Documentation

LoadGraph

Description

  • Distributed build and partitioning of foundation models:

    Loads the OM offline model to the current session and associates the model with GraphId based on the passed graph ID. After LoadGraph is successfully executed, RunGraph can be used to specify GraphId.

  • Asynchronous graph execution:

    Binds a graph with the specified graph ID to the corresponding stream. After LoadGraph is successfully executed, you can use ExecuteGraphWithStreamAsync to execute the graph.

Prototype

  • Distributed build and partitioning of foundation models: 
    Status LoadGraph(const uint32_t graph_id, const std::map<std::string, std::string> &options, const std::string &om_file_path) const;
  • Asynchronous graph execution: 
    Status LoadGraph(const uint32_t graph_id, const std::map<AscendString, AscendString> &options, void *stream) const;

Parameters

Parameter

Input/Output

Description

graph_id

Input

ID of the graph to be executed.

options

Input

Option that may be used in the execution phase. A key-value mapping table, for the graph configuration. key indicates the option type, and value indicates the option value.

Generally, this parameter can be left empty, indicating using the same options configuration passed to GEInitialize.

  • In the distributed build and partitioning scenario of foundation models, the key and value are of the string type. You can use this parameter to configure the current graph separately. For details about the supported configuration options, see Command-Line Options.
  • In the asynchronous graph execution scenario, the key and value are of the AscendString type. You can use this parameter to configure the current graph separately. For details about the supported configuration options, see Command-Line Options > ge.exec.frozenInputIndexes. Currently, only this parameter can be configured.

om_file_path (distributed build and partitioning of foundation models)

Input

Path of the OM offline model.

stream (asynchronous graph execution)

Input

Graph execution flow.

Returns

Parameter

Type

Description

-

Status

GE_CLI_SESS_RUN_FAILED: Serialization failed when a subgraph is executed.

SUCCESS: Subgraph executed successfully.

FAILED: Failed to execute the subgraph.

Restrictions

  • Distributed build and partitioning of foundation models:
    1. Only models used for distributed deployment can be loaded, that is, offline models generated after --distributed_cluster_build=1 is enabled by the ATC tool or DISTRIBUTED_CLUSTER_BUILD is enabled by Graph Build to an Offline Model during model compilation. If this API is used to load other offline models, the models will be verified, and a failure message will be returned.
    2. If offline models loaded by calling this API contain variables, a failure message is returned.
  • Asynchronous graph execution:

    Before calling this API, ensure that the CompileGraph process has been completed and the API must be used together with ExecuteGraphWithStreamAsync.

Parent topic: Graph Running APIs