Agent Platform Deploy

Agent Platform Model Garden Deploy Skill

This skill provides instructions for deploying Open Models from Agent Platform Model Garden to endpoints, and subsequently undeploying them to clean up resources.

1P Tuned Model Copy & Deployment

If you need to copy a 1P (First-Party) Tuned Model from a source project to a destination region or project and deploy it to a newly created endpoint, refer to the 1P Tuned Model Copy & Deployment Guide.

Safety & Confirmation Tiers (CRITICAL)

Before executing any commands on behalf of the user, you MUST adhere to the following safety tiers based on the action requested:

1. Tier R: Read-only (list, describe, list-deployment-config)
   • Rule: No confirmation needed. You may execute these commands immediately to gather information for the user.
2. Tier M: Mutating & Reversible (deploy, undeploy-model)
   • Rule: This requires explicit user confirmation. You MUST present a clear confirmation prompt to the user explaining the proposed command. You MUST wait for their explicit confirmation before executing. For undeploy-model, you MUST first verify that the endpoint and deployed model exist; if describe or list returns a 404 or empty result, you MUST halt and inform the user rather than attempting undeployment.
   • Same-turn restriction: Do not run the command in the same turn as presenting the confirmation prompt. End your turn after asking and wait for the user's reply; only execute after explicit approval. Printing a preview and then calling the tool before the user can answer does not count as obtaining confirmation.
3. Tier D: Destructive & Irreversible (delete)
   • Rule: This requires explicit typed confirmation. You MUST output a text message explaining the irreversible nature of endpoint or model deletion and asking the user to type "I confirm" or "Yes, delete it" before executing the deletion command.

1. Prerequisites

Before deploying, ensure you have the correct project and region set. The commands below use placeholder variables PROJECT_ID and LOCATION_ID.

Ensure you are authenticated:

gcloud auth login
gcloud auth application-default login
gcloud config set project $PROJECT_ID

2. Discovering Deployable Models

You can list models available in Model Garden and check if they can be self-deployed.

gcloud ai model-garden models list

To see what machine types and accelerators are supported for a specific model, pass a MODEL_ID you obtained from the models list output above. Substitute <PUBLISHER>/<FAMILY>@<VERSION-ID> below with the exact string from the catalog output — the placeholder is deliberately not a real model ID:

gcloud ai model-garden models list-deployment-config \
    --model="<PUBLISHER>/<FAMILY>@<VERSION-ID>"

[!NOTE] Some models, especially Hugging Face models, might require a Hugging Face Access Token for deployment.

[!TIP] Model Recommendation Instructions: Whenever you are about to name a specific model version in a response, do NOT recommend from memory. This applies in all of the following situations — not just direct deploy requests:

• The user asks to deploy a model without naming one.
• You are volunteering a next-step suggestion after a list, describe, or undeploy operation (e.g. "Would you like me to deploy <model> to this endpoint?").
• The user asks a general "what should I use?" / "what's a good model for X?" question.
• You are filling in a MODEL_ID value in an example command you are showing the user (as opposed to a placeholder like <PUBLISHER>/<FAMILY>@<VERSION-ID>).

New model versions ship frequently and older ones may be deprecated, so training-corpus knowledge of which models exist is unreliable. Follow this procedure:

1. Clarify the use case if it isn't already clear from context (task type, quality vs. latency vs. cost priorities, hardware/quota constraints, license constraints). Skip if the user has already given enough signal.
2. Query the live catalog with gcloud ai model-garden models list. Narrow with --filter when appropriate (e.g. --filter="name~gemma", --filter="name~llama", --filter="name~qwen", --filter="name~deepseek"). Never name a specific model version to the user until you have seen it in the catalog output for this project.
3. Pick the latest generally-available version in the family that fits the use case. When multiple size variants exist, pick the one that matches the user's hardware/cost tolerance. Prefer a newer major version over an older one unless it is marked preview/experimental and the user explicitly asked for a stable option.
4. Verify the exact model ID is deployable with gcloud ai model-garden models list-deployment-config --model="<publisher>/<family>@<version>" before naming it in your response.
5. Cite the model ID verbatim in your recommendation, exactly as it appears in the catalog. Do not paraphrase to a family label ("Gemma", "Llama").

The MODEL_ID values in the §3 examples below are intentionally non-substantive placeholders (<PUBLISHER>/<FAMILY>@<VERSION-ID>). Do NOT replace them with a remembered model name for a user-facing recommendation — always re-run steps 2-4 first, then cite the exact string from the catalog.

2.1 Region Availability Check for Publisher Endpoints (Gemini + LoRA base)

[!NOTE] Skip this section if the user is asking to deploy an open-weights model from Model Garden (Gemma, Llama, DeepSeek, Qwen, or any user-supplied weights) — i.e. anything served via gcloud ai model-garden models deploy onto a dedicated endpoint. These models have no per-region availability restriction; the Model Garden catalog is global. The real failure modes for an unusual region are (a) the requested accelerator/machine type isn't offered in that region, or (b) the project has no quota — both surface as a clean error at deploy time before any resources are provisioned (§3's cost-confirm gate catches them). Go straight to §3.

Apply this section only if the user is asking to serve a first-party managed Gemini model (google/gemini-*) or a fine-tuned Gemini LoRA adapter — both of which route through a publisher endpoint whose regional availability actually varies.

Before responding to any deploy request that names a specific region for a first-party managed model (google/gemini-*) or a fine-tuned Gemini LoRA adapter, you MUST verify the model is actually available in that region by making a live API call. Do not rely on Google Search, training-corpus knowledge, or publisher documentation for availability claims — regional availability changes frequently and grounded text can be stale or wrong.

Probe only the exact model and region the user asked about. Do not probe other models as a "control" — you cannot infer anything about model A's availability from model B's status, because a different model may itself be unavailable in the reference region for unrelated reasons.

For first-party publisher models (google/*), probe with a real :generateContent call using a minimal valid payload:

curl -sS -o /dev/null -w "%{http_code}\n" \
    -H "Authorization: Bearer $(gcloud auth print-access-token)" \
    -H "Content-Type: application/json" \
    "https://${LOCATION_ID}-aiplatform.googleapis.com/v1/projects/${PROJECT_ID}/locations/${LOCATION_ID}/publishers/google/${MODEL_ID}:generateContent" \
    -d "{\"contents\":{\"role\":\"user\",\"parts\":{\"text\":\"${PROBE_TEXT:-hi}\"}}}"

For fine-tuned Gemini LoRA models (deploying a user-tuned adapter on top of a base Gemini model), probe the base model in the target region using the same :generateContent call above with ${MODEL_ID} set to the base (e.g. gemini-2.5-flash if the adapter was tuned on gemini-2.5-flash). The LoRA adapter cannot serve in a region where its base model isn't available.

Interpret the probe result and act:

• 200 — model is available in that region. Proceed with the deploy.
• 404 — model is not available in that region. STOP. Tell the user plainly that the model isn't offered in that region and list the regions where it is available (from gcloud ai model-garden models list --filter="name~$MODEL_NAME" without --region). Do not silently switch regions. Do not proceed to write deploy code or SDK initialization for the unsupported region. Do not run additional "control" probes to double-check the 404 — the target-region probe is authoritative.
• Any other outcome (permission denied, quota, transient failure, etc.) — do not conclude the model is available or unavailable. Explain the underlying cause in plain language (e.g. "your account doesn't have access to this project's Vertex AI API — enable it in the console or switch projects") and the concrete next action.

3. Deploying a Model

[!WARNING] Deploying models, especially large ones, consumes significant compute resources and incurs costs.

1. You MUST compute an hourly $ estimate for the requested --machine-type before proposing a deploy. Try, in order, and fall through on any failure (tool unavailable, tool returns status != "success", script exits non-zero, script rejects the machine type):

   a. If the estimate_cost tool is available AND returns status == "success", use its result -- it returns live SKU-resolved pricing (machine + accelerator + total) from CostEstimationService rather than a hardcoded snapshot. On any other status (including error), fall through to (b).

   b. Otherwise, run scripts/calculate_cost.py. The accelerator type and count are fixed per machine type in Model Garden and derived automatically. Example:

   python3 scripts/calculate_cost.py \
       --machine-type=g2-standard-48
   

   If the script exits non-zero (unknown --machine-type — a routine state for machines in the Model Garden catalog but not yet in the price snapshot, e.g. A4/B200 today), fall through to (c). Do NOT invent a number.

   c. Fall back to Agent Platform prediction pricing if the tool is unavailable AND the script does not know the requested machine type. Read the accelerator + hourly rate directly off that page and cite the URL in the estimate you present to the user.

2. You MUST present this cost estimation to the user and warn them that this is the list price, which may differ from their actual bill due to potential discounts, reservations, or non-us-central1 regions.

3. You MUST ALWAYS request explicit confirmation from the user agreeing to the estimated cost before executing any deploy command.

To deploy a model, use the deploy command. It is highly recommended to use the --asynchronous flag for long-running deployments, and then poll the status if necessary.

Example: Deploying an open-weights model from Model Garden

Here is a typical bash script to deploy a model. You can run this block directly.

#!/bin/bash
# Example script to deploy an open-weights model from Model Garden.
#
# NOTE: MODEL_ID below is a PLACEHOLDER, not a real model ID. Substitute it
# with a value from a live `gcloud ai model-garden models list` (see §2)
# before running this script, and do NOT quote the placeholder back to the
# user as a recommended model.

PROJECT_ID=$(gcloud config get-value project)
LOCATION_ID="us-central1" # Recommended default region
MODEL_ID="<PUBLISHER>/<FAMILY>@<VERSION-ID>" # PLACEHOLDER — replace with the exact ID from `gcloud ai model-garden models list`

echo "Deploying model $MODEL_ID to project $PROJECT_ID in $LOCATION_ID..."

# Model Garden can automatically select the required hardware based on the list-deployment-config if hardware params are omitted.
# Below is a comprehensive command with all supported parameters:
gcloud ai model-garden models deploy \
    --project=$PROJECT_ID \
    --region=$LOCATION_ID \
    --model=$MODEL_ID \
    --machine-type="g2-standard-48" \
    --accelerator-type="NVIDIA_L4" \
    --accelerator-count=4 \
    --endpoint-display-name="my-open-model-deployment" \
    --hugging-face-access-token="YOUR_HF_TOKEN" \
    --reservation-affinity="reservation-affinity-type=specific-reservation,key=compute.googleapis.com/reservation-name,values=my-reservation" \
    --asynchronous

echo "Deployment initiated asynchronously."

Example: Deploying Custom Weights

To deploy a model using custom weights, you can use the exact same deploy command. Instead of providing the model garden model ID, provide the Google Cloud Storage (GCS) URI to your custom weights folder in the --model flag.

#!/bin/bash
# Example script to deploy a model with custom weights from a GCS bucket

PROJECT_ID=$(gcloud config get-value project)
LOCATION_ID="us-central1"
# Replace with the gs:// URI pointing to your custom weights
MODEL_GCS_URI="gs://your-bucket-name/path/to/custom-weights"

echo "Deploying custom model from $MODEL_GCS_URI to project $PROJECT_ID in $LOCATION_ID..."

gcloud ai model-garden models deploy \
    --project=$PROJECT_ID \
    --region=$LOCATION_ID \
    --model=$MODEL_GCS_URI \
    --machine-type="g2-standard-12" \
    --accelerator-type="NVIDIA_L4" \
    --endpoint-display-name="my-custom-model" \
    --asynchronous

echo "Deployment initiated asynchronously."

4. Checking Deployment Status

When you deploy a model asynchronously using the --asynchronous flag, the deploy command will return an operation ID. You can use this ID to check the ongoing status of the deployment.

gcloud ai operations describe YOUR_OPERATION_ID \
    --region=$LOCATION_ID

[!NOTE] As an agent, you can also offer to check the status of a deployment for the user if they provide an operation ID or if they just initiated the deployment with you.

Alternatively, you can list your endpoints to see if it shows up and check the Cloud Console under the "Online prediction" tab.

gcloud ai endpoints list \
    --region=$LOCATION_ID

Note: Large models (roughly 20B+ parameters) may take 15-20 minutes to fully deploy and start serving.

Verifying Deployment

If the model is successfully deployed, verify by making a prediction call to test. Because Model Garden models are often deployed to Dedicated Endpoints, you shouldn't use gcloud ai endpoints predict. Instead, you must fetch the endpoint's dedicated DNS name and send a curl request.

[!TIP] Ask the user to try using their own prompt to see the results. Otherwise use the default.

Use the following script:

#!/bin/bash
PROJECT_ID=$(gcloud config get-value project)
LOCATION_ID="us-central1"
ENDPOINT_ID="YOUR_ENDPOINT_ID"
PROMPT=${1:-"Explain quantum computing in simple terms."}

echo "Fetching dedicated Endpoint DNS..."
ENDPOINT_URL=$(gcloud ai endpoints describe $ENDPOINT_ID --project=$PROJECT_ID --region=$LOCATION_ID --format="value(dedicatedEndpointDns)")

if [ -z "$ENDPOINT_URL" ]; then
    echo "Error: Could not retrieve a dedicated endpoint URL. Verify your ENDPOINT_ID."
    exit 1
fi

echo "Sending prediction request to $ENDPOINT_URL..."
curl -X POST \
  -H "Authorization: Bearer $(gcloud auth print-access-token)" \
  -H "Content-Type: application/json" \
  "https://${ENDPOINT_URL}/v1beta1/projects/${PROJECT_ID}/locations/${LOCATION_ID}/endpoints/${ENDPOINT_ID}/chat/completions" \
  -d '{
    "model": "'"$ENDPOINT_ID"'",
    "messages": [
      {
        "role": "user",
        "content": "'"$PROMPT"'"
      }
    ]
  }'

5. Undeploying and Cleaning Up

To stop incurring charges, you must undeploy the model from the endpoint. This is a multi-step process if you don't already have the exact endpoint and deployed model IDs.

Example: Finding and Undeploying a Model

Here is a bash script demonstrating how to find the IDs and undeploy the model.

#!/bin/bash
# Example script to undeploy a model

PROJECT_ID=$(gcloud config get-value project)
LOCATION_ID="us-central1"
# The model ID used during deployment (without the provider prefix sometimes, or exactly as listed in describe)
# It's usually easier to find the specific ID via `gcloud ai models list`
# For this example, let's assume we know the exact Endpoint ID and Deployed Model ID.

# 1. Find the Endpoint ID
echo "Listing endpoints in $LOCATION_ID:"
gcloud ai endpoints list --project=$PROJECT_ID --region=$LOCATION_ID

# (Assuming you extracted ENDPOINT_ID from the above output)
# ENDPOINT_ID="your_endpoint_id"

# 2. Find the Deployed Model ID
echo "Listing models in $LOCATION_ID to find model description:"
gcloud ai models list --project=$PROJECT_ID --region=$LOCATION_ID

# (Assuming you found the specific MODEL_ID)
# MODEL_ID="your_model_id"
# gcloud ai models describe $MODEL_ID --project=$PROJECT_ID --region=$LOCATION_ID
# (Extract the deployedModelId from the output)
# DEPLOYED_MODEL_ID="your_deployed_model_id"

# 3. Undeploy
echo "Undeploying model $DEPLOYED_MODEL_ID from endpoint $ENDPOINT_ID..."
gcloud ai endpoints undeploy-model $ENDPOINT_ID \
    --project=$PROJECT_ID \
    --region=$LOCATION_ID \
    --deployed-model-id=$DEPLOYED_MODEL_ID

echo "Model undeployed."

# 4. Delete Endpoint
echo "Deleting endpoint $ENDPOINT_ID..."
gcloud ai endpoints delete $ENDPOINT_ID \
    --project=$PROJECT_ID \
    --region=$LOCATION_ID \
    --quiet
echo "Endpoint deleted."

# 5. Delete Model
echo "Deleting model $MODEL_ID..."
gcloud ai models delete $MODEL_ID \
    --project=$PROJECT_ID \
    --region=$LOCATION_ID \
    --quiet
echo "Model deleted."

[!WARNING] Failing to undeploy a model will result in continuous charges for the allocated compute resources, even if you are not sending prediction requests. Always clean up after testing.

6. Troubleshooting

Deployment Failure: Quota or Resource Exhausted

If your deployment fails (or stays in an error state) due to QUOTA_EXCEEDED or RESOURCE_EXHAUSTED errors, the specific hardware requested (e.g., NVIDIA_L4 or g2-standard-24) is either not available in your chosen region or exceeds your project's quota limits.

Solution: Look closely at the error message returned. It will often recommend an alternative region or machine type that currently has availability. Ask the user for confirmation to retry the deployment using the suggested --region or --machine-type parameters.

[!WARNING] If the alternative suggestions involve changing the machine type or accelerator, you MUST recalculate the estimated cost by re-running scripts/calculate_cost.py with the new params (see §3), warn the user about list prices versus actual billing, and get their explicit confirmation for the new cost before retrying the deployment.