CoinPower·Oct 16Solana’s Shanghai Event Was Halted — Crypto’s Return to Mainland China Just Hit a Reality CheckFor a few days, Solana’s push into mainland China looked surprisingly bold.
Yeonwoo Sung·Sep 11The Contracts Accelerate Actually Owns: Optimizer, Scheduler, and DataLoader WrappersOnce the training loop starts touching the graph that prepare assembled, what Accelerate actually owns is not the model container. It is…
Yeonwoo Sung·Sep 11Accelerator.prepare is not a wrap. It assembles an object graphIt looks like a one-line API, but prepare is an assembler that makes the model, optimizer, scheduler, and dataloader point at the same…
Yeonwoo Sung·Sep 3huggingface accelerate — DDP communicates in backward, not optimizer.stepIn distributed training, communication between processes happens in the AllReduce inside backward(), not in optimizer.step(). Gradient…A response icon1A response icon1
InBug, Blocks and Bedtime StoriesbyAdrian Benko·Jul 13How to Build High-Performance DevOps (via Accelerate)What if I told you that the fastest technology teams in the world are actually the ones with the fewest errors?
Mihailo Zoin·Oct 2, 2025NotebookLM: Three Strategies That Accelerate What You Should Already Be DoingIntroduction: Efficiency, Not Magic
Dr. Shouke Wei·Sep 1, 2025Accelerating Numerical Computing with JAXUnlocking High-Performance Machine Learning and Scientific Simulations Through Composable Transformations
Jonatan Ortiz·Aug 12, 2025Building an Audio-Reactive Haptic Feedback Video Player in SwiftUI(as Apple TV)
Lucas Mrowskovsky Paim·Jul 18, 2025Writing a High-Performance Audio Wave in SwiftUIThis is the first of a series of articles that I'll write based on my learnings from creating a video editor app for the Apple Platform.
kirouane Ayoub·Jul 17, 2025What I Learned Scaling LLMs with DDP and FSDPTraining large language models on a single GPU isn’t realistic anymore — the models are just too big. if you want to scale training across…
