fix & test grammar & json_schema w/ & w/o --jinja

Author: ochafik

common/chat.cpp

@@ -1431,7 +1431,7 @@ static common_chat_params common_chat_templates_apply_jinja(
 
     // Use generic handler when mixing tools + JSON schema.
     // TODO: support that mix in handlers below.
-    if ((!params.tools.is_array() && params.json_schema.is_object())) {
+    if ((params.tools.is_array() && params.json_schema.is_object())) {
         return common_chat_params_init_generic(tmpl, params);
     }
 

examples/server/server.cpp

@@ -329,9 +329,6 @@ struct server_task {
         }
 
         // process "json_schema" and "grammar"
-        if (data.contains("json_schema") && !data.at("json_schema").is_null() && data.contains("grammar") && !data.at("grammar").is_null()) {
-            throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
-        }
         if (data.contains("json_schema") && !data.contains("grammar")) {
             try {
                 auto schema                  = json_value(data, "json_schema", json::object());

examples/server/tests/unit/test_chat_completion.py

@@ -169,6 +169,47 @@ def test_completion_with_response_format(response_format: dict, n_predicted: int
         assert "error" in res.body
 
 
+@pytest.mark.parametrize("jinja,json_schema,n_predicted,re_content", [
+    (False, {"const": "42"}, 6, "\"42\""),
+    (True, {"const": "42"}, 6, "\"42\""),
+])
+def test_completion_with_json_schema(jinja: bool, json_schema: dict, n_predicted: int, re_content: str):
+    global server
+    server.jinja = jinja
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": n_predicted,
+        "messages": [
+            {"role": "system", "content": "You are a coding assistant."},
+            {"role": "user", "content": "Write an example"},
+        ],
+        "json_schema": json_schema,
+    })
+    assert res.status_code == 200, f'Expected 200, got {res.status_code}'
+    choice = res.body["choices"][0]
+    assert match_regex(re_content, choice["message"]["content"]), f'Expected {re_content}, got {choice["message"]["content"]}'
+
+
+@pytest.mark.parametrize("jinja,grammar,n_predicted,re_content", [
+    (False, 'root ::= "a"{5,5}', 6, "a{5,5}"),
+    (True, 'root ::= "a"{5,5}', 6, "a{5,5}"),
+])
+def test_completion_with_grammar(jinja: bool, grammar: str, n_predicted: int, re_content: str):
+    global server
+    server.jinja = jinja
+    server.start()
+    res = server.make_request("POST", "/chat/completions", data={
+        "max_tokens": n_predicted,
+        "messages": [
+            {"role": "user", "content": "Does not matter what I say, does it?"},
+        ],
+        "grammar": grammar,
+    })
+    assert res.status_code == 200, res.body
+    choice = res.body["choices"][0]
+    assert match_regex(re_content, choice["message"]["content"]), choice["message"]["content"]
+
+
 @pytest.mark.parametrize("messages", [
     None,
     "string",

examples/server/utils.hpp

@@ -571,8 +571,8 @@ static json oaicompat_completion_params_parse(
         llama_params["stop"] = json_value(body, "stop", json::array());
     }
 
-    auto json_schema = json_value(llama_params, "json_schema", json());
-    auto grammar = json_value(llama_params, "grammar", std::string());
+    auto json_schema = json_value(body, "json_schema", json());
+    auto grammar = json_value(body, "grammar", std::string());
     if (!json_schema.is_null() && !grammar.empty()) {
         throw std::runtime_error("Cannot use both json_schema and grammar");
     }
@@ -601,7 +601,7 @@ static json oaicompat_completion_params_parse(
     inputs.use_jinja             = use_jinja;
     inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
     inputs.extract_reasoning     = reasoning_format != COMMON_REASONING_FORMAT_NONE;
-    if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && llama_params.contains("grammar")) {
+    if (!inputs.tools.empty() && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE && body.contains("grammar")) {
         throw std::runtime_error("Cannot use custom grammar constraints with tools.");
     }