TensorViz

Appearance

View Markdown ↗

Graph files, walkthroughs and discussions — reference ​

For the short walkthrough, see the user guide. This page keeps the detailed behavior and implementation notes.

Contents ​

A .tensorviz.json file contains an architecture snapshot and optional walkthrough notes and discussions. It preserves collapsed contents, connections, public parameters, available shapes and layout. The architecture is read-only; annotations are editable. Nothing updates from Python automatically or executes the model. Project frames and read-only snapshots cannot be ungrouped; expanding them changes only the view.

Version 4 files contain canonical captured computation. They preserve every retained operation/call, ordered argument and output tree, typed ports, tensor/state identity, aliases, mutation effects and coverage. Select a captured layer and choose View captured details to inspect those facts entirely offline. The header distinguishes a complete capture of one execution path from a partial capture; neither certifies other inputs or branches.

Live Export graph, file import, and offline inspection are available. Export prepares the exact displayed capture through the connected VS Code host without executing the model. Retargeting, eviction, disconnection or invalid/private evidence refuses the export. Version 4 comparison uses canonical arguments, tensor uses, sharing and state, with explicit unresolved correspondence. Static/legacy export, import, notes and comparison retain their existing behavior.

Legacy snapshots preserve trace capture limits. A Depth-limited block has no captured internal trace contents and may carry measured boundary shapes if it was observed; it does not mean the Python module has zero layers. Contents unavailable disables expansion when no members are present, including when restoring a saved open view. Older empty trace snapshots show the generic contents-unavailable message because their capture depth is unknown. Existing captured members still expand locally, and repeat counts still refer to the same canonical representative.

Measured boundary describes shapes observed at that node for the traced input, not validation of every child. Static, declared, unobserved and stale evidence remain distinct. Export retains bounded-capture metadata, shapes and repeat identity; executable custom Python and internal import/source paths remain in the workspace and are excluded by the existing portable-file privacy rules.

Snapshots also preserve optional exact parameterCount and complete outputTensor metadata (shape, dtype and logical bytes). Zero is a known value. The byte overlay is an output tensor footprint, not peak memory; see overlay evidence and limits. Stale exports remain uncoloured. Versions 1, 2 and 3 still open; older files lacking these fields retain their shapes without inventing counts or assuming float32 bytes. The importer validates the optional fields and retains only the closed tensor metadata subset, preserving existing source and private-path filtering.

Export and exchange ​

  1. In Open Canvas in Browser, choose Export graph after the model finishes loading.
  2. Choose Open graph file viewer and open the downloaded file.
  3. Add notes or replies, then Download graph file and send the updated file.

Recipients need the graph file and a static viewer page, without Python, Torch, VS Code, an account or an API key. Your localhost URL is not a share link. Viewer comments travel in files: the viewer has no online storage, simultaneous collaboration or automatic merging of separately edited copies. Use one exchanged copy when collecting replies.

The VS Code Canvas offers Add to tensorViz in Export's place: it uploads the same file to your tensorViz account, private until you choose who can see it on tensorviz.dev; see Add graphs to tensorViz.

Notes remain in the tab until downloaded. Invalid files preserve the current graph and notes. Opening another file asks before discarding edits; reload/close triggers the browser's unsaved-change warning. A newer file choice cancels any earlier pending replacement. Post or discard draft messages before downloading.

Explore the snapshot ​

Navigation, both searches, metric overlays and the connection guide share one vertical controls column. Results and legends move the controls below them; the column scrolls when space is limited, without resizing the graph viewport. Fit, zoom, Blocks and Expand all change presentation only.

Search layers finds a layer or module by name/type; choosing a result or pressing Enter reveals its containing modules. Escape clears this search. Search the graph supports structured queries; its results select the visible containing block without changing expansion. Parameters and Output tensor bytes retain the recorded evidence and legend limits described above. Saved views and discussions remain available beside these controls, and architecture stays read-only.

Share a specific view ​

Choose Saved views, select a layer, block or real connection, and arrange the canvas. Share this view stops any camera motion, captures the view and makes it the opening view. Give it a title and optional note, then Download graph file and send that file.

The recipient opens at that view, with its selection, expanded modules and camera. The camera keeps the same center on a different screen size. Named views can be revisited, reordered, updated or deleted; Make opening view chooses which appears first, and Clear opening view returns to the usual graph overview. Changed or missing targets remain marked after an architecture update. These are views inside a file, not hosted links. Walkthroughs and discussions travel alongside them.

Guided walkthroughs ​

Choose Walkthrough. Select a layer or block, arrange and expand the graph, then Add current view. Edit its title and note; Update saved view recaptures the selection, positions, expanded blocks and camera. Move steps up/down or delete them.

Readers use Start walkthrough, Previous and Next. Each step restores its saved view. They can pan and explore freely, use Show saved view to return, or Exit walkthrough to continue exploring. Notes are plain text, including line breaks.

Discussions on layers and connections ​

Select a layer and choose Discuss layer, or click an expanded connection and choose Discuss connection. The Discussions panel also lists a selected layer's connected edges. Display-only edges created for collapsed modules cannot anchor a discussion; expand the module or select a connection from that list.

Choose Start discussion, enter a name and message, then Add message. Readers can Reply, Resolve discussion, Reopen discussion, and Show current target. Names are unverified labels, not accounts. A draft keeps its original target even if you select another layer. Download and exchange the updated file to share replies.

Update the architecture while keeping notes ​

Update architecture takes architecture from a newer graph file and retains the current walkthrough and discussions. The replacement must contain no annotations; annotated replacements are refused so neither set of notes is silently discarded.

Discussion anchors keep their original snapshot date and semantic context. Parameter, module-content or connection changes are marked Context changed. Deleted targets are Target missing; non-unique connection matches are Target ambiguous. Original details remain available. Layout, expansion and edge ordering alone do not change context. Renamed layers are not guessed or rebound. Walkthroughs also flag changed or missing targets; review the notes before updating a step's saved view.

Compare two snapshots ​

Open the newer file, choose Compare, then Choose Before file to select the older snapshot. For version 4 captures, the report separates computation changes, exact tensor-use changes, runtime observations and execution conditions. Missing correspondence or partial evidence cannot certify equality. Static/legacy comparison lists added/removed layers and connections, parameters and other structure.

Show before and Show after focus a changed layer or connection at a readable scale beside the comparison panel. View before/after switch the canvas version. Its header always labels the side. Canonical switching follows a corresponding selected card or destination operand slot, even if capture IDs or the operand's producer changed. Without a confident counterpart, the destination restores its own view. Inspect before/after opens exact retained records without switching the canvas, including module records and large argument trees. Previews are limited to 4,000 characters; comparison uses the complete facts. Each section initially shows 25 records, with an explicit Show more records action. Closing comparison returns to the current After graph, preserving its notes and message drafts. Invalid comparison files keep the previous comparison available. Comparison is temporary and read-only; neither file's annotations are merged.

Matching and evidence ​

Comparison matches portable layer IDs, which retain source hierarchy when available. It compares operation, kind, parameters, module membership, ports and connections. Connection order and generated edge IDs, layout, annotations and export dates do not count as architecture changes. Duplicate connections are counted individually.

Unmatched IDs remain added/removed layers. Structurally identical unmatched layers are flagged as uncertain possible renames; tensorViz never silently pairs them. Different model names also produce a warning. Recorded shapes and tracing status are separate evidence: different inputs or tracing availability can change those without changing the architecture. Comparison does not execute either model. This matching describes versions 1–3. Capture-scoped event IDs must not be matched using those legacy rules. Version 4 uses conservative source/ownership/dataflow correspondence and exact argument/value-use comparison. Repeated operations and numbered containers can remain unmatched: unmatched does not mean added or removed. A matching captured computation does not prove numerical equality or behavior on other inputs. Mixed canonical/legacy files report comparison unavailable. The rules for a captured computation are in coverage and correspondence.

File format and limits ​

Version 3 uses format: "tensorviz.graph", version, snapshotId, exportedAt, status, warningCount, document and collaboration. The document is a validated portable subset of GraphDocumentV2. Collaboration contains one ordered walkthrough, discussion threads, named saved views and an optional opening-view id. Version 1 and 2 files remain readable; older viewers reject version 3 instead of dropping saved views. Saving notes preserves the architecture export time. Saved views retain their original target context, selected nodes or a real semantic connection, camera, expansion and optional capture viewport dimensions. The opening-view id must reference a saved view in the file. Snapshot labels identify context and do not authenticate a file.

Version 4 adds capture: PortableCapture v1, versioned independently from the file. Optional initializer-entry scalar snapshots survive export/import and appear as observations in comparison. Older captures have no such evidence. Opaque initializer objects carry no payload or type name and do not invalidate complete forward capture. Sensitive numeric parameter names and path-like variadic keys refuse lossless export; benign declared parameters such as Embedding._freeze remain portable. Its document is checked against the payload before rendering: all captured events, exact operand paths and producers, original edge/port identities, invocation ownership, tensor facts and declared metrics must agree. Source paths become deterministic aliases, source bindings follow them, and live workspace/editor identity is removed. Recognized private/path-like argument literals refuse export/import instead of losing an operand. Version 4 does not apply legacy parameter filtering to captured arguments. Its display-only summaries are derived separately from retained evidence. The same Python-generated structural validators are precompiled at build time, so the viewer keeps its existing CSP without unsafe-eval. Relational/privacy checks share a corpus of acceptance and refusal cases with the Python validators. Imported evidence is not authenticated.

Captured discussions include exact record/port/value context, including literals and sharing, within the existing anchor-size bound. A changed argument invalidates that context; layout and notes do not alter the retained computation. Large target contexts can exceed the existing note-anchor bound and must be narrowed to a smaller target.

Imports are limited to 96 MiB, 20,000 nodes, 50,000 edges, 64 hierarchy levels, 8 parameter nesting levels, 50 walkthrough steps, 50 saved views, 200 threads and 1,000 messages. Notes/messages allow 4,000 characters each. Invalid identities, parents, coordinates, unsafe keys and unsupported versions are rejected before rendering. Version 4 permits up to 32,768 declared ports per node within the same total file limit; versions 1–3 keep their 64-port bound. Runtime capture budgets remain independent from file size, document size and display collapse. Limits refuse without truncation.

Source text, local paths, arbitrary metadata, credentials, live-session identities, weights and data files are excluded from architecture export. Internal/source/path parameter fields and absolute local path values are omitted. Ordinary public values, labels, names and annotation text remain visible: review them before sharing. Static/traced/stale status describes the original snapshot, not a new runtime check. Stale live graphs export as stale. Canonical graphs require a live connection to retrieve their retained capture; disconnected static/legacy graphs export as stale. Portable preparation accepts at most 16 MiB of retained input and transfers at most 16 MiB of detached capture. One preparation runs per host; the browser waits at most 15 seconds. Oversized payloads refuse without truncation.

Build and serve the viewer ​

From the repository root, run npm run build:viewer --prefix canvas. Serve canvas/viewer-dist/ with a static HTTP server, for example npx --yes http-server canvas/viewer-dist -p 4175 -c-1. It includes local tldraw fonts/icons/translations and makes no backend or CDN requests. Files are read locally in the browser. This is a static website build, not a self-contained HTML export. The extension bundles the same viewer under media/app/viewer/.

JAX exports keep inferred shape provenance. Import refuses a file that calls those shapes measured. Large graphs lay out with iterative ranking above 1,000 nodes or 4,000 edges so the layout does not recurse without bound. All nodes and edges are retained. See licensing before you ship a viewer build.