Update llama.cpp to v0.3.0 and fit an optional extra model

llama-cpp-bindings-sys/wrapper_chat_apply.cpp

@@ -1,17 +1,16 @@
 #include "wrapper_chat_apply.h"
-#include "nlohmann/json_fwd.hpp"
 #include "wrapper_token_text.h"
 
 #include "llama.cpp/common/chat-auto-parser.h"
 #include "llama.cpp/common/chat.h"
+#include "llama.cpp/common/json.h"
 #include "llama.cpp/include/llama.h"
 #include "wrapper_utils.h"
 
 #include <cstddef>
 #include <exception>
 #include <gsl/span>
 #include <new>
-#include <nlohmann/json.hpp>
 #include <string>
 #include <utility>
 
@@ -60,7 +59,7 @@ extern "C" auto llama_rs_apply_chat_template(
 
         common_chat_template const tmpl(template_src, bos_token, eos_token);
 
-        nlohmann::ordered_json messages = nlohmann::ordered_json::array();
+        common_json messages = common_json::array();
         const gsl::span<const char * const> role_span(roles, n_messages);
         const gsl::span<const char * const> content_span(contents, n_messages);
         for (size_t index = 0; index < n_messages; index++) {
@@ -72,7 +71,7 @@ extern "C" auto llama_rs_apply_chat_template(
 
         autoparser::generation_params inputs;
         inputs.messages              = std::move(messages);
-        inputs.tools                 = nlohmann::ordered_json::array();
+        inputs.tools                 = common_json::array();
         inputs.add_generation_prompt = add_generation_prompt != 0;
         inputs.enable_thinking       = enable_thinking != 0;
 

llama-cpp-bindings-sys/wrapper_chat_parse.cpp

@@ -1,11 +1,10 @@
 #include "wrapper_chat_parse.h"
-#include <nlohmann/json.hpp> // IWYU pragma: keep
-#include <nlohmann/json_fwd.hpp>
 #include "peg-parser.h"
 #include "wrapper_token_text.h"
 
 #include "llama.cpp/common/chat-auto-parser.h"
 #include "llama.cpp/common/chat.h"
+#include "llama.cpp/common/json.h"
 #include "llama.cpp/include/llama.h"
 #include "wrapper_utils.h"
 
@@ -152,9 +151,9 @@ extern "C" auto llama_rs_parse_chat_message(
         autoparser::generation_params inputs;
 
         if ((tools_json != nullptr) && *tools_json != '\0') {
-            inputs.tools = nlohmann::ordered_json::parse(tools_json);
+            inputs.tools = common_json::parse(tools_json);
         } else {
-            inputs.tools = nlohmann::ordered_json::array();
+            inputs.tools = common_json::array();
         }
 
         common_peg_arena const chat_parser = parser->parser.build_parser(inputs, std::string());

llama-cpp-bindings-sys/wrapper_common.cpp

@@ -11,9 +11,8 @@
 #include <string>
 
 #include "llama.cpp/common/json-schema-to-grammar.h"
+#include "llama.cpp/common/json.h"
 #include "llama.cpp/include/llama.h"
-#include <nlohmann/json.hpp> // IWYU pragma: keep
-#include <nlohmann/json_fwd.hpp>
 #include "wrapper_utils.h"
 
 extern "C" auto llama_rs_json_schema_to_grammar(
@@ -38,7 +37,7 @@ extern "C" auto llama_rs_json_schema_to_grammar(
     }
 
     try {
-        const auto schema = nlohmann::ordered_json::parse(schema_json);
+        const auto schema = common_json::parse(schema_json);
         const auto grammar = json_schema_to_grammar(schema, force_gbnf);
         *out_grammar = llama_rs_dup_string(grammar);
         if (*out_grammar == nullptr) {

llama-cpp-bindings-sys/wrapper_fit.cpp

@@ -18,6 +18,7 @@ extern "C" auto llama_rs_fit_params(
     struct llama_model_tensor_buft_override * tensor_buft_overrides,
     size_t * margins,
     uint32_t n_ctx_min,
+    const llama_rs_fit_extra_model * extra,
     enum ggml_log_level log_level,
     int32_t * out_unrecognized_status_code,
     char ** out_error) -> llama_rs_fit_params_status {
@@ -42,11 +43,30 @@ extern "C" auto llama_rs_fit_params(
     if (out_error == nullptr) {
         return LLAMA_RS_FIT_PARAMS_NULL_OUT_ERROR_ARG;
     }
+    if (extra != nullptr) {
+        if (extra->path_model == nullptr) {
+            return LLAMA_RS_FIT_PARAMS_NULL_EXTRA_PATH_MODEL_ARG;
+        }
+        if (extra->mparams == nullptr) {
+            return LLAMA_RS_FIT_PARAMS_NULL_EXTRA_MPARAMS_ARG;
+        }
+        if (extra->cparams == nullptr) {
+            return LLAMA_RS_FIT_PARAMS_NULL_EXTRA_CPARAMS_ARG;
+        }
+    }
 
     try {
+        common_fit_extra_model vendored_extra{};
+        if (extra != nullptr) {
+            vendored_extra.path_model   = extra->path_model;
+            vendored_extra.mparams      = extra->mparams;
+            vendored_extra.cparams      = extra->cparams;
+            vendored_extra.shares_model = extra->shares_model;
+        }
+
         const common_params_fit_status status = common_fit_params(
             path_model, mparams, cparams, tensor_split, tensor_buft_overrides,
-            margins, n_ctx_min, log_level);
+            margins, n_ctx_min, extra == nullptr ? nullptr : &vendored_extra, log_level);
         switch (status) {
             case COMMON_PARAMS_FIT_STATUS_SUCCESS:
                 return LLAMA_RS_FIT_PARAMS_OK;

llama-cpp-bindings-sys/wrapper_fit.h

@@ -3,6 +3,7 @@
 #include "llama.cpp/ggml/include/ggml.h"
 #include "llama.cpp/include/llama.h"
 
+#include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
 
@@ -23,8 +24,18 @@ typedef enum llama_rs_fit_params_status {
     LLAMA_RS_FIT_PARAMS_ERROR_STRING_ALLOCATION_FAILED,
     LLAMA_RS_FIT_PARAMS_VENDORED_OUT_OF_MEMORY,
     LLAMA_RS_FIT_PARAMS_VENDORED_THREW_CXX_EXCEPTION,
+    LLAMA_RS_FIT_PARAMS_NULL_EXTRA_PATH_MODEL_ARG,
+    LLAMA_RS_FIT_PARAMS_NULL_EXTRA_MPARAMS_ARG,
+    LLAMA_RS_FIT_PARAMS_NULL_EXTRA_CPARAMS_ARG,
 } llama_rs_fit_params_status;
 
+typedef struct llama_rs_fit_extra_model {
+    const char * path_model;
+    struct llama_model_params * mparams;
+    struct llama_context_params * cparams;
+    bool shares_model;
+} llama_rs_fit_extra_model;
+
 llama_rs_fit_params_status llama_rs_fit_params(
     const char * path_model,
     struct llama_model_params * mparams,
@@ -33,6 +44,7 @@ llama_rs_fit_params_status llama_rs_fit_params(
     struct llama_model_tensor_buft_override * tensor_buft_overrides,
     size_t * margins,
     uint32_t n_ctx_min,
+    const llama_rs_fit_extra_model * extra,
     enum ggml_log_level log_level,
     int32_t * out_unrecognized_status_code,
     char ** out_error);

llama-cpp-bindings-sys/wrapper_reasoning.cpp

@@ -3,9 +3,8 @@
 
 #include "llama.cpp/common/chat-auto-parser.h"
 #include "llama.cpp/common/chat.h"
+#include "llama.cpp/common/json.h"
 #include "llama.cpp/include/llama.h"
-#include <nlohmann/json.hpp> // IWYU pragma: keep
-#include <nlohmann/json_fwd.hpp>
 #include "wrapper_utils.h"
 
 #include <cstddef>
@@ -36,8 +35,8 @@ auto find_reasoning_markers(
     probe_params.is_inference = false;
     probe_params.add_inference = false;
     probe_params.mark_input = false;
-    probe_params.messages = nlohmann::ordered_json::array({
-        nlohmann::ordered_json{ { "role", "user" }, { "content", "ping" } },
+    probe_params.messages = common_json::array({
+        common_json{ { "role", "user" }, { "content", "ping" } },
     });
 
     const std::string tmpl_src_str = tmpl_src;

llama-cpp-bindings-sys/wrapper_tool_calls.cpp

@@ -1,11 +1,10 @@
 #include "wrapper_tool_calls.h"
-#include <nlohmann/json.hpp> // IWYU pragma: keep
-#include <nlohmann/json_fwd.hpp>
 #include "wrapper_token_text.h"
 
 #include "llama.cpp/common/chat-auto-parser.h"
 #include "llama.cpp/common/chat-auto-parser-helpers.h"
 #include "llama.cpp/common/chat.h"
+#include "llama.cpp/common/json.h"
 #include "llama.cpp/include/llama.h"
 #include "wrapper_utils.h"
 
@@ -18,18 +17,21 @@ using wrapper_helpers::token_text_or_empty;
 
 namespace {
 
-auto detect_tool_call_haystack(
-    const common_chat_template & tmpl,
-    const autoparser::analyze_reasoning & reasoning) -> std::string {
-    nlohmann::ordered_json const user_msg = {
+struct tool_call_probe_params {
+    template_params without_tool_calls;
+    template_params with_tool_calls;
+};
+
+auto build_tool_call_probe_params() -> tool_call_probe_params {
+    common_json const user_msg = {
         { "role",    "user"                },
         { "content", "Please use the tool" }
     };
-    nlohmann::ordered_json const assistant_no_tools = {
+    common_json const assistant_no_tools = {
         { "role",    "assistant"      },
         { "content", "Sure, calling." }
     };
-    nlohmann::ordered_json const first_tool_call = {
+    common_json const first_tool_call = {
         { "id",       "call_001"  },
         { "type",     "function"  },
         { "function", {
@@ -40,12 +42,12 @@ auto detect_tool_call_haystack(
             }}
         }}
     };
-    nlohmann::ordered_json const assistant_with_tools = {
-        { "role",       "assistant"                                                  },
-        { "content",    ""                                                           },
-        { "tool_calls", nlohmann::ordered_json::array({ first_tool_call })           }
+    common_json const assistant_with_tools = {
+        { "role",       "assistant"                              },
+        { "content",    ""                                       },
+        { "tool_calls", common_json::array({ first_tool_call })  }
     };
-    nlohmann::ordered_json const tool_definition = {
+    common_json const tool_definition = {
         { "type",     "function"  },
         { "function", {
             { "name",        "tool_first"           },
@@ -56,23 +58,30 @@ auto detect_tool_call_haystack(
                     { "arg_first", { { "type", "string" }, { "description", "first arg" } } },
                     { "arg_second", { { "type", "string" }, { "description", "second arg" } } },
                 }},
-                { "required", nlohmann::ordered_json::array({ "arg_first", "arg_second" }) },
+                { "required", common_json::array({ "arg_first", "arg_second" }) },
             }}
         }}
     };
 
-    template_params params_no_tools;
-    params_no_tools.messages              = nlohmann::ordered_json::array({ user_msg, assistant_no_tools });
-    params_no_tools.tools                 = nlohmann::ordered_json::array({ tool_definition });
-    params_no_tools.add_generation_prompt = false;
-    params_no_tools.enable_thinking       = true;
+    tool_call_probe_params probe;
+    probe.without_tool_calls.messages              = common_json::array({ user_msg, assistant_no_tools });
+    probe.without_tool_calls.tools                 = common_json::array({ tool_definition });
+    probe.without_tool_calls.add_generation_prompt = false;
+    probe.without_tool_calls.enable_thinking       = true;
+
+    probe.with_tool_calls          = probe.without_tool_calls;
+    probe.with_tool_calls.messages = common_json::array({ user_msg, assistant_with_tools });
+
+    return probe;
+}
 
-    template_params params_with_tools = params_no_tools;
-    params_with_tools.messages =
-        nlohmann::ordered_json::array({ user_msg, assistant_with_tools });
+auto detect_tool_call_haystack(
+    const common_chat_template & tmpl,
+    const autoparser::analyze_reasoning & reasoning) -> std::string {
+    tool_call_probe_params const probe = build_tool_call_probe_params();
 
-    std::string const output_no_tools = autoparser::apply_template(tmpl, params_no_tools);
-    std::string const output_with_tools = autoparser::apply_template(tmpl, params_with_tools);
+    std::string const output_no_tools = autoparser::apply_template(tmpl, probe.without_tool_calls);
+    std::string const output_with_tools = autoparser::apply_template(tmpl, probe.with_tool_calls);
 
     if (output_no_tools.empty() || output_with_tools.empty()) {
         return {};
@@ -210,58 +219,10 @@ extern "C" auto llama_rs_diagnose_tool_call_synthetic_renders(
 
         common_chat_template const tmpl(tmpl_src, bos_token, eos_token);
 
-        nlohmann::ordered_json const user_msg = {
-            { "role",    "user"                },
-            { "content", "Please use the tool" }
-        };
-        nlohmann::ordered_json const assistant_no_tools = {
-            { "role",    "assistant"      },
-            { "content", "Sure, calling." }
-        };
-        nlohmann::ordered_json const first_tool_call = {
-            { "id",       "call_001"  },
-            { "type",     "function"  },
-            { "function", {
-                { "name",      "tool_first" },
-                { "arguments", {
-                    { "arg_first", "XXXX" },
-                    { "arg_second", "YYYY" },
-                }}
-            }}
-        };
-        nlohmann::ordered_json const assistant_with_tools = {
-            { "role",       "assistant"                                                  },
-            { "content",    ""                                                           },
-            { "tool_calls", nlohmann::ordered_json::array({ first_tool_call })           }
-        };
-        nlohmann::ordered_json const tool_definition = {
-            { "type",     "function"  },
-            { "function", {
-                { "name",        "tool_first"           },
-                { "description", "First test tool"      },
-                { "parameters", {
-                    { "type", "object" },
-                    { "properties", {
-                        { "arg_first", { { "type", "string" }, { "description", "first arg" } } },
-                        { "arg_second", { { "type", "string" }, { "description", "second arg" } } },
-                    }},
-                    { "required", nlohmann::ordered_json::array({ "arg_first", "arg_second" }) },
-                }}
-            }}
-        };
-
-        template_params params_no_tools;
-        params_no_tools.messages              = nlohmann::ordered_json::array({ user_msg, assistant_no_tools });
-        params_no_tools.tools                 = nlohmann::ordered_json::array({ tool_definition });
-        params_no_tools.add_generation_prompt = false;
-        params_no_tools.enable_thinking       = true;
-
-        template_params params_with_tools = params_no_tools;
-        params_with_tools.messages =
-            nlohmann::ordered_json::array({ user_msg, assistant_with_tools });
+        tool_call_probe_params const probe = build_tool_call_probe_params();
 
-        std::string const output_a = autoparser::apply_template(tmpl, params_no_tools);
-        std::string const output_b = autoparser::apply_template(tmpl, params_with_tools);
+        std::string const output_a = autoparser::apply_template(tmpl, probe.without_tool_calls);
+        std::string const output_b = autoparser::apply_template(tmpl, probe.with_tool_calls);
 
         std::unique_ptr<char[]> a_dup(llama_rs_dup_string(output_a));
         std::unique_ptr<char[]> b_dup(llama_rs_dup_string(output_b));