Update mlserver-huggingface to support zero-shot-object-detection task

mlserver/parallel/worker.py

@@ -78,7 +78,14 @@ def _ignore_signals(self):
         To avoid this, and be able to properly shut them down, we forcefully
         ignore the signals coming from the main parent process.
         """
-        loop = asyncio.get_event_loop()
+        # Fix for Python 3.10.10+ / 3.11+ compatibility where asyncio.get_event_loop() 
+        # no longer creates a new loop automatically in child process threads, causing 
+        # RuntimeError when uvloop is active in worker processes.
+        try:
+            loop = asyncio.get_event_loop()
+        except RuntimeError:
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
 
         for sign in IGNORED_SIGNALS:
             # Ensure that signal handlers are a no-op, to let the main process

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "mlserver"
-version = "1.6.1"
+version = "1.6.2"
 description = "MLServer"
 authors = ["Seldon Technologies Ltd. <hello@seldon.io>"]
 license = "Apache-2.0"
@@ -43,7 +43,7 @@ asyncio_mode = "auto"
 addopts = "--import-mode=importlib"
 
 [tool.poetry.dependencies]
-python = ">=3.10,<3.12"
+python = ">=3.11,<3.12"
 click = "*"
 fastapi = ">=0.88.0,!=0.89.0,<=0.110.0"
 python-dotenv = "*"
@@ -68,9 +68,11 @@ pydantic-settings = "2.2.1"
 python-multipart = "*"
 h11 = "0.16.*"
 werkzeug = "^3.0.6"
-keras = "3.*"
+# Replaced keras with tf-keras for transformers compatibility
+tf-keras = ">=2.16.0"
 filelock = "3.20.1"
-
+# Explicitly add setuptools pinned below v80 to preserve pkg_resources
+setuptools = ">=68.0.0,<=80.0.0"
 ## The importlib-resources backport is required to use some
 ## functionality added in Python 3.10
 ## https://setuptools.pypa.io/en/latest/userguide/datafiles.html#accessing-data-files-at-runtime
@@ -88,7 +90,7 @@ tox = "4.14.2"
 
 [tool.poetry.group.dev.dependencies]
 datamodel-code-generator = "0.25.5"
-grpcio-tools = "1.48.2"
+grpcio-tools = "^1.62.0"
 pytest = "7.4.4"
 pytest-asyncio = "0.21.1"
 pytest-mock = "3.12.0"
@@ -107,7 +109,7 @@ conda-pack = "0.7.1"
 flake8 = "7.0.0"
 flake8-black = "0.3.6"
 mypy = "1.10.0"
-mypy-protobuf = "3.1.0"
+mypy-protobuf = ">=3.4.0"
 types-protobuf = "5.26.0.20240422"
 types-orjson = "3.6.2"
 types-aiofiles = "23.2.0.20240106"
@@ -119,7 +121,7 @@ filelock = "^3.13.1"
 httpx-sse = "^0.3.1"
 
 [tool.poetry.group.docker.dependencies]
-protobuf = "3.20.3"
+protobuf = "^4.21.6"
 # tensorflow = "<2.15" #TODO(strive): is this really needed?
 
 [tool.poetry.group.all-runtimes]
@@ -144,10 +146,10 @@ requests-mock = "~1.11.0"
 types-requests = ">=2.28.11.5,<2.32.0.0"
 
 ## Dev dependencies from MLflow
-torch = "2.2.1"
+torch = "2.3.1"
 pytorch-lightning = "2.3.1"
 torchmetrics = "1.4.0.post0"
-torchvision = "0.17.1"
+torchvision = "0.18.1"
 mlflow = "2.13.1"
 
 ## Dev dependencies from HuggingFace

runtimes/huggingface/mlserver_huggingface/metadata.py

@@ -280,6 +280,39 @@
         ],
         outputs=[],
     ),
+    "zero-shot-object-detection": dict(
+        inputs=[
+            # File path inputs (e.g., local image paths or URLs)
+            MetadataTensor(
+                name="array_inputs",
+                shape=[-1],
+                datatype="BYTES",
+                parameters=dict(content_type="str"),
+            ),
+            # File content bytes inputs (e.g., raw base64 image strings)
+            MetadataTensor(
+                name="array_inputs",
+                shape=[-1],
+                datatype="BYTES",
+                parameters=dict(content_type="base64"),
+            ),
+            # Raw pixel matrix inputs (e.g., preprocessed NumPy/Torch tensors)
+            MetadataTensor(
+                name="array_inputs",
+                shape=[-1],
+                datatype="FP32",
+                parameters=dict(content_type="nplist"),
+            ),
+            # Candidate text labels/prompts for object detection
+            MetadataTensor(
+                name="candidate_labels",
+                shape=[-1],
+                datatype="BYTES",
+                parameters=dict(content_type="str"),
+            ),
+        ],
+        outputs=[],
+    ),
     "image-classification": dict(
         inputs=[
             # file path inputs

runtimes/huggingface/pyproject.toml

@@ -1,14 +1,14 @@
 [tool.poetry]
 name = "mlserver-huggingface"
-version = "2.1.0"
+version = "2.1.1"
 description = "HuggingFace runtime for MLServer"
 authors = ["Seldon Technologies Ltd. <hello@seldon.io>"]
 license = "Apache-2.0"
 readme = "README.md"
 packages = [{include = "mlserver_huggingface"}]
 
 [tool.poetry.dependencies]
-python = ">=3.10,<3.12"
+python = ">=3.11,<3.12"
 mlserver = "*"
 tensorflow = "*"
 pillow = "*"