feat: Use specific images for SMP v2 jobs

src/sagemaker/image_uri_config/pytorch-smp.json

@@ -0,0 +1,37 @@
+{
+    "training": {
+        "processors": [
+            "gpu"
+        ],
+        "version_aliases": {
+            "2.0": "2.0.1"
+        },
+        "versions": {
+            "2.0.1": {
+                "py_versions": [
+                    "py310"
+                ],
+                "registries": {
+                    "ap-northeast-1": "658645717510",
+                    "ap-northeast-2": "658645717510",
+                    "ap-northeast-3": "658645717510",
+                    "ap-south-1": "658645717510",
+                    "ap-southeast-1": "658645717510",
+                    "ap-southeast-2": "658645717510",
+                    "ca-central-1": "658645717510",
+                    "eu-central-1": "658645717510",
+                    "eu-north-1": "658645717510",
+                    "eu-west-1": "658645717510",
+                    "eu-west-2": "658645717510",
+                    "eu-west-3": "658645717510",
+                    "sa-east-1": "658645717510",
+                    "us-east-1": "658645717510",
+                    "us-east-2": "658645717510",
+                    "us-west-1": "658645717510",
+                    "us-west-2": "658645717510"
+                },
+                "repository": "smdistributed-modelparallel"
+            }
+        }
+    }
+}

src/sagemaker/image_uris.py

@@ -27,7 +27,10 @@
 from sagemaker.jumpstart import artifacts
 from sagemaker.workflow import is_pipeline_variable
 from sagemaker.workflow.utilities import override_pipeline_parameter_var
-from sagemaker.fw_utils import GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY, GRAVITON_ALLOWED_FRAMEWORKS
+from sagemaker.fw_utils import (
+    GRAVITON_ALLOWED_TARGET_INSTANCE_FAMILY,
+    GRAVITON_ALLOWED_FRAMEWORKS,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -343,7 +346,8 @@ def _config_for_framework_and_scope(framework, image_scope, accelerator_type=Non
 
         if image_scope not in ("eia", "inference"):
             logger.warning(
-                "Elastic inference is for inference only. Ignoring image scope: %s.", image_scope
+                "Elastic inference is for inference only. Ignoring image scope: %s.",
+                image_scope,
             )
         image_scope = "eia"
 
@@ -660,6 +664,17 @@ def get_training_image_uri(
         container_version = None
         base_framework_version = None
 
+    # Check for smp library
+    if distribution is not None:
+        if "torch_distributed" in distribution and "smdistributed" in distribution:
+            if "modelparallel" in distribution["smdistributed"]:
+                if distribution["smdistributed"]["modelparallel"].get("enabled", True):
+                    framework = "pytorch-smp"
+                    if "p5" in instance_type:
+                        container_version = "cu121"
+                    else:
+                        container_version = "cu118"
+
     return retrieve(
         framework,
         region,

tests/unit/sagemaker/image_uris/test_smp_v2.py

@@ -0,0 +1,53 @@
+# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License"). You
+# may not use this file except in compliance with the License. A copy of
+# the License is located at
+#
+#     http://aws.amazon.com/apache2.0/
+#
+# or in the "license" file accompanying this file. This file is
+# distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF
+# ANY KIND, either express or implied. See the License for the specific
+# language governing permissions and limitations under the License.
+from __future__ import absolute_import
+
+import pytest
+from sagemaker import image_uris
+from tests.unit.sagemaker.image_uris import expected_uris
+
+CONTAINER_VERSIONS = {"ml.p4d.24xlarge": "cu118", "ml.p5d.24xlarge": "cu121"}
+
+
+@pytest.mark.parametrize("load_config", ["pytorch-smp.json"], indirect=True)
+def test_smp_v2(load_config):
+    VERSIONS = load_config["training"]["versions"]
+    PROCESSORS = load_config["training"]["processors"]
+    distribution = {
+        "torch_distributed": {"enabled": True},
+        "smdistributed": {"modelparallel": {"enabled": True}},
+    }
+    for processor in PROCESSORS:
+        for version in VERSIONS:
+            ACCOUNTS = load_config["training"]["versions"][version]["registries"]
+            PY_VERSIONS = load_config["training"]["versions"][version]["py_versions"]
+            for py_version in PY_VERSIONS:
+                for region in ACCOUNTS.keys():
+                    for instance_type in CONTAINER_VERSIONS.keys():
+                        uri = image_uris.get_training_image_uri(
+                            region,
+                            framework="pytorch",
+                            framework_version=version,
+                            py_version=py_version,
+                            distribution=distribution,
+                            instance_type=instance_type,
+                        )
+                        expected = expected_uris.framework_uri(
+                            repo="smdistributed-modelparallel",
+                            fw_version=version,
+                            py_version=f"{py_version}-{CONTAINER_VERSIONS[instance_type]}",
+                            processor=processor,
+                            region=region,
+                            account=ACCOUNTS[region],
+                        )
+                        assert expected == uri