Skip to content
Docs
Core Concepts
Tensor, matrix, block and scalar

Tensor, Matrix, Block and Scalar

This page continues the address space from where Model, Architecture, Layer and Module leaves off - the fine-grained half of the same eight-step drill ladder:

… → Module → Tensor → Tile/Block → Row/Col/Channel → Scalar

Tensor

A named weight or activation tensor within a module - the level at which shape, dtype, and stride become meaningful (see Shape, Rank, Stride and Layout), and the level a weight diff is usually framed at.

Tile / block

A bounded rectangular region within a tensor - the unit Tensormorph actually streams, computes statistics for, and renders at. A tile isn't a fixed size; it's sized appropriately for the current level of detail, so "one tile" means something coarser when you're viewing a whole layer and something finer once you've zoomed into a specific region. See Level of Detail and Streaming for how tile granularity adapts to zoom.

Row / column / channel

Below a tile, a single row, column, or channel - a one-dimensional slice through a tensor. This is the level the Signal and Distribution Editor's 1D views (Raw Value Plot, Layer Profile, Expert Profile) operate on, distinct from the 2D tile-level views.

Scalar

A single number: one element of one tensor. The Scalar Inspector is the dedicated view for this level, and it's also the finest-grained target for a probe or a runtime breakpoint (see Runtime Traces).

Every level shares one addressing scheme

💡

A scalar's address is a strict refinement of its tensor's address, which is a strict refinement of its module's address, all the way up to the model. There's no separate coordinate system per level - zooming from the whole model to one scalar is narrowing the same address, not switching between unrelated schemes.

This is what makes a TQL expression like model["layer.12"].tensor("*.q_proj.weight")[3, 7] a single coherent path rather than a a join across separate systems, and it's why selection stays valid as you switch between the 3D, 2D, and 1D views at different points along the ladder - the address doesn't change, only which view is rendering it.

Related resources