Highlights
Stars
FlyDSL is the Python front‑end of the project: a Flexible Layout Python DSL for expressing tiling, partitioning, data movement, and kernel structure at a high level.
Ultrafast low latency megakernel for Muse Spark Open - 1000+ tok/s
UCCL is an efficient communication library for GPUs, covering collectives, P2P (e.g., KV cache transfer, RL weight transfer), and EP (e.g., GPU-driven)
RTX 6000 Pro Wiki — Running Large LLMs (Qwen3.5-397B, Kimi-K2.5, GLM-5) on PCIe GPUs without NVLink
Megatron-LM patch for SpectralShift: extending Gated DeltaNet's effective context window via spectral reparameterization and learning-rate scaling.
Bitwise-identical token probabilities between RL rollout (vLLM) and training
🧱 Experimental library for building inference engines
A collection of EAT-discovered, verified GPU kernels across hardware
EAT: Evolving Agent Teams (NeurIPS'26). A supervisor agent evolves teams of coding agents for long-running optimization tasks.
Community registry of LLM serving-path traps that produce confidently wrong measurements: templates, tool parsers, reasoning fields, quant kernel paths, CUDA toolchains, KV allocation, eval harness…
Automatic formal verification for warp-specialized GPU kernels.
DeepInfer: High-performance, large-scale inference framework for DeepSeek models
Making large AI models cheaper, faster and more accessible
InternEvo is an open-sourced lightweight training framework aims to support model pre-training without the need for extensive dependencies.
Public feedback discussions for: GitHub Mobile, GitHub Discussions, GitHub Codespaces, GitHub Sponsors, GitHub Issues and more!
Discover, compare, and contribute to cutting-edge residual connections designed for large-scale deep learning and depth scaling.
State of the art inference performance with modern accelerators on Kubernetes
Reverse engineer anything with agents, from app behavior down to native binaries.
Community maintained hardware plugin for vLLM on MetaX GPU
lithos-metal: high-performance LLM inference kernels and serving for Apple silicon.
🔥 The living map of LLM-driven kernel generation: 280+ agents, models, benchmarks, and industry systems across NVIDIA, AMD, Ascend, TPU, and more. Continuously updated.
Pure Rust + CUDA LLM inference engine — no PyTorch, OpenAI-compatible, serves Qwen3 to Kimi-K2
Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.