A domain-specific language (DSL) based on Triton but providing higher-level abstractions.

Mirror of https://gitcode.com/Ascend/AscendNPU-IR

Triton adapter for Ascend. Mirror of https://gitcode.com/ascend/triton-ascend

Perplexity open source garden for inference technology

[CVPR 2025 Best Paper Award] VGGT: Visual Geometry Grounded Transformer

[CVPR 2023 Highlight] InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions

A framework for few-shot evaluation of language models.

Speed of Light Analysis for ML Model Runtime

A benchmark of real-world DL kernel problems

high-performance linear attention kernel library built on TileLang

Building General-Purpose Robots Based on Embodied Foundation Model

Tritonbench is a collection of PyTorch custom operators with example inputs to measure their performance.

🚀 Efficient implementations for emerging model architectures

Learning TileLang with 10 puzzles!

Community maintained hardware plugin for vLLM on Ascend

Community maintained hardware plugin for vLLM on MetaX GPU

🤖FFPA: Extends FlashAttention-2 via Split-D for large headdims, 1.5x~3×↑🎉 vs SDPA, up to 430T🎉 on H200.

[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.

A framework for efficient model inference with omni-modality models

A fast communication-overlapping library for tensor/expert parallelism on GPUs.

[NeurIPS'25 Spotlight] Adaptive Attention Sparsity with Hierarchical Top-p Pruning

Building the Virtuous Cycle for AI-driven LLM Systems

A collection of memory efficient attention operators implemented in the Triton language.

Ring attention implementation with flash attention

depyf is a tool to help you understand and adapt to PyTorch compiler torch.compile.