Papers

PDF research notes.

4 published

  1. Name the Traffic Before You Name the FabricInfiniBand and RoCEv2 both move RDMA. Neither repairs an inference job that did not need to leave the server. The note separates in-server links from the hop between machines. It does not report a fabric bake-off on the author's single-GPU machine.
  2. Trace the Request Before You Buy the Next GPUA serving endpoint can have requests waiting while the GPU still has room for useful work. The note separates admission, cache pressure, prefill interference, and host overhead. It uses published serving results, not a new benchmark of the author's machine.
  3. Why the Next Token Slows Down as the Chat Gets LongerOn a single machine, a local model can slow down as the chat grows because each new token rereads stored keys and values. The curves are calculated from a formula. They are not a timing run on a particular card.
  4. LLM Inference as a Moving Bottleneck: Memory, Scheduling, and Service-Level ObjectivesA bottleneck-oriented framework for autoregressive LLM inference. Evidence is verified literature and analytical calculations, not new GPU benchmarks.