GPU-architecture compilation

Two kinds of artifact are tied to the GPU they were built on, so oh-my-mlip builds them on your machine instead of downloading them:

• the compiled NequIP and Allegro models (.pt2), built during install;
• the CUDA kernel for CatBench's D3 dispersion correction, built on first use.

Ask your LLM

I'm moving my NequIP jobs to a machine with a different GPU. Prepare the model for it.

GPU architecture names

The architecture is the GPU's CUDA compute capability, written sm<major><minor>:

Arch	Example GPUs
sm80	A100
sm86	RTX A5000, RTX A6000, RTX 30xx
sm89	L40S, RTX 40xx

nvidia-smi --query-gpu=compute_cap --format=csv,noheader prints yours (e.g. 8.9 = sm89). A model compiled for one architecture is not guaranteed to run on another.

NequIP and Allegro models

During install, scripts/prepare_nequip_weights.py and scripts/prepare_allegro_weights.py compile each checkpoint for the GPU in the machine into models/compiled/<arch>/. See Accelerators for the compile options.

At run time resolve() fills in the architecture: it detects the current GPU with nvidia-smi and puts that arch into the calculator line's models/compiled/<arch>/... path. To write jobs for another machine, pass it explicitly:

spec = resolve("NequIP", arch="sm86")

The .pt2 for that arch must exist. Compile it on a machine with that GPU by running the prepare script there (it detects the arch from the GPU); see "Run it yourself" on Accelerators.

D3 dispersion kernel

CatBench's D3 correction (catbench.dispersion) uses a CUDA kernel that is compiled for your GPU and cached. install.sh triggers that build once when nvcc is available, so a later calculation does not stop to compile.

The build needs the CUDA toolkit's nvcc. env.sh finds the toolkit (CUDA_HOME) automatically. Check with:

source env.sh
command -v nvcc && nvcc --version

If nvcc is missing, install a CUDA toolkit and run install.sh again. The models themselves (energy and forces) do not need nvcc; without it, only D3 is unavailable, and install.sh says so.