Architecture

How Lucid works

Six diagrams traced from the source rather than drawn from memory — how a call reaches a kernel, what backward really does, where tensor bytes live. Each opens in a full interactive viewer: guided chapters, search, focus, route tracing, and export all work in place.

01 · Structure

Lucid Layer Map

Every layer an op crosses: the Python API, pybind11, ops and kernels, the backend dispatcher, and the two Apple stream backends.

Architecture12 components

02 · Execution

a + b, End to End

One a + b traced end to end — dtype promotion, SchemaGuard, vDSP or MLX, and the AddBackward node it leaves behind.

Sequence8 participants · 14 messages

03 · Execution

loss.backward(), End to End

What loss.backward() actually runs: the ones seed, the reverse-DFS walk, version checks and gradient accumulation.

Sequence8 participants · 14 messages

04 · Training

One Training Step

A full iteration from DataLoader to optimizer.step, including the AMP branch that skips a step on overflow.

Workflow10 steps

05 · Data & state

How Tensor Bytes Move

Where tensor bytes enter and leave Lucid — the bridges, the three Storage variants, and which moves copy.

Data flow12 nodes

06 · Data & state

A Tensor's Autograd Life

A tensor's autograd states, from recorded to walked to released, plus the version-mismatch recovery loop.

Lifecycle8 states