Research notes · systems

Models, systems, and the details between.

Start with a concrete question, then follow it into the implementation.

Written by Minshen Zhang.

About the author

The archive · by half-year

2026 · Second half

Second half

Interactive dialogue · 2026-10-08

GPU Kernel Tips: From Warps to Tile Sizes

What I picked up about GPU kernels from talking with other people, written as a dialogue tree: from coalescing and bank conflicts to warpgroups, a few hardware numbers that settle FlashAttention's block_KV, and finally FlashFFN and its extension to MoE. Open the follow-ups you care about and skip the rest.

Read the article →

2026 · First half

First half

Featured research · 2026-02-23

FlashMHF: Why Give the FFN Multiple Heads?

Can an FFN gain multi-head expressivity without paying for a larger activation footprint? Follow the parallel design and SRAM-fused path that keep the cost under control.

Read the article →
FlashMHF SRAM fused computation diagram