KV Cache Calculator

Live sizing for transformer KV-cache memory. Supports standard MHA as well as GQA/MQA through the number of KV heads.

Layers (L)

Attention heads

KV heads

Head dimension (dₕ)

Sequence length / cached tokens (N)

Batch / concurrent sequences (B)

KV precision

Optional model hidden size

Used only for comparison with standard MHA.

Results

KV bytes per token / sequence

—

KV cache per sequence

—

Total KV cache for batch

—

KV vector width

—

Quantity	Value
K elements per token per layer
V elements per token per layer
K+V elements per token per layer
K+V bytes per token across all layers
Total elements cached
Total bytes cached

KV bytes = B × N × L × 2 × n_kv × d_h × bytes_per_element 2 = K + V n_kv = number of KV heads d_h = head dimension