# Copyright 2023 Haotian Liu & Qinghao Ye (Modified from LLaVA) # # 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 from typing import List, Optional, Tuple, Union import torch import torch.nn as nn from torch.nn import CrossEntropyLoss from transformers import AutoConfig, AutoModelForCausalLM, LlamaModel, LlamaForCausalLM from transformers.modeling_outputs import CausalLMOutputWithPast from extensions.sd_smartprocess.mplug_owl2.constants import IMAGE_TOKEN_INDEX, IGNORE_INDEX from .configuration_mplug_owl2 import MPLUGOwl2Config, MplugOwlVisionConfig, MplugOwlVisualAbstractorConfig, \ MPLUGOwl2QwenConfig from .modeling_llama2 import replace_llama_modality_adaptive from .modeling_qwen import QWenLMHeadModel, QWenModel from .visual_encoder import MplugOwlVisionModel, MplugOwlVisualAbstractorModel class MPLUGOwl2MetaModel: def __init__(self, config): super(MPLUGOwl2MetaModel, self).__init__(config) self.vision_model = MplugOwlVisionModel( MplugOwlVisionConfig(**config.visual_config["visual_model"]) ) self.visual_abstractor = MplugOwlVisualAbstractorModel( MplugOwlVisualAbstractorConfig(**config.visual_config["visual_abstractor"]), config.hidden_size ) def get_vision_tower(self): vision_model = getattr(self, 'vision_model', None) if type(vision_model) is list: vision_model = vision_model[0] return vision_model def get_visual_abstractor(self): visual_abstractor = getattr(self, 'visual_abstractor', None) if type(visual_abstractor) is list: visual_abstractor = visual_abstractor[0] return visual_abstractor class MPLUGOwl2MetaForCausalLM(ABC): @abstractmethod def get_model(self): pass def encode_images(self, images): image_features = self.get_model().vision_model(images).last_hidden_state image_features = self.get_model().visual_abstractor(encoder_hidden_states=image_features).last_hidden_state return image_features def prepare_inputs_labels_for_multimodal( self, input_ids, attention_mask, past_key_values, labels, images ): if images is None or input_ids.shape[1] == 1: if past_key_values is not None and images is not None and input_ids.shape[1] == 1: # print(attention_mask) if attention_mask is not None: attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device) multiway_indices = torch.zeros_like(input_ids).long().to(self.device) return input_ids, multiway_indices, attention_mask, past_key_values, None, labels if type(images) is list or images.ndim == 5: 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) for x in image_features] else: image_features = self.encode_images(images) new_input_embeds = [] new_modality_indicators = [] new_labels = [] if labels is not None else None cur_image_idx = 0 for batch_idx, cur_input_ids in enumerate(input_ids): if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0: # multimodal LLM, but the current sample is not multimodal # FIXME: this is a hacky fix, for deepspeed zero3 to work half_len = cur_input_ids.shape[0] // 2 cur_image_features = image_features[cur_image_idx] cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids[:half_len]) cur_input_embeds_2 = self.get_model().embed_tokens(cur_input_ids[half_len:]) cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2], dim=0) new_input_embeds.append(cur_input_embeds) cur_modality_indicators = torch.zeros(len(cur_input_embeds)).long().to(self.device) new_modality_indicators.append(cur_modality_indicators) if labels is not None: new_labels.append(labels[batch_idx]) cur_image_idx += 1 continue image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0] cur_new_input_embeds = [] cur_modality_indicators = [] if labels is not None: cur_labels = labels[batch_idx] cur_new_labels = [] assert cur_labels.shape == cur_input_ids.shape while image_token_indices.numel() > 0: cur_image_features = image_features[cur_image_idx] image_token_start = image_token_indices[0] cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start])) cur_new_input_embeds.append(cur_image_features) # Add modality indicator assert image_token_start == len(cur_input_ids[:image_token_start]) cur_modality_indicators.append(torch.zeros(len(cur_input_ids[:image_token_start])).long()) cur_modality_indicators.append(torch.ones(len(cur_image_features)).long()) if labels is not None: cur_new_labels.append(cur_labels[:image_token_start]) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype)) cur_labels = cur_labels[image_token_start + 1:] cur_image_idx += 1 cur_input_ids = cur_input_ids[image_token_start + 1:] image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0] if cur_input_ids.numel() > 0: cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids)) cur_modality_indicators.append(torch.zeros(len(cur_input_ids)).long()) if labels is not None: cur_new_labels.append(cur_labels) cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds] cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0) new_input_embeds.append(cur_new_input_embeds) # Modality cur_modality_indicators = [x.to(device=self.device) for x in cur_modality_indicators] cur_modality_indicators = torch.cat(cur_modality_indicators, dim=0) new_modality_indicators.append(cur_modality_indicators) if labels is not None: cur_new_labels = torch.cat(cur_new_labels, dim=0) new_labels.append(cur_new_labels) if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds): max_len = max(x.shape[0] for x in new_input_embeds) # Embedding new_input_embeds_align = [] for cur_new_embed in new_input_embeds: cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0) new_input_embeds_align.append(cur_new_embed) new_input_embeds = torch.stack(new_input_embeds_align, dim=0) # Modality new_modality_indicators_align = [] for cur_modality_indicator in new_modality_indicators: cur_new_embed = torch.cat((cur_modality_indicator, torch.zeros(max_len - cur_modality_indicator.shape[0], dtype=cur_modality_indicator.dtype, device=cur_modality_indicator.device)), dim=0) new_modality_indicators_align.append(cur_new_embed) new_modality_indicators = torch.stack(new_modality_indicators_align, dim=0) # Label if labels is not None: new_labels_align = [] _new_labels = new_labels for cur_new_label in new_labels: cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0) new_labels_align.append(cur_new_label) new_labels = torch.stack(new_labels_align, dim=0) # Attention Mask if attention_mask is not None: new_attention_mask = [] for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels): new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device) new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device) cur_new_attention_mask = torch.cat( (new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0) new_attention_mask.append(cur_new_attention_mask) attention_mask = torch.stack(new_attention_mask, dim=0) assert attention_mask.shape == new_labels.shape else: new_input_embeds = torch.stack(new_input_embeds, dim=0) new_modality_indicators = torch.stack(new_modality_indicators, dim=0) if labels is not None: new_labels = torch.stack(new_labels, dim=0) if attention_mask is not None: new_attn_mask_pad_left = torch.full( (attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device) attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1) assert attention_mask.shape == new_input_embeds.shape[:2] return None, new_modality_indicators, attention_mask, past_key_values, new_input_embeds, new_labels def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None): """ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. """ bsz, src_len = mask.size() tgt_len = tgt_len if tgt_len is not None else src_len expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype) inverted_mask = 1.0 - expanded_mask return inverted_mask.masked_fill( inverted_mask.to(torch.bool), torch.finfo(dtype).min ) def _make_causal_mask( input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0 ): """ Make causal mask used for bi-directional self-attention. """ bsz, tgt_len = input_ids_shape mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min)) mask_cond = torch.arange(mask.size(-1)) mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0) mask = mask.to(dtype) if past_key_values_length > 0: mask = torch.cat( [torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1 ) return mask[None, None, :, :].expand( bsz, 1, tgt_len, tgt_len + past_key_values_length ) class MPLUGOwl2LlamaModel(MPLUGOwl2MetaModel, LlamaModel): config_class = MPLUGOwl2Config def __init__(self, config: MPLUGOwl2Config): super(MPLUGOwl2LlamaModel, self).__init__(config) def _prepare_decoder_attention_mask( self, attention_mask, input_shape, inputs_embeds, past_key_values_length ): # create causal mask # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] combined_attention_mask = None if input_shape[-1] > 1: combined_attention_mask = _make_causal_mask( input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length, ).to(inputs_embeds.device) if attention_mask is not None: # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len] expanded_attn_mask = _expand_mask( attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1] ).to(inputs_embeds.device) combined_attention_mask = ( expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask ) return combined_attention_mask class MPLUGOwl2QWenModel(MPLUGOwl2MetaModel, QWenModel): config_class = MPLUGOwl2QwenConfig def __init__(self, config: MPLUGOwl2QwenConfig): super(MPLUGOwl2QWenModel, self).__init__(config) class MPLUGOwl2LlamaForCausalLM(LlamaForCausalLM, MPLUGOwl2MetaForCausalLM): config_class = MPLUGOwl2Config def __init__(self, config): super(LlamaForCausalLM, self).__init__(config) self.model = MPLUGOwl2LlamaModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # Initialize weights and apply final processing self.post_init() def encode_images(self, images): image_features = self.get_model().vision_model(images).last_hidden_state image_features = self.get_model().visual_abstractor(encoder_hidden_states=image_features).last_hidden_state return image_features def get_model(self): return self.model def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, images: Optional[torch.FloatTensor] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) return_dict = return_dict if return_dict is not None else self.config.use_return_dict input_ids, modality_indicators, attention_mask, past_key_values, inputs_embeds, labels = \ self.prepare_inputs_labels_for_multimodal(input_ids, attention_mask, past_key_values, labels, images) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.model( input_ids=input_ids, modality_indicators=modality_indicators, attention_mask=attention_mask, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict ) hidden_states = outputs[0] logits = self.lm_head(hidden_states) loss = None if labels is not None: # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model/pipeline parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs ): if past_key_values: input_ids = input_ids[:, -1:] # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and past_key_values is None: model_inputs = {"inputs_embeds": inputs_embeds} else: model_inputs = {"input_ids": input_ids} model_inputs.update( { "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "attention_mask": attention_mask, "images": kwargs.get("images", None), } ) return model_inputs class MPLUGOwl2QWenForCausalLM(QWenLMHeadModel, MPLUGOwl2MetaForCausalLM): config_class = MPLUGOwl2QwenConfig def __init__(self, config): super(QWenLMHeadModel, self).__init__(config) from .modeling_qwen import SUPPORT_BF16, logger, SUPPORT_FP16, SUPPORT_CUDA, _import_flash_attn autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0 if autoset_precision: if SUPPORT_BF16: logger.warn( "The model is automatically converting to bf16 for faster inference. " "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"." ) config.bf16 = True elif SUPPORT_FP16: logger.warn( "The model is automatically converting to fp16 for faster inference. " "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"." ) config.fp16 = True else: config.fp32 = True if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16: logger.warn( "Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".") if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16: logger.warn( "Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster") if config.fp32: if SUPPORT_BF16: logger.warn( "Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".") elif SUPPORT_FP16: logger.warn( "Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".") if config.use_flash_attn == "auto": if config.bf16 or config.fp16: logger.warn("Try importing flash-attention for faster inference...") config.use_flash_attn = True else: config.use_flash_attn = False if config.use_flash_attn and config.fp32: logger.warn("Flash attention will be disabled because it does NOT support fp32.") if config.use_flash_attn: _import_flash_attn() self.transformer = MPLUGOwl2QWenModel(config) self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) if config.bf16: self.transformer.bfloat16() self.lm_head.bfloat16() if config.fp16: self.transformer.half() self.lm_head.half() self.post_init() def get_model(self): return self.transformer def forward( self, input_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, attention_mask: Optional[torch.FloatTensor] = None, token_type_ids: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, encoder_hidden_states: Optional[torch.Tensor] = None, encoder_attention_mask: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, images=None, return_dict: Optional[bool] = None, ) -> Union[Tuple, CausalLMOutputWithPast]: return_dict = ( return_dict if return_dict is not None else self.config.use_return_dict ) input_ids, modality_indicators, attention_mask, past_key_values, inputs_embeds, labels = \ self.prepare_inputs_labels_for_multimodal(input_ids, attention_mask, past_key_values, labels, images) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.transformer( input_ids, modality_indicators=modality_indicators, past_key_values=past_key_values, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = outputs[0] logits = self.lm_head(hidden_states) loss = None if labels is not None: # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model/pipeline parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return CausalLMOutputWithPast( loss=loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) AutoConfig.register("mplug_owl2", MPLUGOwl2Config) AutoModelForCausalLM.register(MPLUGOwl2Config, MPLUGOwl2LlamaForCausalLM) AutoConfig.register("mplug_owl2_1", MPLUGOwl2QwenConfig) AutoModelForCausalLM.register(MPLUGOwl2QwenConfig, MPLUGOwl2QWenForCausalLM) replace_llama_modality_adaptive() if __name__ == "__main__": config = MPLUGOwl2Config.from_pretrained('/cpfs01/shared/public/test/vicuna-7b-v1.5/') from icecream import ic # config = MPLUGOwl2Config() model = MPLUGOwl2LlamaForCausalLM(config) images = torch.randn(2, 3, 448, 448) input_ids = torch.cat([ torch.ones(8).long(), torch.tensor([-1] * 1).long(), torch.ones(8).long(), torch.tensor([-1] * 1).long(), torch.ones(8).long() ], dim=0).unsqueeze(0) labels = input_ids.clone() labels[labels < 0] = -100 # image_feature = model.encode_images(images) # ic(image_feature.shape) output = model(images=images, input_ids=input_ids, labels=labels) ic(output.loss) ic(output.logits.shape) model.save_pretrained('/cpfs01/shared/public/test/tmp_owl')