AI BriefingHuggingFace Blog news
The latest AI and developer news from HuggingFace Blog, with short summaries.
Visit website →Llama 3.1 405B Deployment on Vertex AI
A Hugging Face guide walks through deploying Llama 3.1 405B on Google Cloud Vertex AI.
Infini-Attention Experiment Results: Limits and Alternatives for Long-Context Extension
Infini-attention's performance degrades as memory compression repeats, and Ring Attention or YaRN have been confirmed as more effective alternatives for now.
Getting Started with ggml, the Foundation of Local LLMs
Introduces the features and core concepts of ggml, a low-level C/C++ library optimized for Transformer inference.
Hugging Face Releases Unified Tool Use API
Hugging Face has released a unified Tool Use API that is compatible across various model families.
TII Releases Falcon Mamba 7B Model
TII has released a 7B-scale Mamba-based model that achieves high performance without an attention mechanism.
Hugging Face Acquires XetHub
Hugging Face has acquired XetHub to optimize the management of large-scale data and models.
HuggingFace Unveils Document Image Augmentation Technique
It introduces a multimodal data augmentation technique that simultaneously transforms text and images in document images to improve VLM training efficiency.
2024 Security Feature Highlights
Hugging Face summarized key security features—including fine-grained permission tokens, 2FA, and commit signing—designed to protect Hub users' assets.
Google Releases Gemma 2 2B and Safety/Interpretability Tools
Google has released Gemma 2 2B for on-device use, the safety classifier ShieldGemma, and the interpretability tool Gemma Scope.
Diffusion Memory Optimization Based on Quanto
This presents a method for efficiently reducing the memory usage of Diffusion Transformer models by leveraging the Quanto quantization tool.
LAVE, an LLM-based VQA evaluation metric, proposed
LAVE is proposed as a metric that uses an LLM to evaluate semantic similarity instead of the existing string-matching approach.
Meta Llama 3.1 Release and Ecosystem Integration Support
Meta's Llama 3.1 model series has been released, offering 128K context and multilingual support.
Running Mistral 7B with Core ML
Introduces a method for efficiently running the Mistral 7B model on Mac by leveraging Apple's new Core ML features.
TGI Unveils Multi-LoRA Serving Feature
Hugging Face's TGI has launched a feature that dynamically serves multiple LoRA adapters from a single base model.
HuggingFace Releases Docmatix, a Large-Scale Dataset for DocVQA
HuggingFace has released Docmatix, a DocVQA dataset that is 240 times larger than existing ones, containing 2.4 million images and 9.5 million Q/A pairs.
HuggingFace Releases SmolLM
HuggingFace released the SmolLM series of small language models in 135M, 360M, and 1.7B sizes, trained on high-quality data.
NuminaMath Wins AIMO
The NuminaMath 7B TIR model won the first AI Math Olympiad (AIMO) competition.
TRL Adds DPO Support for VLMs
HuggingFace's TRL library now supports Direct Preference Optimization (DPO) for Vision Language Models (VLMs).
Hugging Face experiments with PII detection for datasets
Hugging Face is experimenting with a report feature that uses Presidio to flag whether datasets contain personally identifiable information (PII).
Hugging Face and KerasHub Integration: Transformers Model Support
Transformers models from Hugging Face can now be loaded and used directly in KerasHub.