Vast Gpu

SSH public key must be uploaded at https://cloud.vast.ai/manage-keys/ BEFORE creating any instance. Keys are baked into instances at creation time — if you add a key after renting, you must destroy and re-create the instance.

State File

All active vast.ai instances are tracked in vast-instances.json at the project root:

[
  {
    "instance_id": 33799165,
    "offer_id": 25831376,
    "gpu_name": "RTX_3060",
    "num_gpus": 1,
    "dph": 0.0414,
    "ssh_url": "ssh://[email protected]:58955",
    "ssh_host": "1.208.108.242",
    "ssh_port": 58955,
    "created_at": "2026-03-29T21:12:00Z",
    "status": "running",
    "experiment": "exp01_baseline",
    "estimated_hours": 4.0,
    "estimated_cost": 0.17
  }
]

This file is the source of truth for /run-experiment and /monitor-experiment to connect to vast.ai instances.

Workflow

Action: Provision (default)

Analyze the task, find the best GPU, and present cost-optimized options. This is the main entry point — called directly or automatically by /run-experiment when gpu: vast is set.

Step 1: Analyze Task Requirements

Read available context to determine what the task needs:

1. From the experiment plan (refine-logs/EXPERIMENT_PLAN.md):

   • Compute budget (total GPU-hours)
   • Hardware hints (e.g., "4x RTX 3090")
   • Model architecture and dataset size
   • Run order and per-milestone cost estimates

2. From experiment scripts (if already written):

   • Model size — scan for model class, num_parameters, config files
   • Batch size, sequence length — estimate VRAM from these
   • Dataset — estimate training time from dataset size + epochs
   • Multi-GPU — check for DataParallel, DistributedDataParallel, accelerate, deepspeed

3. From user description (if no plan/scripts exist):

   • Model name/size (e.g., "fine-tune LLaMA-7B", "train ResNet-50")
   • Dataset scale (e.g., "ImageNet", "10k samples")
   • Estimated duration (e.g., "about 2 hours")

Step 2: Determine GPU Requirements

Based on the task analysis, determine:

Step 3: Search Offers

Search across multiple GPU tiers to find the best value. Always search broadly — do NOT limit to one GPU model:

# Tier 1: Budget GPUs (good for small models, fine-tuning, ablations)
vastai search offers "gpu_ram>=<MIN_VRAM> num_gpus>=<N> reliability>0.95 inet_down>100" -o 'dph+' --storage <DISK> --limit 10

# Tier 2: If VRAM > 24 GB, also search high-VRAM cards specifically
vastai search offers "gpu_ram>=48 num_gpus>=<N> reliability>0.95" -o 'dph+' --storage <DISK> --limit 5

The output is a table with columns: ID, CUDA, N (GPU count), Model, PCIE, cpu_ghz, vCPUs, RAM, Disk, $/hr, DLP (deep learning perf), score, NV Driver, Net_up, Net_down, R (reliability %), Max_Days, mach_id, status, host_id, ports, country.

The first column (ID) is the offer ID needed for vastai create instance.

Step 4: Present Cost-Optimized Options

Present 3 options to the user, ranked by estimated total cost:

Task analysis:
- Model: [model name/size] → estimated VRAM: ~[X] GB
- Training: ~[Y] hours estimated
- Requirements: [N] GPU(s), ≥[X] GB VRAM, ~[Z] GB disk

Recommended options (sorted by estimated total cost):

| # | GPU          | VRAM  | $/hr   | Est. Hours | Est. Total | Reliability | Offer ID  |
|---|-------------|-------|--------|------------|------------|-------------|-----------|
| 1 | RTX 3060    | 12 GB | $0.04  | ~6h        | ~$0.25     | 99.4%       | 25831376  |  ← cheapest
| 2 | RTX 4090    | 24 GB | $0.28  | ~4h        | ~$1.12     | 99.2%       | 6995713   |  ← best value
| 3 | A100 SXM    | 80 GB | $0.95  | ~2h        | ~$1.90     | 99.5%       | 7023456   |  ← fastest

Option 1 is cheapest overall. Option 3 finishes fastest.
Pick a number (or type a different offer ID):

Key presentation rules:

• Always show estimated total cost ($/hr × estimated hours), not just $/hr
• Faster GPUs have shorter estimated hours (scale by relative FLOPS)
• Flag if a cheap option has reliability < 0.97 ("budget pick — 3% chance of interruption")
• If task is small (<1 hour), recommend interruptible pricing for even lower cost
• If no offers meet VRAM requirements, explain why and suggest alternatives (e.g., multi-GPU, quantization)

Relative speed scaling (approximate, for estimating hours across GPU tiers):

Use these to scale the base estimated hours across offers.

Action: Rent

Create an instance from a user-selected offer.

Step 1: Create Instance

vastai create instance <OFFER_ID> \
  --image <DOCKER_IMAGE> \
  --disk <DISK_GB> \
  --ssh \
  --direct \
  --onstart-cmd "apt-get update && apt-get install -y git screen rsync"

Default Docker image: pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel (override via CLAUDE.md image: field if set).

The output looks like:

Started. {'success': True, 'new_contract': 33799165, 'instance_api_key': '...'}

The new_contract value is the instance ID — save this for all subsequent commands.

Step 2: Wait for Instance Ready

Poll instance status every 20 seconds until it's running (typically takes 30-60 seconds, max ~5 minutes):

vastai show instances --raw | python3 -c "
import sys, json
instances = json.load(sys.stdin)
for inst in instances:
    if inst['id'] == <INSTANCE_ID>:
        print(inst['actual_status'])
"

Wait states: loading → running. If stuck in loading for >5 minutes, warn the user — the host may be slow or the image may be large.

Step 3: Get SSH Connection Details

vastai ssh-url <INSTANCE_ID>

This returns a URL in the format: ssh://root@<HOST>:<PORT>

Parse out host and port from this URL. Example:

• Input: ssh://[email protected]:58955
• Host: 1.208.108.242, Port: 58955

Important: Always use vastai ssh-url to get connection details — do NOT rely on ssh_host/ssh_port from vastai show instances, as those may point to proxy servers that differ from the direct connection endpoint.

Step 4: Verify SSH Connectivity

ssh -o StrictHostKeyChecking=no -o ConnectTimeout=15 -p <PORT> root@<HOST> "nvidia-smi && echo 'CONNECTION_OK'"

If SSH fails with "Permission denied (publickey)":

• The user's SSH key was not uploaded to https://cloud.vast.ai/manage-keys/ before the instance was created
• Fix: Destroy this instance, have user upload their key, then create a new instance. Keys are baked in at creation time — there is no way to add keys to a running instance.

If SSH fails with "Connection refused":

• The instance may still be initializing. Retry up to 3 times with 15-second intervals.

Step 5: Update State File

Write/update vast-instances.json with the new instance details including the ssh_url from Step 3, estimated hours and cost.

Step 6: Report

Vast.ai instance ready:
- Instance ID: <ID>
- GPU: <GPU_NAME> x <NUM_GPUS>
- Cost: $<DPH>/hr (estimated total: ~$<TOTAL>)
- SSH: ssh -p <PORT> root@<HOST>
- Docker: <IMAGE>

To deploy: /run-experiment (will auto-detect this instance)
To destroy when done: /vast-gpu destroy <ID>

Action: Setup

Set up the rented instance for a specific experiment. Called automatically by /run-experiment when targeting a vast.ai instance.

Step 1: Install Dependencies

ssh -p <PORT> root@<HOST> "pip install -q wandb tensorboard scipy scikit-learn pandas"

If a requirements.txt exists in the project, install that instead:

scp -P <PORT> requirements.txt root@<HOST>:/workspace/
ssh -p <PORT> root@<HOST> "pip install -q -r /workspace/requirements.txt"

Note: scp uses uppercase -P for port, while ssh uses lowercase -p.

Step 2: Sync Code

rsync -avz -e "ssh -p <PORT>" \
  --include='*.py' --include='*.yaml' --include='*.yml' --include='*.json' \
  --include='*.txt' --include='*.sh' --include='*/' \
  --exclude='*.pt' --exclude='*.pth' --exclude='*.ckpt' \
  --exclude='__pycache__' --exclude='.git' --exclude='data/' \
  --exclude='wandb/' --exclude='outputs/' \
  ./ root@<HOST>:/workspace/project/

Step 3: Verify Setup

ssh -p <PORT> root@<HOST> "cd /workspace/project && python -c 'import torch; print(f\"PyTorch {torch.__version__}, CUDA: {torch.cuda.is_available()}, GPUs: {torch.cuda.device_count()}\")'"

Expected output: PyTorch 2.1.0, CUDA: True, GPUs: 1 (or more GPUs if multi-GPU instance).

Action: Destroy

Tear down a vast.ai instance to stop billing.

Step 1: Confirm Results Collected

Before destroying, check if there are experiment results to download:

ssh -p <PORT> root@<HOST> "ls /workspace/project/results/ 2>/dev/null || echo 'NO_RESULTS_DIR'"

If results exist, download them first:

rsync -avz -e "ssh -p <PORT>" root@<HOST>:/workspace/project/results/ ./results/

Also download logs:

scp -P <PORT> root@<HOST>:/workspace/*.log ./logs/ 2>/dev/null

Step 2: Destroy Instance

vastai destroy instance <INSTANCE_ID>

Output: destroying instance <INSTANCE_ID>.

Destruction is irreversible — all data on the instance is permanently deleted.

Step 3: Update State File

Remove the instance from vast-instances.json or mark its status as destroyed.

Step 4: Report Cost

Calculate actual cost based on creation time and $/hr:

Instance <ID> destroyed.
- Duration: ~X.X hours
- Actual cost: ~$X.XX (estimated was $Y.YY)
- Results downloaded to: ./results/

Action: List

Show all active vast.ai instances:

vastai show instances

Cross-reference with vast-instances.json for experiment associations.

Action: Destroy All

Tear down all active instances (use after all experiments complete):

1. Download results from each instance
2. Destroy all instances
3. Clear vast-instances.json
4. Report total cost

Key Rules

• Task-driven selection — NEVER ask users to pick GPU models. Analyze the task, estimate requirements, present cost-optimized options with total price
• ALWAYS destroy instances when experiments are done — vast.ai bills per second, leaving instances running wastes money
• Download results before destroying — data is lost permanently on destroy
• Prefer on-demand pricing for short experiments (<2 hours). Suggest interruptible/bid pricing for long runs (>4 hours) with checkpointing
• Check reliability > 0.95 — unreliable hosts may crash mid-training
• Use --direct SSH when creating instances — faster than proxy SSH
• Always use vastai ssh-url <ID> to get connection details — the host/port from show instances may differ
• SSH keys must be uploaded BEFORE creating instances — keys are baked in at creation time. If SSH fails with "Permission denied", destroy and recreate after adding the key
• Default Docker image: pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel unless user specifies otherwise
• Working directory on instance: /workspace/ (Docker default). Code syncs to /workspace/project/
• State file vast-instances.json must stay up to date — other skills depend on it
• Show estimated total cost, not just $/hr — a $0.90/hr GPU that finishes in 2h ($1.80) beats a $0.30/hr GPU that takes 8h ($2.40)
• vastai CLI requires Python ≥ 3.10 — if system Python is older, use a conda env

CLAUDE.md Example

Users only need to set gpu: vast — no hardware preferences required:

## Vast.ai
- gpu: vast                  # tells run-experiment to use vast.ai
- auto_destroy: true         # auto-destroy after experiment completes (default: true)
- max_budget: 5.00           # optional: max total $ to spend (skill warns if estimate exceeds this)
- image: pytorch/pytorch:2.1.0-cuda12.1-cudnn8-devel  # optional: override Docker image

The skill analyzes experiment scripts and plans to determine what GPU to rent. No need to specify GPU model, VRAM, or instance count.

Composing with Other Skills

/run-experiment "train model"       ← detects gpu: vast, calls /vast-gpu provision
  ↳ /vast-gpu provision             ← analyzes task, presents options with cost
  ↳ user picks option               ← rent + setup + deploy
  ↳ /vast-gpu destroy               ← auto-destroy when done (if auto_destroy: true)

/vast-gpu provision                 ← manual: analyze task + show options
/vast-gpu rent <offer_id>           ← manual: rent a specific offer
/vast-gpu list                      ← show active instances
/vast-gpu destroy <instance_id>     ← tear down, stop billing
/vast-gpu destroy-all               ← tear down everything