Valohai Utils

0.7.0 verified Mon Apr 27 auth: no python

Utilities for building and running machine learning pipelines on the Valohai platform. Version 0.7.0, stable but occasional updates for Valohai API changes.

pip install valohai-utils

Common errors

error ModuleNotFoundError: No module named 'valohai' ↓

cause The import is 'valohai' not 'valohai_utils'. The package is installed as valohai-utils but import uses valohai.

fix

Install valohai-utils (pip install valohai-utils) then use 'import valohai'.

error KeyError: 'dataset' when accessing ctx.inputs['dataset'] ↓

cause The input name in code does not match the key defined in valohai.yaml inputs section.

fix

Ensure the input key in valohai.yaml matches exactly: e.g., inputs: [name: dataset] then code uses ctx.inputs['dataset'].

Warnings

breaking In v0.5+, task_context no longer auto-uploads outputs; you must explicitly call ctx.upload_outputs() or use write_output which auto-uploads. ↓

fix After writing outputs, call ctx.upload_outputs() or use ctx.write_output() which handles upload.

gotcha Inputs are accessed via comma-separated values in the pipeline YAML. If you define an input as 'dataset', you must use ctx.inputs['dataset'] — case-sensitive and matches exactly the YAML key. ↓

fix Use ctx.inputs['input_name'] exactly as defined in valohai.yaml under inputs:.

gotcha The local execution mode does not validate Valohai API tokens; scripts that work locally may fail on Valohai due to missing environment variables. ↓

fix Always test on Valohai or use valohai-utils in an environment with VALOHAI_TOKEN set.

deprecated valohai.prepare_execution() is deprecated. Use valohai.initialize() or task_context instead. ↓

fix Replace valohai.prepare_execution() with valohai.initialize() and use task_context.

Imports

valohai
wrong
```
import valohai_utils
```
correct
```
import valohai
```
valohai-utils is the PyPI package, but importable as valohai
task_context
wrong
```
from valohai.utils import task_context
```
correct
```
from valohai import task_context
```
Renamed in older versions; always use valohai.task_context
Inputs
wrong
```
from valohai import Inputs
```
correct
```
from valohai.internals.inputs import Inputs
```
Inputs helper is in internals submodule, not direct import

Quickstart

Minimal Valohai pipeline step using task_context for input/output management.

import valohai
from valohai import task_context

if __name__ == '__main__':
    with task_context() as ctx:
        # Access input files
        for file_path in ctx.inputs['dataset']:
            print(f"Processing {file_path}")
        
        # Write outputs
        for output_name in ctx.outputs:
            ctx.write_output(output_name, data=[], output_path='/some/path')