# TileLang

> MIT Pythonic DSL for GPU/CPU/NPU kernels on TVM

**Category:** Python  
**Pricing:** open-source  
**URL:** https://tilelang.com/  
**GitHub:** https://github.com/tile-ai/tilelang  
**Added:** 2026-10-03

---

## The Hook

TileLang is a MIT Pythonic DSL for high-performance GEMM/attention-style kernels — write tiles in Python, lower through TVM to CUDA, ROCm, Metal, Ascend, and CPU.

## What's Good

MIT (tile-ai/tilelang). PyPI `tilelang`; docs and install at tilelang.com. Multi-backend dialect (CUDA/ROCm/Metal/LLVM/Ascend 950), LSP, puzzles, and active kernel examples (FlashAttention, block-scaled MMA). Free path is pip + your own accelerators — no SaaS seat. ★~8,258 / F831.

## The Catch

Not an agent product. Serious kernels need real GPU/NPU hardware and TVM-shaped mental models; CPU paths will not match CUDA blog numbers. Vs NVIDIA Model Optimizer: ModelOpt quantizes/prunes/distills HF models for TensorRT-LLM/vLLM; TileLang authors the kernels themselves. v0.1.x still drops legacy APIs — pin a release before production kernels.

## Verdict

MIT Pythonic kernel DSL when you write GPU/NPU tiles on TVM — not ModelOpt quantization; bring the hardware.