WRITING
Open as folder >Profiler traces, compilation failures, and things I built to understand something. Mostly the second one.
MAR 202645,000 kernel launches and one global flagA picket fence in the Nsight timeline, and the shared CUDA context that explained it.FEB 2026Eleven graph breaks: chasing a 24× inference regression17 ms became 410 ms. Full-graph compilation was never happening.JAN 2026A token-merging patch that made compilation 480× slower25 ms to twelve seconds, inside a Vision Transformer integration.NOV 2025Reading Nsight like a mapThe four shapes I look for before touching any code.SEP 2025Writing a matmul that beats BLASCache-aware SIMD microkernels, packing, and why powers of two hurt.
No posts with that tag yet.