Install model libraries

Checked when you open a model. Recovery after a bad environment is in environment recovery. Opening steps are in open a graph.

1. The first service install asks before it downloads a runtime. Cancel does not consent.
2. Open as Graph checks the model's imports in the capture interpreter before it runs.
3. Declared dependencies in requirements.txt or pyproject.toml become an install plan. Unknown import names are reported, not guessed.
4. Install dependencies is additive and keeps the existing Torch or CUDA build. You approve the versions. A source build is a separate review.
5. Inspect source only stays available without those imports. It does not measure shapes.

First service installation asks Install tensorViz’s runtime to open graphs? The native modal's detail separates Python's download cap, additional dependency downloads and estimated disk use. It provides one Install action and native Cancel; dismissal grants no consent. Runtime download failures retain their sanitized underlying cause and offer Repair runtime/Show Log, with service fallback probes listed separately. Project Python is not repaired by this step. Temporary download failures get at most three attempts with cancellable waits. Checksum, untrusted redirect, oversized archive and certificate failures are not retried; exhausted attempts keep a clear failure and manual recovery action. Setup progress also reports completed stage durations. For local troubleshooting, runtime/last-setup.json in tensorViz's extension storage records stage timings and ready/failed/cancelled outcome. It contains no project paths or source and is not uploaded. Setup duration is separate from time to the first rendered graph.

Normal Open as Graph checks saved model imports in the selected capture interpreter before preparing inputs or running the model. This applies to both an existing project environment and Use a separate tensorViz environment. The managed environment includes TensorViz/PyTorch; graph opening now offers to install the additional model libraries there, without requiring a coding agent. Switching environments with a graph selected reconnects Canvas and resumes that model's opening, including its dependency check. Explicit source-only inspection remains source-only.

The native starting screen checks only tensorViz's owned runtime automatically; project Python and capture-library checks start after the user selects a model. At explicit model opening, saved/configured choices remain authoritative. A compatible VS Code selection or a sole detected environment in the model's project/owning folder can be reused and saved for that workspace without another chooser. Ambiguous choices, missing libraries and broken saved paths still lead to inline recovery; another workspace folder's Python is never adopted automatically. Readiness metadata alone does not certify a capture: model dependencies and the private worker are checked again before execution. See unified onboarding.

TensorViz reads requirements.txt and [project].dependencies in pyproject.toml for names and version constraints. Known mappings cover common PyTorch libraries (including Transformers, Diffusers, timm, Lightning, einops and their usual helper libraries) and import aliases such as PIL → pillow. Unknown import names are reported rather than guessed as PyPI packages: declare the correct distribution in the project requirements, save, then Refresh. URLs, editable/path installs, requirements-file includes and conditional dependency markers need project setup outside this guided installer.

If a missing import is already declared with an unsupported source or condition (for example a pinned Git dependency), the model check says that the project declaration exists rather than asking the user to add it again. A ready tensorViz runtime supplies capture libraries, not every model's dependencies. With no supported installation plan, recovery promotes Use Agent (or Copy MCP investigation request in the standalone page), while Browse… and Refresh remain available for a prepared project environment. Inspect source only is also available without model imports or dependency installation; it does not measure shapes or guarantee that every model can be projected from source.

The bounded static check follows local imports, including src packages and sibling modules and parent package initializers, without importing the model. Imports inside functions, conditions or try blocks remain optional during this check. A structured missing-module error during input preparation or before a capture can confirm an additional required library and offer installation. After installation, opening checks the environment and saved profile again before retrying. A failed forward with an existing partial capture is retained for inspection.

The native review lists exact resolved package versions and the Python executable. Install dependencies uses the same additive, hash-verified installer as MCP: existing packages, including Torch/CUDA, are preserved. Compatible transitive dependencies are included. Source builds require a separate review. Cancellation, version conflicts, unavailable packages and changed source/environment remain visible failures; Refresh retries. Model libraries are verified after installation before opening continues. Static inspection is bounded to 128 files, 256 KiB per file and 2 MiB total; dynamic imports cannot all be predicted up front.

An installed library can still be incompatible with the model. Import-time code errors now report library compatibility or package initialization rather than claiming the package is missing. TensorViz can offer tested compatible versions in a separate copy of the environment, respecting project constraints and keeping Torch/CUDA fixed. One native review approves the resolved versions; the model must import successfully before that environment is selected and graph opening resumes. See guided environment recovery.

Loaded ScriptModules

An ordinary forward capture does not install optional activation or gradient-statistics hooks when those statistics are disabled. This permits a model to call a loaded TorchScript module without failing on unsupported forward-hook registration. It does not enable training diagnostics or add capture coverage for opaque compiled operations.