Add sample_optional, and fix pytest for parameters, optional

Author: kimdwkimdw

.github/workflows/pre-commit_pytest.yml

@@ -37,7 +37,7 @@ jobs:
     runs-on: ubuntu-latest
     needs: pre-commit
     container:
-      image: nvcr.io/nvidia/tritonserver:23.03-pyt-python-py3
+      image: nvcr.io/nvidia/tritonserver:23.08-pyt-python-py3
       options:
         --shm-size=1g
     steps:

README.md

@@ -55,6 +55,7 @@ if __name__ == "__main__":
 
 ## Release Notes
 
+- 23.08.30 Support `optional` with model input, `parameters` on config.pbtxt
 - 23.06.16 Support tritonclient>=2.34.0
 - Loosely modified the requirements related to tritonclient
 

bin/run_triton_tritony_sample.sh

@@ -0,0 +1,18 @@
+#!/bin/bash
+
+HERE=$(dirname "$(readlink -f $0)")
+PARENT_DIR=$(dirname "$HERE")
+
+docker run -it --rm --name triton_tritony \
+    -p8100:8000   \
+    -p8101:8001   \
+    -p8102:8002    \
+    -v "${PARENT_DIR}"/model_repository:/models:ro \
+    -e OMP_NUM_THREADS=2 \
+    -e OPENBLAS_NUM_THREADS=2 \
+    --shm-size=1g  \
+    nvcr.io/nvidia/tritonserver:23.08-pyt-python-py3 \
+    tritonserver --model-repository=/models \
+    --exit-timeout-secs 15 \
+    --min-supported-compute-capability 7.0 \
+    --log-verbose 0  # 0-nothing, 1-info, 2-debug, 3-trace

model_repository/sample/1/model.py

@@ -13,13 +13,11 @@ def initialize(self, args):
             pb_utils.triton_string_to_numpy(output_config["data_type"]) for output_config in output_configs
         ]
 
-        parameters = self.model_config["parameters"]
-
     def execute(self, requests):
         responses = [None for _ in requests]
         for idx, request in enumerate(requests):
             current_add_value = int(json.loads(request.parameters()).get("add", 0))
-            in_tensor = [item.as_numpy() + current_add_value for item in request.inputs()]
+            in_tensor = [item.as_numpy() + current_add_value for item in request.inputs() if item.name() == "model_in"]
             out_tensor = [
                 pb_utils.Tensor(output_name, x.astype(output_dtype))
                 for x, output_name, output_dtype in zip(in_tensor, self.output_name_list, self.output_dtype_list)

model_repository/sample/config.pbtxt

@@ -2,13 +2,6 @@ name: "sample"
 backend: "python"
 max_batch_size: 0
 
-parameters [
-  {
-    key: "add",
-    value: { string_value: "0" }
-  }
-]
-
 input [
 {
     name: "model_in"

model_repository/sample_optional/1/model.py

@@ -0,0 +1,37 @@
+import json
+
+import triton_python_backend_utils as pb_utils
+
+
+class TritonPythonModel:
+    def initialize(self, args):
+        self.model_config = model_config = json.loads(args["model_config"])
+        output_configs = model_config["output"]
+
+        self.output_name_list = [output_config["name"] for output_config in output_configs]
+        self.output_dtype_list = [
+            pb_utils.triton_string_to_numpy(output_config["data_type"]) for output_config in output_configs
+        ]
+
+    def execute(self, requests):
+        responses = [None for _ in requests]
+        for idx, request in enumerate(requests):
+            current_add_value = int(json.loads(request.parameters()).get("add", 0))
+            optional_in_tensor = pb_utils.get_input_tensor_by_name(request, "optional_model_sub")
+            if optional_in_tensor:
+                optional_in_tensor = optional_in_tensor.as_numpy()
+            else:
+                optional_in_tensor = 0
+
+            in_tensor = [
+                item.as_numpy() + current_add_value - optional_in_tensor
+                for item in request.inputs()
+                if item.name() == "model_in"
+            ]
+            out_tensor = [
+                pb_utils.Tensor(output_name, x.astype(output_dtype))
+                for x, output_name, output_dtype in zip(in_tensor, self.output_name_list, self.output_dtype_list)
+            ]
+            inference_response = pb_utils.InferenceResponse(output_tensors=out_tensor)
+            responses[idx] = inference_response
+        return responses

model_repository/sample_optional/config.pbtxt

@@ -0,0 +1,54 @@
+name: "sample_optional"
+backend: "python"
+max_batch_size: 0
+
+parameters [
+  {
+    key: "add",
+    value: { string_value: "0" }
+  }
+]
+
+input [
+{
+    name: "model_in"
+    data_type: TYPE_FP32
+    dims: [ -1 ]
+},
+{
+    name: "optional_model_sub"
+    data_type: TYPE_FP32
+    optional: true
+    dims: [ -1 ]
+}
+]
+
+output [
+{
+    name: "model_out"
+    data_type: TYPE_FP32
+    dims: [ -1 ]
+}
+]
+
+instance_group [{ kind: KIND_CPU, count: 1 }]
+
+model_warmup {
+  name: "RandomSampleInput"
+  batch_size: 1
+  inputs [{
+      key: "model_in"
+      value: {
+        data_type: TYPE_FP32
+        dims: [ 10 ]
+        random_data: true
+      }
+   }, {
+      key: "model_in"
+      value: {
+        data_type: TYPE_FP32
+        dims: [ 10 ]
+        zero_data: true
+      }
+   }]
+}

tests/test_model_call.py

@@ -5,47 +5,72 @@
 
 from tritony import InferenceClient
 
-MODEL_NAME = os.environ.get("MODEL_NAME", "sample")
 TRITON_HOST = os.environ.get("TRITON_HOST", "localhost")
 TRITON_HTTP = os.environ.get("TRITON_HTTP", "8000")
 TRITON_GRPC = os.environ.get("TRITON_GRPC", "8001")
 
 
+EPSILON = 1e-8
+
+
 @pytest.fixture(params=[("http", TRITON_HTTP), ("grpc", TRITON_GRPC)])
 def protocol_and_port(request):
     return request.param
 
 
-def get_client(protocol, port):
+def get_client(protocol, port, model_name):
     print(f"Testing {protocol}", flush=True)
-    return InferenceClient.create_with(MODEL_NAME, f"{TRITON_HOST}:{port}", protocol=protocol)
+    return InferenceClient.create_with(model_name, f"{TRITON_HOST}:{port}", protocol=protocol)
 
 
 def test_swithcing(protocol_and_port):
-    client = get_client(*protocol_and_port)
+    client = get_client(*protocol_and_port, model_name="sample")
 
     sample = np.random.rand(1, 100).astype(np.float32)
     result = client(sample)
     assert {np.isclose(result, sample).all()}
 
     sample_batched = np.random.rand(100, 100).astype(np.float32)
     client(sample_batched, model_name="sample_autobatching")
-    assert {np.isclose(result, sample).all()}
+    assert np.isclose(result, sample).all()
 
 
 def test_with_input_name(protocol_and_port):
-    client = get_client(*protocol_and_port)
+    client = get_client(*protocol_and_port, model_name="sample")
 
     sample = np.random.rand(100, 100).astype(np.float32)
     result = client({client.default_model_spec.model_input[0].name: sample})
-    assert {np.isclose(result, sample).all()}
+    assert np.isclose(result, sample).all()
 
 
 def test_with_parameters(protocol_and_port):
-    client = get_client(*protocol_and_port)
+    client = get_client(*protocol_and_port, model_name="sample")
 
     sample = np.random.rand(1, 100).astype(np.float32)
     ADD_VALUE = 1
     result = client({client.default_model_spec.model_input[0].name: sample}, parameters={"add": f"{ADD_VALUE}"})
 
-    assert {np.isclose(result[0], sample[0] + ADD_VALUE).all()}
+    assert np.isclose(result[0], sample[0] + ADD_VALUE).all()
+
+
+def test_with_optional(protocol_and_port):
+    client = get_client(*protocol_and_port, model_name="sample_optional")
+
+    sample = np.random.rand(1, 100).astype(np.float32)
+
+    result = client({client.default_model_spec.model_input[0].name: sample})
+    assert np.isclose(result[0], sample[0], rtol=EPSILON).all()
+
+    OPTIONAL_SUB_VALUE = np.zeros_like(sample) + 3
+    result = client(
+        {
+            client.default_model_spec.model_input[0].name: sample,
+            "optional_model_sub": OPTIONAL_SUB_VALUE,
+        }
+    )
+    assert np.isclose(result[0], sample[0] - OPTIONAL_SUB_VALUE, rtol=EPSILON).all()
+
+
+if __name__ == "__main__":
+    test_with_parameters(("grpc", "8101"))
+    test_with_optional(("grpc", "8101"))

tritony/tools.py

@@ -115,7 +115,11 @@ async def request_async(protocol: TritonProtocol, model_input: Dict, triton_clie
         loop = asyncio.get_running_loop()
 
         if "parameters" in grpc_get_inference_request.__code__.co_varnames:
-            model_input["parameters"] = None
+            # check tritonclient[all]>=2.34.0, NGC 23.04
+            model_input["parameters"] = model_input.get("parameters", None)
+        else:
+            logger.warning("tritonclient[all]<2.34.0, NGC 21.04")
+            model_input.pop("parameters")
         request = grpc_get_inference_request(
             **model_input,
             priority=0,