> ## Documentation Index
> Fetch the complete documentation index at: https://vastai-80aa3a82-ltxv2-serverless.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Technical FAQ

> Docker configuration, performance, and advanced topics

<script
  type="application/ld+json"
  dangerouslySetInnerHTML={{
__html: JSON.stringify({
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
  {
    "@type": "Question",
    "name": "What Docker options can I use?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Add Docker run arguments in the template configuration. For port mapping: -p 8080:8080 -p 8081:8081. For environment variables: -e TZ=UTC -e CUDA_VISIBLE_DEVICES=0. For shared memory (for PyTorch): --shm-size=32gb."
    }
  },
  {
    "@type": "Question",
    "name": "Can I use my own Docker images?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Yes! When creating a template: Specify your Docker image URL, ensure it's publicly accessible or provide auth, use standard Docker Hub, GHCR, or other registries, and include all dependencies in the image."
    }
  },
  {
    "@type": "Question",
    "name": "Why can't I run Docker inside my instance?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Docker-in-Docker is disabled for security. Alternatives: Use separate instances for different containers, build multi-service images, or use process managers like supervisord."
    }
  },
  {
    "@type": "Question",
    "name": "How can I maximize GPU utilization?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "1. Batch size optimization: Increase until GPU memory is nearly full, monitor with nvidia-smi. 2. Data pipeline: Pre-process data, use multiple data loader workers, cache datasets locally. 3. Mixed precision training using PyTorch autocast or similar frameworks."
    }
  },
  {
    "@type": "Question",
    "name": "Why is my training slower than expected?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Common issues: CPU bottleneck - Check data loading. Network I/O - Download data to local storage first. Wrong GPU mode - Ensure CUDA is enabled. Thermal throttling - Some consumer GPUs throttle. PCIe bandwidth - Multi-GPU setups may be limited."
    }
  },
  {
    "@type": "Question",
    "name": "What's the difference between instance storage and volumes?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Instance Storage: Included with every instance, deleted when instance is destroyed, size set at creation (cannot change), faster performance. Volumes: Persistent across instances, can be attached/detached, additional cost, good for datasets and checkpoints."
    }
  },
  {
    "@type": "Question",
    "name": "How do I install additional packages?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "In Jupyter terminal or SSH: For system packages: apt-get update && apt-get install -y package-name. For Python packages: pip install package-name. For Conda (if available): conda install package-name. Add to /root/onstart.sh for persistence across restarts."
    }
  },
  {
    "@type": "Question",
    "name": "How do I use specific CUDA versions?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "CUDA version depends on the Docker image. To check: nvcc --version or nvidia-smi. To use specific versions, choose appropriate templates or create custom images with your required CUDA version."
    }
  },
  {
    "@type": "Question",
    "name": "My instance won't start - how do I debug?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "1. Check instance logs for errors 2. Verify Docker image exists and is accessible 3. Check if ports are already in use 4. Ensure sufficient disk space requested 5. Try a different provider 6. Contact support with instance ID."
    }
  }
]
})
}}
/>

## Docker Configuration

### What Docker options can I use?

Add Docker run arguments in the template configuration:

```bash theme={null}
# Port mapping
-p 8080:8080 -p 8081:8081

# Environment variables  
-e TZ=UTC -e CUDA_VISIBLE_DEVICES=0

# Shared memory (for PyTorch)
--shm-size=32gb
```

### Can I use my own Docker images?

Yes! When creating a template:

1. Specify your Docker image URL
2. Ensure it's publicly accessible or provide auth
3. Use standard Docker Hub, GHCR, or other registries
4. Include all dependencies in the image

### Why can't I run Docker inside my instance?

Docker-in-Docker is disabled for security. Alternatives:

* Use separate instances for different containers
* Build multi-service images
* Use process managers like supervisord

## Performance Optimization

### How can I maximize GPU utilization?

1. **Batch size optimization:**
   * Increase until GPU memory is nearly full
   * Monitor with `nvidia-smi`

2. **Data pipeline:**
   * Pre-process data
   * Use multiple data loader workers
   * Cache datasets locally

3. **Mixed precision training:**
   ```python theme={null}
   # PyTorch example
   from torch.cuda.amp import autocast
   with autocast():
       output = model(input)
   ```

### Why is my training slower than expected?

Common issues:

* **CPU bottleneck** - Check data loading
* **Network I/O** - Download data to local storage first
* **Wrong GPU mode** - Ensure CUDA is enabled
* **Thermal throttling** - Some consumer GPUs throttle
* **PCIe bandwidth** - Multi-GPU setups may be limited

## Storage and Volumes

### What's the difference between instance storage and volumes?

**Instance Storage:**

* Included with every instance
* Deleted when instance is destroyed
* Size set at creation (cannot change)
* Faster performance

**Volumes:**

* Persistent across instances
* Can be attached/detached
* Additional cost
* Good for datasets and checkpoints

### How do I use volumes?

1. Create a volume in the Volumes section
2. Attach when creating an instance
3. Mount point specified in template
4. Data persists after instance destruction

See [Volumes Guide](/documentation/instances/volumes) for details.

## Environment Setup

### How do I install additional packages?

In Jupyter terminal or SSH:

```bash theme={null}
# System packages
apt-get update && apt-get install -y package-name

# Python packages
pip install package-name

# Conda (if available)
conda install package-name
```

Add to `/root/onstart.sh` for persistence across restarts.

### How do I use specific CUDA versions?

CUDA version depends on the Docker image. To check:

```bash theme={null}
nvcc --version
nvidia-smi
```

To use specific versions, choose appropriate templates or create custom images with your required CUDA version.

## Debugging

### How do I view instance logs?

* Through web console: Click "Logs" on instance card
* Via CLI: `vastai logs INSTANCE_ID`
* Inside instance: Check `/var/log/` directory

### My instance won't start - how do I debug?

1. Check instance logs for errors
2. Verify Docker image exists and is accessible
3. Check if ports are already in use
4. Ensure sufficient disk space requested
5. Try a different provider
6. Contact support with instance ID

### How do I monitor resource usage?

```bash theme={null}
# GPU monitoring
watch -n 1 nvidia-smi

# CPU and memory
htop

# Disk usage
df -h

# Network
iftop or nethogs

# All resources
gpustat -cp
```