CalyxFlow

CalyxFlow is a lightweight agentic artificial intelligent workflow. This workflow demonstrates the use of AI LLMs to generate modeling and simulation inputs for a scientific simulation and manage execution and analysis of a suite of simulations. Open source software release O5016

Building

• Clone with --recursive to download LAMMPS if you don't already have a LAMMPS binary

• Install python dependencies using uv:

uv sync

We primarily use vLLM as our model server. If you want to use Ollama:

./setup_ollama.sh

This also downloads llama3.1:8b-instruct-fp16. Use ./run-scripts/download_ollama_model.sh to download more models.

All scripts will start and kill the ollama server automatically.

LAMMPS

• Build with the script ./build_lammps.sh; This uses NVHPC to build LAMMMPS with the packages needed for the examples we've been testing with.
• Can pass in TAG to script: TAG=<tag> ./build_lammps.sh build dir is build-lammps-${TAG}, build-lammps-grace by default
• When building with NVHPC on Grace nodes, mpirun should be used instead of srun

Running the LAMMPS workflow

In one window, serve your model:

. .venv/bin/activate
./run-scripts/run_vllm.sh <model> <model name> ["<extra vLLM args>"] [profile] [profile dir=./vllm_profiler]

For meta-llama/Llama-3.3-70B-Instruct, I've needed to use "--quantization=fp8 --kv-cache-dtype=fp8 --cpu-offload-gb=100 --max-model-len=64K" as additional args.

In another, run the workflow:

. .venv/bin/activate
./run-scripts/run_lammps_workflow.sh <model name> <model server>=vllm <LAMMPS bindir> <LAMMPS inputs dir> [profile]