Graph Class¶

The Forge Graph class wraps around the igraph.Graph object that is initialized by the graphical encoding provided by the graph-core API.

forge.relay.graph.graph.Graph ¶

Graph(
    relay_expr: obj,
    params: Optional[Union[dict, ParamsDict]] = None,
    sink_name: str = "Sink",
)

Wrapper around igraph.Graph instances for graph-encoded Relay

This class provides a high-level API and context for working with Forge Nodes, enabling a 1-to-1 graphical representation of any Relay expression. It includes methods for traversal, transformation, and manipulation of the graph while maintaining Relay-specific semantics.

Initializes a Graph instance from a Relay expression.

Parameters:

Name	Type	Description	Default
`relay_expr`	`obj`	The Relay expression to encode as a graph.	required
`params`	`Optional[Union[dict, ParamsDict]]`	Optional parameters to associate with the graph.	`None`
`sink_name`	`str`	The name of the artificial sink vertex. Defaults to "Sink".	`'Sink'`

Methods:

Name	Description
`find`	Finds a single node in the graph using `igraph.Graph.vs.find()`.
`select`	Selects multiple nodes in the graph using `igraph.Graph.vs.select()`.
`copy`	Creates a deep copy of the graph.
`get_inference_function`	Generates an inference function for the graph.
`graft_relay_expr`	Grafts a Relay expression onto the graph.
`dfs_iter`	Iterates over the graph using depth-first search (DFS) from some root node.
`get_relay_mod`	Converts the graph rooted at some node into a Relay IRModule.
`get_relay_func`	Converts the graph rooted at some node into a Relay Function.
`get_relay_expr`	Converts the graph rooted at some node into a Relay expression.
`get_unbound_var_nodes`	Retrieves unbound variable nodes in the graph.
`get_bound_var_nodes`	Retrieves bound variable nodes in the graph.
`reload`	Reloads the graph to restore its original state.
`__iter__`	Iterates over all nodes in the graph in DFS from the sink.
`__eq__`	Checks structural equality between two Graph instances.
`__hash__`	Computes a structural hash value for the graph.
`__repr__`	Generates a string representation of the graph.
`__getstate__`	Serializes the graph state for pickling.
`__setstate__`	Restores the graph state from a serialized representation.
`_init_node`	Initializes a Node instance from an igraph.Vertex.
`_init_nodes`	Initializes multiple Node instances from igraph vertices.
`_get_ivertex`	Retrieves the corresponding igraph.Vertex for a given node.

Attributes:

Name	Type	Description
`params`	`ParamsDict`
`igraph`	`Graph`
`calibration`	`Dict[str, CalibrationMetaData]`
`fingerprint`	`str`	Computes a unique fingerprint for the graph, a hash of structure and data
`inputs`	`List[Node]`	Retrieves the input nodes of the graph.
`output`	`Node`	Retrieves the output node of the graph.
`sink`	`Node`	Retrieves the artificial sink node of the graph.
`relay_expr_type`	`str`	Gets the type of the Relay expression represented by the graph.
`mod`	`obj`	Creates the Relay IRModule encoded by the graph.
`typed_mod`	`obj`	Creates the Relay IRModule encoded by the graph with typing.
`operators`	`Dict[str, int]`	Retrieves a dictionary of operators in the graph and their usage counts.

Attributes¶

params `class-attribute` `instance-attribute` ¶

params: ParamsDict = params

igraph `class-attribute` `instance-attribute` ¶

igraph: Graph = from_relay(
    relay_expr, sink_name=sink_name, input_names=input_names
)

calibration `class-attribute` `instance-attribute` ¶

calibration: Dict[str, CalibrationMetaData] = field(
    init=False, factory=dict, repr=False
)

fingerprint `property` ¶

fingerprint: str

Computes a unique fingerprint for the graph, a hash of structure and data

Returns:

Name	Type	Description
`str`	`str`	A string representing the graph's unique fingerprint.

inputs `property` ¶

inputs: List[Node]

Retrieves the input nodes of the graph.

Returns:

Type	Description
`List[Node]`	List[Node]: A list of input nodes.

output `property` ¶

output: Node

Retrieves the output node of the graph.

Returns:

Name	Type	Description
`Node`	`Node`	The output node.

Note

The output node is the node that immediately precedes the sink nodes

sink `property` ¶

sink: Node

Retrieves the artificial sink node of the graph.

Returns:

Name	Type	Description
`Node`	`Node`	The sink node.

relay_expr_type `property` ¶

relay_expr_type: str

Gets the type of the Relay expression represented by the graph.

Returns:

Name	Type	Description
`str`	`str`	The type of the Relay expression.

mod `property` ¶

mod: obj

Creates the Relay IRModule encoded by the graph.

Returns:

Type	Description
`obj`	Relay.IRModule.obj: The represented Relay IRModule.

typed_mod `property` ¶

typed_mod: obj

Creates the Relay IRModule encoded by the graph with typing.

Returns:

Type	Description
`obj`	Relay.IRModule.obj: The typed Relay IRModule.

Note

This is can be used as a 'validation' of expression correctness. An error will be thrown for invalid expressions.

operators `property` ¶

operators: Dict[str, int]

Retrieves a dictionary of operators in the graph and their usage counts.

Returns:

Type	Description
`Dict[str, int]`	Dict[str, int]: A mapping of operator names to their usage counts.

Functions¶

find ¶

find(*args, **kwargs) -> Node

Finds a single node in the graph using igraph.Graph.vs.find().

Parameters:

Name	Type	Description	Default
`*args`		Positional arguments for `igraph.Graph.vs.find()`.	`()`
`**kwargs`		Keyword arguments for `igraph.Graph.vs.find()`.	`{}`

Returns:

Name	Type	Description
`Node`	`Node`	The found node.

Note

This method is direct access to the underlying igraph method, so API matches it directly, please reference the igraph documentation.
If using lambda functions, the functions are expected to interact with igraph.Vertex objects, not Node objects.
This method returns a Forge Node instance

select ¶

select(*args, **kwargs) -> List[Node]

Selects multiple nodes in the graph using igraph.Graph.vs.select().

Parameters:

Name	Type	Description	Default
`*args`		Positional arguments for `igraph.Graph.vs.select()`.	`()`
`**kwargs`		Keyword arguments for `igraph.Graph.vs.select()`.	`{}`

Returns:

Type	Description
`List[Node]`	List[Node]: The selected nodes.

Note

This method is direct access to the underlying igraph method, so API matches it directly, please reference the igraph documentation.
If using lambda functions, the functions are expected to interact with igraph.Vertex objects, not Node objects.
This method returns a Forge Node instance

copy ¶

copy() -> Graph

Creates a deep copy of the graph.

Returns:

Name	Type	Description
`Graph`	`Graph`	A deep copy of the current graph instance.

get_inference_function ¶

get_inference_function(
    root: Optional[Union[str, Node, Vertex]] = None, target: str = "llvm"
) -> Callable

Generates an inference function for the graph.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node for inference. Defaults to None (i.e. the sink node).	`None`
`target`	`str`	The target for inference (e.g., "llvm"). Defaults to "llvm".	`'llvm'`

Returns:

Name	Type	Description
`Callable`	`Callable`	A callable function for inference.

graft_relay_expr ¶

graft_relay_expr(
    relay_expr: obj,
    target_names: List[str],
    target_argidxs: List[int],
    *,
    root_name: Optional[str] = None
) -> str

Grafts a Relay expression onto the graph.

Parameters:

Name	Type	Description	Default
`relay_expr`	`obj`	The Relay expression to graft.	required
`target_names`	`List[str]`	Names of the target nodes in the graph.	required
`target_argidxs`	`List[int]`	Argument indices for the target nodes.	required
`root_name`	`Optional[str]`	Optional custom name for the root of the grafted expression. If None, a random node name will be generated. Defaults to None.	`None`

Returns:

Name	Type	Description
`str`	`str`	The name of the grafted root node.

dfs_iter ¶

dfs_iter(
    root: Optional[Union[str, Node, Vertex]] = None, in_order: bool = True
) -> Iterator

Iterates over the graph using depth-first search (DFS) from some root node.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start the DFS from. If None, the sink node is used as the root.	`None`
`in_order`	`bool`	Whether to perform an in-order DFS traversal. Defaults to True.	`True`

Returns:

Name	Type	Description
`Iterator`	`Iterator`	An iterator over the nodes in DFS order.

get_relay_mod ¶

get_relay_mod(
    root: Optional[Union[str, Node, Vertex]] = None,
    typed: bool = True,
    global_vars: Optional[Dict[str, obj]] = None,
) -> obj

Converts the graph rooted at some node into a Relay IRModule.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start from. If None, the sink node is used. Defaults to None.	`None`
`typed`	`bool`	Whether to produce a typed Relay IRModule. Defaults to True.	`True`
`global_vars`	`Optional[Dict[str, obj]]`	A dictionary of global variables to include. Defaults to None.	`None`

Returns:

Type	Description
`obj`	Relay.IRModule.obj: The corresponding Relay IRModule.

get_relay_func ¶

get_relay_func(
    root: Optional[Union[str, Node, Vertex]] = None,
    typed: bool = True,
    global_vars: Optional[Dict[str, obj]] = None,
) -> obj

Converts the graph rooted at some node into a Relay Function.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start from. If None, the sink node is used. Defaults to None.	`None`
`typed`	`bool`	Whether to produce a typed Relay Function. Defaults to True.	`True`
`global_vars`	`Optional[Dict[str, obj]]`	A dictionary of global variables to include. Defaults to None.	`None`

Returns:

Type	Description
`obj`	Relay.Function.obj: The corresponding Relay Function.

get_relay_expr ¶

get_relay_expr(
    root: Optional[Union[str, Node, Vertex]] = None,
    typed: bool = True,
    global_vars: Optional[Dict[str, obj]] = None,
) -> obj

Converts the graph rooted at some node into a Relay expression.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start from. If None, the sink node is used. Defaults to None.	`None`
`typed`	`bool`	Whether to produce a typed Relay expression. Defaults to True.	`True`
`global_vars`	`Optional[Dict[str, obj]]`	A dictionary of global variables to include. Defaults to None.	`None`

Returns:

Type	Description
`obj`	Relay.Node.obj: The corresponding Relay expression.

get_unbound_var_nodes ¶

get_unbound_var_nodes(
    root: Optional[Union[str, Node, Vertex]] = None, include_inputs: bool = True
) -> List[Node]

Retrieves unbound variable nodes in the graph.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start from. If None, the sink node is used. Defaults to None.	`None`
`include_inputs`	`bool`	Whether to include input nodes in the result. Defaults to True.	`True`

Returns:

Type	Description
`List[Node]`	List[Node]: A list of unbound variable nodes.

get_bound_var_nodes ¶

get_bound_var_nodes(
    root: Optional[Union[str, Node, Vertex]] = None,
) -> List[Node]

Retrieves bound variable nodes in the graph.

Parameters:

Name	Type	Description	Default
`root`	`Optional[Union[str, Node, Vertex]]`	The root node to start from. If None, the sink node is used. Defaults to None.	`None`

Returns:

Type	Description
`List[Node]`	List[Node]: A list of bound variable nodes.

reload ¶

reload()

Reloads the graph to restore its original state.

This function resets the graph to its initial state, restoring the original igraph bytes.

iter ¶

__iter__() -> Iterator

Iterates over all nodes in the graph in DFS from the sink.

Returns:

Name	Type	Description
`Iterator`	`Iterator`	An iterator over the nodes in the graph.

eq ¶

__eq__(other) -> bool

Checks structural equality between two Graph instances.

Parameters:

Name	Type	Description	Default
`other`	`Graph`	The other graph to compare with.	required

Returns:

Name	Type	Description
`bool`	`bool`	True if the graphs are equal, False otherwise.

hash ¶

__hash__() -> int

Computes a structural hash value for the graph.

Returns:

Name	Type	Description
`int`	`int`	The hash value of the graph.

repr ¶

__repr__() -> str

Generates a string representation of the graph.

Returns:

Name	Type	Description
`str`	`str`	A string describing the graph.

getstate ¶

__getstate__() -> dict

Serializes the graph state for pickling.

Returns:

Name	Type	Description
`dict`	`dict`	The serialized state of the graph.

setstate ¶

__setstate__(state) -> None

Restores the graph state from a serialized representation.

Parameters:

Name	Type	Description	Default
`state`	`dict`	The serialized state to restore.	required

_init_node ¶

_init_node(vertex: Vertex) -> Node

Initializes a Node instance from an igraph.Vertex.

Parameters:

Name	Type	Description	Default
`vertex`	`Vertex`	The vertex to initialize a Node from.	required

Returns:

Name	Type	Description
`Node`	`Node`	The initialized Node.

_init_nodes ¶

_init_nodes(vertices: Union[Iterable[Vertex], VertexSeq]) -> List[Node]

Initializes multiple Node instances from igraph vertices.

Parameters:

Name	Type	Description	Default
`vertices`	`Union[Iterable[Vertex], VertexSeq]`	The vertices to initialize Nodes from.	required

Returns:

Type	Description
`List[Node]`	List[Node]: A list of initialized Nodes.

_get_ivertex ¶

_get_ivertex(node: Union[str, Node, Vertex]) -> Vertex

Retrieves the corresponding igraph.Vertex for a given node.

Parameters:

Name	Type	Description	Default
`node`	`Union[str, Node, Vertex]`	The node to retrieve the vertex for.	required

Returns:

Type	Description
`Vertex`	igraph.Vertex: The corresponding igraph.Vertex.

Graph Class¶

forge.relay.graph.graph.Graph ¶

Attributes¶

params class-attribute instance-attribute ¶

igraph class-attribute instance-attribute ¶

calibration class-attribute instance-attribute ¶

fingerprint property ¶

inputs property ¶

output property ¶

sink property ¶

relay_expr_type property ¶

mod property ¶

typed_mod property ¶

operators property ¶

Functions¶

find ¶

select ¶

copy ¶

get_inference_function ¶

graft_relay_expr ¶

dfs_iter ¶

get_relay_mod ¶

get_relay_func ¶

get_relay_expr ¶

get_unbound_var_nodes ¶

get_bound_var_nodes ¶

reload ¶

__iter__ ¶

__eq__ ¶

__hash__ ¶

__repr__ ¶

__getstate__ ¶

__setstate__ ¶

_init_node ¶

_init_nodes ¶

_get_ivertex ¶

params `class-attribute` `instance-attribute` ¶

igraph `class-attribute` `instance-attribute` ¶

calibration `class-attribute` `instance-attribute` ¶

fingerprint `property` ¶

inputs `property` ¶

output `property` ¶

sink `property` ¶

relay_expr_type `property` ¶

mod `property` ¶

typed_mod `property` ¶

operators `property` ¶

iter ¶

eq ¶

hash ¶

repr ¶

getstate ¶

setstate ¶