IrekoGPT: Turning Structured Pruning into Post-Hoc Slimmable LLMs

简体中文

L1 nested-subnetwork contract. The original model and throughput are not reproduced.

Paper information

Field	Details
Paper	arXiv v1
Institution	University of Modena and Reggio Emilia (first author).
First public release	2026-10-01 (arXiv v1)
Author code	aimagelab/IrekoGPT.
Local adapter or method key	irekogpt
Local reproduction code	src/auto_research/foundation_latest_20261003.py

Paper method and evidence

Keep SliceGPT's projection basis rather than discarding it after pruning, so one checkpoint exposes nested widths. Calibrate at multiple compression ratios and use ridge corrections for downstream linear layers.

flowchart LR
  P[Retain SliceGPT basis] --> W[Expose nested widths]
  W --> C[Multi-ratio calibration]
  C --> R[Ridge correction]

Original paper figure

Figure 1 shows the nested-width architecture. From the original paper; copyright remains with the authors.

Core equation

\(W_k=P_{:k}^TWP_{:k}\); widths share one nested basis.

Reported results

Preliminary author experiments on Llama and Qwen report better results than naive PCA slimming at high compression.

Local reproduction

Compare against the mechanism disabled or in its default state. The three-seed diagnostic is diagnostic_only=true; no formal relative gain or capability ranking.