tinymesh¶

tinymesh is a tinygrad-native library for learning over sparse structures through space and time.

A graph is the smallest mesh. Nodes carry tensor fields and edges say which nodes may interact. Coordinates, higher-dimensional cells, and time can extend that structure without replacing the sparse core.

Start here¶

Goal	Read
Run one graph end to end	Quick start
Look up a class or function	API
Understand core theory	Topology, message passing, time
See what the evidence supports	Research
Reproduce an observation	Experiments
Read an implementation source	Papers

Public boundary¶

Import	Owns
`tinymesh`	`Graph`, `StaticGraphTemporalSignal`
`tinymesh.nn`	reusable equations and parameters
`tinymesh.datasets`	pinned source validation and tensor lowering
`experiments`	non-runtime data policy, training, controls, and claims

Neural-network components are direct objects with ordinary tinygrad Tensor attributes and __call__. There is no factory, registry, trainer, or PyTorch compatibility surface.

The stack¶

edge facts             source -> target, optional COO values
    |
    v
topology lowering      COO / graph products -> CSR(A) + CSR(A.T) + edge maps
    |
    v
sparse operations      endpoint fields, target softmax, node / edge sums
    |
    v
spatial composition    position -> displacement -> distance -> weight
    |
    v
temporal alignment     Graph + x[T,N,F] + y[T,N,Y]
    |
    v
model composition      graph convolution, attention, recurrence, diffusion

tinygrad owns tensors, autograd, compilation, and device execution. tinymesh owns sparse topology, mesh semantics, and the model compositions that need them.

Components compose in ordinary Python:

spatial state -----+
spectral state ----+--> same-shaped tensors --> sum / concat / attention --> task head
temporal state ----+
future multiscale -+

Task heads, training, and combination policy stay outside the library until a reusable invariant needs an owner.

Current boundary¶

The fixed-topology core supports unit, scalar-weighted, and edge-vector aggregation, first-order gradients, target-normalized attention, sparse graph products, fixed-graph temporal signals, and direct graph-temporal components on CPU and Metal. It stores sparse topology and never constructs dense adjacency.

The private CSR backend uses alpha tinygrad Tensor.custom_kernel and disables default kernel optimization for its data-dependent loop. Batching different graphs, per-lane edge values, changing topology, higher-order gradients, general coordinate-reference machinery, geodesy, and higher-dimensional cells remain outside the current contract.

Concept pages hold durable theory. Research records bind claims to exact revisions and measurements. Source and tests own current behavior.