-
Notifications
You must be signed in to change notification settings - Fork 0
/
mm_trainer.py
501 lines (455 loc) · 19.7 KB
/
mm_trainer.py
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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
# pyre-strict
import os
from typing import Dict, List, Optional
import torch
import torch.nn as nn
from longvu.mm_datautils import get_mm_adapter_state_maybe_zero_3
from torch.utils.data import DataLoader, Sampler
from transformers import Trainer
from transformers.trainer import ALL_LAYERNORM_LAYERS, get_parameter_names, has_length
from transformers.trainer import SCHEDULER_NAME
VALID_CKPT_FILE="checkpoint_is_complete.flag"
# pyre-fixme[3]: Return type must be annotated.
# pyre-fixme[2]: Parameter must be annotated.
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
# pyre-fixme[3]: Return type must be annotated.
# pyre-fixme[2]: Parameter must be annotated.
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]
# pyre-fixme[3]: Return type must be annotated.
# pyre-fixme[2]: Parameter must be annotated.
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."
if all(l > 0 for l in lengths) or all(l < 0 for l in lengths):
# all samples are in the same modality
return get_length_grouped_indices(
lengths, batch_size, world_size, generator=generator
)
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])
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) > 0:
megabatches.append(sorted(additional_batch))
return [i for megabatch in megabatches for i in megabatch]
# pyre-fixme[3]: Return type must be annotated.
# pyre-fixme[2]: Parameter must be annotated.
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)
def get_mm_adapter_state_maybe_zero_3(
named_params: Dict[str, torch.Tensor], keys_to_match: List[str]
) -> Dict[str, torch.Tensor]:
to_return = {
k: t
for k, t in named_params
if any(key_match in k for key_match in keys_to_match)
}
to_return = {
k: maybe_zero_3(v, ignore_status=True, name=k).cpu() # pyre-ignore
for k, v in to_return.items()
}
return to_return
def maybe_zero_3(
param: torch.Tensor, ignore_status: bool = False, name: Optional[str] = None
) -> torch.Tensor:
return param.detach().cpu().clone()
class LLaVATrainer(Trainer):
def __init__(
self,
train_dataloader: Optional[DataLoader] = None,
# pyre-fixme[2]: Parameter must be annotated.
*args,
# pyre-fixme[2]: Parameter must be annotated.
**kwargs,
) -> None:
self.train_dataloader = train_dataloader
super().__init__(*args, **kwargs)
def get_train_dataloader(self) -> DataLoader:
if self.train_dataloader is not None:
print("Using sonic dataloader")
# pyre-fixme[16]: `LLaVATrainer` has no attribute `accelerator`.
return self.accelerator.prepare(self.train_dataloader)
# pyre-fixme[16]: `Trainer` has no attribute `get_train_dataloader`.
return super().get_train_dataloader()
def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
# pyre-fixme[16]: `LLaVATrainer` has no attribute `train_dataset`.
if self.train_dataset is None or not has_length(self.train_dataset):
return None
# pyre-fixme[16]: `LLaVATrainer` has no attribute `args`.
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:
# pyre-fixme[16]: `Trainer` has no attribute `_get_train_sampler`.
return super()._get_train_sampler()
# pyre-fixme[3]: Return type must be annotated.
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.
"""
# pyre-fixme[16]: `Trainer` has no attribute `model`.
opt_model = self.model
# if self.args.unfreeze_mm_vision_tower:
# opt_model.get_model().vision_tower_aux_list = nn.ModuleList(opt_model.get_vision_tower_aux_list())
# self.param_to_name = map_params_to_module_names([opt_model])
# pyre-fixme[16]: `Trainer` has no attribute `optimizer`.
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]
# pyre-fixme[16]: `Trainer` has no attribute `mm_projector_lr`.
assert not (self.args.mm_projector_lr and self.args.mm_vision_sampler_lr)
if self.args.mm_projector_lr is not None:
projector_parameters = [
name
for name, _ in opt_model.named_parameters()
if "mm_projector" in name
]
optimizer_grouped_parameters = [
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n not in projector_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n not in projector_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n in projector_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
"lr": self.args.mm_projector_lr,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n in projector_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
"lr": self.args.mm_projector_lr,
},
]
elif self.args.mm_vision_sampler_lr is not None:
vision_sampler_parameters = [
name
for name, _ in opt_model.named_parameters()
if ("vision_sampler" in name) or ("vision_query" in name)
]
optimizer_grouped_parameters = [
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n not in vision_sampler_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n not in vision_sampler_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n in vision_sampler_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
"lr": self.args.mm_vision_sampler_lr,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n in vision_sampler_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
"lr": self.args.mm_vision_sampler_lr,
},
]
elif (
self.args.unfreeze_mm_vision_tower
and self.args.mm_vision_tower_lr is not None
):
vision_tower_parameters = [
name
for name, _ in opt_model.named_parameters()
if "vision_tower" in name
]
optimizer_grouped_parameters = [
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n not in vision_tower_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n not in vision_tower_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n in decay_parameters
and n in vision_tower_parameters
and p.requires_grad
)
],
"weight_decay": self.args.weight_decay,
"lr": self.args.mm_vision_tower_lr,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (
n not in decay_parameters
and n in vision_tower_parameters
and p.requires_grad
)
],
"weight_decay": 0.0,
"lr": self.args.mm_vision_tower_lr,
},
]
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,
},
{
"params": [
p
for n, p in opt_model.named_parameters()
if (n not in decay_parameters and p.requires_grad)
],
"weight_decay": 0.0,
},
]
optimizer_cls, optimizer_kwargs = Trainer.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()
)
manager.register_module_override(
module, "weight", {"optim_bits": 32}
)
return self.optimizer
# pyre-fixme[2]: Parameter must be annotated.
def _save_checkpoint(self, model, trial, metrics=None) -> None:
super(LLaVATrainer, self)._save_checkpoint(model, trial, metrics)
# pyre-fixme[16]: `LLaVATrainer` has no attribute `args`.
if(self.args.should_save):
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
checkpoint_folder = (
f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}")
run_dir = self._get_output_dir(trial=trial)
output_dir = os.path.join(run_dir, checkpoint_folder)
torch.save(self.lr_scheduler.state_dict(), os.path.join(
output_dir, SCHEDULER_NAME))
# Only save Adapter
keys_to_match = ["mm_projector", "vision_resampler"]
if getattr(self.args, "use_im_start_end", False):
keys_to_match.extend(["embed_tokens", "embed_in"])
weight_to_save = get_mm_adapter_state_maybe_zero_3(
self.model.named_parameters(),
keys_to_match)
if self.args.local_rank == 0 or self.args.local_rank == -1:
self.model.config.save_pretrained(output_dir)
torch.save(weight_to_save, os.path.join(
output_dir, "mm_projector.bin"))
with(open(os.path.join(output_dir, VALID_CKPT_FILE), "w") as f):
f.write("this is a valid checkpoint")
def _load_optimizer_and_scheduler(self, checkpoint):
#the default function immediately returns when used with deepspeed
# => the lr scheduler state is never loaded
if(checkpoint is None):
return
self.lr_scheduler.load_state_dict(torch.load(os.path.join(
checkpoint, SCHEDULER_NAME)))
# pyre-fixme[2]: Parameter must be annotated.
# def _save(self, output_dir: Optional[str] = None, state_dict=None) -> None:
# # pyre-fixme[16]: `LLaVATrainer` has no attribute `args`.
# if getattr(self.args, "tune_mm_mlp_adapter", False):
# pass
# else:
# # pyre-fixme[16]: `Trainer` has no attribute `_save`.
# super(LLaVATrainer, self)._save(output_dir, state_dict)