long-context-modeling/train_ramba_passkey.py at main · ant-research/long-context-modeling · 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
import torch
import argparse
from transformers import EvalPrediction, AutoTokenizer
from model.model_factory import create_model, load_pretrained
from transformers import AutoConfig, AutoTokenizer, TrainingArguments
# from model.modeling_mamba2 import Mamba2ForCausalLM
from model.modeling_mamba2_nsa import Mamba2ForCausalLM
from reader.dataset_new import TextDataset
from reader.dataset import LongRNNDataset
from reader.lazy_loader import LazyLoader, LazyChunkedLoader
from reader.data_collator import PasskeyRetrievalDataCollator
from trainer.mamba_trainer_passkey import MambaPasskeyTrainer
import json
import numpy as np
from flash_attn.losses.cross_entropy import CrossEntropyLoss
import torch.distributed as dist


def eval_accuracy(eval_pred: EvalPrediction):
    pred, gold_labels = eval_pred.predictions, eval_pred.label_ids
    acc = np.sum(pred == gold_labels) / pred.shape[0]
    return {"eval_accuracy": acc}


def run(args):
    if args.checkpoint_path is not None:
        model = load_pretrained(args.model_type, args.checkpoint_path, config_path=args.config_path)
    else:
        model = create_model(args.model_type, args.config_path)

    torch.set_printoptions(profile='full')
    ds = LazyChunkedLoader(args.train_path,  array_data_type=np.uint16)
    # dataset = LongRNNDataset(
    #     ds,
    #     batch_size=args.batch_size,
    #     segment_len=args.segment_len,
    #     segment_size=1,
    #     num_samples = args.total_steps,
    #     ramdom_sampling=True
    # )
    dataset = TextDataset(
        ds,
        batch_size=args.batch_size,
        segment_len=args.segment_len,
        num_samples = args.total_steps,
        ramdom_sampling=True,
        is_lazy=True,
        reset_state_samples=64
    )

    valid_ds = LazyLoader(args.valid_path)
    # valid_ds = LazyChunkedLoader(args.valid_path, array_data_type=np.uint16)
    valid_dataset = LongRNNDataset(
        valid_ds,
        batch_size=1,
        segment_len=args.segment_len,
        ramdom_sampling=False,
        segment_size=1,
        epochs=1
    )

    tokenizer = AutoTokenizer.from_pretrained(args.vocab_dir)
    data_collator = PasskeyRetrievalDataCollator(tokenizer, chunk_retrieval=False, segment_size=args.segment_size)

    trainer = MambaPasskeyTrainer(
        model=model,
        train_dataset=dataset,
        eval_dataset=valid_dataset,
        data_collator=data_collator.fn,
        compute_metrics=eval_accuracy,
        args=TrainingArguments(
            learning_rate=args.learning_rate,
            weight_decay=args.weight_decay,
            adam_beta1=args.adam_beta1,
            adam_beta2=args.adam_beta2,
            num_train_epochs=1,
            eval_strategy="steps",
            label_names=["labels"],
            metric_for_best_model="eval_accuracy",
            save_strategy="steps",
            save_total_limit=1,
            dataloader_num_workers=args.num_workers,
            per_device_train_batch_size=args.batch_size,
            per_device_eval_batch_size=1,
            gradient_accumulation_steps=args.gradient_accumulation_steps,
            optim=args.optim,
            remove_unused_columns=False,
            warmup_ratio=args.warmup_ratio,
            output_dir=args.output_dir,
            logging_steps=args.log_steps,
            save_steps=args.save_steps,
            eval_steps=args.eval_steps,
            bf16=True,
            save_safetensors=False,
            ddp_find_unused_parameters=False,
            lr_scheduler_type="cosine_with_min_lr",
            lr_scheduler_kwargs={
                "min_lr_rate":args.min_lr_rate
            }
        ),

    )

    trainer.train(resume_from_checkpoint=args.continue_training)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--config_path", type=str, required=True)
    parser.add_argument("--vocab_dir", type=str, required=True)
    parser.add_argument("--train_path", type=str, required=True)
    parser.add_argument("--valid_path", type=str, required=True)
    parser.add_argument("--model_type", type=str, required=True)
    parser.add_argument("--output_dir", type=str, required=True)
    parser.add_argument("--checkpoint_path", type=str, required=False, default=None)
    parser.add_argument("--warmup_ratio", type=float, default=0.02)
    parser.add_argument("--total_steps", type=int, required=True)
    parser.add_argument("--learning_rate", type=float, default=5e-5)
    parser.add_argument("--segment_len", type=int, default=16384)
    parser.add_argument("--segment_size", type=int, default=2, help="cut a sample into s segments")
    parser.add_argument("--batch_size", type=int, default=4)
    parser.add_argument("--min_lr_rate", type=float, default=0.2)
    parser.add_argument("--adam_beta1", type=float, default=0.9)
    parser.add_argument("--adam_beta2", type=float, default=0.95)
    parser.add_argument("--weight_decay", type=float, default=0.001)
    parser.add_argument("--save_steps", type=int, default=1000)
    parser.add_argument("--num_workers", type=int, default=2)
    parser.add_argument("--enabling_ibs", action="store_true")
    parser.add_argument("--continue_training", action="store_true")
    parser.add_argument("--reorg_prob", type=float, default=0.5)
    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
    parser.add_argument("--optim", type=str, default="adamw_torch")
    parser.add_argument("--log_steps", default=50, type=int)
    parser.add_argument("--eval_steps", default=1000, type=int)
    args = parser.parse_args()

    run(args)