# Source figure record

Checked 2026-10-13 UTC (2026-10-12 local). Original PDF remains only in temporary study storage and is not tracked or deployed.

## gqa-head-sharing.png

- Source: Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, Sumit Sanghai. *GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints*.
- Edition: EMNLP 2023 conference edition, December 6–10, 2023; printed pages 4895–4901.
- Canonical page: https://aclanthology.org/2023.emnlp-main.298/
- Exact PDF: https://aclanthology.org/2023.emnlp-main.298.pdf
- Source location: Figure 2, PDF page 2 (printed page 4896).
- Original PDF SHA-256: c52de1d3339174fbf93b588aa4cc9bf36092c865f2358ece3999de152f57455c.
- Asset: gqa-head-sharing.png, 1750 × 592 pixels, PNG.
- Asset SHA-256: 97da46bc4a9dea622d345a68faaf331e537890a79e84267a40c9329c1bf12d41.
- Extraction: Poppler pdftoppm at 300 dpi, page 2, top-left pixel crop x=375, y=292, width=1750, height=592.
- Changes: surrounding prose and whitespace cropped; original panel layout, scientific colors, axes, legends, labels and arrows retained. Caption rewritten with full attribution beside the page figure.
- Review: original complete page and final high-resolution crop visually inspected. No source figure label is clipped. Parent desktop/mobile/theme/Focus, keyboard and complete text-export review passed on October 12, 2026 local time.
- Site placement: src/content/docs/edge-ai/efficient-models.mdx, figure-2; minWidth 760 pixels for mobile panning.

## Reuse basis

© 2023 Association for Computational Linguistics.

CC BY 4.0: https://creativecommons.org/licenses/by/4.0/

License evidence: https://aclanthology.org/faq/copyright/

The ACL conference edition, not the arXiv distribution license, establishes CC BY 4.0 for this reproduced figure.

License review covers this selected original scientific figure. No separately credited third-party image, photograph, logo or artwork is present in the selected crop. Attribution, source and license links are retained in the caption, footnote and NOTICE.md. This source artwork is not relicensed under the project MIT license. No endorsement is implied.

## Explanation boundary

Explain the exact sharing relationships between distinct queries and retained keys/values.

The illustrated 8:4:1 state ratio assumes equal layers, tokens, head width and precision; not a latency or quality ratio. SmolLM2 uses separate nine-query/three-KV dimensions.
