Qwen3.6-27B MTP Hits 2x on Mi50s

This is the kind of benchmark that matters for people trying to keep older ROCm hardware relevant: the speedup is real, but it is workload-sensitive and not free. The headline numbers look great, yet the prefill regression means you should treat MTP as a decode accelerator, not a universal throughput win. Grafting MTP onto an existing Q4_1 quant lowers the barrier for people who already have local GGUF workflows and older AMD cards. The biggest gains show up in short, decode-heavy workloads; the 18k-token prompt shows the real-world win is smaller than the short-benchmark peak. Tensor parallelism appears to do a lot of the heavy lifting, with MTP adding another layer of improvement on top. For local AI builders, this is a useful sign that llama.cpp’s ROCm path is getting more competitive even on aging GPUs like the Mi50. The current caveat is important: prefill performance regressed, so anyone deploying this needs to benchmark against their own prompt mix before drawing conclusions.