Triton Python Packages

liger-kernel

Efficient Triton Kernels for LLM Training

sageattention

[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.

liger-kernel-nightly

Efficient Triton Kernels for LLM Training

grouped-nf4-gemm

4-bit GPU kernels that compute directly on packed Mixture-of-Experts weights, for training and inference.

tritonparse

TritonParse: A Compiler Tracer, Visualizer, and Reproducer for Triton Kernels

dsalt

Official implementation of 'Contraction in Dense Self-Attention, and How DSALT Escapes It'

kernel-lens

Deploy any research model right away

tokenspeed-iris

AMD RAD's multi-GPU Triton-based framework for seamless multi-GPU programming

fastdeploy

Deploy DL/ ML inference pipelines with minimal extra code.

flag-gems

FlagGems is an operator library for large language models implemented in the Triton Language.

g2n

A torch.compile backend that pays attention: custom FX fusion passes, a Triton LayerNorm kernel, and a persistent compile cache. Apache-2.0 free core of the g2n platform.

clearml-serving

ClearML - Model-Serving Orchestration and Repository Solution

turboquant-pro

Consumer-aware compression for embedding indexes and LLM KV caches — compress by the metric the downstream consumer actually uses. PCA-Matryoshka + TurboQuant (27x @ 99.8% recall@10), asymmetric K/V, CUDA/Triton kernels, vLLM plugin, replayable CI-gated claims. MIT.

neurostream

Run transformers larger than memory. Weights stream from NVMe at neuron-block granularity under a hard RAM/VRAM ceiling - 142 GB across 3 shards on an 8 GB GPU. GGUF, MoE selective expert fetch, Triton fused dequant-GEMV kernels.

hyperloom-inference-optimizer

An agentic system that auto-optimizes LLM workloads on AMD GPUs.

gridbook

Out-of-tree vLLM plugin and open format spec for NVFP4-CB / FP8-CB product-codebook weights — 2-6 bit-per-weight LLM quantization served on native Blackwell tensor cores.

flash-sinkhorn

The official repository of FlashSinkhorn [ICML 2026 Oral]

late-interaction-kernels

Fused Triton kernels for late-interaction (MaxSim) scoring

deformops

Deformable sampling operators for PyTorch: multi-scale deformable attention, deformable sampling and deformable convolution

isaac-net

GPU-batched 5G network simulation for massively parallel robot learning: thousands of Isaac Lab / MJX environments, network-in-the-loop on one GPU, validated against ns-3 5G-LENA and OAI

hilbertsfc

Ultra-fast 2D & 3D Hilbert curve kernels in Python. JIT compiled, branchless, L1-cache-friendly lookup tables, loop unrolling, SIMD, and multi-threading.

triton-msl

Apple Silicon (Metal) backend for Triton: write standard @triton.jit kernels on your Mac GPU. The same source runs bit-identical on NVIDIA and AMD, so you develop kernel logic locally and rent a datacenter GPU only for the perf pass.

fdclient

Deploy DL/ ML inference pipelines with minimal extra code.

intentdsl

A programmable kernel DSL and MLIR compiler for GPU, CPU, and accelerator execution models