SMP PT upgrade to 2.1 (#4404)

adtian2 · Andrew Tian · web-flow · commit a9069210aa29 · 2024-02-05T09:07:05.000-08:00
* adding pt 2.1 support for smp

* adding supported PT versions of 2.1.2

* bug fix in smp tests

* adding 2.1.2

* adding 2.1.0 for Herring changes

* changed image_uri generation logic for PT version 2.1.2 for smp

* fixed this unit test to reflect changes for version PT 2.1

---------

Co-authored-by: Andrew Tian &lt;tinandr@amazon.com&gt;
diff --git a/src/sagemaker/fw_utils.py b/src/sagemaker/fw_utils.py
@@ -139,6 +139,8 @@
         "1.13.1",
         "2.0.0",
         "2.0.1",
+        "2.1.0",
+        "2.1.2",
     ],
 }
 
@@ -158,7 +160,7 @@
 ]
 
 
-TORCH_DISTRIBUTED_GPU_SUPPORTED_FRAMEWORK_VERSIONS = ["1.13.1", "2.0.0", "2.0.1", "2.1.0"]
+TORCH_DISTRIBUTED_GPU_SUPPORTED_FRAMEWORK_VERSIONS = ["1.13.1", "2.0.0", "2.0.1", "2.1.0", "2.1.2"]
 
 TRAINIUM_SUPPORTED_DISTRIBUTION_STRATEGIES = ["torch_distributed"]
 TRAINIUM_SUPPORTED_TORCH_DISTRIBUTED_FRAMEWORK_VERSIONS = [
diff --git a/src/sagemaker/image_uri_config/pytorch-smp.json b/src/sagemaker/image_uri_config/pytorch-smp.json
@@ -4,7 +4,8 @@
             "gpu"
         ],
         "version_aliases": {
-            "2.0": "2.0.1"
+            "2.0": "2.0.1",
+            "2.1": "2.1.2"
         },
         "versions": {
             "2.0.1": {
@@ -31,7 +32,32 @@
                     "us-west-2": "658645717510"
                 },
                 "repository": "smdistributed-modelparallel"
+            },
+            "2.1.2": {
+                "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"
             }
         }
     }
-}
+}
diff --git a/src/sagemaker/image_uris.py b/src/sagemaker/image_uris.py
@@ -672,7 +672,7 @@ def get_training_image_uri(
             if "modelparallel" in distribution["smdistributed"]:
                 if distribution["smdistributed"]["modelparallel"].get("enabled", True):
                     framework = "pytorch-smp"
-                    if "p5" in instance_type:
+                    if "p5" in instance_type or "2.1" in framework_version:
                         container_version = "cu121"
                     else:
                         container_version = "cu118"
diff --git a/tests/unit/sagemaker/image_uris/test_smp_v2.py b/tests/unit/sagemaker/image_uris/test_smp_v2.py
@@ -16,7 +16,7 @@
 from sagemaker import image_uris
 from tests.unit.sagemaker.image_uris import expected_uris
 
-CONTAINER_VERSIONS = {"ml.p4d.24xlarge": "cu118", "ml.p5d.24xlarge": "cu121"}
+CONTAINER_VERSIONS = {"ml.p4d.24xlarge": "cu118", "ml.p5.24xlarge": "cu121"}
 
 
 @pytest.mark.parametrize("load_config", ["pytorch-smp.json"], indirect=True)
@@ -34,6 +34,10 @@ def test_smp_v2(load_config):
             for py_version in PY_VERSIONS:
                 for region in ACCOUNTS.keys():
                     for instance_type in CONTAINER_VERSIONS.keys():
+                        cuda_vers = CONTAINER_VERSIONS[instance_type]
+                        if "2.1" in version:
+                            cuda_vers = "cu121"
+
                         uri = image_uris.get_training_image_uri(
                             region,
                             framework="pytorch",
@@ -45,7 +49,7 @@ def test_smp_v2(load_config):
                         expected = expected_uris.framework_uri(
                             repo="smdistributed-modelparallel",
                             fw_version=version,
-                            py_version=f"{py_version}-{CONTAINER_VERSIONS[instance_type]}",
+                            py_version=f"{py_version}-{cuda_vers}",
                             processor=processor,
                             region=region,
                             account=ACCOUNTS[region],

Original file line number	Diff line number	Diff line change
`@@ -139,6 +139,8 @@`
`139`	`139`	`"1.13.1",`
`140`	`140`	`"2.0.0",`
`141`	`141`	`"2.0.1",`
	`142`	`+ "2.1.0",`
	`143`	`+ "2.1.2",`
`142`	`144`	`],`
`143`	`145`	`}`
`144`	`146`
`@@ -158,7 +160,7 @@`
`158`	`160`	`]`
`159`	`161`
`160`	`162`
`161`		`-TORCH_DISTRIBUTED_GPU_SUPPORTED_FRAMEWORK_VERSIONS = ["1.13.1", "2.0.0", "2.0.1", "2.1.0"]`
	`163`	`+TORCH_DISTRIBUTED_GPU_SUPPORTED_FRAMEWORK_VERSIONS = ["1.13.1", "2.0.0", "2.0.1", "2.1.0", "2.1.2"]`
`162`	`164`
`163`	`165`	`TRAINIUM_SUPPORTED_DISTRIBUTION_STRATEGIES = ["torch_distributed"]`
`164`	`166`	`TRAINIUM_SUPPORTED_TORCH_DISTRIBUTED_FRAMEWORK_VERSIONS = [`