sdks/python/apache_beam/ml/inference/gemini_inference.py

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You 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
#
#    http://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.
#

import logging
from collections.abc import Callable
from collections.abc import Iterable
from collections.abc import Sequence
from concurrent.futures import ThreadPoolExecutor
from typing import Any
from typing import Optional
from typing import TypeVar
from typing import Union
from typing import cast

from google import genai
from google.genai import errors
from google.genai.types import HttpOptions
from google.genai.types import Part
from PIL.Image import Image

from apache_beam.ml.inference import utils
from apache_beam.ml.inference.base import PredictionResult
from apache_beam.ml.inference.base import RemoteModelHandler
from apache_beam.ml.inference.base import _validate_batching_args

LOGGER = logging.getLogger("GeminiModelHandler")

# Upper bound on the number of in-flight requests issued concurrently for a
# single batch by the request functions in this module, which send one
# generate_content request per input. Bounds thread usage when adaptive
# batching produces large batches; use max_batch_size to further limit
# per-batch concurrency.
_MAX_CONCURRENT_REQUESTS_PER_BATCH = 32

_InputT = TypeVar('_InputT')
_OutputT = TypeVar('_OutputT')


def _retry_on_appropriate_service_error(exception: Exception) -> bool:
  """
  Retry filter that returns True if a returned HTTP error code is 5xx or 429.
  This is used to retry remote requests that fail, most notably 429
  (throttling by the service)

  Args:
    exception: the returned exception encountered during the request/response
      loop.

  Returns:
    boolean indication whether or not the exception is a ServerError (5xx) or
      a 429 error.
  """
  if not isinstance(exception, errors.APIError):
    return False
  return exception.code == 429 or exception.code >= 500


def _fan_out(fn: Callable[[_InputT], _OutputT],
             items: Sequence[_InputT]) -> list[_OutputT]:
  """Applies fn to each item concurrently and returns results in input order.

  The google-genai client can be shared across threads (its HTTP connection
  pool is thread-safe and credential refreshes are guarded by a lock), so
  per-item requests are dispatched on a bounded thread pool. If any request
  raises, pending requests are cancelled and the first exception (in input
  order) is propagated so Beam's retry logic can handle the batch.
  """
  if len(items) <= 1:
    return [fn(item) for item in items]
  executor = ThreadPoolExecutor(
      max_workers=min(len(items), _MAX_CONCURRENT_REQUESTS_PER_BATCH),
      thread_name_prefix='GeminiModelHandler')
  try:
    return list(executor.map(fn, items))
  finally:
    executor.shutdown(wait=True, cancel_futures=True)


def generate_from_string(
    model_name: str,
    batch: Sequence[str],
    model: genai.Client,
    inference_args: dict[str, Any]):
  """ Request function that expects inputs to be composed of strings, then
  sends requests to Gemini to generate text responses based on the text
  prompts. Each prompt is sent as its own generate_content request (up to
  _MAX_CONCURRENT_REQUESTS_PER_BATCH at a time), because Gemini treats a list
  of strings in a single request as one message and returns one response.

  Args:
    model_name: the Gemini model to use for the request. This model should be
      a text generation model.
    batch: the string inputs to be send to Gemini for text generation.
    model: the genai Client
    inference_args: any additional arguments passed to the generate_content
      call.

  Returns:
    A list with one response per prompt, in the same order as the batch.
  """
  return _fan_out(
      lambda prompt: model.models.generate_content(
          model=model_name, contents=prompt, **inference_args),
      batch)


def generate_image_from_strings_and_images(
    model_name: str,
    batch: Sequence[list[Union[str, Image, Part]]],
    model: genai.Client,
    inference_args: dict[str, Any]):
  """ Request function that expects inputs to be composed of lists of strings
  and PIL Image instances, then sends requests to Gemini to generate images
  based on the text prompts and contextual images. This is currently intended
  to be used with the gemini-2.5-flash-image model (AKA Nano Banana.) Each
  element of the batch is sent as its own generate_content request (up to
  _MAX_CONCURRENT_REQUESTS_PER_BATCH at a time), so every element gets its own
  response.

  Args:
    model_name: the Gemini model to use for the request. This model should be
      an image generation model such as gemini-2.5-flash-image.
    batch: the inputs to be send to Gemini for image generation as prompts.
      Composed of text prompts and contextual pillow Images.
    model: the genai Client
    inference_args: any additional arguments passed to the generate_content
      call.

  Returns:
    A list with one response per element, in the same order as the batch.
  """
  return _fan_out(
      lambda element: model.models.generate_content(
          model=model_name, contents=cast(Any, element), **inference_args),
      batch)


class GeminiModelHandler(RemoteModelHandler[Any, PredictionResult,
                                            genai.Client]):
  def __init__(
      self,
      model_name: str,
      request_fn: Callable[[str, Sequence[Any], genai.Client, dict[str, Any]],
                           Any],
      api_key: Optional[str] = None,
      project: Optional[str] = None,
      location: Optional[str] = None,
      use_vertex_flex_api: Optional[bool] = False,
      *,
      min_batch_size: Optional[int] = None,
      max_batch_size: Optional[int] = None,
      max_batch_duration_secs: Optional[int] = None,
      max_batch_weight: Optional[int] = None,
      element_size_fn: Optional[Callable[[Any], int]] = None,
      batch_length_fn: Optional[Callable[[Any], int]] = None,
      batch_bucket_boundaries: Optional[list[int]] = None,
      **kwargs):
    """Implementation of the ModelHandler interface for Google Gemini.
    **NOTE:** This API and its implementation are under development and
    do not provide backward compatibility guarantees.
    Gemini can be accessed through either the Vertex AI API or the Gemini
    Developer API, and this handler chooses which to connect to based upon
    the arguments provided. As written, this model handler operates solely on
    string input.

    Args:
      model_name: the Gemini model to send the request to
      request_fn: the function to use to send the request. Should take the
        model name and the parameters from request() and return a list with
        one response from Gemini per input, in the same order as the inputs.
        The class will handle bundling the inputs and responses together.
      api_key: the Gemini Developer API key to use for the requests. Setting
        this parameter sends requests for this job to the Gemini Developer API.
        If this paramter is provided, do not set the project or location
        parameters.
      project: the GCP project to use for Vertex AI requests. Setting this
        parameter routes requests to Vertex AI. If this paramter is provided,
        location must also be provided and api_key should not be set.
      location: the GCP project to use for Vertex AI requests. Setting this
        parameter routes requests to Vertex AI. If this paramter is provided,
        project must also be provided and api_key should not be set.
      use_vertex_flex_api: if true, use the Vertex Flex API. This is a
        cost-effective option for accessing Gemini models if you can tolerate
        longer response times and throttling. This is often beneficial for
        data processing workloads which usually have higher latency tolerance
        than live serving paths. See
        https://docs.cloud.google.com/vertex-ai/generative-ai/docs/flex-paygo
        for more details.
      min_batch_size: optional. the minimum batch size to use when batching
        inputs.
      max_batch_size: optional. the maximum batch size to use when batching
        inputs.
      max_batch_duration_secs: optional. the maximum amount of time to buffer
        a batch before emitting; used in streaming contexts.
      max_batch_weight: optional. the maximum total weight of a batch. Cannot
        be combined with batch_length_fn or batch_bucket_boundaries.
      element_size_fn: optional. a function that returns the size (weight)
        of an element.
      batch_length_fn: optional. a callable that returns the length of an
        element for length-aware batching. Cannot be used together with
        max_batch_weight.
      batch_bucket_boundaries: optional. a sorted list of positive boundary
        values for length-aware batching buckets. Cannot be used together with
        max_batch_weight.
    """
    _validate_batching_args(
        max_batch_weight=max_batch_weight,
        element_size_fn=element_size_fn,
        batch_length_fn=batch_length_fn,
        batch_bucket_boundaries=batch_bucket_boundaries)
    self._batching_kwargs = {}
    self._env_vars = kwargs.get('env_vars', {})
    if min_batch_size is not None:
      self._batching_kwargs["min_batch_size"] = min_batch_size
    if max_batch_size is not None:
      self._batching_kwargs["max_batch_size"] = max_batch_size
    if max_batch_duration_secs is not None:
      self._batching_kwargs["max_batch_duration_secs"] = max_batch_duration_secs
    if max_batch_weight is not None:
      self._batching_kwargs["max_batch_weight"] = max_batch_weight
    if element_size_fn is not None:
      self._batching_kwargs['element_size_fn'] = element_size_fn
    if batch_length_fn is not None:
      self._batching_kwargs['length_fn'] = batch_length_fn
    if batch_bucket_boundaries is not None:
      self._batching_kwargs['bucket_boundaries'] = batch_bucket_boundaries

    self.model_name = model_name
    self.request_fn = request_fn

    if api_key:
      if project or location:
        raise ValueError("project and location must be None if api_key is set")
      self.api_key = api_key
      self.use_vertex = False
    else:
      if project is None or location is None:
        raise ValueError(
            "project and location must both be provided if api_key is None")
      self.project = project
      self.location = location
      self.use_vertex = True

    self.use_vertex_flex_api = use_vertex_flex_api

    super().__init__(
        namespace='GeminiModelHandler',
        retry_filter=_retry_on_appropriate_service_error,
        **kwargs)

  def batch_elements_kwargs(self):
    return self._batching_kwargs

  def create_client(self) -> genai.Client:
    """Creates the GenAI client used to send requests. Creates a version for
    the Vertex AI API or the Gemini Developer API based on the arguments
    provided when the GeminiModelHandler class is instantiated.
    """
    if self.use_vertex:
      if self.use_vertex_flex_api:
        return genai.Client(
            vertexai=True,
            project=self.project,
            location=self.location,
            http_options=HttpOptions(
                api_version="v1",
                headers={"X-Vertex-AI-LLM-Request-Type": "flex"},
                # Set timeout in the unit of millisecond.
                timeout=600000))
      else:
        return genai.Client(
            vertexai=True, project=self.project, location=self.location)
    return genai.Client(api_key=self.api_key)

  def request(
      self,
      batch: Sequence[Any],
      model: genai.Client,
      inference_args: Optional[dict[str, Any]] = None
  ) -> Iterable[PredictionResult]:
    """ Sends a batch of inputs to a Gemini service using the request function
    and matches each input with its response as an iterable of
    PredictionResults.

    Args:
      batch: a sequence of any values to be passed to the Gemini service.
        Should be inputs accepted by the provided inference function.
      model: a genai.Client object configured to access the desired service.
      inference_args: any additional arguments to send as part of the
        prediction request.

    Returns:
      An iterable of Predictions, one per input.
    """
    if inference_args is None:
      inference_args = {}
    responses = self.request_fn(self.model_name, batch, model, inference_args)
    if len(batch) == 1 and not isinstance(responses, (list, tuple)):
      # Request functions written for single-input batches may return the
      # response itself rather than a list of responses.
      responses = [responses]
    if not isinstance(responses, (list, tuple)):
      raise ValueError(
          'request_fn must return a list with one response per input, but '
          f'returned a single {type(responses).__name__} for a batch of '
          f'{len(batch)} inputs.')
    if len(responses) != len(batch):
      raise ValueError(
          'request_fn must return one response per input, but returned '
          f'{len(responses)} responses for a batch of {len(batch)} inputs.')
    return utils._convert_to_result(batch, responses, self.model_name)