tensorViz — PyTorch & JAX Model VisualizerFollow tensor shapes through your model, beside the Python you are editing. tensorViz opens a PyTorch
Recorded in 0.1.10; setup not shown. Shapes describe this example input.
PyTorch captured shapes describe one input and execution path. Graph diff compares source without executing either revision; unresolved coverage stays visible. Free to install. Core graph exploration and source forms work without a coding agent. Get started · Watch the walkthrough · Try the Model Explorer · Examples and issues The linked repository contains public examples and support material; the extension's source repository remains private. The browser examples are recorded snapshots and need no installation. Open your own model in VS Code to capture it in your selected Python environment. Your first graph
Capture needs Python 3.10–3.14 with compatible PyTorch and your model's dependencies. The separate tensorViz service needs Python 3.11+; it can use a managed environment after your approval. Inspect source only does not execute your model. No repository checkout, Node.js or coding agent is required. Full setup and troubleshooting · Remote SSH · Python and platform compatibility Use Show Canvas / Hide Canvas or Ctrl+Cmd+G on macOS (Ctrl+Alt+G on Windows/Linux) to toggle the sidebar without losing your selected model. Open Canvas in Browser opens it in a separate browser view. If setup fails, its inline controls let you retry; tensorViz: Select Python Environment changes the model environment and tensorViz: Repair runtime repairs only the separate tensorViz runtime. See the opening guide. JAX and FlaxInstall compatible
An input of shape JAX support currently traces on CPU and labels shapes as inferred. Control flow primitives retain their boundary and report partial coverage; tracing does not claim a branch executed. PyTorch source edits are unavailable for these graphs. The shared capture worker still requires Torch. Explore and editPan, zoom, search, expand nested modules and use the Overview minimap. Select a node and click its file:line link at inspection zoom to open its source. Moving the editor cursor into a mapped layer focuses it on the graph. Parse errors leave the last graph visibly stale until the source is fixed. Turn on Code focus in the Canvas header to open each selected node's source beside the graph; double-click or Enter on a layer moves keyboard focus to it. The chevron beside it can dim unrelated code. With the Canvas maximized, showing source restores the split so the code appears beside it. To use your existing coding agent, run tensorViz: Connect coding agent (MCP).
A status-bar item shows Off, Ready or Connected. The connection panel can configure
a supported agent after approval or provide the URL for a Streamable HTTP client.
The agent's Edits stay within supported constructs in one owning class. Shared definitions affect every instance. Grouping changes parameter paths and checkpoint keys. Source review checks syntax and graph consistency, not runtime tensor correctness. Share a graph fileChoose Export graph in the canvas to save a Capture and settingsAfter changing source, save and choose Refresh graph to capture again. Change inputs edits the saved setup. Generated dimensions are example values, not recovered training settings; a successful capture covers that invocation only. Choose Model creation → Python function in model setup to reuse an existing model builder or return the selected component from a pipeline. Modules can remain in their original files. The factory must return the selected class. See input setup guide. Capture on save is off by default. In a trusted workspace,
The selected project Python runs the model. A separate compatible tool environment runs the tensorViz service and dependency installer; changing model Python keeps that service running. The backend launch token stays in VS Code; the browser receives a limited session credential. The MCP endpoint has no bearer credential and relies on its loopback bind and Host/Origin checks. Optional analyticsWhen analytics are available in your installed release, tensorViz asks before
sharing feature usage, performance and categorized errors. Sharing is off by
default and respects VS Code telemetry permissions. Code,
model contents, paths, prompts and raw exceptions are excluded. A random local
installation ID helps measure returning use; it is not an account or machine ID.
Use TensorViz: Manage Analytics Sharing, View Collected Analytics Data or
Reset Analytics Identity from the command palette. Reset disables sharing;
it does not delete data already received by PostHog. The standalone graph-file
viewer has no analytics collector. The packaged Consented events go to tensorViz's PostHog EU project, with one-year retention under the current plan. IP storage is disabled and person profiles are not created; network infrastructure still processes connection metadata. For private analytics questions or deletion requests, email tensorviz.app@gmail.com. You can obtain your random installation ID from View Collected Analytics Data before resetting it. Send that ID privately if needed to locate your events; never post it in a public question. All model features work with sharing off. SupportThe October 7 compatibility pass verified macOS 26.4 Apple Silicon and Ubuntu 24.04.5 x86_64 on personal RunPod with Python 3.10–3.14 and Torch 2.14.1. Windows, other OS releases/architectures, CUDA and free-threaded Python remain unverified in this pass. Linux installed checks used VS Code with a virtual display; native Linux desktop interaction was not certified. Python 3.15 is experimental. Expression-level source edits require Python 3.11+; Python 3.10 still captures models. See the compatibility details. Managed runtimes are available for macOS Apple Silicon, Windows x64 and Linux x64; other architectures need a compatible existing Python environment. Desktop Remote SSH is supported too: install tensorViz on the SSH host and select Python there. For a coding agent on your desktop, use Forward MCP to This Desktop. See SSH setup and platform details for runtime requirements and other environment limits. TensorViz's own code uses the MIT license. The bundled tldraw SDK and assets
retain their separate licenses. Getting started · Platform compatibility · Questions and bug reports · Troubleshooting |
