Llm

28 May 2026 09:00 AM

A practical look at why giant LLMs do not simply run inside one pod: weight memory math, FP16 and BF16 cost, tensor parallelism, pipeline parallelism, expert parallelism, MoE active parameters, and what Kubernetes is actually scheduling.

21 May 2026 09:00 AM

Request count works for normal web apps, but it breaks down when you serve LLMs on Kubernetes. Prompt length, output length, RAG context, KV cache pressure, GPU capacity, latency, and observability are all driven by tokens, not requests.

14 May 2026 09:00 AM

A practical introduction to why LLM serving breaks the usual web-app scaling playbook: requests become token streams, latency splits into TTFT and TPOT, replicas may span GPUs or nodes, memory becomes KV cache, and autoscaling needs workload-aware signals instead of CPU alone.