LLM inference engines
LLM serving · vLLM · SGLang · Benchmarking
A 25-chapter course in Turkish on serving LLMs with vLLM, SGLang, llama.cpp and Ollama, with lab scripts for benchmarking the engines.
Overview
A course in Turkish that takes an engineer from first principles to expert level in serving large language models with vLLM, SGLang, llama.cpp and Ollama. It has 25 chapters and a Turkish–English glossary, about 138,000 words in all. Each chapter rests on research checked against current sources, and a separate agent reviewed it adversarially before it was accepted.
What it covers
How prefill and decode differ and what the KV cache costs; quantization formats from GGUF K-quants to AWQ, GPTQ and FP8; PagedAttention, continuous batching, RadixAttention and speculative decoding; and parallelism across GPUs. A chapter on benchmark method defines time to first token, time per output token, inter-token latency and goodput, so that engines can be compared fairly.
The labs
The course has two reference machines: a workstation GPU with 8 GB, where the labs run Qwen3-1.7B, and an NVIDIA H200 with 141 GB, sized for 70B models in FP8 and for mixture-of-experts models. Lab scripts drive the load, watch the GPU and turn the measurements into a report.
Private codebase
Built with
vLLM · SGLang · llama.cpp · Ollama · Python · Bash