Skip to content
Docs
Core Concepts
Model, architecture, layer and module

Model, Architecture, Layer and Module

Every view, selection, and query in Tensormorph operates over the same address space - a hierarchy running from a whole model down to a single scalar. This page covers the coarse end of that hierarchy; the fine end (tensor down to scalar) continues in Tensor, Matrix, Block and Scalar.

The drill ladder

Navigating from an overview to a specific value follows eight steps:

Model → Architecture → Layer → Module → Tensor → Tile/Block → Row/Col/Channel → Scalar
💡

This eight-step ladder is the navigation-level breakdown. A separate, coarser grouping - model, layer, module, tensor, tile, row, element - names the seven kinds of addressable entity underlying it; the two aren't in conflict, they answer different questions ("what do I click through" vs. "what fundamentally has an address").

  • Model - the whole loaded checkpoint.
  • Architecture - the recognized structural pattern the model follows (transformer, MoE, multi-tower multimodal, and so on) - see Model Adapters below.
  • Layer - a top-level repeated unit (a transformer block, for instance).
  • Module - a named component within a layer (an attention block, an MLP, a single expert).

From here the ladder continues into individual tensors and below; see Tensor, Matrix, Block and Scalar.

The Semantic Model Graph

Tensormorph doesn't work with a checkpoint's raw file structure directly - on open, it's read into the Semantic Model Graph (SMG), a typed tree where every node has a stable identity and an address. Every view (the Outliner, the Architecture view, the Comparison Editor) renders a projection of the same SMG rather than maintaining its own separate model of the checkpoint, which is why selection stays synchronized across every view at once.

The Model Adapter

Building the SMG from a raw checkpoint is the job of a ModelAdapter - a component written for a specific architecture that knows how to recognize its layer/module structure (which weights form an attention block, which form an MLP, whether there's MoE routing or multiple modality towers). When no specific adapter recognizes a model, the fallback adapter still builds a usable SMG from generic structural cues, so every checkpoint opens and is inspectable - you just don't get architecture-specific views (like the MoE Expert Galaxy) without a matching adapter.

Shared state across the address space

Whatever you've focused, selected, filtered, or pinned lives in a single shared InspectionContext - every view reads from and writes to the same context, rather than each view keeping its own private notion of "what's selected." Selecting a module in the Outliner is why it's simultaneously highlighted in the Architecture view and summarized in the Inspector panel, with no explicit synchronization step.

Related resources

Compare two modelsTensor, matrix, block and scalar