talon-graph 0.0.1

TALON Graph: Graph partitioning, core mapping, and spatial routing for Talon IR

TALON Graph Partitioning and Mapping

Graph partitioning, core mapping, and spatial routing for TALON IR graphs. Takes IR graphs and produces placed graphs with core assignments and routing tables for hardware deployment.

Installation

uv add talon-graph

Features

• Graph Partitioning: EdgeMap, spectral bisection, and greedy algorithms
• Core Placement: Minimize spike-hop distance across hardware mesh
• Spike Routing: Forward and feedback routing tables via RustworkX shortest paths
• Resource Allocation: Per-core memory budgets (weights, neuron state, routing buffers)
• DAG + Cyclic Support: Auto-detects topology, co-locates SCCs on same core

Quick Start

from talon import ir, graph

# Load a TALON IR graph
g = ir.read("model.t1c")

# Define hardware constraints
from talon.graph import HardwareSpec
hw = HardwareSpec(max_neurons_per_core=256, sram_bytes_per_core=65536)

# Partition and route
placed = graph.partition(g, hw, algorithm="edgemap")
placed = graph.route(placed, hw)

# Inspect assignments
for name, node in placed.nodes.items():
    if hasattr(node, 'core_id') and node.core_id is not None:
        print(f"  {name} -> core {node.core_id}")

License

MIT