activation_value = turing_sae_latent_values[layer_index][latent_index].item()
latent_activation_value_distance_from_avg = abs(latent_activation_value - latent_avg_activation_value)
layer_eval_data[i]["activation_distance"] = latent_activation_value_distance_from_avg
if latent_activation_value_distance_from_avg > 0.01:
layer_eval_data[i]["success"] = True
else:
layer_eval_data[i]["success"] = False
# Save Eval Data
with open(f"./latent_data/{layer_index}/eval_results.jsonl", 'w') as f:
for layer_eval_dict in layer_eval_data:
f.write(json.dumps(layer_eval_dict) + '\n')
# Clear for Next Layer
del sae_model
torch.cuda.empty_cache()
gc.collect()
print("")
run()