don’t sit there and tell me you’ve suffered, you’ve never had to write a triton kernel in jinja2

when will AI engineers have an original idea

MatchAnything is an insane framework authors have tried to get every view that they can and dump them to modern keypoint matching models for instance you can match iphone map view to google aerial view, thermal camera views to day view even if the images are warped!

I gues that maybe we still do not know if the data manifold hypothesis is true?

Without further ado, The Parallelism Mesh Zoo

new in CUDA 13 time to update the GPU Glossary again

Poco, using Framework H is good friends with Kutu, who uses Framework X. Kutu prefers Framework X as it apparently delivers better speedups than H when Framework X code is written in a proper manner. But Kutu has never ever demonstrated it IRL. Just buzzword keywords. Poco, on

I spent the past few months building JAXformer: One of the first open source guides on how to scale modern transformers in JAX. Trained entirely on TPUs, it supports distributed ML, Ray tokenization, MoE, n-D parallelism and end-to-end inference. Here’s how to do it:

Someone has finally published nuances of how NCCL algorithms and protocols work. Thank you so much to the authors since documentation is so scarce! arxiv.org/abs/2507.04786

parth sf is a waterloo wrapper

Starting to see a lot of kernels called cutlass* Why Nvidia?

Fun tensor-puzzle in the wild in the recent anthropic blog post. Can anyone do it in 1 line? anthropic.com/engineering/a-…

Took me 4 days to read this blog, but totally worth it. A great detailed guide to post training. Why aren't many people talking about it. Kudos to Han Fang Karthik A Sankararaman 🇮🇳🇺🇸 for sharing their thoughts. tokens-for-thoughts.notion.site/post-training-…

The 🐐 himself just dropped the truth bomb!

This is the release I was waiting for :)

the cuda kernel in question

Today, we're shipping native support for context-parallelism to help make diffusion inference go brrr on multiple GPUs 🚀 Our CP API is made to work with two flavors of distributed attention: Ring & Ulysses. Huge thanks to Aryan V S for shipping this! Deets ⬇️

Feeling so happy that we got accepted to #NeurIPS2025 😭 This was a genuinely fulfilling piece of work, and a lot of knobs needed tinkering with. Check out the thread below for more details!

New paper 📜: Tiny Recursion Model (TRM) is a recursive reasoning approach with a tiny 7M parameters neural network that obtains 45% on ARC-AGI-1 and 8% on ARC-AGI-2, beating most LLMs. Blog: alexiajm.github.io/2025/09/29/tin… Code: github.com/SamsungSAILMon… Paper: arxiv.org/abs/2510.04871

Want to improve GPU compute/comms overlap? We just published a new short tutorial for you! A few small changes to the Pallas:MGPU matmul kernel is all it takes to turn it into an all-gather collective matmul that overlaps NVLINK comms with local compute: docs.jax.dev/en/latest/pall…