zenml-nightly 0.64.0.dev20240809__py3-none-any.whl → 0.65.0.dev20240906__py3-none-any.whl

zenml/integrations/feast/__init__.py

@@ -17,9 +17,8 @@ The Feast integration offers a way to connect to a Feast Feature Store. ZenML
 implements a dedicated stack component that you can access as part of your ZenML
 steps in the usual ways.
 """
-from typing import List, Type
+from typing import List, Type, Optional
 
-from zenml.enums import StackComponentType
 from zenml.integrations.constants import FEAST
 from zenml.integrations.integration import Integration
 from zenml.stack import Flavor
@@ -31,10 +30,9 @@ class FeastIntegration(Integration):
     """Definition of Feast integration for ZenML."""
 
     NAME = FEAST
-    # click is added to keep the feast click version
-    # in sync with ZenML's click
+    # click is added to keep the feast click version in sync with ZenML's click
     REQUIREMENTS = ["feast<=0.37.1", "click>=8.0.1,<8.1.4"]
-    REQUIREMENTS_IGNORED_ON_UNINSTALL = ["click"]
+    REQUIREMENTS_IGNORED_ON_UNINSTALL = ["click", "pandas"]
 
     @classmethod
     def flavors(cls) -> List[Type[Flavor]]:
@@ -47,5 +45,20 @@ class FeastIntegration(Integration):
 
         return [FeastFeatureStoreFlavor]
 
+    @classmethod
+    def get_requirements(cls, target_os: Optional[str] = None) -> List[str]:
+        """Method to get the requirements for the integration.
+
+        Args:
+            target_os: The target operating system to get the requirements for.
+
+        Returns:
+            A list of requirements.
+        """
+        from zenml.integrations.pandas import PandasIntegration
+
+        return cls.REQUIREMENTS + \
+            PandasIntegration.get_requirements(target_os=target_os)
+
 
 FeastIntegration.check_installation()

zenml/integrations/gcp/__init__.py

@@ -55,8 +55,8 @@ class GcpIntegration(Integration):
     ]
     REQUIREMENTS_IGNORED_ON_UNINSTALL = ["kubernetes","kfp"]
 
-    @staticmethod
-    def activate() -> None:
+    @classmethod
+    def activate(cls) -> None:
         """Activate the GCP integration."""
         from zenml.integrations.gcp import service_connectors  # noqa
 

zenml/integrations/gcp/flavors/vertex_orchestrator_flavor.py

@@ -162,6 +162,15 @@ class VertexOrchestratorConfig(
         """
         return self.synchronous
 
+    @property
+    def is_schedulable(self) -> bool:
+        """Whether the orchestrator is schedulable or not.
+
+        Returns:
+            Whether the orchestrator is schedulable or not.
+        """
+        return True
+
 
 class VertexOrchestratorFlavor(BaseOrchestratorFlavor):
     """Vertex Orchestrator flavor."""

zenml/integrations/gcp/orchestrators/vertex_orchestrator.py

@@ -32,7 +32,16 @@
 import os
 import re
 import types
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type, cast
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Dict,
+    List,
+    Optional,
+    Tuple,
+    Type,
+    cast,
+)
 from uuid import UUID
 
 from google.api_core import exceptions as google_exceptions

zenml/integrations/great_expectations/__init__.py

@@ -17,9 +17,8 @@ The Great Expectations integration enables you to use Great Expectations as a
 way of profiling and validating your data.
 """
 
-from typing import List, Type
+from typing import List, Type, Optional
 
-from zenml.enums import StackComponentType
 from zenml.integrations.constants import GREAT_EXPECTATIONS
 from zenml.integrations.integration import Integration
 from zenml.stack import Flavor
@@ -31,12 +30,12 @@ class GreatExpectationsIntegration(Integration):
     """Definition of Great Expectations integration for ZenML."""
 
     NAME = GREAT_EXPECTATIONS
-    REQUIREMENTS = [
-        "great-expectations>=0.17.15,<1.0",
-    ]
+    REQUIREMENTS = ["great-expectations>=0.17.15,<1.0"]
 
-    @staticmethod
-    def activate() -> None:
+    REQUIREMENTS_IGNORED_ON_UNINSTALL = ["pandas"]
+
+    @classmethod
+    def activate(cls) -> None:
         """Activate the Great Expectations integration."""
         from zenml.integrations.great_expectations import materializers  # noqa
 
@@ -53,5 +52,20 @@ class GreatExpectationsIntegration(Integration):
 
         return [GreatExpectationsDataValidatorFlavor]
 
+    @classmethod
+    def get_requirements(cls, target_os: Optional[str] = None) -> List[str]:
+        """Method to get the requirements for the integration.
+
+        Args:
+            target_os: The target operating system to get the requirements for.
+
+        Returns:
+            A list of requirements.
+        """
+        from zenml.integrations.pandas import PandasIntegration
+
+        return cls.REQUIREMENTS + \
+            PandasIntegration.get_requirements(target_os=target_os)
+
 
 GreatExpectationsIntegration.check_installation()

zenml/integrations/huggingface/__init__.py

@@ -12,7 +12,8 @@
 #  or implied. See the License for the specific language governing
 #  permissions and limitations under the License.
 """Initialization of the Huggingface integration."""
-from typing import List, Type
+import sys
+from typing import List, Type, Optional
 
 from zenml.integrations.constants import HUGGINGFACE
 from zenml.integrations.integration import Integration
@@ -26,20 +27,8 @@ class HuggingfaceIntegration(Integration):
     """Definition of Huggingface integration for ZenML."""
 
     NAME = HUGGINGFACE
-    REQUIREMENTS = [
-        "transformers<=4.31",
-        "datasets",
-        "huggingface_hub>0.19.0",
-        "accelerate",
-        "bitsandbytes>=0.41.3",
-        "peft",
-        # temporary fix for CI issue similar to:
-        # - https://github.com/huggingface/datasets/issues/6737
-        # - https://github.com/huggingface/datasets/issues/6697
-        # TODO try relaxing it back going forward
-        "fsspec<=2023.12.0",
-    ]
-    REQUIREMENTS_IGNORED_ON_UNINSTALL = ["fsspec"]
+
+    REQUIREMENTS_IGNORED_ON_UNINSTALL = ["fsspec", "pandas"]
 
     @classmethod
     def activate(cls) -> None:
@@ -47,6 +36,41 @@ class HuggingfaceIntegration(Integration):
         from zenml.integrations.huggingface import materializers  # noqa
         from zenml.integrations.huggingface import services
 
+    @classmethod
+    def get_requirements(cls, target_os: Optional[str] = None) -> List[str]:
+        """Defines platform specific requirements for the integration.
+
+        Args:
+            target_os: The target operating system.
+
+        Returns:
+            A list of requirements.
+        """
+        requirements = [
+            "datasets",
+            "huggingface_hub>0.19.0",
+            "accelerate",
+            "bitsandbytes>=0.41.3",
+            "peft",
+            # temporary fix for CI issue similar to:
+            # - https://github.com/huggingface/datasets/issues/6737
+            # - https://github.com/huggingface/datasets/issues/6697
+            # TODO try relaxing it back going forward
+            "fsspec<=2023.12.0",
+        ]
+
+        # In python 3.8 higher transformers version lead to other packages breaking
+        if sys.version_info.minor > 8:
+            requirements += ["transformers"]
+        else:
+            requirements += ["transformers<=4.31"]
+
+        # Add the pandas integration requirements
+        from zenml.integrations.pandas import PandasIntegration
+
+        return requirements + \
+            PandasIntegration.get_requirements(target_os=target_os)
+
     @classmethod
     def flavors(cls) -> List[Type[Flavor]]:
         """Declare the stack component flavors for the Huggingface integration.

zenml/integrations/huggingface/materializers/__init__.py

@@ -25,3 +25,6 @@ from zenml.integrations.huggingface.materializers.huggingface_tf_model_materiali
 from zenml.integrations.huggingface.materializers.huggingface_tokenizer_materializer import (
     HFTokenizerMaterializer,
 )
+from zenml.integrations.huggingface.materializers.huggingface_t5_materializer import (
+    HFT5Materializer,
+)

zenml/integrations/huggingface/materializers/huggingface_datasets_materializer.py

@@ -31,9 +31,11 @@ from datasets import Dataset, load_from_disk
 from datasets.dataset_dict import DatasetDict
 
 from zenml.enums import ArtifactType, VisualizationType
+from zenml.integrations.pandas.materializers.pandas_materializer import (
+    PandasMaterializer,
+)
 from zenml.io import fileio
 from zenml.materializers.base_materializer import BaseMaterializer
-from zenml.materializers.pandas_materializer import PandasMaterializer
 from zenml.utils import io_utils
 
 if TYPE_CHECKING:

zenml/integrations/huggingface/materializers/huggingface_pt_model_materializer.py

@@ -18,7 +18,7 @@ import os
 from tempfile import TemporaryDirectory
 from typing import Any, ClassVar, Dict, Tuple, Type
 
-from transformers import (  # type: ignore [import-untyped]
+from transformers import (
     AutoConfig,
     PreTrainedModel,
 )

zenml/integrations/huggingface/materializers/huggingface_t5_materializer.py

@@ -0,0 +1,107 @@
+#  Copyright (c) ZenML GmbH 2024. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at:
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+"""Implementation of the Huggingface t5 materializer."""
+
+import os
+import tempfile
+from typing import Any, ClassVar, Type, Union
+
+from transformers import (
+    T5ForConditionalGeneration,
+    T5Tokenizer,
+    T5TokenizerFast,
+)
+
+from zenml.io import fileio
+from zenml.materializers.base_materializer import BaseMaterializer
+
+
+class HFT5Materializer(BaseMaterializer):
+    """Base class for huggingface t5 models."""
+
+    SKIP_REGISTRATION: ClassVar[bool] = False
+    ASSOCIATED_TYPES = (
+        T5ForConditionalGeneration,
+        T5Tokenizer,
+        T5TokenizerFast,
+    )
+
+    def load(
+        self, data_type: Type[Any]
+    ) -> Union[T5ForConditionalGeneration, T5Tokenizer, T5TokenizerFast]:
+        """Reads a T5ForConditionalGeneration model or T5Tokenizer from a serialized zip file.
+
+        Args:
+            data_type: A T5ForConditionalGeneration or T5Tokenizer type.
+
+        Returns:
+            A T5ForConditionalGeneration or T5Tokenizer object.
+
+        Raises:
+            ValueError: Unsupported data type used
+        """
+        filepath = self.uri
+
+        with tempfile.TemporaryDirectory(prefix="zenml-temp-") as temp_dir:
+            # Copy files from artifact store to temporary directory
+            for file in fileio.listdir(filepath):
+                src = os.path.join(filepath, file)
+                dst = os.path.join(temp_dir, file)
+                if fileio.isdir(src):
+                    fileio.makedirs(dst)
+                    for subfile in fileio.listdir(src):
+                        subsrc = os.path.join(src, subfile)
+                        subdst = os.path.join(dst, subfile)
+                        fileio.copy(subsrc, subdst)
+                else:
+                    fileio.copy(src, dst)
+
+            # Load the model or tokenizer from the temporary directory
+            if data_type in [
+                T5ForConditionalGeneration,
+                T5Tokenizer,
+                T5TokenizerFast,
+            ]:
+                return data_type.from_pretrained(temp_dir)
+            else:
+                raise ValueError(f"Unsupported data type: {data_type}")
+
+    def save(
+        self,
+        obj: Union[T5ForConditionalGeneration, T5Tokenizer, T5TokenizerFast],
+    ) -> None:
+        """Creates a serialization for a T5ForConditionalGeneration model or T5Tokenizer.
+
+        Args:
+            obj: A T5ForConditionalGeneration model or T5Tokenizer.
+        """
+        # Create a temporary directory
+        with tempfile.TemporaryDirectory(prefix="zenml-temp-") as temp_dir:
+            # Save the model or tokenizer
+            obj.save_pretrained(temp_dir)
+
+            # Copy the directory to the artifact store
+            filepath = self.uri
+            fileio.makedirs(filepath)
+            for file in os.listdir(temp_dir):
+                src = os.path.join(temp_dir, file)
+                dst = os.path.join(filepath, file)
+                if os.path.isdir(src):
+                    fileio.makedirs(dst)
+                    for subfile in os.listdir(src):
+                        subsrc = os.path.join(src, subfile)
+                        subdst = os.path.join(dst, subfile)
+                        fileio.copy(subsrc, subdst)
+                else:
+                    fileio.copy(src, dst)

zenml/integrations/huggingface/materializers/huggingface_tf_model_materializer.py

@@ -18,7 +18,7 @@ import os
 from tempfile import TemporaryDirectory
 from typing import Any, ClassVar, Dict, Tuple, Type
 
-from transformers import (  # type: ignore [import-untyped]
+from transformers import (
     AutoConfig,
     TFPreTrainedModel,
 )

zenml/integrations/huggingface/materializers/huggingface_tokenizer_materializer.py

@@ -17,8 +17,8 @@ import os
 from tempfile import TemporaryDirectory
 from typing import Any, ClassVar, Tuple, Type
 
-from transformers import AutoTokenizer  # type: ignore [import-untyped]
-from transformers.tokenization_utils_base import (  # type: ignore [import-untyped]
+from transformers import AutoTokenizer
+from transformers.tokenization_utils_base import (
     PreTrainedTokenizerBase,
 )
 

zenml/integrations/huggingface/steps/accelerate_runner.py

@@ -17,14 +17,15 @@
 """Step function to run any ZenML step using Accelerate."""
 
 import functools
-from typing import Any, Callable, Optional, TypeVar, cast
+from typing import Any, Callable, Dict, Optional, TypeVar, Union, cast
 
 import cloudpickle as pickle
-from accelerate.commands.launch import (  # type: ignore[import-untyped]
+from accelerate.commands.launch import (
     launch_command,
     launch_command_parser,
 )
 
+from zenml import get_pipeline_context
 from zenml.logger import get_logger
 from zenml.steps import BaseStep
 from zenml.utils.function_utils import _cli_arg_name, create_cli_wrapped_script
@@ -34,10 +35,9 @@ F = TypeVar("F", bound=Callable[..., Any])
 
 
 def run_with_accelerate(
-    step_function: BaseStep,
-    num_processes: Optional[int] = None,
-    use_cpu: bool = False,
-) -> BaseStep:
+    step_function_top_level: Optional[BaseStep] = None,
+    **accelerate_launch_kwargs: Any,
+) -> Union[Callable[[BaseStep], BaseStep], BaseStep]:
     """Run a function with accelerate.
 
     Accelerate package: https://huggingface.co/docs/accelerate/en/index
@@ -45,6 +45,8 @@ def run_with_accelerate(
         ```python
         from zenml import step, pipeline
         from zenml.integrations.hugginface.steps import run_with_accelerate
+
+        @run_with_accelerate(num_processes=4, multi_gpu=True)
         @step
         def training_step(some_param: int, ...):
             # your training code is below
@@ -52,92 +54,113 @@ def run_with_accelerate(
 
         @pipeline
         def training_pipeline(some_param: int, ...):
-            run_with_accelerate(training_step, num_processes=4)(some_param, ...)
+            training_step(some_param, ...)
         ```
 
     Args:
-        step_function: The step function to run.
-        num_processes: The number of processes to use.
-        use_cpu: Whether to use the CPU.
+        step_function_top_level: The step function to run with accelerate [optional].
+            Used in functional calls like `run_with_accelerate(some_func,foo=bar)()`.
+        accelerate_launch_kwargs: A dictionary of arguments to pass along to the
+            `accelerate launch` command, including hardware selection, resource
+            allocation, and training paradigm options. Visit
+            https://huggingface.co/docs/accelerate/en/package_reference/cli#accelerate-launch
+            for more details.
 
     Returns:
         The accelerate-enabled version of the step.
     """
 
-    def _decorator(entrypoint: F) -> F:
-        @functools.wraps(entrypoint)
-        def inner(*args: Any, **kwargs: Any) -> Any:
-            if args:
-                raise ValueError(
-                    "Accelerated steps do not support positional arguments."
-                )
-
-            if not use_cpu:
-                import torch
-
-                logger.info("Starting accelerate job...")
-
-                device_count = torch.cuda.device_count()
-                if num_processes is None:
-                    _num_processes = device_count
-                else:
-                    if num_processes > device_count:
-                        logger.warning(
-                            f"Number of processes ({num_processes}) is greater than "
-                            f"the number of available GPUs ({device_count}). Using all GPUs."
-                        )
-                        _num_processes = device_count
-                    else:
-                        _num_processes = num_processes
-            else:
-                _num_processes = num_processes or 1
-
-            with create_cli_wrapped_script(
-                entrypoint, flavour="accelerate"
-            ) as (
-                script_path,
-                output_path,
-            ):
-                commands = ["--num_processes", str(_num_processes)]
-                if use_cpu:
-                    commands += [
-                        "--cpu",
-                        "--num_cpu_threads_per_process",
-                        "10",
-                    ]
-                commands.append(str(script_path.absolute()))
-                for k, v in kwargs.items():
-                    k = _cli_arg_name(k)
-                    if isinstance(v, bool):
-                        if v:
-                            commands.append(f"--{k}")
-                    elif type(v) in (list, tuple, set):
-                        for each in v:
-                            commands += [f"--{k}", f"{each}"]
-                    else:
-                        commands += [f"--{k}", f"{v}"]
-
-                logger.debug(commands)
-
-                parser = launch_command_parser()
-                args = parser.parse_args(commands)
-                try:
-                    launch_command(args)
-                except Exception as e:
-                    logger.error(
-                        "Accelerate training job failed... See error message for details."
+    def _decorator(step_function: BaseStep) -> BaseStep:
+        def _wrapper(
+            entrypoint: F, accelerate_launch_kwargs: Dict[str, Any]
+        ) -> F:
+            @functools.wraps(entrypoint)
+            def inner(*args: Any, **kwargs: Any) -> Any:
+                if args:
+                    raise ValueError(
+                        "Accelerated steps do not support positional arguments."
                     )
-                    raise RuntimeError(
-                        "Accelerate training job failed."
-                    ) from e
-                else:
-                    logger.info(
-                        "Accelerate training job finished successfully."
-                    )
-                    return pickle.load(open(output_path, "rb"))
-
-        return cast(F, inner)
-
-    setattr(step_function, "entrypoint", _decorator(step_function.entrypoint))
 
-    return step_function
+                with create_cli_wrapped_script(
+                    entrypoint, flavor="accelerate"
+                ) as (
+                    script_path,
+                    output_path,
+                ):
+                    commands = [str(script_path.absolute())]
+                    for k, v in kwargs.items():
+                        k = _cli_arg_name(k)
+                        if isinstance(v, bool):
+                            if v:
+                                commands.append(f"--{k}")
+                        elif type(v) in (list, tuple, set):
+                            for each in v:
+                                commands += [f"--{k}", f"{each}"]
+                        else:
+                            commands += [f"--{k}", f"{v}"]
+                    logger.debug(commands)
+
+                    parser = launch_command_parser()
+                    args = parser.parse_args(commands)
+                    for k, v in accelerate_launch_kwargs.items():
+                        if k in args:
+                            setattr(args, k, v)
+                        else:
+                            logger.warning(
+                                f"You passed in `{k}` as an `accelerate launch` argument, but it was not accepted. "
+                                "Please check https://huggingface.co/docs/accelerate/en/package_reference/cli#accelerate-launch "
+                                "to find out more about supported arguments and retry."
+                            )
+                    try:
+                        launch_command(args)
+                    except Exception as e:
+                        logger.error(
+                            "Accelerate training job failed... See error message for details."
+                        )
+                        raise RuntimeError(
+                            "Accelerate training job failed."
+                        ) from e
+                    else:
+                        logger.info(
+                            "Accelerate training job finished successfully."
+                        )
+                        return pickle.load(open(output_path, "rb"))
+
+            return cast(F, inner)
+
+        try:
+            get_pipeline_context()
+        except RuntimeError:
+            pass
+        else:
+            raise RuntimeError(
+                f"`{run_with_accelerate.__name__}` decorator cannot be used "
+                "in a functional way with steps, please apply decoration "
+                "directly to a step instead. This behavior will be also "
+                "allowed in future, but now it faces technical limitations.\n"
+                "Example (allowed):\n"
+                f"@{run_with_accelerate.__name__}(...)\n"
+                f"def {step_function.name}(...):\n"
+                "    ...\n"
+                "Example (not allowed):\n"
+                "def my_pipeline(...):\n"
+                f"    run_with_accelerate({step_function.name},...)(...)\n"
+            )
+
+        setattr(
+            step_function, "unwrapped_entrypoint", step_function.entrypoint
+        )
+        setattr(
+            step_function,
+            "entrypoint",
+            _wrapper(
+                step_function.entrypoint,
+                accelerate_launch_kwargs=accelerate_launch_kwargs,
+            ),
+        )
+
+        return step_function
+
+    if step_function_top_level:
+        return _decorator(step_function_top_level)
+    return _decorator

zenml/integrations/hyperai/flavors/hyperai_orchestrator_flavor.py

@@ -81,6 +81,15 @@ class HyperAIOrchestratorConfig(
         """
         return True
 
+    @property
+    def is_schedulable(self) -> bool:
+        """Whether the orchestrator is schedulable or not.
+
+        Returns:
+            Whether the orchestrator is schedulable or not.
+        """
+        return True
+
 
 class HyperAIOrchestratorFlavor(BaseOrchestratorFlavor):
     """Flavor for the HyperAI orchestrator."""

zenml/integrations/kubeflow/flavors/kubeflow_orchestrator_flavor.py

@@ -213,6 +213,15 @@ class KubeflowOrchestratorConfig(
         """
         return self.synchronous
 
+    @property
+    def is_schedulable(self) -> bool:
+        """Whether the orchestrator is schedulable or not.
+
+        Returns:
+            Whether the orchestrator is schedulable or not.
+        """
+        return True
+
 
 class KubeflowOrchestratorFlavor(BaseOrchestratorFlavor):
     """Kubeflow orchestrator flavor."""

zenml/integrations/kubeflow/orchestrators/kubeflow_orchestrator.py

@@ -32,7 +32,16 @@
 
 import os
 import types
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type, cast
+from typing import (
+    TYPE_CHECKING,
+    Any,
+    Dict,
+    List,
+    Optional,
+    Tuple,
+    Type,
+    cast,
+)
 from uuid import UUID
 
 import kfp

zenml/integrations/kubernetes/flavors/kubernetes_orchestrator_flavor.py

@@ -125,6 +125,15 @@ class KubernetesOrchestratorConfig(
         """
         return self.synchronous
 
+    @property
+    def is_schedulable(self) -> bool:
+        """Whether the orchestrator is schedulable or not.
+
+        Returns:
+            Whether the orchestrator is schedulable or not.
+        """
+        return True
+
 
 class KubernetesOrchestratorFlavor(BaseOrchestratorFlavor):
     """Kubernetes orchestrator flavor."""