# Copyright 2023 Haotian Liu # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from abc import ABC, abstractmethod import torch from LLAVA_Biovil.biovil_t.model import ImageModel from LLAVA_Biovil.biovil_t.pretrained import _download_biovil_t_image_model_weights from LLAVA_Biovil.biovil_t.types import ImageEncoderType from LLAVA_Biovil.llava.model.multimodal_encoder.builder import build_vision_tower from LLAVA_Biovil.llava.model.multimodal_projector.builder import build_vision_projector, build_image_pooler from LLAVA_Biovil.llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN class LlavaMetaModel: def __init__(self, config, mv_type='none'): super(LlavaMetaModel, self).__init__(config) if hasattr(config, "mm_vision_tower"): self.vision_tower = build_vision_tower(config, delay_load=True) self.mm_projector = build_vision_projector(config) self.image_pooler = build_image_pooler(config) if "pool" in mv_type else None 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_image_pooler(self): return self.image_pooler 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 mm_vision_select_feature = model_args.mm_vision_select_feature pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter self.config.mm_vision_tower = vision_tower self.config.mv_type = getattr(model_args, 'mv_type', False) if self.get_vision_tower() is None: if self.config.mm_vision_tower == 'biovil': biovilt_checkpoint_path = _download_biovil_t_image_model_weights() model_type = ImageEncoderType.RESNET50_MULTI_IMAGE vision_tower = ImageModel(img_encoder_type=model_type, joint_feature_size=128, pretrained_model_path=biovilt_checkpoint_path) # freeze vision_tower layers for p in vision_tower.parameters(): p.requires_grad = False else: 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() self.config.use_mm_proj = True self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear') self.config.mm_hidden_size = vision_tower.hidden_size if self.config.mm_vision_tower != 'biovil' else vision_tower.feature_size self.config.mm_vision_select_layer = mm_vision_select_layer self.config.mm_vision_select_feature = mm_vision_select_feature if getattr(self, 'mm_projector', None) is None or model_args.vision_tower == 'biovil': #for biovil wrong weights are loaded from model shards, so we need to overwrite the vision projector again self.mm_projector = build_vision_projector(self.config) else: # In case it is frozen by LoRA for p in self.mm_projector.parameters(): p.requires_grad = True # unfreeze image pooler if self.image_pooler is not None: for p in self.image_pooler.parameters(): p.requires_grad = True if pretrain_mm_mlp_adapter is not None: mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu') def get_w(weights, keyword): return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k} self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector')) class LlavaMetaForCausalLM(ABC): @abstractmethod def get_model(self): pass def get_vision_tower(self): return self.get_model().get_vision_tower() def encode_images(self, images): image_features = self.get_model().get_vision_tower()(images) if self.get_model().config.mm_vision_tower == 'biovil': image_features = image_features.patch_embeddings # flatten image_features = image_features.flatten(2).transpose(1,2) image_features = self.get_model().mm_projector(image_features) return image_features def pad_embeddings(self, embeddings, num_imgs_present=None, num_imgs_past=None, padding_value=0): """ Pad the embeddings to have the same number in each batch. Args: - embeddings (List[Tensor]): List of embedding tensors, each with shape (num_images, embedding_dim). - padding_value (float): Value to use for padding. Returns: - Tensor: Padded embeddings with shape (batch_size, max_num_images, embedding_dim). - Tensor: Mask indicating real data (1) and padding (0). """ batch_size = len(embeddings) img_len = embeddings[0].shape[1] embedding_dim = embeddings[0].shape[2] max_num_images = max(emb.shape[0] for emb in embeddings) # Initialize padded embeddings and mask padded_embeddings = torch.full((batch_size, max_num_images, img_len, embedding_dim), padding_value, dtype=embeddings[0].dtype, device=embeddings[0].device) mask = torch.zeros(batch_size, max_num_images*img_len, dtype=torch.bool, device=embeddings[0].device) # create token type ids with 0 for present 1 for past, 2 for padding, of shape (batch_size, max_num_images * img_len) token_type_ids = torch.zeros(batch_size, max_num_images * img_len, dtype=torch.long, device=embeddings[0].device) if num_imgs_present is not None: # set token type ids for present to 1, for past to 2, 0 is padded elements for idx, (present_len, past_len) in enumerate(zip(num_imgs_present, num_imgs_past)): token_type_ids[idx, :present_len*img_len] = 1 token_type_ids[idx, present_len*img_len:(present_len+past_len)*img_len] = 2 # Pad each item in the batch for idx, emb in enumerate(embeddings): num_images = emb.shape[0] padded_embeddings[idx, :num_images] = emb mask[idx, :num_images*img_len] = 1 return padded_embeddings.flatten(1,2), mask, token_type_ids def pad_embeddings_mv(self, embeddings, padding_value=0): """ Pad the embeddings to have the same number in each batch. Args: - embeddings (List[Tensor]): List of embedding tensors, each with shape (num_images, embedding_dim). - padding_value (float): Value to use for padding. Returns: - Tensor: Padded embeddings with shape (batch_size, max_num_images, embedding_dim). - Tensor: Mask indicating real data (1) and padding (0). """ batch_size = len(embeddings) img_len = embeddings[0].shape[1] embedding_dim = embeddings[0].shape[2] max_num_images = max(emb.shape[0] for emb in embeddings) # Initialize padded embeddings and mask padded_embeddings = torch.full((batch_size, max_num_images, img_len, embedding_dim), padding_value, dtype=embeddings[0].dtype, device=embeddings[0].device) mask = torch.zeros(batch_size, max_num_images*img_len, dtype=torch.bool, device=embeddings[0].device) # Pad each item in the batch for idx, emb in enumerate(embeddings): num_images = emb.shape[0] padded_embeddings[idx, :num_images] = emb mask[idx, :num_images*img_len] = 1 return padded_embeddings.flatten(1,2), mask def encode_images_pooled(self, images, split_sizes, num_imgs_present, num_imgs_past, mv_type="pool_all"): image_pooler = self.get_image_pooler() image_features = self.get_model().get_vision_tower()(images) if self.get_model().config.mm_vision_tower == 'biovil': image_features = image_features.patch_embeddings # flatten image_features = image_features.flatten(2).transpose(1,2) if split_sizes is not None: image_features = torch.split(image_features, split_sizes, dim=0) if mv_type == "pool_all": # merge present and past per batch present_features = [image_features[i] for i in range(len(num_imgs_present))] past_features = [] i = 0 for num_imgs_elem in num_imgs_past: if num_imgs_elem != 0: past_features.append(image_features[i+len(num_imgs_present)]) i += 1 else: past_features.append(None) all_img_features = [] for idx, (batch_num_present, batch_num_past) in enumerate(zip(num_imgs_present, num_imgs_past)): if batch_num_past == 0: all_img_features.append(present_features[idx]) else: all_img_features.append(torch.cat((present_features[idx], past_features[idx]), dim=0)) all_img_features, mask, token_type_ids = self.pad_embeddings(all_img_features, num_imgs_present, num_imgs_past) all_img_features = image_pooler(all_img_features, mask, token_type_ids) elif mv_type == "pool_concat": present_features = [image_features[i] for i in range(len(num_imgs_present))] past_features = [image_features[i+len(num_imgs_present)] for i in range(len(image_features)-len(num_imgs_present))] present_features, mask_present, _ = self.pad_embeddings(present_features) past_features, mask_past, _ = self.pad_embeddings(past_features) present_features = image_pooler(present_features, mask_present) past_features = image_pooler(past_features, mask_past) # TODO maybe max pool on past features to save tokens # concat present and past per batch if past is not empty all_img_features = [] idx_present = 0 idx_past = 0 for batch_num_present, batch_num_past in zip(num_imgs_present, num_imgs_past): if batch_num_past == 0: all_img_features.append(present_features[idx_present]) idx_present += 1 else: all_img_features.append(torch.cat((present_features[idx_present], past_features[idx_past]), dim=0)) idx_present += 1 idx_past += 1 else: raise NotImplementedError if type(all_img_features) is list: split_sizes = [image.shape[0] for image in all_img_features] all_img_features = self.get_model().mm_projector(torch.cat(all_img_features, dim=0)) all_img_features = torch.split(all_img_features, split_sizes, dim=0) else: all_img_features = self.get_model().mm_projector(all_img_features) return all_img_features def encode_images_pooled_mv(self, images, split_sizes): image_pooler = self.get_image_pooler() image_features = self.get_model().get_vision_tower()(images) if split_sizes is not None: image_features = torch.split(image_features, split_sizes, dim=0) image_features, mask = self.pad_embeddings_mv(image_features) image_features = image_pooler(image_features, mask) else: mask = torch.ones((image_features.shape[0], image_features.shape[1]), dtype=torch.bool, device=image_features[0].device) image_features = image_pooler(image_features, mask) image_features = self.get_model().mm_projector(image_features) return image_features def get_image_pooler(self): return self.get_model().get_image_pooler() def prepare_inputs_labels_for_multimodal( self, input_ids, position_ids, attention_mask, past_key_values, labels, images, prev_images=None ): vision_tower = self.get_vision_tower() if vision_tower is None or images is None or input_ids.shape[1] == 1: if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1: target_shape = past_key_values[-1][-1].shape[-2] + 1 attention_mask = torch.cat((attention_mask, torch.ones( (attention_mask.shape[0], target_shape - attention_mask.shape[1]), dtype=attention_mask.dtype, device=attention_mask.device )), dim=1) position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1 return input_ids, position_ids, attention_mask, past_key_values, None, labels if type(images) is list or images.ndim == 5: if getattr(self.config, 'mv_type') == "concat": concat_images = torch.cat([image for image in images], dim=0) image_features = self.encode_images(concat_images) split_sizes = [image.shape[0] for image in images] image_features = torch.split(image_features, split_sizes, dim=0) image_features = [x.flatten(0, 1).to(self.device) for x in image_features] if getattr(self.config, 'mv_type') == "pool_all": concat_images = torch.cat((torch.cat([image for image in images], dim=0), torch.cat([image for image in prev_images if image is not None], dim=0))) # first present, then past, all will be merged split_sizes = [image.shape[0] for image in images]+ [image.shape[0] for image in prev_images if image is not None] num_imgs_present = [image.shape[0] if image is not None else 0 for image in images] num_imgs_past = [image.shape[0] if image is not None else 0 for image in prev_images] image_features = self.encode_images_pooled(concat_images, split_sizes, num_imgs_present, num_imgs_past, "pool_all") if getattr(self.config, 'mv_type') == "pool_concat": # TODO make sure to allow empty past -> shorter sequence concat_images = torch.cat((torch.cat([image for image in images], dim=0), torch.cat([image for image in prev_images if image is not None], dim=0))) # first present, then past, all will be merged split_sizes = [image.shape[0] for image in images]+ [image.shape[0] for image in prev_images if image is not None] num_imgs_present = [image.shape[0] if image is not None else 0 for image in images] num_imgs_past = [image.shape[0] if image is not None else 0 for image in prev_images] image_features = self.encode_images_pooled(concat_images, split_sizes, num_imgs_present, num_imgs_past, "pool_concat") if getattr(self.config, 'mv_type') == "pool": #no past images concat_images = torch.cat([image for image in images], dim=0) split_sizes = [image.shape[0] for image in images] image_features = self.encode_images_pooled_mv(concat_images, split_sizes) else: if hasattr(self.config, 'mv_type') and getattr(self.config, 'mv_type') == "pool_all": image_features = self.encode_images_pooled(images, None).to(self.device) elif hasattr(self.config, 'mv_type') and getattr(self.config, 'mv_type') == "pool": image_features = self.encode_images_pooled_mv(images, None).to(self.device) else: image_features = self.encode_images(images).to(self.device) # TODO: image start / end is not implemented here to support pretraining. if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): raise NotImplementedError # 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. _labels = labels _position_ids = position_ids _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) #TODO throws GPU error if labels is None: labels = torch.full_like(input_ids, IGNORE_INDEX) # remove the padding using attention_mask -- TODO: double check 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 for batch_idx, cur_input_ids in enumerate(input_ids): num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum() 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) new_input_embeds.append(cur_input_embeds) new_labels.append(labels[batch_idx]) cur_image_idx += 1 continue 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): 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)) cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0) cur_new_input_embeds = [] cur_new_labels = [] for i in range(num_images + 1): cur_new_input_embeds.append(cur_input_embeds_no_im[i]) cur_new_labels.append(cur_labels_noim[i]) if i < num_images: cur_image_features = image_features[cur_image_idx] cur_image_idx += 1 cur_new_input_embeds.append(cur_image_features) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype)) 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) 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) if tokenizer_model_max_length is not None: max_len_orig = max(x.shape[0] for x in new_input_embeds) if max_len_orig > tokenizer_model_max_length: print(f"Truncating sequences of len {max_len_orig} to {tokenizer_model_max_length} to fit the model's input length") new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds] new_labels = [x[:tokenizer_model_max_length] for x in new_labels] # 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: 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) 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 return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels def initialize_vision_tokenizer(self, model_args, tokenizer): if model_args.mm_use_im_patch_token: tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True) self.resize_token_embeddings(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.resize_token_embeddings(len(tokenizer)) if num_new_tokens > 0: 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 if model_args.tune_mm_mlp_adapter: for p in self.get_input_embeddings().parameters(): p.requires_grad = True for p in self.get_output_embeddings().parameters(): p.requires_grad = False if model_args.pretrain_mm_mlp_adapter: mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu') embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight'] assert num_new_tokens == 2 if input_embeddings.shape == embed_tokens_weight.shape: input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:] elif embed_tokens_weight.shape[0] == num_new_tokens: input_embeddings[-num_new_tokens:] = embed_tokens_weight else: raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.") elif model_args.mm_use_im_patch_token: if model_args.tune_mm_mlp_adapter: for p in self.get_input_embeddings().parameters(): p.requires_grad = False for p in self.get_output_embeddings().parameters(): p.requires_grad = False