Qwen3.6-27B MTP boosts short prompts, drags at 70k+

The hot take is that MTP is helping the easy part of the workload and hurting the hard part; for long-context serving, speculative decoding can become the bottleneck instead of the accelerator.

• –The reported config is aggressive: `--max-model-len 262144`, chunked prefill, prefix caching, and `--speculative-config {"method":"mtp","num_speculative_tokens":3}` all stack pressure onto memory bandwidth and scheduler overhead.
• –The pattern suggests the MTP draft path is paying a larger per-step tax as the KV cache grows, so the speedup vanishes once the prompt gets very long.
• –For agentic coding, where long contexts are routine, a lower speculative token count or disabling MTP entirely is likely the pragmatic choice.
• –The post is useful less as a bug report and more as a serving benchmark: it shows that “faster on short prompts” does not automatically mean “faster in real long-context workloads.”