---
title: "CUDA Graph Capture & Replay"
description: "Record a decode step's kernel sequence once and replay it: stream capture, instantiation, what a graph fixes in place, and the two graphs an engine keeps."
canonical_url: "https://fanout.sh/inference-eng/curriculum/15-04"
md_url: "https://fanout.sh/inference-eng/curriculum/15-04.md"
access: "pro-preview"
---

# CUDA Graph Capture & Replay

Record a decode step's kernel sequence once and replay it: stream capture, instantiation, what a graph fixes in place, and the two graphs an engine keeps.

- Access: Fanout Pro (public preview)

## Public preview

A decode pass issues the same operations in the same order every step, and the host arranges each of them from scratch every time it does. The remaining move is to record the arranged sequence once and hand back the recording ever after. CUDA ships that move as the graph: the mechanism the GPU execution and memory lesson stopped at,…

## In this lesson

- Stream capture
- Instantiation and replay
- What the record fixes in place
- The engine's two graphs
- New work through an old graph

## Continue learning

Pro unlocks the complete lesson, video, practice, and code where available.

---
This representation contains public Fanout content only. Protected Pro lessons, account data, billing, checkout, and pricing are not included.

Browse the public content map: https://fanout.sh/sitemap.md