Module error_analysis 

Error Accumulation Analysis for Graph-Level Tensor Operations

This module provides tools for tracking and analyzing numerical errors that accumulate during graph-level tensor operations like orthogonalization and decomposition.

Purpose

When performing operations like QR decomposition on weights across a graph, numerical errors can accumulate. This module helps:

• Track per-layer error contributions
• Generate statistical reports on error distributions
• Identify problematic layers with high error

Example

use god_gragh::transformer::optimization::error_analysis::ErrorAccumulator;

let mut accumulator = ErrorAccumulator::new();

// Record errors from each layer
accumulator.record_error("layer_0/q_proj", 1.2e-14);
accumulator.record_error("layer_0/k_proj", 1.5e-14);
accumulator.record_error("layer_1/q_proj", 2.1e-14);

// Generate and print report
let report = accumulator.generate_report();
println!("{}", report);

Structs

ErrorAccumulator

Error accumulator for tracking numerical errors across graph operations

ErrorReport

Comprehensive error report

ErrorStatistics

Global error statistics

LayerErrorStats

Per-layer error statistics