100M tokens. Infinite compute. Lowest val loss wins.

The official repo of EurekaClaw

Bash is all you need - A nano claude code–like 「agent harness」, built from 0 to 1

Autoresearch for GPU kernels. Give it any PyTorch model, go to sleep, wake up to optimized Triton kernels.

AI-Driven Scientific and Algorithmic Discovery

45 tips for getting the most out of Claude Code, from basics to advanced - includes a custom status line script, cutting the system prompt in half, using Gemini CLI as Claude Code's minion, and Cla…

World Model Inference Engine

A zero-to-one guide on scaling modern transformers with n-dimensional parallelism.

Advancing Open-source World Models

Efficient Long-context Language Model Training by Core Attention Disaggregation

Miles is an enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolving with slime.

Official JAX implementation of End-to-End Test-Time Training for Long Context

Repo for Qwen Image Finetune

Sutskever 30 implementations inspired by https://papercode.vercel.app/ | For Agents, use https://github.com/pageman/Sutskever-Agent | Polyglot / Multi-Backed version at https://github.com/pageman/s…

Machine Learning Systems

My learning notes for ML SYS.

HY-World 1.5: A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency

Type annotations and runtime checking for shape and dtype of JAX/NumPy/PyTorch/etc. arrays. https://docs.kidger.site/jaxtyping/

Based on Nano-vLLM, a simple replication of vLLM with self-contained paged attention and flash attention implementation

Simple and efficient DeepSeek V3 SFT using pipeline parallel and expert parallel, with both FP8 and BF16 trainings

This repository contains a curated collection of 300+ case studies from over 80 companies, detailing practical applications and insights into machine learning (ML) system design. The contents are o…

Minimal PDF creation library. <400 LOC, zero dependencies, makes real PDFs.

[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.

SR-DiT Speedrunning ImageNet Diffusion

Curate, Annotate, and Manage Your Data in LightlyStudio.

A curated list of papers on reinforcement learning for video generation

Official Implementation of VideoDPO

A simple pip-installable Python tool to generate your HTML citation world map from your Google Scholar ID.