spt-attn-mincut

Min-cut gating attention operator: dynamic graph-based alternative to softmax attention

Overview

spt-attn-mincut replaces uniform softmax attention with a graph-theoretic min-cut gate. It builds a weighted directed graph from Q*K^T logits, computes a minimum cut via Dinic's max-flow algorithm to prune irrelevant edges, then normalises surviving edges with row-softmax before multiplying by V. This fits into the SPT ML pipeline as an alternative attention mechanism for transformer inference.

Key Features

• Graph construction -- builds weighted directed graphs from attention logits
• Dinic's max-flow / min-cut -- solver for gating irrelevant attention edges
• Temporal hysteresis -- stabilises gating decisions over time to avoid oscillation
• Witness logging -- SHA-256 hashed determinism verification for reproducibility
• Configurable defaults -- MinCutConfig with sane out-of-the-box parameters

Dependencies

Key internal dependencies:

• serde / serde_json -- serialization of config and witness entries
• sha2 -- SHA-256 hashing for witness determinism verification

Usage

use spt_attn_mincut::{attn_mincut, MinCutConfig, AttentionOutput};

let config = MinCutConfig::default();
// Build logits matrix, then apply min-cut gated attention
// let output = attn_mincut(&q, &k, &v, &config);

API Reference

Core Types

Type	Description
MinCutConfig	Configuration for min-cut gating thresholds and parameters
AttentionOutput	Result of an attention computation (gated or softmax)
AttentionGraph	Weighted directed graph built from attention logits
Edge	A single weighted edge in the attention graph
DinicSolver	Dinic's max-flow algorithm implementation
CutResult	Result of a min-cut computation
GatingResult	Which edges survived the min-cut gate
HysteresisTracker	Temporal smoothing of gating decisions
WitnessEntry	Logged witness record with SHA-256 hash