ML Workflows without orchestration overhead

Build machine learning workflows from process-train-evaluate pipelines to large fan-out jobs with familiar bash and Python commands.

Use vrun and vbatch to scale cloud training and data processing while keeping your ML workflow reproducible, debuggable, and fast.

Get Started Book a demo

vbatch -P h200-8 --name train --after-success process python train_model.py

Design machine learning pipelines in clear stages

Chain tasks with after-success dependencies so each stage in your ML pipeline starts only when the previous stage passes.

Process

Run cleaning, feature extraction, and validation on CPU pools.

Train

Start GPU training only after preprocessing succeeds.

Evaluate

Gate promotion with evaluation and regression checks.

# Process data
vbatch -P cpu-16 --name process python process_data.py
# Train after processing
vbatch -P h200-8 --name train --after-success process python train_model.py
# Evaluate after training
vbatch -P cpu-8 --name eval --after-success train python evaluate_model.py

Scale with fan-out and recursive ML sub-pipelines

Process thousands of files in parallel, and nest per-file inference plus evaluation as machine learning workflows become more complex.

Fan-out data processing

Use a multiline for loop to dispatch one task per file for parallel data processing in your ML workflow.

for file in *.csv; do
  vbatch --name "$file" -P cpu-8 python process_file.py "$file"
done

Recursive workflow pattern

Create a recursive sub-pipeline for each input item, such as inference followed by evaluation per file.

# process_movie.sh:
FILE=$1
vbatch --name preprocess -P cpu-8 ./pre-process.sh $FILE
vbatch --name inference -P h100-1 --after-success preprocess ./inference $FILE 
vbatch --name upload-data -P cpu-1 ./upload-data $FILE
# Entry point
for file in *.mp4; do
  vbatch --name "$file" ./process_movie.sh "$file"
done

Why this machine learning workflow model works

You keep shell-native control while gaining cloud-native scheduling, dependency management, and reproducibility for MLOps pipelines.

Simple

Build pipelines with bash and task names, no heavy orchestration framework required.

Scalable

Fan out across many files and nodes while retaining one clear parent workflow.

Reproducible

Consistent environments and explicit dependencies reduce workflow drift over time.

Observable

Track task status, inspect logs, and troubleshoot failures at each stage quickly.

Velda Cloud

Managed cloud with instant VSCode + GPU access, plus free credit. Perfect for individual and small teams.

Get Started

Enterprise

Self hosted or dedicated infrastructure, premium support for organizations of any size.

Book a demo