fix: Adding ensemble support for vllm container (#68)

Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

Author: KrishnanPrash

.gitignore

@@ -50,6 +50,7 @@ coverage.xml
 .hypothesis/
 .pytest_cache/
 cover/
+*.out
 
 # Translations
 *.mo

README.md

@@ -100,6 +100,7 @@ export TRITON_CONTAINER_VERSION=<YY.MM>
                 --upstream-container-version=${TRITON_CONTAINER_VERSION}
                 --backend=python:r${TRITON_CONTAINER_VERSION}
                 --backend=vllm:r${TRITON_CONTAINER_VERSION}
+                --backend=ensemble
 ```
 
 ### Option 3. Add the vLLM Backend to the Default Triton Container

ci/L0_backend_vllm/vllm_backend/ensemble_config.pbtxt

@@ -0,0 +1,59 @@
+# Copyright 2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+name: "ensemble_model"
+platform: "ensemble"
+max_batch_size: 1
+input [
+  {
+    name: "text_input"
+    data_type: TYPE_STRING
+    dims: [ -1 ]
+  }
+]
+output [
+  {
+    name: "text_output"
+    data_type: TYPE_STRING
+    dims: [ -1 ]
+  }
+]
+ensemble_scheduling {
+  step [
+    {
+      model_name: "vllm_opt"
+      model_version: -1
+      input_map {
+        key: "text_input"
+        value: "text_input"
+      }
+      output_map {
+        key: "text_output"
+        value: "text_output"
+      }
+    }
+  ]
+}

ci/L0_backend_vllm/vllm_backend/test.sh

@@ -70,6 +70,11 @@ sed -i 's/"disable_log_requests"/"invalid_attribute"/' models/vllm_invalid_1/1/m
 cp -r ${SAMPLE_MODELS_REPO}/vllm_model models/vllm_invalid_2/
 sed -i 's/"facebook\/opt-125m"/"invalid_model"/' models/vllm_invalid_2/1/model.json
 
+
+# Sanity check ensembles are enabled and can successfully be loaded
+mkdir -p models/ensemble_model/1
+cp -r ensemble_config.pbtxt models/ensemble_model/config.pbtxt
+
 RET=0
 
 run_server
@@ -166,4 +171,4 @@ fi
 
 collect_artifacts_from_subdir
 
-exit $RET
+exit $RET

ci/L0_backend_vllm/vllm_backend/vllm_backend_test.py

@@ -48,6 +48,7 @@ def setUp(self):
         self.triton_client = grpcclient.InferenceServerClient(url="localhost:8001")
         self.vllm_model_name = "vllm_opt"
         self.python_model_name = "add_sub"
+        self.ensemble_model_name = "ensemble_model"
         self.vllm_load_test = "vllm_load_test"
 
     def test_vllm_triton_backend(self):
@@ -57,6 +58,13 @@ def test_vllm_triton_backend(self):
         self.triton_client.load_model(self.python_model_name)
         self.assertTrue(self.triton_client.is_model_ready(self.python_model_name))
 
+        # Test to ensure that ensemble models are supported in vllm container.
+        # If ensemble support not present, triton will error out at model loading stage.
+        # Ensemble Model is a pipeline consisting of 1 model (vllm_opt)
+        self.triton_client.load_model(self.ensemble_model_name)
+        self.assertTrue(self.triton_client.is_model_ready(self.ensemble_model_name))
+        self.triton_client.unload_model(self.ensemble_model_name)
+
         # Unload vllm model and test add_sub model
         self.triton_client.unload_model(self.vllm_load_test)
         self.assertFalse(self.triton_client.is_model_ready(self.vllm_load_test))