AttnCache/test_bert.py at main · PASAUCMerced/AttnCache · 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
from transformers import AutoModelForSequenceClassification, AutoTokenizer, AutoConfig
import torch
from datasets import load_dataset
import numpy as np
from sklearn.metrics import accuracy_score
from datasets import load_dataset
import numpy as np
from sklearn.metrics import accuracy_score
import argparse
from statistics import mode
import torch
import numpy as np
import torch
from models.utils import VecDB, Emb, LatencyCollector, register_forward_latency_collector
import pickle
from sklearn.metrics import accuracy_score
from models.modeling_bert import BertForSequenceClassification

def parse_args():
    parser = argparse.ArgumentParser()

    parser.add_argument("--model-path", type=str, default="./trained_bert_sst2")
    parser.add_argument("--database-path", type=str, default="./BertDB")
    parser.add_argument("--dataset", type=str, default="sst2")
    parser.add_argument("--max-length", type=int, default=64)
    parser.add_argument("--threshold", type=float, default=1)
    parser.add_argument("--base-latency", type=float, default=32.4267707824707 )

    return parser.parse_args()

def retrieve(data):
    inputs = tokenizer(data, padding="max_length", truncation=True, max_length=args.max_length, return_tensors="pt")
    inputs = inputs.to(device)

    x = model.bert.embeddings(
            input_ids=inputs["input_ids"],
            token_type_ids=inputs["token_type_ids"],
            position_ids=None,
            inputs_embeds=None,
        )

    feature_vector = feature_projector.embed(x.cpu().detach().numpy())
    sims, idx_list = vecDB.search(feature_vector)

    reuse_tensor_index = np.flatnonzero(1 - sims >= args.threshold)
    hitted_records = idx_list[reuse_tensor_index]
    compute_tensor_index = np.flatnonzero(1 - sims < args.threshold)

    return reuse_tensor_index, hitted_records, compute_tensor_index, inputs


def evaluate(reuse_tensor_index, hitted_records, compute_tensor_index, inputs):

    with torch.no_grad():
        total_tensor_index = np.concatenate((reuse_tensor_index, compute_tensor_index), axis=0)
        attention_cache = torch.empty((config.num_hidden_layers, len(total_tensor_index), config.num_attention_heads, args.max_length, args.max_length))
        if len(reuse_tensor_index) != 0:
            print(f"=========== hit {len(reuse_tensor_index)} APMs")
            for layer_idx in range(config.num_hidden_layers):
                for idx, record in zip(reuse_tensor_index, hitted_records):
                    with open(f"{args.database_path}/APMsDB/{record[0]}.pickle", "rb") as file:
                        attn_weights = pickle.load(file)
                        attention_cache[layer_idx][idx] = torch.from_numpy(attn_weights)

                hitted_records += 1  # index of next layer hitted apms
        else:
            print(f"=========== no hit APMs")

        attention_cache = attention_cache.to(device)
        latency_collector = LatencyCollector()
        register_forward_latency_collector(latency_collector, model.bert.encoder.layer[-1].attention.self)

        for _ in range(100):
            outputs = model(**inputs, attention_cache=attention_cache, compute_tensor_index=compute_tensor_index)

    return outputs, latency_collector


if __name__ == "__main__":

    # =================== model ===================

    args = parse_args()
    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
    config = AutoConfig.from_pretrained(args.model_path)
    model = BertForSequenceClassification.from_pretrained(args.model_path)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model.to(device)
    model.eval()

    # =================== data ===================

    dataset = load_dataset("glue", args.dataset)
    validation_data = dataset["validation"] # sst2: 872
    test_data = dataset["test"] # sst2: 1821
    train_data = dataset["train"] # sst2: 67349


    # =================== feature_projector & vecDB ===================

    feature_projector_save_path = f'{args.database_path}/Embedding_models/mlp_model_attn_cache-epoch3.pth'
    feature_projector = Emb(f"{feature_projector_save_path}")
    vecDB_save_path =  f"{args.database_path}/VectorDB/attn_cache_epoch-3_vectors.faiss"
    vecDB = VecDB().load(f"{vecDB_save_path}")


    # =================== evaluation single validation_data ===================

    for example in validation_data:
        sentence = example["sentence"]

        reuse_tensor_index, hitted_records, compute_tensor_index, inputs = retrieve(sentence)

        if hitted_records.size:
            s1 = train_data[int(hitted_records[0][0] // 12)]
            s2 = example
            t1 = tokenizer.tokenize(s1["sentence"])
            t2 = tokenizer.tokenize(s2["sentence"])
            print("reuse idx in train_data: ", hitted_records[0][0] // 12)
            print("sentence 1 in train_data: ", s1)
            print("token 1: ", t1, "lenth 1: ", len(t1))
            print("sentence 2 in valid_data: ", s2)
            print("token 2: ", t2, "lenth 2: ", len(t2))

            outputs, latency_collector = evaluate(reuse_tensor_index, hitted_records, compute_tensor_index, inputs)

            attn_cache_latency = latency_collector.latency_list
            attn_cache_average_time = np.mean(attn_cache_latency) * 1000
            print(f"self-attn with_attn_cache average time: {attn_cache_average_time} ms")


            # =================== without AttnCache ===================

            inputs = tokenizer(sentence, padding="max_length", truncation=True, max_length=args.max_length, return_tensors="pt")
            inputs = inputs.to(device)
            latency_collector = LatencyCollector()
            register_forward_latency_collector(latency_collector, model.bert.encoder.layer[-1].attention.self)
            for _ in range(100):
                outputs = model(**inputs, attention_cache=None, compute_tensor_index=None)
            latency = latency_collector.latency_list
            average_time = np.mean(latency) * 1000
            print(f"self-attn without_attn_cache average time: {average_time} ms")

            print("average speedup: ", average_time / attn_cache_average_time)
            print('-' * 90)


    # =================== evaluation all validation_data ===================

    print(f"=========== evaluate all valid data")

    val_texts = [example["sentence"] for example in dataset["validation"]]
    val_labels = [example["label"] for example in dataset["validation"]]


    reuse_tensor_index, hitted_records, compute_tensor_index, inputs = retrieve(val_texts)
    outputs, latency_collector = evaluate(reuse_tensor_index, hitted_records, compute_tensor_index, inputs)

    logits = outputs.logits
    val_predictions = torch.argmax(logits, dim=-1).cpu().numpy()
    accuracy = accuracy_score(val_labels, val_predictions)
    print(f"Final Validation Accuracy: {accuracy:.4f}")

    latency = latency_collector.latency_list
    average_time = np.mean(latency) * 1000
    print(f"self-attn average time: {average_time} ms")
    print("average speedup: ", args.base_latency / average_time)