server: add --reasoning-format=disabled to disable thinking (incl. qwen3 w/ enable_thinking:false)

Author: ochafik

common/arg.cpp

@@ -2854,6 +2854,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
+            else if (value == "disabled") { params.reasoning_format = COMMON_REASONING_FORMAT_DISABLED; }
             else { std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));

common/chat.cpp

@@ -133,6 +133,7 @@ struct templates_params {
     bool stream;
     std::string grammar;
     bool add_generation_prompt = true;
+    bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
@@ -591,6 +592,16 @@ std::string common_chat_format_name(common_chat_format format) {
     }
 }
 
+std::string common_reasoning_format_name(common_reasoning_format format) {
+    switch (format) {
+        case COMMON_REASONING_FORMAT_NONE: return "none";
+        case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
+        case COMMON_REASONING_FORMAT_DISABLED: return "disabled";
+        default:
+            throw std::runtime_error("Unknown reasoning format");
+    }
+}
+
 static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) {
     std::string arguments;
     if (builder.is_partial()) {
@@ -918,7 +929,11 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
     data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
-        data.thinking_forced_open = true;
+        if (!inputs.enable_thinking) {
+            data.prompt += "<|END_THINKING|>";
+        } else {
+            data.thinking_forced_open = true;
+        }
     }
 
     data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1186,7 +1201,11 @@ static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_DEEPSEEK_R1;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        data.thinking_forced_open = true;
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
     }
 
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
@@ -1460,10 +1479,18 @@ static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser
 static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    json additional_context = {
+        {"enable_thinking", inputs.enable_thinking},
+    };
+
+    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        data.thinking_forced_open = true;
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>";
+        } else {
+            data.thinking_forced_open = true;
+        }
     }
 
     if (!inputs.tools.is_null()) {
@@ -1671,6 +1698,7 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.messages = common_chat_msgs_to_json_oaicompat<json>(inputs.messages, /* concat_text= */ !tmpl.original_caps().requires_typed_content);
     params.add_generation_prompt = inputs.add_generation_prompt;
     params.tool_choice = inputs.tool_choice;
+    params.enable_thinking = inputs.reasoning_format != COMMON_REASONING_FORMAT_DISABLED;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
     if (!inputs.json_schema.empty()) {

common/chat.h

@@ -123,6 +123,7 @@ struct common_chat_templates_inputs {
     common_chat_tool_choice tool_choice = COMMON_CHAT_TOOL_CHOICE_AUTO;
     bool parallel_tool_calls = false;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_NONE;
+    bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
 };
 
@@ -182,6 +183,7 @@ std::string common_chat_format_example(
     bool use_jinja);
 
 std::string               common_chat_format_name(common_chat_format format);
+std::string               common_reasoning_format_name(common_reasoning_format format);
 common_chat_msg           common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax);
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);

common/common.h

@@ -216,6 +216,7 @@ struct common_params_vocoder {
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
     COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DISABLED, // Disable thinking (causes any thinking tag to be closed, or empty thinking tags to be inserted, depending on the model)
 };
 
 struct common_params {

tools/server/server.cpp

@@ -178,7 +178,7 @@ struct slot_params {
             {"grammar_triggers",          grammar_triggers},
             {"preserved_tokens",          sampling.preserved_tokens},
             {"chat_format",               common_chat_format_name(oaicompat_chat_syntax.format)},
-            {"reasoning_format",          (oaicompat_chat_syntax.reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK ? "deepseek" : "none")},
+            {"reasoning_format",          common_reasoning_format_name(oaicompat_chat_syntax.reasoning_format)},
             {"reasoning_in_content",      oaicompat_chat_syntax.reasoning_in_content},
             {"thinking_forced_open",      oaicompat_chat_syntax.thinking_forced_open},
             {"samplers",                  samplers},

tools/server/tests/unit/test_template.py

@@ -25,6 +25,38 @@ def create_server():
     server.n_slots = 1
 
 
+@pytest.mark.parametrize("template_name,enable_thinking,expected_end", [
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", True,  "<think>\n"),
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", False, "<think>\n</think>"),
+
+    ("Qwen-Qwen3-0.6B", True,  "<|im_start|>assistant\n"),
+    ("Qwen-Qwen3-0.6B", False, "<|im_start|>assistant\n<think>\n\n</think>\n\n"),
+
+    ("Qwen-QwQ-32B", True,  "<|im_start|>assistant\n<think>\n"),
+    ("Qwen-QwQ-32B", False, "<|im_start|>assistant\n<think>\n</think>"),
+
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use-think", True,  "<|START_THINKING|>"),
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use-think", False, "<|START_THINKING|><|END_THINKING|>"),
+])
+def test_enable_thinking(template_name: str, enable_thinking: bool, expected_end: str):
+    global server
+    server.jinja = True
+    server.reasoning_format = 'deepseek' if enable_thinking else 'disabled'
+    server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
+    server.start(timeout_seconds=TIMEOUT_SERVER_START)
+
+    res = server.make_request("POST", "/apply-template", data={
+        "messages": [
+            {"role": "user", "content": "What is today?"},
+        ],
+        "tools": [TEST_TOOL],
+    })
+    assert res.status_code == 200
+    prompt = res.body["prompt"]
+
+    assert prompt.endswith(expected_end), f"Expected prompt to end with '{expected_end}', got '{prompt}'"
+
+
 @pytest.mark.parametrize("tools", [None, [], [TEST_TOOL]])
 @pytest.mark.parametrize("template_name,format", [
     ("meta-llama-Llama-3.3-70B-Instruct",    "%d %b %Y"),

tools/server/tests/utils.py

@@ -84,7 +84,7 @@ class ServerProcess:
     draft_max: int | None = None
     no_webui: bool | None = None
     jinja: bool | None = None
-    reasoning_format: Literal['deepseek', 'none'] | None = None
+    reasoning_format: Literal['deepseek', 'none', 'disabled'] | None = None
     chat_template: str | None = None
     chat_template_file: str | None = None
     server_path: str | None = None