graph-mode-internals - LMdeploy graph mode architecture
Explains how graph mode captures and replays compute sequences for decode steps, covering runner architecture, buffer management, and vendor differences.
Tags
Updated: 2026-06-30Typical Inputs
Typical Outputs
What this skill does
- capture graph
- replay graph
- allocate buffers
- fill buffers
- update context
- get outputs
- compare vendors
Inputs
- batch size
- graph metadata
- model
- device
Outputs
- hidden states
- tokens
Requirements
- CUDA support
- NPU support
- CUDA graph runner
- NPU graph runner
