Qwen3.8 27B Swift-1.5

User reports Qwen3.8-27B (Swift-1.5 fine-tune, IQ3_S GGUF) at 60-70 t/s in short conversations on an RX 7900 XT 20GB, dropping to 45-50 t/s with around 90k words of history. Setup is llama.cpp (RDNA-boosts build) with q8_0 K and V cache, MTP plus n-gram speculation at n-max 2, auto-fit context with a 512MB margin, and all layers on the card. An earlier config that spilled into system RAM ran 27-28 t/s. Compressing the MTP head's memory saved about 1GB but dropped acceptance from about 0.7 to about 0.5 and cut speed; compressing the KV cache to q4 made no speed difference. An external DFlash2 draft model measured about the same as the built-in guesser.