Practical overview
When it fits
Use it when you need intermediate values or a separable graph subpart from a PyTorch model without editing the original model code.
Who it is for
- Python developers debugging or analyzing PyTorch models
- Researchers examining activations, embeddings, or attention-related outputs
- Model developers applying losses to intermediate layers or working with extracted subparts
Documented uses
- List named layers or graph nodes in a PyTorch model.
- Read one or more intermediate layer outputs during a normal forward pass.
- Return selected graph nodes through an extracted model.
- Run or train a model subpart independently for debugging or experimentation.
Requirements & limits
The project is open source under the MIT license. The README documents installation with pip.
Inspectruns the entire model and returns selected layer outputs alongside its normal result.Extractreturns selected graph nodes from an extracted model and can specify new input nodes.- The documented default extraction tracer is PyTorch FX’s
torch.fx.Tracer. - Changing an input node may not remove other dependencies; the extracted model’s summary shows required inputs and outputs.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Inspect selected outputs during a model pass
Return one or more intermediate layer outputs alongside a PyTorch model’s normal output.
Before you start
- Python environment with PyTorch installed
- A PyTorch model and its layer names
Steps
Install the package with the command below.
pip install surgeon-pytorchUse
get_layers(model)to list the model’s layers.Wrap the model with
Inspect(model, layer='layer_name'), or pass a list of layer names.Call the wrapped model with an input; it returns the model result and the selected layer output or outputs.
Expected outcome
The forward call provides the original model output together with selected intermediate activations.
Guide limitations
- Inspect executes the entire model on each forward call.
- The README’s examples assume the requested layer names are present in the model.
Extract a graph node as a model output
Create a model that returns a selected intermediate graph node.
Before you start
- Python environment with PyTorch installed
- A PyTorch model whose computation graph can be traced
Steps
Install the package with the command below.
pip install surgeon-pytorchUse
get_nodes(model)to list intermediate graph nodes.Create an extracted model with
Extract(model, node_out='node_name').Pass an input to the extracted model to obtain the selected node’s output.
Expected outcome
The extracted model returns the requested intermediate node output.
Guide limitations
- The documented default tracer is
torch.fx.Tracer. - Changing an input node may leave other dependencies on earlier inputs; inspect
model_ext.summaryto see required inputs and outputs.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 16 October 2026
- Original website title
- archinetai/surgeon-pytorch · Official homepage title · retained in its source language