import os import random from abc import ABC, abstractmethod import torch import torch.nn as nn from blip3o.constants import ( DEFAULT_IM_END_TOKEN, DEFAULT_IM_START_TOKEN, IGNORE_INDEX, IMAGE_TOKEN_INDEX, ) from blip3o.utils import rank0_print from .multimodal_encoder.builder import build_vision_tower from .multimodal_decoder.builder import build_sana, build_vae from diffusers.models.normalization import RMSNorm from diffusers import AutoencoderDC, FlowMatchEulerDiscreteScheduler, SanaTransformer2DModel import math class blip3oMetaModel: def __init__(self, config): super(blip3oMetaModel, self).__init__(config) if hasattr(config, "mm_vision_tower"): # aka TA-Tok: this logic is used during SFT, not used during pretrain delay_load = getattr(config, "delay_load", False) self.vision_tower = build_vision_tower(config, delay_load=delay_load) self.sana = build_sana(config) self.sana_vae = build_vae(config) norm = RMSNorm(2304, eps=1e-5, elementwise_affine=True) # 2304 is the hidden size of sana with torch.no_grad(): norm.weight.fill_(math.sqrt(5.5)) self.diffusion_connector = nn.Sequential( # llm feature to SANA nn.Linear(config.hidden_size, 2304), # config.hidden_size is the hidden size of the language model (2048) nn.GELU(approximate="tanh"), nn.Linear(2304, 2304), norm, ) self.vae_connector = nn.Sequential( # vae feature to SANA nn.Linear(2240, 2304), # inner_dim(2240) = num_attention_heads * attention_head_dim nn.GELU(approximate="tanh"), nn.Linear(2304, 2304), norm, ) self.noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(config.diffusion_name_or_path, subfolder="scheduler") self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(config.diffusion_name_or_path, subfolder="scheduler") def get_vision_tower(self): vision_tower = getattr(self, "vision_tower", None) if type(vision_tower) is list: vision_tower = vision_tower[0] return vision_tower def get_sana(self): sana = getattr(self, 'sana', None) if type(sana) is list: sana = sana[0] if sana is not None: sana.to(self.device) return sana def get_sana_vae(self): sana_vae = getattr(self, 'sana_vae', None) if type(sana_vae) is list: sana_vae = sana_vae[0] if sana_vae is not None: sana_vae.to(self.device) return sana_vae def initialize_vision_modules(self, model_args, fsdp=None): vision_tower = model_args.vision_tower mm_vision_select_layer = model_args.mm_vision_select_layer # sft -2 mm_vision_select_feature = model_args.mm_vision_select_feature # sft "patch" mm_patch_merge_type = model_args.mm_patch_merge_type # sft 'flat' self.config.mm_vision_tower = vision_tower self.config.vision_tower_pretrained = getattr(model_args, "vision_tower_pretrained", "") if self.get_vision_tower() is None: vision_tower = build_vision_tower(model_args) if fsdp is not None and len(fsdp) > 0: self.vision_tower = [vision_tower] else: self.vision_tower = vision_tower else: if fsdp is not None and len(fsdp) > 0: vision_tower = self.vision_tower[0] else: vision_tower = self.vision_tower vision_tower.load_model() if self.get_sana() is None: sana = build_sana(model_args) self.noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(model_args.diffusion_name_or_path, subfolder="scheduler" ) self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(model_args.diffusion_name_or_path, subfolder="scheduler") if fsdp is not None and len(fsdp) > 0: self.sana = [sana] else: self.sana = sana else: if fsdp is not None and len(fsdp) > 0: sana = self.sana[0] else: sana = self.sana if self.get_sana_vae() is None: sana_vae = build_vae(model_args) if fsdp is not None and len(fsdp) > 0: self.sana_vae = [sana_vae] else: self.sana_vae = sana_vae else: if fsdp is not None and len(fsdp) > 0: sana_vae = self.sana_vae[0] else: sana_vae = self.sana_vae if getattr(self, 'diffusion_connector', None) is None: norm = RMSNorm(2304, eps=1e-5, elementwise_affine=True) with torch.no_grad(): norm.weight.fill_(math.sqrt(5.5)) self.diffusion_connector = nn.Sequential( nn.Linear(self.config.hidden_size, 2304), nn.GELU(approximate="tanh"), nn.Linear(2304, 2304), norm, ) else: for p in self.diffusion_connector.parameters(): p.requires_grad = True if getattr(self, 'vae_connector', None) is None and hasattr(model_args, "vae_connector") and model_args.vae_connector: self.vae_connector = nn.Sequential( nn.Linear(2240, 2304), nn.GELU(approximate="tanh"), nn.Linear(2304, 2304), RMSNorm(2304, eps=1e-5, elementwise_affine=True), ) self.config.use_mm_proj = True self.config.mm_hidden_size = vision_tower.hidden_size self.config.mm_vision_select_layer = mm_vision_select_layer self.config.mm_vision_select_feature = mm_vision_select_feature self.config.mm_patch_merge_type = mm_patch_merge_type class blip3oMetaForCausalLM(ABC): @abstractmethod def get_model(self): pass def get_vision_tower(self): return self.get_model().get_vision_tower() def encode_images(self, images, modalities, pool_scale=None): image_features = self.get_model().get_vision_tower()(images, pool_scale=pool_scale) assert 'tokens' in image_features image_tokens = image_features['tokens'] # discrete features for gen related tasks image_tokens = image_tokens + self.config.image_start_token_id image_features = self.get_model().embed_tokens(image_tokens) return {'image_features': image_features, 'image_tokens': image_tokens} def prepare_inputs_labels_for_multimodal(self, input_ids, position_ids, attention_mask, past_key_values, labels, images, modalities=None, image_sizes=None): vision_tower = self.get_vision_tower() if vision_tower is None or images is None or input_ids.shape[1] == 1: return input_ids, position_ids, attention_mask, past_key_values, None, labels if not isinstance(modalities, list): modalities = [modalities] # random scale for training, but scale 1 for understanding evaluation # if self.training: # pool_scale = random.choice(vision_tower.pool_scales) # vision_tower.pool_scales = [1, 1, 2, 3] # else: # pool_scale = 1 pool_scale = 1 if type(images) is list or images.ndim == 5: if type(images) is list: images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images] # [(1, 3, 384, 384), (1, 3, 384, 384), (1, 3, 384, 384), ......] images_list = [] for image in images: if image.ndim == 4: images_list.append(image) else: images_list.append(image.unsqueeze(0)) concat_images = torch.cat([image for image in images_list], dim=0) # [B, 3, 384, 384] split_sizes = [image.shape[0] for image in images_list] # [1, 1, 1, 1, ......] length is batch size encoded_image_features = self.encode_images(concat_images, modalities, pool_scale=pool_scale) # modalities not used # dict(['image_features': torch.Size([B, 729, 2048]), 'image_tokens': torch.Size([B, 729])]) image_tokens = encoded_image_features['image_tokens'] # torch.Size([B, 729]) encoded_image_features = encoded_image_features['image_features'] # torch.Size([B, 729, 2048]) # This is a list, each element is [num_images, patch * patch, dim] encoded_image_features = torch.split(encoded_image_features, split_sizes) # tuple of length batch size, each element is torch.Size([1, 729, 2048]) if image_tokens is not None: image_tokens = torch.split(image_tokens, split_sizes) # tuple, each element is torch.Size([1, 729]) image_features = [] for idx, image_feat in enumerate(encoded_image_features): image_features.append(image_feat) mm_patch_merge_type = getattr(self.config, "mm_patch_merge_type", "flat") if mm_patch_merge_type == "flat": image_features = [x.flatten(0, 1) for x in image_features] # list of length batch size, each element is torch.Size([729, 2048]) if image_tokens is not None: image_tokens = [x.flatten(0, 1) for x in image_tokens] # list of length batch size, each element is torch.Size([729]) else: raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}") else: image_features = self.encode_images(images, modalities, pool_scale=pool_scale) image_tokens = image_features['image_tokens'] image_features = image_features['image_features'] # Let's just add dummy tensors if they do not exist, # it is a headache to deal with None all the time. # But it is not ideal, and if you have a better idea, # please open an issue / submit a PR, thanks. # import ipdb; ipdb.set_trace() _labels = labels _position_ids = position_ids # None _attention_mask = attention_mask if attention_mask is None: attention_mask = torch.ones_like(input_ids, dtype=torch.bool) else: attention_mask = attention_mask.bool() if position_ids is None: position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device) if labels is None: labels = torch.full_like(input_ids, IGNORE_INDEX) # remove the padding using attention_mask -- FIXME _input_ids = input_ids input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)] labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)] new_input_embeds = [] new_labels = [] cur_image_idx = 0 # rank_print("Inserting Images embedding") for batch_idx, cur_input_ids in enumerate(input_ids): num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum() # rank0_print(num_images) if num_images == 0: # cur_image_features = image_features[cur_image_idx] cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids) # cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0) cur_input_embeds = torch.cat([cur_input_embeds_1, cur_input_embeds_1[0:0]], dim=0) # Unify processing flow: regardless of whether there is an image, follow the same splicing logic to prevent errors during parallel processing. new_input_embeds.append(cur_input_embeds) new_labels.append(labels[batch_idx]) cur_image_idx += 1 continue # [-1, pos1, pos2, ..., posn, cur_input_ids.shape[0]] image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]] cur_input_ids_noim = [] cur_labels = labels[batch_idx] cur_labels_noim = [] for i in range(len(image_token_indices) - 1): # Use IMAGE_TOKEN_INDEX as the splitting point to split input_ids and labels into num_images+1 parts, and then insert the image embedding cur_input_ids_noim.append(cur_input_ids[image_token_indices[i] + 1 : image_token_indices[i + 1]]) cur_labels_noim.append(cur_labels[image_token_indices[i] + 1 : image_token_indices[i + 1]]) split_sizes = [x.shape[0] for x in cur_labels_noim] cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim)) # Text embedding after removing images cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0) # Divide again cur_new_input_embeds = [] cur_new_labels = [] for i in range(num_images + 1): # n images divided into n+1 parts cur_new_input_embeds.append(cur_input_embeds_no_im[i]) cur_new_labels.append(cur_labels_noim[i]) # Adding text tokens before this line, adding image tokens after this line if i < num_images: # Image needs to be inserted after the first n parts try: cur_image_features = image_features[cur_image_idx] except IndexError: rank0_print("Error image_features[cur_image_idx]!") break # [Assisant\n] Logic for the last image. This if logic is mainly for adding pool tokens before the last image if self.config.image_start_tag_id == cur_labels_noim[i][-1] and image_tokens is not None: # Current slice's last token is tag cur_image_tokens = image_tokens[cur_image_idx] if pool_scale is not None: pool_token = self.config.scale_start_token_id + pool_scale - 1 pool_token = torch.tensor([pool_token], dtype=torch.long, device=cur_image_tokens.device) cur_image_tokens = torch.cat([pool_token, cur_image_tokens]) # [pool, 0, 1, 2, ..., 728] total 730 tokens pool_embed = self.get_model().embed_tokens(pool_token) cur_image_features = torch.cat([pool_embed, cur_image_features]) else: # Previous images go through this logic cur_image_tokens = torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype) cur_image_idx += 1 cur_new_input_embeds.append(cur_image_features) cur_new_labels.append(cur_image_tokens) cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds] cur_new_input_embeds = torch.cat(cur_new_input_embeds) cur_new_labels = torch.cat(cur_new_labels) new_input_embeds.append(cur_new_input_embeds) # [757, 2048] new_labels.append(cur_new_labels) # Truncate sequences to max length as image embeddings can make the sequence longer tokenizer_model_max_length = getattr(self.config, "tokenizer_model_max_length", None) # 2048: one image is long enough # import ipdb; ipdb.set_trace() new_input_embeds = [x[:tokenizer_model_max_length] for x, modality in zip(new_input_embeds, modalities)] new_labels = [x[:tokenizer_model_max_length] for x, modality in zip(new_labels, modalities)] # Combine them max_len = max(x.shape[0] for x in new_input_embeds) batch_size = len(new_input_embeds) new_input_embeds_padded = [] new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device) attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device) position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device) for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)): cur_len = cur_new_embed.shape[0] if getattr(self.config, "tokenizer_padding_side", "right") == "left": new_input_embeds_padded.append(torch.cat((torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device), cur_new_embed), dim=0)) if cur_len > 0: new_labels_padded[i, -cur_len:] = cur_new_labels attention_mask[i, -cur_len:] = True position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device) else: # Logic for right padding, pad with 0 new_input_embeds_padded.append(torch.cat((cur_new_embed, torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)) if cur_len > 0: new_labels_padded[i, :cur_len] = cur_new_labels attention_mask[i, :cur_len] = True position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device) new_input_embeds = torch.stack(new_input_embeds_padded, dim=0) # E.g., torch.Size([8, 901, 2048]) # Up to here we have the new input_embeds, labels, attention_mask, position_ids if _labels is None: new_labels = None else: new_labels = new_labels_padded if _attention_mask is None: attention_mask = None else: attention_mask = attention_mask.to(dtype=_attention_mask.dtype) if _position_ids is None: position_ids = None if getattr(self.config, "use_pos_skipping", False) and self.training: # This code implements the **Position Skipping** technique, which is a data augmentation method during training. position_ids = torch.arange(new_input_embeds.size(1), device=new_input_embeds.device).unsqueeze(0).to(new_input_embeds.device) split_position = random.randint(0, new_input_embeds.size(1)) left_add = random.randint(0, self.config.pos_skipping_range) right_add = random.randint(left_add, self.config.pos_skipping_range) position_ids[:, :split_position] += left_add position_ids[:, split_position:] += right_add return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels def initialize_vision_tokenizer(self, model_args, tokenizer): total_num_new_tokens = 0 vocab_size = len(tokenizer) if model_args.mm_use_im_start_end: num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True) self.config.image_start_tag_id = tokenizer.convert_tokens_to_ids(DEFAULT_IM_START_TOKEN) self.config.image_end_tag_id = tokenizer.convert_tokens_to_ids(DEFAULT_IM_END_TOKEN) total_num_new_tokens += num_new_tokens self.resize_token_embeddings(vocab_size + total_num_new_tokens) if model_args.num_scale_tokens > 0: scale_tokens = [model_args.scale_token_format.format(str(i)) for i in range(model_args.num_scale_tokens)] num_new_tokens = tokenizer.add_tokens(scale_tokens, special_tokens=False) self.config.scale_start_token_id = tokenizer.convert_tokens_to_ids(scale_tokens[0]) self.config.scale_end_token_id = tokenizer.convert_tokens_to_ids(scale_tokens[-1]) self.config.num_scale_tokens = model_args.num_scale_tokens total_num_new_tokens += num_new_tokens self.resize_token_embeddings(vocab_size + total_num_new_tokens) if model_args.num_image_tokens > 0: image_tokens = [model_args.image_token_format.format(str(i)) for i in range(model_args.num_image_tokens)] num_new_tokens = tokenizer.add_tokens(image_tokens, special_tokens=False) self.config.image_start_token_id = tokenizer.convert_tokens_to_ids(image_tokens[0]) self.config.image_end_token_id = tokenizer.convert_tokens_to_ids(image_tokens[-1]) self.config.num_image_tokens = model_args.num_image_tokens total_num_new_tokens += num_new_tokens self.resize_token_embeddings(vocab_size + total_num_new_tokens) if num_new_tokens > 0: self.config.num_new_tokens = num_new_tokens input_embeddings = self.get_input_embeddings().weight.data output_embeddings = self.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 vision_tower = self.get_vision_tower() if model_args.load_embeddings_from_vision and vision_tower is not None: vision_embeddings = vision_tower.get_embedding() if model_args.num_image_tokens == vision_embeddings.shape[0] and input_embeddings.shape[1] == vision_embeddings.shape[1]: rank0_print("Load vision embeddings from vision tower.") input_embeddings[self.config.image_start_token_id:self.config.image_end_token_id+1] = vision_embeddings