core

Classes

ConstantTarget	ConstantTarget(name: 'str', value: 'Any')
ExternalTarget	External target for stitched modules.
FunctionTarget	FunctionTarget(name: 'str', function: 'Callable[..., Any]')
IOReducer	IOReducer()
InputDescriptor	InputDescriptor(target: 'Target', input_name: 'str' = '', input_adapter: 'InputAdapter' = <function default_input_adapter_fn at 0x7f8ce6f45bc0>, reducer: 'InputReducer' = <factory>)
InputReducer	InputReducer(reducer_fn: 'Callable[[InputArgs, InputArgs, InputArgs, int, list[InputArgs]], InputArgs]' = <function default_input_reducer_fn at 0x7f8ce6f474c0>)
ModuleTarget	ModuleTarget(name: 'str', module: 'nn.Module')
Needle
Node	Node(target: 'Target', stitches_to: 'list[StitchDescriptor]' = <factory>, stitches_from: 'list[StitchDescriptor]' = <factory>)
OutputDescriptor	OutputDescriptor(target: 'Target', output_name: 'str' = '', output_adapter: 'OutputAdapter' = <function default_output_adapter_fn at 0x7f8ce6f45d00>, reducer: 'OutputReducer' = <factory>)
OutputReducer	OutputReducer(reducer_fn: 'Callable[[OutputValue, OutputValue, Optional[OutputValue], int, list[OutputValue]], OutputValue]' = <function default_output_reducer_fn at 0x7f8ce6fcf6a0>, requires_original_output: 'bool' = False)
RemoteDataDescriptor	RemoteDataDescriptor(key: 'str')
RemotePythonDataDescriptor	RemotePythonDataDescriptor(key: 'str', value: 'Any')
RemoteTarget	RemoteTarget(peer_rank: 'Union[int, Sequence[int]]', process_group: 'Optional[torch.distributed.ProcessGroup]' = None, blocking: 'bool' = True)
RemoteTensorDataDescriptor	RemoteTensorDataDescriptor(key: 'str', device: "Literal['cuda', 'cpu']", dtype: 'torch.dtype', shape: 'torch.Size')
Singleton
StitchDescriptor	StitchDescriptor(source_descriptor: 'IODescriptor', destination_descriptor: 'IODescriptor')
StitchedModule
StitchedModuleOutput	StitchedModuleOutput(captured_inputs: 'dict[str, InputArgs]', captured_outputs: 'dict[str, Any]')
Target	Target()
TargetWithInput	TargetWithInput()
TargetWithNamedInputs	TargetWithNamedInputs()
TargetWithNamedOutputs	TargetWithNamedOutputs()
TargetWithOutput	TargetWithOutput()

Functions

default_input_adapter_fn
default_input_reducer_fn
default_output_adapter_fn
default_output_reducer_fn

exception CantResolveNodeDependenciesException

Bases: StitchedModuleException

class ConstantTarget

Bases: TargetWithOutput

ConstantTarget(name: ‘str’, value: ‘Any’)

__init__(name, value)

Parameters:

• name (str)

• value (Any)

Return type:

None

name: str

value: Any

class ExternalTarget

Bases: TargetWithNamedInputs, TargetWithNamedOutputs

External target for stitched modules.

__init__()

Return type:

None

class FunctionTarget

Bases: TargetWithInput, TargetWithOutput

FunctionTarget(name: ‘str’, function: ‘Callable[…, Any]’)

__init__(name, function)

Parameters:

• name (str)

• function (Callable[[...], Any])

Return type:

None

function: Callable[[...], Any]

name: str

class IOReducer

Bases: object

IOReducer()

__init__()

Return type:

None

class InputDescriptor

Bases: object

InputDescriptor(target: ‘Target’, input_name: ‘str’ = ‘’, input_adapter: ‘InputAdapter’ = <function default_input_adapter_fn at 0x7f8ce6f45bc0>, reducer: ‘InputReducer’ = <factory>)

__init__(target, input_name='', input_adapter=<function default_input_adapter_fn>, reducer=<factory>)

Parameters:

• target (Target)

• input_name (str)

• input_adapter (Callable[[InputArgs], InputArgs])

• reducer (InputReducer)

Return type:

None

input_adapter()

Parameters:

input_values (InputArgs)

Return type:

InputArgs

input_name: str = ''

reducer: InputReducer

target: Target

class InputReducer

Bases: IOReducer

InputReducer(reducer_fn: ‘Callable[[InputArgs, InputArgs, InputArgs, int, list[InputArgs]], InputArgs]’ = <function default_input_reducer_fn at 0x7f8ce6f474c0>)

__init__(reducer_fn=<function default_input_reducer_fn>)

Parameters:

reducer_fn (Callable[[InputArgs, InputArgs, InputArgs, int, list[InputArgs]], InputArgs])

Return type:

None

classmethod default()

Return type:

InputReducer

reducer_fn(input_override, *args)

Parameters:

• acc (InputArgs)

• input_override (InputArgs)

exception InputsLoopFoundException

Bases: LoopFoundException

exception KnotException

Bases: Exception

exception LoopFoundException

Bases: KnotException

class ModuleTarget

Bases: TargetWithNamedInputs, TargetWithNamedOutputs

ModuleTarget(name: ‘str’, module: ‘nn.Module’)

__init__(name, module)

Parameters:

• name (str)

• module (Module)

Return type:

None

module: Module

name: str

exception MultipleExternalNodesException

Bases: KnotException

class Needle

Bases: object

__init__()

Return type:

None

get_node_for_target(target)

Parameters:

target (Target)

Return type:

Node

knot(capture_cache_outputs_predicate=<function always_false_predicate>, early_exit=True, ignore_extra_overrides=False)

Return type:

StitchedModule

stitch(src, dst)

Parameters:

• src (InputDescriptor | OutputDescriptor)

• dst (InputDescriptor | OutputDescriptor)

Return type:

Self

class Node

Bases: object

Node(target: ‘Target’, stitches_to: ‘list[StitchDescriptor]’ = <factory>, stitches_from: ‘list[StitchDescriptor]’ = <factory>)

__init__(target, stitches_to=<factory>, stitches_from=<factory>)

Parameters:

• target (Target)

• stitches_to (list[StitchDescriptor])

• stitches_from (list[StitchDescriptor])

Return type:

None

stitches_from: list[StitchDescriptor]

stitches_to: list[StitchDescriptor]

target: Target

exception OnlyInternalNodesException

Bases: KnotException

class OutputDescriptor

Bases: object

OutputDescriptor(target: ‘Target’, output_name: ‘str’ = ‘’, output_adapter: ‘OutputAdapter’ = <function default_output_adapter_fn at 0x7f8ce6f45d00>, reducer: ‘OutputReducer’ = <factory>)

__init__(target, output_name='', output_adapter=<function default_output_adapter_fn>, reducer=<factory>)

Parameters:

• target (Target)

• output_name (str)

• output_adapter (Callable[[...], Any])

• reducer (OutputReducer)

Return type:

None

output_adapter()

Parameters:

v (Any)

Return type:

Any

output_name: str = ''

reducer: OutputReducer

target: Target

class OutputReducer

Bases: IOReducer

OutputReducer(reducer_fn: ‘Callable[[OutputValue, OutputValue, Optional[OutputValue], int, list[OutputValue]], OutputValue]’ = <function default_output_reducer_fn at 0x7f8ce6fcf6a0>, requires_original_output: ‘bool’ = False)

__init__(reducer_fn=<function default_output_reducer_fn>, requires_original_output=False)

Parameters:

• reducer_fn (Callable[[Any, Any, Any | None, int, list[Any]], Any])

• requires_original_output (bool)

Return type:

None

classmethod default()

Return type:

OutputReducer

reducer_fn(input_override, *args)

Parameters:

• acc (Any)

• input_override (Any)

requires_original_output: bool = False

exception OutputsLoopFoundException

Bases: LoopFoundException

class RemoteDataDescriptor

Bases: ABC

RemoteDataDescriptor(key: ‘str’)

__init__(key)

Parameters:

key (str)

Return type:

None

key: str

class RemotePythonDataDescriptor

Bases: RemoteDataDescriptor

RemotePythonDataDescriptor(key: ‘str’, value: ‘Any’)

__init__(key, value)

Parameters:

• key (str)

• value (Any)

Return type:

None

value: Any

class RemoteTarget

Bases: Target

RemoteTarget(peer_rank: ‘Union[int, Sequence[int]]’, process_group: ‘Optional[torch.distributed.ProcessGroup]’ = None, blocking: ‘bool’ = True)

__init__(peer_rank, process_group=None, blocking=True)

Parameters:

• peer_rank (int | Sequence[int])

• process_group (ProcessGroup | None)

• blocking (bool)

Return type:

None

blocking: bool = True

peer_rank: int | Sequence[int]

process_group: ProcessGroup | None = None

value(name, adapter=<function default_output_adapter_fn>, reducer=OutputReducer(reducer_fn=<function default_output_reducer_fn>, requires_original_output=False))

Parameters:

• name (str)

• adapter (Callable[[...], Any])

• reducer (OutputReducer)

Return type:

OutputDescriptor

class RemoteTensorDataDescriptor

Bases: RemoteDataDescriptor

RemoteTensorDataDescriptor(key: ‘str’, device: “Literal[‘cuda’, ‘cpu’]”, dtype: ‘torch.dtype’, shape: ‘torch.Size’)

__init__(key, device, dtype, shape)

Parameters:

• key (str)

• device (Literal['cuda', 'cpu'])

• dtype (dtype)

• shape (Size)

Return type:

None

device: Literal['cuda', 'cpu']

dtype: dtype

shape: Size

class Singleton

Bases: type

class StitchDescriptor

Bases: object

StitchDescriptor(source_descriptor: ‘IODescriptor’, destination_descriptor: ‘IODescriptor’)

__init__(source_descriptor, destination_descriptor)

Parameters:

• source_descriptor (InputDescriptor | OutputDescriptor)

• destination_descriptor (InputDescriptor | OutputDescriptor)

Return type:

None

destination_descriptor: InputDescriptor | OutputDescriptor

source_descriptor: InputDescriptor | OutputDescriptor

class StitchedModule

Bases: Module

__init__(nodes, capture_cache_outputs_predicate=<function always_false_predicate>, early_exit=True, ignore_extra_overrides=False)

Parameters:

• nodes (dict[Target, Node])

• capture_cache_outputs_predicate (Callable[[str, Module], bool])

Return type:

None

create_input_overrides(values_to_node)

Parameters:

values_to_node (dict[InputDescriptor | OutputDescriptor, Any])

Return type:

PassageInputOverrides

create_output_overrides(values_to_node)

Parameters:

values_to_node (dict[InputDescriptor | OutputDescriptor, Any])

Return type:

PassageOutputOverrides

forward(input_overrides, output_overrides, *args, **kwargs)

Parameters:

• input_overrides (dict[str, Any])

• output_overrides (dict[str, Any])

Return type:

StitchedModuleOutput

exception StitchedModuleException

Bases: Exception

class StitchedModuleOutput

Bases: object

StitchedModuleOutput(captured_inputs: ‘dict[str, InputArgs]’, captured_outputs: ‘dict[str, Any]’)

__init__(captured_inputs, captured_outputs)

Parameters:

• captured_inputs (dict[str, InputArgs])

• captured_outputs (dict[str, Any])

Return type:

None

captured_inputs: dict[str, InputArgs]

captured_outputs: dict[str, Any]

class Target

Bases: object

Target()

__init__()

Return type:

None

class TargetWithInput

Bases: Target

TargetWithInput()

__init__()

Return type:

None

input(adapter=<function default_input_adapter_fn>, reducer=InputReducer(reducer_fn=<function default_input_reducer_fn>))

Parameters:

• adapter (Callable[[InputArgs], InputArgs])

• reducer (InputReducer)

Return type:

InputDescriptor

class TargetWithNamedInputs

Bases: Target

TargetWithNamedInputs()

__init__()

Return type:

None

input(name, adapter=<function default_input_adapter_fn>, reducer=InputReducer(reducer_fn=<function default_input_reducer_fn>))

Parameters:

• name (str)

• adapter (Callable[[InputArgs], InputArgs])

• reducer (InputReducer)

Return type:

InputDescriptor

class TargetWithNamedOutputs

Bases: Target

TargetWithNamedOutputs()

__init__()

Return type:

None

output(name, adapter=<function default_output_adapter_fn>, reducer=OutputReducer(reducer_fn=<function default_output_reducer_fn>, requires_original_output=False))

Parameters:

• name (str)

• adapter (Callable[[...], Any])

• reducer (OutputReducer)

Return type:

OutputDescriptor

class TargetWithOutput

Bases: Target

TargetWithOutput()

__init__()

Return type:

None

output(adapter=<function default_output_adapter_fn>, reducer=OutputReducer(reducer_fn=<function default_output_reducer_fn>, requires_original_output=False))

Parameters:

• adapter (Callable[[...], Any])

• reducer (OutputReducer)

Return type:

OutputDescriptor

default_input_adapter_fn(input_values)

Parameters:

input_values (InputArgs)

Return type:

InputArgs

default_input_reducer_fn(acc, input_override, *args)

Parameters:

• acc (InputArgs)

• input_override (InputArgs)

default_output_adapter_fn(v)

Parameters:

v (Any)

Return type:

Any

default_output_reducer_fn(acc, input_override, *args)

Parameters:

• acc (Any)

• input_override (Any)