API reference

Holds traversal state and emitter interfaces for backward rules.

One node of a parsed configuration document.

The seam every training loop reads through.

Generates concentric rings labelled by ring index.

Generates deterministic, linearly separable multi-class batches.

Assigns every node of a differentiated graph a region key.

Opens a named scope for its own lifetime.

Cuts a graph only where the caller opened or closed a scope.

Splits the nodes traced into one graph into segments, by the scope open at each.

The exception every SushiAI failure surfaces as.

Owns one Storage per arena in a plan and hands out views into them.

Summarises the casts and value retitlings inserted by the pass.

Manages the lowered and compiled execution plan for a graph.

Selects fused forms based on device profile and execution mode.

Records which subtrees collapsed and which were refused.

A single-assignment DAG of tensor operations.

Emits shape-checked IR nodes into a Graph.

Defines the interface for bind-time selection of fused forms.

Interface selecting execution element types per operation.

Device work recorded into the compiled step rather than flushed alone.

Translates IR graph operations into recorded SushiBLAS work.

The planner's output: one Placement per value, plus arena sizes.

Selects lower precision for compute-bound operations.

Owns device memory buffers backing graph values across steps.

Preserves declared precision across all graph operations.

Deserializes checkpoint headers and records from an input stream.

Serializes checkpoint headers and tensor records to an output stream.

One layer of a model whose architecture came from a file.

Adapts a compile-time Module to the runtime ILayer interface.

Holds a runtime sequential list of polymorphic layers.

Holds a learnable tensor's metadata and graph value identifier.

Non-owning type-erased parameter callback wrapper.

Applies layers in sequence and indexes parameter paths.

The update above, applied to a fixed parameter set.

The state machine, its device-resident scale, and the screen.

Host-side state machine tracking dynamic loss scale adjustments.

Common base class providing parameter update recording and lifecycle hooks.

The update rule above, applied to a fixed parameter set.

A tiny device-resident float vector, rewritten from the host per step.

Represents tensor extents as a fixed-capacity value type.

Multi-class cross-entropy over one-hot targets, and accuracy.

What a model is being trained to minimise.

The graph values an objective declared, by role.

Owns the compiled step and drives it over a dataset.

Enables Shape as an unordered-container key.

The knobs, with the defaults the derivation above justifies.

What a check found.

Records, for every node differentiate() appended, the forward node it serves.

One operation's backward rule, paired with its identity.

Where a run's batches come from.

One layer as the file spelled it: a type, and everything else.

A parsed architecture.

One training run, end to end.

One minibatch, resident on the device.

One row of DATASET_REGISTRY.

Where the IDX files are and how much of them to use.

How to shape the concentric-rings problem.

How to shape the generated classification problem.

Holds the extents and the stride of one feature level.

Holds where one level's rows sit in a [B, A, C] array over all anchors.

Assigns every node of one differentiated graph to a region.

Describes one region of a partition.

Holds what two steps must agree on for a region to be reused.

Names a maximal run of consecutively traced nodes with one innermost scope.

One allocation the plan asks for, and what it holds.

Stores inline scalar configuration values for an operation.

Holds the three gradients produced by batch normalisation backward.

Holds normalised output tensor and saved batch statistics.

What the pass is allowed to change.

Defines a 2-D window geometry in eight configuration slots.

One chain that matched structurally and was refused anyway.

One place a graph disagrees with a precision policy.

One chain the pass replaced with a single node.

What the lowering should spell, and why.

What the pass is allowed to do.

One fused node, described in the terms a selector reasons in.

One selector decision, kept so the lowering can be interrogated.

One conversion the pass added.

One operation: an identity, its operands, its results.

Where one value's elements live.

The three knobs the planner has, and nothing else.

Holds pooled tensor output and argmax indices.

What the operation should compute in, and why.

One operation, described in the terms a precision policy reasons in.

Holds the height and the width a spatial operation yields.

One single-assignment tensor-shaped value in the graph.

One tensor's identity in a checkpoint: its name, type and extents.

Averages every window of an NHWC image, dividing by the full window area.

Convolves an NHWC activation with a bank of learnable filters.

Computes mean cross-entropy over a batch from raw logits.

Folds [N, d1, ..., dk] into [N, d1 * ... * dk] without moving data.

The Gaussian error linear unit, tanh approximation.

Averages an [N, H, W, C] activation over H and W, giving [N, C].

How a whole model is initialised.

One row of LAYER_REGISTRY.

Applies x where x is positive and slope * x elsewhere, elementwise.

Represents an affine transformation: Y = X * W^T + b.

Takes the maximum of every window of an NHWC activation.

One entry of a layer's fields() list: a name and a member.

max(0, x), elementwise.

The logistic sigmoid, elementwise.

The hyperbolic tangent, elementwise.

Views an [N, height * width * channels] value as [N, height, width, channels].

Repeats every pixel of an NHWC image over a scale-by-scale block.

Describes an activation, its gradient operation, and gradient source.

One operation's identity: the string and its hash, kept together.

AdamW's hyperparameters, with the reference defaults.

Configuration constants governing dynamic loss scaling.

One row of OPTIMIZER_REGISTRY.

One parameter and its gradient, as the optimizer sees them.

SGD's hyperparameters, with PyTorch's defaults where they have one.

One named number a step produced, whatever it means.

One row of OBJECTIVE_REGISTRY.

A parameter's name and the two graph values behind it.

What a whole fit() produced.

What one step produced.

The four graph values a training loop has to know by name.

How the loop behaves.

A traced, differentiated model, ready to lower.