pytorch

Suvaditya Mukherjee of Hugging Face will present on using the PyTorch Profiler to identify performance bottlenecks and optimize model execution at PyTorch Conference North America 2026 in San Jose, October 20-21.

At the first PyTorch Conference China, PyTorch Foundation Executive Director Mark Collier argued that open source accelerates frontier AI by speeding information flow across hardware, models, and inference, citing Shopify's daily production agent iteration with PyTorch and vLLM, and announcing Alibaba Cloud, Ant Group, and Cambricon as new foundation members.

IBM and Red Hat contributors share a PyTorch community blog detailing how they built Torch Spyre on top of PyTorch's Cross-Repository CI Relay (CRCR), using an agentic test-selection pipeline and declarative YAML test-reuse framework to achieve L2 integration for out-of-tree accelerator backends.

All sessions from PyTorch Conference China 2026 are now available on YouTube, covering technical discussions on models, frameworks, and open source in AI. The upcoming PyTorch Conference North America is announced for October in San Jose.

The article announces the PyTorch Ambassador Program and a presentation at PyTorchCon North America 2026, explaining how ambassadors drive PyTorch adoption and providing conference registration details.

DanLing NestedTensor presents a PyTorch tensor abstraction for composable multi-ragged tensors, reducing padding waste in variable-size deep learning inputs and achieving significant speedups (e.g., 2.74×–3.39×) and memory efficiency improvements on A100 GPUs.

The article promotes a PyTorch Associate Training session on October 19, 2026, providing hands-on experience with PyTorch to prepare for the PTCA certification exam, instructed by Faradawn Yang of NVIDIA AI.

The article announces the PyTorch Conference North America 2026, highlighting a session on deploying hardware-aware AI agentic systems on edge devices using PyTorch and ExecuTorch, along with registration information.

The PyTorch Conference North America is an event bringing together practitioners to share insights on open-source AI and real-world applications, with sessions, networking, and workshops, taking place in San Jose from October 20-21.

MEM v3 is a memory management tool for PyTorch that dynamically adjusts batch size and gradient accumulation to prevent CUDA OOM crashes during training and fine-tuning, featuring chaos resistance, crash-proof checkpoints, and live telemetry.

Tauon is a new optimizer using polynomial and orthogonalization techniques that outperforms Muon and AdamW in initial benchmarks on a small GPT-Mini model, showing lower loss and faster step time.

This article describes Sebastian Raschka's YouTube video episode on reasoning techniques in AI, focusing on log-probability scoring and self-refinement to improve model performance without retraining.

DistribAI v2 is a platform for distributed training of AI models across multiple consumer devices, featuring improved hosting options and robust handling of edge cases.

The PyTorch Foundation is hosting an Introduction Track and PyTorch Associate Training at PyTorch Conference North America 2026 to help developers, researchers, and engineers enhance their technical skills in deep learning and AI.

The ExecuTorch Hackathon is a two-day event in San Francisco where developers form teams to build and deploy PyTorch models on edge hardware using the ExecuTorch framework, with tracks for compute, mobile+XR, and IoT, sponsored by Meta, Qualcomm, and others.

The article explores Dynamic Abliteration, a non-destructive method to suppress refusal behavior in open-weight LLMs like Qwen3-4B at runtime without permanently altering model weights, using multi-layer engram steering via PyTorch forward hooks.

The PyTorch Foundation is inviting academic open source projects to submit for a Day 0 workshop at PyTorchCon North America 2026, where selected projects will present lightning talks and receive hands-on mentorship.

vLLM is introducing hardware-agnostic layers to balance high performance on cutting-edge hardware with portability across different accelerators, addressing compatibility issues with torch.compile.

Introduces hardware-agnostic layers in vLLM to maintain high performance while ensuring portability across diverse hardware, as announced in a PyTorch Foundation blog post.

The article announces PyTorch Conference North America 2026, where Andrey Talman from Meta will discuss modernizing the PyTorch release process, including coordination with Triton and vllm, and the use of AI agentic workflows.