Qwen 27B achieves 50-90 tokens/s on dual GPU

The 27B parameter model runs at 50-90 tokens/s decode speed on a 4090+3090 system with multi-token prediction. Prefill speed reaches 1500-2200 tokens/s using q6kxl quantization.