-
Notifications
You must be signed in to change notification settings - Fork 251
/
mgie_train.py
834 lines (706 loc) · 32.4 KB
/
mgie_train.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
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
#
# For licensing see accompanying LICENSE file.
# Copyright (C) 2024 Apple Inc. All Rights Reserved.
#
# modified from https://github.com/haotian-liu/LLaVA/blob/7ace501183c4bdec6052ec1a30039cdc3242a67c/llava/train/train.py
import os
import copy
from dataclasses import dataclass, field
import json
import logging
import pathlib
from typing import Dict, Optional, Sequence, List
import torch
import transformers
from torch.utils.data import Dataset
from llava.train.llava_trainer import LLaVATrainer
from llava import conversation as conversation_lib
from llava.model import *
from PIL import Image
import torch.nn as nn
# TODO: import and use code from ../data/dataset.py
IGNORE_INDEX = -100
DEFAULT_PAD_TOKEN = "[PAD]"
DEFAULT_EOS_TOKEN = "</s>"
DEFAULT_BOS_TOKEN = "<s>"
DEFAULT_UNK_TOKEN = "<unk>"
DEFAULT_IMAGE_TOKEN = "<image>"
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
DEFAULT_IM_START_TOKEN = "<im_start>"
DEFAULT_IM_END_TOKEN = "<im_end>"
import io, base64, pickle, random
from tqdm import tqdm
import numpy as np
def b2f(b): return Image.open(io.BytesIO(base64.b64decode(b))).convert('RGB')
def resize(f):
w, h = f.size
if w>h:
p = (w-h)//2
f = f.crop([p, 0, p+h, h])
elif h>w:
p = (h-w)//2
f = f.crop([0, p, w, p+w])
f = f.resize([512, 512])
return f
def img2npy(f): return (2.0*np.array(f)/255.0-1.0).transpose((2, 0, 1)).astype(np.float32)
@dataclass
class ModelArguments:
model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
version: Optional[str] = field(default="v0")
freeze_backbone: bool = field(default=False)
tune_mm_mlp_adapter: bool = field(default=False)
vision_tower: Optional[str] = field(default=None)
mm_vision_select_layer: Optional[int] = field(default=-1) # default to the last layer
pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
mm_use_im_start_end: bool = field(default=False)
@dataclass
class DataArguments:
data_path: str = field(default=None,
metadata={"help": "Path to the training data."})
lazy_preprocess: bool = False
is_multimodal: bool = False
sep_image_conv_front: bool = False
image_token_len: int = 0
image_folder: Optional[str] = field(default=None)
image_aspect_ratio: str = 'square'
@dataclass
class TrainingArguments(transformers.TrainingArguments):
cache_dir: Optional[str] = field(default=None)
optim: str = field(default="adamw_torch")
remove_unused_columns: bool = field(default=False)
freeze_mm_mlp_adapter: bool = field(default=False)
force_fsdp: bool = field(default=False)
model_max_length: int = field(
default=512,
metadata={
"help":
"Maximum sequence length. Sequences will be right padded (and possibly truncated)."
},
)
double_quant: bool = field(
default=True,
metadata={"help": "Compress the quantization statistics through double quantization."}
)
quant_type: str = field(
default="nf4",
metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
)
bits: int = field(
default=16,
metadata={"help": "How many bits to use."}
)
lora_enable: bool = False
lora_r: int = 64
lora_alpha: int = 16
lora_dropout: float = 0.05
lora_weight_path: str = ""
lora_bias: str = "none"
def maybe_zero_3(param, ignore_status=False, name=None):
from deepspeed import zero
from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
if hasattr(param, "ds_id"):
if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
if not ignore_status:
logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
with zero.GatheredParameters([param]):
param = param.data.detach().cpu().clone()
else:
param = param.detach().cpu().clone()
return param
# Borrowed from peft.utils.get_peft_model_state_dict
def get_peft_state_maybe_zero_3(named_params, bias):
if bias == "none":
to_return = {k: t for k, t in named_params if "lora_" in k}
elif bias == "all":
to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
elif bias == "lora_only":
to_return = {}
maybe_lora_bias = {}
lora_bias_names = set()
for k, t in named_params:
if "lora_" in k:
to_return[k] = t
bias_name = k.split("lora_")[0] + "bias"
lora_bias_names.add(bias_name)
elif "bias" in k:
maybe_lora_bias[k] = t
for k, t in maybe_lora_bias:
if bias_name in lora_bias_names:
to_return[bias_name] = t
else:
raise NotImplementedError
to_return = {k: maybe_zero_3(v, name=k) for k, v in to_return.items()}
return to_return
def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
to_return = {k: t for k, t in named_params if "lora_" not in k}
if require_grad_only:
to_return = {k: t for k, t in to_return.items() if t.requires_grad}
to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
return to_return
def find_all_linear_names(model):
cls = torch.nn.Linear
lora_module_names = set()
for name, module in model.named_modules():
if isinstance(module, cls):
names = name.split('.')
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
if 'lm_head' in lora_module_names: # needed for 16-bit
lora_module_names.remove('lm_head')
return list(lora_module_names)
def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
output_dir: str):
"""Collects the state dict and dump to disk."""
if trainer.deepspeed:
torch.cuda.synchronize()
trainer.save_model(output_dir)
return
state_dict = trainer.model.state_dict()
if trainer.args.should_save:
cpu_state_dict = {
key: value.cpu()
for key, value in state_dict.items()
}
del state_dict
trainer._save(output_dir, state_dict=cpu_state_dict) # noqa
def smart_tokenizer_and_embedding_resize(
special_tokens_dict: Dict,
tokenizer: transformers.PreTrainedTokenizer,
model: transformers.PreTrainedModel,
):
"""Resize tokenizer and embedding.
Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
"""
num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
model.resize_token_embeddings(len(tokenizer))
if num_new_tokens > 0:
input_embeddings = model.get_input_embeddings().weight.data
output_embeddings = model.get_output_embeddings().weight.data
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
dim=0, keepdim=True)
output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
dim=0, keepdim=True)
input_embeddings[-num_new_tokens:] = input_embeddings_avg
output_embeddings[-num_new_tokens:] = output_embeddings_avg
def _tokenize_fn(strings: Sequence[str],
tokenizer: transformers.PreTrainedTokenizer) -> Dict:
"""Tokenize a list of strings."""
tokenized_list = [
tokenizer(
text,
return_tensors="pt",
padding="longest",
max_length=tokenizer.model_max_length,
truncation=True,
) for text in strings
]
input_ids = labels = [
tokenized.input_ids[0] for tokenized in tokenized_list
]
input_ids_lens = labels_lens = [
tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
for tokenized in tokenized_list
]
return dict(
input_ids=input_ids,
labels=labels,
input_ids_lens=input_ids_lens,
labels_lens=labels_lens,
)
def _mask_targets(target, tokenized_lens, speakers):
# cur_idx = 0
cur_idx = tokenized_lens[0]
tokenized_lens = tokenized_lens[1:]
target[:cur_idx] = IGNORE_INDEX
for tokenized_len, speaker in zip(tokenized_lens, speakers):
if speaker == "human":
target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
cur_idx += tokenized_len
def _add_speaker_and_signal(header, source, get_conversation=True):
"""Add speaker and start/end signal on each round."""
BEGIN_SIGNAL = "### "
END_SIGNAL = "\n"
conversation = header
for sentence in source:
from_str = sentence["from"]
if from_str.lower() == "human":
from_str = conversation_lib.default_conversation.roles[0]
elif from_str.lower() == "gpt":
from_str = conversation_lib.default_conversation.roles[1]
else:
from_str = 'unknown'
sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
sentence["value"] + END_SIGNAL)
if get_conversation:
conversation += sentence["value"]
conversation += BEGIN_SIGNAL
return conversation
def preprocess_multimodal(
sources: Sequence[str],
multimodal_cfg: dict,
cur_token_len: int,
) -> Dict:
is_multimodal = multimodal_cfg['is_multimodal']
# image_token_len = multimodal_cfg['image_token_len']
image_token_len = cur_token_len
if not is_multimodal:
return sources
for source in sources:
if multimodal_cfg['sep_image_conv_front']:
assert DEFAULT_IMAGE_TOKEN in source[0]['value']
source[0]['value'] = source[0]['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip()
source[0]['value'] = DEFAULT_IMAGE_TOKEN + conversation_lib.default_conversation.sep + conversation_lib.default_conversation.roles[0] + ": " + source[0]['value']
for sentence in source:
replace_token = DEFAULT_IMAGE_PATCH_TOKEN * image_token_len
if multimodal_cfg['use_im_start_end']:
replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)
return sources
def preprocess_v1(
sources,
tokenizer: transformers.PreTrainedTokenizer,
) -> Dict:
conv = conversation_lib.default_conversation.copy()
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
# Apply prompt templates
conversations = []
for i, source in enumerate(sources):
if roles[source[0]["from"]] != conv.roles[0]:
# Skip the first one if it is not from human
source = source[1:]
conv.messages = []
for j, sentence in enumerate(source):
role = roles[sentence["from"]]
assert role == conv.roles[j % 2], f"{i}"
conv.append_message(role, sentence["value"])
conversations.append(conv.get_prompt())
# Tokenize conversations
input_ids = tokenizer(
conversations,
return_tensors="pt",
padding="longest",
max_length=tokenizer.model_max_length,
truncation=True,
).input_ids
targets = input_ids.clone()
assert conv.sep_style == conversation_lib.SeparatorStyle.TWO
# Mask targets
sep = conv.sep + conv.roles[1] + ": "
for conversation, target in zip(conversations, targets):
total_len = int(target.ne(tokenizer.pad_token_id).sum())
rounds = conversation.split(conv.sep2)
cur_len = 1
target[:cur_len] = IGNORE_INDEX
for i, rou in enumerate(rounds):
if rou == "":
break
parts = rou.split(sep)
if len(parts) != 2:
break
parts[0] += sep
round_len = len(tokenizer(rou).input_ids)
instruction_len = len(tokenizer(parts[0]).input_ids) - 2
target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
cur_len += round_len
target[cur_len:] = IGNORE_INDEX
if cur_len < tokenizer.model_max_length:
if cur_len != total_len:
target[:] = IGNORE_INDEX
print(
f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
f" (ignored)"
)
return dict(
input_ids=input_ids,
labels=targets,
)
def preprocess_mpt(
sources,
tokenizer: transformers.PreTrainedTokenizer,
) -> Dict:
conv = conversation_lib.default_conversation.copy()
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
# Apply prompt templates
conversations = []
for i, source in enumerate(sources):
if roles[source[0]["from"]] != conv.roles[0]:
# Skip the first one if it is not from human
source = source[1:]
conv.messages = []
for j, sentence in enumerate(source):
role = roles[sentence["from"]]
assert role == conv.roles[j % 2], f"{i}"
conv.append_message(role, sentence["value"])
conversations.append(conv.get_prompt())
# Tokenize conversations
input_ids = tokenizer(
conversations,
return_tensors="pt",
padding="longest",
max_length=tokenizer.model_max_length,
truncation=True,
).input_ids
targets = input_ids.clone()
assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
# Mask targets
sep = conv.sep + conv.roles[1]
for conversation, target in zip(conversations, targets):
total_len = int(target.ne(tokenizer.pad_token_id).sum())
rounds = conversation.split(conv.sep)
re_rounds = [conv.sep.join(rounds[:3])] # system + user + gpt
for conv_idx in range(3, len(rounds), 2):
re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx+2])) # user + gpt
cur_len = 0
target[:cur_len] = IGNORE_INDEX
for i, rou in enumerate(re_rounds):
if rou == "":
break
parts = rou.split(sep)
if len(parts) != 2:
break
parts[0] += sep
round_len = len(tokenizer(rou).input_ids) + len(tokenizer(conv.sep).input_ids)
instruction_len = len(tokenizer(parts[0]).input_ids)
target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
cur_len += round_len
target[cur_len:] = IGNORE_INDEX
if cur_len < tokenizer.model_max_length:
if cur_len != total_len:
target[:] = IGNORE_INDEX
print(
f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
f" (ignored)"
)
return dict(
input_ids=input_ids,
labels=targets,
)
def preprocess(
sources: Sequence[str],
tokenizer: transformers.PreTrainedTokenizer,
) -> Dict:
"""
Given a list of sources, each is a conversation list. This transform:
1. Add signal '### ' at the beginning each sentence, with end signal '\n';
2. Concatenate conversations together;
3. Tokenize the concatenated conversation;
4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
"""
if conversation_lib.default_conversation.version == "v1":
return preprocess_v1(sources, tokenizer)
if conversation_lib.default_conversation.version == "mpt":
return preprocess_mpt(sources, tokenizer)
# add end signal and concatenate together
conversations = []
for source in sources:
header = f"{conversation_lib.default_conversation.system}\n\n"
conversation = _add_speaker_and_signal(header, source)
conversations.append(conversation)
# tokenize conversations
conversations_tokenized = _tokenize_fn(conversations, tokenizer)
input_ids = conversations_tokenized["input_ids"]
targets = copy.deepcopy(input_ids)
for target, source in zip(targets, sources):
tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source],
tokenizer)["input_ids_lens"]
speakers = [sentence["from"] for sentence in source]
_mask_targets(target, tokenized_lens, speakers)
return dict(input_ids=input_ids, labels=targets)
class SupervisedDataset(Dataset):
"""Dataset for supervised fine-tuning."""
def __init__(self, data_path: str,
tokenizer: transformers.PreTrainedTokenizer):
super(SupervisedDataset, self).__init__()
logging.warning("Loading data...")
list_data_dict = json.load(open(data_path, "r"))
logging.warning("Formatting inputs...")
sources = [example["conversations"] for example in list_data_dict]
data_dict = preprocess(sources, tokenizer)
self.input_ids = data_dict["input_ids"]
self.labels = data_dict["labels"]
def __len__(self):
return len(self.input_ids)
def __getitem__(self, i) -> Dict[str, torch.Tensor]:
return dict(input_ids=self.input_ids[i], labels=self.labels[i])
class LazySupervisedDataset(Dataset):
def __init__(self, data_path: str,
tokenizer: transformers.PreTrainedTokenizer,
multimodal_cfg: dict):
super(LazySupervisedDataset, self).__init__()
self.tokenizer, self.multimodal_cfg = tokenizer, multimodal_cfg
self.pkl, self.prompt = pickle.load(open('./_data/ipr2pr.pkl', 'rb'))['task'], json.load(open('./_data/ipr2pr_expressive.json', 'r'))
random.shuffle(self.pkl)
print('--pkl: %d--'%(len(self.pkl)))
def __len__(self):
return len(self.pkl)
def __getitem__(self, i) -> Dict[str, torch.Tensor]:
item = self.pkl[i][0]
tsv = open('./_data/ipr2pr.tsv', 'r')
tsv.seek(item['lineidx'])
b = tsv.readline().strip().split('\t')
image = resize(b2f(b[0]))
processor = self.multimodal_cfg['image_processor']
image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
cur_token_len = (image.shape[1]//14)*(image.shape[2]//14)
query = "what will this image be like if '%s'\n%s"%(item['instruction'], DEFAULT_IMAGE_TOKEN)
ans = '%s [IMG0] [IMG1] [IMG2] [IMG3] [IMG4] [IMG5] [IMG6] [IMG7]'%(self.prompt[item['input']]['expressive'])
sources = preprocess_multimodal(copy.deepcopy([[{'from': 'human', 'value': query}, {'from': 'gpt', 'value': ans}]]),
self.multimodal_cfg, cur_token_len)
data_dict = preprocess(sources, self.tokenizer)
if isinstance(i, int): data_dict = dict(input_ids=data_dict['input_ids'][0],
labels=data_dict['labels'][0])
data_dict['image'] = image
p2p_inp, p2p_ans = img2npy(resize(b2f(b[0])).resize([256, 256])), img2npy(resize(b2f(b[1])).resize([256, 256]))
data_dict['p2p_inp'], data_dict['p2p_ans'] = p2p_inp, p2p_ans
return data_dict
@dataclass
class DataCollatorForSupervisedDataset(object):
"""Collate examples for supervised fine-tuning."""
tokenizer: transformers.PreTrainedTokenizer
def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
input_ids, labels = tuple([instance[key] for instance in instances]
for key in ("input_ids", "labels"))
input_ids = torch.nn.utils.rnn.pad_sequence(
input_ids,
batch_first=True,
padding_value=self.tokenizer.pad_token_id)
labels = torch.nn.utils.rnn.pad_sequence(labels,
batch_first=True,
padding_value=IGNORE_INDEX)
batch = dict(
input_ids=input_ids,
labels=labels,
attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
)
if 'image' in instances[0]:
images = [instance['image'] for instance in instances]
if all(x is not None and x.shape == images[0].shape for x in images):
batch['images'] = torch.stack(images)
else:
batch['images'] = images
batch['p2p_inp'], batch['p2p_ans'] = [torch.cat([torch.from_numpy(d['p2p_inp']).unsqueeze(dim=0) for d in instances], dim=0),
torch.cat([torch.from_numpy(d['p2p_ans']).unsqueeze(dim=0) for d in instances], dim=0)]
return batch
def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
data_args) -> Dict:
"""Make dataset and collator for supervised fine-tuning."""
dataset_cls = (LazySupervisedDataset
if data_args.lazy_preprocess else SupervisedDataset)
train_dataset = dataset_cls(tokenizer=tokenizer,
data_path=data_args.data_path,
multimodal_cfg=dict(
is_multimodal=data_args.is_multimodal,
sep_image_conv_front=data_args.sep_image_conv_front,
image_token_len=data_args.image_token_len,
image_folder=data_args.image_folder,
image_aspect_ratio=data_args.image_aspect_ratio,
use_im_start_end=getattr(data_args, 'mm_use_im_start_end', False),
image_processor=getattr(data_args, 'image_processor', None)))
data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
return dict(train_dataset=train_dataset,
eval_dataset=None,
data_collator=data_collator)
def train():
parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
bnb_model_from_pretrained_args = {}
if training_args.bits in [4, 8]:
from transformers import BitsAndBytesConfig
from peft import prepare_model_for_int8_training
bnb_model_from_pretrained_args.update(dict(
device_map={"": training_args.device},
load_in_4bit=training_args.bits == 4,
load_in_8bit=training_args.bits == 8,
quantization_config=BitsAndBytesConfig(
load_in_4bit=training_args.bits == 4,
load_in_8bit=training_args.bits == 8,
llm_int8_threshold=6.0,
llm_int8_has_fp16_weight=False,
bnb_4bit_compute_dtype=compute_dtype,
bnb_4bit_use_double_quant=training_args.double_quant,
bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'}
)
))
if model_args.vision_tower is not None:
if 'mpt' in model_args.model_name_or_path:
model = LlavaMPTForCausalLM.from_pretrained(
model_args.model_name_or_path,
cache_dir=training_args.cache_dir,
**bnb_model_from_pretrained_args
)
else:
model = LlavaLlamaForCausalLM.from_pretrained(
model_args.model_name_or_path,
cache_dir=training_args.cache_dir,
**bnb_model_from_pretrained_args
)
else:
model = transformers.LlamaForCausalLM.from_pretrained(
model_args.model_name_or_path,
cache_dir=training_args.cache_dir,
**bnb_model_from_pretrained_args
)
model.config.use_cache = False
if model_args.freeze_backbone:
model.model.requires_grad_(False)
if training_args.bits in [4, 8]:
model.config.torch_dtype=(torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
model = prepare_model_for_int8_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing)
if training_args.gradient_checkpointing and model_args.vision_tower is None:
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
if training_args.lora_enable:
from peft import LoraConfig, get_peft_model
lora_config = LoraConfig(
r=training_args.lora_r,
lora_alpha=training_args.lora_alpha,
target_modules=find_all_linear_names(model),
lora_dropout=training_args.lora_dropout,
bias=training_args.lora_bias,
task_type="CAUSAL_LM",
)
if training_args.bits == 16:
if training_args.bf16:
model.to(torch.bfloat16)
if training_args.fp16:
model.to(torch.float16)
logging.warning("Adding LoRA adapters...")
model = get_peft_model(model, lora_config)
if 'mpt' in model_args.model_name_or_path:
tokenizer = transformers.AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=training_args.cache_dir,
model_max_length=training_args.model_max_length,
padding_side="right"
)
else:
tokenizer = transformers.AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
cache_dir=training_args.cache_dir,
model_max_length=training_args.model_max_length,
padding_side="right",
use_fast=False,
)
if model_args.version == "v0":
if tokenizer.pad_token is None:
smart_tokenizer_and_embedding_resize(
special_tokens_dict=dict(pad_token=DEFAULT_PAD_TOKEN),
tokenizer=tokenizer,
model=model,
)
if "llama" in model_args.model_name_or_path:
tokenizer.add_special_tokens({
"eos_token": DEFAULT_EOS_TOKEN,
"bos_token": DEFAULT_BOS_TOKEN,
"unk_token": DEFAULT_UNK_TOKEN,
})
else:
tokenizer.pad_token = tokenizer.unk_token
if "mpt" in model_args.model_name_or_path:
conversation_lib.default_conversation = conversation_lib.conv_templates["mpt"]
else:
conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1_1"]
if model_args.vision_tower is not None:
model_vision_dict = model.get_model().initialize_vision_modules(
vision_tower=model_args.vision_tower,
mm_vision_select_layer=model_args.mm_vision_select_layer,
pretrain_mm_mlp_adapter=model_args.pretrain_mm_mlp_adapter,
fsdp=training_args.fsdp
)
model.get_vision_tower().to(dtype=torch.float16, device=training_args.device)
vision_config = model_vision_dict['vision_config']
data_args.image_token_len = model_vision_dict['image_token_len']
data_args.image_processor = model_vision_dict['image_processor']
data_args.is_multimodal = True
model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
if model_args.tune_mm_mlp_adapter:
model.requires_grad_(False)
for p in model.get_model().mm_projector.parameters():
p.requires_grad = True
model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
if training_args.freeze_mm_mlp_adapter:
for p in model.get_model().mm_projector.parameters():
p.requires_grad = False
if training_args.bits in [4, 8]:
model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
vision_config.use_im_start_end = training_args.use_im_start_end = model_args.mm_use_im_start_end
model.config.sep_image_conv_front = data_args.sep_image_conv_front
model.initialize_vision_tokenizer(mm_use_im_start_end=model_args.mm_use_im_start_end, tokenizer=tokenizer, device=training_args.device,
tune_mm_mlp_adapter=model_args.tune_mm_mlp_adapter, pretrain_mm_mlp_adapter=model_args.pretrain_mm_mlp_adapter)
params_no_grad = [n for n, p in model.named_parameters() if not p.requires_grad]
if len(params_no_grad) > 0:
if training_args.fsdp is not None and len(training_args.fsdp) > 0:
if len(params_no_grad) < 10:
print('[WARNING] Attempting to use FSDP while {} parameters do not require gradients: {}'. format(len(params_no_grad), params_no_grad))
else:
print('[WARNING] Attempting to use FSDP while {} parameters do not require gradients: {}...(omitted)'. format(len(params_no_grad), ', '.join(params_no_grad[:10])))
print("[WARNING] Attempting to use FSDP with partially frozen paramters, this is experimental.")
print("[WARNING] As of 4/30/23, this feature requires PyTorch-nightly build. See here for details: https://github.com/haotian-liu/LLaVA#experimental-use-fsdp-to-save-memory-in-pretraining")
from torch.distributed.fsdp.fully_sharded_data_parallel import FullyShardedDataParallel as FSDP
def patch_FSDP_use_orig_params(func):
def wrap_func(*args, **kwargs):
use_orig_params = kwargs.pop('use_orig_params', True)
return func(*args, **kwargs, use_orig_params=use_orig_params)
return wrap_func
FSDP.__init__ = patch_FSDP_use_orig_params(FSDP.__init__)
if training_args.bits in [4, 8]:
from peft.tuners.lora import LoraLayer
for name, module in model.named_modules():
if isinstance(module, LoraLayer):
if training_args.bf16:
module = module.to(torch.bfloat16)
if 'norm' in name:
module = module.to(torch.float32)
if 'lm_head' in name or 'embed_tokens' in name:
if hasattr(module, 'weight'):
if training_args.bf16 and module.weight.dtype == torch.float32:
module = module.to(torch.bfloat16)
# start for MGIE
os.makedirs('_log', exist_ok=True)
pt = {}
for i in tqdm(range(2)): pt.update(torch.load('./_ckpt/LLaVA-7B-v1/pytorch_model-0000%d-of-00002.bin'%(i+1), map_location='cpu'))
miss, unexp = model.load_state_dict(pt, strict=False)
print('miss:', miss), print('unexp:', unexp)
tokenizer.add_tokens(['[IMG0]', '[IMG1]', '[IMG2]', '[IMG3]', '[IMG4]', '[IMG5]', '[IMG6]', '[IMG7]'], special_tokens=True)
model.resize_token_embeddings(len(tokenizer))
print(tokenizer), json.dump(tokenizer.get_vocab(), open('_log/vocabs.json', 'w'), indent=2)
for n, p in model.named_parameters():
if 'embed_tokens' in n or 'lm_head' in n or 'edit_head' in n or 'unet' in n: p.requires_grad = True
else: p.requires_grad = False
with open('_log/parameters.txt', 'w') as F:
for n, p in model.named_parameters(): F.write('%s %s %s\n'%(n, str(p.shape), str(p.requires_grad)))
with open('_log/args_train.txt', 'w') as F:
for key in vars(training_args): F.write('%s: %s\n'%(str(key), str(vars(training_args)[key])))
# end for MGIE
data_module = make_supervised_data_module(tokenizer=tokenizer,
data_args=data_args)
trainer = LLaVATrainer(model=model,
tokenizer=tokenizer,
args=training_args,
**data_module)
if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
trainer.train(resume_from_checkpoint=True)
else:
trainer.train()
trainer.save_state()
if training_args.lora_enable:
state_dict = get_peft_state_maybe_zero_3(
model.named_parameters(), training_args.lora_bias
)
non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(
model.named_parameters()
)
if training_args.local_rank == 0 or training_args.local_rank == -1:
model.config.save_pretrained(training_args.output_dir)
model.save_pretrained(training_args.output_dir, state_dict=state_dict)
torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin'))
else:
safe_save_model_for_hf_trainer(trainer=trainer,
output_dir=training_args.output_dir)
if __name__ == "__main__":
train()