File size: 4,797 Bytes
cf912f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5158692
cf912f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d028752
cf912f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import sys
import torch
from dotenv import find_dotenv, load_dotenv

found_dotenv = find_dotenv(".env")

if len(found_dotenv) == 0:
    found_dotenv = find_dotenv(".env.example")
print(f"loading env vars from: {found_dotenv}")
load_dotenv(found_dotenv, override=False)

path = os.path.dirname(found_dotenv)
print(f"Adding {path} to sys.path")
sys.path.append(path)

from llm_toolkit.llm_utils import *
from llm_toolkit.logical_reasoning_utils import *

device = check_gpu()
is_cuda = torch.cuda.is_available()

model_name = os.getenv("MODEL_NAME")
adapter_name_or_path = os.getenv("ADAPTER_NAME_OR_PATH")
load_in_4bit = os.getenv("LOAD_IN_4BIT") == "true"
data_path = os.getenv("LOGICAL_REASONING_DATA_PATH")
results_path = os.getenv("LOGICAL_REASONING_RESULTS_PATH")
use_english_datasets = os.getenv("USE_ENGLISH_DATASETS") == "true"
batch_size = int(os.getenv("BATCH_SIZE", 1))
using_llama_factory = os.getenv("USING_LLAMA_FACTORY") == "true"
max_new_tokens = int(os.getenv("MAX_NEW_TOKENS", 2048))
start_num_shots = int(os.getenv("START_NUM_SHOTS", 0))

print(
    model_name,
    adapter_name_or_path,
    load_in_4bit,
    data_path,
    results_path,
    max_new_tokens,
    batch_size,
)

dtype = (
    torch.float32
    if os.getenv("USE_FLOAT32_FOR_INFERENCE") == "true"
    else (
        torch.bfloat16
        if os.getenv("USE_BF16_FOR_INFERENCE") == "true"
        else torch.float16
    )
)

if is_cuda:
    torch.cuda.empty_cache()
    gpu_stats = torch.cuda.get_device_properties(0)
    start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
    max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
    print(f"(0) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
    print(f"{start_gpu_memory} GB of memory reserved.")

    torch.cuda.empty_cache()

model, tokenizer = load_model(
    model_name,
    load_in_4bit=load_in_4bit,
    adapter_name_or_path=adapter_name_or_path,
    using_llama_factory=using_llama_factory,
    dtype=dtype,
)

if is_cuda:
    gpu_stats = torch.cuda.get_device_properties(0)
    start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
    max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
    print(f"(2) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
    print(f"{start_gpu_memory} GB of memory reserved.")


def on_num_shots_step_completed(model_name, dataset, predictions):
    save_results(
        model_name,
        results_path,
        dataset,
        predictions,
    )

    metrics = calc_metrics(dataset["label"], predictions, debug=True)
    print(f"{model_name} metrics: {metrics}")


if adapter_name_or_path is not None:
    model_name += "/" + adapter_name_or_path.split("/")[-1]


def evaluate_model_with_num_shots(
    model,
    tokenizer,
    model_name,
    data_path,
    start_num_shots=0,
    range_num_shots=[0, 5, 10, 20, 30, 40, 50],
    batch_size=1,
    max_new_tokens=2048,
    device="cuda",
):
    print(f"Evaluating model: {model_name} on {device}")

    for num_shots in range_num_shots:
        if num_shots < start_num_shots:
            continue

        print(f"*** Evaluating with num_shots: {num_shots}")

        datasets = load_logical_reasoning_dataset(
            data_path,
            tokenizer=tokenizer,
            chinese_prompt=not use_english_datasets,
            using_p1=False,
            num_shots=num_shots,
        )
        if len(sys.argv) > 1:
            num = int(sys.argv[1])
            if num > 0:
                print(f"--- evaluating {num} entries")
                datasets["test"] = datasets["test"].select(range(num))

        print_row_details(datasets["test"].to_pandas(), indices=[0, -1])

        predictions = eval_model(
            model,
            tokenizer,
            datasets["test"],
            device=device,
            batch_size=batch_size,
            max_new_tokens=max_new_tokens,
        )

        model_name_with_rp = f"{model_name}/shots-{num_shots:02d}"

        try:
            on_num_shots_step_completed(
                model_name_with_rp,
                datasets["test"],
                predictions,
            )
        except Exception as e:
            print(e)


evaluate_model_with_num_shots(
    model,
    tokenizer,
    model_name,
    data_path,
    batch_size=batch_size,
    max_new_tokens=max_new_tokens,
    device=device,
    start_num_shots=start_num_shots,
)

if is_cuda:
    gpu_stats = torch.cuda.get_device_properties(0)
    start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
    max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
    print(f"(3) GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
    print(f"{start_gpu_memory} GB of memory reserved.")