LLM inference in C/C++
C++ ggml
chat : pass reasoning_effort to template
Andy Williams7e4c0a96880d
diff --git a/common/arg.cpp b/common/arg.cppindex 1f70d0ad45..3e63ce0a6d 100644--- a/common/arg.cpp+++ b/common/arg.cpp@@ -3646,6 +3646,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex}}).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING"));+ add_opt(common_arg(+ {"--reasoning-effort"}, "LEVEL",+ "reasoning effort level given to the chat template: 'default' to keep the template default,\n"+ "or a level such as 'minimal', 'low', 'medium', 'high', 'xhigh' or 'max' (default: default)",+ [](common_params & params, const std::string & value) {+ if (value == "default") {+ params.default_template_kwargs.erase("reasoning_effort");+ } else {+ params.default_template_kwargs["reasoning_effort"] = json(value).dump();+ }+ }+ ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_EFFORT"));add_opt(common_arg({"--reasoning-budget"}, "N","token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)",diff --git a/common/chat.cpp b/common/chat.cppindex faf51dcd27..a45c51fa0c 100644--- a/common/chat.cpp+++ b/common/chat.cpp@@ -920,6 +920,10 @@ static std::string common_chat_template_direct_apply_impl(bool enabled = inp["preserve_reasoning"].get<bool>();jinja::caps_apply_preserve_reasoning(ctx, enabled);}+ if (inp.contains("reasoning_effort") && inp["reasoning_effort"].is_string() && !inp["reasoning_effort"].empty()) {+ std::string reasoning_effort = inp["reasoning_effort"].get<std::string>();+ jinja::caps_apply_reasoning_effort(ctx, reasoning_effort);+ }jinja::global_from_json(ctx, inp, inputs.mark_input);diff --git a/common/jinja/caps.cpp b/common/jinja/caps.cppindex cdd7ccfa26..fb97909edf 100644--- a/common/jinja/caps.cpp+++ b/common/jinja/caps.cpp@@ -17,7 +17,7 @@ namespace jinja {using caps_json_fn = std::function<json()>;using caps_ctx_fn = std::function<void(context &)>;-using caps_analyze_fn = std::function<void(bool, value &, value &, const std::string &)>;+using caps_analyze_fn = std::function<void(context &, bool, value &, value &, const std::string &)>;void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) {ctx.set_val("preserve_thinking", mk_val<value_bool>(enabled));@@ -26,6 +26,12 @@ void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) {ctx.set_val("drop_thinking", mk_val<value_bool>(!enabled));}+void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort) {+ value var = mk_val<value_string>(effort); // bind to the same value for stats+ ctx.set_val("reasoning_effort", var);+ ctx.set_val("reasoning_strength", var);+}+static void caps_try_execute(jinja::program & prog,const caps_json_fn & messages_fn,const caps_ctx_fn & ctx_fn,@@ -62,7 +68,7 @@ static void caps_try_execute(jinja::program & prog,// ignore exceptions during capability analysis}- analyze_fn(success, messages, tools, result);+ analyze_fn(ctx, success, messages, tools, result);}// for debugging only@@ -87,6 +93,7 @@ std::map<std::string, bool> caps::to_map() const {{"supports_parallel_tool_calls", supports_parallel_tool_calls},{"supports_system_role", supports_system_role},{"supports_preserve_reasoning", supports_preserve_reasoning},+ {"supports_reasoning_effort", supports_reasoning_effort},{"supports_object_arguments", supports_object_arguments},};}@@ -124,7 +131,7 @@ caps caps_get(jinja::program & prog) {},nullptr, // ctx_fnnullptr, // tools_fn- [&](bool success, value & messages, value &, const std::string &) {+ [&](context &, bool success, value & messages, value &, const std::string &) {auto & content = messages->at(0)->at("content");caps_print_stats(content, "messages[0].content");if (has_op(content, "selectattr") || has_op(content, "array_access")) {@@ -158,7 +165,7 @@ caps caps_get(jinja::program & prog) {},nullptr, // ctx_fnnullptr, // tools_fn- [&](bool, value & messages, value &, const std::string &) {+ [&](context &, bool, value & messages, value &, const std::string &) {auto & content = messages->at(0)->at("content");caps_print_stats(content, "messages[0].content");if (!content->stats.used) {@@ -234,7 +241,7 @@ caps caps_get(jinja::program & prog) {},});},- [&](bool success, value & messages, value & tools, const std::string &) {+ [&](context &, bool success, value & messages, value & tools, const std::string &) {if (!success) {return; // Nothing can be inferred}@@ -327,7 +334,7 @@ caps caps_get(jinja::program & prog) {},});},- [&](bool success, value & messages, value & tools, const std::string &) {+ [&](context &, bool success, value & messages, value & tools, const std::string &) {if (!success) {result.supports_tool_calls = false;result.supports_tools = false;@@ -429,7 +436,7 @@ caps caps_get(jinja::program & prog) {},});},- [&](bool success, value & messages, value &, const std::string &) {+ [&](context &, bool success, value & messages, value &, const std::string &) {if (!success) {result.supports_parallel_tool_calls = false;return;@@ -486,7 +493,7 @@ caps caps_get(jinja::program & prog) {caps_apply_preserve_reasoning(ctx, true);},nullptr, // tools_fn- [&](bool, value &, value &, const std::string & output) {+ [&](context &, bool, value &, value &, const std::string & output) {// note: we cannot use stats here because the reasoning_content may be used for "if" condition test, but not actually outputted in the final resultif (output.find(reasoning_placeholder) != std::string::npos) {result.supports_preserve_reasoning = true;@@ -494,6 +501,32 @@ caps caps_get(jinja::program & prog) {});+ JJ_DEBUG("%s\n", ">>> Running capability check: reasoning effort");++ // case: reasoning effort level+ caps_try_execute(+ prog,+ [&]() {+ // messages+ return json::array({+ {+ {"role", "user"},+ {"content", "User message"}+ },+ });+ },+ [&](context & ctx) {+ ctx.set_val("enable_thinking", mk_val<value_bool>(true));+ caps_apply_reasoning_effort(ctx, "low");+ },+ nullptr, // tools_fn+ [&](context & ctx, bool, value &, value &, const std::string &) {+ value effort = ctx.get_val("reasoning_effort");+ caps_print_stats(effort, "reasoning_effort");+ result.supports_reasoning_effort = effort->stats.used;+ }+ );+JJ_DEBUG("%s\n", result.to_string().c_str());return result;diff --git a/common/jinja/caps.h b/common/jinja/caps.hindex a290cd7da6..b81dd95f2e 100644--- a/common/jinja/caps.h+++ b/common/jinja/caps.h@@ -16,6 +16,9 @@ struct caps {// supports preserve reasoning trace in the full history, not just the last assistant messagebool supports_preserve_reasoning = false;+ // supports reasoning effort levels+ bool supports_reasoning_effort = false;+// one of the 2 content capabilities must be truebool supports_string_content = true;bool supports_typed_content = false;@@ -32,5 +35,6 @@ struct caps {caps caps_get(jinja::program & prog);void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled);+void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort);} // namespace jinjadiff --git a/tests/test-chat.cpp b/tests/test-chat.cppindex de75ad3300..51ad854849 100644--- a/tests/test-chat.cpp+++ b/tests/test-chat.cpp@@ -6955,6 +6955,24 @@ static void test_reasoning_budget_message_per_request() {}}+static void test_reasoning_effort_caps() {+ LOG_DBG("%s\n", __func__);++ auto assert_supports_effort = [](const std::string & path, bool expected) {+ auto tmpls = read_templates(path);+ assert_equals(expected, common_chat_templates_get_caps(tmpls.get()).at("supports_reasoning_effort"));+ };++ assert_supports_effort("models/templates/deepseek-ai-DeepSeek-V4.jinja", true);+ assert_supports_effort("models/templates/muse-glimmer.jinja", true);+ assert_supports_effort("models/templates/tencent-Hy3.jinja", true);+ assert_supports_effort("models/templates/openai-gpt-oss-120b.jinja", true);+ assert_supports_effort("models/templates/upstage-Solar-Open-100B.jinja", true);+ assert_supports_effort("models/templates/Cohere2MoE.jinja", true);+ assert_supports_effort("models/templates/meta-llama-Llama-3.1-8B-Instruct.jinja", false);+ assert_supports_effort("models/templates/Qwen-Qwen3-0.6B.jinja", false);+}+static void test_msg_diffs_compute() {LOG_DBG("%s\n", __func__);{@@ -7114,6 +7132,7 @@ int main(int argc, char ** argv) {test_deepseek_v4_thinking_retention();test_deepseek_v4_tool_result_ordering();test_template_generation_prompt();+ test_reasoning_effort_caps();test_reasoning_budget_tokens_per_request();test_reasoning_budget_message_per_request();test_template_output_peg_parsers(detailed_debug);diff --git a/tools/cli/README.md b/tools/cli/README.mdindex 880c4a5408..b3543ed4da 100644--- a/tools/cli/README.md+++ b/tools/cli/README.md@@ -170,6 +170,7 @@| `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<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`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) || `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |+| `--reasoning-effort LEVEL` | reasoning effort level given to the chat template: 'default' to keep the template default,<br/>or a level such as 'minimal', 'low', 'medium', 'high', 'xhigh' or 'max' (default: default)<br/>(env: LLAMA_ARG_REASONING_EFFORT) || `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) || `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) || `--reasoning-preserve, --no-reasoning-preserve` | preserve reasoning trace in the full history, not just the last assistant message (default: template default)<br/>compatible with certain templates having 'supports_preserve_reasoning' capability<br/>example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking<br/>(env: LLAMA_ARG_REASONING_PRESERVE) |diff --git a/tools/completion/README.md b/tools/completion/README.mdindex c2e52ac066..3bc2d823ec 100644--- a/tools/completion/README.md+++ b/tools/completion/README.md@@ -251,6 +251,7 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1| `--jinja, --no-jinja` | whether to use jinja template engine 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`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) || `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |+| `--reasoning-effort LEVEL` | reasoning effort level given to the chat template: 'default' to keep the template default,<br/>or a level such as 'minimal', 'low', 'medium', 'high', 'xhigh' or 'max' (default: default)<br/>(env: LLAMA_ARG_REASONING_EFFORT) || `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) || `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) || `--reasoning-preserve, --no-reasoning-preserve` | preserve reasoning trace in the full history, not just the last assistant message (default: template default)<br/>compatible with certain templates having 'supports_preserve_reasoning' capability<br/>example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking<br/>(env: LLAMA_ARG_REASONING_PRESERVE) |diff --git a/tools/server/README.md b/tools/server/README.mdindex a2ab872b4c..ec3a33e0c3 100644--- a/tools/server/README.md+++ b/tools/server/README.md@@ -226,6 +226,7 @@ For the full list of features, please refer to [server's changelog](https://gith| `--jinja, --no-jinja` | whether to use jinja template engine for chat (default: enabled)<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`<br/>- deepseek-legacy: keeps `<think>` tags in `message.content` while also populating `message.reasoning_content`<br/>(default: auto)<br/>(env: LLAMA_ARG_THINK) || `-rea, --reasoning [on\|off\|auto]` | Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))<br/>(env: LLAMA_ARG_REASONING) |+| `--reasoning-effort LEVEL` | reasoning effort level given to the chat template: 'default' to keep the template default,<br/>or a level such as 'minimal', 'low', 'medium', 'high', 'xhigh' or 'max' (default: default)<br/>(env: LLAMA_ARG_REASONING_EFFORT) || `--reasoning-budget N` | token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)<br/>(env: LLAMA_ARG_THINK_BUDGET) || `--reasoning-budget-message MESSAGE` | message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)<br/>(env: LLAMA_ARG_THINK_BUDGET_MESSAGE) || `--reasoning-preserve, --no-reasoning-preserve` | preserve reasoning trace in the full history, not just the last assistant message (default: template default)<br/>compatible with certain templates having 'supports_preserve_reasoning' capability<br/>example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking<br/>(env: LLAMA_ARG_REASONING_PRESERVE) |@@ -1250,7 +1251,7 @@ The `response_format` parameter supports both plain JSON output (e.g. `{"type":`chat_template_kwargs`: Allows sending additional parameters to the json templating system. For example: `{"enable_thinking": false}`-`reasoning_effort`: If set to `none`, reasoning will be disabled for this request. Other values (e.g., `low`, `max`) have no effect on reasoning.+`reasoning_effort`: If `none`, reasoning/thinking is disabled. Otherwise, the value is made available to the jinja template.`reasoning_format`: The reasoning format to be parsed. If set to `none`, it will output the raw generated text.diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cppindex 5ff7685bb1..7ed486528c 100644--- a/tools/server/server-common.cpp+++ b/tools/server/server-common.cpp@@ -1292,12 +1292,15 @@ json oaicompat_chat_params_parse(throw std::invalid_argument("invalid type for \"enable_thinking\" (expected boolean, got string)");}- // Parse also the OAI "reasoning_effort": "none" specific value+ // Parse the OAI "reasoning_effort" field; "none" disables reasoning.if (body.contains("reasoning_effort")) {auto reasoning_effort = json_value(body, "reasoning_effort", std::string(""));if (reasoning_effort == "none") {inputs.enable_thinking = false;- } // other reasoning_effort values are model-specific and not yet handled+ inputs.chat_template_kwargs.erase("reasoning_effort");+ } else if (!reasoning_effort.empty()) {+ inputs.chat_template_kwargs["reasoning_effort"] = json(reasoning_effort).dump();+ }}inputs.force_pure_content = opt.force_pure_content;