Gemini Enterprise Agent Platform serverless training¶

Maturity: verified for the bounded submission and accepted one-step/resume evidence.

Gemini Enterprise Agent Platform (formerly Vertex AI) is the Google Cloud offering used for the repository’s serverless training submission path.

L4 preset¶

infra/runtime/submit_vertex_ai_training.sh is the committed provider adapter. It reads configs/launch/vertex_ai_l4.yaml, whose provider value is vertex_ai, uploads that exact file, and calls gcloud ai custom-jobs create to create a CustomJob. This command mutates cloud state and is for an authorized operator, not the local getting-started path.

Launch selection¶

Field	Value
provider/job type	Gemini Enterprise Agent Platform `CustomJob`
project/region	`rf-foundation-models-dev` / `us-central1`
machine	`g2-standard-4`
accelerator	one `NVIDIA_L4`
replicas	one
service account	`rf-fm-training-dev@rf-foundation-models-dev.iam.gserviceaccount.com`

Required submission inputs¶

The script requires DISPLAY_NAME, RUN_ID, ATTEMPT_ID, EXPERIMENT, IMAGE_URI, SOURCE_REVISION, LAUNCH_CONFIG_PATH, and LAUNCH_CONFIG_URI. IMAGE_URI must end in an immutable 64-hex SHA-256 digest; SOURCE_REVISION must be a full 40-hex Git commit.

For resume, set both RESUME_FROM and positive TARGET_MAX_STEPS. Omitting both selects a fresh attempt. Setting only one is rejected.

Submission behavior¶

Before submission, the script validates the exact launch fields, identifiers, image and source immutability, accelerator and replica counts, and resume pair. It uploads the launch file with --if-generation-match=0, computes its SHA-256, and passes source, image, launch, run, attempt, experiment, and mounted /gcs inputs to the application.

The script returns after the provider accepts or rejects job creation. It does not wait, monitor, retry, cancel, verify terminal artifacts, or clean up the job. Those actions remain operator responsibilities.

Accepted execution evidence¶

Issue #62 records accepted Gemini Enterprise Agent Platform L4 evidence for one fresh optimizer step and a separate attempt resumed to step two. That evidence validates the bounded path; it does not establish train/validation, model-quality, exact-replay, or scale claims.