Cherry picking #5283 and #5519 into 2.0.1-verified-patch to fix gRPC for mac m1 (MLA 2259) (#5602)

* Only update the native library and the c# calls that need to be made to clean up channels. (#5283)

* Fix Mac backcompat test (#5519)

Co-authored-by: Chris Goy <[email protected]>

Author: jrupert-unity

.yamato/training-backcompat-tests.yml

@@ -1,19 +1,19 @@
 
-test_mac_backcompat_2020.1:
-  {% capture editor_version %}2020.1{% endcapture %}
+test_mac_backcompat_2019.4:
+  {% capture editor_version %}2019.4{% endcapture %}
   {% capture csharp_backcompat_version %}1.0.0{% endcapture %}
   # This test has to run on mac because it requires the custom build of tensorflow without AVX
   # Test against 2020.1 because 2020.2 has to run against package version 1.2.0
   name: Test Mac Backcompat Training {{ editor_version }}
   agent:
     type: Unity::VM::osx
-    image: ml-agents/ml-agents-bokken-mac:0.1.4-492264
+    image: ml-agents/ml-agents-bokken-mac:0.1.5-853758
     flavor: b1.small
   variables:
     UNITY_VERSION: {{ editor_version }}
   commands:
     - |
-      python3 -m venv venv && source venv/bin/activate
+      python -m venv venv && source venv/bin/activate
       python -m pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
       python -m pip install unity-downloader-cli --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple --upgrade
       unity-downloader-cli -u {{ editor_version }} -c editor --wait --fast
@@ -23,9 +23,9 @@ test_mac_backcompat_2020.1:
       python -u -m ml-agents.tests.yamato.standalone_build_tests --build-target=mac
       python -u -m ml-agents.tests.yamato.training_int_tests --csharp {{ csharp_backcompat_version }}
     - |
-      python3 -m venv venv_old && source venv_old/bin/activate
+      python -m venv venv_old && source venv_old/bin/activate
       python -m pip install pyyaml --index-url https://artifactory.prd.it.unity3d.com/artifactory/api/pypi/pypi/simple
-      python -u -m ml-agents.tests.yamato.training_int_tests --python 0.16.0
+      python -u -m ml-agents.tests.yamato.training_int_tests --python 0.24.0
   triggers:
     cancel_old_ci: true
     recurring:

com.unity.ml-agents/Plugins/ProtoBuffer/runtimes/osx/native/libgrpc_csharp_ext.x64.bundle

@@ -2,7 +2,7 @@
 #define MLA_SUPPORTED_TRAINING_PLATFORM
 #endif
 
-# if MLA_SUPPORTED_TRAINING_PLATFORM
+#if MLA_SUPPORTED_TRAINING_PLATFORM
 using Grpc.Core;
 #if UNITY_EDITOR
 using UnityEditor;
@@ -50,6 +50,7 @@ internal class RpcCommunicator : ICommunicator
 
         /// The Unity to External client.
         UnityToExternalProto.UnityToExternalProtoClient m_Client;
+        Channel m_Channel;
 
         /// <summary>
         /// Initializes a new instance of the RPCCommunicator class.
@@ -140,6 +141,7 @@ out input
                     Debug.Log($"Unexpected exception when trying to initialize communication: {ex}");
                 }
                 initParametersOut = new UnityRLInitParameters();
+                NotifyQuitAndShutDownChannel();
                 return false;
             }
 
@@ -181,6 +183,8 @@ out input
 
             UpdateEnvironmentWithInput(input.RlInput);
             initParametersOut = initializationInput.RlInitializationInput.ToUnityRLInitParameters();
+            // Be sure to shut down the grpc channel when the application is quitting.
+            Application.quitting += NotifyQuitAndShutDownChannel;
             return true;
 #else
             initParametersOut = new UnityRLInitParameters();
@@ -217,9 +221,9 @@ void UpdateEnvironmentWithInput(UnityRLInputProto rlInput)
         UnityInputProto Initialize(int port, UnityOutputProto unityOutput, out UnityInputProto unityInput)
         {
             m_IsOpen = true;
-            var channel = new Channel($"localhost:{port}", ChannelCredentials.Insecure);
+            m_Channel = new Channel($"localhost:{port}", ChannelCredentials.Insecure);
 
-            m_Client = new UnityToExternalProto.UnityToExternalProtoClient(channel);
+            m_Client = new UnityToExternalProto.UnityToExternalProtoClient(m_Channel);
             var result = m_Client.Exchange(WrapMessage(unityOutput, 200));
             var inputMessage = m_Client.Exchange(WrapMessage(null, 200));
             unityInput = inputMessage.UnityInput;
@@ -229,11 +233,24 @@ UnityInputProto Initialize(int port, UnityOutputProto unityOutput, out UnityInpu
             if (result.Header.Status != 200 || inputMessage.Header.Status != 200)
             {
                 m_IsOpen = false;
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
             }
             return result.UnityInput;
         }
 
+        void NotifyQuitAndShutDownChannel()
+        {
+            QuitCommandReceived?.Invoke();
+            try
+            {
+                m_Channel.ShutdownAsync().Wait();
+            }
+            catch (Exception)
+            {
+                // do nothing
+            }
+        }
+
 #endregion
 
 #region Destruction
@@ -269,7 +286,7 @@ void SendCommandEvent(CommandProto command)
             {
                 case CommandProto.Quit:
                     {
-                        QuitCommandReceived?.Invoke();
+                        NotifyQuitAndShutDownChannel();
                         return;
                     }
                 case CommandProto.Reset:
@@ -456,7 +473,7 @@ UnityInputProto Exchange(UnityOutputProto unityOutput)
                 // Not sure if the quit command is actually sent when a
                 // non 200 message is received.  Notify that we are indeed
                 // quitting.
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
                 return message.UnityInput;
             }
             catch (Exception ex)
@@ -488,7 +505,7 @@ UnityInputProto Exchange(UnityOutputProto unityOutput)
                 }
 
                 m_IsOpen = false;
-                QuitCommandReceived?.Invoke();
+                NotifyQuitAndShutDownChannel();
                 return null;
             }
         }

com.unity.ml-agents/Runtime/Communicator/RpcCommunicator.cs

@@ -5,7 +5,6 @@
 import sys
 import subprocess
 import time
-from typing import Any
 
 from .yamato_utils import (
     find_executables,
@@ -14,7 +13,6 @@
     run_standalone_build,
     init_venv,
     override_config_file,
-    override_legacy_config_file,
     checkout_csharp_version,
     undo_git_checkout,
 )
@@ -26,7 +24,7 @@ def run_training(python_version: str, csharp_version: str) -> bool:
     print(
         f"Running training with python={python_version or latest} and c#={csharp_version or latest}"
     )
-    output_dir = "models" if python_version else "results"
+    output_dir = "results"
     onnx_file_expected = f"./{output_dir}/{run_id}/3DBall.onnx"
 
     if os.path.exists(onnx_file_expected):
@@ -70,17 +68,11 @@ def run_training(python_version: str, csharp_version: str) -> bool:
     # Copy the default training config but override the max_steps parameter,
     # and reduce the batch_size and buffer_size enough to ensure an update step happens.
     yaml_out = "override.yaml"
-    if python_version:
-        overrides: Any = {"max_steps": 100, "batch_size": 10, "buffer_size": 10}
-        override_legacy_config_file(
-            python_version, "config/trainer_config.yaml", yaml_out, **overrides
-        )
-    else:
-        overrides = {
-            "hyperparameters": {"batch_size": 10, "buffer_size": 10},
-            "max_steps": 100,
-        }
-        override_config_file("config/ppo/3DBall.yaml", yaml_out, overrides)
+    overrides = {
+        "hyperparameters": {"batch_size": 10, "buffer_size": 10},
+        "max_steps": 100,
+    }
+    override_config_file("config/ppo/3DBall.yaml", yaml_out, overrides)
 
     log_output_path = f"{get_base_output_path()}/training.log"
     env_path = os.path.join(get_base_output_path(), standalone_player_path)

ml-agents/tests/yamato/training_int_tests.py

@@ -134,13 +134,10 @@ def init_venv(
     pip_commands = ["--upgrade pip", "--upgrade setuptools"]
     if mlagents_python_version:
         # install from pypi
-        if platform != "darwin":
-            raise RuntimeError("Yamato can only run tensorflow on mac platforms!")
         pip_commands += [
             f"mlagents=={mlagents_python_version}",
             f"gym-unity=={mlagents_python_version}",
             # TODO build these and publish to internal pypi
-            "~/tensorflow_pkg/tensorflow-2.0.0-cp37-cp37m-macosx_10_14_x86_64.whl",
             "tf2onnx==1.6.1",
         ]
     else: