File size: 39,690 Bytes
e5e24c9 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 |
# Copyright (c) OpenMMLab. All rights reserved.
import math
import copy
import os.path as osp
import warnings
from collections import OrderedDict
import torch
import torch.nn as nn
import numpy as np
from accelerate import init_empty_weights
from mmengine import print_log
from mmengine.config import Config, ConfigDict
from mmengine.model import BaseModel
from peft import get_peft_model, prepare_model_for_kbit_training
from transformers import (AddedToken, AutoConfig, CLIPImageProcessor,
CLIPVisionModel, LlamaForCausalLM,
LlamaTokenizerFast, LlavaConfig,
LlavaForConditionalGeneration, LlavaProcessor
)
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
from transformers import PreTrainedModel, PretrainedConfig
from xtuner.registry import BUILDER
from xtuner.utils import DEFAULT_IMAGE_TOKEN
from .modules import ProjectorConfig, ProjectorModel, dispatch_modules
from .modules.dispatch import SUPPORT_FLASH1, SUPPORT_FLASH2
from .utils import (LoadWoInit, find_all_linear_names,
get_peft_model_state_dict, guess_load_checkpoint,
make_inputs_require_grad,
prepare_inputs_labels_for_multimodal, traverse_dict)
from .torchscale.model.LongNet import make_longnet_from_name
from transformers.models.qwen2.modeling_qwen2 import Qwen2DecoderLayer
import torch.nn.functional as F
from torch.nn.init import trunc_normal_
# ====================== Copied from first file ====================== #
def get_abs_pos(abs_pos, tgt_size):
"""
Interpolates 1D absolute positional embeddings to a target size.
This function is modified to handle 1D positional embeddings, which is
suitable for sequences of tokens that do not form a square grid.
Args:
abs_pos (torch.Tensor): The absolute positional embedding tensor of shape (N, C),
where N is the original sequence length and C is the embedding dim.
tgt_size (int): The target sequence length.
Returns:
torch.Tensor: The interpolated positional embedding tensor of shape (tgt_size, C).
"""
src_size = abs_pos.size(0)
dtype = abs_pos.dtype
if src_size == tgt_size:
return abs_pos
# For 1D interpolation, input tensor to F.interpolate should be (N, C, L)
# We reshape our (L, C) tensor to (1, C, L)
interp_input = abs_pos.float().permute(1, 0).unsqueeze(0)
# Perform linear interpolation
interp_output = F.interpolate(
interp_input,
size=tgt_size,
mode='linear',
align_corners=False,
)
# Reshape back to (L_new, C)
interpolated_pos = interp_output.squeeze(0).permute(1, 0).to(dtype)
return interpolated_pos
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
grid_h = np.arange(grid_size, dtype=np.float32)
grid_w = np.arange(grid_size, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h)
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size, grid_size])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])
emb = np.concatenate([emb_h, emb_w], axis=1)
return emb
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=np.float32)
omega /= embed_dim / 2.
omega = 1. / 10000**omega
pos = pos.reshape(-1)
out = np.einsum('m,d->md', pos, omega)
emb_sin = np.sin(out)
emb_cos = np.cos(out)
emb = np.concatenate([emb_sin, emb_cos], axis=1)
return emb
# Step 1: Create a configuration class for the Resampler
class ResamplerConfig(PretrainedConfig):
"""
Configuration class for the Resampler module.
"""
model_type = "resampler"
_auto_class = 'AutoConfig'
def __init__(
self,
grid_size,
embed_dim,
num_heads,
kv_dim=None,
norm_layer=nn.LayerNorm,
**kwargs
):
self.grid_size = grid_size
self.embed_dim = embed_dim
self.num_heads = num_heads
self.kv_dim = kv_dim
# self.hidden_act = hidden_act
self.norm_layer = norm_layer
super().__init__(**kwargs)
class Resampler(PreTrainedModel):
_auto_class = 'AutoModel'
config_class = ResamplerConfig
base_model_prefix = 'model'
supports_gradient_checkpointing = True
def __init__(
self,
config: ResamplerConfig
):
super().__init__(config)
self.gradient_checkpointing = False
self.num_queries = config.grid_size
self.embed_dim = config.embed_dim
self.num_heads = config.num_heads
kv_dim = config.kv_dim
norm_layer = config.norm_layer
# REMOVED: Positional embedding initialization
self.query = nn.Parameter(torch.zeros(self.num_queries, self.embed_dim))
self.query.data.normal_(mean=0.0, std=0.02)
if kv_dim is not None and kv_dim != self.embed_dim:
self.kv_proj = nn.Linear(kv_dim, self.embed_dim, bias=False)
else:
self.kv_proj = nn.Identity()
self.attn = nn.MultiheadAttention(self.embed_dim, self.num_heads, batch_first=True)
self.ln_q = norm_layer(self.embed_dim)
self.ln_kv = norm_layer(self.embed_dim)
nn.init.constant_(self.ln_q.bias, 0)
nn.init.constant_(self.ln_q.weight, 1.0)
nn.init.constant_(self.ln_kv.bias, 0)
nn.init.constant_(self.ln_kv.weight, 1.0)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.LayerNorm):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def init_weights(self):
self.query.data.normal_(mean=0.0, std=0.02)
nn.init.constant_(self.ln_q.bias, 0)
nn.init.constant_(self.ln_q.weight, 1.0)
nn.init.constant_(self.ln_kv.bias, 0)
nn.init.constant_(self.ln_kv.weight, 1.0)
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, Resampler):
module.gradient_checkpointing = value
def forward(self, x, attn_mask=None, text=None):
Q = self.query
x = self.kv_proj(x)
x = self.ln_kv(x)
Q = self.ln_q(Q)
# REMOVED: Positional embedding interpolation and addition
out, attn = self.attn(
Q.unsqueeze(0).expand(x.size(0), Q.size(0), Q.size(1)),
x,
x,
attn_mask=attn_mask
)
return out, attn
# ====================== End of copied code ====================== #
def convert_state_dict_to_hf(state_dict, mapping):
new_state_dict = {}
for key, value in state_dict.items():
if key.endswith('.inv_freq'):
continue
for key_to_modify, new_key in mapping.items():
if key_to_modify in key:
key = key.replace(key_to_modify, new_key)
new_state_dict[key] = value
return new_state_dict
class LLaVAModel(BaseModel):
def __init__(self,
llm,
freeze_llm=True,
visual_select_layer=-2,
pretrained_pth=None,
projector_depth=2,
llm_lora=None,
visual_encoder_lora=None,
use_activation_checkpointing=True,
max_position_embeddings=None,
hidden_size=512,
train_stage='2',
enable_long_net=True,
compressor_grid_size=2, # New parameter
compressor_embed_dim=512, # New parameter
prefusion_layer_num=4): # New parameter for prefusion layers
super().__init__()
self.enable_long_net = enable_long_net
if enable_long_net:
print('enable long net')
else:
print('disable long net')
self.freeze_llm = freeze_llm
self.freeze_visual_encoder = True
if train_stage == '1':
print('train_stage == 1')
self.freeze_llm = True
self.freeze_long_net = False
elif train_stage == '2':
print('train_stage == 2')
self.freeze_llm = True
self.freeze_long_net = True
with LoadWoInit():
if isinstance(llm, dict):
llm = self._dispatch_lm_model_cfg(llm, max_position_embeddings)
self.llm = self._build_from_cfg_or_module(llm)
self.encoder_name = "LongNet_{}_layers_{}_dim".format(2, 512)
self.LongNet_encoder = make_longnet_from_name(self.encoder_name)
self.llm.config.use_cache = False
dispatch_modules(self.llm)
self.projector_depth = projector_depth
self.compressor_grid_size = compressor_grid_size
self.compressor_embed_dim = compressor_embed_dim
self.prefusion_layer_num = prefusion_layer_num
# Build compressor
compressor_config = ResamplerConfig(
grid_size=compressor_grid_size,
embed_dim=compressor_embed_dim,
kv_dim= None,
num_heads= 8, # Default value, can be adjusted
)
self.compressor = Resampler(
config=compressor_config
)
self.compressor.init_weights()
projector_config = ProjectorConfig(
visual_hidden_size=hidden_size,
llm_hidden_size=self.llm.config.hidden_size,
depth=self.projector_depth)
self.projector = ProjectorModel(projector_config).to(
self.llm.dtype)
# Build prefusion layers
temps = copy.deepcopy(self.llm.model.layers[:prefusion_layer_num])
self.prefusion_layers = nn.ModuleList(temps)
del temps
# self.prefusion_layers=nn.ModuleList([Qwen2DecoderLayer(self.llm.config,layer_idx=i) for i in range(self.prefusion_layer_num)])
self.prefusion_layers.to(self.llm.dtype)
if self.freeze_llm:
print('freeze_llm')
self.llm.requires_grad_(False)
if self.freeze_long_net:
print('freeze_long_net')
self.LongNet_encoder.requires_grad_(False)
# Move to correct dtype and device
self.compressor = self.compressor.to(self.llm.dtype)
if use_activation_checkpointing:
if hasattr(self.llm, 'enable_input_require_grads'):
self.llm.enable_input_require_grads()
else:
self.llm.get_input_embeddings().register_forward_hook(
make_inputs_require_grad)
self.projector.enable_input_require_grads()
# for layer in self.prefusion_layers:
# if hasattr(layer, 'enable_input_require_grads'):
# layer.enable_input_require_grads()
# else:
# layer.get_input_embeddings().register_forward_hook(
# make_inputs_require_grad)
# enable gradient (activation) checkpointing for memory efficiency
self.gradient_checkpointing_enable()
self.use_llm_lora = None
self.use_visual_encoder_lora = None
if self.use_llm_lora:
self._prepare_llm_for_lora(llm_lora, use_activation_checkpointing)
if pretrained_pth is not None: # load the pretrained checkpoint
pretrained_state_dict = guess_load_checkpoint(pretrained_pth)
self.load_state_dict(pretrained_state_dict, strict=False)
print_log(f'Load pretrained weight from {pretrained_pth}', 'current')
self.visual_select_layer = visual_select_layer
self._is_init = True
self.is_first_iter = True
def _parse_lora_config(self, lora_config):
if isinstance(lora_config, dict) or isinstance(
lora_config, Config) or isinstance(lora_config, ConfigDict):
lora_config = BUILDER.build(lora_config)
return lora_config
def _prepare_llm_for_lora(self,
lora_config,
use_activation_checkpointing=True):
lora_config = self._parse_lora_config(lora_config)
self.llm = prepare_model_for_kbit_training(
self.llm, use_activation_checkpointing)
if lora_config.target_modules is None:
modules = find_all_linear_names(self.llm)
lora_config.target_modules = modules
self.llm = get_peft_model(self.llm, lora_config)
def _prepare_visual_encoder_for_lora(self,
lora_config,
use_activation_checkpointing=True):
lora_config = self._parse_lora_config(lora_config)
if lora_config.target_modules is None:
modules = find_all_linear_names(self.visual_encoder)
lora_config.target_modules = modules
self.visual_encoder = get_peft_model(self.visual_encoder, lora_config)
def gradient_checkpointing_enable(self):
self.activation_checkpointing_enable()
def activation_checkpointing_enable(self):
self.llm.gradient_checkpointing_enable()
self.compressor.gradient_checkpointing_enable()
self.projector.gradient_checkpointing_enable()
# for layer in self.prefusion_layers:
# layer.gradient_checkpointing_enable()
def gradient_checkpointing_disable(self):
self.activation_checkpointing_disable()
def activation_checkpointing_disable(self):
self.llm.gradient_checkpointing_disable()
self.compressor.gradient_checkpointing_disable()
self.projector.gradient_checkpointing_disable()
# for layer in self.prefusion_layers:
# layer.gradient_checkpointing_disable()
def init_weights(self):
pass
def state_dict(self, *args, **kwargs):
state_dict = super().state_dict(*args, **kwargs)
to_return = OrderedDict()
# Step 1. visual_encoder
if self.use_visual_encoder_lora:
to_return.update(
get_peft_model_state_dict(
self.visual_encoder, state_dict=state_dict))
elif not self.freeze_visual_encoder:
to_return.update({
k: v
for k, v in state_dict.items() if 'visual_encoder.' in k
})
# Step 2. LLM
if self.use_llm_lora:
to_return.update(
get_peft_model_state_dict(self.llm, state_dict=state_dict))
elif not self.freeze_llm:
to_return.update(
{k: v
for k, v in state_dict.items() if 'llm.' in k})
# Step 3. Compressor and Projector
to_return.update(
{k: v
for k, v in state_dict.items() if 'compressor.' in k})
to_return.update(
{k: v
for k, v in state_dict.items() if 'projector.' in k})
# Step 4. Prefusion layers
to_return.update(
{k: v
for k, v in state_dict.items() if 'prefusion_layers.' in k})
# Step 5. LongNet_encoder
to_return.update(
{k: v
for k, v in state_dict.items() if 'LongNet_encoder.' in k})
return to_return
@staticmethod
def _prepare_for_long_context_training(cfg, llm_cfg,
max_position_embeddings):
orig_rope_scaling = getattr(llm_cfg, 'rope_scaling', None)
if orig_rope_scaling is None:
orig_rope_scaling = {'factor': 1}
orig_rope_scaling_factor = orig_rope_scaling[
'factor'] if 'factor' in orig_rope_scaling.keys() else 1
orig_ctx_len = getattr(llm_cfg, 'max_position_embeddings', None)
if orig_ctx_len:
orig_ctx_len *= orig_rope_scaling_factor
if max_position_embeddings > orig_ctx_len:
scaling_factor = float(
math.ceil(max_position_embeddings / orig_ctx_len))
llm_cfg.rope_scaling = {
'type': 'linear',
'factor': scaling_factor
}
# hardcode for internlm2
llm_cfg.attn_implementation = 'flash_attention_2'
cfg.config = llm_cfg
return cfg, llm_cfg
@staticmethod
def _prepare_for_flash_attn(cfg, llm_cfg):
cls_name = type(llm_cfg).__name__
SUPPORT_SDPA_ATTN = ('LlamaConfig', 'GemmaConfig', 'MistralConfig',
'MixtralConfig', 'Qwen2Config', 'Qwen2MoeConfig',
'Starcoder2Config', 'Starcoder2Config',
'Phi3Config')
SUPPORT_FLASH_ATTN2 = ('InternLM2Config', 'LlamaConfig', 'GemmaConfig',
'MistralConfig', 'MixtralConfig', 'Qwen2Config',
'Qwen2MoeConfig', 'Starcoder2Config',
'Starcoder2Config', 'Phi3Config')
torch_dtype = torch.bfloat16 if (
torch.cuda.is_available() and torch.cuda.is_bf16_supported()) \
else torch.float16
if getattr(cfg, 'attn_implementation', None) is not None:
# Flash Attention 2.0 only supports torch.float16 and
# torch.bfloat16 dtypes
if cfg.attn_implementation == 'flash_attention_2':
cfg.torch_dtype = torch_dtype
elif SUPPORT_FLASH2 and cls_name in SUPPORT_FLASH_ATTN2:
cfg.torch_dtype = torch_dtype
cfg.attn_implementation = 'flash_attention_2'
elif SUPPORT_FLASH1 and cls_name in SUPPORT_SDPA_ATTN:
cfg.attn_implementation = 'sdpa'
return cfg, llm_cfg
@staticmethod
def _prepare_for_qlora_zero3(cfg):
if (not is_deepspeed_zero3_enabled()) or (not hasattr(
cfg, 'quantization_config')):
return cfg
torch_dtype = torch.bfloat16 if (
torch.cuda.is_available() and torch.cuda.is_bf16_supported()) \
else torch.float16
cfg.torch_dtype = torch_dtype
quantization_config = cfg.quantization_config
quantization_config.bnb_4bit_compute_dtype = torch_dtype
quantization_config.bnb_4bit_quant_storage = torch_dtype
return cfg
def _dispatch_lm_model_cfg(self, cfg, max_position_embeddings=None):
cfg = self._prepare_for_qlora_zero3(cfg)
pretrained_model_name_or_path = cfg.pretrained_model_name_or_path
llm_cfg = AutoConfig.from_pretrained(
pretrained_model_name_or_path, trust_remote_code=True)
cfg, llm_cfg = self._prepare_for_flash_attn(cfg, llm_cfg)
if max_position_embeddings is not None:
cfg, llm_cfg = self._prepare_for_long_context_training(
cfg, llm_cfg, max_position_embeddings)
return cfg
def _build_from_cfg_or_module(self, cfg_or_mod):
if isinstance(cfg_or_mod, nn.Module):
return cfg_or_mod
elif isinstance(cfg_or_mod, dict):
traverse_dict(cfg_or_mod)
return BUILDER.build(cfg_or_mod)
else:
raise NotImplementedError
def forward(self, data, data_samples=None, mode='loss'):
if self.is_first_iter:
self.to(data['input_ids'].device)
self.is_first_iter = False
if 'pixel_values' in data:
feat_to_proj = data['pixel_values'].to(self.llm.dtype)
if self.enable_long_net:
long_net_output = self.LongNet_encoder(src_tokens=None, token_embeddings=feat_to_proj.permute(1, 0, 2))["encoder_out"]
feat_to_proj = long_net_output.permute(1, 0, 2)
# Apply compressor
compressed_features, _ = self.compressor(feat_to_proj)
# print_log('compressed_features shape: {}'.format(compressed_features.shape), 'current')
# Apply projector
pixel_values = self.projector(compressed_features)
projected_global_image_features = self.projector(feat_to_proj)
# Apply prefusion layers if any
if self.prefusion_layer_num > 0:
# print_log('Applying prefusion layers', 'current')
input_ids = data['input_ids']
text_embeddings = self.llm.get_input_embeddings()(input_ids.clamp(min=0)).detach()
padding_mask=(input_ids <= 0)
# attention_mask = data['attention_mask']
# position_ids = data.get('position_ids', None)
x = torch.cat([projected_global_image_features, pixel_values, text_embeddings], dim=1)
mask=torch.cat((torch.zeros((padding_mask.size(0),projected_global_image_features.size(1)+pixel_values.size(1)),device=padding_mask.device).bool(),padding_mask),dim=1)
# Prepare attention mask for prefusion layers
if getattr(self.llm, "_use_flash_attention_2", False) or \
getattr(self.llm.config, "_attn_implementation", "") == "flash_attention_2":
attention_mask = (~mask).int()
else:
attention_mask =_prepare_4d_causal_attention_mask(~mask, (x.size(0), x.size(1)), x, 0)
position_ids = (~mask).int().long().cumsum(-1) - 1
position_ids.masked_fill_((~mask).int() == 0, 1)
# Apply prefusion layers
for layer in self.prefusion_layers:
x = layer(
x,
attention_mask=attention_mask,
position_ids=position_ids,
use_cache=False,
)[0]
# data['inputs_embeds'] = x
fusion_text_features=x[:,-1*input_ids.size(1):,:]
pixel_values=x[:,-1*input_ids.size(1)-1*pixel_values.size(1):-1*input_ids.size(1),:]
fusion_text_features=fusion_text_features*(~padding_mask).unsqueeze(-1).int()+ text_embeddings*padding_mask.unsqueeze(-1)
data['text_features'] = fusion_text_features
# print_log('text_features shape: {}'.format(fusion_text_features.shape), 'current')
data['pixel_values'] = pixel_values
data = prepare_inputs_labels_for_multimodal(llm=self.llm, **data)
if mode == 'loss':
return self.compute_loss(data, data_samples)
elif mode == 'predict':
return self.predict(data, data_samples)
elif mode == 'tensor':
return self._forward(data, data_samples)
else:
raise NotImplementedError
def _forward(self, data, data_samples=None):
outputs = self.llm(**data)
return outputs
def predict(self, data, data_samples=None):
outputs = self.llm(**data)
logits_dict = [{'logits': logits} for logits in outputs.logits]
return logits_dict
def compute_loss(self, data, data_samples=None):
outputs = self.llm(**data)
loss_dict = {'loss': outputs.loss}
return loss_dict
def __getattr__(self, name: str):
try:
return super().__getattr__(name)
except AttributeError:
return getattr(self.llm, name)
def to_hf(self,
cfg,
save_dir,
fp32=False,
save_pretrained_kwargs={},
save_format='xtuner',
**kwargs):
if save_format == 'xtuner':
self.to_xtuner_llava(cfg, save_dir, fp32, save_pretrained_kwargs)
elif save_format == 'huggingface':
self.to_huggingface_llava(cfg, save_dir, fp32,
save_pretrained_kwargs)
elif save_format == 'official':
self.to_official_llava(cfg, save_dir, fp32, save_pretrained_kwargs)
else:
raise NotImplementedError
def to_xtuner_llava(self,
cfg,
save_dir,
fp32=False,
save_pretrained_kwargs={}):
self.llm.config.use_cache = True
if not fp32:
print_log('Convert LLM to float16', 'current')
self.llm.half()
if self.use_llm_lora:
llm_path = osp.join(save_dir, 'llm_adapter')
print_log(f'Saving LLM adapter to {llm_path}', 'current')
self.llm.save_pretrained(llm_path, **save_pretrained_kwargs)
elif not self.freeze_llm:
llm_path = save_dir
print_log(f'Saving LLM tokenizer to {llm_path}', 'current')
tokenizer = BUILDER.build(cfg.tokenizer)
tokenizer.save_pretrained(llm_path, **save_pretrained_kwargs)
print_log(f'Saving LLM to {llm_path}', 'current')
self.llm.save_pretrained(llm_path, **save_pretrained_kwargs)
self.llm.config.use_cache = False
# Visual Encoder
if self.use_visual_encoder_lora:
visual_encoder_path = osp.join(save_dir, 'visual_encoder_adapter')
print_log(
f'Saving visual_encoder adapter to {visual_encoder_path}',
'current')
self.visual_encoder.save_pretrained(visual_encoder_path,
**save_pretrained_kwargs)
elif not self.freeze_visual_encoder:
visual_encoder_path = osp.join(save_dir, 'visual_encoder')
print_log(
'Saving visual_encoder image_processor to'
f'{visual_encoder_path}', 'current')
image_processor = BUILDER.build(cfg.image_processor)
image_processor.save_pretrained(visual_encoder_path,
**save_pretrained_kwargs)
print_log(f'Saving visual_encoder to {visual_encoder_path}',
'current')
self.visual_encoder.save_pretrained(visual_encoder_path,
**save_pretrained_kwargs)
# projector
projector_path = osp.join(save_dir, 'projector')
print_log(f'Saving projector to {projector_path}', 'current')
self.projector.save_pretrained(projector_path,
**save_pretrained_kwargs)
# compressor
compressor_path = osp.join(save_dir, 'compressor')
print_log(f'Saving compressor to {compressor_path}', 'current')
self.compressor.save_pretrained(compressor_path,
**save_pretrained_kwargs)
# Prefusion layers
if self.prefusion_layer_num > 0:
prefusion_path = osp.join(save_dir, 'prefusion_layers')
print_log(f'Saving prefusion layers to {prefusion_path}', 'current')
torch.save(self.prefusion_layers.state_dict(),
osp.join(prefusion_path, 'prefusion_layers.bin'))
# LongNet_encoder
LongNet_encoder_path = osp.join(save_dir, 'LongNet_encoder')
print_log(f'Saving LongNet_encoder to {LongNet_encoder_path}', 'current')
self.LongNet_encoder.save_pretrained(LongNet_encoder_path,
**save_pretrained_kwargs)
def to_huggingface_llava(self,
cfg,
save_dir,
fp32=False,
save_pretrained_kwargs={}):
LLM_MAPPING = {
'model': 'language_model.model',
'lm_head': 'language_model.lm_head',
}
VIT_MAPPING = {
'vision_model': 'vision_tower.vision_model',
}
PROJECTOR_MAPPING = {
'model.0': 'multi_modal_projector.linear_1',
'model.2': 'multi_modal_projector.linear_2',
}
COMPRESSOR_MAPPING = {
'compressor': 'compressor'
}
PREFUSION_MAPPING = {
'prefusion_layers': 'prefusion_layers'
}
LONGNET_MAPPING = {
'layers.0': 'LongNet_encoder.layers.0',
'layers.1': 'LongNet_encoder.layers.1',
'layer_norm': 'LongNet_encoder.layer_norm'
}
assert getattr(self.llm, 'hf_quantizer', None) is None, \
'This conversion format does not support quantized LLM.'
# get state_dict
llm = self.llm
if self.use_llm_lora:
llm = self.llm.merge_and_unload()
llm.config.use_cache = True
if not fp32:
print_log('Convert LLM to float16', 'current')
llm.half()
assert isinstance(llm, LlamaForCausalLM), \
'This conversion format only supports LlamaForCausalLM.'
llm_state_dict = llm.state_dict()
llm_state_dict = convert_state_dict_to_hf(llm_state_dict, LLM_MAPPING)
need_visual_encoder = (not self.freeze_visual_encoder
or self.use_visual_encoder_lora)
visual_encoder = self.visual_encoder
if self.use_visual_encoder_lora:
visual_encoder = self.visual_encoder.merge_and_unload()
assert isinstance(visual_encoder, CLIPVisionModel),\
'This conversion format only supports CLIPVisionModel.'
if need_visual_encoder:
visual_encoder_state_dict = visual_encoder.state_dict()
visual_encoder_state_dict = convert_state_dict_to_hf(
visual_encoder_state_dict, VIT_MAPPING)
else:
visual_encoder_state_dict = {}
compressor_state_dict = self.compressor.state_dict()
compressor_state_dict = convert_state_dict_to_hf(
compressor_state_dict, COMPRESSOR_MAPPING)
projector_state_dict = self.projector.state_dict()
projector_state_dict = convert_state_dict_to_hf(
projector_state_dict, PROJECTOR_MAPPING)
prefusion_state_dict = self.prefusion_layers.state_dict()
prefusion_state_dict = convert_state_dict_to_hf(
prefusion_state_dict, PREFUSION_MAPPING)
LongNet_encoder_state_dict = self.LongNet_encoder.state_dict()
LongNet_encoder_state_dict = convert_state_dict_to_hf(
LongNet_encoder_state_dict, LONGNET_MAPPING)
state_dict = {
**projector_state_dict,
**compressor_state_dict,
**prefusion_state_dict,
**llm_state_dict,
**visual_encoder_state_dict,
**LongNet_encoder_state_dict
}
# init model
text_config = llm.config
vision_config = visual_encoder.config
config = LlavaConfig(
text_config=text_config,
vision_config=vision_config,
attn_implementation='eager')
with init_empty_weights():
with warnings.catch_warnings():
warnings.filterwarnings(
'ignore', message='.*non-meta.*', category=UserWarning)
model = LlavaForConditionalGeneration(config)
model.load_state_dict(state_dict, strict=True, assign=True)
# processor
cfg.tokenizer.type = LlamaTokenizerFast.from_pretrained
tokenizer = BUILDER.build(cfg.tokenizer)
tokenizer.add_tokens(
AddedToken(DEFAULT_IMAGE_TOKEN, special=True, normalized=False),
special_tokens=True)
tokenizer.add_special_tokens({'pad_token': '<pad>'})
image_processor = BUILDER.build(cfg.image_processor)
assert isinstance(image_processor, CLIPImageProcessor),\
'This conversion format only supports CLIPImageProcessor.'
processor = LlavaProcessor(
tokenizer=tokenizer, image_processor=image_processor)
# Pad to 64 for performance reasons
pad_shape = 64
pre_expansion_embeddings = \
model.language_model.model.embed_tokens.weight.data
mu = torch.mean(pre_expansion_embeddings, dim=0).float()
n = pre_expansion_embeddings.size()[0]
sigma = ((pre_expansion_embeddings - mu).T
@ (pre_expansion_embeddings - mu)) / n
dist = torch.distributions.multivariate_normal.MultivariateNormal(
mu, covariance_matrix=1e-5 * sigma)
# We add an image token so we need to resize the model
ori_vocab_size = config.text_config.vocab_size
tokenizer_vocab_size = tokenizer.encode('<pad>')[-1]
added_token = tokenizer_vocab_size - ori_vocab_size
if added_token > 0:
model.resize_token_embeddings(ori_vocab_size + added_token,
pad_shape)
model.language_model.model.embed_tokens.weight.data[
ori_vocab_size:] = torch.stack(
tuple(
dist.sample()
for _ in range(model.language_model.model.embed_tokens.
weight.data[ori_vocab_size:].shape[0])),
dim=0,
)
model.language_model.lm_head.weight.data[
ori_vocab_size:] = torch.stack(
tuple(dist.sample()
for _ in range(model.language_model.lm_head.weight.
data[ori_vocab_size:].shape[0])),
dim=0,
)
model.config.image_token_index = tokenizer.encode(
DEFAULT_IMAGE_TOKEN)[-1]
model.config.pad_token_id = tokenizer.encode('<pad>')[-1]
# save
print_log(f'Saving to {save_dir}', 'current')
model.save_pretrained(save_dir, **save_pretrained_kwargs)
processor.save_pretrained(save_dir, **save_pretrained_kwargs)
def to_official_llava(self,
cfg,
save_dir,
fp32=False,
save_pretrained_kwargs={}):
VIT_MAPPING = {
'vision_model': 'model.vision_tower.vision_tower.vision_model',
}
PROJECTOR_MAPPING = {
'model.0': 'model.mm_projector.0',
'model.2': 'model.mm_projector.2',
}
COMPRESSOR_MAPPING = {
'compressor': 'model.compressor'
}
PREFUSION_MAPPING = {
'prefusion_layers': 'model.prefusion_layers'
}
LONGNET_MAPPING = {
'layers.0': 'LongNet_encoder.layers.0',
'layers.1': 'LongNet_encoder.layers.1',
'layer_norm': 'LongNet_encoder.layer_norm'
}
try:
from llava.model import LlavaConfig, LlavaLlamaForCausalLM
except ImportError:
raise ImportError(
'Please install llava with '
'`pip install git+https://github.com/haotian-liu/LLaVA.git '
'--no-deps`.')
assert getattr(self.llm, 'hf_quantizer', None) is None, \
'This conversion format does not support quantized LLM.'
# get state_dict
llm = self.llm
if self.use_llm_lora:
llm = self.llm.merge_and_unload()
llm.config.use_cache = True
if not fp32:
print_log('Convert LLM to float16', 'current')
llm.half()
assert isinstance(llm, LlamaForCausalLM), \
'This conversion format only supports LlamaForCausalLM.'
llm_state_dict = llm.state_dict()
need_visual_encoder = (not self.freeze_visual_encoder
or self.use_visual_encoder_lora)
visual_encoder = self.visual_encoder
if self.use_visual_encoder_lora:
visual_encoder = self.visual_encoder.merge_and_unload()
assert isinstance(visual_encoder, CLIPVisionModel),\
'This conversion format only supports CLIPVisionModel.'
if need_visual_encoder:
visual_encoder_state_dict = visual_encoder.state_dict()
visual_encoder_state_dict = convert_state_dict_to_hf(
visual_encoder_state_dict, VIT_MAPPING)
else:
visual_encoder_state_dict = {}
compressor_state_dict = self.compressor.state_dict()
compressor_state_dict = convert_state_dict_to_hf(
compressor_state_dict, COMPRESSOR_MAPPING)
projector_state_dict = self.projector.state_dict()
projector_state_dict = convert_state_dict_to_hf(
projector_state_dict, PROJECTOR_MAPPING)
prefusion_state_dict = self.prefusion_layers.state_dict()
prefusion_state_dict = convert_state_dict_to_hf(
prefusion_state_dict, PREFUSION_MAPPING)
LongNet_encoder_state_dict = self.LongNet_encoder.state_dict()
LongNet_encoder_state_dict = convert_state_dict_to_hf(
LongNet_encoder_state_dict, LONGNET_MAPPING)
state_dict = {
**projector_state_dict,
**compressor_state_dict,
**prefusion_state_dict,
**llm_state_dict,
**visual_encoder_state_dict,
**LongNet_encoder_state_dict
}
# init model
tokenizer = BUILDER.build(cfg.tokenizer)
image_processor = BUILDER.build(cfg.image_processor)
assert isinstance(image_processor, CLIPImageProcessor),\
'This conversion format only supports CLIPImageProcessor.'
llava_config_dict = llm.config.__dict__.copy()
llava_config_dict.update(
dict(
image_aspect_ratio='pad',
mm_hidden_size=visual_encoder.config.hidden_size,
mm_projector_type=f'mlp{self.projector_depth}x_gelu',
mm_use_im_patch_token=False,
mm_use_im_start_end=False,
mm_vision_select_feature='patch',
mm_vision_select_layer=self.visual_select_layer,
mm_vision_tower=visual_encoder.config.name_or_path,
unfreeze_mm_vision_tower=need_visual_encoder,
model_type='llava',
use_cache=True,
use_mm_proj=True))
llava_config = LlavaConfig(**llava_config_dict)
with init_empty_weights():
with warnings.catch_warnings():
warnings.filterwarnings(
'ignore', message='.*non-meta.*', category=UserWarning)
model = LlavaLlamaForCausalLM(llava_config)
model.load_state_dict(state_dict, strict=True, assign=True)
# save
print_log(f'Saving to {save_dir}', 'current')
model.save_pretrained(save_dir, **save_pretrained_kwargs)
image_processor.save_pretrained(save_dir, **save_pretrained_kwargs)
tokenizer.save_pretrained(save_dir, **save_pretrained_kwargs) |