Graph Execution

Core graph execution and model compilation.

class nabla.core.graph_execution.GraphTracer

Bases: object

Handles computation graph tracing and cache key generation.

static compute_node_hash(node)

Compute a deterministic hash for a computation node.

static get_trace(nodes)

Perform iterative DFS to get computation trace and cache key.

Returns:

List of leaf nodes (have impl) trace: Topological ordering of all nodes cache_key: Hash key for caching compiled models

Return type:

inputs

class nabla.core.graph_execution.ModelFactory

Bases: object

Factory for creating MAX models from computation graphs.

static create_model(inputs, trace, outputs, dynamic_inputs=None, show_graph=False)

Create a MAX model from the computation graph.

nabla.core.graph_execution.realize_(outputs, dynamic_inputs=None, show_graph=False)

Realize (compute) the given output Arrays.

This is the main entry point for executing computation graphs. Uses compilation caching for performance.

Parameters:

• outputs (list[Array]) – List of Arrays to realize

• dynamic_inputs (list[Array] | None) – Optional list of dynamic inputs for model compilation

• return_trace_inputs – If True, return tuple of (model, trace_inputs) instead of just model

Returns:

Model or None If return_trace_inputs is True: tuple of (Model, trace_inputs) or None

Return type:

If return_trace_inputs is False

nabla.core.graph_execution.get_trace(nodes)

nabla.core.graph_execution.compute_node_hash(node)