torchrunx 🔥

By Apoorv Khandelwal and Peter Curtin

The easiest way to run PyTorch on multiple GPUs or machines.

---

torchrunx is a functional utility for distributing PyTorch code across devices. This is a more convenient, robust, and featureful alternative to CLI-based launchers, like torchrun, accelerate launch, and deepspeed.

It enables complex workflows within a single script and has useful features even if only using 1 GPU.

pip install torchrunx

Requires: Linux. If using multiple machines: SSH & shared filesystem.

---

Example: simple training loop

Suppose we have some distributed training function (which needs to run on every GPU):

def distributed_training(model: nn.Module, num_steps: int) -> nn.Module: ...

Implementation of distributed_training (click to expand)

from __future__ import annotations
import os
import torch
import torch.nn as nn

def distributed_training(model: nn.Module, num_steps: int = 10) -> nn.Module | None:
    rank = int(os.environ['RANK'])
    local_rank = int(os.environ['LOCAL_RANK'])

    model.to(local_rank)
    ddp_model = nn.parallel.DistributedDataParallel(model, device_ids=[local_rank])
    optimizer = torch.optim.AdamW(ddp_model.parameters())

    for step in range(num_steps):
        optimizer.zero_grad()

        inputs = torch.randn(5, 10).to(local_rank)
        labels = torch.randn(5, 10).to(local_rank)
        outputs = ddp_model(inputs)

        torch.nn.functional.mse_loss(outputs, labels).backward()
        optimizer.step()

    if rank == 0:
        return model.cpu()

We can distribute and run this function (e.g. on 2 machines x 2 GPUs) using torchrunx!

import logging
import torchrunx

logging.basicConfig(level=logging.INFO)

launcher = torchrunx.Launcher(
    hostnames = ["localhost", "second_machine"],  # or IP addresses
    workers_per_host = "gpu"  # default, or just: 2
)

results = launcher.run(
    distributed_training,
    model = nn.Linear(10, 10),
    num_steps = 10
)

Once completed, you can retrieve the results and process them as you wish.

trained_model: nn.Module = results.rank(0)
                     # or: results.index(hostname="localhost", local_rank=0)

# and continue your script
torch.save(trained_model.state_dict(), "outputs/model.pth")

See more examples where we fine-tune LLMs using:

• Transformers
• DeepSpeed
• PyTorch Lightning
• Accelerate

Refer to our API and Usage for many more capabilities!

---

torchrunx uniquely offers

1. An automatic launcher that "just works" for everyone 🚀

torchrunx is an SSH-based, pure-Python library that is universally easy to install.
No system-specific dependencies and orchestration for automatic multi-node distribution.

2. Conventional CLI commands 🖥️

Run familiar commands, like python my_script.py ..., and customize arguments as you wish.

Other launchers override python in a cumbersome way: e.g. torchrun --nproc_per_node=2 --nnodes=2 --node_rank=0 --master_addr=100.43.331.111 --master_port=1234 my_script.py ....

3. Support for more complex workflows in a single script 🎛️

Your workflow may have steps that are complex (e.g. pre-train, fine-tune, test) or may different parallelizations (e.g. training on 8 GPUs, testing on 1 GPU). In these cases, CLI-based launchers require each step to live in its own script. Our library treats these steps in a modular way, so they can cleanly fit together in a single script!

We clean memory leaks as we go, so previous steps won't crash or adversely affect future steps.

4. Better handling of system failures. No more zombies! 🧟

With torchrun, your "work" is inherently coupled to your main Python process. If the system kills one of your workers (e.g. due to RAM OOM or segmentation faults), there is no way to fail gracefully in Python. Your processes might hang for 10 minutes (the NCCL timeout) or become perpetual zombies.

torchrunx decouples "launcher" and "worker" processes. If the system kills a worker, our launcher immediately raises a WorkerFailure exception, which users can handle as they wish. We always clean up all nodes, so no more zombies!

5. Bonus features 🎁

• Return objects from distributed functions.
• Automatic detection of SLURM environments.
• Start multi-node training from Python notebooks!
• Our library is fully typed!
• Custom, fine-grained handling of logging, environment variables, and exception propagation. We have nice defaults too: no more interleaved logs and irrelevant exceptions!

On our roadmap: higher-order parallelism, support for debuggers, and more!