added test dependencies to wheel, removed typing_extensions

python/setup.py

@@ -72,6 +72,10 @@ def get_tag(self):
     "_c/triton_bindings.pyi",
 ]
 
+gpu_extras = ["cupy-cuda12x"]
+test_extras = ["pytest"]
+all_extras = gpu_extras + test_extras
+
 setup(
     name="tritonserver",
     version=VERSION,
@@ -105,5 +109,5 @@ def get_tag(self):
     cmdclass={"bdist_wheel": bdist_wheel},
     data_files=data_files,
     install_requires=["numpy"],
-    extras_require={"GPU": ["cupy-cuda12x"], "all": ["cupy-cuda12x"]},
+    extras_require={"GPU": gpu_extras, "test": test_extras, "all": all_extras},
 )

python/test/test_api.py

@@ -29,6 +29,7 @@
 import queue
 import time
 import unittest
+import os
 
 import numpy
 import pytest
@@ -44,17 +45,19 @@
 except ImportError:
     torch = None
 
+module_directory = os.path.split(os.path.abspath(__file__))[0]
+test_model_directory = os.path.abspath(os.path.join(module_directory,"test_api_models"))
+
 server_options = tritonserver.Options(
     server_id="TestServer",
-    model_repository="/workspace/test/test_api_models",
+    model_repository=test_model_directory,
     log_verbose=0,
     log_error=True,
     exit_on_error=True,
     strict_model_config=False,
     model_control_mode=tritonserver.ModelControlMode.EXPLICIT,
 )
 
-
 class ModelTests(unittest.TestCase):
     def test_create_request(self):
         server = tritonserver.Server(server_options).start(wait_until_ready=True)

python/tritonserver/_api/_model.py

@@ -44,7 +44,6 @@
     TRITONSERVER_ModelTxnPropertyFlag as ModelTxnPropertyFlag,
 )
 from tritonserver._c.triton_bindings import TRITONSERVER_Server
-from typing_extensions import Unpack
 
 if TYPE_CHECKING:
     from tritonserver._api._server import Server