Pipeline Parameters¶

Pipeline parameters let you change a pipeline’s behavior at submission time without recompiling. In Kale, they come from a dedicated cell type:

# tag: pipeline-parameters
learning_rate = 0.01
batch_size = 128
num_epochs = 10
model_name = "rf"
use_gpu = False

Everything you assign in this cell becomes a top-level KFP pipeline parameter of the matching type.

Supported types¶

Kale infers the parameter type from the Python literal you assign. Supported types:

Python literal	KFP parameter type
`int` (e.g. `10`)	`int`
`float` (e.g. `0.01`)	`float`
`str` (e.g. `"rf"`)	`str`
`bool` (e.g. `True`)	`bool`

Anything else — lists, dicts, NumPy scalars, objects — will either fail at compile time or fall back to string encoding. Stick to the four basic types unless you know what you’re doing.

How parameters are used in steps¶

Once a parameter is declared, you can use it in any step cell as if it were a local variable:

# tags: step:train, prev:load
model = train(df, lr=learning_rate, epochs=num_epochs)

Kale wires learning_rate and num_epochs into the train component’s inputs automatically, using KFP’s pipeline-parameter plumbing.

Multiple parameter cells¶

You can have more than one pipeline-parameters cell. Kale concatenates them in notebook order, so the last definition of a parameter wins. This is sometimes convenient for overriding defaults near where they are used, but it can also make notebooks hard to read — prefer a single parameters cell.

Overriding parameters at submission time¶

When you click Compile and Run (or pass --run_pipeline on the CLI), Kale submits a pipeline run with default parameter values. To override them:

In the KFP UI: after the pipeline is uploaded, start a new run from the Pipelines tab. The KFP UI will prompt you for each parameter.
Via the KFP SDK: load the compiled YAML IR and call client.run_pipeline(..., params={...}). This is the recommended way to kick off runs programmatically.

Viewing parameters in the KFP UI¶

Every run page shows the parameter values it was started with, so you can correlate pipeline outputs with the inputs that produced them. Use this in combination with pipeline-metrics cells to build lightweight experiment tracking without needing a separate tracking service.

Defaults vs. placeholders¶

The value you assign in the pipeline-parameters cell is the default. It’s the value used when no override is provided at submission time. So:

# tag: pipeline-parameters
batch_size = 128   # sensible default

reads naturally as “pipeline parameter batch_size with default 128”, and you can still override it on a per-run basis.