Execute Vast.ai primary workflow: Core Workflow A. Use when implementing primary use case, building main features, or core integration tasks. Trigger with phrases like "vastai main workflow", "primary task with vastai".
npxskills add jeremylongshore/claude-code-plugins-plus-skills--skill vastai-core-workflow-aLoading…
Primary workflow for Vast.ai: search for GPU offers, provision an instance, transfer data, execute a training or inference job, collect artifacts, and destroy the instance to stop billing. This is the money-path operation for every Vast.ai user.
vastai-install-auth setupimport subprocess, json
def search_offers(gpu_name="RTX_4090", min_vram=24, min_reliability=0.95,
max_price=0.50, num_gpus=1):
"""Search Vast.ai marketplace with specific filters."""
query = (
f"num_gpus={num_gpus} gpu_name={gpu_name} "
f"gpu_ram>={min_vram} reliability>{min_reliability} "
f"inet_down>200 dph_total<={max_price} rentable=true"
)
result = subprocess.run(
["vastai", "search", "offers", query, "--order", "dph_total", "--raw"],
capture_output=True, text=True, check=True,
)
offers = json.loads(result.stdout)
print(f"Found {len(offers)} offers matching criteria")
for o in offers[:5]:
print(f" ID {o['id']}: {o['gpu_name']} {o['gpu_ram']}GB "
f"${o['dph_total']:.3f}/hr reliability={o['reliability2']:.3f}")
return offers
def provision_instance(offer_id, image, disk_gb=50, onstart_cmd=""):
"""Create an instance from the best offer."""
cmd = [
"vastai", "create", "instance", str(offer_id),
"--image", image,
"--disk", str(disk_gb),
]
if onstart_cmd:
cmd.extend(["--onstart-cmd", onstart_cmd])
result = subprocess.run(cmd, capture_output=True, text=True, check=True)
instance_info = json.loads(result.stdout)
instance_id = instance_info.get("new_contract")
print(f"Instance {instance_id} provisioning...")
return instance_id
import time
def wait_for_instance(instance_id, timeout=300):
"""Poll until instance status is 'running'."""
start = time.time()
while time.time() - start < timeout:
result = subprocess.run(
["vastai", "show", "instance", str(instance_id), "--raw"],
capture_output=True, text=True,
)
info = json.loads(result.stdout)
status = info.get("actual_status", "unknown")
print(f" Status: {status}")
if status == "running":
ssh_host = info.get("ssh_host")
ssh_port = info.get("ssh_port")
print(f" SSH: ssh -p {ssh_port} root@{ssh_host}")
return info
time.sleep(15)
raise TimeoutError(f"Instance {instance_id} did not start within {timeout}s")
# Upload training data to instance
scp -P $SSH_PORT ./data/training.tar.gz root@$SSH_HOST:/workspace/
# Execute training job remotely
ssh -p $SSH_PORT root@$SSH_HOST << 'REMOTE'
cd /workspace
tar xzf training.tar.gz
python train.py --epochs 10 --batch-size 32 --output /workspace/checkpoints/
REMOTE
def cleanup_instance(instance_id, ssh_host, ssh_port, output_dir="./results"):
"""Download results and destroy instance."""
import os
os.makedirs(output_dir, exist_ok=True)
# Download model checkpoints
subprocess.run([
"scp", "-P", str(ssh_port), "-r",
f"root@{ssh_host}:/workspace/checkpoints/",
output_dir,
], check=True)
print(f"Artifacts saved to {output_dir}")
# CRITICAL: Destroy instance to stop billing
subprocess.run(["vastai", "destroy", "instance", str(instance_id)], check=True)
print(f"Instance {instance_id} destroyed — billing stopped")
# End-to-end: search → provision → run → collect → destroy
offers = search_offers(gpu_name="RTX_4090", max_price=0.30)
if not offers:
raise RuntimeError("No offers available")
instance_id = provision_instance(
offer_id=offers[0]["id"],
image="pytorch/pytorch:2.2.0-cuda12.1-cudnn8-runtime",
disk_gb=50,
onstart_cmd="pip install transformers datasets",
)
info = wait_for_instance(instance_id)
# ... transfer data, run job, collect results ...
cleanup_instance(instance_id, info["ssh_host"], info["ssh_port"])
| Error | Cause | Solution |
|---|---|---|
| No offers match filters | GPU type or price too restrictive | Relax dph_total or try different gpu_name |
Instance stuck in loading | Docker image is very large | Use a smaller base image or pre-cached template |
| SSH timeout after creation | Firewall or key mismatch | Verify SSH key is uploaded at cloud.vast.ai |
| Job OOM killed | Insufficient GPU VRAM | Reduce batch size or search for more VRAM |
| Instance preempted (spot) | Host reclaimed interruptible instance | Use on-demand or implement checkpoint recovery |
For multi-instance orchestration and cost optimization, see vastai-core-workflow-b.
Fine-tune LLM: Search for A100 80GB offers, provision with the huggingface/transformers image, upload a LoRA config, run fine-tuning for 3 epochs, download the adapter weights, destroy the instance.
Batch inference: Provision 4 cheap RTX 4090 instances in parallel, distribute an inference dataset across them, collect results, and destroy all instances in a loop.
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Set up and use 1Password CLI (op). Use when installing the CLI, enabling desktop app integration, signing in (single or multi-account), or reading/injecting/running secrets via op.
CLI to manage emails via IMAP/SMTP. Use `himalaya` to list, read, write, reply, forward, search, and organize emails from the terminal. Supports multiple accounts and message composition with MML (MIME Meta Language).
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Create or update AgentSkills. Use when designing, structuring, or packaging skills with scripts, references, and assets.
Set up and use 1Password CLI (op). Use when installing the CLI, enabling desktop app integration, signing in (single or multi-account), or reading/injecting/running secrets via op.
CLI to manage emails via IMAP/SMTP. Use `himalaya` to list, read, write, reply, forward, search, and organize emails from the terminal. Supports multiple accounts and message composition with MML (MIME Meta Language).