mlmm-env-detect

name: mlmm-env-detect description: Fallback skill for detecting the current compute environment (local / PBS / SLURM, CPU architecture, GPU, CUDA, conda env). Use only when the env is unknown — other skills assume placeholders are filled by the user or context.

mlmm Environment Detection

Purpose

Most mlmm-toolkit skills assume you already know your environment and fill in placeholders like <YOUR_QUEUE>, <NCPU>, <NGPU>, <CUDA_MODULE>, <YOUR_ENV> directly. Use this skill only when those are unknown — for example, the first time you run on a new host, or when an automated agent has no prior context about the cluster.

After running this discovery sequence, you should be able to populate every placeholder used by mlmm-hpc/SKILL.md, mlmm-install-backends/{env-cuda,core}.md, and the per-backend install pages.

Discovery sequence

Run these one by one. Each step only takes a second; the goal is to populate the placeholder table at the bottom.

1. Scheduler

if   command -v qsub   >/dev/null; then SCHED=pbs    # Torque or PBSPro
elif command -v sbatch >/dev/null; then SCHED=slurm
else SCHED=local
fi
echo "scheduler: $SCHED"

2. Architecture and OS

uname -m                          # x86_64 / aarch64
uname -s                          # Linux / Darwin
lscpu | grep -E "^(Architecture|Model name|CPU\(s\)):"

aarch64 (ARM64) means gpu4pyscf-cuda12x is not available — DFT must fall back to CPU PySCF. Other backends (UMA / MACE / Orb / AIMNet2) work on aarch64 if the wheels exist for your driver.

3. GPU

nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv 2>/dev/null \
  || echo "no NVIDIA GPU detected"

If no GPU: stay on CPU (--engine cpu, no gpus=N request in PBS). If a GPU is present, note the driver version and VRAM — both constrain which torch CUDA index and which model size you can use.

4. CUDA toolkit

Three places it might live (skill content does not transcribe paths):

# 4a. HPC modulefile (most common on managed clusters)
command -v module >/dev/null && module avail cuda 2>&1 | head -20

# 4b. System-wide install (apt/yum, /usr/local/cuda)
command -v nvcc && nvcc --version

# 4c. Inside a conda env
ls "$(conda info --base 2>/dev/null)/envs"/*/bin/nvcc 2>/dev/null

If none of the three returns anything, CUDA is not installed locally — either install it or run CPU-only.

5. PBS scheduler details (when SCHED=pbs)

qstat -Q                          # available queues + max walltime
pbsnodes -a 2>/dev/null | grep -E "^[a-z0-9]|^ *(np|properties|gpus)" | head
qstat -u "$USER"                  # your running / queued jobs

The output of pbsnodes -a tells you per-node np (CPU count) and gpus (GPU count) — these become <NCPU> and <NGPU> in PBS preambles.

6. SLURM scheduler details (when SCHED=slurm)

sinfo -o "%P %l %N %G"            # partition, max time, nodes, gres
scontrol show partition           # full partition table
squeue -u "$USER"

7. Conda envs

conda env list                    # all envs
# is mlmm installed somewhere?
for env in $(conda env list | awk '/^[a-zA-Z]/{print $1}'); do
  conda run -n "$env" python -c 'import mlmm; print("'"$env"': ", mlmm.__version__)' 2>/dev/null
done

The env that successfully imports mlmm-toolkit is your <YOUR_ENV>. If none does, see mlmm-install-backends/SKILL.md.

8. Loaded modules

command -v module >/dev/null && module list 2>&1

Mapping discovery output to placeholders

Recipe: full one-shot probe

Paste this into the host once. The output is enough to populate every placeholder used by other mlmm-* skills.

{
  echo "=== Scheduler ==="
  command -v qsub   >/dev/null && echo "PBS"
  command -v sbatch >/dev/null && echo "SLURM"
  command -v qsub >/dev/null || command -v sbatch >/dev/null || echo "local only"

  echo
  echo "=== Architecture ==="
  uname -mrs

  echo
  echo "=== GPU ==="
  nvidia-smi --query-gpu=name,memory.total,driver_version --format=csv 2>&1 || echo "no GPU"

  echo
  echo "=== CUDA modulefiles ==="
  command -v module >/dev/null && module avail cuda 2>&1 | head -20

  echo
  echo "=== PBS queues (if PBS) ==="
  command -v qstat >/dev/null && qstat -Q 2>/dev/null

  echo
  echo "=== SLURM partitions (if SLURM) ==="
  command -v sinfo >/dev/null && sinfo -o "%P %l %N %G" 2>/dev/null

  echo
  echo "=== Conda envs with mlmm ==="
  for env in $(conda env list 2>/dev/null | awk '/^[a-zA-Z]/{print $1}'); do
    conda run -n "$env" python -c \
      'import mlmm; print("'"$env"':", mlmm.__version__)' 2>/dev/null
  done
} 2>&1 | tee env_probe.txt

The resulting env_probe.txt is a self-contained summary you can refer back to for the rest of your session.

See also

• mlmm-hpc/SKILL.md — uses <YOUR_QUEUE>, <NCPU>, <NGPU>, <MEM>, <HH:MM:SS>, <CUDA_MODULE>, <YOUR_ENV> placeholders.
• mlmm-install-backends/env-cuda.md — uses driver version and <CUDA_MODULE> to pick the right torch CUDA wheel.
• mlmm-install-backends/dft.md — uses uname -m to decide between gpu4pyscf-cuda12x (x86_64) and CPU PySCF (aarch64).