PyTorch

Google has announced TimesFM-3, an evolution of its time-series foundation model that moves beyond univariate forecasting to support multivariate scenarios. The model leverages 330 million parameters and a trillion-point dataset to predict multiple coevolving series using both past and future covariates.

This article explores how the Vortex columnar file format optimizes the data pipeline for ML training by reducing latency between cloud storage and GPUs. It highlights techniques like column pruning and GPU-side decompression to eliminate traditional CPU and NVMe bottlenecks.

This article details how AI is unevenly accelerating scientific progress, with major impacts in cybersecurity and minor effects in mathematics. It also highlights two new frameworks, SPADE for synthetic environment generation and Hawkeye for automated GPU kernel optimization.

This article explores a method for creating deterministic arithmetic calculators by compiling algorithmic logic directly into transformer weights. By using the Torchwright compiler, the author bypasses the need for training and achieves exact multiplication, addition, and subtraction results.

TorchJD is a specialized PyTorch library for optimizing neural networks with multiple simultaneous loss functions. It implements advanced methods like Jacobian descent and scalarization to resolve optimization conflicts in multi-task learning.

The article advocates for a proactive 'sniff test' methodology to validate the integrity of SBOMs in containerized environments. It highlights how identifying omissions like unpinned packages or missing dependencies is crucial for preventing supply chain attacks.

Kuma introduces a way to run trained PyTorch models live in the browser by compiling them into a specialized '.iph' format. By leveraging WebGPU and embedded WGSL shaders, it eliminates the need for server-side inference or heavy runtime dependencies.

This article introduces tinytorchcompile, a simplified single-file implementation of the torch.compile mechanism focusing on operator fusion. It provides empirical evidence that fusing operations into a single kernel is essential for bypassing memory bandwidth limitations.