def run_on_cluster(hyperparams):
# enable cluster training
cluster = SlurmCluster(hyperparam_optimizer=hyperparams,
log_path=hyperparams.logs_save_path)
# email results if your hpc supports it
cluster.notify_job_status(email='*****@*****.**',
on_done=True,
on_fail=True)
# any modules for code to run in env
cluster.add_command(f'source activate {hyperparams.conda_env}')
# pick the gpu resources
cluster.per_experiment_nb_gpus = hyperparams.gpus
cluster.per_experiment_nb_cpus = 1
cluster.per_experiment_nb_nodes = 1
#cluster.gpu_type = 'k80'
cluster.job_time = '48:00:00'
cluster.minutes_to_checkpoint_before_walltime = 5
cluster.memory_mb_per_node = 250000 #180000
# come up with a short exp name
job_display_name = hyperparams.tt_name.split('_')[0]
job_display_name = job_display_name[0:4]
# optimize across all gpus
print('submitting jobs...')
cluster.optimize_parallel_cluster_gpu(main,
nb_trials=hyperparams.nb_hopt_trials,
job_name=job_display_name)
def optimize_on_cluster(hyperparams):
# enable cluster training
# log all scripts to the test tube folder
cluster = SlurmCluster(
hyperparam_optimizer=hyperparams,
log_path=hyperparams.slurm_log_path,
)
# email for cluster coms
# cluster.notify_job_status(email='add_email_here', on_done=True, on_fail=True)
# configure cluster
cluster.per_experiment_nb_gpus = hyperparams.per_experiment_nb_gpus
cluster.per_experiment_nb_nodes = hyperparams.nb_gpu_nodes
cluster.per_experiment_nb_cpus = hyperparams.per_experiment_nb_cpus
cluster.job_time = hyperparams.job_time
cluster.gpu_type = hyperparams.gpu_type
cluster.memory_mb_per_node = 0
# any modules for code to run in env
cluster.add_command("source activate dialog")
cluster.add_command(
"export PYTHONPATH=$PYTHONPATH:/private/home/koustuvs/mlp/latentDialogAnalysis"
)
# run only on 32GB voltas
# cluster.add_slurm_cmd(cmd='constraint', value='volta32gb',
# comment='use 32gb gpus')
cluster.add_slurm_cmd(cmd="partition",
value=hyperparams.gpu_partition,
comment="use 32gb gpus")
# run hopt
# creates and submits jobs to slurm
cluster.optimize_parallel_cluster_gpu(
main,
nb_trials=hyperparams.nb_hopt_trials,
job_name=hyperparams.id + "_grid_search",
job_display_name=hyperparams.id,
)
# enable cluster training
cluster = SlurmCluster(hyperparam_optimizer=hyperparams,
log_path=hyperparams.log_path,
python_cmd='python3',
test_tube_exp_name=hyperparams.test_tube_exp_name)
# email results if your hpc supports it
cluster.notify_job_status(email='*****@*****.**', on_done=True, on_fail=True)
# any modules for code to run in env
cluster.load_modules(['python-3', 'anaconda3'])
# add commands to the non slurm portion
cluster.add_command('source activate myCondaEnv')
# can also add custom slurm commands which show up as:
# #comment
# #SBATCH --cmd=value
# ############
# cluster.add_slurm_cmd(cmd='cpus-per-task', value='1', comment='nb cpus per task')
# set job compute details (this will apply PER set of hyperparameters)
cluster.per_experiment_nb_cpus = 20
cluster.per_experiment_nb_nodes = 10
# each job (24 in total here) will use 200 cpus for each set of hyperparams
# if job_display_name is set, it's what will display in the slurm queue
cluster.optimize_parallel_cluster_cpu(train,
nb_trials=24,
job_name='first_tt_job',
job_display_name='short_name')
def run_cluster(parser, fn_main, lt_system):
params = parser.parse_args()
if params.system_mode == "3d" and "3d" not in params.model_name:
params.model_name += "_3d"
if not ':' in params.time:
params.time = f"{int(params.time):02d}:00:00"
arch = platform.uname().processor
loaded_module = ''
partition = params.partition
# if partition is None:
if arch == 'x86_64':
partition = 'npl'
elif arch == 'ppc64le':
partition = 'dcs,rpi'
if partition == 'npl':
loaded_module = "module load gcc cuda openmpi"
else:
loaded_module = "module load spectrum-mpi"
log_path = os.path.join(os.environ['HOME'], params.slurm_log_root)
cluster = SlurmCluster(hyperparam_optimizer=params,
log_path=log_path,
python_cmd="python")
# cluster.notify_job_status(email='',
# on_fail=True,
# on_done=False)
# configure cluster
cluster.per_experiment_nb_gpus = params.n_gpus
cluster.per_experiment_nb_nodes = params.num_nodes
cluster.per_experiment_nb_cpus = 0 # disable this option
cluster.job_time = params.time
cluster.minutes_to_checkpoint_before_walltime = 2 # 2 min walltime
cluster.memory_mb_per_node = int(params.n_gpus) * int(
params.cpus_per_task) * int(params.mem_per_cpu)
if params.partition is not None:
cluster.add_slurm_cmd('partition',
value=params.partition,
comment='cluster partition name')
cluster.add_slurm_cmd('ntasks-per-node',
value=params.n_gpus,
comment='#task per node')
cluster.add_slurm_cmd('cpus-per-task',
value=params.cpus_per_task,
comment='#cpu per task/gpu')
cluster.add_slurm_cmd('mem-per-cpu',
value=params.mem_per_cpu,
comment="memory per cpu")
# cluster.memory_mb_per_node = params.memory # disable this option
cluster.add_command('export PYTHONFAULTHANDLER=1')
# cluster.add_command('export NCCL_DEBUG=INFO')
cluster.add_command(loaded_module)
# Master address for multi-node training
cluster.add_command(
"export SLURM_JOB_NODELIST=$(scontrol show hostnames $SLURM_JOB_NODELIST | tr '\\n' ' ')"
)
cluster.add_command("export SLURM_NODELIST=$SLURM_JOB_NODELIST")
cluster.add_command(
"slurm_nodes=$(scontrol show hostnames $SLURM_JOB_NODELIST)")
cluster.add_command(
"export MASTER_ADDRESS=$(echo $slurm_nodes | cut -d' ' -f1)")
if params.job_name is None:
job_name = params.model_name
else:
job_name = params.job_name
# Each hyperparameter combination will use 8 gpus.
cluster.optimize_parallel_cluster_gpu(
# Function to execute
lambda par, _optimizer: fn_main(par, lt_system, _optimizer),
# Number of hyperparameter combinations to search:
nb_trials=params.nb_trials,
enable_auto_resubmit=params.auto_resubmit,
# This is what will display in the slurm queue:
job_name=job_name)