Tensor Volume Editor
Called the Tensor Volume view in Tensormorph's own View menu, one of eight 3D views (alongside Architecture, Layer Stack, MoE Expert Galaxy, Parameter Map, Diff, Morph, and Runtime Flow). "Editor" and "view" refer to the same panel throughout this documentation.
Where the Architecture view renders a model's structure, the Tensor Volume view renders a single tensor's values as a spatial volume - useful once a tensor is too large, or too high-rank, to read meaningfully as a flat 2D grid.
What it renders
A selected tensor is laid out across the address grid described in Snapping to the Address Grid, scaled to its own shape, with the same shading modes available in the Architecture view (magnitude, sparsity, gradient-flow) applied per-element rather than per-module. For a rank-4 tensor - an attention score tensor, for instance - the view resolves the extra dimensions with a selector (batch/head as a chosen slice, or small-multiples of 2D slices across an axis), since no single 3D projection preserves every axis of a four-or-higher-rank tensor without discarding something.
Reading a matrix multiplication spatially
A MatMul node's inputs and output are Tensormorph's most-reused spatial convention: the two input matrices render as perpendicular planes, and the output renders as a third plane, with the shared contraction dimension highlighted as the common edge between them. Once you've learned to read this once, it generalizes to any MatMul you encounter elsewhere in the Runtime Flow view.
When to use this vs. Matrix/Heatmap
The Matrix and Heatmap Editor is the 2D counterpart to this view - better suited to a single 2D slice you want to read precisely (exact values, precise tile boundaries), where the Tensor Volume view is better suited to seeing a whole tensor's shape and value structure at once, at the cost of some precision per element.
Related resources
- Architecture Editor - the structural (module-level) counterpart to this value-level view.
- Matrix and Heatmap Editor - the 2D counterpart for precise, single-slice reading.
- Shape, Rank, Stride and Layout - the metadata that determines how a tensor's dimensions map onto this view.