spt-profiler

Memory, power, and latency profiling hooks with CSV emitters for benchmarking attention mechanisms

Overview

spt-profiler provides lightweight profiling infrastructure for benchmarking SPT attention mechanisms and inference pipelines. It captures memory snapshots, latency distributions, and power consumption samples, then exports results to CSV for analysis. A config hashing utility ensures benchmark configurations are uniquely identified for reproducibility.

Key Features

• Latency profiling -- record and compute statistics (mean, p50, p95, p99) over timing samples
• Memory tracking -- capture memory snapshots and generate usage reports
• Power measurement -- track energy consumption with pluggable power sources (including mock)
• CSV export -- write latency, memory, and combined results to CSV files
• Config hashing -- deterministic hash of benchmark configurations for reproducibility

Dependencies

Key internal dependencies:

• serde / serde_json -- serialization of profiling records and configs

Usage

use spt_profiler::{LatencyRecord, compute_latency_stats, write_latency_csv};

let records = vec![
    LatencyRecord { label: "attn".into(), duration_ns: 1200 },
    LatencyRecord { label: "attn".into(), duration_ns: 1500 },
];
let stats = compute_latency_stats(&records);
// write_latency_csv("bench.csv", &records).unwrap();

API Reference

Core Types

Type	Description
LatencyRecord	A single timed sample with label
LatencyStats	Computed statistics (mean, p50, p95, p99)
MemoryTracker	Captures memory snapshots over time
MemorySnapshot	Point-in-time memory usage reading
MemoryReport	Aggregated memory usage report
PowerTracker	Tracks energy consumption samples
PowerSample	A single power reading
EnergyResult	Total energy computed from power samples
BenchConfig	Benchmark configuration for hashing
ResultRow	Combined result row for CSV export