Single steam benchmarks (club-3090) model: qwen3.6-27b-autoround-int4 BEFORE: 1x3090 *Their default script recipe for single 3090'*s (4-bit quant and 4-bit kv cache, mtp=2) NARRATIVE decode_TPS: mean = 53 std = 0.6 CODE decode_TPS: mean = 62 std= 1.4 AFTER: 2x3090 Their default script recipe for dua