File size: 10,017 Bytes
74b17e0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
import os
import torch
from torch import nn

from torch.utils.data import Sampler

from transformers import Trainer
from transformers.trainer import (
    is_sagemaker_mp_enabled,
    get_parameter_names,
    has_length,
    ALL_LAYERNORM_LAYERS,
    # ShardedDDPOption,
    logger,
)
from typing import List, Optional

from ..utils.train_utils import *


def split_to_even_chunks(indices, lengths, num_chunks):
    """
    Split a list of indices into `chunks` chunks of roughly equal lengths.
    """

    if len(indices) % num_chunks != 0:
        return [indices[i::num_chunks] for i in range(num_chunks)]

    num_indices_per_chunk = len(indices) // num_chunks

    chunks = [[] for _ in range(num_chunks)]
    chunks_lengths = [0 for _ in range(num_chunks)]
    for index in indices:
        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
        chunks[shortest_chunk].append(index)
        chunks_lengths[shortest_chunk] += lengths[index]
        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
            chunks_lengths[shortest_chunk] = float("inf")

    return chunks


def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None):
    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
    assert all(l != 0 for l in lengths), "Should not have zero length."
    mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0])
    lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0])

    assert len(mm_indices) > 0, "Should have at least one multimodal sample."
    assert len(lang_indices) > 0, "Should have at least one language sample."

    mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)]
    lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)]
    megabatch_size = world_size * batch_size
    mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)]
    lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)]

    last_mm = mm_megabatches[-1]
    last_lang = lang_megabatches[-1]
    additional_batch = last_mm + last_lang
    megabatches = mm_megabatches[:-1] + lang_megabatches[:-1]
    megabatch_indices = torch.randperm(len(megabatches), generator=generator)
    megabatches = [megabatches[i] for i in megabatch_indices]

    if len(additional_batch) >= megabatch_size:
        megabatches = [additional_batch[:megabatch_size]] + megabatches
        additional_batch = additional_batch[megabatch_size:]

    if len(additional_batch) > 0:
        megabatches.append(additional_batch)

    return [i for megabatch in megabatches for i in megabatch]


def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
    indices = torch.randperm(len(lengths), generator=generator)
    megabatch_size = world_size * batch_size
    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]

    return [i for megabatch in megabatches for batch in megabatch for i in batch]


class LengthGroupedSampler(Sampler):
    r"""
    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
    keeping a bit of randomness.
    """

    def __init__(
        self,
        batch_size: int,
        world_size: int,
        lengths: Optional[List[int]] = None,
        generator=None,
        group_by_modality: bool = False,
    ):
        if lengths is None:
            raise ValueError("Lengths must be provided.")

        self.batch_size = batch_size
        self.world_size = world_size
        self.lengths = lengths
        self.generator = generator
        self.group_by_modality = group_by_modality

    def __len__(self):
        return len(self.lengths)

    def __iter__(self):
        if self.group_by_modality:
            indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
        else:
            indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
        return iter(indices)


class LLaVATrainer(Trainer):

    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
        if self.train_dataset is None or not has_length(self.train_dataset):
            return None

        if self.args.group_by_modality_length:
            lengths = self.train_dataset.modality_lengths
            return LengthGroupedSampler(
                # self.args.train_batch_size * self.args.gradient_accumulation_steps, # TODO: seems that we should not have gradient_accumulation_steps
                self.args.train_batch_size,
                world_size=self.args.world_size,
                lengths=lengths,
                group_by_modality=True,
            )
        else:
            return super()._get_train_sampler()

    def create_optimizer(self):
        """
        Setup the optimizer.

        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
        Trainer's init through `optimizers`, or subclass and override this method in a subclass.
        """
        if is_sagemaker_mp_enabled():
            return super().create_optimizer()
        # if self.sharded_ddp == ShardedDDPOption.SIMPLE:
        #     return super().create_optimizer()

        opt_model = self.model

        if self.optimizer is None:
            decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS)
            decay_parameters = [name for name in decay_parameters if "bias" not in name]
            if self.args.mm_projector_lr is not None:
                connector_parameters = [name for name, _ in opt_model.named_parameters() if "connector" in name]
                optimizer_grouped_parameters = [
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and n not in connector_parameters and p.requires_grad)
                        ],
                        "weight_decay": self.args.weight_decay,
                        "name": "decay_no_connector_parameters"
                    },
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n not in connector_parameters and p.requires_grad)
                        ],
                        "weight_decay": 0.0,
                        "name": "no_decay_no_connector_parameters"
                    },
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and n in connector_parameters and p.requires_grad)
                        ],
                        "weight_decay": self.args.weight_decay,
                        "lr": self.args.mm_projector_lr,
                        "name": "decay_connector_parameters"
                    },
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and n in connector_parameters and p.requires_grad)
                        ],
                        "weight_decay": 0.0,
                        "lr": self.args.mm_projector_lr,
                        "name": "no_decay_proj_parameters"
                    },
                ]
            else:
                optimizer_grouped_parameters = [
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n in decay_parameters and p.requires_grad)
                        ],
                        "weight_decay": self.args.weight_decay,
                        "name": "decay_parameters"
                    },
                    {
                        "params": [
                            p for n, p in opt_model.named_parameters() if (n not in decay_parameters and p.requires_grad)
                        ],
                        "weight_decay": 0.0,
                        "name": "no_decay_parameters"
                    },
                ]

            if getattr(self.args, "moe_enable", False):
                from deepspeed.moe.utils import split_params_into_different_moe_groups_for_optimizer
                optimizer_grouped_parameters = split_params_into_different_moe_groups_for_optimizer(optimizer_grouped_parameters)
            optimizer_cls, optimizer_kwargs = self.get_optimizer_cls_and_kwargs(self.args)

            self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
            if optimizer_cls.__name__ == "Adam8bit":
                import bitsandbytes

                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()

                skipped = 0
                for module in opt_model.modules():
                    if isinstance(module, nn.Embedding):
                        skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
                        logger.info(f"skipped {module}: {skipped/2**20}M params")
                        manager.register_module_override(module, "weight", {"optim_bits": 32})
                        logger.debug(f"bitsandbytes: will optimize {module} in fp32")
                logger.info(f"skipped: {skipped/2**20}M params")

        return self.optimizer