switch to --reasoning-budget flag

Author: ochafik

common/arg.cpp

@@ -2851,15 +2851,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:\n"
         "- none: leaves thoughts unparsed in `message.content`\n"
         "- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)\n"
-        "- nothink: prevents generation of thoughts (forcibly closing thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3)\n"
         "(default: deepseek)",
         [](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 == "nothink") {  params.reasoning_format = COMMON_REASONING_FORMAT_NOTHINK; }
-            else { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK"));
+    add_opt(common_arg(
+        {"--reasoning-budget"}, "N",
+        "controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)",
+        [](common_params & params, int value) {
+            if (value != 0 && value != -1) { throw std::invalid_argument("invalid value"); }
+            params.reasoning_budget = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_THINK_BUDGET"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(
@@ -2958,7 +2964,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             /**/ if (value == "jsonl") { params.batched_bench_output_jsonl = true; }
             else if (value == "md") { params.batched_bench_output_jsonl = false; }
-            else { std::invalid_argument("invalid value"); }
+            else { throw std::invalid_argument("invalid value"); }
         }
     ).set_examples({LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(

common/chat.cpp

@@ -596,7 +596,6 @@ const char * 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_NOTHINK:  return "nothink";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }
@@ -1700,7 +1699,7 @@ 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_NOTHINK;
+    params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
     if (!inputs.json_schema.empty()) {

common/common.h

@@ -216,7 +216,6 @@ 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_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 {
@@ -369,6 +368,7 @@ struct common_params {
     bool use_jinja = false;                                                                                 // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int reasoning_budget = -1;
     bool prefill_assistant = true;                                                                          // if true, any trailing assistant message will be prefilled into the response
 
     std::vector<std::string> api_keys;

tools/server/README.md

@@ -173,7 +173,8 @@ 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` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)<br/>- nothink: prevents generation of thoughts (forcibly closing thoughts tag or setting template-specific variables such as `enable_thinking: false` for Qwen3)<br/>(default: deepseek)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-format FORMAT` | controls whether thought tags are allowed and/or extracted from the response, and in which format they're returned; one of:<br/>- none: leaves thoughts unparsed in `message.content`<br/>- deepseek: puts thoughts in `message.reasoning_content` (except in streaming mode, which behaves as `none`)<br/>(default: deepseek)<br/>(env: LLAMA_ARG_THINK) |
+| `--reasoning-budget N` | controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) |
 | `--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/server.cpp

@@ -2089,6 +2089,7 @@ struct server_context {
             /* common_chat_templates */ chat_templates.get(),
             /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
             /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
+            /* enable_thinking       */ params_base.reasoning_budget != 0,
         };
     }
 

tools/server/tests/unit/test_template.py

@@ -26,23 +26,24 @@ def create_server():
 
 
 @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>"),
+@pytest.mark.parametrize("template_name,reasoning_budget,expected_end", [
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B", None, "<think>\n"),
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B",   -1, "<think>\n"),
+    ("deepseek-ai-DeepSeek-R1-Distill-Qwen-32B",    0, "<think>\n</think>"),
 
-    ("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-Qwen3-0.6B", -1, "<|im_start|>assistant\n"),
+    ("Qwen-Qwen3-0.6B",  0, "<|im_start|>assistant\n<think>\n\n</think>\n\n"),
 
-    ("Qwen-QwQ-32B", False,  "<|im_start|>assistant\n<think>\n"),
-    ("Qwen-QwQ-32B", True, "<|im_start|>assistant\n<think>\n</think>"),
+    ("Qwen-QwQ-32B", -1, "<|im_start|>assistant\n<think>\n"),
+    ("Qwen-QwQ-32B",  0, "<|im_start|>assistant\n<think>\n</think>"),
 
-    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use", False,  "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"),
-    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use", True, "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_THINKING|><|END_THINKING|>"),
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use", -1, "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"),
+    ("CohereForAI-c4ai-command-r7b-12-2024-tool_use",  0, "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|><|START_THINKING|><|END_THINKING|>"),
 ])
-def test_nothink(template_name: str, nothink: bool, expected_end: str, tools: list[dict]):
+def test_reasoning_budget(template_name: str, reasoning_budget: int | None, expected_end: str, tools: list[dict]):
     global server
     server.jinja = True
-    server.reasoning_format = 'nothink' if nothink else None
+    server.reasoning_budget = reasoning_budget
     server.chat_template_file = f'../../../models/templates/{template_name}.jinja'
     server.start(timeout_seconds=TIMEOUT_SERVER_START)
 

tools/server/tests/utils.py

@@ -85,6 +85,7 @@ class ServerProcess:
     no_webui: bool | None = None
     jinja: bool | None = None
     reasoning_format: Literal['deepseek', 'none', 'nothink'] | None = None
+    reasoning_budget: int | None = None
     chat_template: str | None = None
     chat_template_file: str | None = None
     server_path: str | None = None
@@ -191,6 +192,8 @@ def start(self, timeout_seconds: int | None = DEFAULT_HTTP_TIMEOUT) -> None:
             server_args.append("--jinja")
         if self.reasoning_format is not None:
             server_args.extend(("--reasoning-format", self.reasoning_format))
+        if self.reasoning_budget is not None:
+            server_args.extend(("--reasoning-budget", self.reasoning_budget))
         if self.chat_template:
             server_args.extend(["--chat-template", self.chat_template])
         if self.chat_template_file:

tools/server/utils.hpp

@@ -568,6 +568,7 @@ struct oaicompat_parser_options {
     common_chat_templates * tmpls;
     bool allow_image;
     bool allow_audio;
+    bool enable_thinking = true;
 };
 
 // used by /chat/completions endpoint
@@ -732,6 +733,7 @@ static json oaicompat_chat_params_parse(
     inputs.use_jinja             = opt.use_jinja;
     inputs.parallel_tool_calls   = json_value(body, "parallel_tool_calls", false);
     inputs.reasoning_format      = opt.reasoning_format;
+    inputs.enable_thinking       = opt.enable_thinking;
     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.");
     }