Merge branch 'allegroai:main' into main

.gitignore

@@ -13,6 +13,9 @@ build/
 
 # JetBrains IDE
 .idea/
+.vscode/
+
+tests/huggingface
 
 # Generated by MacOS
 .DS_Store
@@ -20,6 +23,9 @@ build/
 # Generated by Windows
 Thumbs.db
 
+# Virtual environment
+.venv
+
 # Applications
 *.app
 *.exe

clearml_serving/__main__.py

@@ -6,15 +6,22 @@ from pathlib import Path
 
 from clearml.model import Framework
 
-from clearml_serving.serving.model_request_processor import ModelRequestProcessor, CanaryEP
 from clearml_serving.serving.endpoints import ModelMonitoring, ModelEndpoint, EndpointMetricLogging
+from clearml_serving.serving.model_request_processor import ModelRequestProcessor, CanaryEP
+from clearml_serving.version import __version__
+
+# noinspection PyBroadException
+try:
+    from clearml.backend_api import Session
+    Session.add_client(__package__.partition(".")[0].replace("_", "-"), __version__)  # noqa
+except Exception:
+    pass
 
 verbosity = False
 answer_yes = False
 
 
 def verify_session_version(request_processor):
-    from clearml_serving.version import __version__
     current_v = float('.'.join(str(__version__).split(".")[:2]))
     stored_v = float('.'.join(str(request_processor.get_version()).split(".")[:2]))
     if stored_v != current_v:

clearml_serving/engines/triton/requirements.txt

@@ -3,5 +3,5 @@ clearml-serving
 tritonclient[grpc]>=2.32,<2.33
 starlette
 grpcio
-Pillow>=9.0.1,<10
+Pillow>=10.0.1
 pathlib2

clearml_serving/engines/triton/triton_helper.py

@@ -359,6 +359,22 @@ class TritonHelper(object):
         for i, s in enumerate(endpoint.output_name or []):
             config_dict.put("output.{}.name".format(i), "\"{}\"".format(s))
 
+        # check if we have platform in the auxiliary config pbtxt
+        if platform and final_config_pbtxt:
+            # noinspection PyBroadException
+            try:
+                final_config_pbtxt_dict = ConfigFactory.parse_string(final_config_pbtxt)
+                # if we found it, null the requested platform and use the auxiliary config pbtxt platform `value`
+                if final_config_pbtxt_dict.get("platform", None):
+                    print(
+                        "WARNING: ignoring auto-detecetd `platform={}` "
+                        "and using auxiliary pbtxt `platform={}`".format(
+                            str(platform).lower(), final_config_pbtxt_dict.get("platform")))
+                    platform = None
+            except Exception:
+                # we failed parsing the auxiliary pbtxt
+                pass
+
         if platform and not config_dict.get("platform", None) and not config_dict.get("backend", None):
             platform = str(platform).lower()
             if platform.startswith("tensorflow") or platform.startswith("keras"):
@@ -422,10 +438,42 @@ class TritonHelper(object):
             return "FP32"
         elif np_dtype == np.float64:
             return "FP64"
+        elif np_dtype == str:
+            return "STRING"
         elif np_dtype == np.object_ or np_dtype.type == np.bytes_:
             return "BYTES"
         return None
 
+    @staticmethod
+    def triton_to_np_dtype(dtype):
+        if dtype == "BOOL":
+            return bool
+        elif dtype == "INT8":
+            return np.int8
+        elif dtype == "INT16":
+            return np.int16
+        elif dtype == "INT32":
+            return np.int32
+        elif dtype == "INT64":
+            return np.int64
+        elif dtype == "UINT8":
+            return np.uint8
+        elif dtype == "UINT16":
+            return np.uint16
+        elif dtype == "UINT32":
+            return np.uint32
+        elif dtype == "UINT64":
+            return np.uint64
+        elif dtype == "FP16":
+            return np.float16
+        elif dtype == "FP32":
+            return np.float32
+        elif dtype == "FP64":
+            return np.float64
+        elif dtype == "BYTES":
+            return np.object_
+        return None
+
 
 def main():
     title = 'clearml-serving - Nvidia Triton Engine Controller'

clearml_serving/serving/model_request_processor.py

@@ -1292,6 +1292,63 @@ class ModelRequestProcessor(object):
         Raise exception if validation fails, otherwise return True
         """
         if endpoint.engine_type in ("triton", ):
+            if endpoint.auxiliary_cfg:
+                aux_config_dict = {}
+
+                if isinstance(endpoint.auxiliary_cfg, dict):
+                    aux_config_dict = endpoint.auxiliary_cfg
+                elif isinstance(endpoint.auxiliary_cfg, str):
+                    from clearml.utilities.pyhocon import ConfigFactory
+
+                    # noinspection PyBroadException
+                    try:
+                        aux_config_dict = ConfigFactory.parse_string(endpoint.auxiliary_cfg)
+                    except Exception:
+                        # we failed parsing the auxiliary pbtxt
+                        aux_config_dict = {}
+
+                if aux_config_dict.get("input", None) or aux_config_dict.get("output", None):
+                    # noinspection PyBroadException
+                    try:
+                        from ..engines.triton.triton_helper import TritonHelper
+
+                        suggested_cli_in = {"name": [], "dims": [], "data_type": []}
+                        suggested_cli_out = {"name": [], "dims": [], "data_type": []}
+                        for layer in aux_config_dict.get("input", None) or []:
+                            suggested_cli_in["name"] += ['"{}"'.format(layer["name"])]
+                            suggested_cli_in["data_type"] += [
+                                TritonHelper.triton_to_np_dtype(layer["data_type"].replace("TYPE_", "", 1)).__name__]
+                            suggested_cli_in["dims"] += ['"{}"'.format(layer["dims"])]
+
+                        for layer in aux_config_dict.get("output", None) or []:
+                            suggested_cli_out["name"] += ['"{}"'.format(layer["name"])]
+                            suggested_cli_out["data_type"] += [
+                                TritonHelper.triton_to_np_dtype(layer["data_type"].replace("TYPE_", "", 1)).__name__]
+                            suggested_cli_out["dims"] += ['"{}"'.format(layer["dims"])]
+
+                        suggested_cli = "Add to your command line: "\
+                            "--input-name {} --input-type {} --input-size {} " \
+                            "--output-name {} --output-type {} --output-size {} ".format(
+                                " ".join(suggested_cli_in["name"]),
+                                " ".join(suggested_cli_in["data_type"]),
+                                " ".join(suggested_cli_in["dims"]),
+                                " ".join(suggested_cli_out["name"]),
+                                " ".join(suggested_cli_out["data_type"]),
+                                " ".join(suggested_cli_out["dims"]),
+                            )
+                    except Exception:
+                        suggested_cli = "?"
+
+                    raise ValueError(
+                        "Triton engine requires *manual* input/output specification, "
+                        "You input/output in your pbtxt, please remove them and specify manually.\n"
+                        "{}".format(suggested_cli)
+                    )
+
+                if aux_config_dict.get("default_model_filename", None):
+                    raise ValueError("ERROR: You have `default_model_filename` in your config pbtxt, "
+                                     "please remove it. It will be added automatically by the system.")
+
             # verify we have all the info we need
             d = endpoint.as_dict()
             missing = [
@@ -1300,7 +1357,8 @@ class ModelRequestProcessor(object):
                     'output_type', 'output_size', 'output_name',
                 ] if not d.get(k)
             ]
-            if not endpoint.auxiliary_cfg and missing:
+
+            if missing:
                 raise ValueError("Triton engine requires input description - missing values in {}".format(missing))
         return True
 

clearml_serving/serving/preprocess_service.py

@@ -1,6 +1,7 @@
 import os
 import sys
 import threading
+import traceback
 from pathlib import Path
 from typing import Optional, Any, Callable, List
 
@@ -48,8 +49,8 @@ class BasePreprocessRequest(object):
                 try:
                     self._instantiate_custom_preprocess_cls(task)
                 except Exception as ex:
-                    raise ValueError("Error: Failed loading preprocess code for \'{}\': {}".format(
+                    raise ValueError("Error: Failed loading preprocess code for \'{}\': {}\n\n{}".format(
-                        self.model_endpoint.preprocess_artifact, ex))
+                        self.model_endpoint.preprocess_artifact, ex, traceback.format_exc()))
 
     def _instantiate_custom_preprocess_cls(self, task: Task) -> None:
         path = task.artifacts[self.model_endpoint.preprocess_artifact].get_local_copy(extract_archive=False)

clearml_serving/serving/requirements.txt

@@ -11,9 +11,9 @@ numpy>=1.24,<1.25
 scikit-learn>=1.2.2,<1.3
 pandas>=1.5.3,<1.6
 grpcio
-Pillow>=9.0.1,<10
+Pillow>=10.0.1
 xgboost>=1.7.5,<1.8
 lightgbm>=3.3.2,<3.4
-requests>=2.28.2,<2.29
+requests>=2.31.0,<2.29
 kafka-python>=2.0.2,<2.1
 lz4>=4.0.0,<5

clearml_serving/statistics/metrics.py

@@ -240,13 +240,22 @@ class StatisticsController(object):
                 sleep(30)
 
         # we will never leave this loop
-        for message in consumer:
+        while True:
+            # noinspection PyBroadException
+            try:
+                message = next(consumer)
+            except Exception:
+                print("Warning: failed to pull kafka consumer pipe")
+                sleep(5)
+                continue
+
             # noinspection PyBroadException
             try:
                 list_data = json.loads(message.value.decode("utf-8"))
             except Exception:
                 print("Warning: failed to decode kafka stats message")
                 continue
+
             for data in list_data:
                 try:
                     url = data.pop("_url", None)

clearml_serving/statistics/requirements.txt

@@ -1,6 +1,6 @@
 clearml>=1.3.1
 numpy>=1.20,<1.24
-requests>=2.25.1,<2.26
+requests>=2.31.0,<2.26
 kafka-python>=2.0.2,<2.1
 prometheus_client>=0.13.1,<0.14
 lz4>=4.0.0,<5

examples/ensemble/requirements.txt

@@ -1,2 +1,3 @@
 clearml >= 1.1.6
 scikit-learn >= 1.0.2
+numpy>=1.22.2 # not directly required, pinned by Snyk to avoid a vulnerability

examples/huggingface/requirements.txt

@@ -1,2 +1 @@
-clearml-serving
+clearml-serving
-jupyter