Qwen3.8

User reports Qwen3.8 at Q6_K reaching 30 t/s on a mixed AMD Radeon AI PRO R9700 32GB plus NVIDIA RTX 4500 Ada setup, versus 40 t/s on the R9700 alone. Setup is llama.cpp (self-compiled Vulkan) with layer split, 131072 context, moe-cache-mib 12000, batch-size 16384 and ubatch-size 2304, with n-cpu-moe 999. Gemma 4 QAT gave 32 t/s mixed versus 50 t/s on the R9700 alone, and Qwen Flash Next Q4_K_M gave 9 t/s versus 14 t/s. Tensor split gave inconsistent results; the user suspects the slower Ada 4500 or PCIe latency and asks what else the setup could be used for.