LLM Infrastructure · GPU Inference · Solutions Architecture
Upstream contributions across the LLM-inference stack, paired with production work that turns deployment requirements into reliable, on-premises GenAI systems. Focused on model serving, GPU infrastructure, and solution architecture across Linux, Kubernetes, PyTorch, CUDA, TensorRT-LLM, vLLM, and Triton.
Portfolio · PR wall · LinkedIn · CV
- Upstream inference work: correctness and performance across FlashInfer, vLLM, PyTorch, Dynamo, and the surrounding serving stack. The complete live record is on prs.wayne.is-a.dev.
- GPU infrastructure: model serving, distributed communication, quantization, and kernel paths across vLLM, TensorRT-LLM, SGLang, Triton, Dynamo, and FlashInfer.
- Solution architecture: deployment constraints, preflight validation, troubleshooting, acceptance, and production handover for on-premises LLM systems.
| Project | Evidence |
|---|---|
| trtllm-triton-serving | TensorRT-LLM vs vLLM on H100; 12 controlled studies |
| tensor-core-from-scratch | 10 CUDA matmul kernels from naive to Tensor Cores |
| inference-kernel-cookbook | Flash Attention, KV cache, and paged attention from scratch |
| nccl-collectives-bench | NCCL bandwidth, latency, NVLS, and TP-decode limits |
| nim-agent-blueprint | NVIDIA NIM agentic RAG with evaluation and observability |
| llm-security-lab | Reproducible LLM attacks and defenses |