GPU Infrastructure Inference Optimization Model serving

Tile Kernels

Optimized GPU kernels for LLM operations, built with TileLang.

Tile Kernels is DeepSeek's open-source Python GPU kernel library written in TileLang, providing optimized kernels for LLM operations such as MoE routing, quantization, Engram gating, Manifold HyperConnection, transpose, and high-level PyTorch autograd modeling layers.

Recent stories

1 linked story

releasePRIMARY2026-04-23

DeepSeek releases Tile Kernels with Engram, mHC, and FP4/FP8 ops for SM90 and SM100 GPUs

DeepSeek published Tile Kernels, an open-source TileLang repo covering Engram, mHC, MoE routing, and FP4/FP8 kernels, with claims that some are already used in internal training and inference. That matters because it exposes reusable low-level performance work behind DeepSeek’s stack instead of keeping the kernels fully private.