Deploy an Instant Cluster with PyTorch

This tutorial demonstrates how to use Instant Clusters with PyTorch to run distributed workloads across multiple GPUs. By leveraging PyTorch’s distributed processing capabilities and Runpod’s high-speed networking infrastructure, you can significantly accelerate your training process compared to single-GPU setups.

What you’ll learn

• How to deploy an Instant Cluster with PyTorch
• How to initialize a distributed PyTorch environment using Runpod’s pre-configured environment variables
• How to launch multi-node training with torchrun
• How local and global ranks map to GPUs across your cluster

Requirements

• A Runpod account with sufficient credits for a multi-node cluster
• Basic familiarity with PyTorch and distributed training concepts

Step 1: Deploy an Instant Cluster

1. Open the Instant Clusters page on the Runpod web interface.
2. Click Create Cluster.
3. Use the UI to name and configure your Cluster. For this walkthrough, keep Pod Count at 2 and select the option for 16x H100 SXM GPUs. Keep the Pod Template at its default setting (Runpod PyTorch).
4. Click Deploy Cluster. You should be redirected to the Instant Clusters page after a few seconds.

Step 2: Clone the PyTorch demo into each Pod

1. Click your cluster to expand the list of Pods.

2. Click on a Pod, for example CLUSTERNAME-pod-0, to expand the Pod.

3. Click Connect, then click Web Terminal.

4. In the terminal that opens, run this command to clone a basic main.py file into the Pod’s main directory:

   git clone https://github.com/murat-runpod/torch-demo.git
   

Repeat these steps for each Pod in your cluster.

Step 3: Examine the main.py file

Let’s look at the code in our main.py file:

main.py

import os
import torch
import torch.distributed as dist

def init_distributed():
   """Initialize the distributed training environment"""
   # Initialize the process group
   dist.init_process_group(backend="nccl")

   # Get local rank and global rank
   local_rank = int(os.environ["LOCAL_RANK"])
   global_rank = dist.get_rank()
   world_size = dist.get_world_size()

   # Set device for this process
   device = torch.device(f"cuda:{local_rank}")
   torch.cuda.set_device(device)

   return local_rank, global_rank, world_size, device

def cleanup_distributed():
   """Clean up the distributed environment"""
   dist.destroy_process_group()

def main():
   # Initialize distributed environment
   local_rank, global_rank, world_size, device = init_distributed()

   print(f"Running on rank {global_rank}/{world_size-1} (local rank: {local_rank}), device: {device}")

   # Your code here

   # Clean up distributed environment when done
   cleanup_distributed()

if __name__ == "__main__":
   main()

This is the minimal code necessary for initializing a distributed environment. The main() function prints the local and global rank for each GPU process (this is also where you can add your own code).

Instant Cluster environment variables

PyTorch assigns LOCAL_RANK dynamically to each process. All other environment variables are set automatically by Runpod when you deploy your cluster:

Variable	Description
MASTER_ADDR / PRIMARY_ADDR	Address of the primary node for process coordination
MASTER_PORT / PRIMARY_PORT	Port on the primary node
NUM_NODES	Number of nodes in your cluster
NUM_TRAINERS	Number of GPUs per node
NODE_RANK	This node’s rank in the cluster (0 for primary)
WORLD_SIZE	Total GPUs across all nodes

For a complete list of environment variables, see the configuration reference.

Step 4: Start the PyTorch process on each Pod

Run this command in the web terminal of each Pod to start the PyTorch process:

launcher.sh

export NCCL_DEBUG=INFO
export NCCL_SOCKET_IFNAME=ens1
torchrun \
  --nproc_per_node=$NUM_TRAINERS \
  --nnodes=$NUM_NODES \
  --node_rank=$NODE_RANK \
  --master_addr=$MASTER_ADDR \
  --master_port=$MASTER_PORT \
torch-demo/main.py

This command launches eight main.py processes per node (one per GPU in the Pod).

NCCL network configuration details

The NCCL_SOCKET_IFNAME=ens1 setting tells NCCL to use the high-speed internal network interface (ens1) for GPU-to-GPU communication between nodes. Instant Clusters provide up to 8 high-bandwidth interfaces (ens1-ens8) for inter-node traffic, separate from eth0 which handles external internet traffic.

The NCCL_DEBUG=INFO setting enables detailed logging, which is helpful for troubleshooting communication issues. For more information on NCCL configuration and troubleshooting, see the configuration reference.

The NCCL_SOCKET_IFNAME=ens1 setting is critical for proper inter-node communication. Without this configuration, nodes may attempt to communicate using external IP addresses (172.xxx range) instead of the internal network interface, leading to connection timeouts and failed distributed training jobs.

Expected output

After running the command on the last Pod, you should see output similar to this:

Running on rank 8/15 (local rank: 0), device: cuda:0
Running on rank 15/15 (local rank: 7), device: cuda:7
Running on rank 9/15 (local rank: 1), device: cuda:1
Running on rank 12/15 (local rank: 4), device: cuda:4
Running on rank 13/15 (local rank: 5), device: cuda:5
Running on rank 11/15 (local rank: 3), device: cuda:3
Running on rank 14/15 (local rank: 6), device: cuda:6
Running on rank 10/15 (local rank: 2), device: cuda:2

The first number refers to the global rank of the thread, spanning from 0 to WORLD_SIZE-1 (WORLD_SIZE = the total number of GPUs in the cluster). In our example there are two Pods of eight GPUs, so the global rank spans from 0-15. The second number is the local rank, which defines the order of GPUs within a single Pod (0-7 for this example).

The specific number and order of ranks may be different in your terminal, and the global ranks listed will be different for each Pod.

This diagram illustrates how local and global ranks are distributed across multiple Pods:

Step 5: Clean up

If you no longer need your cluster, make sure you return to the Instant Clusters page and delete your cluster to avoid incurring extra charges.

You can monitor your cluster usage and spending using the Billing Explorer at the bottom of the Billing page section under the Cluster tab.

Next steps

Now that you’ve successfully deployed and tested a PyTorch distributed application on an Instant Cluster, you can:

• Adapt your own PyTorch code to run on the cluster by modifying the distributed initialization in your scripts.
• Scale your training by adjusting the number of Pods in your cluster to handle larger models or datasets.
• Try different frameworks like Axolotl for fine-tuning large language models.
• Optimize performance by experimenting with different distributed training strategies like Data Parallel (DP), Distributed Data Parallel (DDP), or Fully Sharded Data Parallel (FSDP).
• Review the configuration reference for detailed information on environment variables, network interfaces, and troubleshooting.

For more information on distributed training with PyTorch, refer to the PyTorch Distributed Training documentation.