pytorch
30 articles about pytorch in AI news
The 'pytorch_no_powerplant_blowup' env var keeping frontier GPUs alive
A tweet reveals a PyTorch env var that throttles GPU power to prevent grid failures, exposing the real-world constraints of frontier AI infrastructure.
Hugging Face Transfers Safetensors to PyTorch Foundation
Hugging Face is transferring ownership of the Safetensors library to the PyTorch Foundation, shepherded by the Linux Foundation. The move aims to establish it as a neutral, community-driven standard for safe AI model serialization.
Claude MCP GPU Debugging: AI Agent Identifies PyTorch Bottleneck in Kernel
A developer used an AI agent powered by Claude Code and the Model Context Protocol (MCP) to diagnose a severe GPU performance bottleneck. The agent analyzed system kernel traces, pinpointing excessive CPU context switches as the culprit, demonstrating a practical application of agentic AI for complex technical debugging.
NVIDIA's Molt: 9.2K-Line RL Framework Scales to 1T-Parameter MoE Models
NVIDIA released Molt, a 9.2K-line PyTorch RL framework scaling to 1T-parameter MoE models via vLLM, targeting agentic tasks with fully-async rollout.
Pyptx: Write Nvidia PTX Kernels in Python for Hopper and Blackwell
Pyptx lets developers write and launch hand-tuned Nvidia PTX kernels directly from Python, supporting Hopper (sm_90a) and Blackwell (sm_100a). It provides explicit control over registers, shared memory, and advanced features like WGMMA and TMA, with dispatch through JAX, PyTorch eager, and torch.compile.
Demystifying AI: Open-Source Blueprint Reveals How to Build ChatGPT From Scratch
A new GitHub repository called 'LLMs-from-scratch' provides a complete, line-by-line guide to building a GPT model in PyTorch, removing the black-box nature of large language models and empowering developers to understand and create their own AI systems.
Google's TensorFlow 2.21 Revolutionizes Edge AI with Unified LiteRT Framework
Google has launched TensorFlow 2.21, marking LiteRT's transition to a production-ready universal on-device inference framework. This major update delivers faster GPU performance, new NPU acceleration, and seamless PyTorch edge deployment, effectively replacing TensorFlow Lite for mobile and edge applications.
k-dense Ships 150 Open-Source Scientific Agent Skills
k-dense released 150 open-source scientific agent skills covering biology, chemistry, drug discovery.
Microsoft to Deploy AMD Helios Rack-Scale AI at Scale on Azure
Microsoft will deploy AMD's Helios rack-scale AI accelerator at scale on Azure, powered by MI455X GPUs and Epyc Venice CPUs. The move diversifies Azure's AI silicon beyond Nvidia.
NanoEuler: GPT-2-Scale 116M Model Built in Pure C/CUDA From Scratch
NanoEuler is a 116M-parameter GPT-2-scale model built in pure C/CUDA from scratch. It provides a complete educational training pipeline for understanding LLMs at the lowest level.
Grouped Query Experts cuts long-context attention cost 44%
GQE speeds long-context attention prefill 1.7–1.8× by routing tokens to 9 of 16 query heads, matching baseline accuracy at 56.04.
How Simon Willison Ported a 0.2B Image Model to the Browser with Claude
Simon Willison used Claude Code to port a 0.2B image inpainting model to WebGPU, running it as a parallel side project while his main agent worked on Datasette. The technique? Research with Claude.ai, then hand off to Claude Code with research.md.
Cerebras Claims Performance Parity With Nvidia H100 on AI Training
Cerebras claims wafer-scale chips match Nvidia H100 on AI training performance per watt, challenging Nvidia's dominance.
Apple Core AI Runs Models On-Device, Zero Server Calls
Apple launched Core AI for on-device model inference on Apple silicon. Zero server calls, supports Qwen, Mistral, SAM3 across devices.
NVIDIA NVFP4 on Blackwell Cuts JAX Training by 1.8x in MaxText
NVIDIA NVFP4 on Blackwell achieves 1.8x training speedup over FP8 in JAX/MaxText with no claimed accuracy loss for models up to 70B, but larger-scale validation is needed.
mlx-vlm v0.6.2 Adds Gemma 4 QAT Support for Local GPUs
mlx-vlm v0.6.2 adds launch-day support for Google DeepMind's Gemma 4 QAT checkpoints, enabling local inference on consumer GPUs and edge devices with video input for the 12B model.
Claude Code Token Costs Got You Down? Here's How to Cut Usage 40% Without
Claude Code users frustrated by token costs should use /compact, optimize CLAUDE.md, and route cheap models via OpenRouter for simple tasks—no local model matches Claude's quality yet.
xAI Drops JAX, Builds Custom C Training Framework After <10% MFU
xAI dropped JAX for GPU training after <10% MFU, building a custom C framework with Grok Build. NVIDIA's JAX team loses its biggest customer.
train-llm-from-scratch: 1B-Parameter LLM on a Single GPU
train-llm-from-scratch trains billion-parameter LLMs on a single GPU, cutting costs from $10M+ to consumer hardware.
MLX CUDA Backend Passes All Tests, Closing Apple GPU Gap
MLX CUDA backend passes all tests, enabling NVIDIA GPU support. Milestone bridges Apple Silicon and CUDA ecosystems for ML workloads.
Meshwatch GNN Stack Ships Fraud Detection with 17.2% Lift over XGBoost
Meshwatch GNN fraud stack achieves 17.2% recall lift over XGBoost at sub-50ms latency, shipping a custom GraphSAGE variant with online neighbor sampling.
AMD ROCm Performance Jumps 75x in 14 Days Post-DeepSeek v4
AMD ROCm stack improved 75x in 14 days post-DeepSeek v4 via fused operations. Still needs 5x more to match B200 performance.
How a Custom Multimodal Transformer Beat a Fine-Tuned LLM for Attribute
LeBonCoin's ML team built a custom late-fusion transformer that uses pre-computed visual embeddings and character n-gram text vectors to predict ad attributes. It outperformed a fine-tuned VLM while running on CPU with sub-200ms latency, offering calibrated probabilities and 15-minute retraining cycles.
DeepMind’s New VAE Matches Stable Diffusion at 10x Resolution
DeepMind’s new VAE produces 1024x1024 images with quality comparable to Stable Diffusion’s 256x256 output, potentially replacing the standard VAE in generative pipelines. This cuts the token count by 10x, enabling faster generation and lower memory usage.
Apple Releases DFNDR-12M Dataset, Claims 5x CLIP Training Efficiency
Apple has open-sourced DFNDR-12M, a multimodal dataset of 12.8 million image-text pairs with synthetic captions and pre-computed embeddings. The company claims it enables up to 5x training efficiency over standard CLIP datasets.
Prince Canuma's M3 Ultra 512GB & RTX Pro 6000 Setup for MLX Research
Independent developer Prince Canuma has assembled a powerful, community-sponsored home compute cluster for MLX research and model porting, featuring an M3 Ultra with 512GB RAM and an RTX Pro 6000.
Gur Singh Claims 7 M4 MacBooks Match A100, Calls Cloud GPU Training a 'Scam'
Developer Gur Singh posted that seven M4 MacBooks (2.9 TFLOPS each) match an NVIDIA A100's performance, calling cloud GPU training a 'scam' and advocating for distributed, consumer-hardware approaches.
NVIDIA Lyra 2.0 Launches on Hugging Face for Persistent 3D World Generation
NVIDIA has released Lyra 2.0 on Hugging Face, a framework designed to generate persistent, explorable 3D worlds at scale. It specifically addresses the core technical challenges of spatial forgetting and temporal drifting in long-horizon video generation.
OpenAI Open-Sources Agents SDK, Supports 100+ LLMs
OpenAI has open-sourced its internal Agents SDK, a lightweight framework for building multi-agent systems. It features three core primitives, works with over 100 LLMs, and has gained 18.9k GitHub stars immediately.
Jensen Huang: Nvidia is a 'Computing Company,' Not a Car
Nvidia CEO Jensen Huang, in a new interview, argued that Nvidia is a 'computing company' and not a car—a product that can be easily interchanged. This distinction underscores Nvidia's strategy to be the indispensable platform for AI infrastructure.