deepseek-ai/TileKernels

Health v2 · maintenance only

Flags: no_releases young

Adoption not part of the score

1743 stars · 160 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TileKernels is a Python library of optimized GPU kernels for LLM operations, written in the TileLang DSL. It provides kernels for MoE routing, FP8/FP4 quantization, batched transpose, and engram/hyper-connection operations, with high-level torch.autograd wrappers for training.

Use cases

• speed up mixture of experts routing on nvidia gpus
• fp8 and fp4 quantization kernels for llm training
• fused swiglu quantization ops
• custom gpu kernels written in tilelang
• engram gating kernels with rmsnorm fusion
• benchmark gpu kernels near hardware limits

When to choose

• you need near-hardware-limit GPU kernels for MoE or quantization on SM90/SM100
• you already use TileLang and want reusable LLM kernels
• you want autograd-compatible layers for custom LLM ops

When to avoid

• you need portable kernels across AMD or older NVIDIA GPUs
• you want a production-ready, well-documented library
• you don't use PyTorch 2.10+ or CUDA 13.1+

Facets

library · maturity active

machine-learning gpu-computing llm-training llm-inference benchmarking deep-learning large-language-models gpu-computing developer-tools python tilelang gpu-kernels mixture-of-experts fp8 fp4 quantization cuda nvidia deepseek gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="deepseek-ai/TileKernels")