Ibteshamul Haque — CUDA kernels, LLM inference systems, and on-device models for security workloads. Projects, measured results, and notes on benchmarking.https://58c0d383.ibteshamul.pages.dev/https://58c0d383.ibteshamul.pages.dev/writing/edge-llms-and-detection-paths/https://58c0d383.ibteshamul.pages.dev/writing/edge-llms-and-detection-paths/Generation takes about eight and a half seconds. Dispatch takes 16.6 microseconds. Which of those two numbers you report decides whether the design is honest.Fri, 04 Sep 2026 00:00:00 GMThttps://58c0d383.ibteshamul.pages.dev/writing/frameworks-lose-on-tiny-models/https://58c0d383.ibteshamul.pages.dev/writing/frameworks-lose-on-tiny-models/TensorRT FP16 was slower than eager PyTorch. ONNX Runtime on the CPU beat several GPU paths. The model was small enough that none of the usual assumptions applied.Fri, 04 Sep 2026 00:00:00 GMThttps://58c0d383.ibteshamul.pages.dev/writing/on-measuring-kernels/https://58c0d383.ibteshamul.pages.dev/writing/on-measuring-kernels/A kernel I had explained away as floating-point noise was a data race. The explanation was wrong in a way I could have checked in one command.Fri, 04 Sep 2026 00:00:00 GMT