migrate server

common/chat.cpp

@@ -973,7 +973,7 @@ static std::string common_chat_template_direct_apply_impl(
         jinja::caps_apply_reasoning_effort(ctx, reasoning_effort);
     }
 
-    jinja::global_from_json(ctx, common_json_ref_from_raw(inp), inputs.mark_input);
+    jinja::global_from_json(ctx, common_json_from_raw(inp), inputs.mark_input);
 
     // render
     jinja::runtime runtime(ctx);

common/chat.h

@@ -17,7 +17,6 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
-using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 

common/jinja/value.cpp

@@ -1355,7 +1355,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const common_json_ref & j, bool mark_input) {
+static value from_json(const common_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1462,8 +1462,7 @@ void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input) {
     }
 }
 
-template void global_from_json<common_json>    (context &, const common_json &,     bool);
-template void global_from_json<common_json_ref>(context &, const common_json_ref &, bool);
+template void global_from_json<common_json>(context &, const common_json &, bool);
 
 // recursively convert value to JSON string
 // TODO: avoid circular references

common/jinja/value.h

@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be common_json or common_json_ref
+// Note: T_JSON can be common_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 

common/json.cpp

@@ -6,27 +6,30 @@
 #include <nlohmann/json.hpp>
 
 #include <iterator>
+#include <new>
+#include <set>
 #include <vector>
 
 using nlohmann::ordered_json;
 
-// common_json_node is never defined, it only stands for an ordered_json in this file
-static ordered_json & as_json(common_json_node * node) {
-    return *reinterpret_cast<ordered_json *>(node);
+// a common_json is the backing value, so any value of a tree can be used as a common_json
+static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
+static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
+
+static ordered_json & as_json(common_json * self) {
+    return *reinterpret_cast<ordered_json *>(self);
 }
 
-static common_json_node * as_node(ordered_json * json) {
-    return reinterpret_cast<common_json_node *>(json);
+static const ordered_json & as_json(const common_json * self) {
+    return *reinterpret_cast<const ordered_json *>(self);
 }
 
-void common_json_node_deleter::operator()(common_json_node * node) const {
-    delete reinterpret_cast<ordered_json *>(node);
+static common_json & as_common(ordered_json & json) {
+    return *reinterpret_cast<common_json *>(&json);
 }
 
-static common_json make_json(const ordered_json & val) {
-    common_json out;
-    as_json(out.get_node()) = val;
-    return out;
+static const common_json & as_common(const ordered_json & json) {
+    return *reinterpret_cast<const common_json *>(&json);
 }
 
 static ordered_json to_json(const common_json_value & val) {
@@ -37,33 +40,33 @@ static ordered_json to_json(const common_json_value & val) {
         case common_json_value::VAL_UINT:   return val.val_uint;
         case common_json_value::VAL_DOUBLE: return val.val_double;
         case common_json_value::VAL_STRING: return val.val_string;
-        case common_json_value::VAL_JSON:   return as_json(val.val_json->get_node());
+        case common_json_value::VAL_JSON:   return as_json(val.val_json.get());
     }
 
     return nullptr;
 }
 
-template <typename T> T & common_json_raw(common_json_ref & json) {
-    return as_json(json.get_node());
+template <typename T> T & common_json_raw(common_json & json) {
+    return as_json(&json);
 }
 
-template <typename T> const T & common_json_raw(const common_json_ref & json) {
-    return as_json(json.get_node());
+template <typename T> const T & common_json_raw(const common_json & json) {
+    return as_json(&json);
 }
 
 template <typename T> common_json common_json_from_raw(const T & json) {
-    return make_json(json);
+    return common_json(as_common(json));
 }
 
-template <typename T> common_json_ref common_json_ref_from_raw(T & json) {
-    return common_json_ref(as_node(&json));
+template <typename T> common_json & common_json_ref_from_raw(T & json) {
+    return as_common(json);
 }
 
 // the bridge is usable only for the type below
-template ordered_json       & common_json_raw<ordered_json>(common_json_ref &);
-template const ordered_json & common_json_raw<ordered_json>(const common_json_ref &);
+template ordered_json       & common_json_raw<ordered_json>(common_json &);
+template const ordered_json & common_json_raw<ordered_json>(const common_json &);
 template common_json          common_json_from_raw<ordered_json>(const ordered_json &);
-template common_json_ref      common_json_ref_from_raw<ordered_json>(ordered_json &);
+template common_json        & common_json_ref_from_raw<ordered_json>(ordered_json &);
 
 common_json_value::common_json_value(const char * val) {
     if (val) {
@@ -77,86 +80,29 @@ common_json_value::common_json_value(const char * val) {
 common_json_value::common_json_value(const common_json & val) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
 
+common_json_value::common_json_value(const std::vector<std::string> & vals) : type(VAL_JSON) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
 common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
     type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
 
-common_json_value::common_json_value(const common_json_ref & val) :
-    type(VAL_JSON), val_json(std::make_shared<common_json>(make_json(as_json(val.get_node())))) {}
-
-bool common_json_ref::is_null()           const { return as_json(node).is_null(); }
-bool common_json_ref::is_object()         const { return as_json(node).is_object(); }
-bool common_json_ref::is_array()          const { return as_json(node).is_array(); }
-bool common_json_ref::is_string()         const { return as_json(node).is_string(); }
-bool common_json_ref::is_boolean()        const { return as_json(node).is_boolean(); }
-bool common_json_ref::is_number()         const { return as_json(node).is_number(); }
-bool common_json_ref::is_number_integer() const { return as_json(node).is_number_integer(); }
-bool common_json_ref::is_number_float()   const { return as_json(node).is_number_float(); }
-
-bool   common_json_ref::empty() const { return as_json(node).empty(); }
-size_t common_json_ref::size()  const { return as_json(node).size(); }
-
-bool common_json_ref::contains(const std::string & key) const {
-    return as_json(node).contains(key);
+common_json::common_json() {
+    new (storage) ordered_json(ordered_json::object());
 }
 
-bool common_json_ref::operator==(const common_json_value & val) const {
-    return as_json(node) == to_json(val);
+common_json::common_json(const common_json & other) {
+    new (storage) ordered_json(as_json(&other));
 }
 
-bool common_json_ref::operator!=(const common_json_value & val) const {
-    return !(*this == val);
-}
-
-common_json_ref common_json_ref::at(const std::string & key) const {
-    return common_json_ref(as_node(&as_json(node).at(key)));
-}
-
-common_json_ref common_json_ref::operator[](const std::string & key) const {
-    return common_json_ref(as_node(&as_json(node)[key]));
-}
-
-common_json_ref common_json_ref::operator[](size_t idx) const {
-    return common_json_ref(as_node(&as_json(node)[idx]));
-}
-
-void common_json_ref::assign(const common_json_value & val) {
-    as_json(node) = to_json(val);
-}
-
-void common_json_ref::set(const common_json_item & item) {
-    as_json(node)[item.key] = to_json(item.val);
-}
-
-void common_json_ref::push_back(const common_json_value & val) {
-    as_json(node).push_back(to_json(val));
-}
-
-std::string common_json_ref::dump(int indent) const {
-    return as_json(node).dump(indent);
-}
-
-// an array is indexed directly, an object needs a walk from the start
-common_json_ref common_json_ref::iterator::operator*() const {
-    if (as_json(node).is_object()) {
-        return common_json_ref(as_node(&std::next(as_json(node).begin(), idx).value()));
-    }
-
-    return common_json_ref(as_node(&as_json(node)[idx]));
-}
-
-std::string common_json_ref::iterator::key() const {
-    return std::next(as_json(node).begin(), idx).key();
-}
-
-std::pair<std::string, common_json_ref> common_json_ref::items_view::iterator::operator*() const {
-    auto it = std::next(as_json(node).begin(), idx);
-
-    return { it.key(), common_json_ref(as_node(&it.value())) };
-}
-
-common_json::common_json() :
-    common_json_ref(nullptr), pimpl(as_node(new ordered_json(ordered_json::object()))) {
-    node = pimpl.get();
+common_json::common_json(common_json && other) noexcept {
+    new (storage) ordered_json(std::move(as_json(&other)));
 }
 
 common_json::common_json(std::initializer_list<common_json_item> items) : common_json() {
@@ -165,43 +111,40 @@ common_json::common_json(std::initializer_list<common_json_item> items) : common
     }
 }
 
-common_json::common_json(const common_json & other) :
-    common_json_ref(nullptr), pimpl(as_node(new ordered_json(as_json(other.node)))) {
-    node = pimpl.get();
+common_json::common_json(const common_json_value & val) {
+    new (storage) ordered_json(to_json(val));
 }
 
-common_json::common_json(common_json && other) noexcept :
-    common_json_ref(other.node), pimpl(std::move(other.pimpl)) {
-    other.node = nullptr;
+common_json::common_json(std::nullptr_t) {
+    new (storage) ordered_json(nullptr);
 }
 
 common_json & common_json::operator=(const common_json & other) {
-    as_json(node) = as_json(other.node);
+    as_json(this) = as_json(&other);
 
     return *this;
 }
 
 common_json & common_json::operator=(common_json && other) noexcept {
-    pimpl = std::move(other.pimpl);
-    node  = pimpl.get();
-
-    other.node = nullptr;
+    as_json(this) = std::move(as_json(&other));
 
     return *this;
 }
 
-common_json::~common_json() = default;
+common_json::~common_json() {
+    as_json(this).~basic_json();
+}
 
 common_json common_json::parse(const std::string & text) {
     try {
-        return make_json(ordered_json::parse(text));
+        return common_json_from_raw(ordered_json::parse(text));
     } catch (const std::exception & e) {
         throw common_json_error(e.what());
     }
 }
 
 common_json common_json::array() {
-    return make_json(ordered_json::array());
+    return common_json_from_raw(ordered_json::array());
 }
 
 common_json common_json::array(std::initializer_list<common_json_value> vals) {
@@ -211,7 +154,7 @@ common_json common_json::array(std::initializer_list<common_json_value> vals) {
         out.push_back(to_json(val));
     }
 
-    return make_json(out);
+    return common_json_from_raw(out);
 }
 
 common_json common_json::object() {
@@ -219,16 +162,114 @@ common_json common_json::object() {
 }
 
 common_json common_json::make(const common_json_value & val) {
-    return make_json(to_json(val));
+    return common_json(val);
 }
 
-template <typename T> T common_json_ref::get() const {
-    return as_json(node).get<T>();
+bool common_json::is_null()           const { return as_json(this).is_null(); }
+bool common_json::is_object()         const { return as_json(this).is_object(); }
+bool common_json::is_array()          const { return as_json(this).is_array(); }
+bool common_json::is_string()         const { return as_json(this).is_string(); }
+bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
+bool common_json::is_number()         const { return as_json(this).is_number(); }
+bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
+bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
+
+bool   common_json::empty() const { return as_json(this).empty(); }
+size_t common_json::size()  const { return as_json(this).size(); }
+
+bool common_json::contains(const std::string & key) const {
+    return as_json(this).contains(key);
+}
+
+bool common_json::operator==(const common_json_value & val) const {
+    return as_json(this) == to_json(val);
+}
+
+bool common_json::operator!=(const common_json_value & val) const {
+    return !(*this == val);
+}
+
+common_json       & common_json::at(const std::string & key)       { return as_common(as_json(this).at(key)); }
+const common_json & common_json::at(const std::string & key) const { return as_common(as_json(this).at(key)); }
+common_json       & common_json::at(size_t idx)                    { return as_common(as_json(this).at(idx)); }
+const common_json & common_json::at(size_t idx)              const { return as_common(as_json(this).at(idx)); }
+
+common_json       & common_json::operator[](const std::string & key)       { return as_common(as_json(this)[key]); }
+const common_json & common_json::operator[](const std::string & key) const { return as_common(as_json(this).at(key)); }
+common_json       & common_json::operator[](size_t idx)                    { return as_common(as_json(this)[idx]); }
+const common_json & common_json::operator[](size_t idx)              const { return as_common(as_json(this).at(idx)); }
+
+common_json       & common_json::front()       { return as_common(as_json(this).front()); }
+const common_json & common_json::front() const { return as_common(as_json(this).front()); }
+common_json       & common_json::back()        { return as_common(as_json(this).back()); }
+const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
+
+void common_json::erase(const std::string & key) {
+    as_json(this).erase(key);
+}
+
+void common_json::erase(size_t idx) {
+    as_json(this).erase(idx);
+}
+
+void common_json::assign(const common_json_value & val) {
+    as_json(this) = to_json(val);
+}
+
+void common_json::set(const common_json_item & item) {
+    as_json(this)[item.key] = to_json(item.val);
+}
+
+void common_json::push_back(const common_json_value & val) {
+    as_json(this).push_back(to_json(val));
+}
+
+std::string common_json::dump(int indent) const {
+    return as_json(this).dump(indent);
+}
+
+std::string common_json::dump_safe(int indent) const {
+    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
+}
+
+// an array is indexed directly, an object needs a walk from the start
+common_json & common_json::iterator::operator*() const {
+    if (as_json(node).is_object()) {
+        return as_common(std::next(as_json(node).begin(), idx).value());
+    }
+
+    return as_common(as_json(node)[idx]);
+}
+
+std::string common_json::iterator::key() const {
+    return std::next(as_json(node).begin(), idx).key();
+}
+
+common_json::iterator common_json::begin() const {
+    return iterator(const_cast<common_json *>(this), 0);
+}
+
+common_json::iterator common_json::end() const {
+    return iterator(const_cast<common_json *>(this), size());
+}
+
+common_json::items_view::entry common_json::items_view::iterator::operator*() const {
+    auto it = std::next(as_json(node).begin(), idx);
+
+    return { it.key(), as_common(it.value()) };
+}
+
+common_json::items_view common_json::items() const {
+    return items_view(const_cast<common_json *>(this), size());
+}
+
+template <typename T> T common_json::get() const {
+    return as_json(this).get<T>();
 }
 
 // get<T>() is usable only for the types below
 
-#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json_ref::get<__VA_ARGS__>() const;
+#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
 
 COMMON_JSON_GET(bool)
 COMMON_JSON_GET(int)
@@ -241,5 +282,6 @@ COMMON_JSON_GET(float)
 COMMON_JSON_GET(double)
 COMMON_JSON_GET(std::string)
 COMMON_JSON_GET(std::vector<std::string>)
+COMMON_JSON_GET(std::set<std::string>)
 
 #undef COMMON_JSON_GET

common/json.h

@@ -2,6 +2,7 @@
 
 // JSON object, it works without the need to include a JSON library header
 // the underlay library is pimpl, it should never be exposed here
+// the backing value lives inside this object, so at() and the iterators give a real reference to it
 // note: object keys keep the order in which they are added
 
 #include <cstddef>
@@ -12,20 +13,13 @@
 #include <string>
 #include <type_traits>
 #include <utility>
+#include <vector>
 
 class common_json;
-class common_json_ref;
 
 // common_json_value holds a list of these, and each of them holds a value, so one must come first
 struct common_json_item;
 
-// one value of the backing library, only json.cpp knows what it is
-struct common_json_node;
-
-struct common_json_node_deleter {
-    void operator()(common_json_node * node) const;
-};
-
 struct common_json_error : std::runtime_error {
     using std::runtime_error::runtime_error;
 };
@@ -59,7 +53,7 @@ struct common_json_value {
     common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
     common_json_value(const char * val);
     common_json_value(const common_json & val);
-    common_json_value(const common_json_ref & val);
+    common_json_value(const std::vector<std::string> & vals);
 
     // nested object, e.g. {"fn", {{"name", "x"}}}
     common_json_value(std::initializer_list<common_json_item> items);
@@ -90,15 +84,31 @@ struct common_json_item {
         key(std::move(key)), val(items) {}
 };
 
-// view to a value owned by a common_json, it goes stale if the owner gets a new key
-class common_json_ref {
+class common_json {
   public:
-    explicit common_json_ref(common_json_node * node) : node(node) {}
+    common_json();
+    common_json(const common_json & other);
+    common_json(common_json && other) noexcept;
+    common_json(std::initializer_list<common_json_item> items);
+    common_json(const common_json_value & val);
 
-    common_json_ref(const common_json_ref &) = default;
+    // direct, a value would need two conversions in a row
+    common_json(std::nullptr_t);
 
-    // rebinding a view is almost always a write-through by mistake, use assign() to write
-    common_json_ref & operator=(const common_json_ref &) = delete;
+    common_json & operator=(const common_json & other);
+    common_json & operator=(common_json && other) noexcept;
+
+    ~common_json();
+
+    // throws common_json_error if the text is not valid JSON
+    static common_json parse(const std::string & text);
+
+    static common_json array();
+    static common_json array(std::initializer_list<common_json_value> vals);
+    static common_json object();
+
+    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
+    static common_json make(const common_json_value & val);
 
     bool is_null()    const;
     bool is_object()  const;
@@ -118,9 +128,23 @@ class common_json_ref {
     bool operator!=(const common_json_value & val) const;
 
     // at() throws if the key is missing, operator[] adds a null value instead
-    common_json_ref at(const std::string & key) const;
-    common_json_ref operator[](const std::string & key) const;
-    common_json_ref operator[](size_t idx) const;
+    common_json       & at(const std::string & key);
+    const common_json & at(const std::string & key) const;
+    common_json       & at(size_t idx);
+    const common_json & at(size_t idx) const;
+
+    common_json       & operator[](const std::string & key);
+    const common_json & operator[](const std::string & key) const;
+    common_json       & operator[](size_t idx);
+    const common_json & operator[](size_t idx) const;
+
+    common_json       & front();
+    const common_json & front() const;
+    common_json       & back();
+    const common_json & back()  const;
+
+    void erase(const std::string & key);
+    void erase(size_t idx);
 
     // only for the types instantiated in json.cpp, the rest fails at link time
     template <typename T> T get() const;
@@ -138,21 +162,26 @@ class common_json_ref {
     void set(const common_json_item & item);
     void push_back(const common_json_value & val);
 
-    template <typename T>
-    common_json_ref & operator=(T && val) {
+    // a common_json goes through the copy assignment above, everything else becomes a value
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
+    common_json & operator=(T && val) {
         assign(common_json_value(std::forward<T>(val)));
         return *this;
     }
 
     std::string dump(int indent = -1) const;
 
+    // same as dump(), but bad UTF-8 gets replaced instead of throwing
+    std::string dump_safe(int indent = -1) const;
+
     // walks an array by index, or an object in insertion order
     class iterator {
       public:
-        iterator(common_json_node * node, size_t idx) : node(node), idx(idx) {}
+        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
 
-        common_json_ref operator*() const;
-        std::string     key()       const;
+        common_json & operator*() const;
+        common_json & value()     const { return **this; }
+        std::string   key()       const;
 
         iterator & operator++() {
             idx++;
@@ -163,23 +192,32 @@ class common_json_ref {
         bool operator==(const iterator & other) const { return idx == other.idx; }
 
       private:
-        common_json_node * node;
-        size_t             idx;
+        common_json * node;
+        size_t        idx;
     };
 
-    iterator begin() const { return iterator(node, 0); }
-    iterator end()   const { return iterator(node, size()); }
+    iterator begin() const;
+    iterator end()   const;
 
     // allows: for (const auto & [key, val] : obj.items())
     class items_view {
       public:
-        items_view(common_json_node * node, size_t n) : node(node), n(n) {}
+        // the members are public, so an entry also works with structured bindings
+        struct entry {
+            std::string   k;
+            common_json & v;
+
+            const std::string & key()   const { return k; }
+            common_json &       value() const { return v; }
+        };
+
+        items_view(common_json * node, size_t n) : node(node), n(n) {}
 
         class iterator {
           public:
-            iterator(common_json_node * node, size_t idx) : node(node), idx(idx) {}
+            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
 
-            std::pair<std::string, common_json_ref> operator*() const;
+            entry operator*() const;
 
             iterator & operator++() {
                 idx++;
@@ -189,62 +227,35 @@ class common_json_ref {
             bool operator!=(const iterator & other) const { return idx != other.idx; }
 
           private:
-            common_json_node * node;
-            size_t             idx;
+            common_json * node;
+            size_t        idx;
         };
 
         iterator begin() const { return iterator(node, 0); }
         iterator end()   const { return iterator(node, n); }
 
       private:
-        common_json_node * node;
-        size_t             n;
+        common_json * node;
+        size_t        n;
     };
 
-    items_view items() const { return items_view(node, size()); }
-
-    common_json_node * get_node() const { return node; }
-
-  protected:
-    common_json_node * node;
-};
-
-// owns the value it points to
-class common_json : public common_json_ref {
-  public:
-    common_json();
-    common_json(std::initializer_list<common_json_item> items);
-    common_json(const common_json & other);
-    common_json(common_json && other) noexcept;
-
-    common_json & operator=(const common_json & other);
-    common_json & operator=(common_json && other) noexcept;
-
-    // out-of-line, the deleter needs to know the real type
-    ~common_json();
-
-    // throws common_json_error if the text is not valid JSON
-    static common_json parse(const std::string & text);
-
-    static common_json array();
-    static common_json array(std::initializer_list<common_json_value> vals);
-    static common_json object();
-
-    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
-    static common_json make(const common_json_value & val);
+    items_view items() const;
 
   private:
-    std::unique_ptr<common_json_node, common_json_node_deleter> pimpl;
+    // the backing value is built here, json.cpp checks that it fits
+    alignas(8) unsigned char storage[32];
 };
 
+using common_json_entry = common_json::items_view::entry;
+
 // bridge for code that still uses internal component from nlohmann::json
 // usage: common_json_raw<nlohmann::ordered_json>(j)
 // TODO: maybe completely remove this in the future
 
-template <typename T> T       & common_json_raw(common_json_ref & json);
-template <typename T> const T & common_json_raw(const common_json_ref & json);
+template <typename T> T       & common_json_raw(common_json & json);
+template <typename T> const T & common_json_raw(const common_json & json);
 
 template <typename T> common_json common_json_from_raw(const T & json);
 
 // view over a value of the backing library, it does not copy
-template <typename T> common_json_ref common_json_ref_from_raw(T & json);
+template <typename T> common_json & common_json_ref_from_raw(T & json);

tools/server/server-chat.h

@@ -6,9 +6,8 @@
 #include "server-common.h"
 #include "server-http.h"
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
-using json = nlohmann::ordered_json;
 
 // 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

@@ -9,6 +9,9 @@
 
 #include "server-common.h"
 
+// the chat API is not migrated yet, so this file still needs the bridge
+#include <nlohmann/json.hpp>
+
 #include <random>
 #include <sstream>
 #include <fstream>
@@ -977,9 +980,9 @@ static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_co
             // JSON object with prompt and multimodal key.
             std::vector<raw_buffer> files;
             for (const auto & entry : json_prompt.at(JSON_MTMD_DATA_KEY)) {
-                files.push_back(base64_decode(entry));
+                files.push_back(base64_decode(entry.get<std::string>()));
             }
-            return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY), files);
+            return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY).get<std::string>(), files);
         } else {
             // Not multimodal, but contains a subobject.
             llama_tokens tmp = tokenize_mixed(vocab, json_prompt.at(JSON_STRING_PROMPT_KEY), add_special, parse_special);
@@ -1258,8 +1261,8 @@ json oaicompat_chat_params_parse(
     auto caps = common_chat_templates_get_caps(opt.tmpls.get());
 
     common_chat_templates_inputs inputs;
-    inputs.messages               = common_chat_msgs_parse_oaicompat(messages);
-    inputs.tools                  = common_chat_tools_parse_oaicompat(tools);
+    inputs.messages               = common_chat_msgs_parse_oaicompat(common_json_raw<nlohmann::ordered_json>(messages));
+    inputs.tools                  = common_chat_tools_parse_oaicompat(common_json_raw<nlohmann::ordered_json>(tools));
     inputs.tool_choice            = common_chat_tool_choice_parse_oaicompat(tool_choice);
     inputs.json_schema            = json_schema.is_null() ? "" : json_schema.dump();
     inputs.grammar                = grammar;
@@ -1267,7 +1270,7 @@ json oaicompat_chat_params_parse(
     inputs.parallel_tool_calls    = json_value(body, "parallel_tool_calls", caps["supports_parallel_tool_calls"]);
     inputs.add_generation_prompt  = json_value(body, "add_generation_prompt", true);
     inputs.continue_final_message = body.contains("continue_final_message") ?
-        common_chat_continuation_parse(body.at("continue_final_message")) :
+        common_chat_continuation_parse(common_json_raw<nlohmann::ordered_json>(body.at("continue_final_message"))) :
         COMMON_CHAT_CONTINUATION_NONE;
     if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_NONE && opt.prefill_assistant
         && !inputs.messages.empty() && inputs.messages.back().role == "assistant") {
@@ -1300,7 +1303,8 @@ json oaicompat_chat_params_parse(
     }
 
     // parse the "enable_thinking" kwarg to override the default value
-    auto enable_thinking_kwarg = json_value(inputs.chat_template_kwargs, "enable_thinking", std::string(""));
+    const auto   kwarg_it            = inputs.chat_template_kwargs.find("enable_thinking");
+    std::string  enable_thinking_kwarg = kwarg_it == inputs.chat_template_kwargs.end() ? "" : kwarg_it->second;
     if (enable_thinking_kwarg == "true") {
         inputs.enable_thinking = true;
     } else if (enable_thinking_kwarg == "false") {
@@ -1316,7 +1320,7 @@ json oaicompat_chat_params_parse(
             inputs.enable_thinking = false;
             inputs.chat_template_kwargs.erase("reasoning_effort");
         } else if (!reasoning_effort.empty()) {
-            inputs.chat_template_kwargs["reasoning_effort"] = json(reasoning_effort).dump();
+            inputs.chat_template_kwargs["reasoning_effort"] = json::make(reasoning_effort).dump();
         }
     }
 
@@ -1347,7 +1351,7 @@ json oaicompat_chat_params_parse(
         llama_params["chat_parser"] = chat_params.parser;
     }
 
-    llama_params["message_delimiters"] = chat_params.message_delimiters.to_json();
+    llama_params["message_delimiters"] = common_json_from_raw(chat_params.message_delimiters.to_json());
 
     // Reasoning budget: pass parameters through to sampling layer
     {
@@ -1465,7 +1469,10 @@ json format_response_rerank(
     });
 
     elements.resize(std::min(top_n, (int)elements.size()));
-    json results = elements;
+    json results = json::array();
+    for (const auto & el : elements) {
+        results.push_back(el);
+    }
 
     if (is_tei_format) return results;
 
@@ -1540,7 +1547,7 @@ 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
@@ -1790,12 +1797,12 @@ server_tokens format_prompt_rerank(
         std::string prompt = rerank_prompt;
         string_replace_all(prompt, "{query}"   , query);
         string_replace_all(prompt, "{document}", doc  );
-        server_tokens tokens = tokenize_input_subprompt(vocab, mctx, prompt, false, true);
+        server_tokens tokens = tokenize_input_subprompt(vocab, mctx, json::make(prompt), false, true);
         result.push_back(tokens);
     } else {
         // Get EOS token - use SEP token as fallback if EOS is not available
-        server_tokens query_tokens = tokenize_input_subprompt(vocab, mctx, query, false, false);
-        server_tokens doc_tokens   = tokenize_input_subprompt(vocab, mctx, doc,   false, false);
+        server_tokens query_tokens = tokenize_input_subprompt(vocab, mctx, json::make(query), false, false);
+        server_tokens doc_tokens   = tokenize_input_subprompt(vocab, mctx, json::make(doc),   false, false);
         llama_token eos_token = llama_vocab_eos(vocab);
         if (eos_token == LLAMA_TOKEN_NULL) {
             eos_token = llama_vocab_sep(vocab);

tools/server/server-common.h

@@ -7,7 +7,7 @@
 #include "mtmd.h"
 
 #define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <atomic>
 #include <chrono>
@@ -19,7 +19,7 @@
 #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 +42,9 @@ 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 std::exception & 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.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-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,