Hugging Face's logo

Transformers documentation

LLaVA-NeXT

You are viewing v4.40.1 version. A newer version v4.46.2 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

LLaVA-NeXT

Overview

The LLaVA-NeXT model was proposed in LLaVA-NeXT: Improved reasoning, OCR, and world knowledge by Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, Yong Jae Lee. LLaVa-NeXT (also called LLaVa-1.6) improves upon LLaVa by increasing the input image resolution and training on an improved visual instruction tuning dataset to improve OCR and common sense reasoning.

The introduction from the blog is the following:

*In October 2023, we released LLaVA-1.5 with a simple and efficient design along with great performance on a benchmark suite of 12 datasets. It has since served as the foundation of many comprehensive studies of data, model, and capabilities of large multimodal models (LMM), and has enabled various new applications.

Today, we are thrilled to present LLaVA-NeXT, with improved reasoning, OCR, and world knowledge. LLaVA-NeXT even exceeds Gemini Pro on several benchmarks.

Compared with LLaVA-1.5, LLaVA-NeXT has several improvements:

Increasing the input image resolution to 4x more pixels. This allows it to grasp more visual details. It supports three aspect ratios, up to 672x672, 336x1344, 1344x336 resolution. Better visual reasoning and OCR capability with an improved visual instruction tuning data mixture. Better visual conversation for more scenarios, covering different applications. Better world knowledge and logical reasoning. Efficient deployment and inference with SGLang. Along with performance improvements, LLaVA-NeXT maintains the minimalist design and data efficiency of LLaVA-1.5. It re-uses the pretrained connector of LLaVA-1.5, and still uses less than 1M visual instruction tuning samples. The largest 34B variant finishes training in ~1 day with 32 A100s.*

drawing LLaVa-NeXT incorporates a higher input resolution by encoding various patches of the input image. Taken from the original paper.

This model was contributed by nielsr. The original code can be found here.

Usage tips

  • We advise users to use padding_side="left" when computing batched generation as it leads to more accurate results. Simply make sure to call processor.tokenizer.padding_side = "left" before generating.

  • Note that each checkpoint has been trained with a specific prompt format, depending on which large language model (LLM) was used. Below, we list the correct prompt formats to use for the text prompt “What is shown in this image?“:

llava-v1.6-mistral-7b-hf requires the following format:

"[INST] <image>\nWhat is shown in this image? [/INST]"

llava-v1.6-vicuna-7b-hf and llava-v1.6-vicuna-13b-hf require the following format:

"A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions. USER: <image>\nWhat is shown in this image? ASSISTANT:"

llava-v1.6-34b-hf requires the following format:

"<|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nWhat is shown in this image?<|im_end|><|im_start|>assistant\n"

Usage example

Here’s how to load the model and perform inference in half-precision (torch.float16):

from transformers import LlavaNextProcessor, LlavaNextForConditionalGeneration
import torch
from PIL import Image
import requests

processor = LlavaNextProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")

model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", torch_dtype=torch.float16, low_cpu_mem_usage=True) 
model.to("cuda:0")

# prepare image and text prompt, using the appropriate prompt template
url = "https://github.com/haotian-liu/LLaVA/blob/1a91fc274d7c35a9b50b3cb29c4247ae5837ce39/images/llava_v1_5_radar.jpg?raw=true"
image = Image.open(requests.get(url, stream=True).raw)
prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"

inputs = processor(prompt, image, return_tensors="pt").to("cuda:0")

# autoregressively complete prompt
output = model.generate(**inputs, max_new_tokens=100)

print(processor.decode(output[0], skip_special_tokens=True))

Model optimization

Quantization using Bitsandbytes

The model can be loaded in 8 or 4 bits, greatly reducing the memory requirements while maintaining the performance of the original model. First make sure to install bitsandbytes, `pip install bitsandbytes“ and make sure to have access to a CUDA compatible GPU device. Simply change the snippet above with:

from transformers import LlavaNextForConditionalGeneration, BitsAndBytesConfig

# specify how to quantize the model
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf", quantization_config=quantization_config, device_map="auto")

Use Flash-Attention 2 to further speed-up generation

First make sure to install flash-attn. Refer to the original repository of Flash Attention regarding that package installation. Simply change the snippet above with:

from transformers import LlavaNextForConditionalGeneration

model = LlavaNextForConditionalGeneration.from_pretrained(
    model_id, 
    torch_dtype=torch.float16, 
    low_cpu_mem_usage=True,
    use_flash_attention_2=True
).to(0)

LlavaNextConfig

class transformers.LlavaNextConfig

< >

( vision_config = None text_config = None ignore_index = -100 image_token_index = 32000 projector_hidden_act = 'gelu' vision_feature_select_strategy = 'default' vision_feature_layer = -2 image_grid_pinpoints = None **kwargs )

Parameters

  • vision_config (Union[AutoConfig, dict], optional, defaults to CLIPVisionConfig) — The config object or dictionary of the vision backbone.
  • text_config (Union[AutoConfig, dict], optional, defaults to LlamaConfig) — The config object or dictionary of the text backbone.
  • ignore_index (int, optional, defaults to -100) — The ignore index for the loss function.
  • image_token_index (int, optional, defaults to 32000) — The image token index to encode the image prompt.
  • projector_hidden_act (str, optional, defaults to "gelu") — The activation function used by the multimodal projector.
  • vision_feature_select_strategy (str, optional, defaults to "default") — The feature selection strategy used to select the vision feature from the vision backbone. Can be one of "default" or "full". If "default", the CLS token is removed from the vision features. If "full", the full vision features are used.
  • vision_feature_layer (int, optional, defaults to -2) — The index of the layer to select the vision feature.
  • image_grid_pinpoints (List, optional, defaults to [[336, 672], [672, 336], [672, 672], [1008, 336], [336, 1008]]) — A list of possible resolutions to use for processing high resolution images. Each item in the list should be a tuple or list of the form (height, width).

This is the configuration class to store the configuration of a LlavaNextForConditionalGeneration. It is used to instantiate an Llava-NeXT model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the llava-hf/llava-v1.6-mistral-7b-hf model.

Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

Example:

>>> from transformers import LlavaNextForConditionalGeneration, LlavaNextConfig, CLIPVisionConfig, LlamaConfig

>>> # Initializing a CLIP-vision config
>>> vision_config = CLIPVisionConfig()

>>> # Initializing a Llama config
>>> text_config = LlamaConfig()

>>> # Initializing a Llava-Next llava-hf/llava-v1.6-mistral-7b-hf style configuration
>>> configuration = LlavaNextConfig(vision_config, text_config)

>>> # Initializing a model from the llava-hf/llava-v1.6-mistral-7b-hf style configuration
>>> model = LlavaNextForConditionalGeneration(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

LlavaNextImageProcessor

class transformers.LlavaNextImageProcessor

< >

( do_resize: bool = True size: Dict = None image_grid_pinpoints: List = None resample: Resampling = <Resampling.BICUBIC: 3> do_center_crop: bool = True crop_size: Dict = None do_rescale: bool = True rescale_factor: Union = 0.00392156862745098 do_normalize: bool = True image_mean: Union = None image_std: Union = None do_convert_rgb: bool = True **kwargs )

Parameters

  • do_resize (bool, optional, defaults to True) — Whether to resize the image’s (height, width) dimensions to the specified size. Can be overridden by do_resize in the preprocess method.
  • size (Dict[str, int] optional, defaults to {"shortest_edge" -- 224}): Size of the image after resizing. The shortest edge of the image is resized to size[“shortest_edge”], with the longest edge resized to keep the input aspect ratio. Can be overridden by size in the preprocess method.
  • image_grid_pinpoints (List optional, defaults to [[672, 336], [336, 672], [672, 672], [336, 1008], [1008, 336]]) — A list of possible resolutions to use for processing high resolution images. The best resolution is selected based on the original size of the image. Can be overridden by image_grid_pinpoints in the preprocess method.
  • resample (PILImageResampling, optional, defaults to Resampling.BICUBIC) — Resampling filter to use if resizing the image. Can be overridden by resample in the preprocess method.
  • do_center_crop (bool, optional, defaults to True) — Whether to center crop the image to the specified crop_size. Can be overridden by do_center_crop in the preprocess method.
  • crop_size (Dict[str, int] optional, defaults to 224) — Size of the output image after applying center_crop. Can be overridden by crop_size in the preprocess method.
  • do_rescale (bool, optional, defaults to True) — Whether to rescale the image by the specified scale rescale_factor. Can be overridden by do_rescale in the preprocess method.
  • rescale_factor (int or float, optional, defaults to 1/255) — Scale factor to use if rescaling the image. Can be overridden by rescale_factor in the preprocess method.
  • do_normalize (bool, optional, defaults to True) — Whether to normalize the image. Can be overridden by do_normalize in the preprocess method.
  • image_mean (float or List[float], optional, defaults to [0.48145466, 0.4578275, 0.40821073]) — Mean to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by the image_mean parameter in the preprocess method.
  • image_std (float or List[float], optional, defaults to [0.26862954, 0.26130258, 0.27577711]) — Standard deviation to use if normalizing the image. This is a float or list of floats the length of the number of channels in the image. Can be overridden by the image_std parameter in the preprocess method. Can be overridden by the image_std parameter in the preprocess method.
  • do_convert_rgb (bool, optional, defaults to True) — Whether to convert the image to RGB.

Constructs a LLaVa-NeXT image processor. Based on CLIPImageProcessor with incorporation of additional techniques for processing high resolution images as explained in the LLaVa paper.

preprocess

< >

( images: Union do_resize: bool = None size: Dict = None image_grid_pinpoints: List = None resample: Resampling = None do_center_crop: bool = None crop_size: int = None do_rescale: bool = None rescale_factor: float = None do_normalize: bool = None image_mean: Union = None image_std: Union = None do_convert_rgb: bool = None return_tensors: Union = None data_format: Optional = <ChannelDimension.FIRST: 'channels_first'> input_data_format: Union = None )

Parameters

  • images (ImageInput) — Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If passing in images with pixel values between 0 and 1, set do_rescale=False.
  • do_resize (bool, optional, defaults to self.do_resize) — Whether to resize the image.
  • size (Dict[str, int], optional, defaults to self.size) — Size of the image after resizing. Shortest edge of the image is resized to size[“shortest_edge”], with the longest edge resized to keep the input aspect ratio.
  • image_grid_pinpoints (List optional, defaults to self.image_grid_pinpoints) — A list of possible resolutions to use for processing high resolution images. The best resolution is selected based on the original size of the image.
  • resample (int, optional, defaults to self.resample) — Resampling filter to use if resizing the image. This can be one of the enum PILImageResampling. Only has an effect if do_resize is set to True.
  • do_center_crop (bool, optional, defaults to self.do_center_crop) — Whether to center crop the image.
  • crop_size (Dict[str, int], optional, defaults to self.crop_size) — Size of the center crop. Only has an effect if do_center_crop is set to True.
  • do_rescale (bool, optional, defaults to self.do_rescale) — Whether to rescale the image.
  • rescale_factor (float, optional, defaults to self.rescale_factor) — Rescale factor to rescale the image by if do_rescale is set to True.
  • do_normalize (bool, optional, defaults to self.do_normalize) — Whether to normalize the image.
  • image_mean (float or List[float], optional, defaults to self.image_mean) — Image mean to use for normalization. Only has an effect if do_normalize is set to True.
  • image_std (float or List[float], optional, defaults to self.image_std) — Image standard deviation to use for normalization. Only has an effect if do_normalize is set to True.
  • do_convert_rgb (bool, optional, defaults to self.do_convert_rgb) — Whether to convert the image to RGB.
  • return_tensors (str or TensorType, optional) — The type of tensors to return. Can be one of:
    • Unset: Return a list of np.ndarray.
    • TensorType.TENSORFLOW or 'tf': Return a batch of type tf.Tensor.
    • TensorType.PYTORCH or 'pt': Return a batch of type torch.Tensor.
    • TensorType.NUMPY or 'np': Return a batch of type np.ndarray.
    • TensorType.JAX or 'jax': Return a batch of type jax.numpy.ndarray.
  • data_format (ChannelDimension or str, optional, defaults to ChannelDimension.FIRST) — The channel dimension format for the output image. Can be one of:
    • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.
    • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.
    • Unset: Use the channel dimension format of the input image.
  • input_data_format (ChannelDimension or str, optional) — The channel dimension format for the input image. If unset, the channel dimension format is inferred from the input image. Can be one of:
    • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.
    • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.
    • "none" or ChannelDimension.NONE: image in (height, width) format.

LlavaNextProcessor

class transformers.LlavaNextProcessor

< >

( image_processor = None tokenizer = None )

Parameters

Constructs a LLaVa-NeXT processor which wraps a LLaVa-NeXT image processor and a LLaMa tokenizer into a single processor.

LlavaNextProcessor offers all the functionalities of LlavaNextImageProcessor and LlamaTokenizerFast. See the __call__() and decode() for more information.

batch_decode

< >

( *args **kwargs )

This method forwards all its arguments to LlamaTokenizerFast’s batch_decode(). Please refer to the docstring of this method for more information.

decode

< >

( *args **kwargs )

This method forwards all its arguments to LlamaTokenizerFast’s decode(). Please refer to the docstring of this method for more information.

LlavaNextForConditionalGeneration

class transformers.LlavaNextForConditionalGeneration

< >

( config: LlavaNextConfig )

Parameters

  • config (LlavaNextConfig or LlavaNextVisionConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The LLAVA-NeXT model which consists of a vision backbone and a language model. This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

< >

( input_ids: LongTensor = None pixel_values: FloatTensor = None image_sizes: Optional = None attention_mask: Optional = None position_ids: Optional = None past_key_values: Optional = None inputs_embeds: Optional = None vision_feature_layer: Optional = None vision_feature_select_strategy: Optional = None labels: Optional = None use_cache: Optional = None output_attentions: Optional = None output_hidden_states: Optional = None return_dict: Optional = None ) → transformers.models.llava_next.modeling_llava_next.LlavaNextCausalLMOutputWithPast or tuple(torch.FloatTensor)

Parameters

  • input_ids (torch.LongTensor of shape (batch_size, sequence_length)) — Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it.

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    What are input IDs?

  • pixel_values (torch.FloatTensor of shape `(batch_size, num_channels, image_size, image_size)) — The tensors corresponding to the input images. Pixel values can be obtained using AutoImageProcessor. See LlavaNextImageProcessor.call() for details. LlavaProcessor uses LlavaNextImageProcessor for processing images.
  • image_sizes (torch.LongTensor of shape (batch_size, 2), optional) — The sizes of the images in the batch, being (height, width) for each image.
  • 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?

    Indices can be obtained using AutoTokenizer. See PreTrainedTokenizer.encode() and PreTrainedTokenizer.call() for details.

    If past_key_values is used, optionally only the last decoder_input_ids have to be input (see past_key_values).

    If you want to change padding behavior, you should read modeling_opt._prepare_decoder_attention_mask and modify to your needs. See diagram 1 in the paper for more information on the default strategy.

    • 1 indicates the head is not masked,
    • 0 indicates the head is masked.
  • position_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) — Indices of positions of each input sequence tokens in the position embeddings. Selected in the range [0, config.n_positions - 1]. What are position IDs?
  • past_key_values (tuple(tuple(torch.FloatTensor)), optional, returned when use_cache=True is passed or when config.use_cache=True) — Tuple of tuple(torch.FloatTensor) of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head)) and 2 additional tensors of shape (batch_size, num_heads, encoder_sequence_length, embed_size_per_head).

    Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

    If past_key_values are used, the user can optionally input only the last decoder_input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, 1) instead of all decoder_input_ids of shape (batch_size, sequence_length).

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — 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.
  • vision_feature_layer (int, optional, defaults to -2) — The index of the layer to select the vision feature.
  • vision_feature_select_strategy (str, optional, defaults to "default") — The feature selection strategy used to select the vision feature from the vision backbone. Can be one of "default" or "full". If "default", the CLS token is removed from the vision features. If "full", the full vision features are used.
  • use_cache (bool, optional) — If set to True, past_key_values key value states are returned and can be used to speed up decoding (see past_key_values).
  • 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 ModelOutput instead of a plain tuple.

    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

transformers.models.llava_next.modeling_llava_next.LlavaNextCausalLMOutputWithPast or tuple(torch.FloatTensor)

A transformers.models.llava_next.modeling_llava_next.LlavaNextCausalLMOutputWithPast or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (LlavaNextConfig) and inputs.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels is provided) — Language modeling loss (for next-token prediction).

  • logits (torch.FloatTensor of shape (batch_size, sequence_length, config.vocab_size)) — Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).

  • past_key_values (tuple(tuple(torch.FloatTensor)), optional, returned when use_cache=True is passed or when config.use_cache=True) — Tuple of tuple(torch.FloatTensor) of length config.n_layers, with each tuple having 2 tensors of shape (batch_size, num_heads, sequence_length, embed_size_per_head))

    Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see past_key_values input) to speed up sequential decoding.

  • hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings, if the model has an embedding layer, + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size).

    Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.

  • attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length).

    Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.

  • image_hidden_states (tuple(torch.FloatTensor), optional) — Tuple of torch.FloatTensor (one for the output of the image embeddings, (batch_size, num_images, sequence_length, hidden_size).

    image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver

The LlavaNextForConditionalGeneration forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Example:

>>> from PIL import Image
>>> import requests
>>> from transformers import AutoProcessor, LlavaNextForConditionalGeneration

>>> model = LlavaNextForConditionalGeneration.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")
>>> processor = AutoProcessor.from_pretrained("llava-hf/llava-v1.6-mistral-7b-hf")

>>> prompt = "[INST] <image>\nWhat is shown in this image? [/INST]"
>>> 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]
"[INST]  \nWhat is shown in this image? [/INST] The image appears to be a radar chart, which is a type of multi-dimensional plot (...)"
< > Update on GitHub

←Llava LXMERT→