NVIDIA RTX 3090 llama.cpp: Gemma4 31B (Q4_K) token generation at 32k context

THIRD-PARTY MEASURED (Hardware Corner GPU LLM Benchmarks page, tables dated March 2026). llama.cpp llama-bench, GGUF, single GPU, batch-1. NVIDIA RTX 3090. Gemma4 31B (Q4_K). Number is the 32k-context column of the Token Generation table: context depth the run was measured at, not a fixed pp512/tg128. Per-request, NOT aggregate. Page annotates CUDA with -fa 1 on the Qwen3 rows of this card.