@fugood/llama.node 0.3.11 → 0.3.13

package/src/llama.cpp/examples/server/server.cpp

@@ -42,7 +42,7 @@ enum stop_type {
     STOP_TYPE_LIMIT,
 };
 
-// state diagram: https://github.com/ggerganov/llama.cpp/pull/9283
+// state diagram: https://github.com/ggml-org/llama.cpp/pull/9283
 enum slot_state {
     SLOT_STATE_IDLE,
     SLOT_STATE_STARTED, // TODO: this state is only used for setting up the initial prompt processing; maybe merge it with launch_slot_with_task in the future
@@ -173,6 +173,7 @@ struct slot_params {
             {"grammar_trigger_words",     grammar_trigger_words},
             {"grammar_trigger_tokens",    sampling.grammar_trigger_tokens},
             {"preserved_tokens",          sampling.preserved_tokens},
+            {"chat_format",               common_chat_format_name(oaicompat_chat_format)},
             {"samplers",                  samplers},
             {"speculative.n_max",         speculative.n_max},
             {"speculative.n_min",         speculative.n_min},
@@ -334,24 +335,24 @@ struct server_task {
         if (data.contains("json_schema") && !data.contains("grammar")) {
             try {
                 auto schema                  = json_value(data, "json_schema", json::object());
-                LOG_DBG("JSON schema: %s\n", schema.dump(2).c_str());
+                SRV_DBG("JSON schema: %s\n", schema.dump(2).c_str());
                 params.sampling.grammar      = json_schema_to_grammar(schema);
-                LOG_DBG("Converted grammar: %s\n", params.sampling.grammar.c_str());
+                SRV_DBG("Converted grammar: %s\n", params.sampling.grammar.c_str());
             } catch (const std::exception & e) {
                 throw std::runtime_error(std::string("\"json_schema\": ") + e.what());
             }
         } else {
             params.sampling.grammar      = json_value(data, "grammar", defaults.sampling.grammar);
-            LOG_DBG("Grammar: %s\n", params.sampling.grammar.c_str());
+            SRV_DBG("Grammar: %s\n", params.sampling.grammar.c_str());
             params.sampling.grammar_lazy = json_value(data, "grammar_lazy", defaults.sampling.grammar_lazy);
-            LOG_DBG("Grammar lazy: %s\n", params.sampling.grammar_lazy ? "true" : "false");
+            SRV_DBG("Grammar lazy: %s\n", params.sampling.grammar_lazy ? "true" : "false");
         }
 
         {
             auto it = data.find("chat_format");
             if (it != data.end()) {
                 params.oaicompat_chat_format = static_cast<common_chat_format>(it->get<int>());
-                LOG_INF("Chat format: %s\n", common_chat_format_name(params.oaicompat_chat_format).c_str());
+                SRV_INF("Chat format: %s\n", common_chat_format_name(params.oaicompat_chat_format).c_str());
             } else {
                 params.oaicompat_chat_format = defaults.oaicompat_chat_format;
             }
@@ -367,12 +368,12 @@ struct server_task {
 
                     auto ids = common_tokenize(vocab, trigger.word, /* add_special= */ false, /* parse_special= */ true);
                     if (ids.size() == 1) {
-                        LOG_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str());
+                        SRV_DBG("Grammar trigger token: %d (`%s`)\n", ids[0], trigger.word.c_str());
                         params.sampling.grammar_trigger_tokens.push_back(ids[0]);
                         params.sampling.preserved_tokens.insert(ids[0]);
                         continue;
                     }
-                    LOG_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str());
+                    SRV_DBG("Grammar trigger word: `%s`\n", trigger.word.c_str());
                     params.sampling.grammar_trigger_words.push_back(trigger);
                 }
             }
@@ -381,11 +382,11 @@ struct server_task {
                 for (const auto & t : *preserved_tokens) {
                     auto ids = common_tokenize(vocab, t.get<std::string>(), /* add_special= */ false, /* parse_special= */ true);
                     if (ids.size() == 1) {
-                        LOG_DBG("Preserved token: %d\n", ids[0]);
+                        SRV_DBG("Preserved token: %d\n", ids[0]);
                         params.sampling.preserved_tokens.insert(ids[0]);
                     } else {
                         // This may happen when using a tool call style meant for a model with special tokens to preserve on a model without said tokens.
-                        LOG_WRN("Not preserved because more than 1 token (wrong chat template override?): %s\n", t.get<std::string>().c_str());
+                        SRV_WRN("Not preserved because more than 1 token (wrong chat template override?): %s\n", t.get<std::string>().c_str());
                     }
                 }
             }
@@ -717,16 +718,26 @@ struct server_task_result_cmpl_final : server_task_result {
         std::string finish_reason = "length";
         common_chat_msg msg;
         if (stop == STOP_TYPE_WORD || stop == STOP_TYPE_EOS) {
-            LOG_DBG("Parsing chat message: %s\n", content.c_str());
+            SRV_DBG("Parsing chat message: %s\n", content.c_str());
             msg = common_chat_parse(content, oaicompat_chat_format);
             finish_reason = msg.tool_calls.empty() ? "stop" : "tool_calls";
         } else {
             msg.content = content;
         }
 
-        json tool_calls;
+        json message {
+            {"role", "assistant"},
+        };
+        if (!msg.reasoning_content.empty()) {
+            message["reasoning_content"] = msg.reasoning_content;
+        }
+        if (msg.content.empty() && !msg.tool_calls.empty()) {
+            message["content"] = json();
+        } else {
+            message["content"] = msg.content;
+        }
         if (!msg.tool_calls.empty()) {
-            tool_calls = json::array();
+            auto tool_calls = json::array();
             for (const auto & tc : msg.tool_calls) {
                 tool_calls.push_back({
                     {"type", "function"},
@@ -737,15 +748,7 @@ struct server_task_result_cmpl_final : server_task_result {
                     {"id", tc.id},
                 });
             }
-        }
-
-        json message {
-            {"content", msg.content},
-            {"tool_calls", tool_calls},
-            {"role", "assistant"},
-        };
-        if (!msg.tool_plan.empty()) {
-            message["tool_plan"] = msg.tool_plan;
+            message["tool_calls"] = tool_calls;
         }
 
         json choice {
@@ -1600,6 +1603,10 @@ struct server_queue {
 
             while (true) {
                 std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (!running) {
+                    QUE_DBG("%s", "terminate\n");
+                    return;
+                }
                 if (queue_tasks.empty()) {
                     lock.unlock();
                     break;
@@ -1620,11 +1627,11 @@ struct server_queue {
             QUE_DBG("%s", "waiting for new tasks\n");
             {
                 std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (!running) {
+                    QUE_DBG("%s", "terminate\n");
+                    return;
+                }
                 if (queue_tasks.empty()) {
-                    if (!running) {
-                        QUE_DBG("%s", "terminate\n");
-                        return;
-                    }
                     condition_tasks.wait(lock, [&]{
                         return (!queue_tasks.empty() || !running);
                     });
@@ -1885,7 +1892,7 @@ struct server_context {
         }
 
         if (params_base.chat_template.empty() && !validate_builtin_chat_template(params.use_jinja)) {
-            LOG_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
+            SRV_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
             chat_templates = common_chat_templates_from_model(model, "chatml");
         } else {
             chat_templates = common_chat_templates_from_model(model, params_base.chat_template);
@@ -2069,8 +2076,8 @@ struct server_context {
 
         if (slot.n_predict > 0 && slot.params.n_predict > slot.n_predict) {
             // Might be better to reject the request with a 400 ?
+            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.params.n_predict, slot.n_predict);
             slot.params.n_predict = slot.n_predict;
-            SLT_WRN(slot, "n_predict = %d exceeds server configuration, setting to %d", slot.n_predict, slot.n_predict);
         }
 
         if (slot.params.ignore_eos && has_eos_token) {
@@ -2275,7 +2282,7 @@ struct server_context {
             for (size_t i = 0; i < std::min(max_probs, n_probs); i++) {
                 result.probs.push_back({
                     cur_p->data[i].id,
-                    common_detokenize(ctx, {cur_p->data[i].id}, special),
+                    common_token_to_piece(ctx, cur_p->data[i].id, special),
                     cur_p->data[i].p
                 });
             }
@@ -2297,7 +2304,7 @@ struct server_context {
             for (size_t i = 0; i < std::min(n_vocab, n_probs); i++) {
                 result.probs.push_back({
                     cur[i].id,
-                    common_detokenize(ctx, {cur[i].id}, special),
+                    common_token_to_piece(ctx, cur[i].id, special),
                     cur[i].p
                 });
             }
@@ -3355,10 +3362,10 @@ static void log_server_request(const httplib::Request & req, const httplib::Resp
 
     // reminder: this function is not covered by httplib's exception handler; if someone does more complicated stuff, think about wrapping it in try-catch
 
-    LOG_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
+    SRV_INF("request: %s %s %s %d\n", req.method.c_str(), req.path.c_str(), req.remote_addr.c_str(), res.status);
 
-    LOG_DBG("request:  %s\n", req.body.c_str());
-    LOG_DBG("response: %s\n", res.body.c_str());
+    SRV_DBG("request:  %s\n", req.body.c_str());
+    SRV_DBG("response: %s\n", res.body.c_str());
 }
 
 std::function<void(int)> shutdown_handler;
@@ -3649,7 +3656,7 @@ int main(int argc, char ** argv) {
             }, {
                     {"name",  "n_busy_slots_per_decode"},
                     {"help",  "Average number of busy slots per llama_decode() call"},
-                    {"value",  (float) res_metrics->n_busy_slots_total / (float) res_metrics->n_decode_total}
+                    {"value",  (float) res_metrics->n_busy_slots_total / std::max((float) res_metrics->n_decode_total, 1.f)}
             }}},
             {"gauge", {{
                     {"name",  "prompt_tokens_seconds"},
@@ -3860,7 +3867,9 @@ int main(int argc, char ** argv) {
 
         try {
             const auto & prompt = data.at("prompt");
-            LOG_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
+            // TODO: this log can become very long, put it behind a flag or think about a more compact format
+            //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
+
             std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
             tasks.reserve(tokenized_prompts.size());
             for (size_t i = 0; i < tokenized_prompts.size(); i++) {
@@ -4054,7 +4063,7 @@ int main(int argc, char ** argv) {
         }
 
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
+        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
 
         return handle_completions_impl(
             SERVER_TASK_TYPE_COMPLETION,
@@ -4067,7 +4076,7 @@ int main(int argc, char ** argv) {
     // same with handle_chat_completions, but without inference part
     const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
         auto body = json::parse(req.body);
-        json data = oaicompat_completion_params_parse(body, params.use_jinja, ctx_server.chat_templates);
+        json data = oaicompat_completion_params_parse(body, params.use_jinja, params.reasoning_format, ctx_server.chat_templates);
         res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
     };
 
@@ -4376,6 +4385,9 @@ int main(int argc, char ** argv) {
                     res.set_content("Error: gzip is not supported by this browser", "text/plain");
                 } else {
                     res.set_header("Content-Encoding", "gzip");
+                    // COEP and COOP headers, required by pyodide (python interpreter)
+                    res.set_header("Cross-Origin-Embedder-Policy", "require-corp");
+                    res.set_header("Cross-Origin-Opener-Policy", "same-origin");
                     res.set_content(reinterpret_cast<const char*>(index_html_gz), index_html_gz_len, "text/html; charset=utf-8");
                 }
                 return false;
@@ -4425,6 +4437,7 @@ int main(int argc, char ** argv) {
 
     // clean up function, to be called before exit
     auto clean_up = [&svr]() {
+        SRV_INF("%s: cleaning up before exit...\n", __func__);
         svr->stop();
         llama_backend_free();
     };
@@ -4441,10 +4454,6 @@ int main(int argc, char ** argv) {
     }
 
     if (!was_bound) {
-        //LOG_ERROR("couldn't bind HTTP server socket", {
-        //    {"hostname", params.hostname},
-        //    {"port", params.port},
-        //});
         LOG_ERR("%s: couldn't bind HTTP server socket, hostname: %s, port: %d\n", __func__, params.hostname.c_str(), params.port);
         clean_up();
         return 1;
@@ -4461,7 +4470,7 @@ int main(int argc, char ** argv) {
 
     if (!ctx_server.load_model(params)) {
         clean_up();
-        t.join();
+        // t.join(); // FIXME: see below
         LOG_ERR("%s: exiting due to model loading error\n", __func__);
         return 1;
     }
@@ -4485,13 +4494,10 @@ int main(int argc, char ** argv) {
     });
 
     shutdown_handler = [&](int) {
+        // this will unblock start_loop()
         ctx_server.queue_tasks.terminate();
     };
 
-    LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
-
-    ctx_server.queue_tasks.start_loop();
-
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
     struct sigaction sigint_action;
     sigint_action.sa_handler = signal_handler;
@@ -4506,8 +4512,13 @@ int main(int argc, char ** argv) {
     SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
+    LOG_INF("%s: server is listening on http://%s:%d - starting the main loop\n", __func__, params.hostname.c_str(), params.port);
+
+    // this call blocks the main thread until queue_tasks.terminate() is called
+    ctx_server.queue_tasks.start_loop();
+
     clean_up();
-    t.join();
+    // t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
 
     return 0;
 }

package/src/llama.cpp/examples/server/utils.hpp

@@ -367,10 +367,10 @@ inline std::string format_chat(const common_chat_template & tmpl, const std::vec
                     }
                 }
             } else {
-                throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+                throw std::runtime_error("Invalid 'content' type (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
             }
         } else {
-            throw std::runtime_error("Missing 'content' (ref: https://github.com/ggerganov/llama.cpp/issues/8367)");
+            throw std::runtime_error("Missing 'content' (ref: https://github.com/ggml-org/llama.cpp/issues/8367)");
         }
 
         chat.push_back({role, content, /* tool_calls= */ {}});
@@ -578,6 +578,7 @@ static json oaicompat_completion_params_parse(const json & body) {
 static json oaicompat_completion_params_parse(
     const json & body, /* openai api json semantics */
     bool use_jinja,
+    common_reasoning_format reasoning_format,
     const common_chat_templates & chat_templates)
 {
     json llama_params;
@@ -633,9 +634,10 @@ static json oaicompat_completion_params_parse(
             throw std::runtime_error("Cannot use custom grammar constraints with tools.");
         }
         common_chat_inputs inputs;
-        inputs.messages = body.at("messages");
-        inputs.tools = tools;
-        inputs.tool_choice = tool_choice;
+        inputs.extract_reasoning   = reasoning_format != COMMON_REASONING_FORMAT_NONE;
+        inputs.messages            = body.at("messages");
+        inputs.tools               = tools;
+        inputs.tool_choice         = tool_choice;
         inputs.parallel_tool_calls = json_value(body, "parallel_tool_calls", false);
         if (inputs.parallel_tool_calls && !tmpl.original_caps().supports_parallel_tool_calls) {
             LOG_DBG("Disabling parallel_tool_calls because the template does not support it\n");

package/src/llama.cpp/ggml/include/ggml-cpu.h

@@ -8,7 +8,7 @@ extern "C" {
 #endif
 
     // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
+    // since https://github.com/ggml-org/ggml/issues/287
     struct ggml_cplan {
         size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`

package/src/llama.cpp/ggml/include/ggml-metal.h

@@ -45,7 +45,7 @@ GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_DEPRECATED(
         GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
-        "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
+        "obsoleted by the new device interface - https://github.com/ggml-org/llama.cpp/pull/9713");
 
 GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 

package/src/llama.cpp/ggml/include/ggml-vulkan.h

@@ -10,8 +10,6 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
-GGML_BACKEND_API void ggml_vk_instance_init(void);
-
 // backend API
 GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 

package/src/llama.cpp/ggml/include/ggml.h

@@ -198,7 +198,7 @@
 
 #ifndef __GNUC__
 #    define GGML_ATTRIBUTE_FORMAT(...)
-#elif defined(__MINGW32__)
+#elif defined(__MINGW32__) && !defined(__clang__)
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
 #else
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))

package/src/llama.cpp/ggml/src/ggml-common.h

@@ -473,7 +473,6 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
-//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
@@ -508,7 +507,6 @@ GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
     0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
 GGML_TABLE_END()
-//#endif
 
 
 GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h

@@ -360,21 +360,15 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #endif
 
 #if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
 /* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+static __m128 __lsx_vreplfr2vr_s(const float val) {
+    v4f32 res = {val, val, val, val};
+    return (__m128)res;
 }
 
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+static __m256 __lasx_xvreplfr2vr_s(const float val) {
+    v8f32 res = {val, val, val, val, val, val, val, val};
+    return (__m256)res;
 }
 #endif