File size: 6,102 Bytes
fc0f7bd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | ---
id: submitit_launcher
title: Submitit Launcher plugin
sidebar_label: Submitit Launcher plugin
---
[](https://pypi.org/project/hydra-submitit-launcher/)
[](https://pypistats.org/packages/hydra-submitit-launcher)
[](https://github.com/facebookresearch/hydra/tree/master/plugins/hydra_submitit_launcher/example)
[](https://github.com/facebookresearch/hydra/tree/master/plugins/hydra_submitit_launcher)
The Submitit Launcher plugin provides a [SLURM ](https://slurm.schedmd.com/documentation.html) Launcher based on [Submitit](https://github.com/facebookincubator/submitit).
### Installation
This plugin requires Hydra 1.0 (Release candidate)
```commandline
$ pip install hydra-submitit-launcher --pre
```
### Usage
Once installed, add `hydra/launcher=submitit` to your command line. Alternatively, override `hydra/launcher` in your config:
```yaml
defaults:
- hydra/launcher: submitit
```
Note that this plugin expects a valid environment in the target host. usually this means a shared file system between
the launching host and the target host.
Submitit supports 3 types of queues: auto, local and slurm. Its config looks like this
```python
class QueueType(Enum):
auto = "auto"
local = "local"
slurm = "slurm"
@dataclass
class SlurmQueueConf:
# Params are used to configure sbatch, for more info check:
# https://github.com/facebookincubator/submitit/blob/master/submitit/slurm/slurm.py
# maximum time for the job in minutes
time: int = 60
# number of cpus to use for each task
cpus_per_task: int = 10
# number of gpus to use on each node
gpus_per_node: int = 1
# number of tasks to spawn on each node
ntasks_per_node: int = 1
# number of nodes to use for the job
nodes: int = 1
# memory to reserve for the job on each node, in GB
mem: str = "${hydra.launcher.mem_limit}GB"
# slurm partition to use on the cluster
partition: Optional[str] = None
# USR1 signal delay before timeout
signal_delay_s: int = 120
# name of the job
job_name: str = "${hydra.job.name}"
# Maximum number of retries on job timeout.
# Change this only after you confirmed your code can handle re-submission
# by properly resuming from the latest stored checkpoint.
# check the following for more info on slurm_max_num_timeout
# https://github.com/facebookincubator/submitit/blob/master/docs/checkpointing.md
max_num_timeout: int = 0
@dataclass
class LocalQueueConf:
# local executor mocks the behavior of slurm locally
# maximum time for the job in minutes
timeout_min: int = 60
# number of gpus to use on each node
gpus_per_node: int = 1
# number of tasks to spawn on each node (only one node available in local executor)
tasks_per_node: int = 1
@dataclass
class AutoQueueConf:
# auto executor automatically identifies and uses available cluster
# Currently this is only slurm, but local executor can be manually forced
# instead.
# Most parameters are shared between clusters, some can be cluster specific
# cluster to use (currently either "slurm" or "local" are supported,
# None defaults to an available cluster)
cluster: Optional[str] = None
# maximum time for the job in minutes
timeout_min: int = 60
# number of cpus to use for each task
cpus_per_task: int = 1
# number of gpus to use on each node
gpus_per_node: int = 0
# number of tasks to spawn on each node
tasks_per_node: int = 1
# memory to reserve for the job on each node (in GB)
mem_gb: int = 4
# number of nodes to use for the job
nodes: int = 1
# name of the job
name: str = "${hydra.job.name}"
# following parameters are SLURM specific
# Maximum number of retries on job timeout.
# Change this only after you confirmed your code can handle re-submission
# by properly resuming from the latest stored checkpoint.
# check the following for more info on slurm_max_num_timeout
# https://github.com/facebookincubator/submitit/blob/master/docs/checkpointing.md
slurm_max_num_timeout: int = 0
# USR1 signal delay before timeout for the slurm queue
slurm_signal_delay_s: int = 30
# slurm partition to use on the cluster
slurm_partition: Optional[str] = None
@dataclass
class QueueParams:
slurm: SlurmQueueConf = SlurmQueueConf()
local: LocalQueueConf = LocalQueueConf()
auto: AutoQueueConf = AutoQueueConf()
@dataclass
class SubmititConf:
queue: QueueType = QueueType.local
folder: str = "${hydra.sweep.dir}/.${hydra.launcher.params.queue}"
queue_parameters: QueueParams = QueueParams()
```
See [Submitit documentation](https://github.com/facebookincubator/submitit) for full details about the parameters above.
An [example application](https://github.com/facebookresearch/hydra/tree/master/plugins/hydra_submitit_launcher/example) using this launcher is provided in the plugin repository.
Starting the app with `python my_app.py task=1,2,3 -m` will launch 3 executions:
```text
$ python my_app.py task=1,2,3 -m
[HYDRA] Sweep output dir : multirun/2020-05-28/15-05-22
[HYDRA] #0 : task=1
[HYDRA] #1 : task=2
[HYDRA] #2 : task=3
```
You will be able to see the output of the app in the output dir:
```commandline
$ tree
.
├── 0
│ └── my_app.log
├── 1
│ └── my_app.log
├── 2
│ └── my_app.log
└── multirun.yaml
$ cat 0/my_app.log
[2020-05-28 15:05:23,511][__main__][INFO] - Process ID 15887 executing task 1 ...
[2020-05-28 15:05:24,514][submitit][INFO] - Job completed successfully
```
|