PIM configs have major sections like system, ai and partition. Each of these sections is covered in detail below.
Configuration should be filled in a file callled config.ini. In case of multiple AI usecase deployment, different copy of config.ini with unique partition.name values can be passed via --config flag of cli utility.
- name: Name of the P10/P11 host, can be found from HMC console
- host-address: IP address of HMC host
- username: Username of HMC user. If the user is a non-root user, make sure the user has sufficient roles to perform HMC operations as captured in user-roles.md
- password: HMC password
- image: Bootable AI container image pushed on an image registry, the AI partition will get loaded with this image.
- config-json: Configuration for your applications running inside the AI partition. Ex: In case of vLLM based bootable container image, vLLM image and config to vLLM application can be specified. Make sure to use triple quotes to wrap the json value
- auth-json: Image credentials with registry information to pull the bootable AI container image. Make sure to use triple quotes to wrap the json value
The below section of configurations helps provide details of your AI application if you intend to verify at the end of PIM deployment via the launch flow.
- request: 'yes' or 'no' to request to validate the AI app deployed as part of PIM lpar
- url: URL of the AI app to check for completion. Key in the same IP address assigned as part of the networking configuration in the URL
- headers: REST request header for "Content-Type". Use a suitable Content-Type as demanded by the AI application
- payload: payload for the REST request for your AI app. It must include the model and messages. A sample payload using the granite model is shown below.
# provide payload to use in json format inside triple quotes
payload = """
{
"model": "ibm-granite/granite-3.2-2b-instruct",
"messages": [
{
"role": "user",
"content": "What is the capital of France?"
}
]
}
"""This section covers PIM partition-related configurations
- name: name of the partition. Make sure the partition name has only alphanumeric characters and an underscore.
- Flavor: this refers to the partition's instance type. Currently, PIM supports small, medium, large and [custom](# custom-Flavor) types. small, medium and large Flavors have fixed CPU cores and memory sizes.
| Flavor | Memory | CPU |
|---|---|---|
| small | 32GB | 1 |
| medium | 64GB | 3 |
| large | 128GB | 5 |
NOTE: For more details on PIM compute flavors, refer configuration files of respective flavors partition-flavors
- Storage: size of the physical storage which needs to be attached to parititon. Make sure minumum size of the storage is 120GB. PIM attaches the disk automatically with capacity closer to specified storage value.
This section covers networking related configurations
- virtual-switch-name: name of the virtual switch connected to the Power host
- virtual-network-name: name of virtual network connected to Power host
This section covers IP address, DNS and gateway
- address: IP address that will be dedicated to the partition. Make sure the IP address is reserved in the pool
- gateway: IP address of the gateway in the virtual network
- nameserver: DNS server IP address
These are optional configurations; configure them if fine-grained control over AI Partition is needed
Use this section to use a custom flavor to provide different CPU and memory sizes other than the pre-defined flavors
SSH configurations to access the PIM partition
- username: name of the SSH user, keep it as 'pim'
- pub-key-file: Public SSH key file if the keypair is generated by the deployer
- priv-key-file: Private SSH key file if the keypair is generated by the deployer
- If pub-key-file and priv-key-file are kept empty, PIM generates SSH keypair and stores it in folder <PIM_workspace>/keys folder
URL of the bootstrap ISO. No need to override this, it will be built and provided by IBM.