Skip to content

Launching Instances

Launch Lambda Labs GPU instances with the optimal configuration for your workload.

Quick Start

Launch with default settings:

soong start

This uses your configured defaults for model, GPU type, region, and lease duration.

Customizing Launch Parameters

Selecting a Model

Override the default model with --model:

soong start --model deepseek-r1-70b

Available models:

Model Parameters Use Case Min GPU
deepseek-r1-70b 70B (INT4) Complex reasoning A100 80GB
qwen2.5-coder-32b 32B (FP16) Fast coding, long context A6000 48GB
qwen2.5-coder-32b-int4 32B (INT4) Budget coding A10 24GB
llama-3.1-70b 70B (INT4) General purpose A100 80GB
llama-3.1-8b 8B (FP16) Simple tasks A10 24GB

See Model Management for detailed comparisons.

Choosing a GPU Type

Override the default GPU with --gpu:

soong start --model qwen2.5-coder-32b --gpu gpu_1x_a6000

GPU Selection

The configure wizard recommends the cheapest GPU that fits your model. You can always select a larger GPU, but it will cost more.

Common GPU types:

# Budget
gpu_1x_a10          # 24GB VRAM - cheapest
gpu_1x_rtx6000      # 48GB VRAM - mid-range

# High Performance
gpu_1x_a100_sxm4_80gb  # 80GB VRAM - large models
gpu_1x_h100_pcie       # 80GB VRAM - fastest

Selecting a Region

Override the default region with --region:

soong start --region us-west-2

Available regions depend on GPU availability. Common regions:

  • us-west-1 (Northern California)
  • us-west-2 (Oregon)
  • us-east-1 (Virginia)
  • us-south-1 (Texas)

Region Availability

GPU availability varies by region. Use soong available to see current capacity.

Setting Lease Duration

Set lease duration (1-8 hours) with --hours:

soong start --hours 6

Lease Limits

Maximum lease duration is 8 hours. After that, you must explicitly extend the lease or the instance will auto-terminate.

Naming Your Instance

Assign a custom name with --name:

soong start --name "feature-branch-testing"

Names help identify instances when running multiple sessions.

Launch Workflow

sequenceDiagram
    participant User
    participant CLI
    participant Lambda API
    participant Instance

    User->>CLI: soong start
    CLI->>CLI: Load configuration
    CLI->>Lambda API: Get instance type pricing
    CLI->>User: Show cost estimate
    User->>CLI: Confirm (or --yes flag)
    CLI->>Lambda API: Launch instance request
    Lambda API-->>CLI: Instance ID
    CLI->>Lambda API: Poll for status (if --wait)
    Lambda API-->>CLI: Instance ready
    CLI->>User: Instance IP and commands

Cost Confirmation

Before launching, soong shows an estimate:

┌─────────────────────────────────────┐
│ Launch Instance                     │
├─────────────────────────────────────┤
│                                     │
│ Cost Estimate                       │
│                                     │
│ GPU: 1x A6000 (48 GB)              │
│ Rate: $0.80/hour                   │
│ Duration: 4 hours                  │
│                                     │
│ Estimated cost: $3.20              │
│                                     │
└─────────────────────────────────────┘
? Proceed with launch? (Y/n)

Skip confirmation with --yes:

soong start --yes

Waiting for Instance Ready

By default, the CLI waits for the instance to be ready:

soong start --wait  # default behavior

Skip waiting to return immediately:

soong start --no-wait

When waiting is disabled, check status manually:

soong status

Examples

Scenario: Deep Reasoning Task

Launch the most capable reasoning model:

soong start \
  --model deepseek-r1-70b \
  --gpu gpu_1x_a100_sxm4_80gb \
  --hours 4 \
  --name "debug-complex-issue"

Cost: ~$12-16 for 4 hours (depending on region pricing)

Scenario: Fast Iteration Coding

Use a fast, budget-friendly coding model:

soong start \
  --model qwen2.5-coder-32b-int4 \
  --gpu gpu_1x_a10 \
  --hours 2 \
  --name "rapid-prototyping"

Cost: ~$1.20-1.60 for 2 hours

Scenario: Long Context Refactoring

Use Qwen with 32K context window:

soong start \
  --model qwen2.5-coder-32b \
  --gpu gpu_1x_a6000 \
  --hours 6 \
  --name "large-file-refactor"

Cost: ~$4.80-6.00 for 6 hours

After Launch

Once launched, the CLI displays:

Instance launched: a1b2c3d4
Instance ready at 123.45.67.89

SSH: soong ssh
Status: soong status

Next steps:

  1. Start SSH tunnel to access services
  2. Check status to monitor lease time
  3. Extend lease if you need more time

Troubleshooting

No Capacity Available

Error launching instance: No capacity available in us-west-1

Solutions:

  1. Try a different region:

    soong start --region us-west-2
    

  2. Check current availability:

    soong available
    

  3. Try a different GPU type (if model allows):

    soong start --gpu gpu_1x_rtx6000
    

SSH Keys Not Found

Error: No SSH keys found in Lambda account
Add an SSH key at: https://cloud.lambdalabs.com/ssh-keys

Solution: Add your SSH public key to Lambda Labs:

  1. Go to https://cloud.lambdalabs.com/ssh-keys
  2. Click "Add SSH Key"
  3. Paste your public key (~/.ssh/id_rsa.pub)
  4. Try launching again

Instance Launch Timeout

Instance launch timed out

Solutions:

  1. Check instance status manually:

    soong status
    

  2. The instance may still be initializing. Wait 1-2 minutes and check again.

  3. If stuck in "booting" state for >5 minutes, terminate and relaunch:

    soong stop --yes
    soong start
    

Cost Estimate Shows Different Price

Cost estimates are based on current Lambda Labs pricing, which can change. The estimate shown is for the lease duration, not the hourly rate.

Example:

Rate: $1.20/hour
Duration: 4 hours
Estimated cost: $4.80  # 4 × $1.20

Actual Costs

If you extend your lease or the instance runs beyond the lease (before auto-termination), costs will increase. Always monitor with soong status.

Best Practices

Choose the Right Model

  • Complex reasoning: DeepSeek-R1 70B
  • Fast coding: Qwen2.5-Coder 32B
  • Budget coding: Qwen2.5-Coder 32B INT4
  • Simple tasks: Llama 3.1 8B

See Model Management for detailed guidance.

Start with Shorter Leases

Begin with 2-4 hours and extend if needed. This prevents paying for unused time.

# Start short
soong start --hours 2

# Extend later if needed
soong extend 2

Use --yes for Automation

When scripting or automating launches:

soong start --yes --no-wait

Set Meaningful Names

Use descriptive names for multi-instance workflows:

soong start --name "frontend-dev-server"
soong start --name "backend-api-server"
soong start --name "ml-training-job"

Next Steps