[Feature][AMD] support GPU graphs for AMD GPUs

[vybhav-ibr]

Are there any plans to support GPU graphs for Rocm enabled AMD GPUs?
It would be nice to have!

[hughperkins]

Seems that AMD supports graphs. Per Opus:

" Yes, AMD GPUs support HIP Graphs, which are the direct equivalent of CUDA Graphs.

HIP Graphs provide the same core concept: you capture a sequence of GPU operations (kernels, memory copies, etc.) into a graph, then launch the entire graph in a single
operation. This eliminates per-launch CPU overhead, which is especially beneficial for workloads with many small kernels.

Key points:

• API parity: The HIP Graphs API mirrors CUDA Graphs closely. Functions like hipGraphCreate, hipGraphLaunch, hipStreamBeginCapture/hipStreamEndCapture correspond directly to
their CUDA counterparts (cudaGraphCreate, cudaGraphLaunch, cudaStreamBeginCapture, etc.).

• Stream capture: Just like CUDA, you can capture work submitted to a HIP stream into a graph, then instantiate and replay it.

• ROCm support: HIP Graphs are supported in ROCm (AMD's GPU compute platform). Support has been maturing over recent ROCm releases.

• Portability: If you're using HIP's CUDA compatibility layer, code using CUDA Graphs can often be ported to HIP Graphs with minimal changes (largely find-and-replace of cuda
→ hip prefixes).

Caveats:
• Feature parity has lagged slightly behind CUDA in some areas (e.g., conditional/dynamic graph nodes were added to CUDA before HIP).
• Performance characteristics can differ — the overhead reduction is real, but the magnitude depends on the specific AMD hardware and driver version.
• Always check the ROCm release notes for your target GPU to confirm which graph features are supported."

PRs welcome.

[vybhav-ibr]

The ROCm organization page has forked quadrants repo and made some commits, I will ask if they are interested in contributing this feature