GraphBuilder

Creating a GraphBuilder

Function
Creates a GraphBuilder for subsequent graph construction.

Prototype

torch_atb.GraphBuilder (graphName: string) -> torch_atb.GraphBuilder

Parameter
Name of a graph operator. The value is of the string type.
Return Value
GraphBuilder
Notice
If the API fails to be called, an exception is thrown and the API stops running.
Example
1
builder = torch_atb.Builder("Graph")

Setting the Input of a Computational Graph

Function
Sets the input of a computational graph.

Prototype

GraphBuilder.AddInput (name: string) -> torch_atb.GraphBuilder

Parameter
Name of the input. The value is of the string type.
Return Value
string
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

x = builder.add_input("x")
y = builder.add_input("y")

Setting the Node of a Computational Graph

Function
Sets the node of a computational graph.

Prototype

GraphBuilder.AddNode (inputs: vector<string>, param: xxxParam) -> torch_atb.GraphBuilder

Parameter
Name of the input, which is the parameter of the operator created using torch_atb.
Return Value
GraphNode
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

# Create and set the operator parameters.
elewise_add = torch_atb.ElewiseParam()
elewise_add.elewise_type = torch_atb.ElewiseParam.ElewiseType.ELEWISE_ADD
layer1 = builder.add_node([x, y], elewise_add)

Setting the Output of a Computational Graph

Function
Marks the input tensor as the output of the computational graph.

Prototype

GraphBuilder.MarkOutput (outTensor: string) -> torch_atb.GraphBuilder

Parameter
Name of the tensor.
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

builder.mark_output(layer2.get_output(0))

Tensor Transposition

Function
Changes the shape of a tensor.

Prototype

GraphBuilder.Reshape (string, std::function, string) -> torch_atb.GraphBuilder

Parameter
Name of the tensor.
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

layer1 = builder.add_node([x, y], elewise_add)
add_out = layer1.get_output(0)
builder.reshape(add_out, lambda shape: [1, shape[0] * shape[1]], "add_out_")
# Use "add_out_" to bind the tensor to other nodes after transposition.
layer2 = builder.add_node(["add_out_", z], elewise_mul)

Building a Computational Graph

Function
Creates a graph op.

Prototype

GraphBuilder.Build () -> torch_atb.GraphBuilder

Return Value
```
atb::OperationWrapper
```
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

builder = torch_atb.Builder("Graph")
# The process of adding inputs, nodes, and marking outputs is omitted.
Graph = builder.build()

Setting the Actual Execution Streams

Function
Sets the actual execution streams.

Prototype

GraphBuilder.SetExecuteStreams (executeStreams: std::vector<std::uintptr_t>) -> torch_atb.GraphBuilder

Parameter
List of actual execution streams (aclStream).
Notice
If the API fails to be called, an exception is thrown and the API stops running.

Example

def create_streams():
    streams = []
    for i in range(2):
        stream, ret = acl.rt.create_stream()
        if ret != ACL_SUCCESS:
            exit(0)
        streams.append(stream)
    return streams

builder = torch_atb.Builder("Graph")
builder.set_execute_streams(create_streams())

Parent topic: Graph API