deepseek-ai

deepseek-ai / TileKernels

Public

A kernel library written in tilelang

553
28
100% credibility
Found Apr 23, 2026 at 553 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

TileKernels is a collection of high-performance GPU operations optimized for large language model training and inference using the TileLang domain-specific language.

How It Works

1
🔍 Discover faster AI helpers

You hear about a tool that makes AI models run much quicker on powerful computers.

2
📥 Grab the tool

Download and add it to your setup with a simple command, no hassle.

3
🔧 Link it to your AI

Connect it to your existing AI project so it uses the speedy parts.

4
⚡ See the speed boost

Run your AI training or chatting, and watch it fly through tasks super fast.

5
✅ Test and tweak

Try different AI features like smart routing or tiny data tricks, everything works smoothly.

🚀 Supercharged AI ready

Your AI now handles huge tasks quickly, saving time and power for bigger ideas.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 553 to 553 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is TileKernels?

TileKernels delivers a fused kernel library of high-performance GPU kernels for LLM ops like top-k MoE routing, FP8/FP4/E5M6 quantization casting, batched transposes, and gating mechanisms including Engram and Manifold HyperConnection. Written in Python via TileLang DSL on PyTorch, it equips devs with torch.autograd-compatible layers that saturate Hopper/Blackwell hardware bandwidth and compute for training/inference. Users pip-install and drop in custom kernels pushing LLM throughput without CUDA C++.

Why is it gaining traction?

It hits hardware perf ceilings on memory-bound LLM workloads—faster MoE dispatch, fused quant/SwiGLU—while letting Python devs author kernels with auto-optimizations, skipping kernel github source tweaks or math kernel library intel setups. 553 stars reflect appeal for kernel library python seekers tired of slow PyTorch fallbacks; internal DeepSeek use proves real-world speed in kernel hooking library scenarios.

Who should use this?

LLM engineers optimizing MoE models with top-k gating or per-token quant on SM90/SM100 GPUs. Perfect for DeepSeek-style training runs needing fused kernel library dispatch, or inference stacks fusing RMSNorm/gating without latency spikes.

Verdict

Grab it for bleeding-edge LLM perf on supported hardware, but alpha status means expect rough edges—basic docs, ongoing code quality fixes. 1.0% credibility score and 553 stars suggest test rigorously before prod; promising kernel library linux alternative as it matures.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.