File size: 11,511 Bytes
4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b 4e3b08f e043d8b |
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 |
from typing import Optional, Union
from transformers import PreTrainedModel, PretrainedConfig, GenerationMixin, Cache, BatchFeature
from modelscope import AutoConfig, AutoProcessor, AutoModel, AutoTokenizer, AutoModelForCausalLM
import torch
import torch.nn as nn
from transformers.image_utils import ImageInput
from transformers.modeling_outputs import CausalLMOutputWithPast
from liger_kernel.transformers import LigerCrossEntropyLoss
from transformers.processing_utils import Unpack, ProcessorMixin
from transformers.tokenization_utils_base import TextInput, PreTokenizedInput
from transformers.utils import TransformersKwargs
class Qwenov3Config(PretrainedConfig):
model_type = "Qwenov3"
def __init__(self, llm_model_path='Qwen/Qwen3-0.6B',
vision_model_path='facebook/dinov3-vitl16-pretrain-lvd1689m',
freeze_vision_model=False,
freeze_llm_model=False,
image_pad_num=49,
training_scratch=False,
num_hidden_layers=None,
hidden_size=None,
num_attention_heads=None,
vocab_size=None,
**kwargs):
self.vision_model_path = vision_model_path
self.llm_model_path = llm_model_path
self.freeze_vision_model = freeze_vision_model
self.freeze_llm_model = freeze_llm_model
self.image_pad_num = image_pad_num
self.freeze_vision_model = freeze_vision_model
self.training_scratch = training_scratch
self.num_hidden_layers = num_hidden_layers
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.vocab_size = vocab_size
super().__init__(**kwargs)
class Qwenov3Processor(ProcessorMixin):
attributes = ["image_processor", "tokenizer"]
image_processor_class = "AutoImageProcessor"
tokenizer_class = "AutoTokenizer"
def __init__(self, image_processor=None, tokenizer=None, chat_template=None, image_pad_num=49, **kwargs):
self.image_token = "<|image_pad|>"
self.image_pad_num = image_pad_num
if chat_template is None and tokenizer is not None:
chat_template = getattr(tokenizer, "chat_template", None)
super().__init__(image_processor, tokenizer, chat_template=chat_template)
def __call__(
self,
images: Optional[ImageInput] = None,
text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
return_tensors: str = "pt",
**kwargs,
) -> BatchFeature:
image_inputs = {}
if images is not None:
image_inputs = {'pixel_values': self.image_processor(images=images, return_tensors="pt")['pixel_values']}
if not isinstance(text, list):
text = [text]
processed_text = []
for t in text:
replacement = '<|vision_start|>' + '<|image_pad|>' * self.image_pad_num + '<|vision_end|>'
if '<image>' not in t:
t = t.replace('<|im_end|>', '<image><|im_end|>', 1)
processed_text.append(t.replace('<image>', replacement))
tokenizer_kwargs = {k: v for k, v in kwargs.items() if k not in ['images']}
text_inputs = self.tokenizer(processed_text, return_tensors=return_tensors, **tokenizer_kwargs)
return BatchFeature(data={**text_inputs, **image_inputs})
class Qwenov3(GenerationMixin, PreTrainedModel):
config_class = Qwenov3Config
base_model_prefix = "model"
supports_gradient_checkpointing = True
_no_split_modules = ["MoeDecoderLayer"]
_skip_keys_device_placement = ["past_key_values"]
_supports_sdpa = True
_supports_flash_attn = True
_can_compile_fullgraph = False
_supports_attention_backend = True
def __init__(self, config):
super().__init__(config)
self.config = config
if self.config.training_scratch:
self.vision_model = AutoModel.from_pretrained(self.config.vision_model_path, low_cpu_mem_usage=True,
dtype=torch.bfloat16, attn_implementation="flash_attention_2")
self.llm_model = AutoModelForCausalLM.from_pretrained(self.config.llm_model_path, low_cpu_mem_usage=True,
dtype=torch.bfloat16,
attn_implementation="flash_attention_2")
else:
vision_config = AutoConfig.from_pretrained(self.config.vision_model_path)
self.vision_model = AutoModel.from_config(vision_config, attn_implementation="sdpa", dtype=torch.bfloat16)
llm_config = AutoConfig.from_pretrained(self.config.llm_model_path)
self.llm_model = AutoModelForCausalLM.from_config(llm_config, attn_implementation="sdpa", dtype=torch.bfloat16)
if self.config.num_hidden_layers is None:
self.config.num_hidden_layers = self.llm_model.config.num_hidden_layers
if self.config.hidden_size is None:
self.config.hidden_size = self.llm_model.config.hidden_size
if self.config.num_attention_heads is None:
self.config.num_attention_heads = self.llm_model.config.num_attention_heads
if self.config.vocab_size is None:
self.config.vocab_size = self.llm_model.config.vocab_size
self.processor = AutoProcessor.from_pretrained(self.config.vision_model_path)
self.tokenizer = AutoTokenizer.from_pretrained(self.config.llm_model_path, use_fast=True)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
if '<|image_pad|>' not in self.tokenizer.get_vocab():
self.tokenizer.add_tokens(['<|image_pad|>'])
self.llm_model.resize_token_embeddings(len(self.tokenizer), mean_resizing=True)
if '<|vision_start|>' not in self.tokenizer.get_vocab():
self.tokenizer.add_tokens(['<|vision_start|>'])
self.llm_model.resize_token_embeddings(len(self.tokenizer), mean_resizing=True)
if '<|vision_end|>' not in self.tokenizer.get_vocab():
self.tokenizer.add_tokens(['<|vision_end|>'])
self.llm_model.resize_token_embeddings(len(self.tokenizer), mean_resizing=True)
self.adapter = nn.Sequential(
nn.RMSNorm(4096, dtype=torch.bfloat16),
nn.Linear(4096, self.llm_model.config.hidden_size, dtype=torch.bfloat16),
nn.GELU(),
nn.Linear(self.llm_model.config.hidden_size, self.llm_model.config.hidden_size, dtype=torch.bfloat16)
)
if self.config.freeze_vision_model:
for param in self.vision_model.parameters():
param.requires_grad = False
if self.config.freeze_llm_model:
for param in self.llm_model.parameters():
param.requires_grad = False
def forward(
self,
input_ids: Optional[torch.LongTensor] = None,
pixel_values: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Cache] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
**kwargs: Unpack[TransformersKwargs],
):
text_embeds = self.llm_model.get_input_embeddings()(input_ids)
if pixel_values is not None:
image_embeds = self.vision_model(pixel_values).last_hidden_state
patch_embeds = image_embeds[:, 5:, :] # [batch, 196, 1024]
b, num_patches, hidden_dim = patch_embeds.shape
patch_embeds = patch_embeds.view(b, num_patches // 4, hidden_dim * 4) # [batch, 49, 4096]
image_features = self.adapter(patch_embeds)
text_embeds = text_embeds.to(image_features.dtype)
inputs_embeds = self.merge_input_ids_with_image_features(image_features, text_embeds, input_ids)
else:
inputs_embeds = text_embeds
outputs = self.llm_model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
cache_position=cache_position,
**kwargs,
)
hidden_states = outputs.last_hidden_state
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
logits = self.lm_head(hidden_states[:, slice_indices, :])
loss = None
if labels is not None:
loss_fct = LigerCrossEntropyLoss(ignore_index=self.tokenizer.pad_token_id)
loss = loss_fct(logits.view(-1, logits.size(-1)), labels.view(-1).to(logits.device))
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
@torch.inference_mode()
def generate(self, input_ids=None, pixel_values=None, attention_mask=None,
max_new_tokens=512, temperature=0.7, top_p=0.8, top_k=20,
do_sample=True, num_beams=1, use_cache=True, **kwargs):
if pixel_values is not None:
text_embeds = self.llm_model.get_input_embeddings()(input_ids)
image_embeds = self.vision_model(pixel_values).last_hidden_state
patch_embeds = image_embeds[:, 5:, :]
b, num_patches, hidden_dim = patch_embeds.shape
patch_embeds = patch_embeds.view(b, num_patches // 4, hidden_dim * 4)
image_features = self.adapter(patch_embeds)
text_embeds = text_embeds.to(image_features.dtype)
inputs_embeds = self.merge_input_ids_with_image_features(image_features, text_embeds, input_ids)
else:
inputs_embeds = self.llm_model.get_input_embeddings()(input_ids)
return self.llm_model.generate(
input_ids=input_ids,
inputs_embeds=inputs_embeds,
attention_mask=attention_mask,
max_new_tokens=max_new_tokens,
temperature=temperature,
top_p=top_p,
top_k=top_k,
do_sample=do_sample,
num_beams=num_beams,
use_cache=use_cache,
pad_token_id=self.tokenizer.pad_token_id,
eos_token_id=self.tokenizer.eos_token_id,
**kwargs
)
def can_generate(self):
return True
def merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids):
num_images, num_image_patches, embed_dim = image_features.shape
batch_indices, image_indices = torch.where(input_ids == self.tokenizer('<|image_pad|>')['input_ids'][0])
if len(batch_indices) == 0:
return inputs_embeds
inputs_embeds[batch_indices, image_indices] = image_features.view(-1, embed_dim)
return inputs_embeds
|