PyTorch

54 articles in the TruthFoundry index. Each links out to the original.

• KVCache Meetup Korea 2026 showcases LLM inference optimization work · PyTorchKR and XCENA co-hosted the 2026 KVCache technical meetup in Korea on September 30 2026.
• Ray open source project launches accelerator hardware Special Interest Group · The Ray distributed computing project announces a dedicated SIG for standardized AI accelerator support.
• PyTorch 2.15 Released With Distributed Training And Hardware Optimizations · PyTorch 2.15 launches with upgraded distributed training and cross-hardware performance improvements.
• ExecuTorch Optimizes Streaming Speech Recognition on Arm Edge Devices · Case study benchmarks Nemotron 3.5 ASR performance on Google Pixel 11 with ExecuTorch.
• PyTorch Conference 2026 Announces Hardware Backends And New Foundation Members · PyTorch Conference North America 2026 hosted 3500+ attendees in San Jose October 20–21 2026.
• PyTorch Foundation Announces 2026 Global Ambassador Cohort · PyTorch Foundation introduces 24 community ambassadors across 5 global regions for 2026.
• TorchTitan platform update adds TitanRL, SOTA models and scaling performance · PyTorch-native AI training platform TorchTitan releases major annual feature update.
• FlyDSL Accelerates PyTorch Workloads on AMD MI350 Series GPUs · FlyDSL backend delivers measurable speedups for PyTorch operations on AMD MI350 GPUs.
• Google and Meta launch TorchTPU native PyTorch backend for Google TPUs · Google and Meta release TorchTPU, open source native PyTorch support for Google TPUs.
• PyTorch Foundation adds nine new members at 2026 conference · PyTorch Foundation announced nine new members at its 2026 North America conference.
• Nebius joins PyTorch Foundation as Platinum member with governing board seat · AI cloud provider Nebius becomes Platinum member of the PyTorch Foundation.
• IBM Spyre MLIR Dataflow Backend Implements KTIR Intermediate Representation · IBM details KTIR MLIR compiler backend for the Spyre hardware accelerator.
• Nscale Joins PyTorch Foundation as Platinum Member · Nscale has joined the PyTorch Foundation as a Platinum member to support AI infrastructure and open source governance.
• NVIDIA Helion Compiler Autotunes Advanced Controls for GPU Kernels · NVIDIA Helion now jointly autotunes compiler controls alongside kernel configurations to optimize GPU performance.
• AWS announces TorchNeuron native PyTorch backend for Trainium AI chips · AWS launches TorchNeuron open source native PyTorch backend for Trainium hardware.
• Zephon adds elastic deterministic data loading to TorchTitan · Zephon eliminates GPU count variance in TorchTitan machine learning training results.
• Meta releases Muse Spark Open open-weights multimodal AI model · Meta launches Muse Spark Open, an open-weights 816B parameter multimodal AI model.
• ExecuTorch Brings PyTorch to Arm Cortex-M CPUs with CMSIS-NN · ExecuTorch enables direct PyTorch model deployment on Arm Cortex-M CPUs using CMSIS-NN acceleration.
• IBM Extends PyTorch Inductor For Spyre AI Accelerator Compiler · IBM migrated Spyre accelerator's PyTorch compiler backend to extended Inductor.
• NVIDIA Dynamo Introduces Session-Aware Inference for Agentic Workloads · NVIDIA Dynamo adds session-level identifiers to optimize inference for complex agentic workflows.
• IBM Spyre Becomes Native PyTorch Device via torch-spyre Integration · IBM's torch-spyre bridges PyTorch abstractions with the Spyre dataflow AI accelerator runtime.
• FBTriton Table Batched Embedding Kernels Outperform Legacy CUDA · FBTriton TBE GPU kernels deliver improved performance for recommendation system workloads.
• PyTorch Consolidates Media Stack Into Three Specialized Libraries · PyTorch maintainers have restructured media processing into TorchCodec, TorchVision, TorchAudio.
• PyTorch Working Group Standardizes Hardware Integration in H1 2026 · PyTorch's Accelerator Integration Working Group launched new CI and test tools to streamline hardware backend support.
• Helion Linear Backend Outperforms vLLM Defaults on Hopper GPUs · Helion's autotuned linear backend improves vLLM inference performance by over 10% on NVIDIA Hopper GPUs.
• Linux Foundation Launches PyTorch Certified Associate Pathway · Linux Foundation Education introduces a new structured pathway for the PyTorch Certified Associate certification.
• Meta Optimizes Jagged Flash Attention on NVIDIA Blackwell Using TLX · Meta engineers achieve 13% forward and 50% backward pass speedups over FA4 on Blackwell using TLX-optimized Jagged Flash Attention.
• Ray Session Guide for 2026 PyTorch Conference North America · This guide outlines key Ray-related sessions at the 2026 PyTorch Conference in San Jose.
• Torch Spyre achieves L2 PyTorch CRCR integration with agentic test pipelines · Torch Spyre implemented agentic test selection for L2 integration with PyTorch CRCR CI system.
• PyTorch Conference NA 2026 Adds Introduction Track and Associate Training · PyTorch Foundation launches new conference track and full-day training in San Jose.