fix: run sagemaker tests for PR build

Author: Chuyang Deng

README.md

@@ -660,7 +660,6 @@ Multi-Model Endpoint can be used together with Pre/Post-Processing. Each model w
                 |--lib
                     |--external_module
                 |--inference.py
-                |--requirements.txt
 
 ## Contributing
 

buildspec.yml

@@ -29,11 +29,11 @@ phases:
         fi
 
       # run SageMaker tests
-      - |
-        if is-release-build; then
-          tox -e py37 -- -n 8 test/integration/sagemaker/test_tfs.py --versions 1.15.0
-          tox -e py37 -- -n 8 test/integration/sagemaker/test_tfs.py --versions 2.1.0
-        fi
+      - tox -e py37 -- -n 8 test/integration/sagemaker/test_tfs.py --versions 1.15.0
+      - tox -e py37 -- -n 8 test/integration/sagemaker/test_tfs.py --versions 2.1.0
+
+      - tox -e py37 -- test/integration/sagemaker/test_ei.py -n 8 --versions 1.14 --region {region} --registry {aws-id}
+
 
       # write deployment details to file
       # todo sort out eia versioning

docker/build_artifacts/sagemaker/serve.py

@@ -83,6 +83,7 @@ def __init__(self):
 
     def _create_tfs_config(self):
         models = tfs_utils.find_models()
+
         if not models:
             raise ValueError('no SavedModel bundles found!')
 
@@ -255,12 +256,6 @@ def start(self):
         self._state = 'starting'
         signal.signal(signal.SIGTERM, self._stop)
 
-        self._create_nginx_config()
-
-        if self._tfs_enable_batching:
-            log.info('batching is enabled')
-            tfs_utils.create_batching_config(self._tfs_batching_config_path)
-
         if self._tfs_enable_multi_model_endpoint:
             log.info('multi-model endpoint is enabled, TFS model servers will be started later')
         else:
@@ -271,6 +266,12 @@ def start(self):
             self._create_tfs_config()
             self._start_tfs()
 
+        self._create_nginx_config()
+
+        if self._tfs_enable_batching:
+            log.info('batching is enabled')
+            tfs_utils.create_batching_config(self._tfs_batching_config_path)
+
         if self._use_gunicorn:
             self._setup_gunicorn()
             self._start_gunicorn()

docker/build_artifacts/sagemaker/tfs_utils.py

@@ -111,8 +111,6 @@ def create_tfs_config(
         config += '  }\n'
     config += '}\n'
 
-    log.info('tensorflow serving model config: \n%s\n', config)
-
     with open(tfs_config_path, 'w') as f:
         f.write(config)
 

scripts/build-all.sh

@@ -6,7 +6,7 @@ set -euo pipefail
 
 DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-${DIR}/build.sh --version 1.14.0 --arch eia
+${DIR}/build.sh --version 2.0.0 --arch eia
 ${DIR}/build.sh --version 1.15.0 --arch cpu
 ${DIR}/build.sh --version 1.15.0 --arch gpu
 ${DIR}/build.sh --version 2.1.0 --arch cpu

scripts/publish-all.sh

@@ -6,7 +6,7 @@ set -euo pipefail
 
 DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-${DIR}/publish.sh --version 1.14.0 --arch eia
+${DIR}/publish.sh --version 2.0.0 --arch eia
 ${DIR}/publish.sh --version 1.15.0 --arch cpu
 ${DIR}/publish.sh --version 1.15.0 --arch gpu
 ${DIR}/publish.sh --version 2.1.0 --arch cpu

test/integration/local/conftest.py

@@ -53,3 +53,11 @@ def tag(request, framework_version, processor):
     if not image_tag:
         image_tag = '{}-{}'.format(framework_version, processor)
     return image_tag
+
+
+@pytest.fixture(autouse=True)
+def skip_by_device_type(request, processor):
+    is_gpu = processor == "gpu"
+    if (request.node.get_closest_marker('skip_gpu') and is_gpu) or \
+            (request.node.get_closest_marker('skip_cpu') and not is_gpu):
+        pytest.skip('Skipping because running on \'{}\' instance'.format(processor))

test/integration/local/test_container.py

@@ -45,7 +45,6 @@ def container(request, docker_base_name, tag, runtime_config):
         command = (
             'docker run {}--name sagemaker-tensorflow-serving-test -p 8080:8080'
             ' --mount type=volume,source=model_volume,target=/opt/ml/model,readonly'
-            ' -e SAGEMAKER_TFS_DEFAULT_MODEL_NAME=half_plus_three'
             ' -e SAGEMAKER_TFS_NGINX_LOGLEVEL=info'
             ' -e SAGEMAKER_BIND_TO_PORT=8080'
             ' -e SAGEMAKER_SAFE_PORT_RANGE=9000-9999'

test/integration/local/test_multi_model_endpoint.py

@@ -69,11 +69,13 @@ def container(request, docker_base_name, tag, runtime_config):
         subprocess.check_call('docker rm -f sagemaker-tensorflow-serving-test'.split())
 
 
+@pytest.mark.skip_gpu
 def test_ping():
     res = requests.get(PING_URL)
     assert res.status_code == 200
 
 
+@pytest.mark.skip_gpu
 def test_container_start_invocation_fail():
     x = {
         'instances': [1.0, 2.0, 5.0]
@@ -84,13 +86,15 @@ def test_container_start_invocation_fail():
     assert "Model half_plus_three is not loaded yet." in str(y)
 
 
+@pytest.mark.skip_gpu
 def test_list_models_empty():
     code, res = make_list_model_request()
     res = json.loads(res)
     assert code == 200
     assert len(res) == 0
 
 
+@pytest.mark.skip_gpu
 def test_delete_unloaded_model():
     # unloads the given model/version, no-op if not loaded
     model_name = 'non-existing-model'
@@ -99,6 +103,7 @@ def test_delete_unloaded_model():
     assert 'Model {} is not loaded yet'.format(model_name) in res
 
 
+@pytest.mark.skip_gpu
 def test_delete_model():
     model_name = 'half_plus_three'
     model_data = {
@@ -125,6 +130,7 @@ def test_delete_model():
     assert 'Model {} is not loaded yet.'.format(model_name) in str(y2)
 
 
+@pytest.mark.skip_gpu
 def test_load_two_models():
     model_name_1 = 'half_plus_two'
     model_data_1 = {
@@ -165,6 +171,7 @@ def test_load_two_models():
     assert len(res3) == 2
 
 
+@pytest.mark.skip_gpu
 def test_load_one_model_two_times():
     model_name = 'cifar'
     model_data = {
@@ -180,6 +187,7 @@ def test_load_one_model_two_times():
     assert'Model {} is already loaded'.format(model_name) in res2
 
 
+@pytest.mark.skip_gpu
 def test_load_non_existing_model():
     model_name = 'non-existing'
     base_path = '/opt/ml/models/non-existing'
@@ -192,6 +200,7 @@ def test_load_non_existing_model():
     assert 'Could not find valid base path {} for servable {}'.format(base_path, model_name) in str(res)
 
 
+@pytest.mark.skip_gpu
 def test_bad_model_reqeust():
     bad_model_data = {
         'model_name': 'model_name',
@@ -201,6 +210,7 @@ def test_bad_model_reqeust():
     assert code == 500
 
 
+@pytest.mark.skip_gpu
 def test_invalid_model_version():
     model_name = 'invalid_version'
     base_path = '/opt/ml/models/invalid_version'

test/integration/local/test_pre_post_processing.py

@@ -53,7 +53,6 @@ def container(volume, docker_base_name, tag, runtime_config):
         command = (
             'docker run {}--name sagemaker-tensorflow-serving-test -p 8080:8080'
             ' --mount type=volume,source={},target=/opt/ml/model,readonly'
-            ' -e SAGEMAKER_TFS_DEFAULT_MODEL_NAME=half_plus_three'
             ' -e SAGEMAKER_TFS_NGINX_LOGLEVEL=info'
             ' -e SAGEMAKER_BIND_TO_PORT=8080'
             ' -e SAGEMAKER_SAFE_PORT_RANGE=9000-9999'

test/integration/local/test_pre_post_processing_mme.py

@@ -92,18 +92,21 @@ def model():
     return MODEL_NAME
 
 
+@pytest.mark.skip_gpu
 def test_ping_service():
     response = requests.get(PING_URL)
     assert 200 == response.status_code
 
 
+@pytest.mark.skip_gpu
 def test_predict_json(model):
     headers = make_headers()
     data = '{"instances": [1.0, 2.0, 5.0]}'
     response = requests.post(INVOCATION_URL.format(model), data=data, headers=headers).json()
     assert response == {'predictions': [3.5, 4.0, 5.5]}
 
 
+@pytest.mark.skip_gpu
 def test_zero_content():
     headers = make_headers()
     x = ''
@@ -112,6 +115,7 @@ def test_zero_content():
     assert 'document is empty' in response.text
 
 
+@pytest.mark.skip_gpu
 def test_large_input():
     data_file = 'test/resources/inputs/test-large.csv'
 
@@ -123,13 +127,15 @@ def test_large_input():
         assert len(predictions) == 753936
 
 
+@pytest.mark.skip_gpu
 def test_csv_input():
     headers = make_headers(content_type='text/csv')
     data = '1.0,2.0,5.0'
     response = requests.post(INVOCATION_URL.format(MODEL_NAME), data=data, headers=headers).json()
     assert response == {'predictions': [3.5, 4.0, 5.5]}
 
 
+@pytest.mark.skip_gpu
 def test_unsupported_content_type():
     headers = make_headers('unsupported-type', 'predict')
     data = 'aW1hZ2UgYnl0ZXM='

test/integration/local/test_tfs_batching.py

@@ -41,7 +41,6 @@ def test_run_tfs_with_batching_parameters(docker_base_name, tag, runtime_config)
             ' -e SAGEMAKER_TFS_BATCH_TIMEOUT_MICROS=500'
             ' -e SAGEMAKER_TFS_NUM_BATCH_THREADS=100'
             ' -e SAGEMAKER_TFS_MAX_ENQUEUED_BATCHES=1'
-            ' -e SAGEMAKER_TFS_DEFAULT_MODEL_NAME=half_plus_three'
             ' -e SAGEMAKER_TFS_NGINX_LOGLEVEL=info'
             ' -e SAGEMAKER_BIND_TO_PORT=8080'
             ' -e SAGEMAKER_SAFE_PORT_RANGE=9000-9999'

test/integration/sagemaker/conftest.py

@@ -121,3 +121,11 @@ def skip_gpu_instance_restricted_regions(region, instance_type):
     if (region in NO_P2_REGIONS and instance_type.startswith('ml.p2')) or \
             (region in NO_P3_REGIONS and instance_type.startswith('ml.p3')):
         pytest.skip('Skipping GPU test in region {}'.format(region))
+
+
+@pytest.fixture(autouse=True)
+def skip_by_device_type(request, instance_type):
+    is_gpu = instance_type[3] in ['g', 'p']
+    if (request.node.get_closest_marker('skip_gpu') and is_gpu) or \
+            (request.node.get_closest_marker('skip_cpu') and not is_gpu):
+        pytest.skip('Skipping because running on \'{}\' instance'.format(instance_type))

tox.ini

@@ -78,3 +78,7 @@ whitelist_externals =
     jshint
 commands =
     jshint docker/build_artifacts/
+
+[pytest]
+markers =
+    skip_gpu: skip test if running on gpu instance