llama-server flag optimization benchmark

llama-server flags benchmark on RTX 5090, 16K prompt, six models (Qwen3.8-27B, Gemma 4 31B, Qwen3.6-35B-A3B, Nemotron Lightning, Ornith 35B, Qwopus Flash). --flash-attn on: irrelevant, auto already selects Flash Attention. --parallel 1: useful, costs nothing, up to 10% decode improvement on Gemma at depth, frees ~2.3 GB VRAM. KV cache Q8_0: saves 1-2.6 GB on dense models, only 0.1 GB on 3B-active MoE, costs 5-7% decode speed. KV cache Q4_0: decode drops -17%, -23%, -27%, -23%, -29%, -17% across six models; quality barely moves (GPQA: F16 100, Q8 100, Q4 98). MTP draft heads: Qwen3.8-27B ~2.2x, Gemma ~1.5x, Nemotron Lightning 0.7x (slower). Speculative decoding not automatically beneficial above ~300 tok/s. 16K prompt first request: 5-6s processing; cached: 0.1-0.4s. Takeaway: don't copy commands blindly; benchmark your actual workload.