slop: reset some common args to the multi-model applies properly

Author: meh

common/arg.cpp

@@ -3073,6 +3073,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--chat-template-kwargs"}, "STRING",
         "sets additional params for the json template parser, must be a valid json object string, e.g. '{\"key1\":\"value1\",\"key2\":\"value2\"}'",
         [](common_params & params, const std::string & value) {
+            params.default_template_kwargs.clear();
             auto parsed = json::parse(value);
             for (const auto & item : parsed.items()) {
                 if (item.key() == "enable_thinking") {

common/preset.cpp

@@ -166,6 +166,9 @@ void common_preset::merge(const common_preset & other, bool overwrite_existing)
 }
 
 void common_preset::apply_to_params(common_params & params) const {
+    params.default_template_kwargs.clear();
+    params.enable_reasoning = -1;
+    params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
     for (const auto & [opt, val] : options) {
         // apply each option to params
         if (opt.handler_string) {