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.
WAVE, a portable GPU ISA and toolchain that runs across various GPU architectures, has been unveiled.
It explains the relationship between compute, memory bandwidth, and overhead that determines deep learning performance, along with the underlying optimization principles.
HRM-Text has released a method for pretraining a 1B model for about $1000.
The PyTorch library for Hessian/GGN analysis has been rewritten as v1.0.0a1.
PyTorch 2.12 significantly strengthens CUDA performance, export, and ROCm support.
The article explains how NCCL optimizes GPU collective communication on top of AWS EFA.
A PoC that reduces PCIe bottlenecks using NVENC was released along with real-world measurement results for its compression approach.
Built a compiler that lowers PyTorch graphs down to CUDA using 5,000 lines of Python and raw CUDA.
Benchmark results were released for 17 local LLMs and 5 agent harnesses.
Dynamic batching boosted throughput by 3.3x when training NLLB-200 600M on an RTX 5090.
Moonshot released FlashKDA, which is up to 2.22x faster than Triton on H20.
Trained a 235M-parameter LLM from scratch using a single RTX 5080.
Meta pushed ETT% above 90%, significantly reducing training inefficiency.
An open weights model has been released that transcribes overlapping speech from multiple speakers using a single model.
TRELLIS.2 was ported to Apple Silicon's MPS, enabling 3D conversion without an NVIDIA GPU.
Ported TRELLIS.2 to run 3D generation on Apple Silicon without CUDA.
easyaligner has been released, supporting GPU-based Viterbi and text normalization.
LIDARLearn, a unified PyTorch library for 3D point clouds, has been open-sourced.
Explains why multi-timescale advantage routing breaks PPO and proposes a decoupled fix.
An open-source book and repository implementing everything from GPT-2 to Llama 3.2-3B to DeepSeek in code.
An educational PyTorch repository has been released that implements key distributed training techniques such as DP, FSDP, and TP from scratch.
Google has unveiled TorchTPU, which brings PyTorch natively onto TPUs.
Monarch is a Python API that makes it easy to run PyTorch distributed training on large-scale clusters.
Safetensors joins the PyTorch Foundation, signaling core PyTorch integration and a new roadmap.