# Note this class will get deprecated in the future without notice
# Use python/gigl/src/inference/inferencer.py instead
import argparse
import concurrent.futures
import sys
import threading
import traceback
from dataclasses import dataclass
from typing import Any, Dict, List, Optional
from apache_beam.runners.dataflow.dataflow_runner import DataflowPipelineResult
from apache_beam.runners.runner import PipelineState
from google.cloud import bigquery
from gigl.common import GcsUri, Uri, UriFactory
from gigl.common.env_config import get_available_cpus
from gigl.common.logger import Logger
from gigl.common.metrics.decorators import flushes_metrics, profileit
from gigl.common.utils import os_utils
from gigl.env.pipelines_config import get_resource_config
from gigl.src.common.constants.metrics import TIMER_INFERENCER_S
from gigl.src.common.graph_builder.graph_builder_factory import GraphBuilderFactory
from gigl.src.common.types import AppliedTaskIdentifier
from gigl.src.common.types.graph_data import NodeType
from gigl.src.common.types.pb_wrappers.gbml_config import GbmlConfigPbWrapper
from gigl.src.common.utils.bq import BqUtils
from gigl.src.common.utils.metrics_service_provider import (
get_metrics_service_instance,
initialize_metrics,
)
from gigl.src.common.utils.model import load_state_dict_from_uri
from gigl.src.inference.lib.assets import InferenceAssets
from gigl.src.inference.v1.lib.base_inference_blueprint import BaseInferenceBlueprint
from gigl.src.inference.v1.lib.base_inferencer import BaseInferencer
from gigl.src.inference.v1.lib.inference_blueprint_factory import (
InferenceBlueprintFactory,
)
from gigl.src.inference.v1.lib.utils import (
get_inferencer_pipeline_component_for_single_node_type,
)
from snapchat.research.gbml.inference_metadata_pb2 import InferenceOutput
[docs]
MAX_INFERENCER_NUM_WORKERS = 4
@dataclass
[docs]
class InferencerOutputPaths:
"""
Dataclass containing the output path fields from running inference for a single node type. These fields are used to write files from gcs to bigquery.
"""
[docs]
bq_inferencer_output_paths: InferenceOutput
[docs]
temp_predictions_gcs_path: Optional[GcsUri]
[docs]
temp_embedding_gcs_path: Optional[GcsUri]
[docs]
class InferencerV1:
"""
********** WILL BE DEPRECATED **********
Note this class will get deprecated in the future without notice
Use python/gigl/src/inference/inferencer.py instead
********** WILL BE DEPRECATED **********
GiGL Component that runs inference of a trained model on samples generated by the Subgraph Sampler component and outputs embedding and/or prediction assets.
"""
__bq_utils: BqUtils
__gbml_config_pb_wrapper: GbmlConfigPbWrapper
@property
[docs]
def bq_utils(self) -> BqUtils:
if not self.__bq_utils:
raise ValueError(f"bq_utils is not initialized before use.")
return self.__bq_utils
@property
[docs]
def gbml_config_pb_wrapper(self) -> GbmlConfigPbWrapper:
if not self.__gbml_config_pb_wrapper:
raise ValueError(f"gbml_config_pb_wrapper is not initialized before use.")
return self.__gbml_config_pb_wrapper
[docs]
def write_from_gcs_to_bq(
self,
schema: Dict[str, List[Dict[str, str]]],
gcs_uri: GcsUri,
bq_table_uri: str,
) -> None:
"""
Writes embeddings or predictions from gcs folder to bq table with specified schema
Args:
schema (Optional[Dict[str, List[Dict[str, str]]]): BQ Table schema for embeddings or predictions from inference output
gcs_uri (GcsUri): GCS Folder for embeddings or predictions from inference output
bq_table_uri (str): Path to the table for embeddings or predictions output
"""
assert schema is not None
assert "fields" in schema
field_schema = schema["fields"]
logger.info(f"schema = {field_schema}")
logger.info(f"loading from {gcs_uri} to BQ table: {bq_table_uri}")
self.bq_utils.load_file_to_bq(
source_path=GcsUri.join(gcs_uri, "*"),
bq_path=bq_table_uri,
job_config=bigquery.LoadJobConfig(
source_format=bigquery.SourceFormat.NEWLINE_DELIMITED_JSON,
write_disposition=bigquery.WriteDisposition.WRITE_TRUNCATE,
schema=field_schema,
),
retry=True,
)
logger.info(f"Finished loading to BQ table {bq_table_uri}")
[docs]
def generate_inferencer_instance(self) -> BaseInferencer:
kwargs: Dict[str, Any] = {}
inferencer_class_path: str = (
self.gbml_config_pb_wrapper.inferencer_config.inferencer_cls_path
)
kwargs = dict(self.gbml_config_pb_wrapper.inferencer_config.inferencer_args)
inferencer_cls = os_utils.import_obj(inferencer_class_path)
inferencer_instance: BaseInferencer
try:
inferencer_instance = inferencer_cls(**kwargs)
assert isinstance(inferencer_instance, BaseInferencer)
except Exception as e:
logger.error(f"Could not instantiate class {inferencer_cls}: {e}")
raise e
model_save_path_uri = UriFactory.create_uri(
self.gbml_config_pb_wrapper.shared_config.trained_model_metadata.trained_model_uri
)
logger.info(
f"Loading model state dict from: {model_save_path_uri}, for inferencer: {inferencer_instance}"
)
model_state_dict = load_state_dict_from_uri(load_from_uri=model_save_path_uri)
inferencer_instance.init_model(
gbml_config_pb_wrapper=self.gbml_config_pb_wrapper,
state_dict=model_state_dict,
)
return inferencer_instance
def __infer_single_node_type(
self,
inference_blueprint: BaseInferenceBlueprint,
applied_task_identifier: AppliedTaskIdentifier,
custom_worker_image_uri: Optional[str],
node_type: NodeType,
uri_prefix_list: List[Uri],
lock: threading.Lock,
) -> InferencerOutputPaths:
"""
Runs inference on a single node type
Args:
inference_blueprint (BaseInferenceBlueprint): Blueprint for running and saving inference for GBML pipelines
applied_task_identifier (AppliedTaskIdentifier): Identifier for the GiGL job
custom_worker_image_uri (Optional[str]): Uri to custom worker image
node_type (NodeType): Node type being inferred
uri_prefix_list (List[Uri]): List of prefixes for running inference for given node type
lock (threading.Lock): lock to prevent race conditions when starting dataflow pipelines
Returns:
(InferencerOutputPaths): Dataclass with path fields for writing from gcs to bigquery for given node type
"""
node_type_to_inferencer_output_info_map = (
self.gbml_config_pb_wrapper.shared_config.inference_metadata.node_type_to_inferencer_output_info_map
)
# Sanity check that we have some paths defined for intended inferred assets.
if not (
node_type_to_inferencer_output_info_map[node_type].embeddings_path
or node_type_to_inferencer_output_info_map[node_type].predictions_path
):
raise ValueError(
f"Inference metadata for node type {node_type} is missing; must have at least one of "
"embeddings_path or predictions_path defined."
)
should_persist_predictions = bool(
node_type_to_inferencer_output_info_map[node_type].predictions_path
)
should_persist_embeddings = bool(
node_type_to_inferencer_output_info_map[node_type].embeddings_path
)
temp_predictions_gcs_path: Optional[GcsUri]
temp_embeddings_gcs_path: Optional[GcsUri]
if should_persist_predictions:
temp_predictions_gcs_path = InferenceAssets.get_gcs_asset_write_path_prefix(
applied_task_identifier=applied_task_identifier,
bq_table_path=node_type_to_inferencer_output_info_map[
node_type
].predictions_path,
)
else:
temp_predictions_gcs_path = None
if should_persist_embeddings:
temp_embeddings_gcs_path = InferenceAssets.get_gcs_asset_write_path_prefix(
applied_task_identifier=applied_task_identifier,
bq_table_path=node_type_to_inferencer_output_info_map[
node_type
].embeddings_path,
)
else:
temp_embeddings_gcs_path = None
with lock:
logger.debug(f"Node Type {node_type} acquiring lock.")
p = get_inferencer_pipeline_component_for_single_node_type(
gbml_config_pb_wrapper=self.gbml_config_pb_wrapper,
inference_blueprint=inference_blueprint,
applied_task_identifier=applied_task_identifier,
custom_worker_image_uri=custom_worker_image_uri,
node_type=node_type,
uri_prefix_list=uri_prefix_list,
temp_predictions_gcs_path=temp_predictions_gcs_path,
temp_embeddings_gcs_path=temp_embeddings_gcs_path,
)
inferencer_pipeline_result = p.run()
logger.debug(f"Node Type {node_type} releasing lock.")
logger.info(f"Starting Dataflow job to run inference on {node_type} node type")
inferencer_pipeline_result.wait_until_finish()
logger.info(f"Finished Dataflow job to run inference on {node_type} node type")
if isinstance(inferencer_pipeline_result, DataflowPipelineResult):
pipeline_state: str = inferencer_pipeline_result.state
if pipeline_state != PipelineState.DONE:
raise RuntimeError(
f"A dataflow pipeline failed, has state {pipeline_state}: {inferencer_pipeline_result}"
)
return InferencerOutputPaths(
bq_inferencer_output_paths=node_type_to_inferencer_output_info_map[
node_type
],
temp_predictions_gcs_path=temp_predictions_gcs_path,
temp_embedding_gcs_path=temp_embeddings_gcs_path,
)
def __run(
self,
applied_task_identifier: AppliedTaskIdentifier,
task_config_uri: Uri,
custom_worker_image_uri: Optional[str] = None,
):
self.__gbml_config_pb_wrapper = (
GbmlConfigPbWrapper.get_gbml_config_pb_wrapper_from_uri(
gbml_config_uri=task_config_uri
)
)
if self.gbml_config_pb_wrapper.shared_config.should_skip_inference:
logger.info("Skipping inference as flag set in GbmlConfig")
return
inferencer_instance: BaseInferencer = self.generate_inferencer_instance()
graph_builder = GraphBuilderFactory.get_graph_builder(
backend_name=inferencer_instance.model.graph_backend # type: ignore
)
inference_blueprint: BaseInferenceBlueprint = (
InferenceBlueprintFactory.get_inference_blueprint(
gbml_config_pb_wrapper=self.gbml_config_pb_wrapper,
inferencer_instance=inferencer_instance,
graph_builder=graph_builder,
)
)
node_type_to_inferencer_output_paths_map: Dict[
NodeType, InferencerOutputPaths
] = dict()
dataflow_setup_lock = threading.Lock()
# We kick off multiple Inferencer pipelines, each of which kicks off a setup.py sdist run.
# sdist has race-condition issues for simultaneous runs: https://github.com/pypa/setuptools/issues/1222
# We have each thread take a lock when kicking off the pipelines to avoid this issue.
num_workers = min(get_available_cpus(), MAX_INFERENCER_NUM_WORKERS)
with concurrent.futures.ThreadPoolExecutor(max_workers=num_workers) as executor:
logger.info(f"Using up to {num_workers} threads.")
futures: Dict[
concurrent.futures.Future[InferencerOutputPaths], NodeType
] = dict()
for (
node_type,
uri_prefix_list,
) in (
inference_blueprint.get_inference_data_tf_record_uri_prefixes().items()
):
# Launching one beam pipeline per node type
future = executor.submit(
self.__infer_single_node_type,
inference_blueprint=inference_blueprint,
applied_task_identifier=applied_task_identifier,
custom_worker_image_uri=custom_worker_image_uri,
node_type=node_type,
uri_prefix_list=uri_prefix_list,
lock=dataflow_setup_lock,
)
futures.update({future: node_type})
for future in concurrent.futures.as_completed(futures):
node_type = futures[future]
try:
inferencer_output_paths: InferencerOutputPaths = future.result()
node_type_to_inferencer_output_paths_map[
node_type
] = inferencer_output_paths
except Exception as e:
logger.exception(
f"{node_type} inferencer job failed due to a raised exception: {e}"
)
raise e
for (
node_type,
inferencer_output_paths,
) in node_type_to_inferencer_output_paths_map.items():
condensed_node_type = self.__gbml_config_pb_wrapper.graph_metadata_pb_wrapper.node_type_to_condensed_node_type_map[
node_type
]
should_run_unenumeration = bool(
self.__gbml_config_pb_wrapper.preprocessed_metadata_pb_wrapper.preprocessed_metadata_pb.condensed_node_type_to_preprocessed_metadata[
condensed_node_type
].enumerated_node_ids_bq_table
)
temp_predictions_gcs_path = (
inferencer_output_paths.temp_predictions_gcs_path
)
temp_embeddings_gcs_path = inferencer_output_paths.temp_embedding_gcs_path
bq_inferencer_output_paths = (
inferencer_output_paths.bq_inferencer_output_paths
)
if temp_predictions_gcs_path is not None:
self.write_from_gcs_to_bq(
schema=inference_blueprint.get_pred_table_schema(should_run_unenumeration=should_run_unenumeration).schema, # type: ignore
gcs_uri=temp_predictions_gcs_path,
bq_table_uri=bq_inferencer_output_paths.predictions_path,
)
if temp_embeddings_gcs_path is not None:
self.write_from_gcs_to_bq(
schema=inference_blueprint.get_emb_table_schema(should_run_unenumeration=should_run_unenumeration).schema, # type: ignore
gcs_uri=temp_embeddings_gcs_path,
bq_table_uri=bq_inferencer_output_paths.embeddings_path,
)
@flushes_metrics(get_metrics_service_instance_fn=get_metrics_service_instance)
@profileit(
metric_name=TIMER_INFERENCER_S,
get_metrics_service_instance_fn=get_metrics_service_instance,
)
[docs]
def run(
self,
applied_task_identifier: AppliedTaskIdentifier,
task_config_uri: Uri,
custom_worker_image_uri: Optional[str] = None,
):
try:
return self.__run(
applied_task_identifier=applied_task_identifier,
task_config_uri=task_config_uri,
custom_worker_image_uri=custom_worker_image_uri,
)
except Exception as e:
logger.error(
"Inference failed due to a raised exception; which will follow"
)
logger.error(e)
logger.error(traceback.format_exc())
sys.exit(f"System will now exit: {e}")
def __init__(self, bq_gcp_project: str):
self.__bq_utils = BqUtils(project=bq_gcp_project if bq_gcp_project else None)
if __name__ == "__main__":
[docs]
parser = argparse.ArgumentParser(description="Program to run distributed inference")
parser.add_argument(
"--job_name",
type=str,
help="Unique identifier for the job name",
required=True,
)
parser.add_argument(
"--task_config_uri",
type=str,
help="Gbml config uri",
required=True,
)
parser.add_argument(
"--resource_config_uri",
type=str,
help="Runtime argument for resource and env specifications of each component",
required=True,
)
parser.add_argument(
"--custom_worker_image_uri",
type=str,
help="Docker image to use for the worker harness in dataflow",
required=False,
)
parser.add_argument(
"--cpu_docker_uri",
type=str,
help="User Specified or KFP compiled Docker Image for CPU inference",
required=False,
)
parser.add_argument(
"--cuda_docker_uri",
type=str,
help="User Specified or KFP compiled Docker Image for GPU inference",
required=False,
)
args = parser.parse_args()
task_config_uri = UriFactory.create_uri(args.task_config_uri)
resource_config_uri = UriFactory.create_uri(args.resource_config_uri)
custom_worker_image_uri = args.custom_worker_image_uri
initialize_metrics(task_config_uri=task_config_uri, service_name=args.job_name)
applied_task_identifier = AppliedTaskIdentifier(args.job_name)
inferencer = InferencerV1(bq_gcp_project=get_resource_config().project)
inferencer.run(
applied_task_identifier=applied_task_identifier,
task_config_uri=task_config_uri,
custom_worker_image_uri=custom_worker_image_uri,
)
