find-compute

Find Compute

Find the best place to run a job across all configured compute in .raca/clusters.yaml.

Step 1: Read the cluster config

cat .raca/clusters.yaml

Extract all configured clusters. For each, note:

• type (slurm / runpod / local)
• GPU types and VRAM
• For SLURM: partitions and accounts available

Step 2: Parse arguments

From the command arguments, extract:

• --gpus N — number of GPUs needed (default: 1)
• --time Xh — estimated job duration (default: unknown)
• --job-type — training, inference, or eval (for cost/capability reasoning)

Step 3: Check each cluster

For SLURM clusters

For each configured SLURM cluster, run these checks via raca ssh:

# Check queue status — what's running and pending
raca ssh <cluster> "squeue --format='%P %T %G %M %l' --noheader | head -30"

# Check partition availability
raca ssh <cluster> "sinfo --format='%P %G %D %a %t' --noheader"

# Verify actual access with sbatch --test-only
raca ssh <cluster> "sbatch --test-only --partition=<partition> <gpu_directive> --account=<account> --time=00:05:00 --wrap='hostname' 2>&1"

If raca ssh fails (VPN, auth), note the cluster as "unreachable" — do not skip it silently.

From the queue output, estimate wait time:

• 0 jobs in partition → idle (minutes)
• 1-5 jobs → minutes to hours
• 6-20 jobs → hours
• 20+ jobs → days — flag as congested

For RunPod

Check if the API key is available:

echo ${RUNPOD_API_KEY:+set}

If set, estimate cost:

hourly_cost = GPU_cost_per_hour (from .raca/clusters.yaml or reference table)
total_cost = hourly_cost × estimated_hours × gpu_count

RunPod GPU pricing reference (community cloud, approximate):

GPU	$/hr
H200 SXM	~$3.49
H100 SXM	~$2.49
A100 SXM	~$1.64
L40S	~$0.99
RTX 4090	~$0.44
RTX 3090	~$0.22

Note: RunPod has no queue wait — pods start in ~1-3 minutes.

For local

nvidia-smi --query-gpu=name,memory.total,utilization.gpu --format=csv,noheader 2>/dev/null \
  || system_profiler SPDisplaysDataType 2>/dev/null | grep -E "Chipset|VRAM"