Hey, I'm Bonolo 👋 I write GPU kernels.
- 🤖 Before that: robotics and embedded work, including NAO6 robot football controllers and a full build of an automated drink dispenser (APDS9960 gesture driver rewrite).
- SGLang · #45844: stop Triton recompiling the DSA tiled decode kernel per page-table stride (do_not_specialize), with a regression test · in review
| Project | Description |
|---|---|
| baby-vllm | vLLM-style inference engine from scratch: continuous batching (mixed prefill and decode), paged KV cache with preemption and recompute, a vLLM-like LLM / SamplingParams API, and a throughput benchmark against HuggingFace · matches vLLM 0.32.0 on H100 (Qwen2.5-7B): 168.0 vs 163.7 tok/s at batch 1, 1,239 vs 1,198 at batch 8 |
| mistral7B-inference | Running inference on Mistral 7B and comparing it with a baseline HuggingFace model |
| day5-cuda-graphs | CUDA graphs experiments (Python) |
| miniOrca | Python project |
| robotics-tdp-team1 | NAO6 robot football team (University of Glasgow): balance, fall-recovery and side-step controllers |
| FlowFizzy | Automated drink dispenser (C++): APDS9960 gesture/proximity driver rewrite and memory-management refactor |