STARFlow and STARFlow-V are open-source models designed for generating high-quality images and videos from text prompts. They combine autoregressive models with normalizing flows to achieve impressive results in both text-to-image and text-to-video tasks. Users can easily set up the models and start generating content with provided scripts and configurations.

This article explains tensor parallelism (TP) in transformer models, focusing on how it allows for efficient matrix multiplication across multiple GPUs. It details the application of TP in both the Multi-Head Attention and Feed-Forward Network components, highlighting its constraints and practical usage with the Hugging Face library.

This article critiques the use of perplexity as a metric for evaluating machine learning models, particularly Transformers. It argues that a model can achieve low perplexity while failing to predict certain sequences accurately, highlighting the metric's inadequacy in reliably selecting the best model. The authors provide analytical insights into how model confidence and accuracy relate to perplexity.

Pingkit is a toolkit designed for training reproducible, capacity-aware models using transformer activations. It offers features for extracting embeddings, training neural architectures, and creating custom probes tailored to specific research needs. The toolkit is integrated with Hugging Face models and provides various utilities for data processing and model training.

Text-to-LoRA (T2L) is a hypernetwork that enables the instant adaptation of large language models to specific tasks using only natural language descriptions, eliminating the need for extensive fine-tuning and dataset curation. Trained on various pre-existing LoRA adapters, T2L can generate task-specific adapters in a single forward pass, demonstrating performance comparable to traditional methods while significantly reducing computational requirements and allowing zero-shot generalization to new tasks.

OpenAI's GPT-OSS models introduce several efficiency upgrades for transformers, including MXFP4 quantization and specialized kernels that enhance performance during model loading and execution. The updates allow for faster inference and fine-tuning while maintaining compatibility across major models in the transformers library. Additionally, community-contributed kernels are integrated to streamline usage and performance optimization.

Power Attention is an open-source implementation designed to optimize the core operation of symmetric power transformers, enabling efficient training and inference on long-context sequences. It serves as a drop-in replacement for various attention forms, significantly improving performance metrics like loss-per-FLOP compared to traditional and linear attention models. The architecture’s adjustable hyperparameter allows for better balance between weight and state FLOPs, enhancing scalability and learning efficiency.

NN-Former introduces a novel approach to neural architecture representation by combining the strengths of Graph Neural Networks and transformers while addressing their limitations. It emphasizes the importance of sibling nodes in the architecture topology and proposes new mechanisms for predicting accuracy and latency, achieving improved performance in learning Directed Acyclic Graph topology.

Test-Time Training (TTT) layers enhance pre-trained Transformers' ability to generate one-minute videos from text narratives, yielding improved coherence and aesthetics compared to existing methods. Despite notable artifacts and limitations in the current implementation, TTT-MLP shows significant advancements in temporal consistency and motion smoothness, particularly when tested on a dataset of Tom and Jerry cartoons. Future work aims to extend this approach to longer videos and more complex storytelling.

The study introduces a theoretical framework for understanding in-context learning (ICL) in large language models (LLMs) by utilizing hierarchical concept modeling and optimization theory. It demonstrates how nonlinear residual transformers can effectively perform factual-recall tasks through vector arithmetic, proving strong generalization and robustness against concept recombination and distribution shifts. Empirical simulations support these theoretical findings, showcasing the advantages of transformers over traditional static embeddings.

This study investigates how a one-layer transformer learns to recognize regular languages, focusing on tasks such as 'even pairs' and 'parity check'. Through theoretical analysis of training dynamics under gradient descent, it reveals two distinct phases in the learning process, demonstrating how the attention and linear layers interact to achieve effective separation of data sequences. Experimental results confirm the theoretical findings.