Llama models

Llama models on Vertex AI offer fully managed and serverless models as APIs. To use a Llama model on Vertex AI, send a request directly to the Vertex AI API endpoint. Because Llama models use a managed API, there's no need to provision or manage infrastructure.

You can stream your responses to reduce the end-user latency perception. A streamed response uses server-sent events (SSE) to incrementally stream the response.

Available Llama models

The following Llama models are available from Meta to use in Vertex AI. To access a Llama model, go to its Model Garden model card.

Llama 3.2

Preview

This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms. Pre-GA features are available "as is" and might have limited support. For more information, see the launch stage descriptions.

Llama 3.2 lets developers to build and deploy the latest generative AI models and applications that use the latest Llama's capabilities, such as image reasoning. Llama 3.2 is also designed to be more accessible for on-device applications.

Go to the Llama 3.2 model card

There are no charges during the Preview period. If you require a production-ready service, use the self-hosted Llama models.

Considerations

When using llama-3.2-90b-vision-instruct-maas, there are no restriction when you send text-only prompts. However, if you include an image in your prompt, the image must be at beginning of your prompt, and you can include only one image. You cannot, for example, include some text and then an image.

Llama 3.1

To see an example of getting started with Llama on Vertex AI, run the "Getting started with Llama 3.1" Jupyter notebook in one of the following environments:

Open in Colab | Open in Colab Enterprise | Open in Vertex AI Workbench user-managed notebooks | View on GitHub

Llama 3.1 is an auto-regressive language model that uses an optimized transformer architecture. The tuned versions use supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF) to align with human preferences for helpfulness and safety.

Llama 3.1 405B is Generally Available. You are charged as you use the model (pay as you go). For pay-as-you-go pricing, see Llama model pricing on the Vertex AI pricing page.

The other Llama 3.1 models are in Preview. There are no charges for the Preview models. If you require a production-ready service, use the self-hosted Llama models.

Go to the Llama 3.1 model card

Use Llama models

When you send requests to use Llama's models, use the following model names:

• For Llama 3.2 90B (Preview), use llama-3.2-90b-vision-instruct-maas.
• For Llama 3.1 405B (GA), use llama3.1-405b-instruct-maas.
• For Llama 3.1 70B (Preview), use llama3.1-70b-instruct-maas.
• For Llama 3.1 8B (Preview), use llama3.1-8b-instruct-maas.

We recommend that you use the model versions that include a suffix that starts with an @ symbol because of the possible differences between model versions. If you don't specify a model version, the latest version is always used, which can inadvertently affect your workflows when a model version changes.

To see an example of evaluating Llama 3.1 models by using Automatic side-by-side (AutoSxS) evaluation, run the "Evaluate Llama 3.1 models using Vertex AI AutoSxS" Jupyter notebook in one of the following environments:

Open in Colab | Open in Colab Enterprise | Open in Vertex AI Workbench user-managed notebooks | View on GitHub

Before you begin

To use Llama models with Vertex AI, you must perform the following steps. The Vertex AI API (aiplatform.googleapis.com) must be enabled to use Vertex AI. If you already have an existing project with the Vertex AI API enabled, you can use that project instead of creating a new project.

Make sure you have the required permissions to enable and use partner models. For more information, see Grant the required permissions.

1. Sign in to your Google Cloud account. If you're new to Google Cloud, create an account to evaluate how our products perform in real-world scenarios. New customers also get $300 in free credits to run, test, and deploy workloads.

2. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

   Go to project selector

3. Make sure that billing is enabled for your Google Cloud project.

4. Enable the Vertex AI API.

   Enable the API

5. In the Google Cloud console, on the project selector page, select or create a Google Cloud project.

   Go to project selector

6. Make sure that billing is enabled for your Google Cloud project.

7. Enable the Vertex AI API.

   Enable the API

8. Go to one of the following Model Garden model cards, then click enable:
   • Go to the Llama 3.2 model card
   • Go to the Llama 3.1 model card

Make a streaming call to a Llama model

The following sample makes a streaming call to a Llama model.

REST

After you set up your environment, you can use REST to test a text prompt. The following sample sends a request to the publisher model endpoint.

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

• LOCATION: A region that supports Llama models.
• MODEL: The model name you want to use.
• ROLE: The role associated with a message. You can specify a user or an assistant. The first message must use the user role. The models operate with alternating user and assistant turns. If the final message uses the assistant role, then the response content continues immediately from the content in that message. You can use this to constrain part of the model's response.
• CONTENT: The content, such as text, of the user or assistant message.
• MAX_OUTPUT_TOKENS: Maximum number of tokens that can be generated in the response. A token is approximately four characters. 100 tokens correspond to roughly 60-80 words.

  Specify a lower value for shorter responses and a higher value for potentially longer responses.

• STREAM: A boolean that specifies whether the response is streamed or not. Stream your response to reduce the end-use latency perception. Set to true to stream the response and false to return the response all at once.
• ENABLE_LLAMA_GUARD: A boolean that specifies whether to enable Llama Guard on your inputs and outputs. By default, Llama Guard is enabled and flags responses if it determines they are unsafe.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions

Request JSON body:

{
  "model": "meta/MODEL",
  "messages": [
    {
      "role": "ROLE",
      "content": "CONTENT"
    }
  ],
  "max_tokens": MAX_OUTPUT_TOKENS,
  "stream": true,
  "extra_body": {
    "google": {
      "model_safety_settings": {
        "enabled": ENABLE_LLAMA_GUARD,
        "llama_guard_settings": {}
      }
    }
  }
}

To send your request, choose one of these options:

curl

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login , or by using Cloud Shell, which automatically logs you into the gcloud CLI . You can check the currently active account by running gcloud auth list.

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/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions"

PowerShell

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login . You can check the currently active account by running gcloud auth list.

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/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions" | Select-Object -Expand Content

You should receive a JSON response similar to the following.

Response

data: {
  "choices": [
    {
      "delta": {
        "content": "CONTENT",
        "role": "assistant",
        "refusal: "REFUSAL_REASON" #If using Llama Guard and response was flagged by Llama Guard
      },
      "index": 0
    }
  ],
  "model": "meta/MODEL_NAME",
  "object": "chat.completion.chunk"
}

data: {
  "choices": [
    {
      "delta": {
        "content": "CONTENT",
        "role": "assistant"
      },
      "finish_reason": "stop",
      "index": 0
    }
  ],
  "model": "meta/MODEL_NAME",
  "object": "chat.completion.chunk",
  "usage": {
    "completion_tokens": 131,
    "prompt_tokens": 14,
    "total_tokens": 145
  }
}

Make a unary call to a Llama model

The following sample makes a unary call to a Llama model.

REST

After you set up your environment, you can use REST to test a text prompt. The following sample sends a request to the publisher model endpoint.

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

• LOCATION: A region that supports Llama models.
• MODEL: The model name you want to use.
• ROLE: The role associated with a message. You can specify a user or an assistant. The first message must use the user role. The models operate with alternating user and assistant turns. If the final message uses the assistant role, then the response content continues immediately from the content in that message. You can use this to constrain part of the model's response.
• CONTENT: The content, such as text, of the user or assistant message.
• MAX_OUTPUT_TOKENS: Maximum number of tokens that can be generated in the response. A token is approximately four characters. 100 tokens correspond to roughly 60-80 words.

  Specify a lower value for shorter responses and a higher value for potentially longer responses.

• STREAM: A boolean that specifies whether the response is streamed or not. Stream your response to reduce the end-use latency perception. Set to true to stream the response and false to return the response all at once.
• ENABLE_LLAMA_GUARD: A boolean that specifies whether to enable Llama Guard on your inputs and outputs. By default, Llama Guard is enabled and flags responses if it determines they are unsafe.

HTTP method and URL:

POST https://LOCATION-aiplatform.googleapis.com/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions

Request JSON body:

{
  "model": "meta/MODEL",
  "messages": [
    {
      "role": "ROLE",
      "content": "CONTENT"
    }
  ],
  "max_tokens": MAX_OUTPUT_TOKENS,
  "stream": false,
  "extra_body": {
    "google": {
      "model_safety_settings": {
        "enabled": ENABLE_LLAMA_GUARD,
        "llama_guard_settings": {}
      }
    }
  }
}

To send your request, choose one of these options:

curl

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login , or by using Cloud Shell, which automatically logs you into the gcloud CLI . You can check the currently active account by running gcloud auth list.

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/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions"

PowerShell

Note: The following command assumes that you have logged in to the gcloud CLI with your user account by running gcloud init or gcloud auth login . You can check the currently active account by running gcloud auth list.

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/v1beta1/projects/PROJECT_ID/locations/LOCATION/endpoints/openapi/chat/completions" | Select-Object -Expand Content

You should receive a JSON response similar to the following.

Response

{
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "message": {
        "content": "CONTENT",
        "role": "assistant",
        "refusal: "REFUSAL_REASON" #If using Llama Guard and response was flagged by Llama Guard
      }
    }
  ],
  "model": "meta/llama3-405b-instruct-maas",
  "object": "chat.completion",
  "usage": {
    "completion_tokens": 367,
    "prompt_tokens": 14,
    "total_tokens": 381
  }
}

Flagged responses

By default, Llama Guard is enabled on all predictions that you make with Llama 3.1 models. Llama Guard helps safeguard responses by checking inputs and outputs. If Llama Guard determines they are unsafe, it flags the response.

If you want to disable Llama Guard, modify the model safety setting. For more information, see the model_safety_settings field in the streaming or unary example.

Note: Llama Guard isn't available for Llama 3.2.

Llama model region availability and quotas

For Llama models, a quota applies for each region where the model is available. The quota is specified in queries per minute (QPM).

The supported regions, default quotas, and maximum context length for each Llama model is listed in the following tables:

Llama 3.2 90B (Preview)

Region	Quota system	Supported context length
us-central1	30 QPM	128,000 tokens

Llama 3.1 405B (GA)

Region	Quota system	Supported context length
us-central1	60 QPM	128,000 tokens

Llama 3.1 70B (Preview)

Region	Quota system	Supported context length
us-central1	60 QPM	128,000 tokens

Llama 3.1 8B (Preview)

Region	Quota system	Supported context length
us-central1	60 QPM	128,000 tokens

If you want to increase any of your quotas for Generative AI on Vertex AI, you can use the Google Cloud console to request a quota increase. To learn more about quotas, see Work with quotas.