server : support jinja extra template kwargs (Qwen3 enable_thinking feature), from...

* initial commit for handling extra template kwargs

* enable_thinking and assistant prefill cannot be enabled at the same time

* can set chat_template_kwargs in command line

* added doc

* fixed formatting

* add support for extra context in generic template init

* coding standard: common/chat.cpp

Co-authored-by: Georgi Gerganov <redacted>
* coding standard:  common/chat.cpp

Co-authored-by: Georgi Gerganov <redacted>
* Apply suggestions from code review

coding standard: cosmetic changes

Co-authored-by: Georgi Gerganov <redacted>
* fix merge conflict

* chat.cpp: simplify calls to apply to ensure systematic propagation of extra_context (+ the odd existing additional_context)

* normalize environment variable name

* simplify code

* prefill cannot be used with thinking models

* compatibility with the new reasoning-budget parameter

* fix prefill for non thinking models

---------

Co-authored-by: Georgi Gerganov <redacted>
Co-authored-by: Olivier Chafik <redacted>

diff --git a/common/arg.cpp b/common/arg.cpp
index c4ad85c47b61b4ccd5f4624bf4fb6e46d9127b2e..40af7e574830f6b1d08f25d996ec7331612ceda6 100644 (file)
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -2794,6 +2794,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.ssl_file_cert = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_SSL_CERT_FILE"));
+    add_opt(common_arg(
+        {"--chat-template-kwargs"}, "STRING",
+        string_format("sets additional params for the json template parser"),
+        [](common_params & params, const std::string &  value) {
+            auto parsed = json::parse(value);
+            for (const auto & item : parsed.items()) {
+                params.default_template_kwargs[item.key()] = item.value().dump();
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_CHAT_TEMPLATE_KWARGS"));
     add_opt(common_arg(
         {"-to", "--timeout"}, "N",
         string_format("server read/write timeout in seconds (default: %d)", params.timeout_read),

diff --git a/common/chat.cpp b/common/chat.cpp
index 7d9aaeb12a1906115d68fba577740bf679978386..114dbfccdbfe78904325f75eeaceb83ef697ccf9 100644 (file)
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -17,6 +17,8 @@
 #include <string>
 #include <vector>
 
+using json = nlohmann::ordered_json;
+
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -140,6 +142,7 @@ struct templates_params {
     bool add_generation_prompt = true;
     bool enable_thinking = true;
     std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
+    json extra_context;
 };
 
 common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice) {
@@ -720,16 +723,23 @@ static void foreach_function(const json & tools, const std::function<void(const
 
 static std::string apply(
     const common_chat_template & tmpl,
-    const nlohmann::ordered_json & messages,
-    const nlohmann::ordered_json & tools,
-    bool add_generation_prompt,
-    const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
+    const struct templates_params & inputs,
+    const std::optional<json> & messages_override = std::nullopt,
+    const std::optional<json> & tools_override = std::nullopt,
+    const std::optional<json> & additional_context = std::nullopt)
 {
     minja::chat_template_inputs tmpl_inputs;
-    tmpl_inputs.messages = messages;
-    tmpl_inputs.tools = tools;
-    tmpl_inputs.add_generation_prompt = add_generation_prompt;
-    tmpl_inputs.extra_context = extra_context;
+    tmpl_inputs.messages = messages_override ? *messages_override : inputs.messages;
+    if (tools_override) {
+        tmpl_inputs.tools = *tools_override;
+    } else {
+        tmpl_inputs.tools = inputs.tools.empty() ? json() : inputs.tools;
+    }
+    tmpl_inputs.add_generation_prompt = inputs.add_generation_prompt;
+    tmpl_inputs.extra_context = inputs.extra_context;
+    if (additional_context) {
+        tmpl_inputs.extra_context.merge_patch(*additional_context);
+    }
     // TODO: add flag to control date/time, if only for testing purposes.
     // tmpl_inputs.now = std::chrono::system_clock::now();
 
@@ -828,7 +838,7 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp
         inputs.messages,
         "Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
 
-    data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ tweaked_messages);
     data.format = COMMON_CHAT_FORMAT_GENERIC;
     return data;
 }
@@ -904,7 +914,7 @@ static common_chat_params common_chat_params_init_mistral_nemo(const common_chat
     data.preserved_tokens = {
         "[TOOL_CALLS]",
     };
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
     return data;
 }
@@ -934,7 +944,7 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_
             adjusted_messages.push_back(msg);
         }
     }
-    data.prompt = apply(tmpl, adjusted_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {});
+    data.prompt = apply(tmpl, inputs, /* messages_override= */ adjusted_messages);
     data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
     if (string_ends_with(data.prompt, "<|START_THINKING|>")) {
         if (!inputs.enable_thinking) {
@@ -1122,7 +1132,7 @@ static common_chat_params common_chat_params_init_llama_3_x(const common_chat_te
     } else {
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, json {
         {"date_string", format_time(inputs.now, "%d %b %Y")},
         {"tools_in_user_message", false},
         {"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
@@ -1187,7 +1197,7 @@ static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool w
 
 static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
-    auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    auto prompt = apply(tmpl, inputs);
 
     // Hacks to fix the official (broken) prompt.
     // It is advisable to use --chat-template-file models/templates/llama-cpp-deepseek-r1.jinja instead,
@@ -1282,7 +1292,7 @@ static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) {
 static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) {
     LOG_DBG("%s\n", __func__);
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ json(), json {
         {"datetime", format_time(inputs.now, "%b %d %Y %H:%M:%S GMT")},
         {"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
     });
@@ -1338,7 +1348,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
     // Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
     // If the function is python, we also allow raw python code (if the line after `python\n` doesn't start w/ opening `{`), which the model seems to prefer for multiline code.
     common_chat_params data;
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
     if (inputs.tools.is_array() && !inputs.tools.empty()) {
         data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
@@ -1465,7 +1475,7 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con
         data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     }
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
+    data.prompt = apply(tmpl, inputs);
     // TODO: if (has_raw_python)
     return data;
 }
@@ -1498,14 +1508,15 @@ 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;
 
-    json additional_context = {
+    json extra_context = json {
         {"enable_thinking", inputs.enable_thinking},
     };
+    extra_context.update(inputs.extra_context);
 
-    data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, additional_context);
+    data.prompt = apply(tmpl, inputs, /* messages_override =*/ std::nullopt, /* tools_override= */ std::nullopt, extra_context);
     data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
     if (string_ends_with(data.prompt, "<think>\n")) {
-        if (!inputs.enable_thinking) {
+        if (!extra_context["enable_thinking"]) {
             data.prompt += "</think>";
         } else {
             data.thinking_forced_open = true;
@@ -1691,7 +1702,7 @@ static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) {
 
 static common_chat_params common_chat_params_init_without_tools(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);
+    data.prompt = apply(tmpl, inputs);
     data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
     data.grammar_lazy = false;
     if (!inputs.json_schema.is_null()) {
@@ -1722,6 +1733,12 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
+
+    params.extra_context = json::object();
+    for (auto el : inputs.chat_template_kwargs) {
+        params.extra_context[el.first] = json::parse(el.second);
+    }
+
     if (!inputs.json_schema.empty()) {
         params.json_schema = json::parse(inputs.json_schema);
     }

diff --git a/common/chat.h b/common/chat.h
index 9f59e6b08738dcc13e148d4cd5e71f053bd4536b..ca807c145ee82ee66ac0820b0dc7295a2e4dffc5 100644 (file)
--- a/common/chat.h
+++ b/common/chat.h
@@ -7,6 +7,7 @@
 #include <chrono>
 #include <string>
 #include <vector>
+#include <map>
 
 struct common_chat_templates;
 
@@ -125,6 +126,7 @@ struct common_chat_templates_inputs {
     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();
+    std::map<std::string, std::string> chat_template_kwargs;
 };
 
 struct common_chat_params {

diff --git a/common/common.h b/common/common.h
index e08a59eae75438818705ec8974a2918b3cf6f7ab..8922090e7b10d50e52f86f9851afdfbdb27c5015 100644 (file)
--- a/common/common.h
+++ b/common/common.h
@@ -8,6 +8,7 @@
 #include <string>
 #include <string_view>
 #include <vector>
+#include <map>
 #include <sstream>
 
 #ifdef _WIN32
@@ -381,6 +382,8 @@ struct common_params {
     std::string ssl_file_key  = "";                                                                         // NOLINT
     std::string ssl_file_cert = "";                                                                         // NOLINT
 
+    std::map<std::string, std::string> default_template_kwargs;
+
     // "advanced" endpoints are disabled by default for better security
     bool webui            = true;
     bool endpoint_slots   = false;

diff --git a/tools/server/README.md b/tools/server/README.md
index 1a624c13bee9658ea06c512c336afbf2c1ed1127..6f962664f67747f5c4386b646d22fa1dc650342c 100644 (file)
--- a/tools/server/README.md
+++ b/tools/server/README.md
@@ -164,6 +164,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--api-key-file FNAME` | path to file containing API keys (default: none) |
 | `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key<br/>(env: LLAMA_ARG_SSL_KEY_FILE) |
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
+| `--chat-template-kwargs STRING` | JSON object containing additional params for the json template parser. Example: `--chat_template_kwargs "{\"enable_thinking\":false}`"<br/>(env: LLAMA_CHAT_TEMPLATE_KWARGS) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
 | `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>[(card)](https://ggml.ai/f0.png)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
@@ -1118,6 +1119,8 @@ See [OpenAI Chat Completions API documentation](https://platform.openai.com/docs
 
 The `response_format` parameter supports both plain JSON output (e.g. `{"type": "json_object"}`) and schema-constrained JSON (e.g. `{"type": "json_object", "schema": {"type": "string", "minLength": 10, "maxLength": 100}}` or `{"type": "json_schema", "schema": {"properties": { "name": { "title": "Name",  "type": "string" }, "date": { "title": "Date",  "type": "string" }, "participants": { "items": {"type: "string" }, "title": "Participants",  "type": "string" } } } }`), similar to other OpenAI-inspired API providers.
 
+`chat_template_kwargs`: Allows sending additional parameters to the json templating system. For example: `{"enable_thinking": false}`
+
 *Examples:*
 
 You can use either Python `openai` library with appropriate checkpoints:

diff --git a/tools/server/server.cpp b/tools/server/server.cpp
index 852352383bdbe8eb27ac5acaa81a8f76b5036588..d3f6271931f62c5b16323e886fc9bea28d35b3af 100644 (file)
--- a/tools/server/server.cpp
+++ b/tools/server/server.cpp
@@ -2110,6 +2110,7 @@ struct server_context {
             /* use_jinja             */ params_base.use_jinja,
             /* prefill_assistant     */ params_base.prefill_assistant,
             /* reasoning_format      */ params_base.reasoning_format,
+            /* chat_template_kwargs  */ params_base.default_template_kwargs,
             /* common_chat_templates */ chat_templates.get(),
             /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
             /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,

diff --git a/tools/server/utils.hpp b/tools/server/utils.hpp
index f8fab2c86664e16ac89ebb96761dd4560fbc39a0..2ef9a164513c39a8c221ea1b66d29af1c89d0d26 100644 (file)
--- a/tools/server/utils.hpp
+++ b/tools/server/utils.hpp
@@ -579,6 +579,7 @@ struct oaicompat_parser_options {
     bool use_jinja;
     bool prefill_assistant;
     common_reasoning_format reasoning_format;
+    std::map<std::string,std::string> chat_template_kwargs;
     common_chat_templates * tmpls;
     bool allow_image;
     bool allow_audio;
@@ -756,6 +757,13 @@ static json oaicompat_chat_params_parse(
         llama_params["parse_tool_calls"] = true;
     }
 
+    // merge the template args provided from command line with the args provided in the user request
+    auto chat_template_kwargs_object = json_value(body, "chat_template_kwargs", json::object());
+    inputs.chat_template_kwargs = opt.chat_template_kwargs;
+    for (const auto & item : chat_template_kwargs_object.items()) {
+        inputs.chat_template_kwargs[item.key()] = item.value().dump();
+    }
+
     // if the assistant message appears at the end of list, we do not add end-of-turn token
     // for ex. this can be useful to modify the reasoning process in reasoning models
     bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant" && opt.prefill_assistant;
@@ -771,6 +779,11 @@ static json oaicompat_chat_params_parse(
 
         /* TODO: test this properly */
         inputs.reasoning_format = COMMON_REASONING_FORMAT_NONE;
+
+        if ( (!inputs.enable_thinking) || inputs.chat_template_kwargs.find("enable_thinking") != inputs.chat_template_kwargs.end()) {
+            throw std::runtime_error("Assistant response prefill is incompatible with enable_thinking.");
+        }
+
         inputs.add_generation_prompt = true;
     }