20 articles
Hugging Face releases a guide to finetune Stable Diffusion models using DDPO (Diffusion DDPOTrainer) integrated into TRL. The method enables optimizing image generation models with custom reward functions without requiring additional training data.
Practical guide to train a LLaMA 2 chatbot without technical expertise. Covers data preparation, fine-tuning, and deployment using accessible tools.
Hugging Face releases a benchmark of Llama 2 on Amazon SageMaker. The study measures inference and fine-tuning performance of the model on AWS cloud infrastructure, providing latency and throughput metrics.
ChatGPT now integrates vision, audio, and speech capabilities. Users can have voice conversations with the model, show it images, and receive spoken responses. Rolling out gradually on iOS, Android, and web.
OpenAI releases the system card for GPT-4V, documenting the model's computer vision capabilities, limitations, and identified safety risks. The document details performance evaluations and risk mitigation measures.
Hugging Face launches Inference for PROs, a professional inference offering with priority model access, dedicated support and guaranteed SLAs. Cloud service for deploying and serving LLM models in production.
OpenAI announces an open call for its Red Teaming Network, recruiting domain experts to identify vulnerabilities and improve the safety of its models.
Rocket Money and Hugging Face collaborate to scale volatile ML models in production. The solution optimizes inference and resource management for unstable models in critical environments.
Hugging Face launches an object detection leaderboard to benchmark and compare computer vision models' performance on standardized datasets.
Hugging Face introduces a guide to 3D Gaussian Splatting, a rendering technique representing 3D scenes through gaussians. Alternative approach to NeRF offering better performance and visual quality.
Hugging Face publishes a guide to optimizing LLMs in production, covering quantization techniques, distillation, and efficient deployment to reduce inference latency and costs.
OpenAI opens an office in Dublin, Ireland, to expand its presence in Europe.
Hugging Face introduces Würstchen, a diffusion model for image generation optimized for speed. The model uses a two-stage architecture reducing latency while maintaining visual quality. Code and weights available open-source.
Hugging Face releases a comprehensive guide for fine-tuning Llama 2 70B using PyTorch FSDP. The method leverages distributed sharding to optimize GPU memory and accelerate training across multiple nodes.
Hugging Face outlines natively supported quantization schemes in Transformers: GPTQ, AWQ, GGUF, bitsandbytes (8-bit, 4-bit). Each method trades off compression versus precision, with direct integration into the library.
Hugging Face compares SafeCoder, its open-source code assistant, against closed-source solutions. The study evaluates security, code quality, and performance on standard benchmarks.
Hugging Face introduces T2I-Adapters, lightweight modules for precise control over SDXL image generation without model fine-tuning. Adapters enable guided generation via conditions (sketches, poses, depth) with minimal computational overhead.
OpenAI hosts its first developer conference on November 6 in San Francisco. In-person registration opens in coming weeks, with livestream access for the keynote.
Hugging Face announces the release of Falcon 180B, an open-source large language model with 180 billion parameters. The model is available in base and instruction-tuned versions, designed for complex text generation and reasoning tasks.
Fetch reduces ML processing latency by 50% using Amazon SageMaker and Hugging Face. The optimization focuses on model inference and data pipeline acceleration for production applications.