SGLang MLX backend
Lead contributor to the MLX backend. MoE performance optimization for Apple Silicon, with merged kernels for fused SwiGLU and quantized matvec.
JL:About#
Interested in building GPU kernels and ML Systems.
Cross-substrate work: NVFP4 on NVIDIA's Blackwell, MXFP4 on AMD's MI355X, and a custom MoE Metal kernel on Apple silicon. Working across all three, I built mental models of how each architecture ticks and where they diverge and align.
Open source contributor to SGLang's MLX backend.
JL:Projects#
Amd-mxfp4-mm kernel
A fused Triton quant and shuffle kernel on MI355X, ranked on the AMD x GPU MODE leaderboard through per shape tile selection and hardcoded dispatch.
Serverless GPU inference
A small inference engine built on consumer hardware, benchmarking the path from cold start to first token.
JL:Writing#
Notes and reading, kept as a Zettelkasten.