StableDRL/math_metrics_arpo.py at master · cure-lab/StableDRL · 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
import click
from tqdm import tqdm
from typing import Sequence
import torch
import torch.distributed as dist
from torch.utils.data import DataLoader, DistributedSampler
from transformers import AutoTokenizer
from datasets import load_dataset

from networks.lladou_arpo import LLaDOUARPOModel, sample_arpo
from dataloaders.collate_fn_math import collate_fn_math, extract_answer_gsm8k, collate_fn_gsm8k
from dataloaders.math import _math_verify_equal


def judge_answer_MATH(answers: Sequence[str], responses: Sequence[str], counts):
    counts[1] += len(answers)
    for ans, res in zip(answers, responses):
        if _math_verify_equal(ans, res):
            counts[0] += 1
    return counts


def judge_answer_GSM8K(answers: Sequence[str], responses: Sequence[str], counts):
    ext_ans = [extract_answer_gsm8k(ans) for ans in answers]
    counts[1] += len(ext_ans)
    for ans, res in zip(ext_ans, responses):
        if _math_verify_equal(ans, res):
            counts[0] += 1
    return counts


@click.command()
@click.option("--ckpt_path", type=str, default="")
@click.option("--local_data_path", type=str, default="openai/gsm8k")
@click.option("--batch_size", type=int, default=1)
@click.option("--num_workers", type=int, default=1)
@click.option("--steps", type=int, default=256)
@click.option("--gen_length", type=int, default=256)
@click.option("--block_length", type=int, default=8)
@click.option("--task", type=str, default="gsm8k")
@click.option("--seed", type=int, default=42)
@click.option("--no_sample", type=bool, default=False)
def main(
    ckpt_path: str = "",
    local_data_path: str = "openai/gsm8k",
    batch_size: int = 1,
    num_workers: int = 1,
    steps: int = 256,
    gen_length: int = 256,
    _block_length: int = 8,
    no_sample: bool = False,
    seed: int = 42,
    **_kwargs,
):
    # Distributed setup
    torch.distributed.init_process_group(backend="nccl")
    torch.cuda.set_device(dist.get_rank())
    torch.manual_seed(seed)

    device = "cuda"
    tokenizer = AutoTokenizer.from_pretrained("GSAI-ML/LLaDA-8B-Instruct")
    tokenizer.pad_token_id = 126081

    model = LLaDOUARPOModel.from_pretrained(
        pretrained_model_name_or_path=ckpt_path,
        trust_remote_code=True,
        torch_dtype=torch.bfloat16,
    )
    model.eval().requires_grad_(False).to(device)

    # load data
    if "MATH" in local_data_path:
        ds = load_dataset(local_data_path, split="test").with_format("torch")
    elif "gsm8k" in local_data_path:
        ds = load_dataset(local_data_path, split="test", data_dir="main").with_format("torch")
    else:
        raise ValueError(f"Invalid data path: {local_data_path}")

    sampler = DistributedSampler(ds, rank=dist.get_rank(), num_replicas=dist.get_world_size(), shuffle=False)
    collate_fn = collate_fn_math if "MATH" in local_data_path else collate_fn_gsm8k
    dl = DataLoader(
        ds,
        batch_size=batch_size,
        collate_fn=collate_fn,
        num_workers=num_workers,
        pin_memory=True,
        sampler=sampler,
    )

    pbar = tqdm(dl, disable=dist.get_rank() != 0)
    counts = torch.tensor([0, 0], device=device)  # correct, total

    for _, batch in enumerate(pbar):
        answers = batch["answers"]

        outputs = sample_arpo(
            model,
            batch,
            tokenizer,
            device=device,
            inference=no_sample,
            max_steps=steps,
            gen_length=gen_length,
            temperature=0.3,
        )
        responses = tokenizer.batch_decode(outputs["final_tokens"], skip_special_tokens=True)

        if "MATH" in local_data_path:
            counts = judge_answer_MATH(answers, responses, counts)
        elif "gsm8k" in local_data_path:
            counts = judge_answer_GSM8K(answers, responses, counts)

        if dist.get_rank() == 0:
            counts_list = [counts.clone() for _ in range(dist.get_world_size())]
        else:
            counts_list = None

        # gather acc
        torch.distributed.gather(counts, counts_list, dst=0)
        if dist.get_rank() == 0:
            counts_list = torch.stack(counts_list, dim=0).sum(dim=0)
            acc = counts_list[0] / counts_list[1]
            pbar.set_description(f"acc: {acc.item() * 100:.2f}%")

    if dist.get_rank() == 0:
        print(counts_list)
        print("Final Acc: ", acc)


if __name__ == "__main__":
    main()