Lucas Chow

Projects

Inference research, benchmarks, and kernels.

  1. Zero-Free Conditioning for Ideogram 4

    Jul 2026, OptimizationIdeogram 4 rebuilds a 53,248-wide conditioning tensor that is mostly zeros at every denoising step. Keeping only the real text rows saves 9.46 GiB at 2048² and is bitwise identical.

  2. Memory-Efficient LM-Head Inference for LLaDA

    Jul 2026, KernelsFused CUDA and Triton kernels combine the LM-head projection with a streaming logsumexp reduction, cutting peak memory 4.5× without ever materializing the logits tensor.