OfflineCompile

Product Support Status

Product	Support Status
Atlas A3 Training Series Products/Atlas A3 Inference Series Products	√
Atlas A2 Training Series Products/Atlas A2 Inference Series Products	√

Module Import

from ge.offline_compile import (
    build_initialize, build_finalize, build_model, save_model,
    bundle_build_model, bundle_save_model, GraphWithOptions, ModelBuffer
)
from ge.error import GeError

Functionality Description

Offline compilation module provides the capability to compile computation graphs into offline models (.om files). Supports single graph compilation (build_model) and multi-graph bundle compilation (bundle_build_model).

Usage flow:

1. Call build_initialize to initialize offline compilation resources.
2. Call build_model to compile single graph, or call bundle_build_model to compile multiple graphs into bundle offline model.
3. Call save_model or bundle_save_model to save compilation result as .om file.
4. Call build_finalize to release offline compilation resources.

Configuration in build_options has higher priority than global_options in build_initialize.

When GE underlying interface execution fails, build_initialize, build_model, save_model, bundle_build_model, bundle_save_model will throw GeError. GeError inherits from RuntimeError, exception information contains GE internal error information and interface context.

Data Classes

GraphWithOptions

A combination of a graph and its corresponding compilation options, used in bundle compilation scenarios.

@dataclass
class GraphWithOptions:
    graph: Graph
    build_options: Dict[str, str]

Attribute	Type	Description
graph	Graph	Computation graph object to be compiled
build_options	Dict[str, str]	Compilation options dictionary corresponding to this graph, keys and values are strings, default is empty dictionary

Classes

ModelBuffer

Offline model buffer class, used to store compiled offline model data. This class does not support direct instantiation, returned by build_model or bundle_build_model.

Method/Attribute	Description
length (property)	Gets model buffer length (bytes), returns int
get_length()	Gets model buffer length (bytes), returns int

Function Prototypes

build_initialize

def build_initialize(global_options: Optional[dict] = None) -> None

Initializes resources required for offline compilation.

build_finalize

def build_finalize() -> None

Releases offline compilation resources.

build_model

def build_model(graph: Graph, build_options: Optional[dict] = None) -> ModelBuffer

Compiles single computation graph into offline model, model data saved in memory.

save_model

def save_model(output_file: str, model: ModelBuffer) -> None

Serializes offline model in memory into .om file.

bundle_build_model

def bundle_build_model(graph_with_options: List[GraphWithOptions]) -> ModelBuffer

Compiles multiple graphs into bundle offline model.

bundle_save_model

def bundle_save_model(output_file: str, model: ModelBuffer) -> None

Serializes bundle offline model in memory into .om file.

Parameter Description

build_initialize

Parameter	Type	Required	Description
global_options	Optional[dict]	No	Global compilation configuration dictionary, keys and values are strings. Default is None, means no additional configuration passed

build_model

Parameter	Type	Required	Description
graph	Graph	Yes	Computation graph object to be compiled
build_options	Optional[dict]	No	Graph-level compilation configuration dictionary, keys and values are strings. Configuration priority see constraint description. Default is None

save_model

Parameter	Type	Required	Description
output_file	str	Yes	Base name of output model file. Generated offline model file name automatically has .om suffix. If file name contains OS and architecture information, then this OM file can only be used in corresponding OS and architecture runtime environment
model	ModelBuffer	Yes	Offline model buffer object

bundle_build_model

Parameter	Type	Required	Description
graph_with_options	List[GraphWithOptions]	Yes	List containing multiple GraphWithOptions, each element contains a computation graph and its compilation options. List must contain 2 or more graphs

bundle_save_model

Parameter	Type	Required	Description
output_file	str	Yes	Base name of output model file. Generated offline model file name automatically has .om suffix. If file name contains OS and architecture information, then this OM file can only be used in corresponding OS and architecture runtime environment
model	ModelBuffer	Yes	Bundle offline model buffer object

Return Value Description

build_model

Return Value	Type	Description
model	ModelBuffer	Buffer object containing compiled offline model data

bundle_build_model

Return Value	Type	Description
model	ModelBuffer	Buffer object containing compiled bundle offline model data

Other Functions

build_initialize, build_finalize, save_model, bundle_save_model have no return value.

Constraint Description

• Before using offline compilation module, must call build_initialize to initialize resources, after use must call build_finalize to release resources.
• Keys and values in build_options must be string type.
• graph_with_options list in bundle_build_model must contain 2 or more graphs.
• ModelBuffer objects do not support copy and deep copy, and do not support direct instantiation.
• output_file parameter in save_model and bundle_save_model is base name of output file, system automatically adds .om suffix.
• Configuration in build_options has higher priority than global_options in build_initialize.
• When GE underlying interface execution fails, will throw GeError, exception information contains GE internal error information and interface context.

Usage Example

from ge.offline_compile import (
    build_initialize, build_finalize,
    build_model, save_model,
    bundle_build_model, bundle_save_model,
    GraphWithOptions
)
from ge.graph import Graph

# Initialize offline compilation resources
build_initialize({"log_level": "0"})

# Single graph compilation
model = build_model(graph, {"input_format": "ND"})
save_model("output_model", model)
print(f"Model size: {model.length} bytes")

# Multi-graph bundle compilation
gwo_list = [
    GraphWithOptions(graph1, {"input_format": "ND"}),
    GraphWithOptions(graph2, {"input_format": "NCHW"}),
]
bundle_model = bundle_build_model(gwo_list)
bundle_save_model("bundle_output", bundle_model)

# Release offline compilation resources
build_finalize()