tile-ai/tilelang

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

7288 stars · 714 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

TileLang is a Pythonic domain-specific language and compiler (built on TVM) for writing high-performance GPU/CPU/accelerator kernels such as GEMM, FlashAttention, and linear attention. It lets developers express tiled compute kernels concisely while the compiler handles low-level optimizations across CUDA, Metal, and other backends.

Use cases

• write high-performance CUDA kernels in Python
• implement FlashAttention or FlashMLA efficiently
• generate optimized GEMM and dequant GEMM kernels
• compile kernels for NVIDIA, AMD, and Apple Metal GPUs
• autotune tile sizes and layouts for peak performance
• prototype deep learning operators without writing raw CUDA

When to choose

• you need near-SOTA kernel performance with Pythonic productivity
• you target multiple GPU backends (CUDA, Metal, ROCm) from one codebase
• you are building custom attention or matrix-multiply operators
• you want TVM-based compilation with autotuning

When to avoid

• you only need off-the-shelf operators from PyTorch or cuBLAS
• you want a general-purpose programming language rather than a kernel DSL
• your workloads are CPU-only and simple
• you need long-term API stability - the project removes legacy APIs between releases

Facets

library · maturity active

compiler programming-language gpu-computing machine-learning deep-learning benchmarking compilers programming-languages gpu-computing deep-learning machine-learning performance python windows cpp dsl tvm kernel-authoring gemm flashattention cuda metal autotuning code-generation gpu linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="tile-ai/tilelang")