Starred repositories
A theoretical reconstruction of the Claude Mythos architecture, built from first principles using the available research literature.
cuTile is a programming model for writing parallel kernels for NVIDIA GPUs
xDiT: A Scalable Inference Engine for Diffusion Transformers (DiTs) with Massive Parallelism
Qwen-Image is a powerful image generation foundation model capable of complex text rendering and precise image editing.
Offline optimization of your disaggregated Dynamo graph
Cost-efficient and pluggable Infrastructure components for GenAI inference
VeOmni: Scaling Any Modality Model Training with Model-Centric Distributed Recipe Zoo
Allow torch tensor memory to be released and resumed later
Train speculative decoding models effortlessly and port them smoothly to SGLang serving.
An extremely fast Python package and project manager, written in Rust.
Powerful system-level package manager for Linux, macOS and Windows written in Rust – building on top of the Conda ecosystem.
DeepGEMM: clean and efficient FP8 GEMM kernels with fine-grained scaling
A version of verl to support diverse tool use [TMLR 2026]
Qwen3-Coder is the code version of Qwen3, the large language model series developed by Qwen team.
Training an LLM to use a calculator with multi-turn reinforcement learning, achieving a **62% absolute increase in evaluation accuracy**.
Implementation for FP8/INT8 Rollout for RL training without performence drop.
NVIDIA GPU metrics exporter for Prometheus leveraging DCGM
slime is an LLM post-training framework for RL Scaling.
The official repo of Pai-Megatron-Patch for LLM & VLM large scale training developed by Alibaba Cloud.
Manage multiple AI terminal agents like Claude Code, Codex, OpenCode, and Amp.
FB (Facebook) + GEMM (General Matrix-Matrix Multiplication) - https://code.fb.com/ml-applications/fbgemm/
[ICLR 2025] COAT: Compressing Optimizer States and Activation for Memory-Efficient FP8 Training
Fast, Flexible and Portable Structured Generation
Interactive visualization and analytics on ADS-B data with ClickHouse