<!--Copyright 2023 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Quantization
Quantization techniques reduce memory and computational costs by representing weights and activations with lower-precision data types like 8-bit integers (int8). This enables loading larger models you normally wouldn't be able to fit into memory, and speeding up inference. Transformers supports the AWQ and GPTQ quantization algorithms and it supports 8-bit and 4-bit quantization with bitsandbytes.
Quantization techniques that aren't supported in Transformers can be added with the [`HfQuantizer`] class.
<Tip>
Learn how to quantize models in the [Quantization](../quantization) guide.
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Generation
Each framework has a generate method for text generation implemented in their respective `GenerationMixin` class:
- PyTorch [`~generation.GenerationMixin.generate`] is implemented in [`~generation.GenerationMixin`].
You can parameterize the generate method with a [`~generation.GenerationConfig`] class instance. Please refer to this class for the complete list of generation parameters, which control the behavior of the generation method.
To learn how to inspect a model's generation configuration, what are the defaults, how to change the parameters ad hoc,
and how to create and save a customized generation configuration, refer to the
[text generation strategies guide](../generation_strategies). The guide also explains how to use related features,
`encode_plus` and `batch_encode_plus`) and is derived from a Python dictionary. When the tokenizer is a pure python
tokenizer, this class behaves just like a standard python dictionary and holds the various model inputs computed by
these methods (`input_ids`, `attention_mask`...). When the tokenizer is a "Fast" tokenizer (i.e., backed by
HuggingFace [tokenizers library](https://github.com/huggingface/tokenizers)), this class provides in addition
several advanced alignment methods which can be used to map between the original string (character and words) and the
token space (e.g., getting the index of the token comprising a given character or the span of characters corresponding
to a given token).
## Multimodal Tokenizer
Apart from that each tokenizer can be a "multimodal" tokenizer which means that the tokenizer will hold all relevant special tokens
as part of tokenizer attributes for easier access. For example, if the tokenizer is loaded from a vision-language model like LLaVA, you will
be able to access `tokenizer.image_token_id` to obtain the special image token used as a placeholder.
To enable extra special tokens for any type of tokenizer, you have to add the following lines and save the tokenizer. Extra special tokens do not
have to be modality related and can be anything that the model often needs access to. In the below code, tokenizer at `output_dir` will have direct access
The [`PreTrainedTokenizerFast`] depend on the [tokenizers](https://huggingface.co/docs/tokenizers) library. The tokenizers obtained from the 🤗 tokenizers library can be
loaded very simply into 🤗 transformers. Take a look at the [Using tokenizers from 🤗 tokenizers](../fast_tokenizers) page to understand how this is done.
<!--Copyright 2020 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Trainer
The [`Trainer`] class provides an API for feature-complete training in PyTorch, and it supports distributed training on multiple GPUs/TPUs, mixed precision for [NVIDIA GPUs](https://nvidia.github.io/apex/), [AMD GPUs](https://rocm.docs.amd.com/en/latest/rocm.html), and [`torch.amp`](https://pytorch.org/docs/stable/amp.html) for PyTorch. [`Trainer`] goes hand-in-hand with the [`TrainingArguments`] class, which offers a wide range of options to customize how a model is trained. Together, these two classes provide a complete training API.
[`Seq2SeqTrainer`] and [`Seq2SeqTrainingArguments`] inherit from the [`Trainer`] and [`TrainingArguments`] classes and they're adapted for training models for sequence-to-sequence tasks such as summarization or translation.
<Tipwarning={true}>
The [`Trainer`] class is optimized for 🤗 Transformers models and can have surprising behaviors
when used with other models. When using it with your own model, make sure:
- your model always return tuples or subclasses of [`~utils.ModelOutput`]
- your model can compute the loss if a `labels` argument is provided and that loss is returned as the first
element of the tuple (if your model returns tuples)
- your model can accept multiple label arguments (use `label_names` in [`TrainingArguments`] to indicate their name to the [`Trainer`]) but none of them should be named `"label"`
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Video Processor
A **Video Processor** is a utility responsible for preparing input features for video models, as well as handling the post-processing of their outputs. It provides transformations such as resizing, normalization, and conversion into PyTorch. Along ith transformations the `VideoProcessor` class handles video decoding from local paths or URLs (requires [`torchcodec`](https://pypi.org/project/torchcodec/)) and frame sampling according to model-specific strategies.
The video processor extends the functionality of image processors by allowing Vision Large Language Models (VLMs) to handle videos with a distinct set of arguments compared to images. It serves as the bridge between raw video data and the model, ensuring that input features are optimized for the VLM.
When adding a new VLM or updating an existing one to enable distinct video preprocessing, saving and reloading the processor configuration will store the video related arguments in a dedicated file named `video_preprocessing_config.json`. Don't worry if you haven't updated your VLM, the processor will try to load video related configurations from a file named `preprocessing_config.json`.
### Usage Example
Here's an example of how to load a video processor with [`llava-hf/llava-onevision-qwen2-0.5b-ov-hf`](https://huggingface.co/llava-hf/llava-onevision-qwen2-0.5b-ov-hf) model:
Currently, if using base image processor for videos, it processes video data by treating each frame as an individual image and applying transformations frame-by-frame. While functional, this approach is not highly efficient. Using `AutoVideoProcessor` allows us to take advantage of **fast video processors**, leveraging the [torchvision](https://pytorch.org/vision/stable/index.html) library. Fast processors handle the whole batch of videos at once, without iterating over each video or frame. These updates introduce GPU acceleration and significantly enhance processing speed, especially for tasks requiring high throughput.
Fast video processors are available for all models and are loaded by default when an `AutoVideoProcessor` is initialized. When using a fast video processor, you can also set the `device` argument to specify the device on which the processing should be done. By default, the processing is done on the same device as the inputs if the inputs are tensors, or on the CPU otherwise. For even more speed improvement, we can compile the processor when using 'cuda' as device.
The video processor can also sample video frames using the technique best suited for the given model. Sampling behavior is controlled with the `do_sample_frames` argument and can be configured through model-specific parameters such as `num_frames` or `fps` (the rate at which the video will be sampled). If the input video is given as a local path or URL (`str`), the processor will decode it automatically. To obtain metadata about the decoded video, such as sampled frame indices, original dimensions, duration, and fps, pass `return_metadata=True` to the processor.
<Tipwarning={false}>
- Specifying `num_frames` does not guarantee the output will contain exactly that number of frames. Depending on the model, the sampler may enforce minimum or maximum frame limits.
- The default decoder is [`torchcodec`](https://pypi.org/project/torchcodec/), which must be installed.
If you pass an already decoded video array but still want to enable model-specific frame sampling, it is strongly recommended to provide video_metadata. This allows the sampler to know the original video’s duration and FPS. You can pass metadata as a `VideoMetadata` object or as a plain dict.
<!--Copyright 2025 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
*This model was released on 2024-11-21 and added to Hugging Face Transformers on 2025-07-08.*
# AIMv2
## Overview
The AIMv2 model was proposed in [Multimodal Autoregressive Pre-training of Large Vision Encoders](https://huggingface.co/papers/2411.14402) by Enrico Fini, Mustafa Shukor, Xiujun Li, Philipp Dufter, Michal Klein, David Haldimann, Sai Aitharaju, Victor Guilherme Turrisi da Costa, Louis Béthune, Zhe Gan, Alexander T Toshev, Marcin Eichner, Moin Nabi, Yinfei Yang, Joshua M. Susskind, Alaaeldin El-Nouby.
The abstract from the paper is the following:
*We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.*
This model was contributed by [Yaswanth Gali](https://huggingface.co/yaswanthgali).
The original code can be found [here](https://github.com/apple/ml-aim).
## Usage Example
Here is an example of Image Feature Extraction using specific checkpoints on resized images and native resolution images:
[ALBERT](https://huggingface.co/papers/1909.11942) is designed to address memory limitations of scaling and training of [BERT](./bert). It adds two parameter reduction techniques. The first, factorized embedding parametrization, splits the larger vocabulary embedding matrix into two smaller matrices so you can grow the hidden size without adding a lot more parameters. The second, cross-layer parameter sharing, allows layer to share parameters which keeps the number of learnable parameters lower.
ALBERT was created to address problems like -- GPU/TPU memory limitations, longer training times, and unexpected model degradation in BERT. ALBERT uses two parameter-reduction techniques to lower memory consumption and increase the training speed of BERT:
-**Factorized embedding parameterization:** The large vocabulary embedding matrix is decomposed into two smaller matrices, reducing memory consumption.
-**Cross-layer parameter sharing:** Instead of learning separate parameters for each transformer layer, ALBERT shares parameters across layers, further reducing the number of learnable weights.
ALBERT uses absolute position embeddings (like BERT) so padding is applied at right. Size of embeddings is 128 While BERT uses 768. ALBERT can processes maximum 512 token at a time.
You can find all the original ALBERT checkpoints under the [ALBERT community](https://huggingface.co/albert) organization.
> [!TIP]
> Click on the ALBERT models in the right sidebar for more examples of how to apply ALBERT to different language tasks.
The example below demonstrates how to predict the `[MASK]` token with [`Pipeline`], [`AutoModel`], and from the command line.
<hfoptionsid="usage">
<hfoptionid="Pipeline">
```py
importtorch
fromtransformersimportpipeline
pipeline=pipeline(
task="fill-mask",
model="albert-base-v2",
dtype=torch.float16,
device=0
)
pipeline("Plants create [MASK] through a process known as photosynthesis.",top_k=5)
echo-e"Plants create [MASK] through a process known as photosynthesis." | transformers run --task fill-mask --model albert-base-v2 --device 0
```
</hfoption>
</hfoptions>
## Notes
- Inputs should be padded on the right because BERT uses absolute position embeddings.
- The embedding size `E` is different from the hidden size `H` because the embeddings are context independent (one embedding vector represents one token) and the hidden states are context dependent (one hidden state represents a sequence of tokens). The embedding matrix is also larger because `V x E` where `V` is the vocabulary size. As a result, it's more logical if `H >> E`. If `E < H`, the model has less parameters.
## Resources
The resources provided in the following sections consist of a list of official Hugging Face and community (indicated by 🌎) resources to help you get started with AlBERT. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
<PipelineTagpipeline="text-classification"/>
-[`AlbertForSequenceClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/text-classification).
- Check the [Text classification task guide](../tasks/sequence_classification) on how to use the model.
<PipelineTagpipeline="token-classification"/>
-[`AlbertForTokenClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/token-classification).
-[Token classification](https://huggingface.co/course/chapter7/2?fw=pt) chapter of the 🤗 Hugging Face Course.
- Check the [Token classification task guide](../tasks/token_classification) on how to use the model.
<PipelineTagpipeline="fill-mask"/>
-[`AlbertForMaskedLM`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/language-modeling#robertabertdistilbert-and-masked-language-modeling) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/language_modeling.ipynb).
-[Masked language modeling](https://huggingface.co/course/chapter7/3?fw=pt) chapter of the 🤗 Hugging Face Course.
- Check the [Masked language modeling task guide](../tasks/masked_language_modeling) on how to use the model.
<PipelineTagpipeline="question-answering"/>
-[`AlbertForQuestionAnswering`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/question_answering.ipynb).
-[Question answering](https://huggingface.co/course/chapter7/7?fw=pt) chapter of the 🤗 Hugging Face Course.
- Check the [Question answering task guide](../tasks/question_answering) on how to use the model.
**Multiple choice**
-[`AlbertForMultipleChoice`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/multiple-choice) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/multiple_choice.ipynb).
- Check the [Multiple choice task guide](../tasks/multiple_choice) on how to use the model.
[ALIGN](https://huggingface.co/papers/2102.05918) is pretrained on a noisy 1.8 billion alt‑text and image pair dataset to show that scale can make up for the noise. It uses a dual‑encoder architecture, [EfficientNet](./efficientnet) for images and [BERT](./bert) for text, and a contrastive loss to align similar image–text embeddings together while pushing different embeddings apart. Once trained, ALIGN can encode any image and candidate captions into a shared vector space for zero‑shot retrieval or classification without requiring extra labels. This scale‑first approach reduces dataset curation costs and powers state‑of‑the‑art image–text retrieval and zero‑shot ImageNet classification.
You can find all the original ALIGN checkpoints under the [Kakao Brain](https://huggingface.co/kakaobrain?search_models=align) organization.
> [!TIP]
> Click on the ALIGN models in the right sidebar for more examples of how to apply ALIGN to different vision and text related tasks.
The example below demonstrates zero-shot image classification with [`Pipeline`] or the [`AutoModel`] class.
- ALIGN projects the text and visual features into latent space and the dot product between the projected image and text features is used as the similarity score. The example below demonstrates how to calculate the image-text similarity score with [`AlignProcessor`] and [`AlignModel`].
```py
# Example of using ALIGN for image-text similarity
print(f"Most similar text: '{texts[most_similar_idx]}'")
```
## Resources
- Refer to the [Kakao Brain’s Open Source ViT, ALIGN, and the New COYO Text-Image Dataset](https://huggingface.co/blog/vit-align) blog post for more details.
[AltCLIP](https://huggingface.co/papers/2211.06679) replaces the [CLIP](./clip) text encoder with a multilingual XLM-R encoder and aligns image and text representations with teacher learning and contrastive learning.
You can find all the original AltCLIP checkpoints under the [AltClip](https://huggingface.co/collections/BAAI/alt-clip-diffusion-66987a97de8525205f1221bf) collection.
> [!TIP]
> Click on the AltCLIP models in the right sidebar for more examples of how to apply AltCLIP to different tasks.
The examples below demonstrates how to calculate similarity scores between an image and one or more captions with the [`AutoModel`] class.
inputs=processor(text=["a photo of a cat","a photo of a dog"],images=image,return_tensors="pt",padding=True)
outputs=model(**inputs)
logits_per_image=outputs.logits_per_image# this is the image-text similarity score
probs=logits_per_image.softmax(dim=1)# we can take the softmax to get the label probabilities
labels=["a photo of a cat","a photo of a dog"]
forlabel,probinzip(labels,probs[0]):
print(f"{label}: {prob.item():.4f}")
```
</hfoption>
</hfoptions>
Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
The example below uses [torchao](../quantization/torchao) to only quantize the weights to int4.
[Arcee](https://www.arcee.ai/blog/deep-dive-afm-4-5b-the-first-arcee-foundational-model) is a decoder-only transformer model based on the Llama architecture with a key modification: it uses ReLU² (ReLU-squared) activation in the MLP blocks instead of SiLU, following recent research showing improved training efficiency with squared activations. This architecture is designed for efficient training and inference while maintaining the proven stability of the Llama design.
The Arcee model is architecturally similar to Llama but uses `x * relu(x)` in MLP layers for improved gradient flow and is optimized for efficiency in both training and inference scenarios.
> [!TIP]
> The Arcee model supports extended context with RoPE scaling and all standard transformers features including Flash Attention 2, SDPA, gradient checkpointing, and quantization support.
The example below demonstrates how to generate text with Arcee using [`Pipeline`] or the [`AutoModel`].
[Aria](https://huggingface.co/papers/2410.05993) is a multimodal mixture-of-experts (MoE) model. The goal of this model is to open-source a training recipe for creating a multimodal native model from scratch. Aria has 3.9B and 3.5B activated parameters per visual and text token respectively. Text is handled by a MoE decoder and visual inputs are handled by a lightweight visual encoder. It is trained in 4 stages, language pretraining, multimodal pretraining, multimodal long-context pretraining, and multimodal post-training.
You can find all the original Aria checkpoints under the [Aria](https://huggingface.co/rhymes-ai?search_models=aria) organization.
> [!TIP]
> Click on the Aria models in the right sidebar for more examples of how to apply Aria to different multimodal tasks.
The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
The example below uses [torchao](../quantization/torchao) to only quantize the weights to int4 and the [rhymes-ai/Aria-sequential_mlp](https://huggingface.co/rhymes-ai/Aria-sequential_mlp) checkpoint. This checkpoint replaces grouped GEMM with `torch.nn.Linear` layers for easier quantization.
The Audio Spectrogram Transformer model was proposed in [AST: Audio Spectrogram Transformer](https://huggingface.co/papers/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
The Audio Spectrogram Transformer applies a [Vision Transformer](vit) to audio, by turning audio into an image (spectrogram). The model obtains state-of-the-art results
for audio classification.
The abstract from the paper is the following:
*In the past decade, convolutional neural networks (CNNs) have been widely adopted as the main building block for end-to-end audio classification models, which aim to learn a direct mapping from audio spectrograms to corresponding labels. To better capture long-range global context, a recent trend is to add a self-attention mechanism on top of the CNN, forming a CNN-attention hybrid model. However, it is unclear whether the reliance on a CNN is necessary, and if neural networks purely based on attention are sufficient to obtain good performance in audio classification. In this paper, we answer the question by introducing the Audio Spectrogram Transformer (AST), the first convolution-free, purely attention-based model for audio classification. We evaluate AST on various audio classification benchmarks, where it achieves new state-of-the-art results of 0.485 mAP on AudioSet, 95.6% accuracy on ESC-50, and 98.1% accuracy on Speech Commands V2.*
<small> Audio Spectrogram Transformer architecture. Taken from the <ahref="https://huggingface.co/papers/2104.01778">original paper</a>.</small>
This model was contributed by [nielsr](https://huggingface.co/nielsr).
The original code can be found [here](https://github.com/YuanGongND/ast).
## Usage tips
- When fine-tuning the Audio Spectrogram Transformer (AST) on your own dataset, it's recommended to take care of the input normalization (to make
sure the input has mean of 0 and std of 0.5). [`ASTFeatureExtractor`] takes care of this. Note that it uses the AudioSet
mean and std by default. You can check [`ast/src/get_norm_stats.py`](https://github.com/YuanGongND/ast/blob/master/src/get_norm_stats.py) to see how
the authors compute the stats for a downstream dataset.
- Note that the AST needs a low learning rate (the authors use a 10 times smaller learning rate compared to their CNN model proposed in the
[PSLA paper](https://huggingface.co/papers/2102.01243)) and converges quickly, so please search for a suitable learning rate and learning rate scheduler for your task.
### Using Scaled Dot Product Attention (SDPA)
PyTorch includes a native scaled dot-product attention (SDPA) operator as part of `torch.nn.functional`. This function
encompasses several implementations that can be applied depending on the inputs and the hardware in use. See the
For the best speedups, we recommend loading the model in half-precision (e.g. `torch.float16` or `torch.bfloat16`).
On a local benchmark (A100-40GB, PyTorch 2.3.0, OS Ubuntu 22.04) with `float32` and `MIT/ast-finetuned-audioset-10-10-0.4593` model, we saw the following speedups during inference.
| Batch size | Average inference time (ms), eager mode | Average inference time (ms), sdpa model | Speed up, Sdpa / Eager (x) |
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with the Audio Spectrogram Transformer.
<PipelineTagpipeline="audio-classification"/>
- A notebook illustrating inference with AST for audio classification can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/AST).
-[`ASTForAudioClassification`] is supported by this [example script](https://github.com/huggingface/transformers/tree/main/examples/pytorch/audio-classification) and [notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/audio_classification.ipynb).
- See also: [Audio classification](../tasks/audio_classification).
If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
The Autoformer model was proposed in [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://huggingface.co/papers/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
This model augments the Transformer as a deep decomposition architecture, which can progressively decompose the trend and seasonal components during the forecasting process.
The abstract from the paper is the following:
*Extending the forecasting time is a critical demand for real applications, such as extreme weather early warning and long-term energy consumption planning. This paper studies the long-term forecasting problem of time series. Prior Transformer-based models adopt various self-attention mechanisms to discover the long-range dependencies. However, intricate temporal patterns of the long-term future prohibit the model from finding reliable dependencies. Also, Transformers have to adopt the sparse versions of point-wise self-attentions for long series efficiency, resulting in the information utilization bottleneck. Going beyond Transformers, we design Autoformer as a novel decomposition architecture with an Auto-Correlation mechanism. We break with the pre-processing convention of series decomposition and renovate it as a basic inner block of deep models. This design empowers Autoformer with progressive decomposition capacities for complex time series. Further, inspired by the stochastic process theory, we design the Auto-Correlation mechanism based on the series periodicity, which conducts the dependencies discovery and representation aggregation at the sub-series level. Auto-Correlation outperforms self-attention in both efficiency and accuracy. In long-term forecasting, Autoformer yields state-of-the-art accuracy, with a 38% relative improvement on six benchmarks, covering five practical applications: energy, traffic, economics, weather and disease.*
This model was contributed by [elisim](https://huggingface.co/elisim) and [kashif](https://huggingface.co/kashif).
The original code can be found [here](https://github.com/thuml/Autoformer).
## Resources
A list of official Hugging Face and community (indicated by 🌎) resources to help you get started. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
- Check out the Autoformer blog-post in HuggingFace blog: [Yes, Transformers are Effective for Time Series Forecasting (+ Autoformer)](https://huggingface.co/blog/autoformer)
[Aya Vision](https://huggingface.co/papers/2505.08751) is a family of open-weight multimodal vision-language models from Cohere Labs. It is trained with a synthetic annotation framework that generates high-quality multilingual image captions, improving Aya Vision's generated responses. In addition, a cross-modal model merging technique is used to prevent the model from losing its text capabilities after adding vision capabilities. The model combines a CommandR-7B language model with a SigLIP vision encoder.
You can find all the original Aya Vision checkpoints under the [Aya Vision](https://huggingface.co/collections/CohereLabs/cohere-labs-aya-vision-67c4ccd395ca064308ee1484) collection.
> [!TIP]
> This model was contributed by [saurabhdash](https://huggingface.co/saurabhdash) and [yonigozlan](https://huggingface.co/yonigozlan).
>
> Click on the Aya Vision models in the right sidebar for more examples of how to apply Aya Vision to different image-to-text tasks.
The example below demonstrates how to generate text based on an image with [`Pipeline`] or the [`AutoModel`] class.
Quantization reduces the memory footprint of large models by representing weights at lower precision. Refer to the [Quantization](../quantization/overview) overview for supported backends.
The example below uses [bitsandbytes](../quantization/bitsandbytes) to only quantize the weights to 4-bits.
[Bamba](https://huggingface.co/blog/bamba) is a 9B parameter decoder-only language model built on the [Mamba-2](./mamba2) architecture. It is pretrained in two stages - it starts by training on 2T tokens from the [Dolma v1.7](https://huggingface.co/datasets/allenai/dolma) dataset and then trained on an additional 200B tokens from [FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) and [Cosmopedia](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia).
You can find all the original Bamba checkpoints under the [Bamba](https://huggingface.co/collections/ibm-ai-platform/bamba-674f1388b9bbc98b413c7bab) collection.
> [!TIP]
> This model was contributed by [ani300](https://github.com/ani300) and [fabianlim](https://github.com/fabianlim).
>
> Click on the Bamba models in the right sidebar for more examples of how to apply Bamba to different text generation tasks.
The example below demonstrates how to generate text with [`Pipeline`], [`AutoModel`], and from the command line.
<hfoptionsid="usage">
<hfoptionid="Pipeline">
```python
importtorch
fromtransformersimportpipeline
pipeline=pipeline(
task="text-generation",
model="ibm-ai-platform/Bamba-9B-v2",
dtype=torch.bfloat16,
device=0
)
pipeline("Plants create energy through a process known as")
echo"Plants create energy through a process known as" | transformers run --task text-generation --model ibm-ai-platform/Bamba-9B-v2 --device 0
```
</hfoption>
</hfoptions>
Quantization reduces the memory burden of large models by representing the weights in a lower precision. Refer to the [Quantization](../quantization/overview) overview for more available quantization backends.
The example below uses [torchao](../quantization/torchao) to only quantize the weights to int4.
- Bamba supports padding-free training which concatenates distinct training examples while still processing inputs as separate batches. It can significantly accelerate inference by [~2x](https://github.com/huggingface/transformers/pull/35861#issue-2807873129)(depending on model and data distribution) and reduce memory-usage if there are examples of varying lengths by avoiding unnecessary compute and memory overhead from padding tokens.
Padding-free training requires the `flash-attn`, `mamba-ssm`, and `causal-conv1d` packages and the following arguments must be passed to the model in addition to `input_ids` and `labels`.
-`position_ids: torch.LongTensor`: the position index of each token in each sequence.
-`seq_idx: torch.IntTensor`: the index of each sequence in the batch.
- Each of the [`FlashAttentionKwargs`]
-`cu_seq_lens_q: torch.LongTensor`: the cumulative sequence lengths of all queries.
-`cu_seq_lens_k: torch.LongTensor`: the cumulative sequence lengths of all keys.
-`max_length_q: int`: the longest query length in the batch.
-`max_length_k: int`: the longest key length in the batch.
The `attention_mask` inputs should not be provided. The [`DataCollatorWithFlattening`] programmatically generates the set of additional arguments above using `return_seq_idx=True` and `return_flash_attn_kwargs=True`. See the [Improving Hugging Face Training Efficiency Through Packing with Flash Attention](https://huggingface.co/blog/packing-with-FA2) blog post for additional information.
[Bark](https://huggingface.co/suno/bark) is a transformer-based text-to-speech model proposed by Suno AI in [suno-ai/bark](https://github.com/suno-ai/bark).
Bark is made of 4 main models:
- [`BarkSemanticModel`] (also referred to as the 'text' model): a causal auto-regressive transformer model that takes as input tokenized text, and predicts semantic text tokens that capture the meaning of the text.
- [`BarkCoarseModel`] (also referred to as the 'coarse acoustics' model): a causal autoregressive transformer, that takes as input the results of the [`BarkSemanticModel`] model. It aims at predicting the first two audio codebooks necessary for EnCodec.
- [`BarkFineModel`] (the 'fine acoustics' model), this time a non-causal autoencoder transformer, which iteratively predicts the last codebooks based on the sum of the previous codebooks embeddings.
- having predicted all the codebook channels from the [`EncodecModel`], Bark uses it to decode the output audio array.
It should be noted that each of the first three modules can support conditional speaker embeddings to condition the output sound according to specific predefined voice.
This model was contributed by [Yoach Lacombe (ylacombe)](https://huggingface.co/ylacombe) and [Sanchit Gandhi (sanchit-gandhi)](https://github.com/sanchit-gandhi).
The original code can be found [here](https://github.com/suno-ai/bark).
### Optimizing Bark
Bark can be optimized with just a few extra lines of code, which **significantly reduces its memory footprint** and **accelerates inference**.
#### Using half-precision
You can speed up inference and reduce memory footprint by 50% simply by loading the model in half-precision.
As mentioned above, Bark is made up of 4 sub-models, which are called up sequentially during audio generation. In other words, while one sub-model is in use, the other sub-models are idle.
If you're using a CUDA GPU or Intel XPU, a simple solution to benefit from an 80% reduction in memory footprint is to offload the submodels from device to CPU when they're idle. This operation is called *CPU offloading*. You can use it with one line of code as follows:
```python
model.enable_cpu_offload()
```
Note that 🤗 Accelerate must be installed before using this feature. [Here's how to install it.](https://huggingface.co/docs/accelerate/basic_tutorials/install)
#### Using Flash Attention 2
Flash Attention 2 is an even faster, optimized version of the previous optimization.
##### Installation
First, check whether your hardware is compatible with Flash Attention 2. The latest list of compatible hardware can be found in the [official documentation](https://github.com/Dao-AILab/flash-attention#installation-and-features).
Next, [install](https://github.com/Dao-AILab/flash-attention#installation-and-features) the latest version of Flash Attention 2:
```bash
pip install-U flash-attn --no-build-isolation
```
##### Usage
To load a model using Flash Attention 2, we can pass the `attn_implementation="flash_attention_2"` flag to [`.from_pretrained`](https://huggingface.co/docs/transformers/main/en/main_classes/model#transformers.PreTrainedModel.from_pretrained). We'll also load the model in half-precision (e.g. `torch.float16`), since it results in almost no degradation to audio quality but significantly lower memory usage and faster inference:
The following diagram shows the latency for the native attention implementation (no optimisation) against Flash Attention 2. In all cases, we generate 400 semantic tokens on a 40GB A100 GPU with PyTorch 2.1:
To put this into perspective, on an NVIDIA A100 and when generating 400 semantic tokens with a batch size of 16, you can get 17 times the [throughput](https://huggingface.co/blog/optimizing-bark#throughput) and still be 2 seconds faster than generating sentences one by one with the native model implementation. In other words, all the samples will be generated 17 times faster.
#### Combining optimization techniques
You can combine optimization techniques, and use CPU offload, half-precision and Flash Attention 2 all at once.
Find out more on inference optimization techniques [here](https://huggingface.co/docs/transformers/perf_infer_gpu_one).
### Usage tips
Suno offers a library of voice presets in a number of languages [here](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c).
These presets are also uploaded in the hub [here](https://huggingface.co/suno/bark-small/tree/main/speaker_embeddings) or [here](https://huggingface.co/suno/bark/tree/main/speaker_embeddings).
[BART](https://huggingface.co/papers/1910.13461) is a sequence-to-sequence model that combines the pretraining objectives from BERT and GPT. It's pretrained by corrupting text in different ways like deleting words, shuffling sentences, or masking tokens and learning how to fix it. The encoder encodes the corrupted document and the corrupted text is fixed by the decoder. As it learns to recover the original text, BART gets really good at both understanding and generating language.
You can find all the original BART checkpoints under the [AI at Meta](https://huggingface.co/facebook?search_models=bart) organization.
The example below demonstrates how to predict the `[MASK]` token with [`Pipeline`], [`AutoModel`], and from the command line.
<hfoptionsid="usage">
<hfoptionid="Pipeline">
```py
importtorch
fromtransformersimportpipeline
pipeline=pipeline(
task="fill-mask",
model="facebook/bart-large",
dtype=torch.float16,
device=0
)
pipeline("Plants create <mask> through a process known as photosynthesis.")
print(f"The predicted token is: {predicted_token}")
```
</hfoption>
<hfoptionid="transformers CLI">
```bash
echo-e"Plants create <mask> through a process known as photosynthesis." | transformers run --task fill-mask --model facebook/bart-large --device 0
```
</hfoption>
</hfoptions>
## Notes
- Inputs should be padded on the right because BERT uses absolute position embeddings.
- The [facebook/bart-large-cnn](https://huggingface.co/facebook/bart-large-cnn) checkpoint doesn't include `mask_token_id` which means it can't perform mask-filling tasks.
- BART doesn't use `token_type_ids` for sequence classification. Use [`BartTokenizer`] or [`~PreTrainedTokenizerBase.encode`] to get the proper splitting.
- The forward pass of [`BartModel`] creates the `decoder_input_ids` if they're not passed. This can be different from other model APIs, but it is a useful feature for mask-filling tasks.
- Model predictions are intended to be identical to the original implementation when `forced_bos_token_id=0`. This only works if the text passed to `fairseq.encode` begins with a space.
- [`~GenerationMixin.generate`] should be used for conditional generation tasks like summarization.
[BARThez](https://huggingface.co/papers/2010.12321) is a [BART](./bart) model designed for French language tasks. Unlike existing French BERT models, BARThez includes a pretrained encoder-decoder, allowing it to generate text as well. This model is also available as a multilingual variant, mBARThez, by continuing pretraining multilingual BART on a French corpus.
You can find all of the original BARThez checkpoints under the [BARThez](https://huggingface.co/collections/dascim/barthez-670920b569a07aa53e3b6887) collection.
> [!TIP]
> This model was contributed by [moussakam](https://huggingface.co/moussakam).
> Refer to the [BART](./bart) docs for more usage examples.
The example below demonstrates how to predict the `<mask>` token with [`Pipeline`], [`AutoModel`], and from the command line.
<hfoptionsid="usage">
<hfoptionid="Pipeline">
```py
importtorch
fromtransformersimportpipeline
pipeline=pipeline(
task="fill-mask",
model="moussaKam/barthez",
dtype=torch.float16,
device=0
)
pipeline("Les plantes produisent <mask> grâce à un processus appelé photosynthèse.")
print(f"The predicted token is: {predicted_token}")
```
</hfoption>
<hfoptionid="transformers CLI">
```bash
echo-e"Les plantes produisent <mask> grâce à un processus appelé photosynthèse." | transformers run --task fill-mask --model moussaKam/barthez --device 0
