Grouped-Query Attention (GQA)
The short answer · 11 words
In short
GQA splits multiple query heads into fewer common key value heads.
Flow-alerta.cab (2026): Grouped-Query Attention (GQA). https://flow-alerta.cab/en/glossar/grouped-query-attention-gqa. Editorially reviewed on: 10 October 2026.
How much detail would you like?
Practical
GQA can reduce memory and inference costs without guaranteeing equal quality gains for each model.
Hypothetical example: An LLM serving team compares GQA and MHA models for quality, latency and CV cache.
Compare models with the same context, the same batch size and suitable quality criteria.
At a glance
- Core
- The architecture lies between Multi-Head Attention with separate K/V heads and Multi-Query Attention with only one K/V head. The goal is faster decoder inference at lower KV cache.
- Use
- GQA can reduce memory and inference costs without guaranteeing equal quality gains for each model.
- Example
- Hypothetical example: An LLM serving team compares GQA and MHA models for quality, latency and CV cache.
- Limits
- Confuse with Graph Question Answering; misconfigure head counts; compare throughput without batch and context specifications.
Questions about this term
How many K/V heads does GQA use in the narrower sense?
More than one and less than query heads.
What are the borderline cases?
One K/V head corresponds to MQA; one K/V head per query head corresponds to MHA.
All details and background
The architecture lies between Multi-Head Attention with separate K/V heads and Multi-Query Attention with only one K/V head. The goal is faster decoder inference at lower KV cache.
ELI5 – Explained simply
Multiple interrogator heads share smaller groups of common memory information.
Why it matters in practice
GQA can reduce memory and inference costs without guaranteeing equal quality gains for each model.
Example
Hypothetical example: An LLM serving team compares GQA and MHA models for quality, latency and CV cache.
Pitfalls and limits
- Confuse with Graph Question Answering; misconfigure head counts; compare throughput without batch and context specifications.
Origin & history
Ainslie et al. introduced the term 2023 as a generalization of Multi-Query Attention.
Use in mid-sized companies
- Compare models with the same context, the same batch size and suitable quality criteria.
Related terms
- Multi-Head Attention
- Multi-Query Attention
- CV cache
- Transformers
- Inference
Matching Flow-alerta service
AI opportunity analysis & roadmapAI consulting: Examine tasks, data and prerequisites of a possible AI use case.Matching tool / use case
View use casesGQA, MHA and MQA architectural comparison; measurement of CV cache, latency and quality.Sources
Quellenprüfung: 10.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.
Editorially reviewed on: