
Graft is a command-line and terminal tool that ports CUDA code to AMD's ROCm/HIP stack and then proves the port actually works. It scans each source file to classify what translates cleanly and what genuinely needs manual porting, so you know your real risk before touching a kernel. Migration is assisted by an AI model that rewrites kernels to HIP, but Graft never asks you to trust a claim: every converted file is executed and compared numerically against the original CUDA result on a real GPU, with tolerances reported per case. Verification targets are reached over SSH or a Jupyter server, including AMD Instinct accelerators in the cloud, and a TUI guides you through the whole flow. Because a host may hide PyTorch inside a container, Graft reports what it can actually see rather than guessing. The result is a migration you can defend: real measurements, evidence, and a per-file report instead of a hopeful port.
18 Oct 2026