}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+ add_opt(common_arg(
+ {"--spec-default"},
+ string_format("enable default speculative decoding config"),
+ [](common_params & params) {
+ params.speculative.type = COMMON_SPECULATIVE_TYPE_NGRAM_MOD;
+ params.speculative.ngram_size_n = 24;
+ params.speculative.n_min = 48;
+ params.speculative.n_max = 64;
+ }
+ ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_CLI}));
+
return ctx_arg;
}