Hugging Face's logo

Transformers documentation

FastVLM

You are viewing main version, which requires installation from source. If you'd like regular pip install, checkout the latest stable version (v5.0.0rc0).
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

This model was released on 2025-05-06 and added to Hugging Face Transformers on 2025-12-04.

FastVLM

PyTorch FlashAttention SDPA

Overview

FastVLM is an open-source vision-language model featuring a novel hybrid vision encoder, FastViTHD. Leveraging reparameterizable convolutional layers, scaled input resolution, and a reduced number of visual tokens, FastVLM delivers high accuracy with exceptional efficiency. Its optimized architecture enables deployment even on edge devices, achieving ultra-low TTFT (time to first token) without sacrificing performance.

The model was proposed in FastVLM: Efficient Vision Encoding for Vision Language Models by Pavan Kumar Anasosalu Vasu, Fartash Faghri, Chun-Liang Li, Cem Koc, Nate True, Albert Antony, Gokul Santhanam, James Gabriel, Peter Grasch, Oncel Tuzel and Hadi Pouransari.

The abstract from the paper is the following:

Scaling the input image resolution is essential for enhancing the performance of Vision Language Models (VLMs), particularly in text-rich image understanding tasks. However, popular visual encoders such as ViTs become inefficient at high resolutions due to the large number of tokens and high encoding latency. At different operational resolutions, the vision encoder of a VLM can be optimized along two axes: reducing encoding latency and minimizing the number of visual tokens passed to the LLM, thereby lowering overall latency. Based on a comprehensive efficiency analysis of the interplay between image resolution, vision latency, token count, and LLM size, we introduce FastVLM—a model that achieves an optimized trade-off between resolution, latency, and accuracy. FastVLM incorporates FastViTHD, a novel hybrid vision encoder designed to output fewer tokens and significantly reduce encoding time for high-resolution images. Unlike previous methods, FastVLM achieves the optimal balance between visual token count and image resolution solely by scaling the input image, eliminating the need for additional token pruning and simplifying the model design. In the LLaVA-1.5 setup, FastVLM achieves 3.2× improvement in time-to-first-token (TTFT) while maintaining similar performance on VLM benchmarks compared to prior works. Compared to LLaVa-OneVision at the highest resolution (1152×1152), FastVLM achieves better performance on key benchmarks like SeedBench, MMMU and DocVQA, using the same 0.5B LLM, but with 85× faster TTFT and a vision encoder that is 3.4× smaller.

This model was contributed by Kamila. 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 the model has not been explicitly trained to process multiple images in the same prompt, although this is technically possible, you may experience inaccurate results.

Important:

Hugging Face models use SDPA by default; however, this model’s visual backbone supports only eager attention, so it automatically falls back to "eager".

If you want to use a different attention implementation in the language decoder, make sure to set it explicitly, for example:

model = FastVlmForConditionalGeneration.from_pretrained("KamilaMila/FastVLM-0.5B", attn_implementation={"text_config": "flash_attention_2"})

Setting it for the entire model, e.g.

model = FastVlmForConditionalGeneration.from_pretrained("KamilaMila/FastVLM-0.5B", attn_implementation="flash_attention_2")

will result in an error.

Formatting Prompts with Chat Templates

Each checkpoint is trained with a specific prompt format, depending on the underlying large language model backbone. To ensure correct formatting, use the processor’s apply_chat_template method.

Important:

  • You must construct a conversation history — passing a plain string won’t work.
  • Each message should be a dictionary with "role" and "content" keys.
  • The "content" should be a list of dictionaries for different modalities like "text" and "image".

Usage examples

Single input inference

import torch
from transformers import AutoProcessor, FastVlmForConditionalGeneration

# Load the model in half-precision
model = FastVlmForConditionalGeneration.from_pretrained("KamilaMila/FastVLM-0.5B", dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained("KamilaMila/FastVLM-0.5B")

conversation = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
            {"type": "text", "text": "What is shown in this image?"},
        ],
    },
]

inputs = processor.apply_chat_template(
    conversation,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device, torch.bfloat16)

# Generate
generate_ids = model.generate(**inputs, max_new_tokens=30)
processor.batch_decode(generate_ids, skip_special_tokens=True)

Batched inference

FastVLM also supports batched inference. Here is how you can do it:

import torch
from transformers import AutoProcessor, FastVlmForConditionalGeneration

# Load the model in half-precision
model = FastVlmForConditionalGeneration.from_pretrained("KamilaMila/FastVLM-0.5B", dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained("KamilaMila/FastVLM-0.5B")


# Prepare a batch of two prompts
conversation_1 = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
            {"type": "text", "text": "What is shown in this image?"},
        ],
    },
]

conversation_2 = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "http://images.cocodataset.org/val2017/000000039769.jpg"},
            {"type": "text", "text": "What is shown in this image?"},
        ],
    },
]

inputs = processor.apply_chat_template(
    [conversation_1, conversation_2],
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    padding=True,
    return_tensors="pt"
).to(model.device, torch.bfloat16)


# Generate
generate_ids = model.generate(**inputs, max_new_tokens=30)
processor.batch_decode(generate_ids, skip_special_tokens=True)

Note regarding reproducing original implementation

In order to match the logits of the original implementation, one needs to use float32. In half precision the logit difference is higher due to tiny differences in how some ops are implemented in timm.

Using Flash Attention 2

Flash Attention 2 is an even faster, optimized version of the previous optimization, please refer to the Flash Attention 2 section of performance docs.

FastVlmConfig

class transformers.FastVlmConfig

< >

( vision_config = None text_config = None image_token_id = 151646 projector_hidden_act = 'gelu' vision_feature_select_strategy = 'full' vision_feature_layer = -1 multimodal_projector_bias = True **kwargs )

Parameters

  • vision_config (Union[AutoConfig, dict], optional, defaults to TimmWrapperConfig for fastvit_mci3) — The config object or dictionary of the vision backbone.
  • text_config (Union[AutoConfig, dict], optional, defaults to Qwen2Config) — The config object or dictionary of the text backbone.
  • image_token_id (int, optional, defaults to 151646) — 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 "full") — The feature selection strategy used to select the vision feature from the vision backbone. Only “full” supported.
  • vision_feature_layer (Union[int, list[int]], optional, defaults to -1) — The index of the layer to select the vision feature. If multiple indices are provided, the vision feature of the corresponding indices will be concatenated to form the vision features. Only -1 supported.
  • multimodal_projector_bias (bool, optional, defaults to True) — Whether to use bias in the multimodal projector.

This is the configuration class to store the configuration of a FastVlmForConditionalGeneration. It is used to instantiate a FastVLM model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield the same configuration as the one of FastVLM-7B.

e.g. KamilaMila/FastVLM-7B

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 FastVlmForConditionalGeneration, FastVlmConfig

>>> # Initializing a FastVLM-7B style configuration
>>> configuration = FastVlmConfig()

>>> # Initializing a model from the FastVLM-7B style configuration
>>> model = FastVlmForConditionalGeneration(configuration)

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

FastVlmModel

class transformers.FastVlmModel

< >

( config: FastVlmConfig )

Parameters

  • config (FastVlmConfig) — 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 FastVlm model which consists of a vision backbone and a language model, without a language modeling head.

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: typing.Optional[torch.LongTensor] = None pixel_values: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.LongTensor] = None past_key_values: typing.Optional[transformers.cache_utils.Cache] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None vision_feature_layer: typing.Union[int, list[int], NoneType] = None vision_feature_select_strategy: typing.Optional[str] = None cache_position: typing.Optional[torch.LongTensor] = None image_sizes: typing.Optional[torch.Tensor] = None **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] ) → transformers.models.fast_vlm.modeling_fast_vlm.FastVlmModelOutputWithPast or tuple(torch.FloatTensor)

Parameters

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

    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), optional) — The tensors corresponding to the input images. Pixel values can be obtained using image_processor_class. See image_processor_class.__call__ for details (processor_class uses image_processor_class for processing images).
  • 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?

  • 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 (~cache_utils.Cache, optional) — Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

    Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

    The model will output the same cache format that is fed as input.

    If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all 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 (Union[int, list[int], NoneType], optional) — The index of the layer to select the vision feature. If multiple indices are provided, the vision feature of the corresponding indices will be concatenated to form the vision features. Only -1 supported.
  • vision_feature_select_strategy (str, optional) — The feature selection strategy used to select the vision feature from the vision backbone. Only “full” supported.
  • cache_position (torch.LongTensor of shape (sequence_length), optional) — Indices depicting the position of the input sequence tokens in the sequence. Contrarily to position_ids, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length.
  • image_sizes (torch.Tensor of shape (batch_size, 2), optional) — The sizes of the images in the batch, being (height, width) for each image.

Returns

transformers.models.fast_vlm.modeling_fast_vlm.FastVlmModelOutputWithPast or tuple(torch.FloatTensor)

A transformers.models.fast_vlm.modeling_fast_vlm.FastVlmModelOutputWithPast 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 (FastVlmConfig) and inputs.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the model.

  • past_key_values (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) — It is a Cache instance. For more details, see our kv cache guide.

    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 (torch.FloatTensor, optional) — A torch.FloatTensor of size (batch_size, num_images, sequence_length, hidden_size). image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.

The FastVlmModel 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.

get_image_features

< >

( pixel_values: FloatTensor vision_feature_layer: typing.Union[int, list[int], NoneType] = None vision_feature_select_strategy: typing.Optional[str] = None **kwargs ) → image_features (torch.Tensor)

Parameters

  • pixel_values (torch.FloatTensor] of shape (batch_size, channels, height, width)) — The tensors corresponding to the input images.
  • vision_feature_layer (Union[int, list[int]], optional) — The index/indices of the layer to select the vision feature. Only -1 supported.
  • vision_feature_select_strategy (str, optional) — The feature selection strategy used to select the vision feature from the vision backbone. Only “full” supported.

Returns

image_features (torch.Tensor)

Image feature tensor of shape (num_images, image_length, embed_dim)).

Obtains image last hidden states from the vision tower and apply multimodal projection.

get_placeholder_mask

< >

( input_ids: LongTensor inputs_embeds: FloatTensor image_features: FloatTensor )

Obtains multimodal placeholder mask from input_ids or inputs_embeds, and checks that the placeholder token count is equal to the length of multimodal features. If the lengths are different, an error is raised.

FastVlmForConditionalGeneration

class transformers.FastVlmForConditionalGeneration

< >

( config: FastVlmConfig )

Parameters

  • config (FastVlmConfig) — 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 FastVlm 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: typing.Optional[torch.LongTensor] = None pixel_values: typing.Optional[torch.FloatTensor] = None attention_mask: typing.Optional[torch.Tensor] = None position_ids: typing.Optional[torch.LongTensor] = None past_key_values: typing.Optional[transformers.cache_utils.Cache] = None inputs_embeds: typing.Optional[torch.FloatTensor] = None vision_feature_layer: typing.Union[int, list[int], NoneType] = None vision_feature_select_strategy: typing.Optional[str] = None labels: typing.Optional[torch.LongTensor] = None cache_position: typing.Optional[torch.LongTensor] = None logits_to_keep: typing.Union[int, torch.Tensor] = 0 image_sizes: typing.Optional[torch.Tensor] = None **kwargs: typing_extensions.Unpack[transformers.utils.generic.TransformersKwargs] ) → transformers.models.fast_vlm.modeling_fast_vlm.FastVlmCausalLMOutputWithPast or tuple(torch.FloatTensor)

Parameters

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

    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), optional) — The tensors corresponding to the input images. Pixel values can be obtained using image_processor_class. See image_processor_class.__call__ for details (processor_class uses image_processor_class for processing images).
  • 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?

  • 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 (~cache_utils.Cache, optional) — Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used to speed up sequential decoding. This typically consists in the past_key_values returned by the model at a previous stage of decoding, when use_cache=True or config.use_cache=True.

    Only Cache instance is allowed as input, see our kv cache guide. If no past_key_values are passed, DynamicCache will be initialized by default.

    The model will output the same cache format that is fed as input.

    If past_key_values are used, the user is expected to input only unprocessed input_ids (those that don’t have their past key value states given to this model) of shape (batch_size, unprocessed_length) instead of all 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 (Union[int, list[int], NoneType], optional) — The index of the layer to select the vision feature. If multiple indices are provided, the vision feature of the corresponding indices will be concatenated to form the vision features. Only -1 supported.
  • vision_feature_select_strategy (str, optional) — The feature selection strategy used to select the vision feature from the vision backbone. Only “full” supported.
  • 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].
  • cache_position (torch.LongTensor of shape (sequence_length), optional) — Indices depicting the position of the input sequence tokens in the sequence. Contrarily to position_ids, this tensor is not affected by padding. It is used to update the cache in the correct position and to infer the complete sequence length.
  • logits_to_keep (Union[int, torch.Tensor], defaults to 0) — If an int, compute logits for the last logits_to_keep tokens. If 0, calculate logits for all input_ids (special case). Only last token logits are needed for generation, and calculating them only for that token can save memory, which becomes pretty significant for long sequences or large vocabulary size. If a torch.Tensor, must be 1D corresponding to the indices to keep in the sequence length dimension. This is useful when using packed tensor format (single dimension for batch and sequence length).
  • image_sizes (torch.Tensor of shape (batch_size, 2), optional) — The sizes of the images in the batch, being (height, width) for each image.

Returns

transformers.models.fast_vlm.modeling_fast_vlm.FastVlmCausalLMOutputWithPast or tuple(torch.FloatTensor)

A transformers.models.fast_vlm.modeling_fast_vlm.FastVlmCausalLMOutputWithPast 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 (FastVlmConfig) 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 (Cache, optional, returned when use_cache=True is passed or when config.use_cache=True) — It is a Cache instance. For more details, see our kv cache guide.

    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 (torch.FloatTensor, optional) — A torch.FloatTensor of size (batch_size, num_images, sequence_length, hidden_size). image_hidden_states of the model produced by the vision encoder and after projecting the last hidden state.

The FastVlmForConditionalGeneration 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, AutoModelForImageTextToText
>>> import torch

>>> device = "cuda" if torch.cuda.is_available() else "cpu"

>>> model = AutoModelForImageTextToText.from_pretrained("KamilaMila/FastVLM-0.5B").to(device)
>>> processor = AutoProcessor.from_pretrained("KamilaMila/FastVLM-0.5B")

>>> conversation = [
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "What are these?"},
                {"type": "image"}
            ]
        }
    ]

>>> prompt = processor.apply_chat_template(conversation, add_generation_prompt=True)
>>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> inputs = processor(images=image, text=prompt, return_tensors="pt").to(device)

>>> # Generate
>>> generated_ids = model.generate(**inputs, max_new_tokens=15)
>>> print(processor.batch_decode(generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0])
system\n You are a helpful assistant.\n user\n What are these?\n assistant\n The image depicts a traditional Chinese street...
Update on GitHub

←Evolla FLAVA→