nnsight.envoy#
- class nnsight.envoy.Envoy(module: Module, module_path: str = '')[source]#
Envoy object act as proxies for torch modules within a model’s module tree in order to add nnsight functionality. Proxies of the underlying module’s output and input are accessed by .output and .input respectively.
- path#
String representing the attribute path of this Envoy’s module relative the the root model. Separated by ‘.’ e.x (‘transformer.h.0.mlp’). Set by NNsight on initialization of meta model.
- Type:
str
- _fake_outputs#
List of ‘meta’ tensors built from the outputs most recent _scan. Is list as there can be multiple shapes for a module called more than once.
- Type:
List[torch.Tensor]
- _fake_inputs#
List of ‘meta’ tensors built from the inputs most recent _scan. Is list as there can be multiple shapes for a module called more than once.
- Type:
List[torch.Tensor]
- output#
Proxy object representing the output of this Envoy’s module. Reset on forward pass.
- input#
Proxy object representing the input of this Envoy’s module. Reset on forward pass.
- _call_iter#
Integer representing the current iteration of this Envoy’s module’s inputs/outputs.
- Type:
int
- _tracer#
Object which adds this Envoy’s module’s output and input proxies to an intervention graph. Must be set on Envoys objects manually by the Tracer.
- Type:
nnsight.context.Tracer.Tracer
- property input: InterventionProxy#
Getting the first positional argument input of the model’s module.
- Returns:
Input proxy.
- Return type:
- property inputs: InterventionProxy#
Calling denotes the user wishes to get the input of the underlying module and therefore we create a Proxy of that request. Only generates a proxy the first time it is references otherwise return the already set one.
- Returns:
Input proxy.
- Return type:
- modules(include_fn: Callable[[Envoy], bool] = None, names: bool = False, envoys: List = None) List[Envoy][source]#
Returns all Envoys in the Envoy tree.
- Parameters:
include_fn (Callable, optional) – Optional function to be ran against all Envoys to check if they should be included in the final collection of Envoys. Defaults to None.
names (bool, optional) – If to include the name/module_path of returned Envoys along with the Envoy itself. Defaults to False.
- Returns:
Included Envoys
- Return type:
List[Envoy]
- named_modules(*args, **kwargs) List[Tuple[str, Envoy]][source]#
Returns all Envoys in the Envoy tree along with their name/module_path.
- Parameters:
include_fn (Callable, optional) – Optional function to be ran against all Envoys to check if they should be included in the final collection of Envoys. Defaults to None.
- Returns:
Included Envoys and their names/module_paths.
- Return type:
List[Tuple[str, Envoy]]
- property output: InterventionProxy#
Calling denotes the user wishes to get the output of the underlying module and therefore we create a Proxy of that request. Only generates a proxy the first time it is references otherwise return the already set one.
- Returns:
Output proxy.
- Return type: