build(deps): Align protobuf and related dependencies with the gRPC v1.81.1 bump

CMakeLists.txt

@@ -167,6 +167,7 @@ if (OPENSSL_ROOT_DIR)
 endif()
 
 set(_FINDPACKAGE_PROTOBUF_CONFIG_DIR "${TRITON_THIRD_PARTY_INSTALL_PREFIX}/protobuf/${LIB_DIR}/cmake/protobuf")
+set(_FINDPACKAGE_UTF8_RANGE_CONFIG_DIR "${TRITON_THIRD_PARTY_INSTALL_PREFIX}/protobuf/${LIB_DIR}/cmake/utf8_range")
 
 set(_FINDPACKAGE_OPENTELEMETRY_CONFIG_DIR "${TRITON_THIRD_PARTY_INSTALL_PREFIX}/opentelemetry-cpp/${LIB_DIR}/cmake/opentelemetry-cpp")
 
@@ -199,6 +200,7 @@ ExternalProject_Add(triton-server
   BINARY_DIR "${CMAKE_CURRENT_BINARY_DIR}/triton-server"
   CMAKE_CACHE_ARGS
     -DProtobuf_DIR:PATH=${_FINDPACKAGE_PROTOBUF_CONFIG_DIR}
+    -Dutf8_range_DIR:PATH=${_FINDPACKAGE_UTF8_RANGE_CONFIG_DIR}
     ${_CMAKE_ARGS_OPENSSL_ROOT_DIR}
     ${_CMAKE_ARGS_CMAKE_TOOLCHAIN_FILE}
     ${_CMAKE_ARGS_VCPKG_TARGET_TRIPLET}

Dockerfile.QA

@@ -352,8 +352,8 @@ RUN apt-get update && apt-get install -y --no-install-recommends \
 RUN rm -f /usr/bin/python && \
     ln -s /usr/bin/python3 /usr/bin/python
 
-RUN pip3 install --upgrade "numpy<2" pillow attrdict future "grpcio>=1.81.1" requests gsutil \
-                           "awscli<=1.36.40" six grpcio-channelz prettytable virtualenv \
+RUN pip3 install --upgrade "numpy<2" pillow attrdict future "grpcio==1.81.1" requests gsutil \
+                           "awscli<=1.36.40" six "grpcio-channelz==1.81.1" prettytable virtualenv \
                            check-jsonschema
 
 # go needed for example go client test.

Dockerfile.sdk

@@ -83,7 +83,7 @@ RUN apt-get update && \
             software-properties-common \
             vim \
             wget && \
-    pip3 install --upgrade grpcio-tools cmake==4.0.3 auditwheel
+    pip3 install --upgrade "grpcio-tools==1.81.1" cmake==4.0.3 auditwheel
 
 ENV CMAKE_POLICY_MINIMUM_REQUIRED=3.5
 
@@ -182,7 +182,7 @@ RUN apt-get update && \
             python3-wheel \
             vim \
             wget && \
-    pip3 install "grpcio>=1.81.1" grpcio-tools && \
+    pip3 install "grpcio==1.81.1" "grpcio-tools==1.81.1" && \
     apt-get clean && \
     rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*;
 

qa/L0_lifecycle/lifecycle_test.py

@@ -1,6 +1,6 @@
 #!/usr/bin/env python3
 
-# Copyright 2018-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright 2018-2026, 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
@@ -26,6 +26,8 @@
 # (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
+# Local test modules are imported after adjusting sys.path for ../common.
+# flake8: noqa: E402
 import sys
 
 sys.path.append("../common")
@@ -308,7 +310,7 @@ def test_parse_error_no_model_config(self):
                 self.assertTrue(triton_client.is_server_live())
                 self.assertTrue(triton_client.is_server_ready())
 
-                md = triton_client.get_model_metadata(model_name, "1")
+                triton_client.get_model_metadata(model_name, "1")
                 self.assertTrue(
                     False,
                     "expected model '"
@@ -2256,7 +2258,6 @@ def test_multiple_model_repository_control_startup_models(self):
     def test_model_repository_index(self):
         # use model control EXPLICIT and --load-model to load a subset of models
         # in model repository
-        tensor_shape = (1, 16)
         model_bases = ["plan", "libtorch", "simple_libtorch"]
 
         # Sanity check on loaded models
@@ -2578,8 +2579,8 @@ def test_file_override_security(self):
         self.assertTrue(os.path.exists(os.path.join(model_basepath, existing_file_rel)))
 
         # Symlinks
-        ## No easy way to inject symlink into generated temp model dir, so for
-        ## testing sake, make a fixed symlink path in /tmp.
+        # No easy way to inject symlink into generated temp model dir, so for
+        # testing sake, make a fixed symlink path in /tmp.
         escape_dir_symlink_rel = os.path.join("..", "escape_symlink")
         escape_dir_symlink_full = "/tmp/escape_symlink"
         self.assertEqual(
@@ -2670,7 +2671,7 @@ def callback(user_data, result, error):
         except InferenceServerException as ex:
             self.assertIn(
                 "failed to connect to all addresses; last error: UNKNOWN: ipv4:127.0.0.1:8001: "
-                + "Failed to connect to remote host: connect: Connection refused (111)",
+                + "Failed to connect to remote host: Connection refused",
                 ex.message(),
             )
 
@@ -2682,7 +2683,7 @@ def callback(user_data, result, error):
 
         # Previous requests should succeed
         for result in async_results:
-            if type(result) == InferenceServerException:
+            if isinstance(result, InferenceServerException):
                 raise result
             output_data = result.as_numpy("OUTPUT0")
             np.testing.assert_allclose(
@@ -2732,17 +2733,24 @@ def callback(user_data, result, error):
             )
             self.assertTrue(False, "expected error for new inference during shutdown")
         except InferenceServerException as ex:
-            # The first request received by the gRPC endpoint while shutting down returns CANCELLED
-            # each subsequent request returns Connection refused
-            self.assertIn("CANCELLED", ex.message())
+            # The first request received by the gRPC endpoint while shutting down
+            # historically returned CANCELLED; gRPC >= 1.81 drops the connection
+            # immediately, so it now returns "Connection refused" like the
+            # subsequent requests. Accept either.
+            self.assertTrue(
+                "CANCELLED" in ex.message()
+                or "Failed to connect to remote host: Connection refused"
+                in ex.message(),
+                f"unexpected shutdown error: {ex.message()}",
+            )
         # 2: New sequence with existing sequence ID
         try:
             triton_client.infer(model_name, inputs, sequence_id=1, sequence_start=True)
             self.assertTrue(False, "expected error for new inference during shutdown")
         except InferenceServerException as ex:
             self.assertIn(
                 "failed to connect to all addresses; last error: UNKNOWN: ipv4:127.0.0.1:8001: "
-                + "Failed to connect to remote host: connect: Connection refused (111)",
+                + "Failed to connect to remote host: Connection refused",
                 ex.message(),
             )
         # 3: Continuing sequence after shutdown
@@ -2752,7 +2760,7 @@ def callback(user_data, result, error):
         except InferenceServerException as ex:
             self.assertIn(
                 "failed to connect to all addresses; last error: UNKNOWN: ipv4:127.0.0.1:8001: "
-                + "Failed to connect to remote host: connect: Connection refused (111)",
+                + "Failed to connect to remote host: Connection refused",
                 ex.message(),
             )
 
@@ -2764,7 +2772,7 @@ def callback(user_data, result, error):
 
         # Previous requests should succeed
         for result in async_results:
-            if type(result) == InferenceServerException:
+            if isinstance(result, InferenceServerException):
                 raise result
             output_data = result.as_numpy("OUTPUT")
             np.testing.assert_allclose(
@@ -2815,7 +2823,7 @@ def callback(user_data, result, error):
         except InferenceServerException as ex:
             self.assertIn(
                 "failed to connect to all addresses; last error: UNKNOWN: ipv4:127.0.0.1:8001: "
-                + "Failed to connect to remote host: connect: Connection refused (111)",
+                + "Failed to connect to remote host: Connection refused",
                 ex.message(),
             )
 
@@ -2827,7 +2835,7 @@ def callback(user_data, result, error):
 
         # Previous requests should succeed
         for result in async_results:
-            if type(result) == InferenceServerException:
+            if isinstance(result, InferenceServerException):
                 raise result
             output_data = result.as_numpy("OUTPUT0")
             np.testing.assert_allclose(
@@ -3073,7 +3081,7 @@ def _load_unload():
         # This test can replicate a load while async unloading on machines with
         # sufficient concurrency. Regardless on whether it is replicated or not,
         # the server must not crash.
-        if load_before_unload_finish[0] == False:
+        if load_before_unload_finish[0] is False:
             # Track non-replication on test printout via statistics.
             warning_msg = "Cannot replicate a load while async unloading. CPU count: {}. num_threads: {}.".format(
                 multiprocessing.cpu_count(), num_threads
@@ -3299,7 +3307,9 @@ def test_model_config_overwite(self):
 """
 
         # Ensure the model has been loaded w/ the expected (different from override) config.
-        self.assertTrue(original_config != None and original_config != override_config)
+        self.assertTrue(
+            original_config is not None and original_config != override_config
+        )
 
         # Reload the model with the overriding configuration value.
         triton_client.load_model(model_name, config=override_config)
@@ -3362,7 +3372,6 @@ def test_shutdown_while_loading(self):
     def test_shutdown_with_live_connection(self):
         model_name = "add_sub"
         model_shape = (16,)
-        from geventhttpclient.response import HTTPConnectionClosed
 
         input_data = np.ones(shape=model_shape, dtype=np.float32)
         inputs = [

src/CMakeLists.txt

@@ -90,6 +90,17 @@ endif()
 #
 find_package(re2 REQUIRED)
 
+# protobuf
+#
+# protobuf v33's libprotobuf.a calls into abseil internally, and the
+# tritonserver executable statically links libprotobuf.a (via the gRPC
+# endpoint). Linking the protobuf package target directly ensures its own
+# abseil dependencies are pulled onto the executable's link line, so builds
+# that do not enable metrics/GPU (which otherwise pull abseil incidentally)
+# still resolve those symbols. protobuf carries its own abseil deps, so no
+# abseil target is referenced here.
+find_package(Protobuf CONFIG REQUIRED)
+
 #
 # tritonserver executable
 #
@@ -150,6 +161,7 @@ target_link_libraries(
     triton-common-logging           # from repo-common
     triton-core-serverapi           # from repo-core
     triton-core-serverstub          # from repo-core
+    protobuf::libprotobuf           # brings its own abseil deps
 )
 
 if(${TRITON_ENABLE_ASAN})

src/grpc/grpc_server.cc

@@ -866,10 +866,15 @@ CommonHandler::RegisterModelConfig()
         err = TRITONSERVER_MessageSerializeToJson(
             model_config_message, &buffer, &byte_size);
         if (err == nullptr) {
-          ::google::protobuf::util::JsonStringToMessage(
-              ::google::protobuf::stringpiece_internal::StringPiece(
-                  buffer, (int)byte_size),
-              response->mutable_config());
+          const auto parse_status =
+              ::google::protobuf::util::JsonStringToMessage(
+                  absl::string_view(buffer, byte_size),
+                  response->mutable_config());
+          if (!parse_status.ok()) {
+            err = TRITONSERVER_ErrorNew(
+                TRITONSERVER_ERROR_INTERNAL,
+                std::string(parse_status.message()).c_str());
+          }
         }
         TRITONSERVER_MessageDelete(model_config_message);
       }