Run on an on-prem cluster (advanced)¶

Run on a SLURM managed cluster¶

Lightning automates the details behind training on a SLURM-powered cluster. In contrast to the general purpose cluster above, the user does not start the jobs manually on each node and instead submits it to SLURM which schedules the resources and time for which the job is allowed to run.

Design your training script¶

To train a model using multiple nodes, do the following:

Design your LightningModule (no need to add anything specific here).

Enable DDP in the trainer

# train on 32 GPUs across 4 nodes
trainer = Trainer(accelerator="gpu", devices=8, num_nodes=4, strategy="ddp")

It’s a good idea to structure your training script like this:

# train.py
def main(args):
    model = YourLightningModule(args)

    trainer = Trainer(accelerator="gpu", devices=8, num_nodes=4, strategy="ddp")

    trainer.fit(model)


if __name__ == "__main__":
    args = ...  # you can use your CLI parser of choice, or the `LightningCLI`
    # TRAIN
    main(args)

Create the appropriate SLURM job:

# (submit.sh)
#!/bin/bash -l

# SLURM SUBMIT SCRIPT
#SBATCH --nodes=4             # This needs to match Trainer(num_nodes=...)
#SBATCH --gres=gpu:8
#SBATCH --ntasks-per-node=8   # This needs to match Trainer(devices=...)
#SBATCH --mem=0
#SBATCH --time=0-02:00:00

# activate conda env
source activate $1

# debugging flags (optional)
export NCCL_DEBUG=INFO
export PYTHONFAULTHANDLER=1

# on your cluster you might need these:
# set the network interface
# export NCCL_SOCKET_IFNAME=^docker0,lo

# might need the latest CUDA
# module load NCCL/2.4.7-1-cuda.10.0

# run script from above
srun python3 train.py

If you want auto-resubmit (read below), add this line to the submit.sh script
```
#SBATCH --signal=SIGUSR1@90
```
Submit the SLURM job
```
sbatch submit.sh
```

Enable auto wall-time resubmitions¶

When you use Lightning in a SLURM cluster, it automatically detects when it is about to run into the wall time and does the following:

Saves a temporary checkpoint.
Requeues the job.
When the job starts, it loads the temporary checkpoint.

To get this behavior make sure to add the correct signal to your SLURM script

# 90 seconds before training ends
SBATCH --signal=SIGUSR1@90

You can change this signal if your environment requires the use of a different one, for example

#SBATCH --signal=SIGHUP@90

Then, when you make your trainer, pass the requeue_signal option to the SLURMEnvironment plugin:

trainer = Trainer(plugins=[SLURMEnvironment(requeue_signal=signal.SIGHUP)])

If auto-resubmit is not desired, it can be turned off in the SLURMEnvironment plugin:

from pytorch_lightning.plugins.environments import SLURMEnvironment

trainer = Trainer(plugins=[SLURMEnvironment(auto_requeue=False)])

Build your SLURM script¶

Instead of manually building SLURM scripts, you can use the SlurmCluster object to do this for you. The SlurmCluster can also run a grid search if you pass in a HyperOptArgumentParser.

Here is an example where you run a grid search of 9 combinations of hyperparameters. See also the multi-node examples here.

# grid search 3 values of learning rate and 3 values of number of layers for your net
# this generates 9 experiments (lr=1e-3, layers=16), (lr=1e-3, layers=32),
# (lr=1e-3, layers=64), ... (lr=1e-1, layers=64)
parser = HyperOptArgumentParser(strategy="grid_search", add_help=False)
parser.opt_list("--learning_rate", default=0.001, type=float, options=[1e-3, 1e-2, 1e-1], tunable=True)
parser.opt_list("--layers", default=1, type=float, options=[16, 32, 64], tunable=True)
hyperparams = parser.parse_args()

# Slurm cluster submits 9 jobs, each with a set of hyperparams
cluster = SlurmCluster(
    hyperparam_optimizer=hyperparams,
    log_path="/some/path/to/save",
)

# OPTIONAL FLAGS WHICH MAY BE CLUSTER DEPENDENT
# which interface your nodes use for communication
cluster.add_command("export NCCL_SOCKET_IFNAME=^docker0,lo")

# see the output of the NCCL connection process
# NCCL is how the nodes talk to each other
cluster.add_command("export NCCL_DEBUG=INFO")

# setting a main port here is a good idea.
cluster.add_command("export MASTER_PORT=%r" % PORT)

# ************** DON'T FORGET THIS ***************
# MUST load the latest NCCL version
cluster.load_modules(["NCCL/2.4.7-1-cuda.10.0"])

# configure cluster
cluster.per_experiment_nb_nodes = 12
cluster.per_experiment_nb_gpus = 8

cluster.add_slurm_cmd(cmd="ntasks-per-node", value=8, comment="1 task per gpu")

# submit a script with 9 combinations of hyper params
# (lr=1e-3, layers=16), (lr=1e-3, layers=32), (lr=1e-3, layers=64), ... (lr=1e-1, layers=64)
cluster.optimize_parallel_cluster_gpu(
    main, nb_trials=9, job_name="name_for_squeue"  # how many permutations of the grid search to run
)

Troubleshooting¶

The Trainer is stuck initializing at startup, what is causing this?

You are seeing a message like this in the logs but nothing happens:

Initializing distributed: GLOBAL_RANK: 0, MEMBER: 1/4

The most likely reasons and how to fix it:

You forgot to run the python train.py command with srun: Please have a look at the SLURM template script above which includes the srun at the botton of the script.
The number of nodes or number of devices per node is configured incorrectly: There are two parametres in the SLURM submission script that determine how many processes will run your training, the #SBATCH --nodes=X setting and #SBATCH --ntasks-per-node=Y settings. The numbers there need to match what is configured in your Trainer in the code: Trainer(num_nodes=X, devices=Y). If you change the numbers, update them in BOTH places.