GPU systems engineer focused on LLM inference, CUDA kernels and compilers. Based in Vancouver.
I'm building FindTensor, an experimental compiler and runtime for LLM inference in Rust, C++ and Python. Previously, I worked with Tor Aamodt at UBC on GPU architecture simulation.
I've contributed code and bug reports to FlashInfer. My SM120 dispatch diagnostics and configuration API were incorporated upstream. I also reported and reproduced an FP8 KV calibration bug that was fixed upstream.
Open PRs cover Hopper MLA decode, FP8 prefill, attention output transforms and ALiBi for tensor parallelism.
- B200 kernels: CUDA C++ and CuTe kernels for NVIDIA SOL-ExecBench, with source and validation reports.
- KernelIndex: A GPU benchmark index that compares matching workloads and links each result to its source.
- H100 serving estimator: GPU time per request modeled from 91 published vLLM runs on two H100s, with held-out evaluation.
- SmolLM2 conformance: CPU implementations of full-sequence and KV-cache inference, with 48 numerical conformance cases.
- Tensor parallel reference: A CPU decoder block across 1, 2 and 4 worker processes, with numerical checks and 96 fault-injection cases.