PR #1952

open

Non-record: Energy as the missing leaderboard axis - Wh per bpb drop across 5 configs

val_bpb

1.1055

Architecture

Transformer

Optimizer

—

Artifact Size

~1.5 MB

Training Techniques

Evaluation

sliding window eval

parameters: {"stride":64}

Test-Time Training

LoRA TTT

parameters: {"rank":8}

Weight Averaging

EMA

parameters: null

Quantization

GPTQ-lite

bits: null

scope: all

Compression

zlib

level: null

Architecture

depth recurrence

Mini depth recurrence used in the frontier config.

parameters: null

ParallelResiduals

Parallel residual architecture variant used in the frontier config.

parameters: null

LR Schedule

warmdown

parameters: {"warmdown_steps":3500}

Adds energy as a missing leaderboard axis for parameter-golf submissions.
Reproduces five published submissions with NVML hardware-counter energy measurement on 8×H100 SXM5.
Shows that two configs can achieve similar val_bpb with nearly identical energy via different mechanisms.
Quantifies that the frontier config is far less energy-efficient than the engineering tier per unit of BPB improvement.
Measures post-training overhead and shows it can dominate total run energy.