s2r - SLURM to Run:ai Converter

Convert SLURM batch scripts to Run:ai configurations using AI.

Installation

pip install s2r

No AWS account or credentials required — a public hosted endpoint handles conversion (rate-limited to 1000 requests per IP per day).

pip install 's2r[iam-auth]'   # only needed if you self-host behind IAM auth

Quick Start

First run — setup wizard

On the first run with no arguments s2r detects your Run:ai environment and saves the configuration to ~/.runai/runai.env:

$ s2r
s2r — SLURM to Run:ai converter

runai CLI detected in PATH.
Auto-detect project, cluster, and datasources from 'runai' CLI? [Y/n]:
Detected:
  project = osu-default
  datasources:
    cache  (HostPath)
    my-bucket  (S3)

Configure environment variables now? [Y/n]:
RUNAI_PROJECT [osu-default]:
RUNAI_BUCKET  (optional, S3 datasource name or s3://uri) [my-bucket]:
RUNAI_CACHE  (datasource name for cache) [cache]:

Saved to /home/user/.runai/runai.env:
  RUNAI_PROJECT=osu-default
  RUNAI_BUCKET=my-bucket
  RUNAI_CACHE=cache

These will be loaded automatically next time you run s2r.

Run s2r --config at any time to update the configuration.

Converting scripts

# Convert file — prints the runai shell script to stdout
s2r job.slurm

# Convert and submit in one step (the printed script is pipe-safe)
s2r job.slurm | bash

# Convert from stdin
s2r < my_slurm_script.sh
cat job.slurm | s2r

# Print the assembled prompt that would be sent to the LLM (no LLM call,
# no rate-limit charge) — useful for debugging conversion behaviour
s2r --prompt job.slurm

Note: YAML manifest output (runai workload submit --file) is not yet implemented — Run:ai 2.25 only accepts standard Kubernetes/Kubeflow kinds (Job, PyTorchJob, etc.) via that path. s2r currently emits the imperative runai training standard submit shell command instead.

Library usage

from s2r import convert_slurm_to_runai

result = convert_slurm_to_runai("""
#!/bin/bash
#SBATCH --job-name=my-job
#SBATCH --gres=gpu:2
#SBATCH --mem=32G

python train.py
""")
print(result)

How It Works

1. s2r loads ~/.runai/runai.env and injects your project/bucket/cache as context into the SLURM script
2. Signs the request with HMAC-SHA256 (no AWS credentials needed on the client)
3. POSTs to an AWS API Gateway HTTP endpoint
4. A Lambda behind the API calls Bedrock (Claude Sonnet 4.6 by default; opt into Opus 4.7 via RUNAI_MODEL=opus) to convert the script
5. Returns two fenced blocks: a TrainingWorkload YAML manifest + a runai shell script

Configuration

~/.runai/runai.env (auto-loaded on every run)

Created by s2r --config or on first run. Supports standard .env format. Shell environment variables always take precedence.

Other environment variables

export S2R_API_ENDPOINT=https://...   # custom endpoint (self-hosted)
export S2R_AWS_REGION=us-west-2       # region for SigV4 (self-hosted IAM auth)
export S2R_USE_IAM_AUTH=true          # enable IAM auth (requires s2r[iam-auth])
export S2R_VERBOSE=1                  # show detection/API warnings

Example output

Given this SLURM script:

#!/bin/bash
#SBATCH --job-name=train
#SBATCH --cpus-per-task=8
#SBATCH --mem=32G
#SBATCH --gres=gpu:2

python train.py --epochs 100

s2r produces a train.yaml (TrainingWorkload CRD) and prints the equivalent runai training standard submit command to stdout, with RUNAI_PROJECT, --s3, and --datasource filled in from your saved configuration.

S3 datasource auto-provisioning

If you enter a bucket name that is not yet registered in Run:ai, the setup wizard offers to create it automatically using your AWS credentials from ~/.aws/credentials (or AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY):

  'my-new-bucket' is not yet registered as a Run:ai datasource.
  Using AWS default profile (key: AKIAV44S...)
  Create S3 datasource 'my-new-bucket' in Run:ai now? [Y/n]:
  S3 endpoint URL [https://s3.amazonaws.com]:
  Datasource 'my-new-bucket' created successfully.

Current deployment

Self-hosting

git clone https://github.com/dirkpetersen/slurm2runai.git
cd slurm2runai/lambda
sam deploy --guided   # creates API Gateway + Lambda + DynamoDB
# then set S2R_API_ENDPOINT to the output URL

See CLAUDE.md and SETUP_JOURNEY.md for full deployment and architecture details.

Examples

See examples/ for reproduction-ready Run:ai job examples (GPU detection, R statistics, PyTorch training).

Development

pip install -e ".[dev]"
pytest
ruff check .

License

MIT — see LICENSE