Get batch predictions from a self-deployed Model Garden model

Some of the models that are available in Model Garden can be self-deployed in your own Google Cloud project and used to provide batch predictions. Batch predictions let you efficiently use a model to process multiple text-only prompts that aren't latency sensitive.

Prepare input

Before you begin, prepare your inputs in a BigQuery table or as a JSONL file in Cloud Storage. The input for both sources must follow the OpenAI API schema JSON format, as shown in the following example:

{"body": {"messages": [{"role": "user", "content": "Give me a recipe for banana bread"}], "max_tokens": 1000}}

BigQuery

Your BigQuery input table must adhere to the following schema:

Column name Description
custom_id An ID for each request to match the input with the output.
method The request method.
url The request endpoint.
body(JSON) Your input prompt.
  • Your input table can have other columns, which are ignored by the batch job and passed directly to the output table.
  • Batch prediction jobs reserve two column names for the batch prediction output: response(JSON) and id. Don't use these columns in the input table.
  • The method and url columns are dropped and not included in the output table.

Cloud Storage

For Cloud Storage, the input file must be a JSONL file that is located in a Cloud Storage bucket.

Get the required resources for a model

Choose a model and query its resource requirements. The required resources appear in the response, in the dedicatedResources field, which you specify in the configuration of your batch prediction job.

REST

Before using any of the request data, make the following replacements:

  • PUBLISHER: The model publisher, for example, meta, google, mistral-ai, or deepseek-ai.
  • PUBLISHER_MODEL_ID: The publisher's model ID for the model, for example, llama3_1.
  • VERSION_ID: The publisher's version ID for the model, for example, llama-3.1-8b-instruct.

HTTP method and URL: 

GET "https://us-central1-aiplatform.googleapis.com/ui/publishers/PUBLISHER/models/PUBLISHER_MODEL_ID@VERSION_ID" | jq '.supportedActions.multiDeployVertex'

To send your request, choose one of these options:

curl

Execute the following command: 

curl -X GET \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "x-goog-user-project: PROJECT_ID" \
     ""https://us-central1-aiplatform.googleapis.com/ui/publishers/PUBLISHER/models/PUBLISHER_MODEL_ID@VERSION_ID" | jq '.supportedActions.multiDeployVertex'"

PowerShell

Execute the following command: 

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred"; "x-goog-user-project" = "PROJECT_ID" }

Invoke-WebRequest `
    -Method GET `
    -Headers $headers `
    -Uri ""https://us-central1-aiplatform.googleapis.com/ui/publishers/PUBLISHER/models/PUBLISHER_MODEL_ID@VERSION_ID" | jq '.supportedActions.multiDeployVertex'" | Select-Object -Expand Content

You should receive a successful status code (2xx) and an empty response.

Request a batch prediction

Make a batch prediction against a self-deployed Model Garden model by using input from BigQuery or Cloud Storage. You can independently choose to output predictions to either a BigQuery table or a JSONL file in a Cloud Storage bucket.

BigQuery

Specify your BigQuery input table, model, and output location. The batch prediction job and your table must be in the same region.

REST

Before using any of the request data, make the following replacements:

  • LOCATION: A region that supports Model Garden self-deployed models.
  • PROJECT_ID: Your project ID.
  • MODEL: The name of the model to tune, for example, llama-3.1-8b-instruct.
  • PUBLISHER: The model publisher, for example, meta, google, mistral-ai, or deepseek-ai.
  • INPUT_URI: The BigQuery table where your batch prediction input is located such as myproject.mydataset.input_table.
  • OUTPUT_FORMAT: To output to a BigQuery table, specify bigquery. To output to a Cloud Storage bucket, specify jsonl.
  • DESTINATION: For BigQuery, specify bigqueryDestination. For Cloud Storage, specify gcsDestination.
  • OUTPUT_URI_FIELD_NAME: For BigQuery, specify outputUri. For Cloud Storage, specify outputUriPrefix.
  • OUTPUT_URI: For BigQuery, specify the table location such as myproject.mydataset.output_result. For Cloud Storage, specify the bucket and folder location such as gs://mybucket/path/to/outputfile.
  • MACHINE_TYPE: Defines the set of resources to deploy for your model, for example, g2-standard-4.
  • ACC_TYPE: Specifies accelerators to add to your batch prediction job to help improve performance when working with intensive workloads, for example, NVIDIA_L4.
  • ACC_COUNT: The number of accelerators to use in your batch prediction job.

HTTP method and URL: 

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs

Request JSON body: 

'{
  "displayName": "JOB_NAME",
  "model": "publishers/PUBLISHER/models/MODEL",
  "inputConfig": {
    "instancesFormat":"bigquery",
    "bigquerySource":{
      "inputUri" : "INPUT_URI"
    }
  },
  "outputConfig": {
    "predictionsFormat":"OUTPUT_FORMAT",
    "DESTINATION":{
      "OUTPUT_URI_FIELD_NAME": "OUTPUT_URI"
    }
  },
  "dedicated_resources": {
    "machine_spec": {
      "machine_type": "MACHINE_TYPE",
      "accelerator_type": "ACC_TYPE",
      "accelerator_count": ACC_COUNT,
    },
    "starting_replica_count": 1,
  },
}'

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command: 

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs"

PowerShell

Save the request body in a file named request.json, and execute the following command: 

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs" | Select-Object -Expand Content

You should receive a JSON response similar to the following.

Cloud Storage

Specify your JSONL file's Cloud Storage location, model, and output location.

REST

Before using any of the request data, make the following replacements:

  • LOCATION: A region that supports Model Garden self-deployed models.
  • PROJECT_ID: Your project ID.
  • MODEL: The name of the model to tune, for example, llama-3.1-8b-instruct.
  • PUBLISHER: The model publisher, for example, meta, google, mistral-ai, or deepseek-ai.
  • INPUT_URI: The Cloud Storage location of your JSONL batch prediction input such as gs://bucketname/path/to/jsonl.
  • OUTPUT_FORMAT: To output to a BigQuery table, specify bigquery. To output to a Cloud Storage bucket, specify jsonl.
  • DESTINATION: For BigQuery, specify bigqueryDestination. For Cloud Storage, specify gcsDestination.
  • OUTPUT_URI_FIELD_NAME: For BigQuery, specify outputUri. For Cloud Storage, specify outputUriPrefix.
  • OUTPUT_URI: For BigQuery, specify the table location such as myproject.mydataset.output_result. For Cloud Storage, specify the bucket and folder location such as gs://mybucket/path/to/outputfile.
  • MACHINE_TYPE: Defines the set of resources to deploy for your model, for example, g2-standard-4.
  • ACC_TYPE: Specifies accelerators to add to your batch prediction job to help improve performance when working with intensive workloads, for example, NVIDIA_L4.
  • ACC_COUNT: The number of accelerators to use in your batch prediction job.

HTTP method and URL: 

POST https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs

Request JSON body: 

'{
  "displayName": "JOB_NAME",
  "model": "publishers/PUBLISHER/models/MODEL",
  "inputConfig": {
    "instancesFormat":"jsonl",
    "gcsDestination":{
      "uris" : "INPUT_URI"
    }
  },
  "outputConfig": {
    "predictionsFormat":"OUTPUT_FORMAT",
    "DESTINATION":{
      "OUTPUT_URI_FIELD_NAME": "OUTPUT_URI"
    }
  },
  "dedicated_resources": {
    "machine_spec": {
        "machine_type": "MACHINE_TYPE",
        "accelerator_type": "ACC_TYPE",
        "accelerator_count": ACC_COUNT,
    },
    "starting_replica_count": 1,
  },
}'

To send your request, choose one of these options:

curl

Save the request body in a file named request.json, and execute the following command: 

curl -X POST \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     -H "Content-Type: application/json; charset=utf-8" \
     -d @request.json \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs"

PowerShell

Save the request body in a file named request.json, and execute the following command: 

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method POST `
    -Headers $headers `
    -ContentType: "application/json; charset=utf-8" `
    -InFile request.json `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs" | Select-Object -Expand Content

You should receive a JSON response similar to the following.

Get the status of a batch prediction job

Get the state of your batch prediction job to check whether it has completed successfully. The job length depends on the number of input items that you submitted.

REST

Before using any of the request data, make the following replacements:

  • PROJECT_ID: Your project ID.
  • LOCATION: The region where your batch job is located.
  • JOB_ID: The batch job ID that was returned when you created the job.

HTTP method and URL: 

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs/JOB_ID

To send your request, choose one of these options:

curl

Execute the following command: 

curl -X GET \
     -H "Authorization: Bearer $(gcloud auth print-access-token)" \
     "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs/JOB_ID"

PowerShell

Execute the following command: 

$cred = gcloud auth print-access-token
$headers = @{ "Authorization" = "Bearer $cred" }

Invoke-WebRequest `
    -Method GET `
    -Headers $headers `
    -Uri "https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/batchPredictionJobs/JOB_ID" | Select-Object -Expand Content

You should receive a JSON response similar to the following.

Retrieve output

When a batch prediction job completes, retrieve the output from the location that you specified:

  • For BigQuery, the output is in the response(JSON) column of your destination BigQuery table.
  • For Cloud Storage, the output is saved as a JSONL file in the output Cloud Storage location.

Supported models

Vertex AI supports batch predictions for the following self-deployed models:

  • Llama
    • publishers/meta/models/llama3_1@llama-3.1-8b-instruct
    • publishers/meta/models/llama3_1@llama-3.1-70b-instruct
    • publishers/meta/models/llama3_1@llama-3.1-405b-instruct-fp8
    • publishers/meta/models/llama3-2@llama-3.2-1b-instruct
    • publishers/meta/models/llama3-2@llama-3.2-3b-instruct
    • publishers/meta/models/llama3-2@llama-3.2-90b-vision-instruct
  • Gemma
    • publishers/google/models/gemma@gemma-1.1-2b-it
    • publishers/google/models/gemma@gemma-7b-it
    • publishers/google/models/gemma@gemma-1.1-7b-it
    • publishers/google/models/gemma@gemma-2b-it
    • publishers/google/models/gemma2@gemma-2-2b-it
    • publishers/google/models/gemma2@gemma-2-9b-it
    • publishers/google/models/gemma2@gemma-2-27b-it
  • Mistral
    • publishers/mistral-ai/models/mistral@mistral-7b-instruct-v0.2
    • publishers/mistral-ai/models/mistral@mistral-7b-instruct-v0.3
    • publishers/mistral-ai/models/mistral@mistral-7b-instruct-v0.1
    • publishers/mistral-ai/models/mistral@mistral-nemo-instruct-2407
  • Deepseek
    • publishers/deepseek-ai/models/deepseek-r1@deepseek-r1-distill-llama-8b