Configure Amazon SageMaker Deployments

Deepgram on Amazon SageMaker supports runtime configuration through environment variables. When your SageMaker Endpoint starts, the container reads these variables and applies them to the appropriate configuration files (api.toml and engine.toml) before launching Deepgram services.

Use environment variables to tune settings for a specific SageMaker deployment, such as specifying the maximum number of streams for Flux or adjusting the step parameter for interim results (Nova-3).

Each environment variable targets either the API server (api.toml) or the speech engine (engine.toml) based on its prefix.

The suffix after the prefix (for example, 01, 02) is arbitrary and distinguishes multiple variables targeting the same file. Variables are applied in alphabetical order by name.

DEEPGRAM_ENGINE_<suffix>="<dotted.key.path>=<value>"
DEEPGRAM_API_<suffix>="<dotted.key.path>=<value>"

• Dotted key path: Maps to TOML section hierarchy. For example, chunking.streaming.step sets the step key inside [chunking.streaming].
• Value types: Integers, floats, booleans (true/false), and quoted strings.
• Multiple settings: Use separate environment variables with different suffixes.

DEEPGRAM_ENGINE_01=max_active_requests=120

DEEPGRAM_API_01=features.entity_detection=false

This is separate from max_active_requests and is specific to the Flux streaming transcription model.

DEEPGRAM_ENGINE_01=flux.max_streams=25

DEEPGRAM_ENGINE_01=chunking.streaming.step=0.5

DEEPGRAM_ENGINE_01=health.gpu_required=true

Use incrementing suffixes to apply multiple settings to the same file:

DEEPGRAM_ENGINE_01=chunking.streaming.step=0.5
DEEPGRAM_ENGINE_02=health.gpu_required=true
DEEPGRAM_API_01=features.listen_v2=true
DEEPGRAM_API_02=features.topic_detection=false

Environment variables are set at the Model level in SageMaker and passed to the container at runtime.

Create the SageMaker Model resource from a Model Package Amazon Resource Name (ARN). See Find the Model Package ARN for how to obtain it.

If you create the model directly from an ECR image, use --primary-container with Image=... instead.

import boto3
sagemaker = boto3.client("sagemaker")
sagemaker.create_model(
    ModelName="my-deepgram-model",
    ExecutionRoleArn="arn:aws:iam::123456789012:role/SageMakerRole",
    Containers=[
        {
            "ModelPackageName": "arn:aws:sagemaker:us-east-1:123456789012:model-package/my-model-package/1",
            "Environment": {
                "DEEPGRAM_ENGINE_01": "flux.max_streams=25",
                "DEEPGRAM_API_01": "features.listen_v2=true",
            },
        }
    ],
)

After your Endpoint reaches InService status, check Amazon CloudWatch Logs for the Endpoint to confirm variables were applied. Look for log lines similar to:

INFO Starting Deepgram SageMaker configuration...
INFO Applying to engine.toml: DEEPGRAM_ENGINE_01=flux.max_streams=25
INFO Successfully updated engine.toml
INFO Applying to api.toml: DEEPGRAM_API_01=features.listen_v2=true
INFO Successfully updated api.toml
INFO Configuration complete.

Deepgram supports a maximum of 8 environment variables each for the API server and engine. Suffixes range from 01 to 08.

Check CloudWatch Logs for warning messages indicating a variable was skipped due to a parse error.

• Quote string values within the value expression: driver_pool.standard.url="https://engine:8080/v2"
• Boolean values do not require quoting: features.listen_v2=true

AWS requires whitelisting support for environment variables on SageMaker product listings. If you receive the error:

An error occurred (ValidationException) when calling the CreateModel operation: Environment variable map cannot be specified when using a ModelPackage subscribed from AWS Marketplace.

Contact Deepgram Support for help enabling it for that product.

For a complete reference of available TOML configuration keys, refer to the api.toml and engine.toml files in the Deepgram self-hosted GitHub repository, or contact Deepgram Support.