Repository navigation
tensorViz 0.1.43
-
Restore mixed module training modes after inference capture, including custom training-mode side effects, shared submodules, and failure cleanup.
-
Harden graph opener allocation and retired-session handling so repeated graph opens do not exhaust capacity or replay stale updates.
-
Show searchable model and supported effort choices beside the chat composer. Add a model/effort split menu to Explain this view, with separate saved preferences and the actual execution choice visible during and after each pass.
-
Keep graph cards compact at every zoom, wrap full technical names, and move ordinary parameter lists into selection details. Size Input/Output cards from actual rows instead of leaving empty bodies; keep SVG and connection geometry aligned.
-
Resolve selected-node signatures and author docstrings from source without an agent. Prefer source documentation over standard-layer reference text, preserve full details and citations, and keep ambiguous or dynamic targets unresolved.
-
Explain selected nodes with compact source facts, standard-layer descriptions and optional source-grounded agent interpretations. Keep exact callables, execution evidence and linked sources inspectable; invalidate explanations when source, capture or agent settings change.
Published on the VS Code Marketplace. This download is the exact Marketplace package.
Source: edf5600
SHA-256: 521d0c753e921d29ce820cd4ef3fd0fa5fcb3b58295f297826c0b0c90b3c1139