Skip to content

Open a model ​

Open a project… offers three tabs: Your code, Hugging Face, and Templates. Home’s Get started actions take you straight to the matching tab.

Your code ​

Use Choose local folder on desktop. In a browser, enter an absolute path on the service machine, then choose Open local folder. For GitHub, paste a repository URL and choose Open repository.

The GitHub repository is cloned to a managed folder on the machine running klars. Attaching it to a chat gives that chat access to the source; it does not execute the model by itself.

When the model needs configuration, use Graph actions → Model setup. Check the Python interpreter, model entry point, and example inputs. Existing saved profiles can supply this information.

Capture executes your code

A local-code capture runs the model in your chosen Python environment. Review an unfamiliar project and its dependencies before you run it.

Hugging Face ​

Search the public Hub from the Hugging Face tab, or paste a model URL or owner/name ID. Search looks for Transformers models with safetensors weights; public search needs no token.

Select a result and review its inferred inputs. If klars cannot infer them, enter each forward argument’s name, shape, type, and fill value.

Shapes only ​

Choose Open model to trace the architecture from its configuration using fake tensors. This path does not download model weights or capture real tensor values. The graph is labelled Shapes only.

klars uses the installed Transformers implementation and does not run Python code from the model repository. Unsupported architectures may need to be opened from their implementation project instead.

Pretrained weights ​

Open with pretrained weights downloads a safetensors file, up to 2 GB, and captures real values. This takes more time, disk space, and memory. A random-weight build estimated above 4 GiB is refused.

For gated models, accept the model’s conditions on Hugging Face first, then provide a read token in the picker. It is used for that model and is not saved.

Transformers must be installed in the Python environment klars uses. A missing installation is named in the failure card, with an install offered for review.

Templates ​

Search or browse Templates. Select a model, check its declared inputs, and choose Create project to generate a new project and capture its model on the CPU.

When adding a template to an existing project, review the generated Python and CPU profile before saving.

Understand the capture label ​

LabelWhat you can infer
Recorded sampleA saved example graph. Opening it runs nothing.
Shapes onlyArchitecture, shapes, and connections; no real tensor values.
A completed model captureEvidence from the selected model, environment, and inputs. Inspect its provenance before comparing results.

Next, explore the graph and inspector.

Your models. Your machine. Your next idea.