Favorita: Vertex AI Custom Training Jobs

The job-submission layer for Vertex AI: package a Python training script, submit it as a custom job or batch run, and wire it to GCP — the plumbing between your model code and Google's managed training.

Primary Outcome

A reusable way to submit any Python model as a managed Vertex AI training or batch job.

Solution

A shared job-submission layer that packages any Python model and submits it as a Vertex custom job, batch run, or pipeline.

Deliverables

  • A job submission module for Vertex custom training jobs

  • A batch-run entry point for scoring over a range

  • A pipeline submission helper for KFP runs

  • GCP client setup that the model scripts share

Strategic Context

The gap between 'model works locally' and 'model runs in production' is mostly submission plumbing — packaging, auth, config, and entry points. Getting this layer right once means every future model rides the same rails instead of reinventing them.

Technical Architecture

The gap between 'model works locally' and 'model runs in production' is mostly submission plumbing — packaging, auth, config, and entry points. Getting this layer right once means every future model rides the same rails instead of reinventing them.

Problem Statement

Working model code has no clean path to managed infrastructure — every model reinvents packaging, auth, and submission plumbing.

Links

What's Included

The vertex/jobs directory: submit.py and submit_pipeline.py for job and pipeline submission, run.py and run_batch.py entry points, and gcp.py for shared client setup.

FAQs

Does this train the model itself?

No — it submits and runs the model scripts. The model logic lives in the XGBoost, random forest, and time-series templates.

Custom jobs or AutoML?

This is the custom-training path, so you control the model code end to end.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4

Primary Outcome

A reusable way to submit any Python model as a managed Vertex AI training or batch job.

Problem Statement

Working model code has no clean path to managed infrastructure — every model reinvents packaging, auth, and submission plumbing.

Solution

A shared job-submission layer that packages any Python model and submits it as a Vertex custom job, batch run, or pipeline.

Links

Deliverables

  • A job submission module for Vertex custom training jobs

  • A batch-run entry point for scoring over a range

  • A pipeline submission helper for KFP runs

  • GCP client setup that the model scripts share

What's Included

The vertex/jobs directory: submit.py and submit_pipeline.py for job and pipeline submission, run.py and run_batch.py entry points, and gcp.py for shared client setup.

Strategic Context

The gap between 'model works locally' and 'model runs in production' is mostly submission plumbing — packaging, auth, config, and entry points. Getting this layer right once means every future model rides the same rails instead of reinventing them.

FAQs

Does this train the model itself?

No — it submits and runs the model scripts. The model logic lives in the XGBoost, random forest, and time-series templates.

Custom jobs or AutoML?

This is the custom-training path, so you control the model code end to end.

Technical Architecture

The gap between 'model works locally' and 'model runs in production' is mostly submission plumbing — packaging, auth, config, and entry points. Getting this layer right once means every future model rides the same rails instead of reinventing them.

Tech Stack

Tool 1

Tool 4

Tool 4

Tool 3

Tool 2

Tool 4