Revision History
Grouped-Query Attention · 13 revisions
Sizes are character counts of the article source. The signed number is the change from the previous revision.
Recent edit summaries
Detailed summaries recorded by editors. Generic maintenance summaries are omitted here; every recorded revision remains below. Dates describe the edit, not necessarily the event it covers.
- Correction: Kimi K3 was released in July 2026 (July 16), not June.
Version 13 · Oct 31, 2026, 03:44 AM
- Correction: fixed Llama 2 GQA sizes (34B and 70B, not 70B only), removed a fabricated Llama 2 quote, corrected misquotes of the MQA and Qwen2 papers, GQA paper uptraining and Figure 6 descriptions, Table 1 'ROUGE-1' label, MLA-vs-GQA compression ratio, Phi-3/Gemma 2/Qwen/Mistral adoption claims, PyTorch code layout and tensor-parallel claims; added sourced adoption tables, developer ablations, 2025-2026 GQA vs MLA models and conversion methods.
Version 12 · Oct 31, 2026, 03:15 AM
- Add NVIDIA's July 31, 2026 co-design guidance on K/V head count as a decode-efficiency and parallelism lever, with new reference [10]
Version 10 · Aug 5, 2026, 10:09 AM
- Added 6 contextual internal links
Version 9 · Jul 23, 2026, 03:04 PM
- Render formulas with LaTeX math notation
Version 8 · Jul 11, 2026, 07:13 PM