base¶
Bases: Envoy
Base NNsight wrapper around any torch.nn.Module.
Wraps a PyTorch module with NNsight's intervention capabilities, enabling access to and modification of intermediate activations during model execution via the tracing context.
This is the simplest entry point for wrapping arbitrary models.
For HuggingFace language models, use :class:LanguageModel instead.
Example::
import torch
from nnsight import NNsight
net = torch.nn.Sequential(
torch.nn.Linear(5, 10),
torch.nn.Linear(10, 2),
)
model = NNsight(net)
with model.trace(torch.rand(1, 5)):
hidden = model[0].output.save()
| PARAMETER | DESCRIPTION |
|---|---|
*args
|
Positional arguments forwarded to :class:
DEFAULT:
|
**kwargs
|
Keyword arguments forwarded to :class:
DEFAULT:
|