update --reasoning-format={disabled -> nothink} as suggested

Author: ochafik

common/arg.cpp

@@ -2848,12 +2848,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
     add_opt(common_arg(
         {"--reasoning-format"}, "FORMAT",
-        "reasoning format (default: deepseek; allowed values: deepseek, none, disabled)\n"
-        "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only), 'disabled' prevents generation of thoughts (by closing any thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3).",
+        "reasoning format (default: deepseek; allowed values: deepseek, none, nothink)\n"
+        "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only), 'nothink' prevents generation of thoughts (by closing any thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3).",
         [](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 if (value == "nothink") {  params.reasoning_format = COMMON_REASONING_FORMAT_NOTHINK; }
             else { std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));

common/chat.cpp

@@ -594,9 +594,9 @@ 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_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
-        case COMMON_REASONING_FORMAT_DISABLED: return "disabled";
+        case COMMON_REASONING_FORMAT_NOTHINK:  return "nothink";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
@@ -1698,7 +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.enable_thinking = inputs.reasoning_format != COMMON_REASONING_FORMAT_NOTHINK;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
     if (!inputs.json_schema.empty()) {

common/common.h

@@ -216,7 +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)
+    COMMON_REASONING_FORMAT_NOTHINK,  // Forcibly disables thinking (causes any thinking tag to be closed, empty thinking tags to be inserted, or template specific variables to be set, depending on the chat format)
 };
 
 struct common_params {

tools/server/README.md

@@ -173,7 +173,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--no-slots` | disables slots monitoring endpoint<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
 | `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
 | `--jinja` | use jinja template for chat (default: disabled)<br/>(env: LLAMA_ARG_JINJA) |
-| `--reasoning-format FORMAT` | reasoning format (default: deepseek; allowed values: deepseek, none, disabled)<br/>controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only), 'disabled' prevents generation of thoughts (by closing any thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3).<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-format FORMAT` | reasoning format (default: deepseek; allowed values: deepseek, none, nothink)<br/>controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned. 'none' leaves thoughts unparsed in `message.content`, 'deepseek' puts them in `message.reasoning_content` (for DeepSeek R1 & Command R7B only), 'nothink' prevents generation of thoughts (by closing any thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3).<br/>(env: LLAMA_ARG_THINK) |
 | `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
 | `--chat-template-file JINJA_TEMPLATE_FILE` | set custom jinja chat template file (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted (unless --jinja is set before this flag):<br/>list of built-in templates:<br/>bailing, chatglm3, chatglm4, chatml, command-r, deepseek, deepseek2, deepseek3, exaone3, falcon3, gemma, gigachat, glmedge, granite, llama2, llama2-sys, llama2-sys-bos, llama2-sys-strip, llama3, llama4, megrez, minicpm, mistral-v1, mistral-v3, mistral-v3-tekken, mistral-v7, mistral-v7-tekken, monarch, openchat, orion, phi3, phi4, rwkv-world, smolvlm, vicuna, vicuna-orca, yandex, zephyr<br/>(env: LLAMA_ARG_CHAT_TEMPLATE_FILE) |
 | `--no-prefill-assistant` | whether to prefill the assistant's response if the last message is an assistant message (default: prefill enabled)<br/>when this flag is set, if the last message is an assistant message then it will be treated as a full message and not prefilled<br/>(env: LLAMA_ARG_NO_PREFILL_ASSISTANT) |

tools/server/tests/unit/test_template.py

@@ -25,31 +25,32 @@ 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>"),
+@pytest.mark.parametrize("tools", [None, [], [TEST_TOOL]])
+@pytest.mark.parametrize("template_name,nothink,expected_end", [
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", False,  "<think>\n"),
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", True, "<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-Qwen3-0.6B", False,  "<|im_start|>assistant\n"),
+    ("Qwen-Qwen3-0.6B", True, "<|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>"),
+    ("Qwen-QwQ-32B", False,  "<|im_start|>assistant\n<think>\n"),
+    ("Qwen-QwQ-32B", True, "<|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|>"),
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use-think", False,  "<|START_THINKING|>"),
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use-think", True, "<|START_THINKING|><|END_THINKING|>"),
 ])
-def test_enable_thinking(template_name: str, enable_thinking: bool, expected_end: str):
+def test_nothink(template_name: str, nothink: bool, expected_end: str, tools: list[dict]):
     global server
     server.jinja = True
-    server.reasoning_format = 'deepseek' if enable_thinking else 'disabled'
+    server.reasoning_format = 'nothink' if nothink else None
     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],
+        "tools": tools,
     })
     assert res.status_code == 200
     prompt = res.body["prompt"]

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', 'disabled'] | None = None
+    reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None
     chat_template: str | None = None
     chat_template_file: str | None = None
     server_path: str | None = None