from contextlib import nullcontext from dataclasses import dataclass from typing import List, Optional, Tuple, Union import torch import torch.utils.checkpoint from torch import nn from transformers import PreTrainedModel from transformers.activations import ACT2FN from transformers.cache_utils import Cache from transformers.modeling_outputs import ModelOutput from transformers.models.clip.configuration_clip import CLIPConfig from transformers.utils import ( add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings, ) from transformers import AutoModel, AutoModelForCausalLM from transformers.models.llava.configuration_llava import LlavaConfig from transformers.models.llava.modeling_llava import ( LlavaCausalLMOutputWithPast, LlavaMultiModalProjector, LlavaPreTrainedModel, LLAVA_START_DOCSTRING, LLAVA_INPUTS_DOCSTRING, LlavaForConditionalGeneration, ) from transformers.models.blip_2.configuration_blip_2 import ( Blip2Config, Blip2QFormerConfig, ) import os from transformers.models.blip_2.modeling_blip_2 import ( Blip2Config, Blip2QFormerModel, Blip2PreTrainedModel, BLIP_2_INPUTS_DOCSTRING, ) from transformers.utils.import_utils import is_flash_attn_greater_or_equal_2_10 # from .configuration_sealmm import SeaLMMConfig logger = logging.get_logger(__name__) # _CONFIG_FOR_DOC = "LlavaConfig" _CONFIG_FOR_DOC = "SeaLMMConfig" class SeaLMMConfig(LlavaConfig): def __init__(self, *args, **kwargs): self.projector_num_layers = kwargs.get("projector_num_layers", 1) super().__init__(*args, **kwargs) """ Llava vision_config.num_hidden_layers = vision_config.num_hidden_layers + config.vision_feature_layer + 1 # "num_hidden_layers": 24, """ IMAGE_TOKEN = "<|image|>" DEBUG = bool(int(os.environ.get("DEBUG", "0"))) def by_sample_merge_input_ids_with_image_features( self, image_features, inputs_embeds, input_ids, attention_mask=None, position_ids=None ): """ input_ids: [tlen] input_embeds: [tlen, dt] img_embeds: [ilen, ifeat, di] e.g: input_ids: [ a b c d e f X g h i j k X l m ] img_embeds: [3, ifeat, id] # img_embeds has padding """ num_images, num_image_patches, embed_dim = image_features.shape sequence_length = input_ids.size(0) left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id)) assert not left_padding, f'should only use right padding' # 1. Create a mask to know where special image tokens are special_image_token_mask = input_ids == self.config.image_token_index num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1) # Compute the maximum embed dimension max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length from transformers.models.clip.configuration_clip import CLIPConfig, CLIPTextConfig, CLIPVisionConfig from transformers.models.clip.modeling_clip import ( contrastive_loss, clip_loss, CLIPVisionModelOutput, CLIPTextModelOutput, CLIPOutput, CLIPTextEmbeddings, CLIPVisionEmbeddings, CLIPAttention, CLIPMLP, CLIPEncoderLayer, CLIPPreTrainedModel, CLIPTextTransformer, CLIPTextModel, CLIPVisionTransformer, CLIPVisionModel, CLIPModel, CLIPEncoder, CLIPTextModelWithProjection, CLIPVisionModelWithProjection, CLIP_START_DOCSTRING, CLIP_TEXT_INPUTS_DOCSTRING, CLIP_VISION_INPUTS_DOCSTRING, CLIP_INPUTS_DOCSTRING, ) from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling # Copied from transformers.models.llama.modeling_llama._get_unpad_data def _get_unpad_data(attention_mask): import torch.nn.functional as F seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() max_seqlen_in_batch = seqlens_in_batch.max().item() cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0)) return ( indices, cu_seqlens, max_seqlen_in_batch, ) class CLIPFlashAttention2(CLIPAttention): """ CLIP flash attention module. This module inherits from `CLIPAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ def __init__(self, config, is_causal=False): super().__init__(config) self.is_causal = is_causal def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, causal_attention_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = False, ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]: """Input shape: Batch x Time x Channel""" if output_attentions: raise ValueError("CLIPFlashAttention2 does not support output_attentions") if self.is_causal and causal_attention_mask is None: raise ValueError("CLIPFlashAttention2 has causal=True but no causal_attention_mask provided") bsz, tgt_len, embed_dim = hidden_states.size() # [batch_size, tgt_len, embed_dim] query_states = self.q_proj(hidden_states) key_states = self.k_proj(hidden_states) value_states = self.v_proj(hidden_states) # [batch_size, tgt_len, embed_dim] -> [batch_size, tgt_len, num_heads, head_dim] query_states = query_states.view(bsz, tgt_len, self.num_heads, self.head_dim).contiguous() key_states = key_states.view(bsz, tgt_len, self.num_heads, self.head_dim).contiguous() value_states = value_states.view(bsz, tgt_len, self.num_heads, self.head_dim).contiguous() attn_output = self._flash_attention_forward( query_states=query_states, key_states=key_states, value_states=value_states, attention_mask=attention_mask, query_length=tgt_len, dropout=self.dropout, softmax_scale=self.scale, ) # [batch_size, tgt_len, num_heads, head_dim] -> [batch_size, tgt_len, embed_dim] attn_output = attn_output.view(bsz, tgt_len, embed_dim) attn_output = self.out_proj(attn_output) return attn_output, None # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2._flash_attention_forward def _flash_attention_forward( self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None ) -> torch.Tensor: """ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token first unpad the input, then computes the attention scores and pad the final attention scores. Args: query_states (`torch.Tensor`): Input query states to be passed to Flash Attention API key_states (`torch.Tensor`): Input key states to be passed to Flash Attention API value_states (`torch.Tensor`): Input value states to be passed to Flash Attention API attention_mask (`torch.Tensor`): The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the position of padding tokens and 1 for the position of non-padding tokens. dropout (`int`, *optional*): Attention dropout softmax_scale (`float`, *optional*): The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) """ from flash_attn import flash_attn_func, flash_attn_varlen_func from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa # Contains at least one padding token in the sequence if attention_mask is not None: batch_size = query_states.shape[0] query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states, attention_mask, query_length ) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=self.is_causal, ) attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) else: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=self.is_causal ) return attn_output def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape key_layer = index_first_axis( key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k ) value_layer = index_first_axis( value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k ) if query_length == kv_seq_len: query_layer = index_first_axis( query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k ) cu_seqlens_q = cu_seqlens_k max_seqlen_in_batch_q = max_seqlen_in_batch_k indices_q = indices_k elif query_length == 1: max_seqlen_in_batch_q = 1 # There is a memcpy here, that is very bad. cu_seqlens_q = torch.arange(batch_size + 1, dtype=torch.int32, device=query_layer.device) indices_q = cu_seqlens_q[:-1] query_layer = query_layer.squeeze(1) else: # The :q_len slice assumes right padding. attention_mask = attention_mask[:, :query_length] query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) return ( query_layer, key_layer, value_layer, indices_q, (cu_seqlens_q, cu_seqlens_k), (max_seqlen_in_batch_q, max_seqlen_in_batch_k), ) class SeaLMMCLIPEncoderLayer(CLIPEncoderLayer): def __init__(self, config: CLIPConfig): super(CLIPEncoderLayer, self).__init__() self.embed_dim = config.hidden_size # self.self_attn = LlavaCLIPFlashAttention(config) if is_flash_attn_greater_or_equal_2_10(): self.self_attn = CLIPFlashAttention2(config) else: self.self_attn = CLIPAttention(config) self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) self.mlp = CLIPMLP(config) self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps) class SeaLMMCLIPEncoder(CLIPEncoder): """ Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a [`CLIPEncoderLayer`]. Args: config: CLIPConfig """ def __init__(self, config: CLIPConfig): super(CLIPEncoder, self).__init__() self.config = config self.layers = nn.ModuleList([SeaLMMCLIPEncoderLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False def forward( self, inputs_embeds, attention_mask: Optional[torch.Tensor] = None, causal_attention_mask: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutput]: r""" Args: inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert `input_ids` indices into associated vectors than the model's internal embedding lookup matrix. attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) causal_attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): Causal mask for the text model. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. """ 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 output_hidden_states = False output_attentions = False # return_dict = False encoder_states = () if output_hidden_states else None all_attentions = () if output_attentions else None hidden_states = inputs_embeds for idx, encoder_layer in enumerate(self.layers): if output_hidden_states: encoder_states = encoder_states + (hidden_states,) # if self.gradient_checkpointing and self.training: # layer_outputs = self._gradient_checkpointing_func( # encoder_layer.__call__, # hidden_states, # attention_mask, # causal_attention_mask, # output_attentions, # ) # else: # ! enforce no checkpointing here layer_outputs = encoder_layer( hidden_states, attention_mask, causal_attention_mask, output_attentions=output_attentions, ) hidden_states = layer_outputs[0] if output_attentions: all_attentions = all_attentions + (layer_outputs[1],) if output_hidden_states: encoder_states = encoder_states + (hidden_states,) if not return_dict: return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None) return BaseModelOutput( last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions ) class SeaLMMVisionTransformer(nn.Module): def __init__(self, config: CLIPVisionConfig): super().__init__() self.config = config embed_dim = config.hidden_size self.embeddings = CLIPVisionEmbeddings(config) self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) # self.encoder = CLIPEncoder(config) self.encoder = SeaLMMCLIPEncoder(config) # self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps) @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig) def forward( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutputWithPooling]: r""" Returns: """ assert output_attentions is None assert output_hidden_states is None # assert return_dict is None 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 if pixel_values is None: raise ValueError("You have to specify pixel_values") hidden_states = self.embeddings(pixel_values) hidden_states = self.pre_layrnorm(hidden_states) encoder_outputs = self.encoder( inputs_embeds=hidden_states, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) last_hidden_state = encoder_outputs[0] if not return_dict: raise ValueError(f'Not support return_dict') return BaseModelOutputWithPooling( last_hidden_state=last_hidden_state, # pooler_output=pooled_output, pooler_output=None, hidden_states=encoder_outputs.hidden_states, attentions=encoder_outputs.attentions, ) @add_start_docstrings( """The vision model from CLIP without any head or projection on top.""", CLIP_START_DOCSTRING, ) class SeaLMMCLIPVisionModel(CLIPPreTrainedModel): config_class = CLIPVisionConfig main_input_name = "pixel_values" _no_split_modules = ["SeaLMMCLIPEncoderLayer"] def __init__(self, config: CLIPVisionConfig): super().__init__(config) self.vision_model = SeaLMMVisionTransformer(config) # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Module: return self.vision_model.embeddings.patch_embedding @add_start_docstrings_to_model_forward(CLIP_VISION_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=CLIPVisionConfig) def forward( self, pixel_values: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, BaseModelOutputWithPooling]: r""" Returns: Examples: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, CLIPVisionModel >>> model = CLIPVisionModel.from_pretrained("openai/clip-vit-base-patch32") >>> processor = AutoProcessor.from_pretrained("openai/clip-vit-base-patch32") >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> last_hidden_state = outputs.last_hidden_state >>> pooled_output = outputs.pooler_output # pooled CLS states ```""" # return_dict = return_dict if return_dict is not None else self.config.use_return_dict return self.vision_model( pixel_values=pixel_values, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) class SeaLMMMultiModalProjector(SeaLMMCLIPEncoder): def __init__(self, config: SeaLMMConfig): super(CLIPEncoder, self).__init__() self.config = config self.projector_num_layers = getattr(config, "projector_num_layers", 2) self.vision_config = config.vision_config self.num_vision_feature_layer = int(0 - config.vision_feature_layer) - 1 assert self.num_vision_feature_layer > 0 self.layers = nn.ModuleList([ # LlavaCLIPFasterEncoderLayer(self.vision_config) SeaLMMCLIPEncoderLayer(self.vision_config) for _ in range(self.projector_num_layers)] ) projector_layernorm_eps = getattr(config, "projector_layernorm_eps", 1e-05) self.projector_layernorm = nn.LayerNorm( # len(config.vision_feature_layers) * config.vision_config.hidden_size, eps=projector_layernorm_eps config.vision_config.hidden_size, eps=projector_layernorm_eps ) self.linear_1 = nn.Linear( # len(config.vision_feature_layers) * config.vision_config.hidden_size, config.vision_config.hidden_size, config.text_config.hidden_size, bias=True, ) # self.act = ACT2FN[config.projector_hidden_act] # self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True) self.gradient_checkpointing = False def forward(self, hidden_states, attention_mask=None, causal_attention_mask=None): """ hidden_states must not be striped """ output_attentions = False for idx, encoder_layer in enumerate(self.layers): # if output_hidden_states: # encoder_states = encoder_states + (hidden_states,) # if self.gradient_checkpointing and self.training: # layer_outputs = self._gradient_checkpointing_func( # encoder_layer.__call__, # hidden_states, # attention_mask, # causal_attention_mask, # output_attentions, # ) # else: # ! turn off checkpointing layer_outputs = encoder_layer( hidden_states, attention_mask, causal_attention_mask, output_attentions=output_attentions, ) hidden_states = layer_outputs[0] hidden_states = hidden_states[:, 1:] hidden_states = self.projector_layernorm(hidden_states) hidden_states = self.linear_1(hidden_states) # hidden_states = self.act(hidden_states) # hidden_states = self.linear_2(hidden_states) return hidden_states @add_start_docstrings( """The CLip- LLAVA model which consists of a vision backbone and a language model.""", LLAVA_START_DOCSTRING, ) class SeaLMMForCausalLM(LlavaPreTrainedModel): def __init__(self, config: SeaLMMConfig, vision_tower=None, language_model=None): super().__init__(config) # self.vision_tower = AutoModel.from_config(config.vision_config) # self.vision_tower = vision_tower or LlavaCLIPVisionModel(config=config.vision_config) self.vision_tower = vision_tower or SeaLMMCLIPVisionModel(config=config.vision_config) self.multi_modal_projector = SeaLMMMultiModalProjector(config) # self.vocab_size = config.text_config.vocab_size self.language_model = language_model or AutoModelForCausalLM.from_config( config.text_config, attn_implementation=config._attn_implementation ) self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1 self.post_init() self.freeze_vision_tower = True def unfreeze_vision_tower(self): logger.info(f'UNFREEZE {self.freeze_vision_tower=}') self.freeze_vision_tower = False def freeze_vision_tower(self): logger.info(f'FREEZE {self.freeze_vision_tower=}') self.freeze_vision_tower = True @classmethod def create_model_config_from_components( cls, lm_config=None, vision_config=None, tokenizer=None, vision_feature_layer=None, projector_num_layers=1, **kwargs, ) -> SeaLMMConfig: # self.projector_num_layers = kwargs.get("projector_num_layers", 1) config = SeaLMMConfig(vision_config, lm_config, projector_num_layers=projector_num_layers, **kwargs) config.vision_feature_layer = config.vision_feature_layer if vision_feature_layer is None else vision_feature_layer if config.vision_feature_layer < 0: config.vision_config.num_hidden_layers = config.vision_config.num_hidden_layers + config.vision_feature_layer + 1 else: config.vision_config.num_hidden_layers = config.vision_feature_layer + 1 if IMAGE_TOKEN not in tokenizer.get_vocab(): tokenizer.add_special_tokens({"cls_token": IMAGE_TOKEN}) config.image_token_index = tokenizer.cls_token_id config.vocab_size = config.text_config.vocab_size config.architectures = ["SeaLMMForCausalLM"] return config def get_input_embeddings(self): return self.language_model.get_input_embeddings() def set_input_embeddings(self, value): self.language_model.set_input_embeddings(value) def get_output_embeddings(self): return self.language_model.get_output_embeddings() def set_output_embeddings(self, new_embeddings): self.language_model.set_output_embeddings(new_embeddings) def set_decoder(self, decoder): self.language_model.set_decoder(decoder) def get_decoder(self): return self.language_model.get_decoder() def tie_weights(self): return self.language_model.tie_weights() def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding: model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of) # update vocab size self.config.text_config.vocab_size = model_embeds.num_embeddings self.config.vocab_size = model_embeds.num_embeddings self.vocab_size = model_embeds.num_embeddings return model_embeds # @torch.no_grad def _merge_input_ids_with_image_features( self, image_features, inputs_embeds, input_ids, attention_mask, position_ids, labels=None ): """ input_ids: [b, tlen] input_embeds: [b, tlen, dt] image_features: [b, ilen, ifeat, di] labels: None or [b, tlen] --> must extend labels to input_ids, # in input_ids, there may be image_token_index, number of image_token_index <= ilen input_ids: [ a b c d e f X g h i j k X l m o p q r X s t u v _ _ _ _ _ _ ] input_ids should be: [ a b c d e f X X X X X g h i j k X X X X X l m o p q r X X X X X s t u v _ _ _ _ _ _ _ _ _ _ ] labels should be: [ a b c d e f _ _ _ _ _ g h i j k _ _ _ _ _ l m o p q r _ _ _ _ _ s t u v _ _ _ _ _ _ _ _ _ _ ] # mask replace image onto it # Use torch.vmap for simplicy def sample_merge(): input_ids: [tlen] input_embeds: [tlen, dt] img_embeds: [ilen, ifeat, di] e.g: input_ids: [ a b c d e f X g h i j k X l m ] img_embeds: [3, ifeat, id] # img_embeds has padding """ with torch.no_grad(): num_images, num_image_patches, embed_dim = image_features.shape batch_size, sequence_length = input_ids.shape # left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id)) left_padding = torch.any(attention_mask[:, 0] == 0) # assert not left_padding or batch_size == 1 # 1. Create a mask to know where special image tokens are special_image_token_mask = input_ids == self.config.image_token_index num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1) # Reserve for padding of num_images total_num_special_image_tokens = torch.sum(special_image_token_mask) assert total_num_special_image_tokens == num_images, f'{total_num_special_image_tokens=} != {num_images=} | {image_features.shape} {input_ids}' # Compute the maximum embed dimension max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index) # 2. Compute the positions where text should be written # Calculate new positions for text tokens in merged image-text sequence. # `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens. # `torch.cumsum` computes how each image token shifts subsequent text token positions. # - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one. new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1 nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1] if left_padding: new_token_positions += nb_image_pad[:, None] # offset for left padding text_to_overwrite = new_token_positions[batch_indices, non_image_indices] # 3. Create the full embedding, already padded to the maximum position final_embedding = torch.zeros( batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device ) final_attention_mask = torch.zeros( batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device ) final_labels = None if labels is not None: final_labels = torch.full_like(final_attention_mask, self.config.ignore_index).to(torch.long) # In case the Vision model or the Language model has been offloaded to CPU, we need to manually # set the corresponding tensors into their correct target device. target_device = inputs_embeds.device batch_indices, non_image_indices, text_to_overwrite = ( batch_indices.to(target_device), non_image_indices.to(target_device), text_to_overwrite.to(target_device), ) attention_mask = attention_mask.to(target_device) # 4. Fill the embeddings based on the mask. If we have ["hey" "", "how", "are"] # we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices] final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices] if labels is not None: final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices] # 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling image_to_overwrite = torch.all(final_embedding == 0, dim=-1) # image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device) if left_padding: image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device) else: val = torch.arange(max_embed_dim).unsqueeze(0).to(target_device).expand(batch_size, max_embed_dim) < new_token_positions[:, -1:].to(target_device) image_to_overwrite &= val if image_to_overwrite.sum() != image_features.shape[:-1].numel(): raise ValueError( f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while" f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation." ) final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device) final_attention_mask |= image_to_overwrite position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1) if not left_padding: # Making sure its the same seq_lens = final_attention_mask.sum(-1) for i, (mask, seq_len) in enumerate(zip(final_attention_mask, seq_lens)): # seq_len = mask.sum(-1) assert torch.all(mask[:seq_len] == 1), f'final 1 mask[{i}]: {seq_len} {final_attention_mask.tolist()=}' assert torch.all(mask[seq_len:] == 0), f'final 0 mask[{i}]: {seq_len} {final_attention_mask.tolist()=}' # if DEBUG: # print(f'final_attention_mask=\n{final_attention_mask.tolist()}') # print(f'text_to_overwrite=\n{text_to_overwrite.int().tolist()}') # print(f'image_to_overwrite=\n{image_to_overwrite.int().tolist()}') # print(f'position_ids=\n{position_ids.tolist()}') # print(f'labels=\n{labels.tolist()}') # print(f'final_labels=\n{final_labels.tolist()}') return final_embedding, final_attention_mask, position_ids, final_labels def extract_image_features(self, pixel_values, vision_feature_select_strategy=None): vision_feature_select_strategy = ( vision_feature_select_strategy if vision_feature_select_strategy is not None else self.config.vision_feature_select_strategy ) with (torch.no_grad() if self.freeze_vision_tower else nullcontext()): image_outputs = self.vision_tower(pixel_values) hiddent_states = image_outputs.last_hidden_state image_features = self.multi_modal_projector(hiddent_states) return image_features @add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING) @replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) def forward( self, input_ids: torch.LongTensor = None, pixel_values: torch.FloatTensor = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, vision_feature_layer: Optional[int] = None, vision_feature_select_strategy: Optional[str] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, LlavaCausalLMOutputWithPast]: r""" Args: labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Returns: Example: ```python >>> from PIL import Image >>> import requests >>> from transformers import AutoProcessor, LlavaForConditionalGeneration >>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf") >>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf") >>> prompt = "\nUSER: What's the content of the image?\nASSISTANT:" >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg" >>> image = Image.open(requests.get(url, stream=True).raw) >>> inputs = processor(text=prompt, images=image, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(**inputs, max_length=30) >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "\nUSER: What's the content of the image?\nASSISTANT: The image features a stop sign on a street corner" ```""" 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 vision_feature_layer = ( vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer ) vision_feature_select_strategy = ( vision_feature_select_strategy if vision_feature_select_strategy is not None else self.config.vision_feature_select_strategy ) if inputs_embeds is None: # 1. Extra the input embeddings for_inputs_embeds_ids = input_ids.clone() for_inputs_embeds_ids[(input_ids == self.config.image_token_index)] = 0 # inputs_embeds = self.get_input_embeddings()(input_ids) inputs_embeds = self.get_input_embeddings()(for_inputs_embeds_ids) # 2. Merge text and images if pixel_values is not None and input_ids.shape[1] != 1 and pixel_values.size(0) > 0: num_images = pixel_values.size(0) batch_size, sequence_length = input_ids.shape special_image_token_mask = input_ids == self.config.image_token_index # Reserve for padding of num_images total_num_special_image_tokens = torch.sum(special_image_token_mask) assert num_images == total_num_special_image_tokens, ( f'{num_images} < {total_num_special_image_tokens} | {special_image_token_mask}' ) # pixel_values = pixel_values[:total_num_special_image_tokens] # image_outputs = self.vision_tower(pixel_values, output_hidden_states=True) # with (torch.no_grad() if self.freeze_vision_tower else nullcontext()): # image_outputs = self.vision_tower(pixel_values) # # this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated. # # selected_image_feature = image_outputs.hidden_states[vision_feature_layer] # selected_image_feature = image_outputs.last_hidden_state # if vision_feature_select_strategy == "default": # selected_image_feature = selected_image_feature[:, 1:] # elif vision_feature_select_strategy == "full": # selected_image_feature = selected_image_feature # else: # raise ValueError( # f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}" # ) # image_features = self.multi_modal_projector(selected_image_feature) # print(f"{pixel_values.size()=}") # ! extract_image_features will handle all image features extraction image_features = self.extract_image_features(pixel_values) # if DEBUG: # image_features = image_features[:, :3] inputs_embeds, attention_mask, position_ids, labels = self._merge_input_ids_with_image_features( image_features, inputs_embeds, input_ids, attention_mask, position_ids, labels=labels ) # if labels is None: # # ! this is wrong! # labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long) # print(inputs_embeds.size()) elif pixel_values is not None and input_ids.shape[1] != 1 and pixel_values.size(0) == 0: # there is no images pass else: # In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of # generation with cache # ! (phi) why do we need to do this? # if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1: # # ! it can possible the bug because if mistral, from the first layer_key like this # # ! MUST UNDERSTAND and fix error # # Retrieve the first layer to inspect the logits and mask out the hidden states # # that are set to 0 # first_layer_past_key_value = past_key_values[0][0][:, 0, :, 0] # batch_index, non_attended_tokens = torch.where(first_layer_past_key_value == 0) # # Get the target length # target_seqlen = first_layer_past_key_value.shape[-1] + 1 # extended_attention_mask = torch.ones( # (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]), # dtype=attention_mask.dtype, # device=attention_mask.device, # ) # # print(f'{extended_attention_mask.shape} | {batch_index=} | {non_attended_tokens=}') # # Zero-out the places where we don't need to attend # extended_attention_mask[batch_index, non_attended_tokens] = 0 # attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1) # position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1 # ! fix: https://github.com/huggingface/transformers/blob/c90268de7560c3fef21a927e0bfcf2b611a8711e/src/transformers/models/llava/modeling_llava.py # https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941 if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1: # Retrieve the first layer to inspect the logits and mask out the hidden states # that are set to 0 first_layer_past_key_value = past_key_values[0][0][:, :, :, 0] # Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941 batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0) # Get the target length target_seqlen = first_layer_past_key_value.shape[-1] + 1 extended_attention_mask = torch.ones( (attention_mask.shape[0], target_seqlen - attention_mask.shape[1]), dtype=attention_mask.dtype, device=attention_mask.device, ) # Filter out only the tokens that can be un-attended, this can happen # in the case one uses Llava + Fused modules where the cache on the # first iteration is already big enough, or if one passes custom cache valid_indices = non_attended_tokens < extended_attention_mask.size(-1) new_batch_index = batch_index[valid_indices] new_non_attended_tokens = non_attended_tokens[valid_indices] # Zero-out the places where we don't need to attend extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0 attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1) position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1 outputs = self.language_model( attention_mask=attention_mask, position_ids=position_ids, 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, ) logits = outputs[0] loss = None if labels is not None: # Shift so that tokens < n predict n if attention_mask is not None: shift_attention_mask = attention_mask[..., 1:] shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous() shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous() else: shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = nn.CrossEntropyLoss() loss = loss_fct( shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device) ) if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return LlavaCausalLMOutputWithPast( 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, inputs_embeds=None, pixel_values=None, attention_mask=None, **kwargs ): if past_key_values is not None: if isinstance(past_key_values, Cache): cache_length = past_key_values.get_seq_length() past_length = past_key_values.seen_tokens else: cache_length = past_length = past_key_values[0][0].shape[2] # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where # some of the inputs are exclusivelly passed as part of the cache (e.g. when passing input_embeds as # input) if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard # input_ids based on the past_length. elif past_length < input_ids.shape[1]: input_ids = input_ids[:, past_length:] # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. elif self.config.image_token_index in input_ids: input_ids = input_ids[:, input_ids.shape[1] - 1 :] # If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the # older attention values, as their corresponding values are not part of the input. if cache_length < past_length and attention_mask is not None: attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :] position_ids = kwargs.get("position_ids", None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) if past_key_values: position_ids = position_ids[:, -input_ids.shape[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( { "position_ids": position_ids, "past_key_values": past_key_values, "use_cache": kwargs.get("use_cache"), "attention_mask": attention_mask, "pixel_values": pixel_values, } ) return model_inputs def _reorder_cache(self, *args, **kwargs): return self.language_model._reorder_cache(*args, **kwargs)