Qwen3.6-27B benchmark favors NVLink pairs

This is really a topology benchmark disguised as a model benchmark: on consumer GPUs, tensor parallelism is often limited by the slowest link rather than raw FLOPS. NVLink helped more at higher concurrency because communication overhead scales with batch size, so the bandwidth gap widens under load. TP=4 lost because the all-reduce had to traverse mostly PCIe links; sharding across more GPUs only helps when the fabric can keep up. The speculative MTP setup stayed stable across configs, which points to interconnect, not draft quality, as the bottleneck. For two-pair 3090 boxes, two separate TP=2 services is the practical layout; TP=4 looks like a trap unless every GPU pair has fast connectivity.