Feature request: enable speculative decoding by default

[ggerganov]

With recent advances in prompt-based speculative decoding in llama.cpp, there is a new option that enables a "default" setup for speculative decoding. The setup is compatible with all models and does not require extra resources. In certain use cases, it leads to significant performance improvements:

• Agentic workflows
• Editing files
• Summarization

Recommend to add the new --spec-default flag to the llama-server startup command. This flag was introduced in ggml-org/llama.cpp#22223, so it requires the latest version of llama.cpp.

More info:

• arg : add --spec-default llama.cpp#22223
• server : speculative checkpointing llama.cpp#19493
• spec : add ngram-mod llama.cpp#19164