AI Briefing#pytorch
The latest AI and developer news about #pytorch, with the original source and a short summary.
The latest AI and developer news about #pytorch, with the original source and a short summary.
New kernel fusion techniques for LayerNorm and RMSNorm hide up to 90% of normalization latency in GEMM kernels and achieve up to 35% speedup in attention kernels on NVIDIA B200 GPUs.
The team optimized FSDP and Expert Parallelism on 384 GH200 GPUs to reach 1,300 tokens/s per GPU, overcoming inter-node bandwidth bottlenecks.
PyTorch's Helion now uses compile-time heuristics to select fast kernel configurations without autotuning, achieving 1.18x speedup over Flash Linear Attention and 1.88x over torch.compile for matmuls on H100.
The new backend executes torch.export .pte files directly in Chrome, Safari, and other WebGPU-enabled browsers without format conversion.
The algorithm uses exponential moving averages and requires no additional hyperparameters, achieving equivalent loss in roughly 10 fewer epochs than Lloyd's k-means.
The new backend allows a single .pte file to run on Chrome, Safari, and other WebGPU browsers without conversion to other formats.
NVIDIA has disclosed an FP32 technique that accelerates MLIP molecular dynamics by up to 1.7x without retraining.
Meta has released the RecIR compiler in alpha version, enabling PyTorch recommendation models to be trained on both GPUs and TPUs.
Confirmed that removing the KL term is essential for preventing performance collapse and improving efficiency when training Qwen3 0.6B with GRPO.
The new Space verifies if popular Hugging Face models run unmodified on Trainium2, providing execution scripts and a public dataset of results.
The open-source backend uses AI agents to benchmark 841 representative models, ensuring accurate execution and reducing time to the first training step by about a third.
PyTorch 2.15 improves Mac GPU MoE decoding by 5-6.5x and supports FP8.
The open-source PyTorch TPU backend demonstrates linear scaling efficiency and high throughput on large-scale LLM training workloads.
A Hugging Face blog post details system-level profiling methods for TorchNeuron on AWS Trainium, showing that operator fusion reduces RMSNorm device time from 2,355 µs to 375 µs.
Transformers, Diffusers, and Kernels now run natively on TPUs using the TorchTPU backend, eliminating the need for torch_xla or JAX.
The update introduces native FP8 support and optimizes Metal kernels, reducing driver memory peaks from 270.135 GiB to 8.108 GiB for video decoding.
The new vllm-torchtpu plugin allows standard PyTorch models to run on TPU v7x with no code changes, using custom JAX/Pallas kernels for attention and quantization.
The tool captures runtime data to reconstruct modules locally in seconds, supporting both human debugging and autonomous AI agent workflows.
The new backend enables eager execution and native distributed APIs in Transformers, Diffusers, and Kernels, replacing the previous Optimum Neuron workflow.
The open-source library supports non-commutative geometric spaces like hyperbolic and SO(n) rotations to preserve event order in sequence modeling.
The benchmark tests nine models across three access levels, finding that the performance gap between open and closed models largely depends on their ability to diagnose issues using only config files and code.
rGPU is an open-source tool that enables PyTorch applications to execute operations on remote NVIDIA GPUs while keeping the client application local, supporting users on systems like Macs without CUDA.
TorchSharpVisual, a .NET library that converts PyTorch model architectures to PNG or Graphviz, has been released.
OpenArch, an open-source project implementing major LLM architectures like Llama and DeepSeek with a focus on readability, has been released.