RDNA4 performance meta with Qwen 3.8: user questions about architecture-specific inference engines and potential gains against RDNA4

As it says, really - I'm currently running vllm-radiance on dual R9700s, with Qwen 3.8 27B FP8 (or, rather, Swift 1.5 FP8). Performance is great an' all (5000t/s prefill, 130t/s+ code gen), but I'm just...