feature: Adding support for SageMaker Training Compiler PyTorch 1.13

src/sagemaker/image_uri_config/pytorch-training-compiler.json

@@ -4,7 +4,8 @@
             "gpu"
         ],
         "version_aliases": {
-            "1.12": "1.12.0"
+            "1.12": "1.12.0",
+            "1.13": "1.13.1"
         },
         "versions": {
             "1.12.0": {
@@ -35,6 +36,35 @@
                     "us-west-2": "763104351884"
                 },
                 "repository": "pytorch-trcomp-training"
+            },
+            "1.13.1": {
+                "py_versions": [
+                    "py39"
+                ],
+                "registries": {
+                    "af-south-1": "626614931356",
+                    "ap-east-1": "871362719292",
+                    "ap-northeast-1": "763104351884",
+                    "ap-northeast-2": "763104351884",
+                    "ap-northeast-3": "364406365360",
+                    "ap-south-1": "763104351884",
+                    "ap-southeast-1": "763104351884",
+                    "ap-southeast-2": "763104351884",
+                    "ca-central-1": "763104351884",
+                    "eu-central-1": "763104351884",
+                    "eu-north-1": "763104351884",
+                    "eu-west-1": "763104351884",
+                    "eu-west-2": "763104351884",
+                    "eu-west-3": "763104351884",
+                    "eu-south-1": "692866216735",
+                    "me-south-1": "217643126080",
+                    "sa-east-1": "763104351884",
+                    "us-east-1": "763104351884",
+                    "us-east-2": "763104351884",
+                    "us-west-1": "763104351884",
+                    "us-west-2": "763104351884"
+                },
+                "repository": "pytorch-trcomp-training"
             }
         }
     }

tests/conftest.py

@@ -318,6 +318,16 @@ def huggingface_pytorch_latest_training_py_version(
     )
 
 
+@pytest.fixture(scope="module")
+def pytorch_training_compiler_py_version(
+    pytorch_training_compiler_version,
+):
+    return "py39" if Version(pytorch_training_compiler_version) > Version("1.12") else "py38"
+
+
+# TODO: Create a fixture to get the latest py version from TRCOMP image_uri.
+
+
 @pytest.fixture(scope="module")
 def huggingface_pytorch_latest_inference_py_version(
     huggingface_inference_pytorch_latest_version,

tests/integ/test_training_compiler.py

@@ -150,7 +150,7 @@ def test_pytorch(
     with timeout(minutes=TRAINING_DEFAULT_TIMEOUT_MINUTES):
 
         hf = PyTorch(
-            py_version="py38",
+            py_version="py39",
             source_dir=os.path.join(DATA_DIR, "huggingface_byoc"),
             entry_point="run_glue.py",
             role="SageMakerRole",
@@ -216,7 +216,10 @@ def test_huggingface_tensorflow(
 
 @pytest.mark.release
 def test_tensorflow(
-    sagemaker_session, gpu_instance_type, tensorflow_training_latest_version, imagenet_val_set
+    sagemaker_session,
+    gpu_instance_type,
+    tensorflow_training_latest_version,
+    imagenet_val_set,
 ):
     """
     Test the TensorFlow estimator

tests/unit/sagemaker/training_compiler/test_pytorch_compiler.py

@@ -83,22 +83,26 @@ def fixture_sagemaker_session():
     return session
 
 
-def _get_full_gpu_image_uri(version, instance_type, training_compiler_config):
+def _get_full_gpu_image_uri(version, instance_type, training_compiler_config, py_version):
     return image_uris.retrieve(
         "pytorch-training-compiler",
         REGION,
         version=version,
-        py_version="py38",
+        py_version=py_version,
         instance_type=instance_type,
         image_scope="training",
         container_version=None,
         training_compiler_config=training_compiler_config,
     )
 
 
-def _create_train_job(version, instance_type, training_compiler_config, instance_count=1):
+def _create_train_job(
+    version, instance_type, training_compiler_config, py_version, instance_count=1
+):
     return {
-        "image_uri": _get_full_gpu_image_uri(version, instance_type, training_compiler_config),
+        "image_uri": _get_full_gpu_image_uri(
+            version, instance_type, training_compiler_config, py_version
+        ),
         "input_mode": "File",
         "input_config": [
             {
@@ -303,15 +307,20 @@ def test_unsupported_distribution(
 @patch("time.time", return_value=TIME)
 @pytest.mark.parametrize("instance_class", SUPPORTED_GPU_INSTANCE_CLASSES)
 def test_pytorchxla_distribution(
-    time, name_from_base, sagemaker_session, pytorch_training_compiler_version, instance_class
+    time,
+    name_from_base,
+    sagemaker_session,
+    pytorch_training_compiler_version,
+    instance_class,
+    pytorch_training_compiler_py_version,
 ):
     if Version(pytorch_training_compiler_version) < Version("1.12"):
         pytest.skip("This test is intended for PyTorch 1.12 and above")
     compiler_config = TrainingCompilerConfig()
     instance_type = f"ml.{instance_class}.xlarge"
 
     pt = PyTorch(
-        py_version="py38",
+        py_version=pytorch_training_compiler_py_version,
         entry_point=SCRIPT_PATH,
         role=ROLE,
         sagemaker_session=sagemaker_session,
@@ -333,7 +342,11 @@ def test_pytorchxla_distribution(
     assert boto_call_names == ["resource"]
 
     expected_train_args = _create_train_job(
-        pytorch_training_compiler_version, instance_type, compiler_config, instance_count=2
+        pytorch_training_compiler_version,
+        instance_type,
+        compiler_config,
+        pytorch_training_compiler_py_version,
+        instance_count=2,
     )
     expected_train_args["input_config"][0]["DataSource"]["S3DataSource"]["S3Uri"] = inputs
     expected_train_args["enable_sagemaker_metrics"] = False
@@ -357,13 +370,17 @@ def test_pytorchxla_distribution(
 @patch("time.time", return_value=TIME)
 @pytest.mark.parametrize("instance_class", SUPPORTED_GPU_INSTANCE_CLASSES)
 def test_default_compiler_config(
-    time, name_from_base, sagemaker_session, pytorch_training_compiler_version, instance_class
+    time,
+    name_from_base,
+    sagemaker_session,
+    pytorch_training_compiler_version,
+    instance_class,
+    pytorch_training_compiler_py_version,
 ):
     compiler_config = TrainingCompilerConfig()
     instance_type = f"ml.{instance_class}.xlarge"
-
     pt = PyTorch(
-        py_version="py38",
+        py_version=pytorch_training_compiler_py_version,
         entry_point=SCRIPT_PATH,
         role=ROLE,
         sagemaker_session=sagemaker_session,
@@ -384,7 +401,10 @@ def test_default_compiler_config(
     assert boto_call_names == ["resource"]
 
     expected_train_args = _create_train_job(
-        pytorch_training_compiler_version, instance_type, compiler_config
+        pytorch_training_compiler_version,
+        instance_type,
+        compiler_config,
+        pytorch_training_compiler_py_version,
     )
     expected_train_args["input_config"][0]["DataSource"]["S3DataSource"]["S3Uri"] = inputs
     expected_train_args["enable_sagemaker_metrics"] = False
@@ -406,12 +426,16 @@ def test_default_compiler_config(
 @patch("sagemaker.estimator.name_from_base", return_value=JOB_NAME)
 @patch("time.time", return_value=TIME)
 def test_debug_compiler_config(
-    time, name_from_base, sagemaker_session, pytorch_training_compiler_version
+    time,
+    name_from_base,
+    sagemaker_session,
+    pytorch_training_compiler_version,
+    pytorch_training_compiler_py_version,
 ):
     compiler_config = TrainingCompilerConfig(debug=True)
 
     pt = PyTorch(
-        py_version="py38",
+        py_version=pytorch_training_compiler_py_version,
         entry_point=SCRIPT_PATH,
         role=ROLE,
         sagemaker_session=sagemaker_session,
@@ -432,7 +456,10 @@ def test_debug_compiler_config(
     assert boto_call_names == ["resource"]
 
     expected_train_args = _create_train_job(
-        pytorch_training_compiler_version, INSTANCE_TYPE, compiler_config
+        pytorch_training_compiler_version,
+        INSTANCE_TYPE,
+        compiler_config,
+        pytorch_training_compiler_py_version,
     )
     expected_train_args["input_config"][0]["DataSource"]["S3DataSource"]["S3Uri"] = inputs
     expected_train_args["enable_sagemaker_metrics"] = False
@@ -454,12 +481,16 @@ def test_debug_compiler_config(
 @patch("sagemaker.estimator.name_from_base", return_value=JOB_NAME)
 @patch("time.time", return_value=TIME)
 def test_disable_compiler_config(
-    time, name_from_base, sagemaker_session, pytorch_training_compiler_version
+    time,
+    name_from_base,
+    sagemaker_session,
+    pytorch_training_compiler_version,
+    pytorch_training_compiler_py_version,
 ):
     compiler_config = TrainingCompilerConfig(enabled=False)
 
     pt = PyTorch(
-        py_version="py38",
+        py_version=pytorch_training_compiler_py_version,
         entry_point=SCRIPT_PATH,
         role=ROLE,
         sagemaker_session=sagemaker_session,
@@ -480,7 +511,10 @@ def test_disable_compiler_config(
     assert boto_call_names == ["resource"]
 
     expected_train_args = _create_train_job(
-        pytorch_training_compiler_version, INSTANCE_TYPE, compiler_config
+        pytorch_training_compiler_version,
+        INSTANCE_TYPE,
+        compiler_config,
+        pytorch_training_compiler_py_version,
     )
     expected_train_args["input_config"][0]["DataSource"]["S3DataSource"]["S3Uri"] = inputs
     expected_train_args["enable_sagemaker_metrics"] = False
@@ -508,7 +542,10 @@ def test_attach(sagemaker_session, compiler_enabled, debug_enabled):
         "py38-cu113-ubuntu20.04"
     )
     returned_job_description = {
-        "AlgorithmSpecification": {"TrainingInputMode": "File", "TrainingImage": training_image},
+        "AlgorithmSpecification": {
+            "TrainingInputMode": "File",
+            "TrainingImage": training_image,
+        },
         "HyperParameters": {
             "sagemaker_submit_directory": '"s3://some/sourcedir.tar.gz"',
             "sagemaker_program": '"iris-dnn-classifier.py"',
@@ -530,7 +567,10 @@ def test_attach(sagemaker_session, compiler_enabled, debug_enabled):
         "TrainingJobName": "trcomp",
         "TrainingJobStatus": "Completed",
         "TrainingJobArn": "arn:aws:sagemaker:us-west-2:336:training-job/trcomp",
-        "OutputDataConfig": {"KmsKeyId": "", "S3OutputPath": "s3://place/output/trcomp"},
+        "OutputDataConfig": {
+            "KmsKeyId": "",
+            "S3OutputPath": "s3://place/output/trcomp",
+        },
         "TrainingJobOutput": {"S3TrainingJobOutput": "s3://here/output.tar.gz"},
     }
     sagemaker_session.sagemaker_client.describe_training_job = Mock(