Article series

kaggle-vllm Engineering Notes

An evidence-first record of adapting upstream vLLM to constrained, managed GPU environments.

1. Why I Built kaggle-vllm: Reproducible vLLM Inference on Dual NVIDIA T4 GPUs
2. Tensor Parallelism on Dual T4 (planned)
3. Measuring NCCL Across PCIe/PHB (planned)
4. Concurrency Crossover: TP=1 vs TP=2 (planned)
5. Moving the Runtime to AWS g4dn (planned)
6. Ray Multi-Node vLLM (planned)
7. Pipeline Parallelism Across Two Nodes (planned)

Future series

• LLM Inference from First Principles — memory, scheduling, latency, and throughput.
• GPU Topology and Distributed Inference — collectives, links, and placement.
• From Kaggle T4 to Multi-Node Serving — controlled steps from notebook constraints to clustered serving.

Planned titles are a research roadmap, not claims that experiments have already run.