tensi for VS Code
Hover any PyTorch tensor and see its shape, with every dimension named after the variable that sized it.
No annotations needed, and your files are never modified.

That's nanoGPT, unmodified. tensi works out that config.n_embd is a dimension and follows x in from
the modules that call it. Shape errors are underlined in place, in the same names:
Argument `Tensor[[b, 7, v]]` is not assignable to parameter `x` with type `Tensor[[b, 7, d]]`
Where a size really comes from the data (the batch, the image), tensi says so and suggests the one
jaxtyping annotation that would name it, e.g. x: Shaped[Tensor, "batch seq in_features"].
Getting started
- Open a folder with PyTorch code. tensi uses the project's own
.venv (what uv sync creates), else
the interpreter picked in the Python extension.
- If tensi isn't installed there yet, click Install tensi. In a uv project that runs
uv add --dev tensi-shapes, so uv sync keeps it; otherwise it installs into the project's
environment. Install for all projects (uv tool install tensi-shapes) is the other option: tensi
then works everywhere and still reads torch and your packages from each project's environment.
- Hover a tensor. The tensi item in the status bar spins while it works out a file (a few seconds
the first time) and settles when shapes are ready.
Switching projects, changing the interpreter, or running uv sync is picked up on its own: there is
nothing to set and no server to restart. Any Python from 3.10 up works.
If something looks off, run tensi: Check setup from the Command Palette. It checks Pyrefly, the stubs
and the interpreter, runs one real probe, and prints a fix for anything that fails.
tensi runs beside Pylance. By default it only adds hovers that carry a tensor shape, and only reports
shape errors, so nothing is duplicated.
Settings
| Setting |
Default |
|
tensi.pythonPath |
empty |
leave empty to follow each project; set it only to pin an interpreter |
tensi.hover |
tensors |
only hovers that carry a shape; all for every Pyrefly hover |
tensi.diagnostics |
shapes |
shapes, all or off |
tensi.inlayHints |
off |
shapes inline, after each variable |
tensi.shadowWorkspace |
on |
shapes survive imports between your files |
Commands
- tensi: Check setup runs
tensi doctor on the workspace.
- tensi: Restart server
Requirements
Python 3.10 or newer with torch installed. The same package gives you a CLI: tensi peek model.py
lists every tensor's shape, and tensi check src/ exits 1 on a shape error, for CI.
tensi is built on Pyrefly's experimental tensor-shape support. Source, docs and issues:
github.com/tristan-farran/tensi.