Alvis Lu · Notes
LLM Inference Systems
Tracing requests through serving engines, checking what the code says against what the GPU actually does, and measuring where the time goes. Each post pins its source revision and states its environment.
Multimodal Serving
How image + text requests move through SGLang, and what caching and transport change.
Mini-SGLang Code Reading
Following one request through Mini-SGLang, down to individual KV slots on the GPU.
Serving Infrastructure
Running GPU inference on real clusters.
