Add parallel tools constraint and update parsing

Author: aldehir

common/chat-parser.cpp

@@ -108,7 +108,6 @@ bool common_chat_msg_parser::try_consume_partial_literal(const std::string & lit
             }
         }
     }
-
     return try_consume_literal(literal);
 }
 

common/chat.cpp

@@ -1723,6 +1723,9 @@ static common_chat_params common_chat_params_init_llama_3_x_nemotron(const commo
             {"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
             {"minItems", 1},
         };
+        if (!inputs.parallel_tool_calls) {
+            schema["maxItems"] = 1;
+        }
         builder.add_rule("root", "\"<TOOLCALL>\" " + builder.add_schema("tool_calls", schema) + " \"</TOOLCALL>\"");
     });
     data.grammar_triggers.push_back({COMMON_GRAMMAR_TRIGGER_TYPE_WORD, "<TOOLCALL>"});
@@ -1740,14 +1743,17 @@ static void common_chat_parse_llama_3_x_nemotron(common_chat_msg_parser & builde
         builder.add_content(builder.consume_rest());
         return;
     }
-
-    static const common_regex prefix(regex_escape("<TOOLCALL>"));
-    static const common_regex suffix(regex_escape("</TOOLCALL>"));
-
-    parse_prefixed_json_tool_call_array(builder, prefix);
-    if (!builder.try_find_regex(suffix)) {
-        builder.consume_rest();
+    static const common_regex toolcall_regex("<TOOLCALL>");
+    static const common_regex close_regex("</TOOLCALL>");
+    static const std::vector<std::vector<std::string>> args_paths = {{"arguments"}};
+    if (builder.try_find_regex(toolcall_regex)) {
+        auto tool_calls = builder.consume_json_with_dumped_args(args_paths);
+        if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) {
+            throw common_chat_msg_partial_exception("incomplete tool call array");
+        }
+        builder.consume_regex(close_regex);
     }
+    builder.add_content(builder.consume_rest());
 }
 
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
@@ -1852,7 +1858,7 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_llama_3_x(tmpl, params, allow_python_tag_builtin_tools);
     }
 
-    // Llama 3.3 Nemo (w/ tools)
+    // Llama 3.x Nemotron (w/ tools)
     if (src.find("<TOOLCALL>") != std::string::npos) {
         return common_chat_params_init_llama_3_x_nemotron(tmpl, params);
     }