common: add json.h abstraction (#27511)

* add common/json

* migrate common

* adapt jinja

* migrate server

* big wip

* migrate tests

* wip

* revert some excessive changes

* wip

* wip 2

* revert redundant changes

* fix server crash

* various fixes

* fix ci

* harden a bit

* clean up

* rm json-shim

* add some comments

* rm redundant decl

tools/cli/cli-context.cpp

@@ -6,8 +6,7 @@
 #include "log.h"
 #include "console.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <cctype>
@@ -16,7 +15,7 @@
 #include <map>
 #include <set>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 struct cli_context_impl {
     json messages      = json::array();
@@ -73,7 +72,7 @@ static std::string format_error_message(const json & err) {
 
 // err is the raw response body of a failed request; it may or may not be JSON
 static std::string format_error_message(const std::string & err) {
-    json parsed = json::parse(err, nullptr, false);
+    json parsed = json::parse_no_throw(err);
     if (!parsed.is_discarded()) {
         return format_error_message(parsed);
     }
@@ -157,7 +156,7 @@ bool cli_context::init() {
             if (!list_and_ask_models()) {
                 return false;
             }
-        } catch (const json::parse_error & e) {
+        } catch (const common_json_error & e) {
             ui::show_error(e.what());
             ui::show_message("This might be caused by an incorrect server-base endpoint URL");
             return false;
@@ -364,7 +363,7 @@ bool cli_context::generate_completion(generated_content & content_out, cli_timin
     ui::assistant_turn a;
 
     std::string err = client.post_sse("/v1/chat/completions", body.dump(), should_stop, [&](const std::string & payload) {
-        json chunk = json::parse(payload, nullptr, false);
+        json chunk = json::parse_no_throw(payload);
         if (chunk.is_discarded()) {
             return;
         }

tools/parser/debug-template-parser.cpp

@@ -5,7 +5,7 @@
 #include "gguf.h"
 #include "jinja/runtime.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
+#include "json.h"
 #include "peg-parser.h"
 
 #include <fstream>
@@ -15,7 +15,7 @@
 #include <sstream>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 enum class output_mode {
     ANALYSIS,  // Only output analysis results (default)

tools/parser/template-analysis.cpp

@@ -11,9 +11,9 @@
 #include <vector>
 #include <algorithm>
 
-#include "nlohmann/json.hpp"
+#include "json.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // ANSI color codes - using 256-color palette for brighter colors (all bold)
 #define ANSI_RESET       "\033[0m"

tools/server/server-chat.cpp

@@ -153,7 +153,7 @@ json server_chat_convert_responses_to_chatcmpl(const json & response_body) {
                         prev_msg["content"] = json::array();
                     }
                     auto & prev_content = prev_msg["content"];
-                    prev_content.insert(prev_content.end(), chatcmpl_content.begin(), chatcmpl_content.end());
+                    prev_content.insert(chatcmpl_content);
                 } else {
                     item.erase("status");
                     item.erase("type");

tools/server/server-chat.h

@@ -6,9 +6,7 @@
 #include "server-common.h"
 #include "server-http.h"
 
-#include <nlohmann/json_fwd.hpp>
-
-using json = nlohmann::ordered_json;
+#include "json.h"
 
 // Convert OpenAI Responses API format to OpenAI Chat Completions API format
 json server_chat_convert_responses_to_chatcmpl(const json & body);

tools/server/server-common.cpp

@@ -1540,7 +1540,7 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
 }
 
 std::string safe_json_to_str(const json & data) {
-    return data.dump(-1, ' ', false, json::error_handler_t::replace);
+    return data.dump_safe();
 }
 
 // TODO: reuse llama_detokenize

tools/server/server-common.h

@@ -6,8 +6,7 @@
 #include "chat.h"
 #include "mtmd.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <atomic>
 #include <chrono>
@@ -19,7 +18,7 @@
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 #define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
 #define SLT_TRC(slot, fmt, ...) LOG_TRC("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
@@ -42,9 +41,9 @@ static T json_value(const json & body, const std::string & key, const T & defaul
     // Fallback null to default value
     if (body.contains(key) && !body.at(key).is_null()) {
         try {
-            return body.at(key);
-        } catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const & err) {
-            LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value: %s\n", key.c_str(), json(default_value).type_name(), err.what());
+            return body.at(key).get<T>();
+        } catch (const common_json_error & err) {
+            LOG_WRN("Wrong type supplied for parameter '%s', using default value: %s\n", key.c_str(), err.what());
             return default_value;
         }
     } else {

tools/server/server-context.cpp

@@ -35,8 +35,6 @@
 #include <windows.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 constexpr int HTTP_POLLING_SECONDS = 1;
 
 static common_speculative_output_limits server_output_limits(const common_params & params) {
@@ -657,14 +655,14 @@ struct server_slot {
             res["n_prompt_tokens_processed"] = stats.n_prompt_processed;
             res["n_prompt_tokens_cache"]     = stats.n_prompt_cached;
             res["params"] = ptask->params.to_json(only_metrics);
-            res["next_token"] = {
+            res["next_token"] = json::array({
                 {
                     {"has_next_token", has_next_token},
                     {"has_new_line",   has_new_line},
                     {"n_remain",       n_remaining()},
                     {"n_decoded",      stats.n_gen},
                 }
-            };
+            });
 
             if (!only_metrics) {
                 res["prompt"] = ptask->tokens.detokenize(ctx_tgt, true);
@@ -4165,7 +4163,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         // tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks
 
         // message delimiters for checkpointing
-        auto delimiters = common_chat_msg_delimiters_parse(json_value(data, "message_delimiters", json::array()));
+        json delims = json_value(data, "message_delimiters", json::array());
+        auto delimiters = common_chat_msg_delimiters_parse(delims);
         delimiters.tokenize(ctx_server.vocab);
 
         for (size_t i = 0; i < inputs.size(); i++) {
@@ -4428,8 +4427,8 @@ static json get_res_model_info(const server_context_meta & meta) {
 static json get_res_models(const server_context_meta & meta) {
     // note: do NOT use ctx_server here, otherwise it's not possible to use this during sleep
 
-    return {
-        {"models", {
+    return json{
+        {"models", json::array({
             {
                 {"name",  meta.model_name},
                 {"model", meta.model_name},
@@ -4438,23 +4437,23 @@ static json get_res_models(const server_context_meta & meta) {
                 {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
                 {"type", "model"},
                 {"description", ""},
-                {"tags", {""}},
-                {"capabilities", meta.has_mtmd ? json({"completion","multimodal"}) : json({"completion"})},
+                {"tags", json::array({""})},
+                {"capabilities", meta.has_mtmd ? json::array({"completion","multimodal"}) : json::array({"completion"})},
                 {"parameters", ""},
                 {"details", {
                     {"parent_model", ""},
                     {"format", "gguf"},
                     {"family", ""},
-                    {"families", {""}},
+                    {"families", json::array({""})},
                     {"parameter_size", ""},
                     {"quantization_level", ""}
                 }}
             }
-        }},
+        })},
         {"object", "list"},
-        {"data", {
+        {"data", json::array({
             get_res_model_info(meta),
-        }}
+        })}
     };
 }
 
@@ -4990,7 +4989,7 @@ void server_routes::init_routes() {
 
         std::string content;
         if (body.count("tokens") != 0) {
-            const llama_tokens tokens = body.at("tokens");
+            const llama_tokens tokens = body.at("tokens").get<llama_tokens>();
             content = tokens_to_str(ctx_server.vocab, tokens);
         }
 
@@ -5297,7 +5296,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
 
     int embd_normalize = params.embd_normalize;
     if (body.count("embd_normalize") != 0) {
-        embd_normalize = body.at("embd_normalize");
+        embd_normalize = body.at("embd_normalize").get<int>();
         if (meta->pooling_type == LLAMA_POOLING_TYPE_NONE) {
             SRV_DBG("embd_normalize is not supported by pooling type %d, ignoring it\n", meta->pooling_type);
         }

tools/server/server-context.h

@@ -4,7 +4,7 @@
 #include "server-task.h"
 #include "server-queue.h"
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <cstddef>
 #include <memory>

tools/server/server-models.cpp

@@ -2462,7 +2462,7 @@ server_http_proxy::server_http_proxy(
     bool has_files = !files.empty();
 
     if (has_files) {
-        json form_fields = json::parse(body, nullptr, false);
+        json form_fields = json::parse_no_throw(body);
         if (!form_fields.is_discarded()) {
             auto boundary = generate_multipart_boundary();
             effective_body = build_multipart_body(form_fields, files, boundary);

tools/server/server-schema.cpp

@@ -503,7 +503,7 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
         ->set_handler([&](field_eval_context & ctx, const json & data) {
             const auto & samplers = data.at("samplers");
             if (samplers.is_array()) {
-                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers);
+                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers.get<std::vector<std::string>>());
             } else if (samplers.is_string()) {
                 ctx.params.sampling.samplers = common_sampler_types_from_chars(samplers.get<std::string>());
             }
@@ -580,8 +580,7 @@ static void handle_with_catch(const char * name, std::function<void()> func) {
 
 // treat a null value as absent so clients can send null to request the server default
 static bool has_value(const json & data, const char * n) {
-    auto it = data.find(n);
-    return it != data.end() && !it->is_null();
+    return data.contains(n) && !data.at(n).is_null();
 }
 
 template <typename T>

tools/server/server-task.cpp

@@ -12,8 +12,6 @@
 
 #include <sstream>
 
-using json = nlohmann::ordered_json;
-
 //
 // task_params
 //
@@ -304,7 +302,7 @@ json completion_token_output::probs_vector_to_json(const std::vector<completion_
 }
 
 float completion_token_output::logarithm(float x) {
-    // nlohmann::json converts -inf to null, so we need to prevent that
+    // the JSON library converts -inf to null, so we need to prevent that
     return x == 0.0f ? std::numeric_limits<float>::lowest() : std::log(x);
 }
 
@@ -407,7 +405,7 @@ json server_task_result_cmpl_final::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -455,7 +453,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -516,7 +514,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
     }
 
     if (stats.is_set()) {
-        deltas.back().push_back({"timings", stats.to_json()});
+        deltas.back()["timings"] = stats.to_json();
     }
 
     // extra fields for debugging purposes
@@ -709,7 +707,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
     });
 
     if (stats.is_set()) {
-        server_sent_events.back().at("data").push_back({"timings", stats.to_json()});
+        server_sent_events.back().at("data")["timings"] = stats.to_json();
     }
 
     return server_sent_events;
@@ -1061,10 +1059,10 @@ json server_task_result_cmpl_partial::to_json_non_oaicompat() {
     };
     // populate the timings object when needed (usually for the last response or with timings_per_token enabled)
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
     if (!prob_output.probs.empty()) {
         res["completion_probabilities"] = completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs);
@@ -1101,10 +1099,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
 
     return res;
@@ -1155,10 +1153,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
         }
 
         if (stats.is_set()) {
-            last_json.push_back({"timings", stats.to_json()});
+            last_json["timings"] = stats.to_json();
         }
         if (is_progress) {
-            last_json.push_back({"prompt_progress", progress.to_json()});
+            last_json["prompt_progress"] = progress.to_json();
         }
     }
 
@@ -1305,10 +1303,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
     if (!events.empty()) {
         json & data = events.back().at("data");
         if (stats.is_set()) {
-            data.push_back({"timings", stats.to_json()});
+            data["timings"] = stats.to_json();
         }
         if (is_progress) {
-            data.push_back({"prompt_progress", progress.to_json()});
+            data["prompt_progress"] = progress.to_json();
         }
     }
 

tools/server/server-task.h

@@ -11,7 +11,6 @@
 // TODO: prevent including the whole server-common.h as we only use server_tokens
 #include "server-common.h"
 
-using json = nlohmann::ordered_json;
 
 enum server_task_type {
     SERVER_TASK_TYPE_COMPLETION,

tools/server/server-tools.cpp

@@ -2156,7 +2156,7 @@ void server_tools::setup(const std::vector<std::string> & enabled_tools,
                 res->status = 200;
                 res->data   = safe_json_to_str(result);
             }
-        } catch (const json::exception & e) {
+        } catch (const common_json_error & e) {
             res->status = 400;
             res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
         } catch (const std::invalid_argument & e) {