google/gemma-3-27b-it-lora

VRAM is computed from parameter count (~20% added for KV cache + runtime) — an estimate; tune for your context length and serving stack.