OpenArch – PyTorch implementations of modern LLM architectures

Products·launched

Text LLM training from scratch with PyTorch

A clean, readable codebase that implements the full LLM training pipeline (pretraining, SFT, DPO, GRPO/RL) using only PyTorch primitives, avoiding high-level abstractions. It helps developers understand the underlying math and mechanics of LLM training.

A clean, readable codebase that implements the full LLM training pipeline (pretraining, SFT, DPO, GRPO/RL) using only PyTorch primitives, avoiding high-level abstractions. It helps developers understand the underlying math and mechanics of LLM training.

Releases·Tool Release

xllm: GitHub releases light-weight infrastructure for express LLM pre-training

xllm: lightweight infrastructure for express LLM pre-training

xllm: lightweight infrastructure for express LLM pre-training

A developer released open-source AMDGCN kernels aimed at optimizing LLM inference. The kernels target AMD hardware to improve performance.

Social·Community

Developer publishes free open-source book on LLMs with practical code

A developer released a free, open-source book on Large Language Models that bridges high-level tutorials and academic papers. The book focuses on practical code and concepts, and is available on GitHub for feedback.

A developer released a free, open-source book on Large Language Models that bridges high-level tutorials and academic papers. The book focuses on practical code and concepts, and is available on GitHub for feedback.

WebLLM: high-performance in-browser LLM inference engine shared on Hacker News