llama.cpp self-spec decoding hits Qwen snags

This is less a tuning success story than a reality check on local inference ergonomics: if acceptance rate stays low and the server crashes, the bottleneck is probably implementation maturity, not one magic flag.

• –The reported settings (`--spec-type ngram-mod --spec-ngram-size-n 24 --draft-min 48 --draft-max 64`) are fairly aggressive, so repeated low-acceptance resets suggest the draft strategy is not matching the model’s token patterns well.
• –A recent `llama.cpp` GitHub issue documents `ngram-mod` crashes on `Qwen3-Next-80B-A3B-Instruct`, including invalid input batch errors after accepted drafted tokens.
• –The current `llama.cpp` server docs expose speculative decoding controls, but they do not provide model-specific guidance for Qwen3.x A3B MoE checkpoints.
• –For local AI developers, the practical lesson is to confirm architecture support first, then tune draft bounds conservatively on short prompts before trusting speculative decoding in long chat sessions.