Archives
All the articles I've archived.
2026 2
October 1
-
A Tiny VAE for FLUX 3 Klein: 30× Smaller, 11× Faster Decoding
Distilling the FLUX 3 Image Klein autoencoder into a 2.7M-parameter decoder: what the latent space really is, how the model was trained, and how it compares with the full 84M-parameter VAE on real FLUX 3 generations.
June 1
-
Inside a Blackwell 2-CTA GEMM: A Gluon Kernel Tour
A line-by-line, interactive walk through a 2-CTA Gluon GEMM for NVIDIA Blackwell (B200): TMA, Tensor Memory, tcgen05 MMA, mbarriers, and software pipelining.
2025 1
November 1
-
RMSNorm Backward: From Derivation to a Triton Kernel
Derive RMSNorm backward step by step and implement a Triton kernel with PyTorch, numerics tips