Merge branch 'master' into feat/s3-prefix-model-data-for-jumpstart-model

Author: evakravi

src/sagemaker/environment_variables.py

@@ -20,6 +20,7 @@
 from sagemaker.jumpstart import utils as jumpstart_utils
 from sagemaker.jumpstart import artifacts
 from sagemaker.jumpstart.constants import DEFAULT_JUMPSTART_SAGEMAKER_SESSION
+from sagemaker.jumpstart.enums import JumpStartScriptScope
 from sagemaker.session import Session
 
 logger = logging.getLogger(__name__)
@@ -33,6 +34,8 @@ def retrieve_default(
     tolerate_deprecated_model: bool = False,
     include_aws_sdk_env_vars: bool = True,
     sagemaker_session: Session = DEFAULT_JUMPSTART_SAGEMAKER_SESSION,
+    instance_type: Optional[str] = None,
+    script: JumpStartScriptScope = JumpStartScriptScope.INFERENCE,
 ) -> Dict[str, str]:
     """Retrieves the default container environment variables for the model matching the arguments.
 
@@ -58,6 +61,10 @@ def retrieve_default(
             object, used for SageMaker interactions. If not
             specified, one is created using the default AWS configuration
             chain. (Default: sagemaker.jumpstart.constants.DEFAULT_JUMPSTART_SAGEMAKER_SESSION).
+        instance_type (str): An instance type to optionally supply in order to get environment
+            variables specific for the instance type.
+        script (JumpStartScriptScope): The JumpStart script for which to retrieve environment
+            variables.
     Returns:
         dict: The variables to use for the model.
 
@@ -78,4 +85,6 @@ def retrieve_default(
         tolerate_deprecated_model,
         include_aws_sdk_env_vars,
         sagemaker_session=sagemaker_session,
+        instance_type=instance_type,
+        script=script,
     )

src/sagemaker/image_uris.py

@@ -270,20 +270,6 @@ def retrieve(
     return ECR_URI_TEMPLATE.format(registry=registry, hostname=hostname, repository=repo)
 
 
-def _get_instance_type_family(instance_type):
-    """Return the family of the instance type.
-
-    Regex matches either "ml.<family>.<size>" or "ml_<family>. If input is None
-    or there is no match, return an empty string.
-    """
-    instance_type_family = ""
-    if isinstance(instance_type, str):
-        match = re.match(r"^ml[\._]([a-z\d]+)\.?\w*$", instance_type)
-        if match is not None:
-            instance_type_family = match[1]
-    return instance_type_family
-
-
 def _get_image_tag(
     container_version,
     distribution,
@@ -297,7 +283,7 @@ def _get_image_tag(
     version,
 ):
     """Return image tag based on framework, container, and compute configuration(s)."""
-    instance_type_family = _get_instance_type_family(instance_type)
+    instance_type_family = utils.get_instance_type_family(instance_type)
     if framework in (XGBOOST_FRAMEWORK, SKLEARN_FRAMEWORK):
         if instance_type_family and final_image_scope == INFERENCE_GRAVITON:
             _validate_arg(
@@ -385,7 +371,7 @@ def _config_for_framework_and_scope(framework, image_scope, accelerator_type=Non
 
 def _validate_instance_deprecation(framework, instance_type, version):
     """Check if instance type is deprecated for a certain framework with a certain version"""
-    if _get_instance_type_family(instance_type) == "p2":
+    if utils.get_instance_type_family(instance_type) == "p2":
         if (framework == "pytorch" and Version(version) >= Version("1.13")) or (
             framework == "tensorflow" and Version(version) >= Version("2.12")
         ):
@@ -409,7 +395,7 @@ def _validate_for_suppported_frameworks_and_instance_type(framework, instance_ty
     # Validate for Graviton allowed frameowrks
     if (
         instance_type is not None
-        and _get_instance_type_family(instance_type) in GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY
+        and utils.get_instance_type_family(instance_type) in GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY
         and framework not in GRAVITON_ALLOWED_FRAMEWORKS
     ):
         _validate_framework(framework, GRAVITON_ALLOWED_FRAMEWORKS, "framework", "Graviton")
@@ -426,7 +412,7 @@ def _get_final_image_scope(framework, instance_type, image_scope):
     """Return final image scope based on provided framework and instance type."""
     if (
         framework in GRAVITON_ALLOWED_FRAMEWORKS
-        and _get_instance_type_family(instance_type) in GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY
+        and utils.get_instance_type_family(instance_type) in GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY
     ):
         return INFERENCE_GRAVITON
     if image_scope is None and framework in (XGBOOST_FRAMEWORK, SKLEARN_FRAMEWORK):
@@ -441,7 +427,7 @@ def _get_final_image_scope(framework, instance_type, image_scope):
 def _get_inference_tool(inference_tool, instance_type):
     """Extract the inference tool name from instance type."""
     if not inference_tool:
-        instance_type_family = _get_instance_type_family(instance_type)
+        instance_type_family = utils.get_instance_type_family(instance_type)
         if instance_type_family.startswith("inf") or instance_type_family.startswith("trn"):
             return "neuron"
     return inference_tool
@@ -529,7 +515,7 @@ def _processor(instance_type, available_processors, serverless_inference_config=
         processor = "neuron"
     else:
         # looks for either "ml.<family>.<size>" or "ml_<family>"
-        family = _get_instance_type_family(instance_type)
+        family = utils.get_instance_type_family(instance_type)
         if family:
             # For some frameworks, we have optimized images for specific families, e.g c5 or p3.
             # In those cases, we use the family name in the image tag. In other cases, we use
@@ -559,7 +545,7 @@ def _should_auto_select_container_version(instance_type, distribution):
     p4d = False
     if instance_type:
         # looks for either "ml.<family>.<size>" or "ml_<family>"
-        family = _get_instance_type_family(instance_type)
+        family = utils.get_instance_type_family(instance_type)
         if family:
             p4d = family == "p4d"
 

src/sagemaker/jumpstart/artifacts/environment_variables.py

@@ -34,6 +34,8 @@ def _retrieve_default_environment_variables(
     tolerate_deprecated_model: bool = False,
     include_aws_sdk_env_vars: bool = True,
     sagemaker_session: Session = DEFAULT_JUMPSTART_SAGEMAKER_SESSION,
+    instance_type: Optional[str] = None,
+    script: JumpStartScriptScope = JumpStartScriptScope.INFERENCE,
 ) -> Dict[str, str]:
     """Retrieves the inference environment variables for the model matching the given arguments.
 
@@ -59,6 +61,10 @@ def _retrieve_default_environment_variables(
             object, used for SageMaker interactions. If not
             specified, one is created using the default AWS configuration
             chain. (Default: sagemaker.jumpstart.constants.DEFAULT_JUMPSTART_SAGEMAKER_SESSION).
+        instance_type (str): An instance type to optionally supply in order to get
+            environment variables specific for the instance type.
+        script (JumpStartScriptScope): The JumpStart script for which to retrieve
+            environment variables.
     Returns:
         dict: the inference environment variables to use for the model.
     """
@@ -69,17 +75,37 @@ def _retrieve_default_environment_variables(
     model_specs = verify_model_region_and_return_specs(
         model_id=model_id,
         version=model_version,
-        scope=JumpStartScriptScope.INFERENCE,
+        scope=script,
         region=region,
         tolerate_vulnerable_model=tolerate_vulnerable_model,
         tolerate_deprecated_model=tolerate_deprecated_model,
         sagemaker_session=sagemaker_session,
     )
 
     default_environment_variables: Dict[str, str] = {}
-    for environment_variable in model_specs.inference_environment_variables:
-        if include_aws_sdk_env_vars or environment_variable.required_for_model_class:
-            default_environment_variables[environment_variable.name] = str(
-                environment_variable.default
+    if script == JumpStartScriptScope.INFERENCE:
+        for environment_variable in model_specs.inference_environment_variables:
+            if include_aws_sdk_env_vars or environment_variable.required_for_model_class:
+                default_environment_variables[environment_variable.name] = str(
+                    environment_variable.default
+                )
+
+    if instance_type:
+        if script == JumpStartScriptScope.INFERENCE and getattr(
+            model_specs, "hosting_instance_type_variants", None
+        ):
+            default_environment_variables.update(
+                model_specs.hosting_instance_type_variants.get_instance_specific_environment_variables(  # noqa E501  # pylint: disable=c0301
+                    instance_type
+                )
+            )
+        elif script == JumpStartScriptScope.TRAINING and getattr(
+            model_specs, "training_instance_type_variants", None
+        ):
+            default_environment_variables.update(
+                model_specs.training_instance_type_variants.get_instance_specific_environment_variables(  # noqa E501  # pylint: disable=c0301
+                    instance_type
+                )
             )
+
     return default_environment_variables

src/sagemaker/jumpstart/artifacts/image_uris.py

@@ -118,10 +118,35 @@ def _retrieve_image_uri(
     )
 
     if image_scope == JumpStartScriptScope.INFERENCE:
+        hosting_instance_type_variants = model_specs.hosting_instance_type_variants
+        if hosting_instance_type_variants:
+            image_uri = hosting_instance_type_variants.get_image_uri(
+                instance_type=instance_type, region=region
+            )
+            if image_uri is not None:
+                return image_uri
         ecr_specs = model_specs.hosting_ecr_specs
+        if ecr_specs is None:
+            raise ValueError(
+                f"No inference ECR configuration found for JumpStart model ID '{model_id}' "
+                f"with {instance_type} instance type in {region}. "
+                "Please try another instance type or region."
+            )
     elif image_scope == JumpStartScriptScope.TRAINING:
+        training_instance_type_variants = model_specs.training_instance_type_variants
+        if training_instance_type_variants:
+            image_uri = training_instance_type_variants.get_image_uri(
+                instance_type=instance_type, region=region
+            )
+            if image_uri is not None:
+                return image_uri
         ecr_specs = model_specs.training_ecr_specs
-
+        if ecr_specs is None:
+            raise ValueError(
+                f"No training ECR configuration found for JumpStart model ID '{model_id}' "
+                f"with {instance_type} instance type in {region}. "
+                "Please try another instance type or region."
+            )
     if framework is not None and framework != ecr_specs.framework:
         raise ValueError(
             f"Incorrect container framework '{framework}' for JumpStart model ID '{model_id}' "

src/sagemaker/jumpstart/factory/model.py

@@ -304,6 +304,8 @@ def _add_env_to_kwargs(kwargs: JumpStartModelInitKwargs) -> JumpStartModelInitKw
         tolerate_deprecated_model=kwargs.tolerate_deprecated_model,
         tolerate_vulnerable_model=kwargs.tolerate_vulnerable_model,
         sagemaker_session=kwargs.sagemaker_session,
+        script=JumpStartScriptScope.INFERENCE,
+        instance_type=kwargs.instance_type,
     )
 
     for key, value in extra_env_vars.items():