viable/strict/1775203675: Add support for twice-differentiable DTensor redistribution (#160509)
Hey there, little explorer! 👋
Imagine you have a super special toy train 🚂 that can move its wagons around to different tracks.
Sometimes, the train needs to be extra, extra careful when it moves its wagons, like when it's carrying a stack of wobbly blocks! 🧱🧱
This news is about making sure our smart computer friends (like our train driver!) can move those wobbly blocks very, very smoothly and carefully, not just once, but twice! So nothing falls down!
A kind helper named 'sladyn98' checked it, and said "Yay! It works!" 🎉 Now our computer friends can be even smarter and gentler with their special toy blocks.
Fixes #160313 Pull Request resolved: #160509 Approved by: https://github.com/sladyn98
Fixes #160313
Pull Request resolved: https://github.com/pytorch/pytorch/pull/160509 Approved by: https://github.com/sladyn98`
Assets 2
Sign in to highlight and annotate this article
Conversation starters
Daily AI Digest
Get the top 5 AI stories delivered to your inbox every morning.
More about
github![[P] Cadenza: Connect Wandb logs to agents easily for autonomous research.](https://d2xsxph8kpxj0f.cloudfront.net/310419663032563854/konzwo8nGf8Z4uZsMefwMr/default-img-robot-assembly-8kCUTaKVejALbAywaCQBiC.webp)
[P] Cadenza: Connect Wandb logs to agents easily for autonomous research.
Wandb CLI and MCP is atrocious to use with agents for full autonomous research loops. They are slow, clunky, and result in context rot. So I built a CLI tool and a Python SDK to make it easy to connect your Wandb projects and runs to your agent (clawed or otherwise). The cli tool works by allowing you to import your wandb projects and structures your runs in a way that makes it easy for agents to get a sense of the solution space of your research project. When projects are imported, only the configs and metrics are analyzed to index and store your runs. When an agent samples from this index, only the most high performing experiments are returned which reduces context rot. You can also change the behavior of the index and your agent to trade-off exploration with exploitation. Open sourcing
I built a jewelry size database for women with tiny fingers in one day
The problem If your ring size is 2, 3, or 4, most jewelry brands don't make your size. I'm 153cm with size 3 ring fingers and a 13.5cm wrist. Standard rings start at size 5. Standard bracelets are 16-18cm. They literally fall off. I got tired of googling "small ring" and finding nothing useful, so I built the database I wish existed. ## What I built A free, filterable database of 31 jewelry brands verified to carry truly small sizes. Filter by ring size, bracelet length, price, material, adjustable/not 31 brand pages with detailed size info 8 size-filter pages (e.g. "all brands with ring size 2") Ring size conversion chart (US/JP/EU/UK) Printable ring sizer tool Site: https://humancronadmin.github.io/tiny-fit-jewelry/ ## The Japanese brand discovery This was the biggest surprise. Japanese
Basic PSA. PocketPal got updated, so runs Gemma 4.
Just because I've seen a couple of "I want this on Android" questions, PocketPal got updated a few hours ago, and runs Gemma 4 2B and 4B fine. At least on my hardware (crappy little moto g84 workhorse phone). Love an app that gets regular updates. I'm going to try and squeak 26B a4 iq2 quantization into 12gigs of ram, on a fresh boot, but I'm almost certain it can't be done due to Android bloat. But yeah, 2B and 4B work fine and quickly under PocketPal. Hopefully their next one is 7-8B (not 9B), because the new Qwen 3.5 models just skip over memory caps, but the old ones didn't. Super numbers are great, running them with OS overhead and context size needs a bit smaller, to be functional on a 12gig RAM phone. Bring on the GemmaSutra 4 4B though, as another gold standard of thinking's and qu
Knowledge Map
Connected Articles — Knowledge Graph
This article is connected to other articles through shared AI topics and tags.
More in Products
Building eCourses: A Community‑First LMS SaaS (and Why You Should Build in Public)
I’m building a Learning Management System SaaS called eCourses , designed specifically for small communities and independent educators who feel priced out or over‑engineered by existing platforms. This post is the first in a series where I’ll walk through the architecture, decisions, and “lessons learned” from shipping an LMS from scratch — in public, open source, and on a tight budget. Why I Built eCourses Most LMS platforms are either: Too expensive for solo creators and small communities. Too complex for simple “course + modules + lessons + live sessions” workflows. Too rigid to let instructors experiment with their own teaching style. I wanted something that: Feels native to communities (not just single instructors). Scales technically and financially under $10/month at reasonable load
Qodo vs Cody (Sourcegraph): AI Code Review Compared (2026)
Quick Verdict Qodo and Sourcegraph Cody are both AI tools for software teams, but they solve fundamentally different problems. Qodo is a code quality platform - it reviews pull requests automatically, finds bugs through a multi-agent architecture, and generates tests to fill coverage gaps without being asked. Cody is a codebase-aware AI coding assistant - it understands your entire repository and helps developers navigate, generate, and understand code through conversation and inline completions. Choose Qodo if: your team needs automated PR review that runs on every pull request without prompting, you want proactive test generation that closes coverage gaps systematically, you work on GitLab or Azure DevOps alongside GitHub, or the open-source transparency of PR-Agent matters to your organ
XYRONIX
🌌 X Y R O N I X // NEURAL CONCEPTUALIZATION MATRIX // v1.0.0 // "Pure, decentralized creation space. Infinite boundaries. Zero latency." XYRONIX is a high-performance, infinite-plane rendering engine designed for rapid ideation, tactical diagramming, and free-form vector sketching. Operating entirely within the computational bounds of your browser, XYRONIX provides a secure, lightweight environment to map out your digital universe without relying on external server telemetry. 🚀 LIVE DEPLOYMENT: ESTABLISH UPLINK XYRONIX requires zero installation, zero server configurations, and zero package managers. It is hosted directly on the global grid via GitHub Pages. 🔗 Access the XYRONIX Interface Here THE UI To Engage Systems: Initialize: Click the uplink above to open the matrix in your prefer
Discussion
Sign in to join the discussion
No comments yet — be the first to share your thoughts!