Polber commented on code in PR #33406:
URL: https://github.com/apache/beam/pull/33406#discussion_r1894313682
##########
sdks/python/apache_beam/yaml/yaml_ml.py:
##########
@@ -33,11 +40,411 @@
tft = None # type: ignore
+class ModelHandlerProvider:
+ handler_types: Dict[str, Callable[..., "ModelHandlerProvider"]] = {}
+
+ def __init__(
+ self,
+ handler,
+ preprocess: Optional[Dict[str, str]] = None,
+ postprocess: Optional[Dict[str, str]] = None):
+ self._handler = handler
+ self._preprocess_fn = self.parse_processing_transform(
+ preprocess, 'preprocess') or self.default_preprocess_fn()
+ self._postprocess_fn = self.parse_processing_transform(
+ postprocess, 'postprocess') or self.default_postprocess_fn()
+
+ def inference_output_type(self):
+ return Any
+
+ @staticmethod
+ def parse_processing_transform(processing_transform, typ):
+ def _parse_config(callable=None, path=None, name=None):
+ if callable and (path or name):
+ raise ValueError(
+ f"Cannot specify 'callable' with 'path' and 'name' for {typ} "
+ f"function.")
+ if path and name:
+ return python_callable.PythonCallableWithSource.load_from_script(
+ FileSystems.open(path).read().decode(), name)
+ elif callable:
+ return python_callable.PythonCallableWithSource(callable)
+ else:
+ raise ValueError(
+ f"Must specify one of 'callable' or 'path' and 'name' for {typ} "
+ f"function.")
+
+ if processing_transform:
+ if isinstance(processing_transform, dict):
+ return _parse_config(**processing_transform)
+ else:
+ raise ValueError("Invalid model_handler specification.")
+
+ def underlying_handler(self):
+ return self._handler
+
+ @staticmethod
+ def default_preprocess_fn():
+ raise ValueError(
+ 'Handler does not implement a default preprocess '
+ 'method. Please define a preprocessing method using the '
+ '\'preprocess\' tag.')
+
+ def _preprocess_fn_internal(self):
+ return lambda row: (row, self._preprocess_fn(row))
+
+ @staticmethod
+ def default_postprocess_fn():
+ return lambda x: x
+
+ def _postprocess_fn_internal(self):
+ return lambda result: (result[0], self._postprocess_fn(result[1]))
+
+ @staticmethod
+ def validate(model_handler_spec):
+ raise NotImplementedError(type(ModelHandlerProvider))
+
+ @classmethod
+ def register_handler_type(cls, type_name):
+ def apply(constructor):
+ cls.handler_types[type_name] = constructor
+ return constructor
+
+ return apply
+
+ @classmethod
+ def create_handler(cls, model_handler_spec) -> "ModelHandlerProvider":
+ typ = model_handler_spec['type']
+ config = model_handler_spec['config']
+ try:
+ result = cls.handler_types[typ](**config)
+ if not hasattr(result, 'to_json'):
+ result.to_json = lambda: model_handler_spec
+ return result
+ except Exception as exn:
+ raise ValueError(
+ f'Unable to instantiate model handler of type {typ}. {exn}')
+
+
[email protected]_handler_type('VertexAIModelHandlerJSON')
+class VertexAIModelHandlerJSONProvider(ModelHandlerProvider):
+ def __init__(
+ self,
+ endpoint_id: str,
+ project: str,
+ location: str,
+ preprocess: Dict[str, str],
+ experiment: Optional[str] = None,
+ network: Optional[str] = None,
+ private: bool = False,
+ min_batch_size: Optional[int] = None,
+ max_batch_size: Optional[int] = None,
+ max_batch_duration_secs: Optional[int] = None,
+ env_vars: Optional[Dict[str, Any]] = None,
+ postprocess: Optional[Dict[str, str]] = None):
+ """
+ ModelHandler for Vertex AI.
+
+ For example: ::
+
+ - type: RunInference
+ config:
+ inference_tag: 'my_inference'
+ model_handler:
+ type: VertexAIModelHandlerJSON
+ config:
+ endpoint_id: 9876543210
+ project: my-project
+ location: us-east1
+ preprocess:
+ callable: 'lambda x: {"prompt": x.prompt, "max_tokens": 50}'
+
+ Args:
+ endpoint_id: the numerical ID of the Vertex AI endpoint to query.
+ project: the GCP project name where the endpoint is deployed.
+ location: the GCP location where the endpoint is deployed.
+ experiment: Experiment label to apply to the
+ queries. See
+
https://cloud.google.com/vertex-ai/docs/experiments/intro-vertex-ai-experiments
+ for more information.
+ network: The full name of the Compute Engine
+ network the endpoint is deployed on; used for private
+ endpoints. The network or subnetwork Dataflow pipeline
+ option must be set and match this network for pipeline
+ execution.
+ Ex: "projects/12345/global/networks/myVPC"
+ private: If the deployed Vertex AI endpoint is
+ private, set to true. Requires a network to be provided
+ as well.
+ min_batch_size: The minimum batch size to use when batching
+ inputs.
+ max_batch_size: The maximum batch size to use when batching
+ inputs.
+ max_batch_duration_secs: The maximum amount of time to buffer
+ a batch before emitting; used in streaming contexts.
+ env_vars: Environment variables.
+ preprocess: A python callable, defined either inline, or using a file,
Review Comment:
Ah yes, that was the order before I made it required, thanks!
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
