Create a new Paperspace job
{% hint style="info" %}
Note: if a project is not defined for the current working directory, and you are running in command line mode, a project configuration settings file will be created. Use--init false
or specify--project <your-project-name>
to override this behavior.
{% endhint %}
$ gradient jobs create <namespace> <command> [options...]
- Machine Type: Such as
--P100
or--C7
or--TPU
- Container: Such as
--tensorflow/tensorflow:1.5.1-gpu
- Command: Such as
"./do.sh"
$ gradient jobs create \
--name "my job" \
--container "http://dockerhub.com/mycontainer" \
--machineType "P5000" \
--command "/paperspace/run.sh"
Argument | Default | Description |
---|---|---|
name
|
[required] | Job name |
machineType
|
K80 | An optional machine type to run the job on: either 'GPU+', 'P4000', 'P5000', 'P6000', 'V100', 'K80', 'P100', or 'TPU'. |
container
|
paperspace/tensorflow-python | A reference to a docker image in a public or private docker registry, or a container name provided by Paperspace. Docker image repository references must be in lowercase and may include a tag and a hostname prefix followed by a slash; if committed the hostname defaults to that of the public Docker Hub registry. An example docker image reference: docker.io/mynamespace/myimage:mytag. A container name may be mixed case. (Designated container names are currently only provided as part of various Gradient tutorials and samples.) |
command
|
Job command/entrypoint | |
ports
|
An optional list of port mappings to open on the job cluster machine while the job is running. The port mappings are specified as 'XXXX:YYYY' where XXXX is an external port number and YYYY is an internal port number. Multiple port mappings can be provided as a comma separated list. Port numbers must be greater than 1023. Note: only /tcp protocol usage is supported. | |
workspace
|
Path to workspace directory. (Soon also will support a path to a workspace archive or git repository URL.) | |
workspaceArchive
|
Path to workspace archive. (Currently being deprecated in an upcoming version.) | |
workspaceUrl
|
Project git repository URL. (Currently being deprecated in an upcoming version.) | |
workingDirectory
|
Working directory for the experiment | |
experimentId
|
Experiment Id | |
jobEnv
|
Environmental variables in JSON String Format. Example: { "HKS_EPOCHS": 1, "HKS_MAX_EVALS": 4, "DATASET_SIZE": 10000 } |
|
project
|
$CWD | The name of the project for this job. If not provided, this is taken from the .ps_project/config.json file, or the current directory name. |
projectID
|
Project ID | |
apiKey
|
API key to use this time only | |
ignoreFiles
|
Ignore certain files from uploading |
$PS_HOST_PUBLIC_IP_ADDRESS
- the public IP address of the host machine running the job
$PS_HOST_PRIVATE_IP_ADDRESS
- the private IP address of the host machine running the job
$PS_HOSTNAME
- the hostname of the host machine running the job
These can be used in conjunction with the ports
option to send HTTP traffic to the job while it's in progress for example.
{% endhint %}
Note: to run jobs from Dockerfiles, use paperspace-node, or gradient-cli.
Gradient job containers can be created from a Dockerfile. Three options are available:
1) The job can build the container image and push it to a remote registry only. This is useful in cases where you want access to a GPU to build a GPU-enabled container but do not have one on-hand.
2) The job can build the container image and run commands against a running instance of the container without uploading to a remote registry. Useful for experimenting with gradient jobs defined by Dockerfiles and cases where you are only interested in the results of the job.
3) The job can build the container image, upload the image to a remote registry, and then run commands against a running instance of the container . Useful for building images and experimenting with gradient jobs defined by Dockerfiles while retaining the original image used.
The following new job fields are available:
Name | Type | Attributes | Description |
---|---|---|---|
useDockerfile |
boolean | <optional> | determines whether to build from Dockerfile (default false). Do not include a --container argument when using this flag. |
buildOnly |
boolean | <optional> | determines whether to only build and not run image (default false) |
registryTarget |
string | <optional> | registry location to push image to |
registryTargetUsername |
string | <optional> | registry username |
registryTargetPassword |
string | <optional> | registry password |
relDockerfilePath |
string | <optional> | relative location of dockerfile in workspace (default "./Dockerfile") |
For example, to run a job that only builds a container image and pushes to a remote registry:
gradient jobs create --apiKey XXXXXXXXXXXX --workspace https://github.com/ianmiell/simple-dockerfile --useDockerfile true --buildOnly true --registryTarget my-registry/image:0.1-test --registryTargetUsername myusername --registryTargetPassword 123456
Note that if you selected buildOnly
you should supply always a registry target and credentials.
To run a job that builds a container image, pushes to a remote registry, and then runs a command inside an instance of the running container:
gradient jobs create --apiKey XXXXXXXXXXXX --workspace https://github.com/ianmiell/simple-dockerfile --useDockerfile true --buildOnly false --command "echo hello" --registryTarget my-registry/image:0.1-test --registryTargetUsername myusername --registryTargetPassword 123456
To run a job that builds the container image and then runs an instance of the container, without pushing to a remote registry:
gradient jobs create --apiKey XXXXXXXXXXXX --workspace https://github.com/ianmiell/simple-dockerfile --useDockerfile true --buildOnly false --command "echo hello"
{% hint style="info" %}
Note: commands run during the build step from the Dockerfile, like CMD ["command", "arg1"...] are run inside the image layers as it's being built. Once the container image is ready to run, the --command
argument is used to determine what command to run against the built image.
{% endhint %}