HEZQEN LABS / MODEL INFRASTRUCTURE

LLM VRAM sizing calculator

Estimate model-weight memory, key-value cache and runtime headroom before choosing a GPU. Adjust the architecture and serving assumptions to model your workload.

CONFIGURATION

Model and serving parameters

Model parameters (billions)Model weight formatTransformer layersKV heads (GQA/MQA)Head dimensionContext length (tokens)Concurrent batch sizeKV cache formatRuntime reserve (%)

All numeric inputs must be positive whole numbers except model size and runtime reserve.

ESTIMATED MEMORY FOOTPRINT

GPU memory requirement

4.79 GiB

Plan for at least 5 GiB of available VRAM.

Model weights3.67 GiB

KV cache0.50 GiB

Runtime reserve0.63 GiB

Weight storage uses 0.5625 bytes per parameter. KV cache uses 2 bytes per element. Runtime reserve: 15%.

METHODOLOGY & LIMITATIONS

What this estimate includes

Weight memory is parameter count multiplied by effective bytes per parameter. KV-cache memory is estimated as 2 × layers × batch × context × KV heads × head dimension × cache bytes. A configurable reserve is then applied to the combined estimate.

INT8 and INT4 use conservative effective storage assumptions that include a metadata allowance. This is a planning estimate, not a hardware benchmark or a guarantee of fit. Actual allocation varies with architecture, quantization implementation, attention backend, tensor parallelism, fragmentation, CUDA graphs, runtime and serving framework. It does not include every allocator or system-level overhead.