# Profiling Stack: torch.profiler + Nsight Systems + Nsight Compute + MFU Hesabı

> Source: https://sukruyusufkaya.com/learn/fine-tuning-cookbook/ftc-profiling-stack-torch-profiler-nsys-ncu-mfu
> Updated: 2026-08-19T00:09:25.822Z
> Category: Fine-Tuning Cookbook (Model-by-Model)
> Module: Part I — Hardware & Memory Engineering
**TLDR:** Optimization olmadan profiling = boş laf. torch.profiler ile Python-level timing, Nsight Systems (nsys) ile kernel-level timeline, Nsight Compute (ncu) ile kernel-internal metrics, MFU (Model FLOPs Utilization) hesabı: senin Llama 3.1 8B QLoRA Lab'ın RTX 4090'da teorik 165 TFLOPs'un %ne kaçında çalışıyor? Cookbook'un sertifika gereksinimi: her Lab MFU > %35.