The Hugging Face Hub has over 850,000 public models, with around 50,000 new ones added every month. The platform offers an Enterprise Hub subscription that provides compliance, security, governance features, and additional compute capacities for production-level inference. The Hub allows users to run diverse hardware and Docker Spaces, making it suitable for broad usage.
AIHuggingFace | Rating: 73 | 2024-08-22 01:37:54 PM |
Hugging Face has improved training efficiency through packing with Flash Attention 2, which provides up to 2x improvement in training throughput while maintaining convergence quality. This is achieved by packing examples without padding and using token position information. The new DataCollatorWithFlattening is compatible with Flash Attention 2 and is available on GitHub with training examples.
AIHuggingFace | Rating: 62 | 2024-08-21 02:05:14 PM |
Meta Llama 3.1 405B, the latest open LLM from Meta, has been deployed on Google Cloud Vertex AI. The 405B variant is designed for synthetic data, LLM as a Judge or distillation, and other use cases. It features a large context length of 128K tokens, multilingual capabilities, tool usage capabilities, and a more permissive license. The deployment was done on a Google Cloud A3 node with 8 x H100 NVIDIA GPUs.
AIMeta | Rating: 60 | 2024-08-19 02:25:13 PM |
A failed experiment called Infini-Attention was conducted, which aimed to increase the context length of language models. However, the experiment's performance worsened as the memory was compressed multiple times. As a result, ring attention and rope scaling are still considered the best methods for extending a pretrained model to longer context lengths.
AIHuggingFace | Rating: 42 | 2024-08-14 03:25:16 PM |
ggml is a machine learning library written in C and C++ with a focus on Transformer inference. It is an open-source project being actively developed by a growing community. The library is similar to PyTorch and TensorFlow, but is still in its early stages of development. ggml has gained popularity alongside other projects like llama.cpp and whisper.cpp, and is used by many other projects to enable on-device LLM.
AIHuggingFace | Rating: 62 | 2024-08-13 11:25:33 AM |
The Technology Innovation Institute (TII) in Abu Dhabi has released FalconMamba, a new 7B model under the TII Falcon License 2.0. The model is open access and available within the Hugging Face ecosystem for research or application purposes. It is the first general-purpose large-scale pure Mamba model based on the attention mechanism.
AIHuggingFace | Rating: 60 | 2024-08-12 02:16:08 PM |
Transformers now includes unified tool use across several popular families of models, making it portable and easy to use with Mistral NousResearch or Llama models. The tool use feature allows giving callable functions to an LLM, enabling it to decide when to call them to help respond to user queries. The feature is supported by complete documentation and examples, and more models will be added in the near future.
AIHuggingFace | Rating: 62 | 2024-08-12 01:36:02 PM |
Hugging Face acquired XetHub, a Seattle-based company founded by Yucheng Low, Ajit Banerjee, and Rajat Arya, who previously worked at Apple. XetHub's mission is to enable software engineering best practices for AI development, and they have developed technologies to scale Git to TB repositories and enable teams to work together on large evolving datasets and models.
AIHuggingFace | Rating: 62 | 2024-08-08 01:25:56 PM |
Hugging Face prioritizes security, developing features to safeguard users and their assets. As of August 6th, 2024, essential security features are available to all users, including those on the Hugging Face Hub. Advanced controls are available to Enterprise users.
AIHuggingFace | Rating: 62 | 2024-08-06 09:45:57 PM |
Introducing Multimodal TextImage Augmentation for Document Images, developed in collaboration with Albumentations AI. This technique is designed to improve the performance of Vision Language Models (VLMs) on datasets containing document images with high textual content. The goal is to enable models to learn to read properly by interacting with both text and image modalities during training.
AIHuggingFace | Rating: 61 | 2024-08-06 01:15:57 PM |