Upload mllava/modeling_llava.py with huggingface_hub
Browse files- mllava/modeling_llava.py +770 -0
mllava/modeling_llava.py
ADDED
|
@@ -0,0 +1,770 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# coding=utf-8
|
| 2 |
+
# Copyright 2023 the HuggingFace Inc. team. All rights reserved.
|
| 3 |
+
#
|
| 4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 |
+
# you may not use this file except in compliance with the License.
|
| 6 |
+
# You may obtain a copy of the License at
|
| 7 |
+
#
|
| 8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 9 |
+
#
|
| 10 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 |
+
# See the License for the specific language governing permissions and
|
| 14 |
+
# limitations under the License.
|
| 15 |
+
""" PyTorch Llava model."""
|
| 16 |
+
from dataclasses import dataclass
|
| 17 |
+
from typing import List, Optional, Tuple, Union
|
| 18 |
+
|
| 19 |
+
import torch
|
| 20 |
+
import torch.utils.checkpoint
|
| 21 |
+
from torch import nn
|
| 22 |
+
|
| 23 |
+
# from ... import PreTrainedModel
|
| 24 |
+
# from ...activations import ACT2FN
|
| 25 |
+
# from ...cache_utils import Cache
|
| 26 |
+
# from ...modeling_outputs import ModelOutput
|
| 27 |
+
# from ...utils import (
|
| 28 |
+
# add_start_docstrings,
|
| 29 |
+
# add_start_docstrings_to_model_forward,
|
| 30 |
+
# logging,
|
| 31 |
+
# replace_return_docstrings,
|
| 32 |
+
# )
|
| 33 |
+
# from ..auto import AutoModel, AutoModelForCausalLM
|
| 34 |
+
|
| 35 |
+
from .configuration_llava import LlavaConfig
|
| 36 |
+
|
| 37 |
+
from transformers import PreTrainedModel
|
| 38 |
+
from transformers.activations import ACT2FN
|
| 39 |
+
from transformers.cache_utils import Cache
|
| 40 |
+
from transformers.modeling_outputs import ModelOutput
|
| 41 |
+
from transformers.utils import (
|
| 42 |
+
add_start_docstrings,
|
| 43 |
+
add_start_docstrings_to_model_forward,
|
| 44 |
+
logging,
|
| 45 |
+
replace_return_docstrings,
|
| 46 |
+
)
|
| 47 |
+
from transformers.models.auto import AutoModel, AutoModelForCausalLM
|
| 48 |
+
from .configuration_llava import LlavaConfig
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
logger = logging.get_logger(__name__)
|
| 52 |
+
|
| 53 |
+
_CONFIG_FOR_DOC = "LlavaConfig"
|
| 54 |
+
|
| 55 |
+
LLAVA_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
| 56 |
+
"llava-hf/llava-1.5-7b-hf",
|
| 57 |
+
"llava-hf/llava-1.5-13b-hf",
|
| 58 |
+
"llava-hf/bakLlava-v1-hf",
|
| 59 |
+
# See all Llava models at https://huggingface.co/models?filter=llava
|
| 60 |
+
]
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
@dataclass
|
| 64 |
+
# Copied from transformers.models.idefics.modeling_idefics.IdeficsCausalLMOutputWithPast with Idefics->Llava
|
| 65 |
+
class LlavaCausalLMOutputWithPast(ModelOutput):
|
| 66 |
+
"""
|
| 67 |
+
Base class for Llava causal language model (or autoregressive) outputs.
|
| 68 |
+
|
| 69 |
+
Args:
|
| 70 |
+
loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
|
| 71 |
+
Language modeling loss (for next-token prediction).
|
| 72 |
+
logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
|
| 73 |
+
Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
| 74 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 75 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
| 76 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`)
|
| 77 |
+
|
| 78 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
|
| 79 |
+
`past_key_values` input) to speed up sequential decoding.
|
| 80 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 81 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 82 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 83 |
+
|
| 84 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 85 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
| 86 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
| 87 |
+
sequence_length)`.
|
| 88 |
+
|
| 89 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
| 90 |
+
heads.
|
| 91 |
+
image_hidden_states (`tuple(torch.FloatTensor)`, *optional*):
|
| 92 |
+
Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images,
|
| 93 |
+
sequence_length, hidden_size)`.
|
| 94 |
+
|
| 95 |
+
image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver
|
| 96 |
+
"""
|
| 97 |
+
|
| 98 |
+
loss: Optional[torch.FloatTensor] = None
|
| 99 |
+
logits: torch.FloatTensor = None
|
| 100 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None
|
| 101 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
| 102 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
| 103 |
+
image_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
class LlavaMultiModalProjector(nn.Module):
|
| 107 |
+
def __init__(self, config: LlavaConfig):
|
| 108 |
+
super().__init__()
|
| 109 |
+
|
| 110 |
+
self.linear_1 = nn.Linear(config.vision_config.hidden_size, config.text_config.hidden_size, bias=True)
|
| 111 |
+
self.act = ACT2FN[config.projector_hidden_act]
|
| 112 |
+
self.linear_2 = nn.Linear(config.text_config.hidden_size, config.text_config.hidden_size, bias=True)
|
| 113 |
+
|
| 114 |
+
def forward(self, image_features):
|
| 115 |
+
hidden_states = self.linear_1(image_features)
|
| 116 |
+
hidden_states = self.act(hidden_states)
|
| 117 |
+
hidden_states = self.linear_2(hidden_states)
|
| 118 |
+
return hidden_states
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
LLAVA_START_DOCSTRING = r"""
|
| 122 |
+
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
| 123 |
+
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
| 124 |
+
etc.)
|
| 125 |
+
|
| 126 |
+
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
| 127 |
+
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
| 128 |
+
and behavior.
|
| 129 |
+
|
| 130 |
+
Parameters:
|
| 131 |
+
config ([`LlavaConfig`] or [`LlavaVisionConfig`]):
|
| 132 |
+
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
| 133 |
+
load the weights associated with the model, only the configuration. Check out the
|
| 134 |
+
[`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
| 135 |
+
"""
|
| 136 |
+
|
| 137 |
+
|
| 138 |
+
@add_start_docstrings(
|
| 139 |
+
"The bare LLaMA Model outputting raw hidden-states without any specific head on top.",
|
| 140 |
+
LLAVA_START_DOCSTRING,
|
| 141 |
+
)
|
| 142 |
+
class LlavaPreTrainedModel(PreTrainedModel):
|
| 143 |
+
config_class = LlavaConfig
|
| 144 |
+
base_model_prefix = "model"
|
| 145 |
+
supports_gradient_checkpointing = True
|
| 146 |
+
_no_split_modules = ["LlavaVisionAttention"]
|
| 147 |
+
_skip_keys_device_placement = "past_key_values"
|
| 148 |
+
_supports_flash_attn_2 = True
|
| 149 |
+
|
| 150 |
+
def _init_weights(self, module):
|
| 151 |
+
# important: this ported version of Llava isn't meant for training from scratch - only
|
| 152 |
+
# inference and fine-tuning - so the proper init weights code has been removed - the original codebase
|
| 153 |
+
# https://github.com/haotian-liu/LLaVA/tree/main/llava should serve for that purpose
|
| 154 |
+
std = (
|
| 155 |
+
self.config.initializer_range
|
| 156 |
+
if hasattr(self.config, "initializer_range")
|
| 157 |
+
else self.config.text_config.initializer_range
|
| 158 |
+
)
|
| 159 |
+
|
| 160 |
+
if hasattr(module, "class_embedding"):
|
| 161 |
+
module.class_embedding.data.normal_(mean=0.0, std=std)
|
| 162 |
+
|
| 163 |
+
if isinstance(module, (nn.Linear, nn.Conv2d)):
|
| 164 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 165 |
+
if module.bias is not None:
|
| 166 |
+
module.bias.data.zero_()
|
| 167 |
+
elif isinstance(module, nn.Embedding):
|
| 168 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 169 |
+
if module.padding_idx is not None:
|
| 170 |
+
module.weight.data[module.padding_idx].zero_()
|
| 171 |
+
|
| 172 |
+
@property
|
| 173 |
+
def _supports_sdpa(self):
|
| 174 |
+
"""
|
| 175 |
+
Retrieve language_model's attribute to check whether the model supports
|
| 176 |
+
SDPA or not.
|
| 177 |
+
"""
|
| 178 |
+
return self.language_model._supports_sdpa
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
LLAVA_INPUTS_DOCSTRING = r"""
|
| 182 |
+
Args:
|
| 183 |
+
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
| 184 |
+
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
| 185 |
+
it.
|
| 186 |
+
|
| 187 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 188 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 189 |
+
|
| 190 |
+
[What are input IDs?](../glossary#input-ids)
|
| 191 |
+
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)):
|
| 192 |
+
The tensors corresponding to the input images. Pixel values can be obtained using
|
| 193 |
+
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details ([]`LlavaProcessor`] uses
|
| 194 |
+
[`CLIPImageProcessor`] for processing images).
|
| 195 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 196 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 197 |
+
|
| 198 |
+
- 1 for tokens that are **not masked**,
|
| 199 |
+
- 0 for tokens that are **masked**.
|
| 200 |
+
|
| 201 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 202 |
+
|
| 203 |
+
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
| 204 |
+
[`PreTrainedTokenizer.__call__`] for details.
|
| 205 |
+
|
| 206 |
+
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
| 207 |
+
`past_key_values`).
|
| 208 |
+
|
| 209 |
+
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
| 210 |
+
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
| 211 |
+
information on the default strategy.
|
| 212 |
+
|
| 213 |
+
- 1 indicates the head is **not masked**,
|
| 214 |
+
- 0 indicates the head is **masked**.
|
| 215 |
+
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 216 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
| 217 |
+
config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
|
| 218 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
|
| 219 |
+
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
| 220 |
+
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
| 221 |
+
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
| 222 |
+
|
| 223 |
+
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
| 224 |
+
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
| 225 |
+
|
| 226 |
+
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
| 227 |
+
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
| 228 |
+
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
| 229 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
| 230 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
| 231 |
+
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
| 232 |
+
model's internal embedding lookup matrix.
|
| 233 |
+
use_cache (`bool`, *optional*):
|
| 234 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
| 235 |
+
`past_key_values`).
|
| 236 |
+
output_attentions (`bool`, *optional*):
|
| 237 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
| 238 |
+
tensors for more detail.
|
| 239 |
+
output_hidden_states (`bool`, *optional*):
|
| 240 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
| 241 |
+
more detail.
|
| 242 |
+
return_dict (`bool`, *optional*):
|
| 243 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 244 |
+
"""
|
| 245 |
+
|
| 246 |
+
|
| 247 |
+
@add_start_docstrings(
|
| 248 |
+
"""The LLAVA model which consists of a vision backbone and a language model.""",
|
| 249 |
+
LLAVA_START_DOCSTRING,
|
| 250 |
+
)
|
| 251 |
+
class LlavaForConditionalGeneration(LlavaPreTrainedModel):
|
| 252 |
+
def __init__(self, config: LlavaConfig, vision_tower=None, language_model=None):
|
| 253 |
+
super().__init__(config)
|
| 254 |
+
self.vision_tower = AutoModel.from_config(config.vision_config) if vision_tower is None else vision_tower
|
| 255 |
+
|
| 256 |
+
self.multi_modal_projector = LlavaMultiModalProjector(config)
|
| 257 |
+
self.vocab_size = config.vocab_size
|
| 258 |
+
self.language_model = AutoModelForCausalLM.from_config(
|
| 259 |
+
config.text_config, attn_implementation=config._attn_implementation
|
| 260 |
+
) if language_model is None else language_model
|
| 261 |
+
self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
|
| 262 |
+
self.post_init()
|
| 263 |
+
|
| 264 |
+
def get_input_embeddings(self):
|
| 265 |
+
return self.language_model.get_input_embeddings()
|
| 266 |
+
|
| 267 |
+
def set_input_embeddings(self, value):
|
| 268 |
+
self.language_model.set_input_embeddings(value)
|
| 269 |
+
|
| 270 |
+
def get_output_embeddings(self):
|
| 271 |
+
return self.language_model.get_output_embeddings()
|
| 272 |
+
|
| 273 |
+
def set_output_embeddings(self, new_embeddings):
|
| 274 |
+
self.language_model.set_output_embeddings(new_embeddings)
|
| 275 |
+
|
| 276 |
+
def set_decoder(self, decoder):
|
| 277 |
+
self.language_model.set_decoder(decoder)
|
| 278 |
+
|
| 279 |
+
def get_decoder(self):
|
| 280 |
+
return self.language_model.get_decoder()
|
| 281 |
+
|
| 282 |
+
def tie_weights(self):
|
| 283 |
+
return self.language_model.tie_weights()
|
| 284 |
+
|
| 285 |
+
def resize_token_embeddings(self, new_num_tokens: Optional[int] = None, pad_to_multiple_of=None) -> nn.Embedding:
|
| 286 |
+
model_embeds = self.language_model.resize_token_embeddings(new_num_tokens, pad_to_multiple_of)
|
| 287 |
+
# update vocab size
|
| 288 |
+
self.config.text_config.vocab_size = model_embeds.num_embeddings
|
| 289 |
+
self.config.vocab_size = model_embeds.num_embeddings
|
| 290 |
+
self.vocab_size = model_embeds.num_embeddings
|
| 291 |
+
return model_embeds
|
| 292 |
+
|
| 293 |
+
def _merge_input_ids_with_image_features(self, image_features, inputs_embeds, input_ids, attention_mask, labels):
|
| 294 |
+
num_images, num_image_patches, embed_dim = image_features.shape
|
| 295 |
+
batch_size, sequence_length = input_ids.shape
|
| 296 |
+
left_padding = not torch.sum(input_ids[:, -1] == torch.tensor(self.pad_token_id))
|
| 297 |
+
# 1. Create a mask to know where special image tokens are
|
| 298 |
+
special_image_token_mask = input_ids == self.config.image_token_index
|
| 299 |
+
num_special_image_tokens = torch.sum(special_image_token_mask, dim=-1)
|
| 300 |
+
# Compute the maximum embed dimension
|
| 301 |
+
max_embed_dim = (num_special_image_tokens.max() * (num_image_patches - 1)) + sequence_length
|
| 302 |
+
batch_indices, non_image_indices = torch.where(input_ids != self.config.image_token_index)
|
| 303 |
+
|
| 304 |
+
# 2. Compute the positions where text should be written
|
| 305 |
+
# Calculate new positions for text tokens in merged image-text sequence.
|
| 306 |
+
# `special_image_token_mask` identifies image tokens. Each image token will be replaced by `nb_text_tokens_per_images - 1` text tokens.
|
| 307 |
+
# `torch.cumsum` computes how each image token shifts subsequent text token positions.
|
| 308 |
+
# - 1 to adjust for zero-based indexing, as `cumsum` inherently increases indices by one.
|
| 309 |
+
new_token_positions = torch.cumsum((special_image_token_mask * (num_image_patches - 1) + 1), -1) - 1
|
| 310 |
+
nb_image_pad = max_embed_dim - 1 - new_token_positions[:, -1]
|
| 311 |
+
if left_padding:
|
| 312 |
+
new_token_positions += nb_image_pad[:, None] # offset for left padding
|
| 313 |
+
text_to_overwrite = new_token_positions[batch_indices, non_image_indices]
|
| 314 |
+
|
| 315 |
+
# 3. Create the full embedding, already padded to the maximum position
|
| 316 |
+
final_embedding = torch.zeros(
|
| 317 |
+
batch_size, max_embed_dim, embed_dim, dtype=inputs_embeds.dtype, device=inputs_embeds.device
|
| 318 |
+
)
|
| 319 |
+
final_attention_mask = torch.zeros(
|
| 320 |
+
batch_size, max_embed_dim, dtype=attention_mask.dtype, device=inputs_embeds.device
|
| 321 |
+
)
|
| 322 |
+
if labels is not None:
|
| 323 |
+
final_labels = torch.full(
|
| 324 |
+
(batch_size, max_embed_dim), self.config.ignore_index, dtype=input_ids.dtype, device=input_ids.device
|
| 325 |
+
)
|
| 326 |
+
# In case the Vision model or the Language model has been offloaded to CPU, we need to manually
|
| 327 |
+
# set the corresponding tensors into their correct target device.
|
| 328 |
+
target_device = inputs_embeds.device
|
| 329 |
+
batch_indices, non_image_indices, text_to_overwrite = (
|
| 330 |
+
batch_indices.to(target_device),
|
| 331 |
+
non_image_indices.to(target_device),
|
| 332 |
+
text_to_overwrite.to(target_device),
|
| 333 |
+
)
|
| 334 |
+
attention_mask = attention_mask.to(target_device)
|
| 335 |
+
|
| 336 |
+
# 4. Fill the embeddings based on the mask. If we have ["hey" "<image>", "how", "are"]
|
| 337 |
+
# we need to index copy on [0, 577, 578, 579] for the text and [1:576] for the image features
|
| 338 |
+
final_embedding[batch_indices, text_to_overwrite] = inputs_embeds[batch_indices, non_image_indices]
|
| 339 |
+
final_attention_mask[batch_indices, text_to_overwrite] = attention_mask[batch_indices, non_image_indices]
|
| 340 |
+
if labels is not None:
|
| 341 |
+
final_labels[batch_indices, text_to_overwrite] = labels[batch_indices, non_image_indices]
|
| 342 |
+
|
| 343 |
+
# 5. Fill the embeddings corresponding to the images. Anything that is still zeros needs filling
|
| 344 |
+
image_to_overwrite = torch.all(final_embedding == 0, dim=-1)
|
| 345 |
+
image_to_overwrite &= image_to_overwrite.cumsum(-1) - 1 >= nb_image_pad[:, None].to(target_device)
|
| 346 |
+
|
| 347 |
+
if image_to_overwrite.sum() != image_features.shape[:-1].numel():
|
| 348 |
+
raise ValueError(
|
| 349 |
+
f"The input provided to the model are wrong. The number of image tokens is {torch.sum(special_image_token_mask)} while"
|
| 350 |
+
f" the number of image given to the model is {num_images}. This prevents correct indexing and breaks batch generation."
|
| 351 |
+
)
|
| 352 |
+
|
| 353 |
+
final_embedding[image_to_overwrite] = image_features.contiguous().reshape(-1, embed_dim).to(target_device)
|
| 354 |
+
final_attention_mask |= image_to_overwrite
|
| 355 |
+
position_ids = (final_attention_mask.cumsum(-1) - 1).masked_fill_((final_attention_mask == 0), 1)
|
| 356 |
+
|
| 357 |
+
if labels is None:
|
| 358 |
+
final_labels = None
|
| 359 |
+
|
| 360 |
+
return final_embedding, final_attention_mask, final_labels, position_ids
|
| 361 |
+
|
| 362 |
+
@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
|
| 363 |
+
@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
| 364 |
+
def forward(
|
| 365 |
+
self,
|
| 366 |
+
input_ids: torch.LongTensor = None,
|
| 367 |
+
pixel_values: torch.FloatTensor = None,
|
| 368 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 369 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 370 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 371 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 372 |
+
vision_feature_layer: Optional[int] = None,
|
| 373 |
+
vision_feature_select_strategy: Optional[str] = None,
|
| 374 |
+
labels: Optional[torch.LongTensor] = None,
|
| 375 |
+
use_cache: Optional[bool] = None,
|
| 376 |
+
output_attentions: Optional[bool] = None,
|
| 377 |
+
output_hidden_states: Optional[bool] = None,
|
| 378 |
+
return_dict: Optional[bool] = None,
|
| 379 |
+
) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
|
| 380 |
+
r"""
|
| 381 |
+
Args:
|
| 382 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 383 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 384 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 385 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 386 |
+
|
| 387 |
+
Returns:
|
| 388 |
+
|
| 389 |
+
Example:
|
| 390 |
+
|
| 391 |
+
```python
|
| 392 |
+
>>> from PIL import Image
|
| 393 |
+
>>> import requests
|
| 394 |
+
>>> from transformers import AutoProcessor, LlavaForConditionalGeneration
|
| 395 |
+
|
| 396 |
+
>>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
| 397 |
+
>>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
| 398 |
+
|
| 399 |
+
>>> prompt = "<image>\nUSER: What's the content of the image?\nASSISTANT:"
|
| 400 |
+
>>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
|
| 401 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 402 |
+
|
| 403 |
+
>>> inputs = processor(text=prompt, images=image, return_tensors="pt")
|
| 404 |
+
|
| 405 |
+
>>> # Generate
|
| 406 |
+
>>> generate_ids = model.generate(**inputs, max_length=30)
|
| 407 |
+
>>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 408 |
+
"\nUSER: What's the content of the image?\nASSISTANT: The image features a stop sign on a street corner"
|
| 409 |
+
```"""
|
| 410 |
+
|
| 411 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 412 |
+
output_hidden_states = (
|
| 413 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 414 |
+
)
|
| 415 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 416 |
+
vision_feature_layer = (
|
| 417 |
+
vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
|
| 418 |
+
)
|
| 419 |
+
vision_feature_select_strategy = (
|
| 420 |
+
vision_feature_select_strategy
|
| 421 |
+
if vision_feature_select_strategy is not None
|
| 422 |
+
else self.config.vision_feature_select_strategy
|
| 423 |
+
)
|
| 424 |
+
|
| 425 |
+
if inputs_embeds is None:
|
| 426 |
+
# 1. Extra the input embeddings
|
| 427 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
| 428 |
+
|
| 429 |
+
# 2. Merge text and images
|
| 430 |
+
if pixel_values is not None and input_ids.shape[1] != 1:
|
| 431 |
+
if isinstance(pixel_values, list):
|
| 432 |
+
pixel_values = torch.cat([x for x in pixel_values if x is not None], dim=0)
|
| 433 |
+
# for siglip, need to transform the pixel_values to the right data type
|
| 434 |
+
if pixel_values.dtype != self.vision_tower.dtype:
|
| 435 |
+
pixel_values = pixel_values.type(self.vision_tower.dtype)
|
| 436 |
+
image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
|
| 437 |
+
# this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated.
|
| 438 |
+
selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
|
| 439 |
+
|
| 440 |
+
if vision_feature_select_strategy == "default":
|
| 441 |
+
selected_image_feature = selected_image_feature[:, 1:]
|
| 442 |
+
elif vision_feature_select_strategy == "full":
|
| 443 |
+
selected_image_feature = selected_image_feature
|
| 444 |
+
else:
|
| 445 |
+
raise ValueError(
|
| 446 |
+
f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
|
| 447 |
+
)
|
| 448 |
+
|
| 449 |
+
image_features = self.multi_modal_projector(selected_image_feature)
|
| 450 |
+
inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
|
| 451 |
+
image_features, inputs_embeds, input_ids, attention_mask, labels
|
| 452 |
+
)
|
| 453 |
+
if labels is None:
|
| 454 |
+
labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long)
|
| 455 |
+
else:
|
| 456 |
+
# In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
|
| 457 |
+
# generation with cache
|
| 458 |
+
if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
|
| 459 |
+
# Retrieve the first layer to inspect the logits and mask out the hidden states
|
| 460 |
+
# that are set to 0
|
| 461 |
+
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
|
| 462 |
+
|
| 463 |
+
# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
|
| 464 |
+
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)
|
| 465 |
+
|
| 466 |
+
# Get the target length
|
| 467 |
+
target_seqlen = first_layer_past_key_value.shape[-1] + 1
|
| 468 |
+
|
| 469 |
+
extended_attention_mask = torch.ones(
|
| 470 |
+
(attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),
|
| 471 |
+
dtype=attention_mask.dtype,
|
| 472 |
+
device=attention_mask.device,
|
| 473 |
+
)
|
| 474 |
+
|
| 475 |
+
# Filter out only the tokens that can be un-attended, this can happen
|
| 476 |
+
# if one uses Llava + Fused modules where the cache on the
|
| 477 |
+
# first iteration is already big enough, or if one passes custom cache
|
| 478 |
+
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
|
| 479 |
+
new_batch_index = batch_index[valid_indices]
|
| 480 |
+
new_non_attended_tokens = non_attended_tokens[valid_indices]
|
| 481 |
+
|
| 482 |
+
# Zero-out the places where we don't need to attend
|
| 483 |
+
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
|
| 484 |
+
|
| 485 |
+
attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)
|
| 486 |
+
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
|
| 487 |
+
|
| 488 |
+
outputs = self.language_model(
|
| 489 |
+
attention_mask=attention_mask,
|
| 490 |
+
position_ids=position_ids,
|
| 491 |
+
past_key_values=past_key_values,
|
| 492 |
+
inputs_embeds=inputs_embeds,
|
| 493 |
+
use_cache=use_cache,
|
| 494 |
+
output_attentions=output_attentions,
|
| 495 |
+
output_hidden_states=output_hidden_states,
|
| 496 |
+
return_dict=return_dict,
|
| 497 |
+
)
|
| 498 |
+
|
| 499 |
+
logits = outputs[0]
|
| 500 |
+
|
| 501 |
+
loss = None
|
| 502 |
+
if labels is not None:
|
| 503 |
+
# Shift so that tokens < n predict n
|
| 504 |
+
if attention_mask is not None:
|
| 505 |
+
shift_attention_mask = attention_mask[..., 1:]
|
| 506 |
+
shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
|
| 507 |
+
shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
|
| 508 |
+
else:
|
| 509 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 510 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 511 |
+
# Flatten the tokens
|
| 512 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 513 |
+
loss = loss_fct(
|
| 514 |
+
shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
|
| 515 |
+
)
|
| 516 |
+
|
| 517 |
+
if not return_dict:
|
| 518 |
+
output = (logits,) + outputs[1:]
|
| 519 |
+
return (loss,) + output if loss is not None else output
|
| 520 |
+
|
| 521 |
+
return LlavaCausalLMOutputWithPast(
|
| 522 |
+
loss=loss,
|
| 523 |
+
logits=logits,
|
| 524 |
+
past_key_values=outputs.past_key_values,
|
| 525 |
+
hidden_states=outputs.hidden_states,
|
| 526 |
+
attentions=outputs.attentions,
|
| 527 |
+
)
|
| 528 |
+
|
| 529 |
+
def prepare_inputs_for_generation(
|
| 530 |
+
self, input_ids, past_key_values=None, inputs_embeds=None, pixel_values=None, attention_mask=None, **kwargs
|
| 531 |
+
):
|
| 532 |
+
if past_key_values is not None:
|
| 533 |
+
if isinstance(past_key_values, Cache):
|
| 534 |
+
cache_length = past_key_values.get_seq_length()
|
| 535 |
+
past_length = past_key_values.seen_tokens
|
| 536 |
+
else:
|
| 537 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
| 538 |
+
|
| 539 |
+
# Keep only the unprocessed tokens:
|
| 540 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
| 541 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
| 542 |
+
# input)
|
| 543 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
| 544 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
| 545 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
| 546 |
+
# input_ids based on the past_length.
|
| 547 |
+
elif past_length < input_ids.shape[1]:
|
| 548 |
+
input_ids = input_ids[:, past_length:]
|
| 549 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
| 550 |
+
elif self.config.image_token_index in input_ids:
|
| 551 |
+
input_ids = input_ids[:, input_ids.shape[1] - 1 :]
|
| 552 |
+
# If the cache has seen more tokens than it can hold, then the cache has a size limit. Let's discard the
|
| 553 |
+
# older attention values, as their corresponding values are not part of the input.
|
| 554 |
+
if cache_length < past_length and attention_mask is not None:
|
| 555 |
+
attention_mask = attention_mask[:, -(cache_length + input_ids.shape[1]) :]
|
| 556 |
+
|
| 557 |
+
position_ids = kwargs.get("position_ids", None)
|
| 558 |
+
if attention_mask is not None and position_ids is None:
|
| 559 |
+
# create position_ids on the fly for batch generation
|
| 560 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 561 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 562 |
+
if past_key_values:
|
| 563 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 564 |
+
|
| 565 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 566 |
+
if inputs_embeds is not None and past_key_values is None:
|
| 567 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 568 |
+
else:
|
| 569 |
+
model_inputs = {"input_ids": input_ids}
|
| 570 |
+
|
| 571 |
+
model_inputs.update(
|
| 572 |
+
{
|
| 573 |
+
"position_ids": position_ids,
|
| 574 |
+
"past_key_values": past_key_values,
|
| 575 |
+
"use_cache": kwargs.get("use_cache"),
|
| 576 |
+
"attention_mask": attention_mask,
|
| 577 |
+
"pixel_values": pixel_values,
|
| 578 |
+
}
|
| 579 |
+
)
|
| 580 |
+
return model_inputs
|
| 581 |
+
|
| 582 |
+
def _reorder_cache(self, *args, **kwargs):
|
| 583 |
+
return self.language_model._reorder_cache(*args, **kwargs)
|
| 584 |
+
|
| 585 |
+
|
| 586 |
+
|
| 587 |
+
|
| 588 |
+
from transformers.models.clip.modeling_clip import CLIPEncoderLayer, CLIPEncoder
|
| 589 |
+
@add_start_docstrings(
|
| 590 |
+
"""The MLLAVA model which consists of a vision backbone and a language model.""",
|
| 591 |
+
LLAVA_START_DOCSTRING,
|
| 592 |
+
)
|
| 593 |
+
class MLlavaForConditionalGeneration(LlavaForConditionalGeneration):
|
| 594 |
+
def __init__(self, config: LlavaConfig):
|
| 595 |
+
super().__init__(config)
|
| 596 |
+
config.vision_config.type_vocab_size = 144
|
| 597 |
+
self.image_type_embeddings = nn.Embedding(config.vision_config.type_vocab_size, config.vision_config.hidden_size)
|
| 598 |
+
# self.vision_xatten_layers = nn.ModuleList([CLIPEncoderLayer(config.vision_config) for _ in range(config.vision_config.num_hidden_layers)])
|
| 599 |
+
self.vision_xatten_layers = CLIPEncoder(config.vision_config)
|
| 600 |
+
|
| 601 |
+
|
| 602 |
+
@add_start_docstrings_to_model_forward(LLAVA_INPUTS_DOCSTRING)
|
| 603 |
+
@replace_return_docstrings(output_type=LlavaCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
|
| 604 |
+
def forward(
|
| 605 |
+
self,
|
| 606 |
+
input_ids: torch.LongTensor = None,
|
| 607 |
+
pixel_values: torch.FloatTensor = None,
|
| 608 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 609 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 610 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 611 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 612 |
+
vision_feature_layer: Optional[int] = None,
|
| 613 |
+
vision_feature_select_strategy: Optional[str] = None,
|
| 614 |
+
labels: Optional[torch.LongTensor] = None,
|
| 615 |
+
use_cache: Optional[bool] = None,
|
| 616 |
+
output_attentions: Optional[bool] = None,
|
| 617 |
+
output_hidden_states: Optional[bool] = None,
|
| 618 |
+
return_dict: Optional[bool] = None,
|
| 619 |
+
) -> Union[Tuple, LlavaCausalLMOutputWithPast]:
|
| 620 |
+
r"""
|
| 621 |
+
Args:
|
| 622 |
+
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 623 |
+
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
| 624 |
+
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
| 625 |
+
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
| 626 |
+
|
| 627 |
+
Returns:
|
| 628 |
+
|
| 629 |
+
Example:
|
| 630 |
+
|
| 631 |
+
```python
|
| 632 |
+
>>> from PIL import Image
|
| 633 |
+
>>> import requests
|
| 634 |
+
>>> from transformers import AutoProcessor, LlavaForConditionalGeneration
|
| 635 |
+
|
| 636 |
+
>>> model = LlavaForConditionalGeneration.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
| 637 |
+
>>> processor = AutoProcessor.from_pretrained("llava-hf/llava-1.5-7b-hf")
|
| 638 |
+
|
| 639 |
+
>>> prompt = "<image>\nUSER: What's the content of the image?\nASSISTANT:"
|
| 640 |
+
>>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
|
| 641 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 642 |
+
|
| 643 |
+
>>> inputs = processor(text=prompt, images=image, return_tensors="pt")
|
| 644 |
+
|
| 645 |
+
>>> # Generate
|
| 646 |
+
>>> generate_ids = model.generate(**inputs, max_length=30)
|
| 647 |
+
>>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
| 648 |
+
"\nUSER: What's the content of the image?\nASSISTANT: The image features a stop sign on a street corner"
|
| 649 |
+
```"""
|
| 650 |
+
|
| 651 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 652 |
+
output_hidden_states = (
|
| 653 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 654 |
+
)
|
| 655 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 656 |
+
vision_feature_layer = (
|
| 657 |
+
vision_feature_layer if vision_feature_layer is not None else self.config.vision_feature_layer
|
| 658 |
+
)
|
| 659 |
+
vision_feature_select_strategy = (
|
| 660 |
+
vision_feature_select_strategy
|
| 661 |
+
if vision_feature_select_strategy is not None
|
| 662 |
+
else self.config.vision_feature_select_strategy
|
| 663 |
+
)
|
| 664 |
+
|
| 665 |
+
if inputs_embeds is None:
|
| 666 |
+
# 1. Extra the input embeddings
|
| 667 |
+
inputs_embeds = self.get_input_embeddings()(input_ids)
|
| 668 |
+
|
| 669 |
+
# 2. Merge text and images
|
| 670 |
+
if pixel_values is not None and input_ids.shape[1] != 1:
|
| 671 |
+
image_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
|
| 672 |
+
# this is not memory efficient at all (output_hidden_states=True) will save all the hidden stated.
|
| 673 |
+
selected_image_feature = image_outputs.hidden_states[vision_feature_layer]
|
| 674 |
+
|
| 675 |
+
if vision_feature_select_strategy == "default":
|
| 676 |
+
selected_image_feature = selected_image_feature[:, 1:]
|
| 677 |
+
elif vision_feature_select_strategy == "full":
|
| 678 |
+
selected_image_feature = selected_image_feature
|
| 679 |
+
else:
|
| 680 |
+
raise ValueError(
|
| 681 |
+
f"Unexpected select feature strategy: {self.config.vision_feature_select_strategy}"
|
| 682 |
+
)
|
| 683 |
+
|
| 684 |
+
# added by Dongfu
|
| 685 |
+
num_images, num_image_patches, embed_dim = selected_image_feature.shape
|
| 686 |
+
image_type_embeddings = self.image_type_embeddings(torch.arange(num_images, device=selected_image_feature.device))
|
| 687 |
+
selected_image_feature += image_type_embeddings.unsqueeze(1)
|
| 688 |
+
xatten_output = self.vision_xatten_layers(selected_image_feature, attention_mask=None, causal_attention_mask=None)
|
| 689 |
+
selected_image_feature = xatten_output[0]
|
| 690 |
+
# end of added by Dongfu
|
| 691 |
+
|
| 692 |
+
image_features = self.multi_modal_projector(selected_image_feature)
|
| 693 |
+
inputs_embeds, attention_mask, labels, position_ids = self._merge_input_ids_with_image_features(
|
| 694 |
+
image_features, inputs_embeds, input_ids, attention_mask, labels
|
| 695 |
+
)
|
| 696 |
+
if labels is None:
|
| 697 |
+
labels = torch.full_like(attention_mask, self.config.ignore_index).to(torch.long)
|
| 698 |
+
else:
|
| 699 |
+
# In case input_ids.shape[1] == 1 & pixel_values==None & past_key_values != None, we are in the case of
|
| 700 |
+
# generation with cache
|
| 701 |
+
if past_key_values is not None and pixel_values is not None and input_ids.shape[1] == 1:
|
| 702 |
+
# Retrieve the first layer to inspect the logits and mask out the hidden states
|
| 703 |
+
# that are set to 0
|
| 704 |
+
first_layer_past_key_value = past_key_values[0][0][:, :, :, 0]
|
| 705 |
+
|
| 706 |
+
# Sum all dimensions of head_dim (-2) to avoid random errors such as: https://github.com/huggingface/transformers/pull/28032#issuecomment-1863691941
|
| 707 |
+
batch_index, non_attended_tokens = torch.where(first_layer_past_key_value.float().sum(-2) == 0)
|
| 708 |
+
|
| 709 |
+
# Get the target length
|
| 710 |
+
target_seqlen = first_layer_past_key_value.shape[-1] + 1
|
| 711 |
+
|
| 712 |
+
extended_attention_mask = torch.ones(
|
| 713 |
+
(attention_mask.shape[0], target_seqlen - attention_mask.shape[1]),
|
| 714 |
+
dtype=attention_mask.dtype,
|
| 715 |
+
device=attention_mask.device,
|
| 716 |
+
)
|
| 717 |
+
|
| 718 |
+
# Filter out only the tokens that can be un-attended, this can happen
|
| 719 |
+
# if one uses Llava + Fused modules where the cache on the
|
| 720 |
+
# first iteration is already big enough, or if one passes custom cache
|
| 721 |
+
valid_indices = non_attended_tokens < extended_attention_mask.size(-1)
|
| 722 |
+
new_batch_index = batch_index[valid_indices]
|
| 723 |
+
new_non_attended_tokens = non_attended_tokens[valid_indices]
|
| 724 |
+
|
| 725 |
+
# Zero-out the places where we don't need to attend
|
| 726 |
+
extended_attention_mask[new_batch_index, new_non_attended_tokens] = 0
|
| 727 |
+
|
| 728 |
+
attention_mask = torch.cat((attention_mask, extended_attention_mask), dim=1)
|
| 729 |
+
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
|
| 730 |
+
|
| 731 |
+
outputs = self.language_model(
|
| 732 |
+
attention_mask=attention_mask,
|
| 733 |
+
position_ids=position_ids,
|
| 734 |
+
past_key_values=past_key_values,
|
| 735 |
+
inputs_embeds=inputs_embeds,
|
| 736 |
+
use_cache=use_cache,
|
| 737 |
+
output_attentions=output_attentions,
|
| 738 |
+
output_hidden_states=output_hidden_states,
|
| 739 |
+
return_dict=return_dict,
|
| 740 |
+
)
|
| 741 |
+
|
| 742 |
+
logits = outputs[0]
|
| 743 |
+
|
| 744 |
+
loss = None
|
| 745 |
+
if labels is not None:
|
| 746 |
+
# Shift so that tokens < n predict n
|
| 747 |
+
if attention_mask is not None:
|
| 748 |
+
shift_attention_mask = attention_mask[..., 1:]
|
| 749 |
+
shift_logits = logits[..., :-1, :][shift_attention_mask.to(logits.device) != 0].contiguous()
|
| 750 |
+
shift_labels = labels[..., 1:][shift_attention_mask.to(labels.device) != 0].contiguous()
|
| 751 |
+
else:
|
| 752 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 753 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 754 |
+
# Flatten the tokens
|
| 755 |
+
loss_fct = nn.CrossEntropyLoss()
|
| 756 |
+
loss = loss_fct(
|
| 757 |
+
shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1).to(shift_logits.device)
|
| 758 |
+
)
|
| 759 |
+
|
| 760 |
+
if not return_dict:
|
| 761 |
+
output = (logits,) + outputs[1:]
|
| 762 |
+
return (loss,) + output if loss is not None else output
|
| 763 |
+
|
| 764 |
+
return LlavaCausalLMOutputWithPast(
|
| 765 |
+
loss=loss,
|
| 766 |
+
logits=logits,
|
| 767 |
+
past_key_values=outputs.past_key_values,
|
| 768 |
+
hidden_states=outputs.hidden_states,
|
| 769 |
+
attentions=outputs.attentions,
|
| 770 |
+
)
|